diff --git "a/build/webgpu/bench.json" "b/build/webgpu/bench.json"
--- "a/build/webgpu/bench.json"
+++ "b/build/webgpu/bench.json"
@@ -227,7 +227,7 @@
       "preset": "smoke",
       "provenance": {
         "source": "synthetic",
-        "notes": "A biased 1x1 convolution produces 64 register tiles, measuring the register-tiled route at its admission floor."
+        "notes": "Measures a biased, Relu-activated 1x1 FusedConv from 256 to 256 channels over a 32x32 float32 input."
       },
       "vars": {
         "batch": 1,
@@ -667,7 +667,7 @@
       "preset": "smoke",
       "provenance": {
         "source": "synthetic",
-        "notes": "A biased 3x3 convolution produces 64 register tiles, measuring the register-tiled im2col route at its admission floor."
+        "notes": "Measures a biased, Relu-activated 3x3 FusedConv (pad 1) from 256 to 256 channels over a 32x32 float32 input."
       },
       "vars": {
         "batch": 1,
@@ -933,7 +933,7 @@
       },
       "provenance": {
         "source": "synthetic",
-        "notes": "Measures the 32-workgroup implicit split-K preference boundary with output channels, spatial extent, or batch contributing to the grid."
+        "notes": "Measures Relu-activated convolution over output channels growing to 128, with 64 input channels, batch 1 and a fixed 32x32 spatial size."
       }
     },
     {
@@ -952,7 +952,7 @@
       },
       "provenance": {
         "source": "synthetic",
-        "notes": "Measures the 32-workgroup implicit split-K preference boundary with output channels, spatial extent, or batch contributing to the grid."
+        "notes": "Measures Relu-activated convolution over spatial width growing to 64, with 64 input and output channels and batch 1."
       }
     },
     {
@@ -971,7 +971,7 @@
       },
       "provenance": {
         "source": "synthetic",
-        "notes": "Measures the 32-workgroup implicit split-K preference boundary with output channels, spatial extent, or batch contributing to the grid."
+        "notes": "Measures Relu-activated convolution over batch growing to 2, with 64 input and output channels and a fixed 32x32 spatial size."
       }
     },
     {
@@ -1369,6 +1369,5629 @@
         "source": "synthetic",
         "notes": "Row-tile boundary sweep: full, partial, and device-dependent complete row bands share the same reduction and spatial extent."
       }
+    },
+    {
+      "name": "halfcols_float16_b1_c1_m512_h16_w32_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 1,
+        "outChannels": 512,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 1, 16, 32], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 1, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 16, 32], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c1_m512_h17_w31_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 1,
+        "outChannels": 512,
+        "inH": 17,
+        "inW": 31,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 1, 17, 31], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 1, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [512], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 17, 31], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c3_m512_h16_w32_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 3,
+        "outChannels": 512,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 3, 16, 32], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 3, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 16, 32], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c3_m512_h17_w31_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 3,
+        "outChannels": 512,
+        "inH": 17,
+        "inW": 31,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 3, 17, 31], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 3, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [512], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 17, 31], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c7_m512_h16_w32_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 7,
+        "outChannels": 512,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 7, 16, 32], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 7, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 16, 32], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c7_m512_h17_w31_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 7,
+        "outChannels": 512,
+        "inH": 17,
+        "inW": 31,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 7, 17, 31], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 7, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [512], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 17, 31], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c16_m512_h16_w32_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 16,
+        "outChannels": 512,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 16, 16, 32], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 16, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 16, 32], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c16_m512_h17_w31_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 16,
+        "outChannels": 512,
+        "inH": 17,
+        "inW": 31,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 16, 17, 31], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 16, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [512], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 17, 31], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c31_m512_h16_w32_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 31,
+        "outChannels": 512,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 31, 16, 32], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 31, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 16, 32], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c31_m512_h17_w31_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 31,
+        "outChannels": 512,
+        "inH": 17,
+        "inW": 31,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 31, 17, 31], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 31, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [512], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 17, 31], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c32_m512_h16_w32_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 32,
+        "outChannels": 512,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 32, 16, 32], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 32, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 16, 32], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c32_m512_h17_w31_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 32,
+        "outChannels": 512,
+        "inH": 17,
+        "inW": 31,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 32, 17, 31], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 32, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [512], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 17, 31], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c33_m512_h16_w32_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 33,
+        "outChannels": 512,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 33, 16, 32], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 33, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 16, 32], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c33_m512_h17_w31_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 33,
+        "outChannels": 512,
+        "inH": 17,
+        "inW": 31,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 33, 17, 31], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 33, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [512], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 17, 31], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c63_m512_h16_w32_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 63,
+        "outChannels": 512,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 63, 16, 32], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 63, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 16, 32], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c63_m512_h17_w31_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 63,
+        "outChannels": 512,
+        "inH": 17,
+        "inW": 31,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 63, 17, 31], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 63, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [512], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 17, 31], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c64_m512_h16_w32_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 64,
+        "outChannels": 512,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 64, 16, 32], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 64, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 16, 32], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c64_m512_h17_w31_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 64,
+        "outChannels": 512,
+        "inH": 17,
+        "inW": 31,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 64, 17, 31], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 64, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [512], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 17, 31], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c65_m512_h16_w32_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 65,
+        "outChannels": 512,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 65, 16, 32], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 65, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 16, 32], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c65_m512_h17_w31_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 65,
+        "outChannels": 512,
+        "inH": 17,
+        "inW": 31,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 65, 17, 31], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 65, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [512], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 17, 31], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c96_m512_h16_w32_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 96,
