| { |
| "cases": [ |
| { |
| "name": "lowered_prefix_sum_256x512", |
| "preset": "smoke", |
| "inputs": { |
| "initial_state": { "dtype": "float32", "shape": [512], "dist": "normal", "seed": 741, "scale": 0.1 }, |
| "scan_input": { "dtype": "float32", "shape": [256, 512], "dist": "normal", "seed": 742, "scale": 0.1 } |
| }, |
| "outputs": { |
| "final_state": { "dtype": "float32", "shape": [512] }, |
| "scan_output": { "dtype": "float32", "shape": [256, 512] } |
| }, |
| "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "256 * 512 * 4 * 2" }] } |
| }, |
| { |
| "name": "lowered_prefix_sum_2048x512", |
| "inputs": { |
| "initial_state": { "dtype": "float32", "shape": [512], "dist": "normal", "seed": 743, "scale": 0.1 }, |
| "scan_input": { "dtype": "float32", "shape": [2048, 512], "dist": "normal", "seed": 744, "scale": 0.1 } |
| }, |
| "outputs": { |
| "final_state": { "dtype": "float32", "shape": [512] }, |
| "scan_output": { "dtype": "float32", "shape": [2048, 512] } |
| }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2048 * 512 * 4 * 2" }] } |
| }, |
| { |
| "name": "lowered_prefix_sum_256x128", |
| "inputs": { |
| "initial_state": { "dtype": "float32", "shape": [128] }, |
| "scan_input": { "dtype": "float32", "shape": [256, 128] } |
| }, |
| "outputs": { |
| "final_state": { "dtype": "float32", "shape": [128] }, |
| "scan_output": { "dtype": "float32", "shape": [256, 128] } |
| } |
| }, |
| { |
| "name": "lowered_fewlane_serial_dim32_steps4096", |
| "preset": "smoke", |
| "inputs": { |
| "initial_state": { "dtype": "float32", "shape": [32], "dist": "normal", "seed": 751, "scale": 0.1 }, |
| "scan_input": { "dtype": "float32", "shape": [4096, 32], "dist": "normal", "seed": 752, "scale": 0.1 } |
| }, |
| "outputs": { |
| "final_state": { "dtype": "float32", "shape": [32] }, |
| "scan_output": { "dtype": "float32", "shape": [4096, 32] } |
| }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2 * 4 * numel(shapes.scan_input)" }] } |
| }, |
| { |
| "name": "coop_control_dim512_steps4096", |
| "preset": "smoke", |
| "inputs": { |
| "initial_state": { "dtype": "float32", "shape": [512], "dist": "normal", "seed": 753, "scale": 0.1 }, |
| "scan_input": { "dtype": "float32", "shape": [4096, 512], "dist": "normal", "seed": 754, "scale": 0.1 } |
| }, |
| "outputs": { |
| "final_state": { "dtype": "float32", "shape": [512] }, |
| "scan_output": { "dtype": "float32", "shape": [4096, 512] } |
| }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2 * 4 * numel(shapes.scan_input)" }] } |
| }, |
| { |
| "name": "lowered_largedim_fallback_dim4096_steps1024", |
| "preset": "smoke", |
| "inputs": { |
| "initial_state": { "dtype": "float32", "shape": [4096], "dist": "normal", "seed": 755, "scale": 0.1 }, |
| "scan_input": { "dtype": "float32", "shape": [1024, 4096], "dist": "normal", "seed": 756, "scale": 0.1 } |
| }, |
| "outputs": { |
| "final_state": { "dtype": "float32", "shape": [4096] }, |
| "scan_output": { "dtype": "float32", "shape": [1024, 4096] } |
| }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2 * 4 * numel(shapes.scan_input)" }] } |
| }, |
| { |
| "name": "lowered_smalldim2_singlelane_steps8192", |
| "preset": "smoke", |
| "inputs": { |
| "initial_state": { "dtype": "float32", "shape": [2], "dist": "normal", "seed": 757, "scale": 0.1 }, |
