diff --git "a/build/webgpu/test.json" "b/build/webgpu/test.json"
--- "a/build/webgpu/test.json"
+++ "b/build/webgpu/test.json"
@@ -959,7 +959,7 @@
     {
       "name": "relu_3x3_dense_c256m256_8x8",
       "provenance": {
-        "notes": "A compact 3x3 convolution with 256 input and output channels preserves the wide im2col reduction and output-channel dimensions of the subgroup-matrix fallback regime."
+        "notes": "A 3x3 Relu-activated convolution with 256 input and 256 output channels over an 8x8 spatial output checks a 2304-element reduction."
       },
       "attrs": { "activation": "Relu", "strides": [1, 1], "pads": [1, 1, 1, 1] },
       "tolerance": 0.001,
@@ -1016,7 +1016,7 @@
     {
       "name": "same_upper_stride2_autopad",
       "provenance": {
-        "source": "ONNX Runtime com.microsoft.FusedConv CPUExecutionProvider",
+        "source": "ONNX Runtime's CPU provider (com.microsoft.FusedConv)",
         "notes": "FusedConv inherits Conv auto_pad semantics. SAME_UPPER with stride 2 requires trailing padding and a 2x2 output here, exercising padding derived from the official auto_pad and strides attributes."
       },
       "attrs": { "auto_pad": "SAME_UPPER", "activation": "Relu", "strides": [2, 2] },
@@ -1040,7 +1040,7 @@
     {
       "name": "conv1d_relu_direct",
       "provenance": {
-        "source": "ONNX Runtime com.microsoft.FusedConv (CPUExecutionProvider)",
+        "source": "ONNX Runtime's CPU provider (com.microsoft.FusedConv)",
         "notes": "A compact rank-3 NCW case accepted by ORT evaluates `Relu(Conv1D(X, W))` on the direct one-dimensional path."
       },
       "attrs": { "activation": "Relu" },
@@ -1298,7 +1298,7 @@
     {
       "name": "relu_2x2_im2col_padded_z_n49_c8m32",
       "provenance": {
-        "notes": "outputPixels=49 breaks the %64 N alignment the exact im2col subgroup-matrix path demands, so the zero-padded subgroup-matrix GEMM (N padded 49 -> 64) takes the no-bias z-residual route."
+        "notes": "A 2x2 convolution with 49 output positions, residual input, and fused Relu checks the final partial output tile."
       },
       "attrs": { "activation": "Relu", "strides": [1, 1] },
       "tolerance": 0.002,
@@ -1375,7 +1375,7 @@
     {
       "name": "group2_3x3_direct_bias_z_inchan4_4x4",
       "provenance": {
-        "notes": "Grouped conv carrying both a bias and a z-residual: pins the direct unrolled kernel's full epilogue (bias then residual then activation)."
+        "notes": "Grouped 3x3 convolution with bias and z-residual checks the complete epilogue order: bias, residual, then activation."
       },
       "attrs": { "group": 2, "activation": "Relu", "strides": [1, 1], "pads": [1, 1, 1, 1] },
       "inputs": {
@@ -1405,7 +1405,7 @@
     {
       "name": "same_lower_stride2_autopad",
       "provenance": {
-        "source": "ONNX Runtime com.microsoft.FusedConv CPUExecutionProvider",
+        "source": "ONNX Runtime's CPU provider (com.microsoft.FusedConv)",
         "notes": "ONNX `SAME_LOWER` places an odd padding element at the start of the axis. The nonzero leading pad must be derived from `auto_pad` and strides rather than from explicit pad attributes."
       },
       "attrs": { "auto_pad": "SAME_LOWER", "activation": "Relu", "strides": [2, 2] },
@@ -1487,7 +1487,7 @@
     {
       "name": "implicit_sgmat_identity_tails",
       "provenance": {
-        "notes": "Direct-gather boundary lock: K27 pads to32, M tail where nonmultiple, N4097 ends in a one-column output tile; optional bias and residual precede activation in f32 before the single output cast. Nonzero residual/bias prevents a missing epilogue from matching. Uses the narrowed default spatial floor, without forcing tunables."
+        "notes": "A 27-tap (3x3x3) reduction produces 33 output channels, one more than a multiple of 32, over a 4097-element (17x241) spatial output; no bias, residual or activation is applied."
       },
       "inputs": {
         "x": {
@@ -1508,7 +1508,7 @@
     {
       "name": "implicit_sgmat_relu_bias_direct_tail",
       "provenance": {
-        "notes": "Direct-gather boundary lock: K27 pads to32, M tail where nonmultiple, N4097 ends in a one-column output tile; optional bias and residual precede activation in f32 before the single output cast. Nonzero residual/bias prevents a missing epilogue from matching. Uses the narrowed default spatial floor, without forcing tunables."
+        "notes": "A Relu-activated 3x3 convolution over 3 input channels produces 64 output channels, a multiple of 32, over a 4097-element (17x241) spatial output; a 64-element bias is added before the activation, with no residual."
       },
       "inputs": {
         "x": {
@@ -1535,7 +1535,7 @@
     {
       "name": "implicit_sgmat_tanh_residual_tails",
       "provenance": {
-        "notes": "Direct-gather boundary lock: K27 pads to32, M tail where nonmultiple, N4097 ends in a one-column output tile; optional bias and residual precede activation in f32 before the single output cast. Nonzero residual/bias prevents a missing epilogue from matching. Uses the narrowed default spatial floor, without forcing tunables."
+        "notes": "A Tanh-activated 3x3 convolution over 3 input channels produces 33 output channels, one more than a multiple of 32, over a 4097-element (17x241) spatial output; a matching-shape residual is added before the activation, with no bias."
       },
       "inputs": {
         "x": {
@@ -1562,7 +1562,7 @@
     {
       "name": "implicit_sgmat_leaky_bias_residual_batch_tails",
       "provenance": {
-        "notes": "Direct-gather boundary lock: K27 pads to32, M tail where nonmultiple, N4097 ends in a one-column output tile; optional bias and residual precede activation in f32 before the single output cast. Nonzero residual/bias prevents a missing epilogue from matching. Uses the narrowed default spatial floor, without forcing tunables."
+        "notes": "A LeakyRelu (alpha 0.25) 3x3 convolution over 3 input channels and batch 2 produces 65 output channels, one more than a multiple of 64, over a 4097-element (17x241) spatial output; both a 65-element bias and a matching-shape residual are added before the activation."
       },
       "inputs": {
         "x": {
@@ -1594,7 +1594,7 @@
     {
       "name": "implicit_sgmat_f16_relu_tails",
       "provenance": {
-        "notes": "Direct-gather boundary lock: K27 pads to32, M tail where nonmultiple, N4097 ends in a one-column output tile; optional bias and residual precede activation in f32 before the single output cast. Nonzero residual/bias prevents a missing epilogue from matching. Uses the narrowed default spatial floor, without forcing tunables."
+        "notes": "A float16 Relu-activated 3x3 convolution over 3 input channels produces 33 output channels, one more than a multiple of 32, over a 4097-element (17x241) spatial output; no bias or residual is applied."
       },
       "inputs": {
         "x": {
@@ -1614,7 +1614,7 @@
     {
       "name": "implicit_sgmat_relu_bias_direct_tail_f16",
       "provenance": {
-        "notes": "Float16 bias/residual counterpart for the dtype-specific implicit subgroup-matrix config route. M33 stays within the two-tile f16 gate and exercises guarded output storage."
+        "notes": "A float16 Relu-activated 3x3 convolution over 3 input channels produces 33 output channels, one more than a multiple of 32, over a 4097-element (17x241) spatial output; a 33-element bias is added before the activation, with no residual."
       },
       "inputs": {
         "x": {
@@ -1639,7 +1639,7 @@
     {
       "name": "implicit_sgmat_tanh_residual_tails_f16",
       "provenance": {
-        "notes": "Float16 bias/residual counterpart for the dtype-specific implicit subgroup-matrix config route. M33 stays within the two-tile f16 gate and exercises guarded output storage."
+        "notes": "A float16 Tanh-activated 3x3 convolution over 3 input channels produces 33 output channels, one more than a multiple of 32, over a 4097-element (17x241) spatial output; a matching-shape residual is added before the activation, with no bias."
       },
       "inputs": {
         "x": {
@@ -1664,7 +1664,7 @@
     {
       "name": "implicit_sgmat_leaky_bias_residual_batch_tails_f16",
       "provenance": {
-        "notes": "Float16 bias/residual counterpart for the dtype-specific implicit subgroup-matrix config route. M33 stays within the two-tile f16 gate and exercises guarded output storage."
+        "notes": "A float16 LeakyRelu (alpha 0.25) 3x3 convolution over 3 input channels and batch 2 produces 33 output channels, one more than a multiple of 32, over a 4097-element (17x241) spatial output; both a 33-element bias and a matching-shape residual are added before the activation."
       },
       "inputs": {
         "x": {
@@ -1708,9 +1708,7 @@
       },
       "outputs": { "y": { "dtype": "float32", "shape": [1, 96, 3, 43], "tolerance": 0.0001, "relTolerance": 0.00002 } },
       "tunables": { "IMPLICIT_SGMAT_MAX_M_TILES": 4, "IMPLICIT_SGMAT_MIN_OUTPUT_PIXELS": 128 },
-      "provenance": {
-        "notes": "Compact multi-tile implicit gather with K/N tails. Tunables exercise the kernel without allocating the large memory-boundary workload; selector unit tests separately pin the actual device-limit boundary."
-      }
+      "provenance": { "notes": "Compact convolution with partial reduction and output-channel tiles." }
     },
     {
       "name": "implicit_buffer_fallback_compact_float16",
@@ -1729,9 +1727,7 @@
       },
       "outputs": { "y": { "dtype": "float16", "shape": [1, 96, 3, 43], "tolerance": 0.001, "relTolerance": 0.001 } },
       "tunables": { "IMPLICIT_SGMAT_MAX_M_TILES": 4, "IMPLICIT_SGMAT_MIN_OUTPUT_PIXELS": 128 },
-      "provenance": {
-        "notes": "Compact multi-tile implicit gather with K/N tails. Tunables exercise the kernel without allocating the large memory-boundary workload; selector unit tests separately pin the actual device-limit boundary."
-      }
+      "provenance": { "notes": "Compact convolution with partial reduction and output-channel tiles." }
     },
     {
       "name": "implicit_splitk_preferred_grid16_k144",
@@ -1750,7 +1746,7 @@
       },
       "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 32], "tolerance": 0.0001, "relTolerance": 0.0001 } },
       "provenance": {
-        "notes": "Selects the preferred implicit split-K route at its output-grid boundary; nonuniform data exercises reduction and fused activation."
+        "notes": "Nonuniform 3x3 convolution data at a compact output-grid boundary checks reduction and fused activation."