+        "outChannels": 512,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 96, 16, 32], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 96, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 16, 32], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c96_m512_h17_w31_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 96,
+        "outChannels": 512,
+        "inH": 17,
+        "inW": 31,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 96, 17, 31], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 96, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [512], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 17, 31], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c128_m512_h16_w32_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 128,
+        "outChannels": 512,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 128, 16, 32], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 128, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 16, 32], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c128_m512_h17_w31_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 128,
+        "outChannels": 512,
+        "inH": 17,
+        "inW": 31,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 128, 17, 31], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 128, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [512], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 17, 31], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c64_m384_h16_w32_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 64,
+        "outChannels": 384,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 64, 16, 32], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [384, 64, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 384, 16, 32], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c33_m384_h1_w513_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 33,
+        "outChannels": 384,
+        "inH": 1,
+        "inW": 513,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 33, 1, 513], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [384, 33, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [384], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [1, 384, 1, 513], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c33_m512_h1_w513_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 33,
+        "outChannels": 512,
+        "inH": 1,
+        "inW": 513,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 33, 1, 513], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 33, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [512], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 1, 513], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c64_m640_h16_w32_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 64,
+        "outChannels": 640,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 64, 16, 32], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [640, 64, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 640, 16, 32], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c33_m640_h1_w513_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 33,
+        "outChannels": 640,
+        "inH": 1,
+        "inW": 513,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 33, 1, 513], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [640, 33, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [640], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [1, 640, 1, 513], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b2_c64_m384_h16_w32_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 2,
+        "inChannels": 64,
+        "outChannels": 384,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [2, 64, 16, 32], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [384, 64, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [2, 384, 16, 32], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b2_c33_m384_h1_w513_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 2,
+        "inChannels": 33,
+        "outChannels": 384,
+        "inH": 1,
+        "inW": 513,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [2, 33, 1, 513], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [384, 33, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [384], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [2, 384, 1, 513], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b2_c64_m512_h16_w32_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 2,
+        "inChannels": 64,
+        "outChannels": 512,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [2, 64, 16, 32], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 64, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [2, 512, 16, 32], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b2_c33_m512_h1_w513_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 2,
+        "inChannels": 33,
+        "outChannels": 512,
+        "inH": 1,
+        "inW": 513,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [2, 33, 1, 513], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 33, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [512], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [2, 512, 1, 513], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b2_c64_m640_h16_w32_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 2,
+        "inChannels": 64,
+        "outChannels": 640,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [2, 64, 16, 32], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [640, 64, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [2, 640, 16, 32], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b2_c33_m640_h1_w513_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 2,
+        "inChannels": 33,
+        "outChannels": 640,
+        "inH": 1,
+        "inW": 513,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [2, 33, 1, 513], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [640, 33, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [640], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [2, 640, 1, 513], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b2_c1_m512_h1_w513_k3_bias0_none",
+      "preset": "model",
+      "vars": {
+        "batch": 2,
+        "inChannels": 1,
+        "outChannels": 512,
+        "inH": 1,
+        "inW": 513,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1] },
+      "inputs": {
+        "x": { "shape": [2, 1, 1, 513], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 1, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [2, 512, 1, 513], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c512_m512_h128_w128_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 512,
+        "outChannels": 512,
+        "inH": 128,
+        "inW": 128,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 512, 128, 128], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 512, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 128, 128], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c128_m512_h32_w64_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 128,
+        "outChannels": 512,
+        "inH": 32,
+        "inW": 64,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 128, 32, 64], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 128, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 32, 64], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float32_b1_c64_m512_h16_w32_k3_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 64,
+        "outChannels": 512,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 64, 16, 32], "dtype": "float32", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 64, 3, 3], "dtype": "float32", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 16, 32], "dtype": "float32" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c64_m512_h16_w32_k1_bias0_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 64,
+        "outChannels": 512,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 1,
+        "kernelW": 1,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 0,
+        "padW": 0
+      },
+      "attrs": { "strides": [1, 1], "pads": [0, 0, 0, 0], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 64, 16, 32], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 64, 1, 1], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 16, 32], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b2_c1_m512_h1_w513_k3_bias1_none",
+      "preset": "model",
+      "vars": {
+        "batch": 2,
+        "inChannels": 1,
+        "outChannels": 512,
+        "inH": 1,
+        "inW": 513,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1] },
+      "inputs": {
+        "x": { "shape": [2, 1, 1, 513], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 1, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [512], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [2, 512, 1, 513], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c512_m512_h128_w128_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 512,
+        "outChannels": 512,
+        "inH": 128,
+        "inW": 128,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 512, 128, 128], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 512, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [512], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 128, 128], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c128_m512_h32_w64_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 128,
+        "outChannels": 512,
+        "inH": 32,
+        "inW": 64,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 128, 32, 64], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 128, 3, 3], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [512], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 32, 64], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float32_b1_c64_m512_h16_w32_k3_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 64,