| "scan_input": { "dtype": "float32", "shape": [8192, 2], "dist": "normal", "seed": 758, "scale": 0.1 } |
| }, |
| "outputs": { |
| "final_state": { "dtype": "float32", "shape": [2] }, |
| "scan_output": { "dtype": "float32", "shape": [8192, 2] } |
| }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2 * 4 * numel(shapes.scan_input)" }] } |
| }, |
| { |
| "name": "coop_reverse_partialchunk_dim512_steps300", |
| "preset": "smoke", |
| "attrs": { "reverse": 1 }, |
| "inputs": { |
| "initial_state": { "dtype": "float32", "shape": [512], "dist": "normal", "seed": 759, "scale": 0.1 }, |
| "scan_input": { "dtype": "float32", "shape": [300, 512], "dist": "normal", "seed": 760, "scale": 0.1 } |
| }, |
| "outputs": { |
| "final_state": { "dtype": "float32", "shape": [512] }, |
| "scan_output": { "dtype": "float32", "shape": [300, 512] } |
| }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2 * 4 * numel(shapes.scan_input)" }] } |
| }, |
| { |
| "name": "lowered_singlelane_dim1_steps65536", |
| "preset": "stress", |
| "inputs": { |
| "initial_state": { "dtype": "float32", "shape": [1], "dist": "normal", "seed": 811, "scale": 0.1 }, |
| "scan_input": { "dtype": "float32", "shape": [65536, 1], "dist": "normal", "seed": 812, "scale": 0.1 } |
| }, |
| "outputs": { |
| "final_state": { "dtype": "float32", "shape": [1] }, |
| "scan_output": { "dtype": "float32", "shape": [65536, 1] } |
| }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2 * 4 * numel(shapes.scan_input)" }] } |
| }, |
| { |
| "name": "lowered_dim63_steps32768_one_workgroup_launchbound", |
| "preset": "stress", |
| "inputs": { |
| "initial_state": { "dtype": "float32", "shape": [63], "dist": "normal", "seed": 813, "scale": 0.1 }, |
| "scan_input": { "dtype": "float32", "shape": [32768, 63], "dist": "normal", "seed": 814, "scale": 0.1 } |
| }, |
| "outputs": { |
| "final_state": { "dtype": "float32", "shape": [63] }, |
| "scan_output": { "dtype": "float32", "shape": [32768, 63] } |
| }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "2 * 4 * numel(shapes.scan_input)" }] } |
| }, |
| { |
| "name": "streaming_recurrence_dim8_steps131072_multichunk_pathology", |
| "preset": "stress", |
| "inputs": { |
| "initial_state": { "dtype": "float32", "shape": [8], "dist": "normal", "seed": 815, "scale": 0.1 }, |
| "scan_input": { "dtype": "float32", "shape": [131072, 8], "dist": "normal", "seed": 816, "scale": 0.01 } |
| }, |
| "outputs": { |
| "final_state": { "dtype": "float32", "shape": [8] }, |
| "scan_output": { "dtype": "float32", "shape": [131072, 8] } |
| }, |
| "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "2 * 4 * numel(shapes.scan_input)" }] } |
| }, |
| { |
| "name": "vector_channel_boundary_t63_d12_reverse0", |
| "attrs": { "reverse": 0 }, |
| "inputs": { |
| "initial_state": { "dtype": "float32", "shape": [12], "dist": "normal", "seed": 1301, "scale": 0.1 }, |
| "scan_input": { "dtype": "float32", "shape": [63, 12], "dist": "normal", "seed": 1302, "scale": 0.1 } |
| }, |
| "outputs": { |
| "final_state": { "dtype": "float32", "shape": [12] }, |
| "scan_output": { "dtype": "float32", "shape": [63, 12] } |
| }, |
| "preset": "edge", |
| "bench": { |
| "metrics": [{ "type": "bandwidth", "value": "8 * numel(shapes.scan_input) + 8 * dim(shapes.initial_state, 0)" }] |
