{
  "cases": [
    {
      "name": "nchw_8_32_64_64_k3",
      "preset": "smoke",
      "attrs": { "p": 2, "kernel_shape": [3, 3], "strides": [2, 2] },
      "inputs": { "x": { "dtype": "float32", "shape": [8, 32, 64, 64] } },
      "outputs": { "y": { "dtype": "float32", "shape": [8, 32, 31, 31] } }
    },
    {
      "name": "nchw_8x64x112x112_k2s2_p2_unrolled",
      "preset": "smoke",
      "vars": { "batch": 8, "channels": 64, "inH": 112, "inW": 112 },
      "attrs": { "p": 2, "kernel_shape": [2, 2], "strides": [2, 2] },
      "inputs": { "x": { "dtype": "float32", "shape": [8, 64, 112, 112], "dist": "normal", "seed": 206 } },
      "outputs": { "y": { "dtype": "float32", "shape": [8, 64, 56, 56], "dist": "empty" } },
      "bench": {
        "primary": true,
        "metrics": [{ "type": "bandwidth", "value": "args.batch * args.channels * args.inH * args.inW * 4" }]
      }
    },
    {
      "name": "nchw_f16_8x64x112x112_k2s2_p2_unrolled",
      "preset": "smoke",
      "vars": { "batch": 8, "channels": 64, "inH": 112, "inW": 112 },
      "attrs": { "p": 2, "kernel_shape": [2, 2], "strides": [2, 2] },
      "inputs": { "x": { "dtype": "float16", "shape": [8, 64, 112, 112], "dist": "normal", "seed": 207 } },
      "outputs": { "y": { "dtype": "float16", "shape": [8, 64, 56, 56], "dist": "empty" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.batch * args.channels * args.inH * args.inW * 2" }] }
    },
    {
      "name": "nchw_f16_8x64x112x110_k2s2_p2_unaligned_width",
      "preset": "smoke",
      "vars": { "batch": 8, "channels": 64, "inH": 112, "inW": 110 },
      "attrs": { "p": 2, "kernel_shape": [2, 2], "strides": [2, 2] },
      "inputs": { "x": { "dtype": "float16", "shape": [8, 64, 112, 110], "dist": "normal", "seed": 311 } },
      "outputs": { "y": { "dtype": "float16", "shape": [8, 64, 56, 55], "dist": "empty" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.batch * args.channels * args.inH * args.inW * 2" }] }
    },
    {
      "name": "nchw_f16_8x64x113x112_k2s2_p2_unaligned_height",
      "preset": "smoke",
      "vars": { "batch": 8, "channels": 64, "inH": 113, "inW": 112 },
      "attrs": { "p": 2, "kernel_shape": [2, 2], "strides": [2, 2] },
      "inputs": { "x": { "dtype": "float16", "shape": [8, 64, 113, 112], "dist": "normal", "seed": 312 } },
      "outputs": { "y": { "dtype": "float16", "shape": [8, 64, 56, 56], "dist": "empty" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.batch * args.channels * args.inH * args.inW * 2" }] }
    },
    {
      "name": "ncl1d_f32_8x256x4000_k4s2_p2_seqpool",
      "preset": "smoke",
      "vars": { "batch": 8, "channels": 256, "inW": 4000 },
      "attrs": { "p": 2, "kernel_shape": [4], "strides": [2] },
      "inputs": { "x": { "dtype": "float32", "shape": [8, 256, 4000], "dist": "normal", "seed": 313 } },
      "outputs": { "y": { "dtype": "float32", "shape": [8, 256, 1999], "dist": "empty" } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "(args.batch * args.channels * args.inW + numel(shapes.y)) * 4" }]
      }
    },
    {
      "name": "ncdhw3d_f32_2x16x16x16x16_k2s2_p2_volumetric",
      "preset": "smoke",
      "vars": { "batch": 2, "channels": 16, "inD": 16, "inH": 16, "inW": 16 },
      "attrs": { "p": 2, "kernel_shape": [2, 2, 2], "strides": [2, 2, 2] },
      "inputs": { "x": { "dtype": "float32", "shape": [2, 16, 16, 16, 16], "dist": "normal", "seed": 314 } },
      "outputs": { "y": { "dtype": "float32", "shape": [2, 16, 8, 8, 8], "dist": "empty" } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "args.batch * args.channels * args.inD * args.inH * args.inW * 4" }]