       }
     },
     {
@@ -1775,13 +1771,13 @@
       },
       "outputs": { "y": { "dtype": "float32", "shape": [1, 65, 32, 32], "tolerance": 0.0001, "relTolerance": 0.0001 } },
       "provenance": {
-        "notes": "Selects the preferred implicit split-K route at its output-grid boundary; nonuniform data exercises reduction and fused activation."
+        "notes": "Nonuniform 3x3 convolution data at a compact output-grid boundary checks reduction and fused activation."
       }
     },
     {
       "name": "direct_matrix_inputs_m32_c32_h8_w8_mode0",
       "provenance": {
-        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. Original portable and staged paths remain the oracle controls."
+        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. The portable and staged paths serve as controls."
       },
       "inputs": {
         "x": {
@@ -1807,7 +1803,7 @@
     {
       "name": "direct_matrix_inputs_m32_c32_h8_w8_mode1",
       "provenance": {
-        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. Original portable and staged paths remain the oracle controls."
+        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. The portable and staged paths serve as controls."
       },
       "inputs": {
         "x": {
@@ -1838,7 +1834,7 @@
     {
       "name": "direct_matrix_inputs_m32_c32_h8_w8_mode2",
       "provenance": {
-        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. Original portable and staged paths remain the oracle controls."
+        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. The portable and staged paths serve as controls."
       },
       "inputs": {
         "x": {
@@ -1875,7 +1871,7 @@
     {
       "name": "direct_matrix_inputs_m32_c32_h8_w8_mode3",
       "provenance": {
-        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. Original portable and staged paths remain the oracle controls."
+        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. The portable and staged paths serve as controls."
       },
       "inputs": {
         "x": {
@@ -1917,7 +1913,7 @@
     {
       "name": "direct_matrix_inputs_m64_c32_h8_w8_mode0",
       "provenance": {
-        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. Original portable and staged paths remain the oracle controls."
+        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. The portable and staged paths serve as controls."
       },
       "inputs": {
         "x": {
@@ -1943,7 +1939,7 @@
     {
       "name": "direct_matrix_inputs_m64_c32_h8_w8_mode1",
       "provenance": {
-        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. Original portable and staged paths remain the oracle controls."
+        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. The portable and staged paths serve as controls."
       },
       "inputs": {
         "x": {
@@ -1974,7 +1970,7 @@
     {
       "name": "direct_matrix_inputs_m64_c32_h8_w8_mode2",
       "provenance": {
-        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. Original portable and staged paths remain the oracle controls."
+        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. The portable and staged paths serve as controls."
       },
       "inputs": {
         "x": {
@@ -2011,7 +2007,7 @@
     {
       "name": "direct_matrix_inputs_m64_c32_h8_w8_mode3",
       "provenance": {
-        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. Original portable and staged paths remain the oracle controls."
+        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. The portable and staged paths serve as controls."
       },
       "inputs": {
         "x": {
@@ -2053,7 +2049,7 @@
     {
       "name": "direct_matrix_inputs_m128_c32_h8_w8_mode0",
       "provenance": {
-        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. Original portable and staged paths remain the oracle controls."
+        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. The portable and staged paths serve as controls."
       },
       "inputs": {
         "x": {
@@ -2079,7 +2075,7 @@
     {
       "name": "direct_matrix_inputs_m128_c32_h8_w8_mode1",
       "provenance": {
-        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. Original portable and staged paths remain the oracle controls."
+        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. The portable and staged paths serve as controls."
       },
       "inputs": {
         "x": {
@@ -2110,7 +2106,7 @@
     {
       "name": "direct_matrix_inputs_m128_c32_h8_w8_mode2",
       "provenance": {
-        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. Original portable and staged paths remain the oracle controls."
+        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. The portable and staged paths serve as controls."
       },
       "inputs": {
         "x": {
@@ -2147,7 +2143,7 @@
     {
       "name": "direct_matrix_inputs_m128_c32_h8_w8_mode3",
       "provenance": {
-        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. Original portable and staged paths remain the oracle controls."
+        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. The portable and staged paths serve as controls."
       },
       "inputs": {
         "x": {
@@ -2189,7 +2185,7 @@
     {
       "name": "direct_matrix_inputs_m31_c32_h8_w8_mode0",
       "provenance": {
-        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. Original portable and staged paths remain the oracle controls."
+        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. The portable and staged paths serve as controls."
       },
       "inputs": {
         "x": {
@@ -2215,7 +2211,7 @@
     {
       "name": "direct_matrix_inputs_m33_c32_h8_w8_mode0",
       "provenance": {
-        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. Original portable and staged paths remain the oracle controls."
+        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. The portable and staged paths serve as controls."
       },
       "inputs": {
         "x": {
@@ -2241,7 +2237,7 @@
     {
       "name": "direct_matrix_inputs_m65_c32_h8_w8_mode0",
       "provenance": {
-        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. Original portable and staged paths remain the oracle controls."
+        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. The portable and staged paths serve as controls."
       },
       "inputs": {
         "x": {
@@ -2267,7 +2263,7 @@
     {
       "name": "direct_matrix_inputs_m64_c31_h8_w8_mode0",
       "provenance": {
-        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. Original portable and staged paths remain the oracle controls."
+        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. The portable and staged paths serve as controls."
       },
       "inputs": {
         "x": {
@@ -2293,7 +2289,7 @@
     {
       "name": "direct_matrix_inputs_m64_c32_h7_w9_mode0",
       "provenance": {
-        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. Original portable and staged paths remain the oracle controls."
+        "notes": "Aligned materialized-input matrix geometry and adjacent rejection boundaries; bias and residual are nonzero and precede activation in f32. The portable and staged paths serve as controls."
       },
       "inputs": {
         "x": {
@@ -2319,7 +2315,7 @@
     {
       "name": "half_direct_matrix_inputs_m32_c32_h8_w8_mode0",
       "provenance": {
-        "notes": "Aligned f16 materialized columns select direct subgroup-matrix inputs with half-sized scratch. Exercise row groups, batch indexing and each fused epilogue while retaining the existing f32 accumulation and f16 oracle tolerance."
+        "notes": "Aligned float16 3x3 convolution with 32 output channels, batch one, and no bias, residual, or activation checks float32 accumulation at float16 reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2345,7 +2341,7 @@
     {
       "name": "half_direct_matrix_inputs_m32_c32_h8_w8_mode1",
       "provenance": {
-        "notes": "Aligned f16 materialized columns select direct subgroup-matrix inputs with half-sized scratch. Exercise row groups, batch indexing and each fused epilogue while retaining the existing f32 accumulation and f16 oracle tolerance."
+        "notes": "Aligned float16 3x3 convolution with 32 output channels, batch one, channel bias, and fused Relu checks float32 accumulation at float16 reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2376,7 +2372,7 @@
     {
       "name": "half_direct_matrix_inputs_m32_c32_h8_w8_mode2",
       "provenance": {
-        "notes": "Aligned f16 materialized columns select direct subgroup-matrix inputs with half-sized scratch. Exercise row groups, batch indexing and each fused epilogue while retaining the existing f32 accumulation and f16 oracle tolerance."
+        "notes": "Aligned float16 3x3 convolution with 32 output channels, batch one, a residual input, and fused Tanh checks float32 accumulation at float16 reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2413,7 +2409,7 @@
     {
       "name": "half_direct_matrix_inputs_m32_c32_h8_w8_mode3",
       "provenance": {
-        "notes": "Aligned f16 materialized columns select direct subgroup-matrix inputs with half-sized scratch. Exercise row groups, batch indexing and each fused epilogue while retaining the existing f32 accumulation and f16 oracle tolerance."
+        "notes": "Aligned float16 convolution inputs with fused epilogues check row groups, batch indexing, and float32 accumulation at float16 reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2455,7 +2451,7 @@
     {
       "name": "half_direct_matrix_inputs_m64_c32_h8_w8_mode0",
       "provenance": {
-        "notes": "Aligned f16 materialized columns select direct subgroup-matrix inputs with half-sized scratch. Exercise row groups, batch indexing and each fused epilogue while retaining the existing f32 accumulation and f16 oracle tolerance."
+        "notes": "Aligned float16 3x3 convolution with 64 output channels, batch one, and no bias, residual, or activation checks float32 accumulation at float16 reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2481,7 +2477,7 @@
     {
       "name": "half_direct_matrix_inputs_m64_c32_h8_w8_mode1",
       "provenance": {
-        "notes": "Aligned f16 materialized columns select direct subgroup-matrix inputs with half-sized scratch. Exercise row groups, batch indexing and each fused epilogue while retaining the existing f32 accumulation and f16 oracle tolerance."
+        "notes": "Aligned float16 3x3 convolution with 64 output channels, batch one, channel bias, and fused Relu checks float32 accumulation at float16 reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2512,7 +2508,7 @@
     {
       "name": "half_direct_matrix_inputs_m64_c32_h8_w8_mode2",
       "provenance": {
-        "notes": "Aligned f16 materialized columns select direct subgroup-matrix inputs with half-sized scratch. Exercise row groups, batch indexing and each fused epilogue while retaining the existing f32 accumulation and f16 oracle tolerance."
+        "notes": "Aligned float16 3x3 convolution with 64 output channels, batch one, a residual input, and fused Tanh checks float32 accumulation at float16 reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2549,7 +2545,7 @@
     {
       "name": "half_direct_matrix_inputs_m64_c32_h8_w8_mode3",
       "provenance": {
-        "notes": "Aligned f16 materialized columns select direct subgroup-matrix inputs with half-sized scratch. Exercise row groups, batch indexing and each fused epilogue while retaining the existing f32 accumulation and f16 oracle tolerance."
+        "notes": "Aligned float16 convolution inputs with fused epilogues check row groups, batch indexing, and float32 accumulation at float16 reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2591,7 +2587,7 @@
     {
       "name": "half_direct_matrix_inputs_m96_c32_h8_w8_mode0",
       "provenance": {
-        "notes": "Aligned f16 materialized columns select direct subgroup-matrix inputs with half-sized scratch. Exercise row groups, batch indexing and each fused epilogue while retaining the existing f32 accumulation and f16 oracle tolerance."
+        "notes": "Aligned float16 3x3 convolution with 96 output channels, batch one, and no bias, residual, or activation checks float32 accumulation at float16 reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2617,7 +2613,7 @@
     {
       "name": "half_direct_matrix_inputs_m96_c32_h8_w8_mode1",
       "provenance": {
-        "notes": "Aligned f16 materialized columns select direct subgroup-matrix inputs with half-sized scratch. Exercise row groups, batch indexing and each fused epilogue while retaining the existing f32 accumulation and f16 oracle tolerance."
+        "notes": "Aligned float16 3x3 convolution with 96 output channels, batch one, channel bias, and fused Relu checks float32 accumulation at float16 reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2648,7 +2644,7 @@
     {
       "name": "half_direct_matrix_inputs_m96_c32_h8_w8_mode2",
       "provenance": {
-        "notes": "Aligned f16 materialized columns select direct subgroup-matrix inputs with half-sized scratch. Exercise row groups, batch indexing and each fused epilogue while retaining the existing f32 accumulation and f16 oracle tolerance."