+        "outChannels": 512,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 3,
+        "kernelW": 3,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 1,
+        "padW": 1
+      },
+      "attrs": { "strides": [1, 1], "pads": [1, 1, 1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 64, 16, 32], "dtype": "float32", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 64, 3, 3], "dtype": "float32", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float32", "shape": [512], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 16, 32], "dtype": "float32" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "halfcols_float16_b1_c64_m512_h16_w32_k1_bias1_Relu",
+      "preset": "model",
+      "vars": {
+        "batch": 1,
+        "inChannels": 64,
+        "outChannels": 512,
+        "inH": 16,
+        "inW": 32,
+        "kernelH": 1,
+        "kernelW": 1,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 0,
+        "padW": 0
+      },
+      "attrs": { "strides": [1, 1], "pads": [0, 0, 0, 0], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 64, 16, 32], "dtype": "float16", "dist": "normal", "seed": 356, "scale": 0.2 },
+        "w": { "shape": [512, 64, 1, 1], "dtype": "float16", "dist": "normal", "seed": 357, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [512], "dist": "normal", "seed": 9833, "scale": 0.1 }
+      },
+      "outputs": { "y": { "shape": [1, 512, 16, 32], "dtype": "float16" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "gflops",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * dim(shapes.w, 1) * args.kernelH * args.kernelW"
+          }
+        ]
+      },
+      "tunableSpace": {}
+    },
+    {
+      "name": "fused-conv-f32-group8-k9-256c-64x64-nchw-fallback",
+      "preset": "stress",
+      "vars": {
+        "batch": 1,
+        "inChannels": 256,
+        "outChannels": 256,
+        "inH": 64,
+        "inW": 64,
+        "kernelH": 9,
+        "kernelW": 9,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 4,
+        "padW": 4,
+        "group": 8
+      },
+      "attrs": { "group": 8, "strides": [1, 1], "pads": [4, 4, 4, 4], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 256, 64, 64], "dtype": "float32", "dist": "normal", "seed": 4101, "scale": 0.2 },
+        "w": { "shape": [256, 32, 9, 9], "dtype": "float32", "dist": "normal", "seed": 4102, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 256, 64, 64], "dtype": "float32", "dist": "empty" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "bandwidth",
+            "name": "one-touch lower-bound BW",
+            "value": "4 * (numel(shapes.x) + numel(shapes.w) + numel(shapes.y))"
+          },
+          {
+            "type": "gflops",
+            "name": "nominal Conv FLOP/s",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * args.inChannels * args.kernelH * args.kernelW / args.group"
+          }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-conv-f16-group8-k9-256c-64x64-nchw-f16-fallback",
+      "preset": "stress",
+      "vars": {
+        "batch": 1,
+        "inChannels": 256,
+        "outChannels": 256,
+        "inH": 64,
+        "inW": 64,
+        "kernelH": 9,
+        "kernelW": 9,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 4,
+        "padW": 4,
+        "group": 8
+      },
+      "attrs": { "group": 8, "strides": [1, 1], "pads": [4, 4, 4, 4], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 256, 64, 64], "dtype": "float16", "dist": "normal", "seed": 4111, "scale": 0.2 },
+        "w": { "shape": [256, 32, 9, 9], "dtype": "float16", "dist": "normal", "seed": 4112, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 256, 64, 64], "dtype": "float16", "dist": "empty" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "bandwidth",
+            "name": "one-touch lower-bound BW",
+            "value": "2 * (numel(shapes.x) + numel(shapes.w) + numel(shapes.y))"
+          },
+          {
+            "type": "gflops",
+            "name": "nominal Conv FLOP/s",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * args.inChannels * args.kernelH * args.kernelW / args.group"
+          }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-conv-f32-group4-k11-dilated-128c-48x48-nchw-fallback",
+      "preset": "stress",
+      "vars": {
+        "batch": 1,
+        "inChannels": 128,
+        "outChannels": 128,
+        "inH": 48,
+        "inW": 48,
+        "kernelH": 11,
+        "kernelW": 11,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 10,
+        "padW": 10,
+        "dilationH": 2,
+        "dilationW": 2,
+        "group": 4
+      },
+      "attrs": { "group": 4, "strides": [1, 1], "dilations": [2, 2], "pads": [10, 10, 10, 10], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 128, 48, 48], "dtype": "float32", "dist": "normal", "seed": 4121, "scale": 0.2 },
+        "w": { "shape": [128, 32, 11, 11], "dtype": "float32", "dist": "normal", "seed": 4122, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 128, 48, 48], "dtype": "float32", "dist": "empty" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "bandwidth",
+            "name": "one-touch lower-bound BW",
+            "value": "4 * (numel(shapes.x) + numel(shapes.w) + numel(shapes.y))"
+          },
+          {
+            "type": "gflops",
+            "name": "nominal Conv FLOP/s",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * args.inChannels * args.kernelH * args.kernelW / args.group"
+          }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 32 input and 32 output channels per group, dilation 2x2, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-conv-f32-group8-k9-256c-64x63-w4-boundary",
+      "preset": "stress",
+      "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."
+      },
+      "vars": {
+        "batch": 1,
+        "inChannels": 256,
+        "outChannels": 256,
+        "inH": 64,
+        "inW": 63,
+        "kernelH": 9,
+        "kernelW": 9,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 4,
+        "padW": 4,
+        "group": 8
+      },
+      "attrs": { "group": 8, "strides": [1, 1], "pads": [4, 4, 4, 4], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 256, 64, 63], "dtype": "float32", "dist": "normal", "seed": 9131, "scale": 0.2 },
+        "w": { "shape": [256, 32, 9, 9], "dtype": "float32", "dist": "normal", "seed": 9132, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 256, 64, 63], "dtype": "float32", "dist": "empty" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "bandwidth",
+            "name": "one-touch lower-bound BW",
+            "value": "4 * (numel(shapes.x) + numel(shapes.w) + numel(shapes.y))"
+          },
+          {
+            "type": "gflops",
+            "name": "nominal Conv FLOP/s",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * args.inChannels * args.kernelH * args.kernelW / args.group"
+          }
+        ]
+      }
+    },
+    {
+      "name": "fused-conv-f32-group4-k11-dilated6-128c-48x48-register-span-cap",
+      "preset": "stress",
+      "vars": {
+        "batch": 1,
+        "inChannels": 128,
+        "outChannels": 128,
+        "inH": 48,
+        "inW": 48,
+        "kernelH": 11,
+        "kernelW": 11,
+        "strideH": 1,
+        "strideW": 1,
+        "padH": 30,
+        "padW": 30,
+        "dilationH": 6,
+        "dilationW": 6,
+        "group": 4
+      },
+      "attrs": { "group": 4, "strides": [1, 1], "dilations": [6, 6], "pads": [30, 30, 30, 30], "activation": "Relu" },
+      "inputs": {
+        "x": { "shape": [1, 128, 48, 48], "dtype": "float32", "dist": "normal", "seed": 4121, "scale": 0.2 },
+        "w": { "shape": [128, 32, 11, 11], "dtype": "float32", "dist": "normal", "seed": 4122, "scale": 0.02 }
+      },
+      "outputs": { "y": { "shape": [1, 128, 48, 48], "dtype": "float32", "dist": "empty" } },
+      "bench": {
+        "metrics": [
+          {
+            "type": "bandwidth",
+            "name": "one-touch lower-bound BW",
+            "value": "4 * (numel(shapes.x) + numel(shapes.w) + numel(shapes.y))"
+          },
+          {
+            "type": "gflops",
+            "name": "nominal Conv FLOP/s",
+            "value": "2 * args.batch * args.outChannels * dim(shapes.y, 2) * dim(shapes.y, 3) * args.inChannels * args.kernelH * args.kernelW / args.group"
+          }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 32 input and 32 output channels per group, dilation 6x6, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m16_32x33_k3x3_d1x1_s1x1_bias0_asym0-Tanh",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [1, 1, 1, 1], "strides": [1, 1], "dilations": [1, 1], "activation": "Tanh" },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 8, 3, 3], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 3x3 convolution with 8 input and 16 output channels per group, dilation 1x1, stride 1x1, Tanh activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m16_32x33_k3x3_d2x2_s1x1_bias0_asym0-HardSigmoid",
+      "preset": "model",
+      "attrs": {
+        "group": 4,
+        "pads": [2, 2, 2, 2],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "HardSigmoid",
+        "activation_params": [0.2, 0.5]
+      },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 8, 3, 3], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 3x3 convolution with 8 input and 16 output channels per group, dilation 2x2, stride 1x1, HardSigmoid activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m16_32x33_k3x3_d3x3_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [3, 3, 3, 3], "strides": [1, 1], "dilations": [3, 3], "activation": "HardSwish" },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 8, 3, 3], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 3x3 convolution with 8 input and 16 output channels per group, dilation 3x3, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m16_32x33_k5x5_d1x1_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 8, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m16_32x33_k5x5_d2x2_s1x1_bias0_asym0-Clip",
+      "preset": "model",
+      "attrs": {