      }
    },
    {
      "name": "nchw2d_fallback_bigkernel_lowocc_p2_128x128window",
      "preset": "stress",
      "vars": { "batch": 1, "channels": 64, "inH": 128, "inW": 128 },
      "attrs": { "p": 2, "kernel_shape": [128, 128] },
      "inputs": { "x": { "dtype": "float32", "shape": [1, 64, 128, 128], "dist": "normal", "seed": 741, "scale": 2 } },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 1, 1], "dist": "empty" } },
      "bench": {
        "primary": true,
        "metrics": [{ "type": "bandwidth", "value": "args.batch * args.channels * args.inH * args.inW * 4" }]
      }
    },
    {
      "name": "nchw2d_fallback_bigkernel_p3_pow_transcendental_96x96window",
      "preset": "stress",
      "vars": { "batch": 1, "channels": 32, "inH": 96, "inW": 96 },
      "attrs": { "p": 3, "kernel_shape": [96, 96] },
      "inputs": { "x": { "dtype": "float32", "shape": [1, 32, 96, 96], "dist": "normal", "seed": 742, "scale": 2 } },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 1, 1], "dist": "empty" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.batch * args.channels * args.inH * args.inW * 4" }] }
    },
    {
      "name": "ncl1d_bigkernel_lowocc_p2_seqglobalpool",
      "preset": "stress",
      "vars": { "batch": 1, "channels": 64, "inW": 16384 },
      "attrs": { "p": 2, "kernel_shape": [16384] },
      "inputs": { "x": { "dtype": "float32", "shape": [1, 64, 16384], "dist": "normal", "seed": 743, "scale": 2 } },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 1], "dist": "empty" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.batch * args.channels * args.inW * 4" }] }
    },
    {
      "name": "global-lppool-p2-rank5-serial-reduction-pathology",
      "preset": "stress",
      "provenance": {
        "source": "synthetic benchmark",
        "notes": "A realistic rank-5 global L2 pool measures the parallel global-window reduction across feature tiers."
      },
      "vars": { "batch": 1, "channels": 32, "inD": 32, "inH": 64, "inW": 64 },
      "attrs": { "p": 2, "kernel_shape": [32, 64, 64] },
      "inputs": {
        "x": { "dtype": "float32", "shape": [1, 32, 32, 64, 64], "dist": "normal", "seed": 5103, "scale": 0.2 }
      },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 1, 1, 1], "dist": "empty" } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "args.batch * args.channels * args.inD * args.inH * args.inW * 4" }]
      }
    },
    {
      "name": "nchw_f32_1x64x224x224_k7s2p3_generic_bigwindow",
      "preset": "stress",
      "attrs": { "p": 2, "kernel_shape": [7, 7], "strides": [2, 2], "pads": [3, 3, 3, 3] },
      "inputs": { "x": { "dtype": "float32", "shape": [1, 64, 224, 224], "dist": "normal", "seed": 401, "scale": 2 } },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 112, 112], "dist": "empty" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "numel(shapes.x) * 4" }] }
    },
    {
      "name": "nchw_f32_1x64x224x224_k6s2pad2_separable_breakeven",
      "preset": "stress",
      "attrs": { "p": 2, "kernel_shape": [6, 6], "strides": [2, 2], "pads": [2, 2, 2, 2] },
      "inputs": { "x": { "dtype": "float32", "shape": [1, 64, 224, 224], "dist": "normal", "seed": 409, "scale": 2 } },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 112, 112], "dist": "empty" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "numel(shapes.x) * 4" }] }
    },
    {
      "name": "lppool-f32-b1c512-20x20-k5s1p2-separable-floor",
      "provenance": {
        "notes": "Measures L2 pooling (p=2) with a 5x5 window and padding 2 over 512 channels of a 20x20 plane, keeping the spatial size unchanged."