+        "notes": "Aligned float16 3x3 convolution with 96 output channels, batch one, a residual input, and fused Tanh checks float32 accumulation at float16 reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2685,7 +2681,7 @@
     {
       "name": "half_direct_matrix_inputs_m96_c32_h8_w8_mode3",
       "provenance": {
-        "notes": "Aligned f16 materialized columns select direct subgroup-matrix inputs with half-sized scratch. Exercise row groups, batch indexing and each fused epilogue while retaining the existing f32 accumulation and f16 oracle tolerance."
+        "notes": "Aligned float16 convolution inputs with fused epilogues check row groups, batch indexing, and float32 accumulation at float16 reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2727,7 +2723,7 @@
     {
       "name": "half_direct_matrix_inputs_m128_c32_h8_w8_mode0",
       "provenance": {
-        "notes": "Aligned f16 materialized columns select direct subgroup-matrix inputs with half-sized scratch. Exercise row groups, batch indexing and each fused epilogue while retaining the existing f32 accumulation and f16 oracle tolerance."
+        "notes": "Aligned float16 3x3 convolution with 128 output channels, batch one, and no bias, residual, or activation checks float32 accumulation at float16 reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2753,7 +2749,7 @@
     {
       "name": "half_direct_matrix_inputs_m128_c32_h8_w8_mode1",
       "provenance": {
-        "notes": "Aligned f16 materialized columns select direct subgroup-matrix inputs with half-sized scratch. Exercise row groups, batch indexing and each fused epilogue while retaining the existing f32 accumulation and f16 oracle tolerance."
+        "notes": "Aligned float16 3x3 convolution with 128 output channels, batch one, channel bias, and fused Relu checks float32 accumulation at float16 reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2784,7 +2780,7 @@
     {
       "name": "half_direct_matrix_inputs_m128_c32_h8_w8_mode2",
       "provenance": {
-        "notes": "Aligned f16 materialized columns select direct subgroup-matrix inputs with half-sized scratch. Exercise row groups, batch indexing and each fused epilogue while retaining the existing f32 accumulation and f16 oracle tolerance."
+        "notes": "Aligned float16 3x3 convolution with 128 output channels, batch one, a residual input, and fused Tanh checks float32 accumulation at float16 reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2821,7 +2817,7 @@
     {
       "name": "half_direct_matrix_inputs_m128_c32_h8_w8_mode3",
       "provenance": {
-        "notes": "Aligned f16 materialized columns select direct subgroup-matrix inputs with half-sized scratch. Exercise row groups, batch indexing and each fused epilogue while retaining the existing f32 accumulation and f16 oracle tolerance."
+        "notes": "Aligned float16 convolution inputs with fused epilogues check row groups, batch indexing, and float32 accumulation at float16 reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2863,7 +2859,7 @@
     {
       "name": "half_direct_matrix_inputs_m31_c32_h8_w8_mode0",
       "provenance": {
-        "notes": "The direct f16 column path rejects incomplete output rows, reduction tiles and spatial tiles; retain the existing guarded fallback and oracle tolerance."
+        "notes": "The direct f16 column path rejects incomplete output rows, reduction tiles and spatial tiles, which take the guarded fallback at the same reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2889,7 +2885,7 @@
     {
       "name": "half_direct_matrix_inputs_m33_c32_h8_w8_mode0",
       "provenance": {
-        "notes": "The direct f16 column path rejects incomplete output rows, reduction tiles and spatial tiles; retain the existing guarded fallback and oracle tolerance."
+        "notes": "The direct f16 column path rejects incomplete output rows, reduction tiles and spatial tiles, which take the guarded fallback at the same reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2915,7 +2911,7 @@
     {
       "name": "half_direct_matrix_inputs_m65_c32_h8_w8_mode0",
       "provenance": {
-        "notes": "The direct f16 column path rejects incomplete output rows, reduction tiles and spatial tiles; retain the existing guarded fallback and oracle tolerance."
+        "notes": "The direct f16 column path rejects incomplete output rows, reduction tiles and spatial tiles, which take the guarded fallback at the same reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2941,7 +2937,7 @@
     {
       "name": "half_direct_matrix_inputs_m64_c31_h8_w8_mode0",
       "provenance": {
-        "notes": "The direct f16 column path rejects incomplete output rows, reduction tiles and spatial tiles; retain the existing guarded fallback and oracle tolerance."
+        "notes": "The direct f16 column path rejects incomplete output rows, reduction tiles and spatial tiles, which take the guarded fallback at the same reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2967,7 +2963,7 @@
     {
       "name": "half_direct_matrix_inputs_m64_c32_h7_w9_mode0",
       "provenance": {
-        "notes": "The direct f16 column path rejects incomplete output rows, reduction tiles and spatial tiles; retain the existing guarded fallback and oracle tolerance."
+        "notes": "The direct f16 column path rejects incomplete output rows, reduction tiles and spatial tiles, which take the guarded fallback at the same reference tolerance."
       },
       "inputs": {
         "x": {
@@ -2993,7 +2989,7 @@
     {
       "name": "direct_1x1_float32_m32_k32_n64_mode0",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0] },
       "inputs": {
@@ -3013,7 +3009,7 @@
     {
       "name": "direct_1x1_float32_m32_k32_n64_mode1",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0], "activation": "Relu" },
       "inputs": {
@@ -3038,7 +3034,7 @@
     {
       "name": "direct_1x1_float32_m32_k32_n64_mode2",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0], "activation": "Tanh" },
       "inputs": {
@@ -3063,7 +3059,7 @@
     {
       "name": "direct_1x1_float32_m32_k32_n64_mode3",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0], "activation": "LeakyRelu", "activation_params": [0.25] },
       "inputs": {
@@ -3093,7 +3089,7 @@
     {
       "name": "direct_1x1_float32_m64_k32_n64_mode0",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0] },
       "inputs": {
@@ -3113,7 +3109,7 @@
     {
       "name": "direct_1x1_float32_m64_k32_n64_mode1",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0], "activation": "Relu" },
       "inputs": {
@@ -3138,7 +3134,7 @@
     {
       "name": "direct_1x1_float32_m64_k32_n64_mode2",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0], "activation": "Tanh" },
       "inputs": {
@@ -3163,7 +3159,7 @@
     {
       "name": "direct_1x1_float32_m64_k32_n64_mode3",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0], "activation": "LeakyRelu", "activation_params": [0.25] },
       "inputs": {
@@ -3193,7 +3189,7 @@
     {
       "name": "direct_1x1_float32_m33_k32_n64_mode0",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0] },
       "inputs": {
@@ -3213,7 +3209,7 @@
     {
       "name": "direct_1x1_float32_m65_k32_n64_mode0",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0] },
       "inputs": {
@@ -3233,7 +3229,7 @@
     {
       "name": "direct_1x1_float32_m96_k32_n64_mode0",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0] },
       "inputs": {
@@ -3253,7 +3249,7 @@
     {
       "name": "direct_1x1_float32_m64_k31_n64_mode0",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0] },
       "inputs": {
@@ -3273,7 +3269,7 @@
     {
       "name": "direct_1x1_float32_m64_k33_n64_mode0",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0] },
       "inputs": {
@@ -3293,7 +3289,7 @@
     {
       "name": "direct_1x1_float32_m64_k32_n63_mode0",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0] },
       "inputs": {
@@ -3313,7 +3309,7 @@
     {
       "name": "direct_1x1_float32_m64_k32_n65_mode0",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0] },
       "inputs": {
@@ -3333,7 +3329,7 @@
     {
       "name": "direct_1x1_float16_m32_k32_n64_mode0",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0] },
       "inputs": {
@@ -3353,7 +3349,7 @@
     {
       "name": "direct_1x1_float16_m32_k32_n64_mode1",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0], "activation": "Relu" },
       "inputs": {
@@ -3378,7 +3374,7 @@
     {
       "name": "direct_1x1_float16_m32_k32_n64_mode2",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0], "activation": "Tanh" },
       "inputs": {
@@ -3403,7 +3399,7 @@
     {
       "name": "direct_1x1_float16_m32_k32_n64_mode3",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0], "activation": "LeakyRelu", "activation_params": [0.25] },
       "inputs": {
@@ -3433,7 +3429,7 @@
     {
       "name": "direct_1x1_float16_m64_k32_n64_mode0",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0] },
       "inputs": {
@@ -3453,7 +3449,7 @@
     {
       "name": "direct_1x1_float16_m64_k32_n64_mode1",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0], "activation": "Relu" },
       "inputs": {
@@ -3478,7 +3474,7 @@
     {
       "name": "direct_1x1_float16_m64_k32_n64_mode2",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0], "activation": "Tanh" },
       "inputs": {
@@ -3503,7 +3499,7 @@
     {
       "name": "direct_1x1_float16_m64_k32_n64_mode3",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0], "activation": "LeakyRelu", "activation_params": [0.25] },
       "inputs": {
@@ -3533,7 +3529,7 @@
     {
       "name": "direct_1x1_float16_m33_k32_n64_mode0",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0] },
       "inputs": {
@@ -3553,7 +3549,7 @@
     {
       "name": "direct_1x1_float16_m65_k32_n64_mode0",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0] },
       "inputs": {
@@ -3573,7 +3569,7 @@
     {
       "name": "direct_1x1_float16_m96_k32_n64_mode0",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0] },
       "inputs": {
@@ -3593,7 +3589,7 @@
     {
       "name": "direct_1x1_float16_m64_k31_n64_mode0",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0] },
       "inputs": {
@@ -3613,7 +3609,7 @@
     {
       "name": "direct_1x1_float16_m64_k33_n64_mode0",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0] },
       "inputs": {
@@ -3633,7 +3629,7 @@
     {
       "name": "direct_1x1_float16_m64_k32_n63_mode0",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0] },
       "inputs": {
@@ -3653,7 +3649,7 @@
     {
       "name": "direct_1x1_float16_m64_k32_n65_mode0",
       "provenance": {
-        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles retain guarded staging or the existing fallback; device limits determine the row tile."
+        "notes": "Batched direct 1x1 matrix loads cover both scalar widths and fused epilogues. Neighboring incomplete M/K/N tiles take guarded staging or the fallback route; device limits determine the row tile."
       },
       "attrs": { "pads": [0, 0, 0, 0] },
       "inputs": {
@@ -3669,6 +3665,5959 @@
         }
       },
       "outputs": { "y": { "dtype": "float16", "shape": [2, 64, 1, 65], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "half_columns_short_k_batched",
+      "provenance": {
+        "notes": "Float16 convolution with spatial and channel tails, batches, padding and epilogues checks full-output parity against the CPU reference."
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "dilations": [1, 1] },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 1, 1, 513],
+          "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [512, 1, 3, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.03, "scale": 0.1, "offset": 0.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 512, 1, 513], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "half_columns_batch_channel_tails",
+      "provenance": {
+        "notes": "Float16 convolution with spatial and channel tails, batches, padding and epilogues checks full-output parity against the CPU reference."
