| { |
| "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] }, |
| "cases": [ |
| { |
| "name": "rank4_image", |
| "attrs": { "pads": [0, 0, 1, 1, 0, 0, 1, 1] }, |
| "inputs": { "data": { "dtype": "float32", "shape": [1, 16, 128, 128], "dist": "normal", "seed": 1 } }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 16, 130, 130], "dist": "empty" } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.data) + numel(shapes.output)) * 4" }] |
| } |
| }, |
| { |
| "name": "rank4_image_inner_aligned", |
| "attrs": { "pads": [0, 0, 1, 1, 0, 0, 1, 1] }, |
| "inputs": { "data": { "dtype": "float32", "shape": [1, 16, 126, 126], "dist": "normal", "seed": 2 } }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 16, 128, 128], "dist": "empty" } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.data) + numel(shapes.output)) * 4" }] |
| }, |
| "preset": "smoke" |
| }, |
| { |
| "name": "rank3_seq_pad_llm", |
| "attrs": { "pads": [0, 0, 0, 0, 64, 0] }, |
| "inputs": { "data": { "dtype": "float32", "shape": [1, 448, 768], "dist": "normal", "seed": 3 } }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 512, 768], "dist": "empty" } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.data) + numel(shapes.output)) * 4" }] |
| } |
| }, |
| { |
| "name": "rank4_image_reflect_inner_aligned", |
| "attrs": { "pads": [0, 0, 2, 2, 0, 0, 2, 2], "mode": "reflect" }, |
| "inputs": { "data": { "dtype": "float32", "shape": [1, 16, 128, 128], "dist": "normal", "seed": 4 } }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 16, 132, 132], "dist": "empty" } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.data) + numel(shapes.output)) * 4" }] |
| } |
| }, |
| { |
| "name": "rank4_image_f16_inner_aligned", |
| "attrs": { "pads": [0, 0, 1, 1, 0, 0, 1, 1] }, |
| "inputs": { "data": { "dtype": "float16", "shape": [1, 16, 126, 126], "dist": "normal", "seed": 5 } }, |
| "outputs": { "output": { "dtype": "float16", "shape": [1, 16, 128, 128], "dist": "empty" } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.data) + numel(shapes.output)) * 2" }] |
| } |
| }, |
| { |
| "name": "rank4_image_value_input_inner_aligned", |
| "attrs": { "pads": [0, 0, 1, 1, 0, 0, 1, 1] }, |
| "inputs": { |
| "data": { "dtype": "float32", "shape": [1, 16, 126, 126], "dist": "normal", "seed": 6 }, |
| "constant_value": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [0.0] } } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 16, 128, 128], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.data) + numel(shapes.output)) * 4" }] } |
| }, |
| { |
| "name": "rank4_image_edge_inner_aligned", |
| "attrs": { "pads": [0, 0, 1, 1, 0, 0, 1, 1], "mode": "edge" }, |
| "inputs": { "data": { "dtype": "float32", "shape": [1, 16, 126, 126], "dist": "normal", "seed": 7 } }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 16, 128, 128], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.data) + numel(shapes.output)) * 4" }] } |
| }, |
| { |
| "name": "rank4_image_f16_reflect_inner_aligned", |
| "attrs": { "pads": [0, 0, 2, 2, 0, 0, 2, 2], "mode": "reflect" }, |
| "inputs": { "data": { "dtype": "float16", "shape": [1, 16, 128, 128], "dist": "normal", "seed": 8 } }, |
| "outputs": { "output": { "dtype": "float16", "shape": [1, 16, 132, 132], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.data) + numel(shapes.output)) * 2" }] } |
| }, |
| { |
| "name": "rank3_value_input_scalar_path_large", |
| "preset": "stress", |
| "attrs": { "mode": "constant", "pads": [0, 0, 0, 0, 64, 0] }, |
| "inputs": { |
| "data": { "dtype": "float32", "shape": [8, 448, 768], "dist": "normal", "seed": 11, "scale": 2 }, |
| "constant_value": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [0.0] } } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [8, 512, 768], "dist": "empty" } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.data) + numel(shapes.output)) * 4" }] |
| } |
| }, |
| { |
| "name": "rank4_edge_inner_unaligned_scalar_path", |
| "preset": "stress", |
| "attrs": { "mode": "edge", "pads": [0, 0, 1, 1, 0, 0, 1, 1] }, |
| "inputs": { "data": { "dtype": "float32", "shape": [1, 64, 127, 127], "dist": "normal", "seed": 12, "scale": 2 } }, |
| "outputs": { "output": { "dtype": "float32", "shape": [1, 64, 129, 129], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.data) + numel(shapes.output)) * 4" }] } |
| }, |
| { |
| "name": "rank2_constant_inner_unaligned_scalar_path", |
| "preset": "stress", |
| "attrs": { "mode": "constant", "pads": [1, 1, 1, 0] }, |
| "inputs": { |
| "data": { "dtype": "float32", "shape": [2046, 1024], "dist": "normal", "seed": 13, "scale": 2 }, |
| "constant_value": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [3.0] } } |
| }, |
| "outputs": { "output": { "dtype": "float32", "shape": [2048, 1025], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.data) + numel(shapes.output)) * 4" }] } |
| } |
| ] |
| } |
|
|