+        "group": 4,
+        "pads": [4, 4, 4, 4],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "Clip",
+        "activation_params": [-0.15, 0.2]
+      },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 8, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 8 input and 16 output channels per group, dilation 2x2, stride 1x1, Clip activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m16_32x33_k5x5_d3x3_s1x1_bias0_asym0-identity",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [6, 6, 6, 6], "strides": [1, 1], "dilations": [3, 3] },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 8, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 8 input and 16 output channels per group, dilation 3x3, stride 1x1, identity activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m16_32x33_k7x7_d1x1_s1x1_bias0_asym0-identity",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [3, 3, 3, 3], "strides": [1, 1], "dilations": [1, 1] },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 8, 7, 7], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 8 input and 16 output channels per group, dilation 1x1, stride 1x1, identity activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m16_32x33_k7x7_d2x2_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [6, 6, 6, 6], "strides": [1, 1], "dilations": [2, 2], "activation": "Relu" },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 8, 7, 7], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 8 input and 16 output channels per group, dilation 2x2, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m16_32x33_k7x7_d3x3_s1x1_bias0_asym0-LeakyRelu",
+      "preset": "model",
+      "attrs": {
+        "group": 4,
+        "pads": [9, 9, 9, 9],
+        "strides": [1, 1],
+        "dilations": [3, 3],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 8, 7, 7], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 8 input and 16 output channels per group, dilation 3x3, stride 1x1, LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m16_32x33_k9x9_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 8, 9, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m16_32x33_k9x9_d2x2_s1x1_bias0_asym0-Sigmoid",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [8, 8, 8, 8], "strides": [1, 1], "dilations": [2, 2], "activation": "Sigmoid" },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 8, 9, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 9x9 convolution with 8 input and 16 output channels per group, dilation 2x2, stride 1x1, Sigmoid activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m16_32x33_k9x9_d3x3_s1x1_bias0_asym0-Tanh",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [12, 12, 12, 12], "strides": [1, 1], "dilations": [3, 3], "activation": "Tanh" },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 8, 9, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 9x9 convolution with 8 input and 16 output channels per group, dilation 3x3, stride 1x1, Tanh activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m16_32x33_k11x11_d1x1_s1x1_bias0_asym0-Tanh",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [5, 5, 5, 5], "strides": [1, 1], "dilations": [1, 1], "activation": "Tanh" },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 8, 11, 11], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 8 input and 16 output channels per group, dilation 1x1, stride 1x1, Tanh activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m16_32x33_k11x11_d2x2_s1x1_bias0_asym0-HardSigmoid",
+      "preset": "model",
+      "attrs": {
+        "group": 4,
+        "pads": [10, 10, 10, 10],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "HardSigmoid",
+        "activation_params": [0.2, 0.5]
+      },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 8, 11, 11], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 8 input and 16 output channels per group, dilation 2x2, stride 1x1, HardSigmoid activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m16_32x33_k11x11_d3x3_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "attrs": {
+        "group": 4,
+        "pads": [15, 15, 15, 15],
+        "strides": [1, 1],
+        "dilations": [3, 3],
+        "activation": "HardSwish"
+      },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 8, 11, 11], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 8 input and 16 output channels per group, dilation 3x3, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m16_32x33_k3x3_d1x1_s1x1_bias0_asym0-Tanh",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [1, 1, 1, 1], "strides": [1, 1], "dilations": [1, 1], "activation": "Tanh" },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 8, 3, 3], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 3x3 convolution with 8 input and 16 output channels per group, dilation 1x1, stride 1x1, Tanh activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m16_32x33_k3x3_d2x2_s1x1_bias0_asym0-HardSigmoid",
+      "preset": "model",
+      "attrs": {
+        "group": 4,
+        "pads": [2, 2, 2, 2],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "HardSigmoid",
+        "activation_params": [0.2, 0.5]
+      },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 8, 3, 3], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 3x3 convolution with 8 input and 16 output channels per group, dilation 2x2, stride 1x1, HardSigmoid activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m16_32x33_k3x3_d3x3_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [3, 3, 3, 3], "strides": [1, 1], "dilations": [3, 3], "activation": "HardSwish" },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 8, 3, 3], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 3x3 convolution with 8 input and 16 output channels per group, dilation 3x3, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m16_32x33_k5x5_d1x1_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 8, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m16_32x33_k5x5_d2x2_s1x1_bias0_asym0-Clip",
+      "preset": "model",
+      "attrs": {
+        "group": 4,
+        "pads": [4, 4, 4, 4],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "Clip",
+        "activation_params": [-0.15, 0.2]
+      },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 8, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 8 input and 16 output channels per group, dilation 2x2, stride 1x1, Clip activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m16_32x33_k5x5_d3x3_s1x1_bias0_asym0-identity",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [6, 6, 6, 6], "strides": [1, 1], "dilations": [3, 3] },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 8, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 8 input and 16 output channels per group, dilation 3x3, stride 1x1, identity activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m16_32x33_k7x7_d1x1_s1x1_bias0_asym0-identity",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [3, 3, 3, 3], "strides": [1, 1], "dilations": [1, 1] },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 8, 7, 7], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 8 input and 16 output channels per group, dilation 1x1, stride 1x1, identity activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m16_32x33_k7x7_d2x2_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [6, 6, 6, 6], "strides": [1, 1], "dilations": [2, 2], "activation": "Relu" },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 8, 7, 7], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 8 input and 16 output channels per group, dilation 2x2, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m16_32x33_k7x7_d3x3_s1x1_bias0_asym0-LeakyRelu",
+      "preset": "model",
+      "attrs": {
+        "group": 4,
+        "pads": [9, 9, 9, 9],
+        "strides": [1, 1],
+        "dilations": [3, 3],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 8, 7, 7], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 8 input and 16 output channels per group, dilation 3x3, stride 1x1, LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m16_32x33_k9x9_d1x1_s1x1_bias0_asym0-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 8, 9, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m16_32x33_k9x9_d2x2_s1x1_bias0_asym0-Sigmoid",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [8, 8, 8, 8], "strides": [1, 1], "dilations": [2, 2], "activation": "Sigmoid" },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 8, 9, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 9x9 convolution with 8 input and 16 output channels per group, dilation 2x2, stride 1x1, Sigmoid activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m16_32x33_k9x9_d3x3_s1x1_bias0_asym0-Tanh",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [12, 12, 12, 12], "strides": [1, 1], "dilations": [3, 3], "activation": "Tanh" },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 8, 9, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 9x9 convolution with 8 input and 16 output channels per group, dilation 3x3, stride 1x1, Tanh activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m16_32x33_k11x11_d1x1_s1x1_bias0_asym0-Tanh",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [5, 5, 5, 5], "strides": [1, 1], "dilations": [1, 1], "activation": "Tanh" },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 8, 11, 11], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 8 input and 16 output channels per group, dilation 1x1, stride 1x1, Tanh activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m16_32x33_k11x11_d2x2_s1x1_bias0_asym0-HardSigmoid",
+      "preset": "model",
+      "attrs": {
+        "group": 4,
+        "pads": [10, 10, 10, 10],
+        "strides": [1, 1],
+        "dilations": [2, 2],
+        "activation": "HardSigmoid",
+        "activation_params": [0.2, 0.5]
+      },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 8, 11, 11], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 8 input and 16 output channels per group, dilation 2x2, stride 1x1, HardSigmoid activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m16_32x33_k11x11_d3x3_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "attrs": {