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 3, 17, 31],
+          "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [257, 3, 3, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.03, "scale": 0.1, "offset": 0.0 }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [257],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.020999999999999998,
+            "cosStep": 0.033,
+            "scale": 0.05,
+            "offset": 0.03
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 257, 17, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "half_columns_leakyrelu_odd_k",
+      "provenance": {
+        "notes": "Float16 convolution with spatial and channel tails, batches, padding and epilogues checks full-output parity against the CPU reference."
+      },
+      "attrs": {
+        "strides": [1, 1],
+        "pads": [1, 1, 1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.25]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 7, 17, 31],
+          "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [512, 7, 3, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.03, "scale": 0.1, "offset": 0.0 }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [512],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.020999999999999998,
+            "cosStep": 0.033,
+            "scale": 0.05,
+            "offset": 0.03
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 512, 17, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "half_columns_even_k",
+      "provenance": {
+        "notes": "Float16 convolution with spatial and channel tails, batches, padding and epilogues checks full-output parity against the CPU reference."
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "dilations": [1, 1] },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 32, 16, 32],
+          "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [512, 32, 3, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.03, "scale": 0.1, "offset": 0.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 512, 16, 32], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "half_columns_odd_k",
+      "provenance": {
+        "notes": "Float16 convolution with spatial and channel tails, batches, padding and epilogues checks full-output parity against the CPU reference."
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 33, 17, 31],
+          "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [512, 33, 3, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.03, "scale": 0.1, "offset": 0.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 512, 17, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "half_columns_stride2_asymmetric_pad",
+      "provenance": {
+        "notes": "Float16 convolution with spatial and channel tails, batches, padding and epilogues checks full-output parity against the CPU reference."
+      },
+      "attrs": { "strides": [2, 2], "pads": [1, 2, 1, 0], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 3, 33, 61],
+          "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [257, 3, 3, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.03, "scale": 0.1, "offset": 0.0 }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [257],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.020999999999999998,
+            "cosStep": 0.033,
+            "scale": 0.05,
+            "offset": 0.03
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 257, 17, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "half_columns_dilation2",
+      "provenance": {
+        "notes": "Float16 convolution with spatial and channel tails, batches, padding and epilogues checks full-output parity against the CPU reference."
+      },
+      "attrs": { "strides": [1, 1], "pads": [2, 2, 2, 2], "dilations": [2, 2], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 3, 17, 31],
+          "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [512, 3, 3, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.03, "scale": 0.1, "offset": 0.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 512, 17, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "half_columns_hardswish_short_k",
+      "provenance": {
+        "notes": "Float16 convolution with spatial and channel tails, batches, padding and epilogues checks full-output parity against the CPU reference."
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "dilations": [1, 1], "activation": "HardSwish" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 1, 1, 513],
+          "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [512, 1, 3, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.03, "scale": 0.1, "offset": 0.0 }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [512],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.020999999999999998,
+            "cosStep": 0.033,
+            "scale": 0.05,
+            "offset": 0.03
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 512, 1, 513], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-conv-f32-group8-k9-256c-64x64-nchw-fallback",
+      "provenance": {
+        "notes": "Grouped 9x9 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 8, "strides": [1, 1], "pads": [4, 4, 4, 4], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 256, 64, 64],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [256, 32, 9, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "shape": [1, 256, 64, 64], "dtype": "float32", "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-conv-f16-group8-k9-256c-64x64-nchw-f16-fallback",
+      "provenance": {
+        "notes": "Grouped 9x9 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 8, "strides": [1, 1], "pads": [4, 4, 4, 4], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 256, 64, 64],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [256, 32, 9, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "shape": [1, 256, 64, 64], "dtype": "float16", "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-conv-f32-group8-k9-256c-64x63-w4-boundary",
+      "provenance": {
+        "notes": "Grouped 9x9 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 8, "strides": [1, 1], "pads": [4, 4, 4, 4], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 256, 64, 63],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [256, 32, 9, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "shape": [1, 256, 64, 63], "dtype": "float32", "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m16_32x33_k5x5_d1x1_s1x1_bias0_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 8 input and 16 output channels per group, dilation 1x1, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "HardSwish" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 32, 32, 33],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [64, 8, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m16_32x33_k9x9_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 9x9 convolution with 8 input and 16 output channels per group, dilation 1x1, stride 1x1, LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [4, 4, 4, 4],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 32, 32, 33],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [64, 8, 9, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m16_32x33_k5x5_d1x1_s1x1_bias0_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 8 input and 16 output channels per group, dilation 1x1, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "HardSwish" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 32, 32, 33],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [64, 8, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m16_32x33_k9x9_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 9x9 convolution with 8 input and 16 output channels per group, dilation 1x1, stride 1x1, LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [4, 4, 4, 4],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 32, 32, 33],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [64, 8, 9, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-identity",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 8 input and 32 output channels per group, dilation 1x1, stride 1x1, identity activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1] },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 32, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 8, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 8 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 32, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 8, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 8 input and 32 output channels per group, dilation 1x1, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "HardSwish" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 32, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 8, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-identity",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, identity activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1] },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 16, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 16, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "HardSwish" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 16, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-identity",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, identity activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1] },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 32, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 32, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "HardSwish" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 32, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-identity",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 8 input and 32 output channels per group, dilation 1x1, stride 1x1, identity activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1] },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 32, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 8, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 8 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 32, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 8, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 8 input and 32 output channels per group, dilation 1x1, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "HardSwish" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 32, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 8, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-identity",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, identity activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1] },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 16, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 16, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "HardSwish" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 16, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-identity",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, identity activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1] },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 32, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 32, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "HardSwish" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 32, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-identity",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, identity activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 3, "pads": [10, 10, 11, 12], "strides": [1, 1], "dilations": [2, 2] },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [18, 3, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-Relu",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 3, "pads": [10, 10, 11, 12], "strides": [1, 1], "dilations": [2, 2], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [18, 3, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [10, 10, 11, 12],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [18, 3, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-Sigmoid",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, Sigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 3, "pads": [10, 10, 11, 12], "strides": [1, 1], "dilations": [2, 2], "activation": "Sigmoid" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [18, 3, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-Tanh",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, Tanh activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 3, "pads": [10, 10, 11, 12], "strides": [1, 1], "dilations": [2, 2], "activation": "Tanh" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [18, 3, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-HardSigmoid",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, HardSigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [10, 10, 11, 12],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "HardSigmoid",
+        "activation_params": [0.2, 0.5]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [18, 3, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-HardSwish",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [10, 10, 11, 12],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "HardSwish"
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [18, 3, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-Clip",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, Clip activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [10, 10, 11, 12],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "Clip",
+        "activation_params": [-0.15, 0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [18, 3, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-identity",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, identity activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 3, "pads": [10, 10, 11, 12], "strides": [1, 1], "dilations": [2, 2] },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [18, 3, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-Relu",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 3, "pads": [10, 10, 11, 12], "strides": [1, 1], "dilations": [2, 2], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [18, 3, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [10, 10, 11, 12],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [18, 3, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-Sigmoid",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, Sigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 3, "pads": [10, 10, 11, 12], "strides": [1, 1], "dilations": [2, 2], "activation": "Sigmoid" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [18, 3, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-Tanh",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, Tanh activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 3, "pads": [10, 10, 11, 12], "strides": [1, 1], "dilations": [2, 2], "activation": "Tanh" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [18, 3, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-HardSigmoid",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, HardSigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [10, 10, 11, 12],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "HardSigmoid",
+        "activation_params": [0.2, 0.5]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [18, 3, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-HardSwish",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [10, 10, 11, 12],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "HardSwish"
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [18, 3, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-Clip",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, Clip activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [10, 10, 11, 12],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "Clip",
+        "activation_params": [-0.15, 0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [18, 3, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g16_c1_m2_33x35_k11x11_d2x2_s1x1_bias0_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 1 input and 2 output channels per group, dilation 2x2, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 16,
+        "pads": [10, 10, 10, 10],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "HardSwish"
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 16, 33, 35],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [32, 1, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 33, 35], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g16_c1_m2_33x35_k11x11_d3x3_s1x1_bias0_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 1 input and 2 output channels per group, dilation 3x3, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 16,
+        "pads": [15, 15, 15, 15],
+        "strides": [1, 1],
+        "dilations": [3, 3],
+        "activation": "HardSwish"
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 16, 33, 35],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [32, 1, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 33, 35], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g16_c1_m2_33x35_k11x11_d2x2_s1x1_bias0_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 1 input and 2 output channels per group, dilation 2x2, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 16,
+        "pads": [10, 10, 10, 10],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "HardSwish"
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 16, 33, 35],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [32, 1, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 33, 35], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g16_c1_m2_33x35_k11x11_d3x3_s1x1_bias0_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 1 input and 2 output channels per group, dilation 3x3, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 16,
+        "pads": [15, 15, 15, 15],
+        "strides": [1, 1],
+        "dilations": [3, 3],
+        "activation": "HardSwish"
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 16, 33, 35],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [32, 1, 11, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 33, 35], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_33x35_k3x11_d1x2_s1x1_bias0_asym1-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 3x11 convolution with 3 input and 6 output channels per group, dilation 1x2, stride 1x1, LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [1, 10, 2, 12],
+        "strides": [1, 1],
+        "dilations": [1, 2],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 33, 35],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [24, 3, 3, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 34, 37], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_33x35_k11x3_d2x1_s1x1_bias0_asym1-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 11x3 convolution with 3 input and 6 output channels per group, dilation 2x1, stride 1x1, LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [10, 1, 11, 3],
+        "strides": [1, 1],
+        "dilations": [2, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 33, 35],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [24, 3, 11, 3],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 34, 37], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_33x35_k5x9_d1x3_s2x1_bias0_asym1-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 5x9 convolution with 3 input and 6 output channels per group, dilation 1x3, stride 2x1, LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [2, 12, 3, 14],
+        "strides": [2, 1],
+        "dilations": [1, 3],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 33, 35],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [24, 3, 5, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 17, 37], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_33x35_k7x5_d2x2_s1x2_bias0_asym1-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 7x5 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x2, LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [6, 4, 7, 6],
+        "strides": [1, 2],
+        "dilations": [2, 2],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 33, 35],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [24, 3, 7, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 34, 19], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c3_m6_33x35_k3x11_d1x2_s1x1_bias0_asym1-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 3x11 convolution with 3 input and 6 output channels per group, dilation 1x2, stride 1x1, LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [1, 10, 2, 12],
+        "strides": [1, 1],
+        "dilations": [1, 2],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 12, 33, 35],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [24, 3, 3, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 24, 34, 37], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c3_m6_33x35_k11x3_d2x1_s1x1_bias0_asym1-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 11x3 convolution with 3 input and 6 output channels per group, dilation 2x1, stride 1x1, LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [10, 1, 11, 3],
+        "strides": [1, 1],
+        "dilations": [2, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 12, 33, 35],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [24, 3, 11, 3],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 24, 34, 37], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c3_m6_33x35_k5x9_d1x3_s2x1_bias0_asym1-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 5x9 convolution with 3 input and 6 output channels per group, dilation 1x3, stride 2x1, LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [2, 12, 3, 14],
+        "strides": [2, 1],
+        "dilations": [1, 3],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 12, 33, 35],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [24, 3, 5, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 24, 17, 37], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c3_m6_33x35_k7x5_d2x2_s1x2_bias0_asym1-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 7x5 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x2, LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [6, 4, 7, 6],
+        "strides": [1, 2],
