| { |
| "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] }, |
| "cases": [ |
| { |
| "name": "nchw_1x16_32x32_k3", |
| "attrs": {}, |
| "inputs": { |
| "input": { "dtype": "float32", "shape": [1, 144, 900] }, |
| "image_shape": { "dtype": "uint32", "shape": [2], "data": { "kind": "values", "values": [32, 32] } }, |
| "block_shape": { "dtype": "uint32", "shape": [2], "data": { "kind": "values", "values": [3, 3] } } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 16, 32, 32] } } |
| }, |
| { |
| "name": "nchw_4m_k3_overlap_control", |
| "preset": "smoke", |
| "attrs": {}, |
| "inputs": { |
| "input": { "dtype": "float32", "shape": [1, 144, 260100], "dist": "normal", "seed": 411, "scale": 0.2 }, |
| "image_shape": { "dtype": "uint32", "shape": [2], "data": { "kind": "values", "values": [512, 512] } }, |
| "block_shape": { "dtype": "uint32", "shape": [2], "data": { "kind": "values", "values": [3, 3] } } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 16, 512, 512] } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.input) + numel(shapes.output)) * 4" }] |
| } |
| }, |
| { |
| "name": "nchw_dispatch_cliff_8m_gidy_fold", |
| "preset": "smoke", |
| "attrs": {}, |
| "inputs": { |
| "input": { "dtype": "float32", "shape": [1, 32, 262144], "dist": "normal", "seed": 412, "scale": 0.2 }, |
| "image_shape": { "dtype": "uint32", "shape": [2], "data": { "kind": "values", "values": [512, 512] } }, |
| "block_shape": { "dtype": "uint32", "shape": [2], "data": { "kind": "values", "values": [1, 1] } } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 32, 512, 512] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.input) + numel(shapes.output)) * 4" }] } |
| }, |
| { |
| "name": "nchw_k1x1_copy_baseline", |
| "preset": "smoke", |
| "attrs": {}, |
| "inputs": { |
| "input": { "dtype": "float32", "shape": [1, 16, 262144], "dist": "normal", "seed": 413, "scale": 0.2 }, |
| "image_shape": { "dtype": "uint32", "shape": [2], "data": { "kind": "values", "values": [512, 512] } }, |
| "block_shape": { "dtype": "uint32", "shape": [2], "data": { "kind": "values", "values": [1, 1] } } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 16, 512, 512] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.input) + numel(shapes.output)) * 4" }] } |
| }, |
| { |
| "name": "nchw-k1-fold-f32-b1c320-128x128-diffusion-feature", |
| "preset": "stress", |
| "attrs": {}, |
| "inputs": { |
| "input": { "dtype": "float32", "shape": [1, 320, 16384], "dist": "normal", "seed": 416, "scale": 0.2 }, |
| "image_shape": { "dtype": "uint32", "shape": [2], "data": { "kind": "values", "values": [128, 128] } }, |
| "block_shape": { "dtype": "uint32", "shape": [2], "data": { "kind": "values", "values": [1, 1] } } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 320, 128, 128] } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.input) + numel(shapes.output)) * 4" }] |
| } |
| }, |
| { |
| "name": "nchw_k5x5_overlap_heavy_gather", |
| "preset": "stress", |
| "provenance": { |
| "notes": "Stress-only capacity case: the overlap-heavy gather declares 409 MiB of GPU-resident tensors." |
| }, |
| "attrs": {}, |
| "inputs": { |
| "input": { "dtype": "float32", "shape": [1, 400, 258064], "dist": "normal", "seed": 414, "scale": 0.1 }, |
| "image_shape": { "dtype": "uint32", "shape": [2], "data": { "kind": "values", "values": [512, 512] } }, |
| "block_shape": { "dtype": "uint32", "shape": [2], "data": { "kind": "values", "values": [5, 5] } } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 16, 512, 512] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.input) + numel(shapes.output)) * 4" }] } |
| }, |
| { |
| "name": "ncdhw_3d_2m_copy_variant", |
| "preset": "smoke", |
| "attrs": {}, |
| "inputs": { |
| "input": { "dtype": "float32", "shape": [1, 16, 131072], "dist": "normal", "seed": 415, "scale": 0.2 }, |
| "image_shape": { "dtype": "uint32", "shape": [3], "data": { "kind": "values", "values": [32, 64, 64] } }, |
| "block_shape": { "dtype": "uint32", "shape": [3], "data": { "kind": "values", "values": [1, 1, 1] } } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 16, 32, 64, 64] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.input) + numel(shapes.output)) * 4" }] } |
| } |
| ] |
| } |
|
|