| } |
| }, |
| { |
| "name": "vector_channel_boundary_t64_d12_reverse0", |
| "attrs": { "reverse": 0 }, |
| "inputs": { |
| "initial_state": { "dtype": "float32", "shape": [12], "dist": "normal", "seed": 1301, "scale": 0.1 }, |
| "scan_input": { "dtype": "float32", "shape": [64, 12], "dist": "normal", "seed": 1302, "scale": 0.1 } |
| }, |
| "outputs": { |
| "final_state": { "dtype": "float32", "shape": [12] }, |
| "scan_output": { "dtype": "float32", "shape": [64, 12] } |
| }, |
| "preset": "edge", |
| "bench": { |
| "metrics": [{ "type": "bandwidth", "value": "8 * numel(shapes.scan_input) + 8 * dim(shapes.initial_state, 0)" }] |
| } |
| }, |
| { |
| "name": "vector_channel_boundary_t65_d12_reverse1", |
| "attrs": { "reverse": 1 }, |
| "inputs": { |
| "initial_state": { "dtype": "float32", "shape": [12], "dist": "normal", "seed": 1301, "scale": 0.1 }, |
| "scan_input": { "dtype": "float32", "shape": [65, 12], "dist": "normal", "seed": 1302, "scale": 0.1 } |
| }, |
| "outputs": { |
| "final_state": { "dtype": "float32", "shape": [12] }, |
| "scan_output": { "dtype": "float32", "shape": [65, 12] } |
| }, |
| "preset": "edge", |
| "bench": { |
| "metrics": [{ "type": "bandwidth", "value": "8 * numel(shapes.scan_input) + 8 * dim(shapes.initial_state, 0)" }] |
| } |
| }, |
| { |
| "name": "vector_channel_boundary_t255_d4_reverse0", |
| "attrs": { "reverse": 0 }, |
| "inputs": { |
| "initial_state": { "dtype": "float32", "shape": [4], "dist": "normal", "seed": 1301, "scale": 0.1 }, |
| "scan_input": { "dtype": "float32", "shape": [255, 4], "dist": "normal", "seed": 1302, "scale": 0.1 } |
| }, |
| "outputs": { |
| "final_state": { "dtype": "float32", "shape": [4] }, |
| "scan_output": { "dtype": "float32", "shape": [255, 4] } |
| }, |
| "preset": "edge", |
| "bench": { |
| "metrics": [{ "type": "bandwidth", "value": "8 * numel(shapes.scan_input) + 8 * dim(shapes.initial_state, 0)" }] |
| } |
| }, |
| { |
| "name": "vector_channel_boundary_t257_d68_reverse1", |
| "attrs": { "reverse": 1 }, |
| "inputs": { |
| "initial_state": { "dtype": "float32", "shape": [68], "dist": "normal", "seed": 1301, "scale": 0.1 }, |
| "scan_input": { "dtype": "float32", "shape": [257, 68], "dist": "normal", "seed": 1302, "scale": 0.1 } |
| }, |
| "outputs": { |
| "final_state": { "dtype": "float32", "shape": [68] }, |
| "scan_output": { "dtype": "float32", "shape": [257, 68] } |
| }, |
| "preset": "edge", |
| "bench": { |
| "metrics": [{ "type": "bandwidth", "value": "8 * numel(shapes.scan_input) + 8 * dim(shapes.initial_state, 0)" }] |
| } |
| }, |
| { |
| "name": "vector_channel_boundary_t257_d69_reverse1", |
| "attrs": { "reverse": 1 }, |
| "inputs": { |
| "initial_state": { "dtype": "float32", "shape": [69], "dist": "normal", "seed": 1301, "scale": 0.1 }, |
| "scan_input": { "dtype": "float32", "shape": [257, 69], "dist": "normal", "seed": 1302, "scale": 0.1 } |
| }, |
| "outputs": { |
| "final_state": { "dtype": "float32", "shape": [69] }, |
| "scan_output": { "dtype": "float32", "shape": [257, 69] } |
| }, |
| "preset": "edge", |
| "bench": { |
| "metrics": [{ "type": "bandwidth", "value": "8 * numel(shapes.scan_input) + 8 * dim(shapes.initial_state, 0)" }] |
| } |
| } |
| ] |
| } |
|
|