+        "dilations": [2, 2],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 12, 33, 35],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [24, 3, 7, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 24, 34, 19], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x7_k5x5_d1x2_s1x1_bias0_asym0-Sigmoid",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 3 input and 6 output channels per group, dilation 1x2, stride 1x1, Sigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 4, 2, 4], "strides": [1, 1], "dilations": [1, 2], "activation": "Sigmoid" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 40, 7],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [24, 3, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 7], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x8_k5x5_d1x2_s1x1_bias0_asym0-Sigmoid",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 3 input and 6 output channels per group, dilation 1x2, stride 1x1, Sigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 4, 2, 4], "strides": [1, 1], "dilations": [1, 2], "activation": "Sigmoid" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 40, 8],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [24, 3, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 8], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x9_k5x5_d1x2_s1x1_bias0_asym0-Sigmoid",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 3 input and 6 output channels per group, dilation 1x2, stride 1x1, Sigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 4, 2, 4], "strides": [1, 1], "dilations": [1, 2], "activation": "Sigmoid" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 40, 9],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [24, 3, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 9], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x11_k5x5_d1x3_s1x1_bias0_asym0-Sigmoid",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 3 input and 6 output channels per group, dilation 1x3, stride 1x1, Sigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 6, 2, 6], "strides": [1, 1], "dilations": [1, 3], "activation": "Sigmoid" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 40, 11],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [24, 3, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 11], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x12_k5x5_d1x3_s1x1_bias0_asym0-Sigmoid",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 3 input and 6 output channels per group, dilation 1x3, stride 1x1, Sigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 6, 2, 6], "strides": [1, 1], "dilations": [1, 3], "activation": "Sigmoid" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 40, 12],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [24, 3, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 12], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x13_k5x5_d1x3_s1x1_bias0_asym0-Sigmoid",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 3 input and 6 output channels per group, dilation 1x3, stride 1x1, Sigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 6, 2, 6], "strides": [1, 1], "dilations": [1, 3], "activation": "Sigmoid" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 40, 13],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [24, 3, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 13], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m16_32x33_k3x3_d1x1_s1x1_bias1_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 3x3 convolution with 16 input and 16 output channels per group, dilation 1x1, stride 1x1, bias and HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 1, 1, 1], "strides": [1, 1], "dilations": [1, 1], "activation": "HardSwish" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 64, 32, 33],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [64, 16, 3, 3],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [64],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m16_32x33_k3x3_d2x2_s1x1_bias1_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 3x3 convolution with 16 input and 16 output channels per group, dilation 2x2, stride 1x1, bias and HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [2, 2], "activation": "HardSwish" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 64, 32, 33],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [64, 16, 3, 3],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [64],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m16_32x33_k5x5_d1x1_s1x1_bias1_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 16 input and 16 output channels per group, dilation 1x1, stride 1x1, bias and Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 64, 32, 33],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [64, 16, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [64],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m16_32x33_k5x5_d2x2_s1x1_bias1_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 16 input and 16 output channels per group, dilation 2x2, stride 1x1, bias and Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [4, 4, 4, 4], "strides": [1, 1], "dilations": [2, 2], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 64, 32, 33],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [64, 16, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [64],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m16_32x33_k7x7_d1x1_s1x1_bias1_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 16 input and 16 output channels per group, dilation 1x1, stride 1x1, bias and Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [3, 3, 3, 3], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 64, 32, 33],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [64, 16, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [64],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m16_32x33_k7x7_d2x2_s1x1_bias1_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 16 input and 16 output channels per group, dilation 2x2, stride 1x1, bias and Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [6, 6, 6, 6], "strides": [1, 1], "dilations": [2, 2], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 64, 32, 33],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [64, 16, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [64],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m16_32x33_k3x3_d1x1_s1x1_bias1_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 3x3 convolution with 16 input and 16 output channels per group, dilation 1x1, stride 1x1, bias and HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 1, 1, 1], "strides": [1, 1], "dilations": [1, 1], "activation": "HardSwish" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 64, 32, 33],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [64, 16, 3, 3],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [64],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m16_32x33_k3x3_d2x2_s1x1_bias1_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 3x3 convolution with 16 input and 16 output channels per group, dilation 2x2, stride 1x1, bias and HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [2, 2], "activation": "HardSwish" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 64, 32, 33],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [64, 16, 3, 3],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [64],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m16_32x33_k5x5_d1x1_s1x1_bias1_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 16 input and 16 output channels per group, dilation 1x1, stride 1x1, bias and Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 64, 32, 33],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [64, 16, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [64],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m16_32x33_k5x5_d2x2_s1x1_bias1_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 16 input and 16 output channels per group, dilation 2x2, stride 1x1, bias and Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [4, 4, 4, 4], "strides": [1, 1], "dilations": [2, 2], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 64, 32, 33],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [64, 16, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [64],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m16_32x33_k7x7_d1x1_s1x1_bias1_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 16 input and 16 output channels per group, dilation 1x1, stride 1x1, bias and Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [3, 3, 3, 3], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 64, 32, 33],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [64, 16, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [64],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m16_32x33_k7x7_d2x2_s1x1_bias1_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 16 input and 16 output channels per group, dilation 2x2, stride 1x1, bias and Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [6, 6, 6, 6], "strides": [1, 1], "dilations": [2, 2], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 64, 32, 33],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [64, 16, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [64],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c7_m32_48x48_k4x6_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 4x6 convolution with 7 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 2, 1, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 28, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 7, 4, 6],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 47], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c7_m32_48x48_k3x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 3x9 convolution with 7 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 4, 1, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 28, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 7, 3, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m32_48x48_k4x6_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 4x6 convolution with 8 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 2, 1, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 32, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 8, 4, 6],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 47], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m32_48x48_k3x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 3x9 convolution with 8 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 4, 1, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 32, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 8, 3, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c9_m32_48x48_k4x6_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 4x6 convolution with 9 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 2, 1, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 36, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 9, 4, 6],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 47], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c9_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 9 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 36, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 9, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c9_m32_48x48_k3x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 3x9 convolution with 9 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 4, 1, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 36, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 9, 3, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k4x6_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 4x6 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 2, 1, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 16, 4, 6],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 47], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k3x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 3x9 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 4, 1, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 16, 3, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k4x6_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 4x6 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 2, 1, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 32, 4, 6],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 47], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k3x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 3x9 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 4, 1, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 32, 3, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c15_m32_48x48_k4x6_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 4x6 convolution with 15 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 2, 1, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 60, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 15, 4, 6],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 47], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c15_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 15 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 60, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 15, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c15_m32_48x48_k3x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 3x9 convolution with 15 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 4, 1, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 60, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 15, 3, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k4x6_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 4x6 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 2, 1, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 16, 4, 6],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 47], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k3x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 3x9 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 4, 1, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 16, 3, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c17_m32_48x48_k4x6_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 4x6 convolution with 17 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 2, 1, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 68, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 17, 4, 6],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 47], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c17_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 17 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 68, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 17, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c17_m32_48x48_k3x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 3x9 convolution with 17 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 4, 1, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 68, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 17, 3, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k4x6_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 4x6 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 2, 1, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 32, 4, 6],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 47], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k3x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 3x9 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [1, 4, 1, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 32, 3, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-identity",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, bias and identity activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 3, "pads": [6, 6, 7, 8], "strides": [1, 1], "dilations": [2, 2] },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [18, 3, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [18],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.5, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-Relu",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, bias and Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 3, "pads": [6, 6, 7, 8], "strides": [1, 1], "dilations": [2, 2], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [18, 3, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [18],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.5, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, bias and LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [6, 6, 7, 8],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [18, 3, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [18],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.5, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-Sigmoid",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, bias and Sigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 3, "pads": [6, 6, 7, 8], "strides": [1, 1], "dilations": [2, 2], "activation": "Sigmoid" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [18, 3, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [18],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.5, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-Tanh",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, bias and Tanh activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 3, "pads": [6, 6, 7, 8], "strides": [1, 1], "dilations": [2, 2], "activation": "Tanh" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [18, 3, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [18],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.5, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-HardSigmoid",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, bias and HardSigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [6, 6, 7, 8],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "HardSigmoid",
+        "activation_params": [0.2, 0.5]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [18, 3, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [18],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.5, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-HardSwish",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, bias and HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 3, "pads": [6, 6, 7, 8], "strides": [1, 1], "dilations": [2, 2], "activation": "HardSwish" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [18, 3, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [18],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.5, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-Clip",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, bias and Clip activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [6, 6, 7, 8],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "Clip",
+        "activation_params": [-0.15, 0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [18, 3, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [18],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.5, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-identity",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, bias and identity activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 3, "pads": [6, 6, 7, 8], "strides": [1, 1], "dilations": [2, 2] },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [18, 3, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [18],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.5, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-Relu",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, bias and Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 3, "pads": [6, 6, 7, 8], "strides": [1, 1], "dilations": [2, 2], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [18, 3, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [18],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.5, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, bias and LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [6, 6, 7, 8],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [18, 3, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [18],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.5, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-Sigmoid",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, bias and Sigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 3, "pads": [6, 6, 7, 8], "strides": [1, 1], "dilations": [2, 2], "activation": "Sigmoid" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [18, 3, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [18],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.5, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-Tanh",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, bias and Tanh activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 3, "pads": [6, 6, 7, 8], "strides": [1, 1], "dilations": [2, 2], "activation": "Tanh" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [18, 3, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [18],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.5, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-HardSigmoid",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, bias and HardSigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [6, 6, 7, 8],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "HardSigmoid",
+        "activation_params": [0.2, 0.5]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [18, 3, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [18],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.5, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-HardSwish",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, bias and HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 3, "pads": [6, 6, 7, 8], "strides": [1, 1], "dilations": [2, 2], "activation": "HardSwish" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [18, 3, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [18],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.5, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-Clip",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 3 input and 6 output channels per group, dilation 2x2, stride 1x1, bias and Clip activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [6, 6, 7, 8],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "Clip",
+        "activation_params": [-0.15, 0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 2.0, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [18, 3, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.4,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [18],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.5, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g16_c1_m1_33x35_k3x3_d2x2_s1x1_bias1_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 3x3 convolution with 1 input and 1 output channels per group, dilation 2x2, stride 1x1, bias and HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 16, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [2, 2], "activation": "HardSwish" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 16, 33, 35],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [16, 1, 3, 3],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [16],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 16, 33, 35], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g16_c1_m2_33x35_k3x3_d2x2_s1x1_bias1_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 3x3 convolution with 1 input and 2 output channels per group, dilation 2x2, stride 1x1, bias and HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 16, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [2, 2], "activation": "HardSwish" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 16, 33, 35],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [32, 1, 3, 3],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [32],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 33, 35], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g16_c1_m1_33x35_k3x3_d2x2_s1x1_bias1_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 3x3 convolution with 1 input and 1 output channels per group, dilation 2x2, stride 1x1, bias and HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 16, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [2, 2], "activation": "HardSwish" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 16, 33, 35],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [16, 1, 3, 3],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [16],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 16, 33, 35], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g16_c1_m2_33x35_k3x3_d2x2_s1x1_bias1_asym0-HardSwish",
+      "provenance": {