+        "group": 4,
+        "pads": [15, 15, 15, 15],
+        "strides": [1, 1],
+        "dilations": [3, 3],
+        "activation": "HardSwish"
+      },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 32, 32, 33], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 8, 11, 11], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 8 input and 16 output channels per group, dilation 3x3, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c2_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-identity",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1] },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 8, 48, 48], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 2, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 2 input and 32 output channels per group, dilation 1x1, stride 1x1, identity activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c2_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 8, 48, 48], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 2, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 2 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c2_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "HardSwish" },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 8, 48, 48], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 2, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 2 input and 32 output channels per group, dilation 1x1, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-identity",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1] },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 32, 48, 48], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 8, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 8, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 8, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-identity",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1] },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 64, 48, 48], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 16, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 16, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 16, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-identity",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1] },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 128, 48, 48], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 32, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 32, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 32, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c2_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-identity",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1] },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 8, 48, 48], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 2, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 2 input and 32 output channels per group, dilation 1x1, stride 1x1, identity activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c2_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 8, 48, 48], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 2, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 2 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c2_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "HardSwish" },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 8, 48, 48], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 2, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 2 input and 32 output channels per group, dilation 1x1, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-identity",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1] },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 32, 48, 48], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 8, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 8, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c8_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 8, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-identity",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1] },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 64, 48, 48], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 16, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 16, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 16, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-identity",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1] },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 128, 48, 48], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 32, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 32, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 32, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-identity",
+      "preset": "model",
+      "attrs": { "group": 3, "pads": [10, 10, 11, 12], "strides": [1, 1], "dilations": [2, 2] },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [2, 9, 17, 29], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float32", "shape": [18, 3, 11, 11], "dist": "normal", "seed": 602, "scale": 0.4 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float32", "shape": [18, 3, 11, 11], "dist": "normal", "seed": 602, "scale": 0.4 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float32", "shape": [18, 3, 11, 11], "dist": "normal", "seed": 602, "scale": 0.4 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-Sigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float32", "shape": [18, 3, 11, 11], "dist": "normal", "seed": 602, "scale": 0.4 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-Tanh",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float32", "shape": [18, 3, 11, 11], "dist": "normal", "seed": 602, "scale": 0.4 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-HardSigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float32", "shape": [18, 3, 11, 11], "dist": "normal", "seed": 602, "scale": 0.4 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float32", "shape": [18, 3, 11, 11], "dist": "normal", "seed": 602, "scale": 0.4 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-Clip",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float32", "shape": [18, 3, 11, 11], "dist": "normal", "seed": 602, "scale": 0.4 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-identity",
+      "preset": "model",
+      "attrs": { "group": 3, "pads": [10, 10, 11, 12], "strides": [1, 1], "dilations": [2, 2] },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [2, 9, 17, 29], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float16", "shape": [18, 3, 11, 11], "dist": "normal", "seed": 602, "scale": 0.4 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float16", "shape": [18, 3, 11, 11], "dist": "normal", "seed": 602, "scale": 0.4 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float16", "shape": [18, 3, 11, 11], "dist": "normal", "seed": 602, "scale": 0.4 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-Sigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float16", "shape": [18, 3, 11, 11], "dist": "normal", "seed": 602, "scale": 0.4 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-Tanh",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float16", "shape": [18, 3, 11, 11], "dist": "normal", "seed": 602, "scale": 0.4 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-HardSigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float16", "shape": [18, 3, 11, 11], "dist": "normal", "seed": 602, "scale": 0.4 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float16", "shape": [18, 3, 11, 11], "dist": "normal", "seed": 602, "scale": 0.4 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k11x11_d2x2_s1x1_bias0_asym1-Clip",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float16", "shape": [18, 3, 11, 11], "dist": "normal", "seed": 602, "scale": 0.4 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g16_c1_m2_33x35_k3x3_d2x2_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [32, 1, 3, 3], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 33, 35] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 3x3 convolution with 1 input and 2 output channels per group, dilation 2x2, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g16_c1_m2_33x35_k3x3_d3x3_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "attrs": { "group": 16, "pads": [3, 3, 3, 3], "strides": [1, 1], "dilations": [3, 3], "activation": "HardSwish" },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 16, 33, 35], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [32, 1, 3, 3], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 33, 35] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 3x3 convolution with 1 input and 2 output channels per group, dilation 3x3, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g16_c1_m2_33x35_k7x7_d2x2_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "attrs": { "group": 16, "pads": [6, 6, 6, 6], "strides": [1, 1], "dilations": [2, 2], "activation": "HardSwish" },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 16, 33, 35], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [32, 1, 7, 7], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 33, 35] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 1 input and 2 output channels per group, dilation 2x2, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g16_c1_m2_33x35_k7x7_d3x3_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "attrs": { "group": 16, "pads": [9, 9, 9, 9], "strides": [1, 1], "dilations": [3, 3], "activation": "HardSwish" },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 16, 33, 35], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [32, 1, 7, 7], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 33, 35] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 1 input and 2 output channels per group, dilation 3x3, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g16_c1_m2_33x35_k11x11_d2x2_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [32, 1, 11, 11], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 33, 35] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g16_c1_m2_33x35_k11x11_d3x3_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [32, 1, 11, 11], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 33, 35] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g16_c1_m2_33x35_k3x3_d2x2_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [32, 1, 3, 3], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 33, 35] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 3x3 convolution with 1 input and 2 output channels per group, dilation 2x2, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g16_c1_m2_33x35_k3x3_d3x3_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "attrs": { "group": 16, "pads": [3, 3, 3, 3], "strides": [1, 1], "dilations": [3, 3], "activation": "HardSwish" },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 16, 33, 35], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [32, 1, 3, 3], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 33, 35] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 3x3 convolution with 1 input and 2 output channels per group, dilation 3x3, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g16_c1_m2_33x35_k7x7_d2x2_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "attrs": { "group": 16, "pads": [6, 6, 6, 6], "strides": [1, 1], "dilations": [2, 2], "activation": "HardSwish" },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 16, 33, 35], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [32, 1, 7, 7], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 33, 35] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 1 input and 2 output channels per group, dilation 2x2, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g16_c1_m2_33x35_k7x7_d3x3_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "attrs": { "group": 16, "pads": [9, 9, 9, 9], "strides": [1, 1], "dilations": [3, 3], "activation": "HardSwish" },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 16, 33, 35], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [32, 1, 7, 7], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 33, 35] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 7x7 convolution with 1 input and 2 output channels per group, dilation 3x3, stride 1x1, HardSwish activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g16_c1_m2_33x35_k11x11_d2x2_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [32, 1, 11, 11], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 33, 35] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g16_c1_m2_33x35_k11x11_d3x3_s1x1_bias0_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [32, 1, 11, 11], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 33, 35] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_33x35_k3x11_d1x2_s1x1_bias0_asym1-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [24, 3, 3, 11], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 34, 37] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_33x35_k11x3_d2x1_s1x1_bias0_asym1-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [24, 3, 11, 3], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 34, 37] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_33x35_k5x9_d1x3_s2x1_bias0_asym1-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [24, 3, 5, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 17, 37] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_33x35_k7x5_d2x2_s1x2_bias0_asym1-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [24, 3, 7, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 34, 19] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_33x35_k11x11_d6x6_s1x1_bias0_asym1-LeakyRelu",
+      "preset": "model",
+      "attrs": {
+        "group": 4,
+        "pads": [30, 30, 31, 32],
+        "strides": [1, 1],
+        "dilations": [6, 6],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 12, 33, 35], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [24, 3, 11, 11], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 34, 37] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 3 input and 6 output channels per group, dilation 6x6, stride 1x1, LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c3_m6_33x35_k3x11_d1x2_s1x1_bias0_asym1-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [24, 3, 3, 11], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 24, 34, 37] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c3_m6_33x35_k11x3_d2x1_s1x1_bias0_asym1-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [24, 3, 11, 3], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 24, 34, 37] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c3_m6_33x35_k5x9_d1x3_s2x1_bias0_asym1-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [24, 3, 5, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 24, 17, 37] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c3_m6_33x35_k7x5_d2x2_s1x2_bias0_asym1-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [24, 3, 7, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 24, 34, 19] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c3_m6_33x35_k11x11_d6x6_s1x1_bias0_asym1-LeakyRelu",
+      "preset": "model",
+      "attrs": {
+        "group": 4,
+        "pads": [30, 30, 31, 32],
+        "strides": [1, 1],
+        "dilations": [6, 6],
+        "activation": "LeakyRelu",
+        "activation_params": [0.2]
+      },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [1, 12, 33, 35], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [24, 3, 11, 11], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 24, 34, 37] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 11x11 convolution with 3 input and 6 output channels per group, dilation 6x6, stride 1x1, LeakyRelu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x7_k5x5_d1x2_s1x1_bias0_asym0-Sigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [24, 3, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 7] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x8_k5x5_d1x2_s1x1_bias0_asym0-Sigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [24, 3, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 8] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      },
+      "tunableSpace": {
+        "GROUPED_ROW_LOOP_MIN_KERNEL_AREA": [0, 64, 121],
+        "GROUPED_ROW_LOOP_MAX_SMALL_CHANNEL_BYTES": [0, 32, 64],
+        "GROUPED_DILATED_MIN_PLAIN_SPAN": [0, 32]
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x9_k5x5_d1x2_s1x1_bias0_asym0-Sigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [24, 3, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 9] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x11_k5x5_d1x3_s1x1_bias0_asym0-Sigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [24, 3, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 11] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x12_k5x5_d1x3_s1x1_bias0_asym0-Sigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [24, 3, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 12] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x13_k5x5_d1x3_s1x1_bias0_asym0-Sigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [24, 3, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 13] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m16_32x33_k3x3_d1x1_s1x1_bias1_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 16, 3, 3], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float32", "shape": [64], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m16_32x33_k3x3_d2x2_s1x1_bias1_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 16, 3, 3], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float32", "shape": [64], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m16_32x33_k5x5_d1x1_s1x1_bias1_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 16, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float32", "shape": [64], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m16_32x33_k5x5_d2x2_s1x1_bias1_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 16, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float32", "shape": [64], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m16_32x33_k7x7_d1x1_s1x1_bias1_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 16, 7, 7], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float32", "shape": [64], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m16_32x33_k7x7_d2x2_s1x1_bias1_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 16, 7, 7], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float32", "shape": [64], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m16_32x33_k3x3_d1x1_s1x1_bias1_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 16, 3, 3], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [64], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m16_32x33_k3x3_d2x2_s1x1_bias1_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 16, 3, 3], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [64], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m16_32x33_k5x5_d1x1_s1x1_bias1_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 16, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [64], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m16_32x33_k5x5_d2x2_s1x1_bias1_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 16, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [64], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m16_32x33_k7x7_d1x1_s1x1_bias1_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 16, 7, 7], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [64], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m16_32x33_k7x7_d2x2_s1x1_bias1_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 16, 7, 7], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [64], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 33] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c7_m32_48x48_k4x6_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 7, 4, 6], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c7_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 28, 48, 48], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 7, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 