      },
      "attrs": { "p": 2, "kernel_shape": [5, 5], "strides": [1, 1], "pads": [2, 2, 2, 2] },
      "inputs": { "x": { "dtype": "float32", "shape": [1, 512, 20, 20], "dist": "uniform" } },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 512, 20, 20], "dist": "empty" } }
    },
    {
      "name": "lp-pool-f32-b1c16-24x24-k12s1p2-p2-window-parallel-starved",
      "preset": "stress",
      "provenance": {
        "source": "synthetic benchmark",
        "notes": "Measures L2 pooling (p=2) with a 12x12 window and padding 2 over 16 channels of a 24x24 input, producing a 17x17 (289-element) output per channel."
      },
      "attrs": { "p": 2, "kernel_shape": [12, 12], "strides": [1, 1], "pads": [2, 2, 2, 2] },
      "inputs": { "x": { "dtype": "float32", "shape": [1, 16, 24, 24], "dist": "normal", "seed": 4321, "scale": 0.2 } },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 16, 17, 17], "dist": "empty" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] }
    },
    {
      "name": "lp-pool-f32-b1c16-56x56-k28s28-p2-window-parallel-extreme-starved",
      "preset": "stress",
      "provenance": {
        "source": "synthetic benchmark",
        "notes": "A 784-tap non-overlapping L2 window leaving 64 outputs."
      },
      "attrs": { "p": 2, "kernel_shape": [28, 28], "strides": [28, 28], "pads": [0, 0, 0, 0] },
      "inputs": { "x": { "dtype": "float32", "shape": [1, 16, 56, 56], "dist": "normal", "seed": 4322, "scale": 0.2 } },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 16, 2, 2], "dist": "empty" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] }
    },
    {
      "name": "occupancy-window_parallel_nchw2d_p2_k12p2_small_output",
      "preset": "stress",
      "attrs": { "p": 2, "kernel_shape": [12, 12], "strides": [2, 2], "pads": [2, 2, 2, 2] },
      "inputs": { "x": { "dtype": "float32", "shape": [1, 1, 24, 24], "dist": "normal", "seed": 5482, "scale": 0.2 } },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 9, 9] } }
    },
    {
      "name": "occupancy-window_parallel_nchw2d_p1_k12p2_small_output",
      "preset": "stress",
      "attrs": { "p": 1, "kernel_shape": [12, 12], "strides": [2, 2], "pads": [2, 2, 2, 2] },
      "inputs": { "x": { "dtype": "float32", "shape": [1, 1, 24, 24], "dist": "normal", "seed": 5482, "scale": 0.2 } },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 9, 9] } }
    },
    {
      "name": "occupancy-window_parallel_nchw2d_p3_k12p2_small_output",
      "preset": "stress",
      "attrs": { "p": 3, "kernel_shape": [12, 12], "strides": [2, 2], "pads": [2, 2, 2, 2] },
      "inputs": { "x": { "dtype": "float32", "shape": [1, 1, 24, 24], "dist": "normal", "seed": 5482, "scale": 0.2 } },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 9, 9] } }
    },
    {
      "name": "occupancy-window_parallel_outputs_to_taps_boundary_128",
      "preset": "stress",
      "attrs": { "kernel_shape": [128], "strides": [1] },
      "inputs": { "x": { "dtype": "float32", "shape": [1, 1, 255], "dist": "normal", "seed": 5482, "scale": 0.2 } },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 128] } }
    },
    {
      "name": "occupancy-window_parallel_outputs_to_taps_boundary_129",
      "preset": "stress",
      "attrs": { "kernel_shape": [128], "strides": [1] },
      "inputs": { "x": { "dtype": "float32", "shape": [1, 1, 256], "dist": "normal", "seed": 5482, "scale": 0.2 } },
      "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 129] } }
    }
  ],
  "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] }
}