+        "notes": "Grouped 3x3 convolution with 1 input and 2 output channels per group, dilation 2x2, stride 1x1, bias and HardSwish activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 16, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [2, 2], "activation": "HardSwish" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 16, 33, 35],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [32, 1, 3, 3],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [32],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 33, 35], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m1_17x29_k5x5_d1x1_s1x1_bias1_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 3 input and 1 output channels per group, dilation 1x1, stride 1x1, bias and LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [2, 2, 2, 2],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [3, 3, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 17, 29], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m3_17x29_k5x5_d1x1_s1x1_bias1_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 3 input and 3 output channels per group, dilation 1x1, stride 1x1, bias and LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [2, 2, 2, 2],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [9, 3, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [9],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 9, 17, 29], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m5_17x29_k5x5_d1x1_s1x1_bias1_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 3 input and 5 output channels per group, dilation 1x1, stride 1x1, bias and LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [2, 2, 2, 2],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [15, 3, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [15],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 15, 17, 29], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m1_17x29_k5x5_d1x1_s1x1_bias1_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 3 input and 1 output channels per group, dilation 1x1, stride 1x1, bias and LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [2, 2, 2, 2],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [3, 3, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 3, 17, 29], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m3_17x29_k5x5_d1x1_s1x1_bias1_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 3 input and 3 output channels per group, dilation 1x1, stride 1x1, bias and LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [2, 2, 2, 2],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [9, 3, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [9],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 9, 17, 29], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m5_17x29_k5x5_d1x1_s1x1_bias1_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 3 input and 5 output channels per group, dilation 1x1, stride 1x1, bias and LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": {
+        "group": 3,
+        "pads": [2, 2, 2, 2],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 9, 17, 29],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [15, 3, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [15],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 15, 17, 29], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x7_k5x5_d1x2_s1x1_bias1_asym0-Sigmoid",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 3 input and 6 output channels per group, dilation 1x2, stride 1x1, bias and Sigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 4, 2, 4], "strides": [1, 1], "dilations": [1, 2], "activation": "Sigmoid" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 40, 7],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [24, 3, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [24],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 7], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x8_k5x5_d1x2_s1x1_bias1_asym0-Sigmoid",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 3 input and 6 output channels per group, dilation 1x2, stride 1x1, bias and Sigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 4, 2, 4], "strides": [1, 1], "dilations": [1, 2], "activation": "Sigmoid" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 40, 8],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [24, 3, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [24],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 8], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x9_k5x5_d1x2_s1x1_bias1_asym0-Sigmoid",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 3 input and 6 output channels per group, dilation 1x2, stride 1x1, bias and Sigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 4, 2, 4], "strides": [1, 1], "dilations": [1, 2], "activation": "Sigmoid" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 40, 9],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [24, 3, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [24],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 9], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x11_k5x5_d1x3_s1x1_bias1_asym0-Sigmoid",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 3 input and 6 output channels per group, dilation 1x3, stride 1x1, bias and Sigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 6, 2, 6], "strides": [1, 1], "dilations": [1, 3], "activation": "Sigmoid" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 40, 11],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [24, 3, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [24],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 11], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x12_k5x5_d1x3_s1x1_bias1_asym0-Sigmoid",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 3 input and 6 output channels per group, dilation 1x3, stride 1x1, bias and Sigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 6, 2, 6], "strides": [1, 1], "dilations": [1, 3], "activation": "Sigmoid" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 40, 12],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [24, 3, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [24],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 12], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x13_k5x5_d1x3_s1x1_bias1_asym0-Sigmoid",
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 3 input and 6 output channels per group, dilation 1x3, stride 1x1, bias and Sigmoid activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 6, 2, 6], "strides": [1, 1], "dilations": [1, 3], "activation": "Sigmoid" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 40, 13],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [24, 3, 5, 5],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [24],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 13], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k7x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 7x9 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [3, 4, 3, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 16, 7, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k8x8_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 8x8 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [3, 3, 3, 3], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 16, 8, 8],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 47], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k6x11_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 6x11 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 5, 2, 5], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 16, 6, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k8x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 8x9 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [3, 4, 3, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 16, 8, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k9x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 9x9 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [4, 4, 4, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 16, 9, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k7x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 7x9 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [3, 4, 3, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 32, 7, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k8x8_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 8x8 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [3, 3, 3, 3], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 32, 8, 8],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 47], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k6x11_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 6x11 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 5, 2, 5], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 32, 6, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k8x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 8x9 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [3, 4, 3, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 32, 8, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k9x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 9x9 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [4, 4, 4, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [128, 32, 9, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k7x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 7x9 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [3, 4, 3, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 16, 7, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k8x8_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 8x8 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [3, 3, 3, 3], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 16, 8, 8],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 47], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k6x11_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 6x11 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 5, 2, 5], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 16, 6, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k8x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 8x9 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [3, 4, 3, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 16, 8, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k9x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 9x9 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [4, 4, 4, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 64, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 16, 9, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k7x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 7x9 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [3, 4, 3, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 32, 7, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k8x8_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 8x8 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [3, 3, 3, 3], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 32, 8, 8],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 47], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k6x11_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 6x11 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [2, 5, 2, 5], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 32, 6, 11],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k8x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 8x9 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [3, 4, 3, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 32, 8, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k9x9_d1x1_s1x1_bias0_asym0-Relu",
+      "provenance": {
+        "notes": "Grouped 9x9 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      },
+      "attrs": { "group": 4, "pads": [4, 4, 4, 4], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 128, 48, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [128, 32, 9, 9],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float32_b1_g4_c1_m2_16x15_k7x7_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [3, 3, 3, 3],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 4, 16, 15],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 1, 7, 7],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 15], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float32_b1_g4_c1_m2_16x16_k7x7_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [3, 3, 3, 3],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 4, 16, 16],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 1, 7, 7],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 16], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float32_b1_g4_c1_m2_16x17_k7x7_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [3, 3, 3, 3],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 4, 16, 17],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 1, 7, 7],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 17], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float32_b1_g4_c1_m2_16x15_k9x9_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [4, 4, 4, 4],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 4, 16, 15],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 1, 9, 9],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 15], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float32_b1_g4_c1_m2_16x16_k9x9_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [4, 4, 4, 4],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 4, 16, 16],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 1, 9, 9],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 16], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float32_b1_g4_c1_m2_16x17_k9x9_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [4, 4, 4, 4],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 4, 16, 17],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 1, 9, 9],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 17], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float32_b1_g4_c1_m2_16x15_k11x11_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [5, 5, 5, 5],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 4, 16, 15],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 1, 11, 11],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 15], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float32_b1_g4_c1_m2_16x16_k11x11_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [5, 5, 5, 5],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 4, 16, 16],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 1, 11, 11],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 16], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float32_b1_g4_c1_m2_16x17_k11x11_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [5, 5, 5, 5],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 4, 16, 17],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 1, 11, 11],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 17], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float32_b1_g4_c3_m2_16x15_k7x7_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [3, 3, 3, 3],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 16, 15],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 3, 7, 7],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 15], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float32_b1_g4_c3_m2_16x16_k7x7_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [3, 3, 3, 3],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 16, 16],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 3, 7, 7],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 16], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float32_b1_g4_c3_m2_16x17_k7x7_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [3, 3, 3, 3],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 16, 17],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 3, 7, 7],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 17], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float32_b1_g4_c3_m2_16x15_k9x9_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [4, 4, 4, 4],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 16, 15],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 3, 9, 9],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 15], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float32_b1_g4_c3_m2_16x16_k9x9_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [4, 4, 4, 4],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 16, 16],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 3, 9, 9],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 16], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float32_b1_g4_c3_m2_16x17_k9x9_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [4, 4, 4, 4],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 16, 17],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 3, 9, 9],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 17], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float32_b1_g4_c3_m2_16x15_k11x11_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [5, 5, 5, 5],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 16, 15],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 3, 11, 11],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 15], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float32_b1_g4_c3_m2_16x16_k11x11_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [5, 5, 5, 5],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 16, 16],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 3, 11, 11],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 16], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float32_b1_g4_c3_m2_16x17_k11x11_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [5, 5, 5, 5],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 12, 16, 17],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 3, 11, 11],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 17], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float16_b1_g4_c1_m2_16x15_k7x7_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [3, 3, 3, 3],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 4, 16, 15],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [8, 1, 7, 7],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 16, 15], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float16_b1_g4_c1_m2_16x16_k7x7_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [3, 3, 3, 3],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 4, 16, 16],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [8, 1, 7, 7],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 16, 16], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float16_b1_g4_c1_m2_16x17_k7x7_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [3, 3, 3, 3],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 4, 16, 17],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [8, 1, 7, 7],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 16, 17], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float16_b1_g4_c1_m2_16x15_k9x9_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [4, 4, 4, 4],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 4, 16, 15],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [8, 1, 9, 9],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 16, 15], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float16_b1_g4_c1_m2_16x16_k9x9_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [4, 4, 4, 4],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 4, 16, 16],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [8, 1, 9, 9],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 16, 16], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float16_b1_g4_c1_m2_16x17_k9x9_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [4, 4, 4, 4],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 4, 16, 17],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [8, 1, 9, 9],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 16, 17], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float16_b1_g4_c1_m2_16x15_k11x11_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [5, 5, 5, 5],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 4, 16, 15],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [8, 1, 11, 11],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 16, 15], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float16_b1_g4_c1_m2_16x16_k11x11_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [5, 5, 5, 5],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 4, 16, 16],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [8, 1, 11, 11],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 16, 16], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float16_b1_g4_c1_m2_16x17_k11x11_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [5, 5, 5, 5],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 4, 16, 17],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [8, 1, 11, 11],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 16, 17], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float16_b1_g4_c3_m2_16x15_k7x7_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [3, 3, 3, 3],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 12, 16, 15],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [8, 3, 7, 7],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 16, 15], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float16_b1_g4_c3_m2_16x16_k7x7_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [3, 3, 3, 3],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 12, 16, 16],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [8, 3, 7, 7],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 16, 16], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float16_b1_g4_c3_m2_16x17_k7x7_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [3, 3, 3, 3],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 12, 16, 17],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [8, 3, 7, 7],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 16, 17], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float16_b1_g4_c3_m2_16x15_k9x9_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [4, 4, 4, 4],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 12, 16, 15],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [8, 3, 9, 9],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 16, 15], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float16_b1_g4_c3_m2_16x16_k9x9_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [4, 4, 4, 4],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 12, 16, 16],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [8, 3, 9, 9],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 16, 16], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float16_b1_g4_c3_m2_16x17_k9x9_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [4, 4, 4, 4],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 12, 16, 17],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [8, 3, 9, 9],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 16, 17], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float16_b1_g4_c3_m2_16x15_k11x11_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [5, 5, 5, 5],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 12, 16, 15],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [8, 3, 11, 11],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 16, 15], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float16_b1_g4_c3_m2_16x16_k11x11_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [5, 5, 5, 5],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 12, 16, 16],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [8, 3, 11, 11],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 16, 16], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "coverage-fused-grouped_float16_b1_g4_c3_m2_16x17_k11x11_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "provenance": {
+        "notes": "Grouped convolution at output widths 15, 16 and 17 with height 16. Checks complete output coverage for multi-channel and depthwise groups with 7x7, 9x9 and 11x11 filters."