5x5 convolution with 7 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c7_m32_48x48_k3x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 7, 3, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c7_m32_48x48_k6x6_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "attrs": { "group": 4, "pads": [2, 2, 2, 2], "strides": [1, 1], "dilations": [1, 1], "activation": "Relu" },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 28, 48, 48], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 7, 6, 6], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 6x6 convolution with 7 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m32_48x48_k4x6_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 8, 4, 6], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m32_48x48_k3x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 8, 3, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c8_m32_48x48_k6x6_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 8, 6, 6], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 6x6 convolution with 8 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c9_m32_48x48_k4x6_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 9, 4, 6], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c9_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 9, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c9_m32_48x48_k3x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 9, 3, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c9_m32_48x48_k6x6_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 9, 6, 6], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 6x6 convolution with 9 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k4x6_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 16, 4, 6], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k3x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 16, 3, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k6x6_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 16, 6, 6], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 6x6 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k4x6_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 32, 4, 6], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k3x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 32, 3, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k6x6_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 32, 6, 6], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 6x6 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c15_m32_48x48_k4x6_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 15, 4, 6], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c15_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 15, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c15_m32_48x48_k3x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 15, 3, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c15_m32_48x48_k6x6_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 15, 6, 6], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 6x6 convolution with 15 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k4x6_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 16, 4, 6], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k3x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 16, 3, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k6x6_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 16, 6, 6], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 6x6 convolution with 16 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c17_m32_48x48_k4x6_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 17, 4, 6], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c17_m32_48x48_k5x5_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 17, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c17_m32_48x48_k3x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 17, 3, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c17_m32_48x48_k6x6_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 17, 6, 6], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 6x6 convolution with 17 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k4x6_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 32, 4, 6], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k3x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 32, 3, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k6x6_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 32, 6, 6], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "provenance": {
+        "notes": "Grouped 6x6 convolution with 32 input and 32 output channels per group, dilation 1x1, stride 1x1, Relu activation. Checks shared row-window reuse and output coverage."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-identity",
+      "preset": "model",
+      "attrs": { "group": 3, "pads": [6, 6, 7, 8], "strides": [1, 1], "dilations": [2, 2] },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [2, 9, 17, 29], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float32", "shape": [18, 3, 7, 7], "dist": "normal", "seed": 602, "scale": 0.4 },
+        "bias": { "dtype": "float32", "shape": [18], "dist": "normal", "seed": 603, "scale": 0.5 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float32", "shape": [18, 3, 7, 7], "dist": "normal", "seed": 602, "scale": 0.4 },
+        "bias": { "dtype": "float32", "shape": [18], "dist": "normal", "seed": 603, "scale": 0.5 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float32", "shape": [18, 3, 7, 7], "dist": "normal", "seed": 602, "scale": 0.4 },
+        "bias": { "dtype": "float32", "shape": [18], "dist": "normal", "seed": 603, "scale": 0.5 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-Sigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float32", "shape": [18, 3, 7, 7], "dist": "normal", "seed": 602, "scale": 0.4 },
+        "bias": { "dtype": "float32", "shape": [18], "dist": "normal", "seed": 603, "scale": 0.5 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-Tanh",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float32", "shape": [18, 3, 7, 7], "dist": "normal", "seed": 602, "scale": 0.4 },
+        "bias": { "dtype": "float32", "shape": [18], "dist": "normal", "seed": 603, "scale": 0.5 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-HardSigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float32", "shape": [18, 3, 7, 7], "dist": "normal", "seed": 602, "scale": 0.4 },
+        "bias": { "dtype": "float32", "shape": [18], "dist": "normal", "seed": 603, "scale": 0.5 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float32", "shape": [18, 3, 7, 7], "dist": "normal", "seed": 602, "scale": 0.4 },
+        "bias": { "dtype": "float32", "shape": [18], "dist": "normal", "seed": 603, "scale": 0.5 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-Clip",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float32", "shape": [18, 3, 7, 7], "dist": "normal", "seed": 602, "scale": 0.4 },
+        "bias": { "dtype": "float32", "shape": [18], "dist": "normal", "seed": 603, "scale": 0.5 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-identity",
+      "preset": "model",
+      "attrs": { "group": 3, "pads": [6, 6, 7, 8], "strides": [1, 1], "dilations": [2, 2] },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [2, 9, 17, 29], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float16", "shape": [18, 3, 7, 7], "dist": "normal", "seed": 602, "scale": 0.4 },
+        "bias": { "dtype": "float16", "shape": [18], "dist": "normal", "seed": 603, "scale": 0.5 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float16", "shape": [18, 3, 7, 7], "dist": "normal", "seed": 602, "scale": 0.4 },
+        "bias": { "dtype": "float16", "shape": [18], "dist": "normal", "seed": 603, "scale": 0.5 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float16", "shape": [18, 3, 7, 7], "dist": "normal", "seed": 602, "scale": 0.4 },
+        "bias": { "dtype": "float16", "shape": [18], "dist": "normal", "seed": 603, "scale": 0.5 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-Sigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float16", "shape": [18, 3, 7, 7], "dist": "normal", "seed": 602, "scale": 0.4 },
+        "bias": { "dtype": "float16", "shape": [18], "dist": "normal", "seed": 603, "scale": 0.5 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-Tanh",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float16", "shape": [18, 3, 7, 7], "dist": "normal", "seed": 602, "scale": 0.4 },
+        "bias": { "dtype": "float16", "shape": [18], "dist": "normal", "seed": 603, "scale": 0.5 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-HardSigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float16", "shape": [18, 3, 7, 7], "dist": "normal", "seed": 602, "scale": 0.4 },
+        "bias": { "dtype": "float16", "shape": [18], "dist": "normal", "seed": 603, "scale": 0.5 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float16", "shape": [18, 3, 7, 7], "dist": "normal", "seed": 602, "scale": 0.4 },
+        "bias": { "dtype": "float16", "shape": [18], "dist": "normal", "seed": 603, "scale": 0.5 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m6_17x29_k7x7_d2x2_s1x1_bias1_asym1-Clip",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 2 },
+        "w": { "dtype": "float16", "shape": [18, 3, 7, 7], "dist": "normal", "seed": 602, "scale": 0.4 },