+      },
+      "attrs": {
+        "group": 4,
+        "pads": [5, 5, 5, 5],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 12, 16, 17],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [8, 3, 11, 11],
+          "data": { "kind": "fillFloat32", "sinStep": 0.20800000000000002, "cosStep": 0.246, "scale": 0.02 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 16, 17], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "grouped_residual_direct_control",
+      "provenance": {
+        "notes": "Grouped convolution with a residual input checks that Z is included before activation at a large spatial extent."
+      },
+      "attrs": { "group": 2, "activation": "Relu", "pads": [1, 1, 1, 1], "strides": [1, 1] },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 8, 16, 17],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 4, 3, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.071, "cosStep": 0.153, "scale": 0.02 }
+        },
+        "zResidual": {
+          "dtype": "float32",
+          "shape": [1, 8, 16, 17],
+          "data": { "kind": "fillFloat32", "sinStep": 0.173, "cosStep": 0.113, "scale": 0.2 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 17], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "grouped_bias_residual_direct_control",
+      "provenance": {
+        "notes": "Grouped convolution with a residual input checks that Z is included before activation at a large spatial extent."
+      },
+      "attrs": { "group": 2, "activation": "Relu", "pads": [1, 1, 1, 1], "strides": [1, 1] },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 8, 16, 17],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 4, 3, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.071, "cosStep": 0.153, "scale": 0.02 }
+        },
+        "zResidual": {
+          "dtype": "float32",
+          "shape": [1, 8, 16, 17],
+          "data": { "kind": "fillFloat32", "sinStep": 0.173, "cosStep": 0.113, "scale": 0.2 }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [8],
+          "data": { "kind": "fillFloat32", "sinStep": 0.071, "cosStep": 0.193, "scale": 0.03 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 16, 17], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-bias-adjacent-float32-b1-c64-m64-g4-32x32-k7",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 16 input and output channels per group, width 32, unit dilation, bias and ReLU. Checks the adjacent four-column layout with a complete output vector."
+      },
+      "attrs": { "group": 4, "pads": [3, 3, 3, 3], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 64, 32, 32],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [64, 16, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [64],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 32], "tolerance": 0.0001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "fused-bias-adjacent-float16-b1-c64-m64-g4-32x32-k7",
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 16 input and output channels per group, width 32, unit dilation, bias and ReLU. Checks the adjacent four-column layout with a complete output vector."
+      },
+      "attrs": { "group": 4, "pads": [3, 3, 3, 3], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 64, 32, 32],
+          "data": { "kind": "fillFloat32", "sinStep": 0.137, "cosStep": 0.193, "scale": 0.2, "offset": 0.0 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [64, 16, 7, 7],
+          "data": {
+            "kind": "fillFloat32",
+            "sinStep": 0.20800000000000002,
+            "cosStep": 0.246,
+            "scale": 0.02,
+            "offset": 0.0
+          }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [64],
+          "data": { "kind": "fillFloat32", "sinStep": 0.279, "cosStep": 0.299, "scale": 0.03, "offset": 0.017 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 32], "tolerance": 0.001, "relTolerance": 0.001 } }
+    },
+    {
+      "name": "ort_conv2d_quickgelu_fusion_parity",
+      "provenance": {
+        "source": "onnxruntime/test/optimizer/graph_transform_test.cc",
+        "test": "GraphTransformationTests.WebGpuIm2ColConvQuickGeluFusionMatchesUnfusedResults",
+        "notes": "ORT builds x*sigmoid(alpha*x) after a Conv and checks the fused result against the unfused graph on random data in [-3,3]; transcribed here with the same shapes and a deterministic fill. alpha 1.4 as in the upstream test."
+      },
+      "attrs": {
+        "activation": "QuickGelu",
+        "activation_params": [1.4],
+        "kernel_shape": [3, 3],
+        "pads": [1, 1, 1, 1],
+        "strides": [1, 1],
+        "dilations": [1, 1]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 4, 14, 14],
+          "data": { "kind": "fillFloat32", "sinStep": 0.23, "cosStep": 0.41, "scale": 3.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 4, 3, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.019, "cosStep": 0.031, "scale": 0.5 }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [8],
+          "data": { "kind": "fillFloat32", "sinStep": 0.007, "cosStep": 0.017, "scale": 0.5 }
+        }
+      },
+      "outputs": {
+        "y": { "dtype": "float32", "shape": [1, 8, 14, 14], "tolerance": 0.000001, "relTolerance": 0.00001 }
+      }
+    },
+    {
+      "name": "ort_conv2d_quickgelu_unit_alpha_silu_parity",
+      "provenance": {
+        "source": "onnxruntime/test/optimizer/graph_transform_test.cc",
+        "test": "GraphTransformationTests.WebGpuIm2ColConvQuickGeluUnitAlphaFusionMatchesUnfusedResults",
+        "notes": "With alpha 1, QuickGelu computes x*sigmoid(x); the expected values distinguish it from other activation settings."
+      },
+      "attrs": {
+        "activation": "QuickGelu",
+        "activation_params": [1],
+        "kernel_shape": [3, 3],
+        "pads": [1, 1, 1, 1],
+        "strides": [1, 1],
+        "dilations": [1, 1]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 4, 14, 14],
+          "data": { "kind": "fillFloat32", "sinStep": 0.29, "cosStep": 0.37, "scale": 3.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 4, 3, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.013, "scale": 0.5 }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [8],
+          "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.019, "scale": 0.5 }
+        }
+      },
+      "outputs": {
+        "y": { "dtype": "float32", "shape": [1, 8, 14, 14], "tolerance": 0.000001, "relTolerance": 0.00001 }
+      }
+    },
+    {
+      "name": "ort_conv2d_fused_quickgelu_parity",
+      "provenance": {
+        "source": "onnxruntime/core/providers/webgpu/nn/fuse_utils.cc",
+        "test": "GetActivationSnippet / ConvFp16Test.WebGpuSubgroupPointwiseNhwcFusedConvNewActivations",
+        "notes": "Transcribed from ORT's fused activation set. Formula: value * sigmoid(alpha*value), alpha default 1.702; alpha==1 is SiLU/Swish. alpha == 1 is a distinct case because the multiply drops out. Shapes follow RunWebGpuConvActivationParity (x [1,4,14,14], w [8,4,3,3], bias [8], input spanning [-3,3] so asymmetric activations see both signs); the upstream test uses random data, replaced here with a deterministic fill."
+      },
+      "attrs": {
+        "activation": "QuickGelu",
+        "kernel_shape": [3, 3],
+        "pads": [1, 1, 1, 1],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation_params": [1.702]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 4, 14, 14],
+          "data": { "kind": "fillFloat32", "sinStep": 0.23, "cosStep": 0.41, "scale": 3.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 4, 3, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.019, "cosStep": 0.031, "scale": 0.5 }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [8],
+          "data": { "kind": "fillFloat32", "sinStep": 0.007, "cosStep": 0.017, "scale": 0.5 }
+        }
+      },
+      "outputs": {
+        "y": { "dtype": "float32", "shape": [1, 8, 14, 14], "tolerance": 0.000001, "relTolerance": 0.00001 }
+      }
+    },
+    {
+      "name": "ort_conv2d_fused_elu_parity",
+      "provenance": {
+        "source": "onnxruntime/core/providers/webgpu/nn/fuse_utils.cc",
+        "test": "GetActivationSnippet / ConvFp16Test.WebGpuSubgroupPointwiseNhwcFusedConvNewActivations",
+        "notes": "Transcribed from ORT's fused activation set. Formula: select(alpha*(exp(v)-1), v, v >= 0). Shapes follow RunWebGpuConvActivationParity (x [1,4,14,14], w [8,4,3,3], bias [8], input spanning [-3,3] so asymmetric activations see both signs); the upstream test uses random data, replaced here with a deterministic fill."
+      },
+      "attrs": {
+        "activation": "Elu",
+        "kernel_shape": [3, 3],
+        "pads": [1, 1, 1, 1],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation_params": [0.5]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 4, 14, 14],
+          "data": { "kind": "fillFloat32", "sinStep": 0.29, "cosStep": 0.37, "scale": 3.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 4, 3, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.019, "cosStep": 0.031, "scale": 0.5 }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [8],
+          "data": { "kind": "fillFloat32", "sinStep": 0.007, "cosStep": 0.017, "scale": 0.5 }
+        }
+      },
+      "outputs": {
+        "y": { "dtype": "float32", "shape": [1, 8, 14, 14], "tolerance": 0.000001, "relTolerance": 0.00001 }
+      }
+    },
+    {
+      "name": "ort_conv2d_fused_gelu_parity",
+      "provenance": {
+        "source": "onnxruntime/core/providers/webgpu/nn/fuse_utils.cc",
+        "test": "GetActivationSnippet / ConvFp16Test.WebGpuSubgroupPointwiseNhwcFusedConvNewActivations",
+        "notes": "Transcribed from ORT's fused activation set. Formula: 0.5*v*(1+erf(v/sqrt2)); ORT encodes the approximation mode as activation_params[0] = 0 (erf) or 1 (tanh) and emits an Abramowitz-Stegun 7.1.26 erf helper as a module-scope declaration. Shapes follow RunWebGpuConvActivationParity (x [1,4,14,14], w [8,4,3,3], bias [8], input spanning [-3,3] so asymmetric activations see both signs); the upstream test uses random data, replaced here with a deterministic fill."
+      },
+      "attrs": {
+        "activation": "Gelu",
+        "kernel_shape": [3, 3],
+        "pads": [1, 1, 1, 1],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation_params": [0]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 4, 14, 14],
+          "data": { "kind": "fillFloat32", "sinStep": 0.31, "cosStep": 0.43, "scale": 3.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 4, 3, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.019, "cosStep": 0.031, "scale": 0.5 }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [8],
+          "data": { "kind": "fillFloat32", "sinStep": 0.007, "cosStep": 0.017, "scale": 0.5 }
+        }
+      },
+      "outputs": {
+        "y": { "dtype": "float32", "shape": [1, 8, 14, 14], "tolerance": 0.000001, "relTolerance": 0.00001 }
+      }
+    },
+    {
+      "name": "ort_conv2d_fused_gelutanh_parity",
+      "provenance": {
+        "source": "onnxruntime/core/providers/webgpu/nn/fuse_utils.cc",
+        "test": "GetActivationSnippet / ConvFp16Test.WebGpuSubgroupPointwiseNhwcFusedConvNewActivations",
+        "notes": "Transcribed from ORT's fused activation set. Formula: v*(0.5 + 0.5*tanh_safe(v*(0.035677408136300125*v*v + 0.79788456080286535))); ORT reaches it as Gelu with activation_params[0] == 1 or as FastGelu, and uses an overflow-safe tanh because WGSL tanh returns NaN for large inputs. Shapes follow RunWebGpuConvActivationParity (x [1,4,14,14], w [8,4,3,3], bias [8], input spanning [-3,3] so asymmetric activations see both signs); the upstream test uses random data, replaced here with a deterministic fill."
+      },
+      "attrs": {
+        "activation": "Gelu",
+        "kernel_shape": [3, 3],
+        "pads": [1, 1, 1, 1],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation_params": [1]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 4, 14, 14],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.47, "scale": 3.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 4, 3, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.019, "cosStep": 0.031, "scale": 0.5 }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [8],
+          "data": { "kind": "fillFloat32", "sinStep": 0.007, "cosStep": 0.017, "scale": 0.5 }
+        }
+      },
+      "outputs": {
+        "y": { "dtype": "float32", "shape": [1, 8, 14, 14], "tolerance": 0.000001, "relTolerance": 0.00001 }
+      }
+    },
+    {
+      "name": "ort_conv2d_fused_softplus_parity",
+      "provenance": {
+        "source": "onnxruntime/core/providers/webgpu/nn/fuse_utils.cc",
+        "test": "GetActivationSnippet / ConvFp16Test.WebGpuSubgroupPointwiseNhwcFusedConvNewActivations",
+        "notes": "Transcribed from ORT's fused activation set. Formula: max(v,0) + log(1 + exp(-abs(v))) -- the overflow-safe form. Shapes follow RunWebGpuConvActivationParity (x [1,4,14,14], w [8,4,3,3], bias [8], input spanning [-3,3] so asymmetric activations see both signs); the upstream test uses random data, replaced here with a deterministic fill."
+      },
+      "attrs": {
+        "activation": "Softplus",
+        "kernel_shape": [3, 3],
+        "pads": [1, 1, 1, 1],
+        "strides": [1, 1],
+        "dilations": [1, 1]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 4, 14, 14],
+          "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.53, "scale": 3.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 4, 3, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.019, "cosStep": 0.031, "scale": 0.5 }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [8],
+          "data": { "kind": "fillFloat32", "sinStep": 0.007, "cosStep": 0.017, "scale": 0.5 }
+        }
+      },
+      "outputs": {
+        "y": { "dtype": "float32", "shape": [1, 8, 14, 14], "tolerance": 0.000001, "relTolerance": 0.00001 }
+      }
+    },
+    {
+      "name": "ort_conv2d_fused_thresholdedrelu_parity",
+      "provenance": {
+        "source": "onnxruntime/core/providers/webgpu/nn/fuse_utils.cc",
+        "test": "GetActivationSnippet / ConvFp16Test.WebGpuSubgroupPointwiseNhwcFusedConvNewActivations",
+        "notes": "Transcribed from ORT's fused activation set. Formula: select(0, v, v > alpha), alpha default 1.0. Shapes follow RunWebGpuConvActivationParity (x [1,4,14,14], w [8,4,3,3], bias [8], input spanning [-3,3] so asymmetric activations see both signs); the upstream test uses random data, replaced here with a deterministic fill."
+      },
+      "attrs": {
+        "activation": "ThresholdedRelu",
+        "kernel_shape": [3, 3],
+        "pads": [1, 1, 1, 1],
+        "strides": [1, 1],
+        "dilations": [1, 1],
+        "activation_params": [0.6]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 4, 14, 14],
+          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.59, "scale": 3.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 4, 3, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.019, "cosStep": 0.031, "scale": 0.5 }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [8],
+          "data": { "kind": "fillFloat32", "sinStep": 0.007, "cosStep": 0.017, "scale": 0.5 }
+        }
+      },
+      "outputs": {
+        "y": { "dtype": "float32", "shape": [1, 8, 14, 14], "tolerance": 0.000001, "relTolerance": 0.00001 }
+      }
+    },
+    {
+      "name": "ort_conv2d_fused_erf_parity",
+      "provenance": {
+        "source": "onnxruntime/core/providers/webgpu/nn/fuse_utils.cc",
+        "test": "GetActivationSnippet / ConvFp16Test.WebGpuSubgroupPointwiseNhwcFusedConvNewActivations",
+        "notes": "Transcribed from ORT's fused activation set. Formula: erf(v), sharing the same A&S 7.1.26 helper as Gelu. Shapes follow RunWebGpuConvActivationParity (x [1,4,14,14], w [8,4,3,3], bias [8], input spanning [-3,3] so asymmetric activations see both signs); the upstream test uses random data, replaced here with a deterministic fill."
+      },
+      "attrs": {
+        "activation": "Erf",
+        "kernel_shape": [3, 3],
+        "pads": [1, 1, 1, 1],
+        "strides": [1, 1],
+        "dilations": [1, 1]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 4, 14, 14],
+          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.61, "scale": 3.0 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 4, 3, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.019, "cosStep": 0.031, "scale": 0.5 }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [8],
+          "data": { "kind": "fillFloat32", "sinStep": 0.007, "cosStep": 0.017, "scale": 0.5 }
+        }
+      },
+      "outputs": {
+        "y": { "dtype": "float32", "shape": [1, 8, 14, 14], "tolerance": 0.000001, "relTolerance": 0.00001 }
+      }
+    },
+    {
+      "name": "tile_partial_float32_b1c4m8_14x14_k3",
+      "provenance": {
+        "notes": "Checks partial K-tile accumulation for fused 3x3 and pointwise convolution, including odd output channels and spatial tails. ELU retains both signs and bias is added after the dot product."
+      },
+      "attrs": {
+        "activation": "Elu",
+        "activation_params": [0.5],
+        "kernel_shape": [3, 3],
+        "pads": [1, 1, 1, 1],
+        "strides": [1, 1],
+        "dilations": [1, 1]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 4, 14, 14],
+          "data": { "kind": "fillFloat32", "scale": 0.3, "sinStep": 0.17, "cosStep": 0.23 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 4, 3, 3],
+          "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.28, "cosStep": 0.30000000000000004 }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [8],
+          "data": { "kind": "fillFloat32", "scale": 0.3, "sinStep": 0.39, "cosStep": 0.37 }
+        }
+      },
+      "outputs": {
+        "y": { "dtype": "float32", "shape": [1, 8, 14, 14], "tolerance": 0.000001, "relTolerance": 0.00001 }
+      }
+    },
+    {
+      "name": "tile_partial_float32_b1c36m8_14x14_k1",
+      "provenance": {
+        "notes": "Checks partial K-tile accumulation for fused 3x3 and pointwise convolution, including odd output channels and spatial tails. ELU retains both signs and bias is added after the dot product."
+      },
+      "attrs": {
+        "activation": "Elu",
+        "activation_params": [0.5],
+        "kernel_shape": [1, 1],
+        "pads": [0, 0, 0, 0],
+        "strides": [1, 1],
+        "dilations": [1, 1]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 36, 14, 14],
+          "data": { "kind": "fillFloat32", "scale": 0.3, "sinStep": 0.17, "cosStep": 0.23 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [8, 36, 1, 1],
+          "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.28, "cosStep": 0.30000000000000004 }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [8],
+          "data": { "kind": "fillFloat32", "scale": 0.3, "sinStep": 0.39, "cosStep": 0.37 }
+        }
+      },
+      "outputs": {
+        "y": { "dtype": "float32", "shape": [1, 8, 14, 14], "tolerance": 0.000001, "relTolerance": 0.00001 }
+      }
+    },
+    {
+      "name": "tile_partial_float32_b1c5m17_13x15_k3",
+      "provenance": {
+        "notes": "Checks partial K-tile accumulation for fused 3x3 and pointwise convolution, including odd output channels and spatial tails. ELU retains both signs and bias is added after the dot product."
+      },
+      "attrs": {
+        "activation": "Elu",
+        "activation_params": [0.5],
+        "kernel_shape": [3, 3],
+        "pads": [1, 1, 1, 1],
+        "strides": [1, 1],
+        "dilations": [1, 1]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 5, 13, 15],
+          "data": { "kind": "fillFloat32", "scale": 0.3, "sinStep": 0.17, "cosStep": 0.23 }
+        },
+        "w": {
+          "dtype": "float32",
+          "shape": [17, 5, 3, 3],
+          "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.28, "cosStep": 0.30000000000000004 }
+        },
+        "bias": {
+          "dtype": "float32",
+          "shape": [17],
+          "data": { "kind": "fillFloat32", "scale": 0.3, "sinStep": 0.39, "cosStep": 0.37 }
+        }
+      },
+      "outputs": {
+        "y": { "dtype": "float32", "shape": [1, 17, 13, 15], "tolerance": 0.000001, "relTolerance": 0.00001 }
+      }
+    },
+    {
+      "name": "tile_partial_float16_b1c4m8_14x14_k3",
+      "provenance": {
+        "notes": "Checks partial K-tile accumulation for fused 3x3 and pointwise convolution, including odd output channels and spatial tails. ELU retains both signs and bias is added after the dot product."
+      },
+      "attrs": {
+        "activation": "Elu",
+        "activation_params": [0.5],
+        "kernel_shape": [3, 3],
+        "pads": [1, 1, 1, 1],
+        "strides": [1, 1],
+        "dilations": [1, 1]
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [1, 4, 14, 14],
+          "data": { "kind": "fillFloat32", "scale": 0.3, "sinStep": 0.17, "cosStep": 0.23 }
+        },
+        "w": {
+          "dtype": "float16",
+          "shape": [8, 4, 3, 3],
+          "data": { "kind": "fillFloat32", "scale": 0.2, "sinStep": 0.28, "cosStep": 0.30000000000000004 }
+        },
+        "bias": {
+          "dtype": "float16",
+          "shape": [8],
+          "data": { "kind": "fillFloat32", "scale": 0.3, "sinStep": 0.39, "cosStep": 0.37 }
+        }
+      },
+      "outputs": {
+        "y": { "dtype": "float16", "shape": [1, 8, 14, 14], "tolerance": 0.00001, "relTolerance": 0.00048828125 }
+      }
     }
   ]
 }