+        "bias": { "dtype": "float16", "shape": [18], "dist": "normal", "seed": 603, "scale": 0.5 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 18, 18, 31] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g16_c1_m1_33x35_k3x3_d2x2_s1x1_bias1_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [16, 1, 3, 3], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float32", "shape": [16], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 16, 33, 35] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g16_c1_m2_33x35_k3x3_d2x2_s1x1_bias1_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [32, 1, 3, 3], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float32", "shape": [32], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 33, 35] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g16_c1_m1_33x35_k3x3_d2x2_s1x1_bias1_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [16, 1, 3, 3], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [16], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 16, 33, 35] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g16_c1_m2_33x35_k3x3_d2x2_s1x1_bias1_asym0-HardSwish",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [32, 1, 3, 3], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [32], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 33, 35] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m1_17x29_k5x5_d1x1_s1x1_bias1_asym0-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [3, 3, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float32", "shape": [3], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 17, 29] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m3_17x29_k5x5_d1x1_s1x1_bias1_asym0-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [9, 3, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float32", "shape": [9], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 9, 17, 29] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b2_g3_c3_m5_17x29_k5x5_d1x1_s1x1_bias1_asym0-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [15, 3, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float32", "shape": [15], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 15, 17, 29] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m1_17x29_k5x5_d1x1_s1x1_bias1_asym0-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [3, 3, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [3], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 3, 17, 29] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m3_17x29_k5x5_d1x1_s1x1_bias1_asym0-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [9, 3, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [9], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 9, 17, 29] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b2_g3_c3_m5_17x29_k5x5_d1x1_s1x1_bias1_asym0-LeakyRelu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [15, 3, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [15], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 15, 17, 29] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x7_k5x5_d1x2_s1x1_bias1_asym0-Sigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [24, 3, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float32", "shape": [24], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 7] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x8_k5x5_d1x2_s1x1_bias1_asym0-Sigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [24, 3, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float32", "shape": [24], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 8] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x9_k5x5_d1x2_s1x1_bias1_asym0-Sigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [24, 3, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float32", "shape": [24], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 9] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x11_k5x5_d1x3_s1x1_bias1_asym0-Sigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [24, 3, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float32", "shape": [24], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 11] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x12_k5x5_d1x3_s1x1_bias1_asym0-Sigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [24, 3, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float32", "shape": [24], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 12] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c3_m6_40x13_k5x5_d1x3_s1x1_bias1_asym0-Sigmoid",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [24, 3, 5, 5], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float32", "shape": [24], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 40, 13] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k7x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 16, 7, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k8x8_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 16, 8, 8], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k6x11_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 16, 6, 11], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k8x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 16, 8, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c16_m32_48x48_k9x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 16, 9, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k7x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 32, 7, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k8x8_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 32, 8, 8], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k6x11_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 32, 6, 11], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k8x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 32, 8, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 47, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float32_b1_g4_c32_m32_48x48_k9x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [128, 32, 9, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k7x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 16, 7, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k8x8_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 16, 8, 8], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k6x11_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 16, 6, 11], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k8x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 16, 8, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c16_m32_48x48_k9x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 16, 9, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k7x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 32, 7, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k8x8_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 32, 8, 8], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 47] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k6x11_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 32, 6, 11], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k8x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 32, 8, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 47, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-grouped_float16_b1_g4_c32_m32_48x48_k9x9_d1x1_s1x1_bias0_asym0-Relu",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [128, 32, 9, 9], "dist": "normal", "seed": 602, "scale": 0.02 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 48, 48] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-bias-adjacent-float32-b1-c64-m64-g4-32x32-k7",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float32", "shape": [64, 16, 7, 7], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float32", "shape": [64], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 32, 32] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "fused-bias-adjacent-float16-b1-c64-m64-g4-32x32-k7",
+      "preset": "model",
+      "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], "dist": "normal", "seed": 601, "scale": 0.2 },
+        "w": { "dtype": "float16", "shape": [64, 16, 7, 7], "dist": "normal", "seed": 602, "scale": 0.02 },
+        "bias": { "dtype": "float16", "shape": [64], "dist": "normal", "seed": 603, "scale": 0.03 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 32, 32] } },
+      "bench": {
+        "metrics": [
+          { "type": "gflops", "value": "2 * numel(shapes.y) * dim(shapes.w, 1) * dim(shapes.w, 2) * dim(shapes.w, 3)" }
+        ]
+      },
+      "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."
+      }
+    },
+    {
+      "name": "tile-partial-float32-b1c4m8-14x14-k3",
+      "preset": "edge",
+      "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], "dist": "normal", "seed": 1111, "scale": 0.3 },
+        "w": { "dtype": "float32", "shape": [8, 4, 3, 3], "dist": "normal", "seed": 1112, "scale": 0.2 },
+        "bias": { "dtype": "float32", "shape": [8], "dist": "normal", "seed": 1113, "scale": 0.3 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 14, 14] } },
+      "bench": { "metrics": [{ "type": "gflops", "value": 112896 }] }
+    },
+    {
+      "name": "tile-partial-float32-b1c36m8-14x14-k1",
+      "preset": "edge",
+      "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], "dist": "normal", "seed": 1111, "scale": 0.3 },
+        "w": { "dtype": "float32", "shape": [8, 36, 1, 1], "dist": "normal", "seed": 1112, "scale": 0.2 },
+        "bias": { "dtype": "float32", "shape": [8], "dist": "normal", "seed": 1113, "scale": 0.3 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 14, 14] } },
+      "bench": { "metrics": [{ "type": "gflops", "value": 112896 }] }
+    },
+    {
+      "name": "tile-partial-float32-b1c5m17-13x15-k3",
+      "preset": "edge",
+      "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], "dist": "normal", "seed": 1111, "scale": 0.3 },
+        "w": { "dtype": "float32", "shape": [17, 5, 3, 3], "dist": "normal", "seed": 1112, "scale": 0.2 },
+        "bias": { "dtype": "float32", "shape": [17], "dist": "normal", "seed": 1113, "scale": 0.3 }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 17, 13, 15] } },
+      "bench": { "metrics": [{ "type": "gflops", "value": 298350 }] }
+    },
+    {
+      "name": "tile-partial-float16-b1c4m8-14x14-k3",
+      "preset": "edge",
+      "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], "dist": "normal", "seed": 1111, "scale": 0.3 },
+        "w": { "dtype": "float16", "shape": [8, 4, 3, 3], "dist": "normal", "seed": 1112, "scale": 0.2 },
+        "bias": { "dtype": "float16", "shape": [8], "dist": "normal", "seed": 1113, "scale": 0.3 }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 14, 14] } },
+      "bench": { "metrics": [{ "type": "gflops", "value": 112896 }] }
     }
   ]
 }