{ "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] }, "cases": [ { "name": "small", "attrs": { "layout": 0, "direction": "forward", "linear_before_reset": 0, "hidden_size": 32 }, "inputs": { "x": { "dtype": "float32", "shape": [16, 1, 32], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02 } }, "w": { "dtype": "float32", "shape": [1, 96, 32], "data": { "kind": "fillFloat32", "sinStep": 0.03, "cosStep": 0.04, "scale": 0.08 } }, "r": { "dtype": "float32", "shape": [1, 96, 32], "data": { "kind": "fillFloat32", "sinStep": 0.05, "cosStep": 0.06, "scale": 0.08 } }, "b": { "dtype": "float32", "shape": [1, 192], "data": { "kind": "constant", "value": 0.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [16, 1, 1, 32] }, "y_h": { "dtype": "float32", "shape": [1, 1, 32] } }, "bench": { "metrics": [ { "type": "gflops", "value": "2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) * (dim(shapes.x, 2) + dim(shapes.r, 2))" }, { "type": "bandwidth", "name": "recurrent lower-bound BW", "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))" } ] } }, { "name": "h256-in64-b8-seq32-vec4", "preset": "smoke", "attrs": { "layout": 0, "direction": "forward", "linear_before_reset": 0, "hidden_size": 256 }, "vars": { "seq": 32, "batch": 8, "inputSize": 64, "hidden": 256 }, "inputs": { "x": { "dtype": "float32", "shape": [32, 8, 64], "dist": "normal", "seed": 421, "scale": 0.3 }, "w": { "dtype": "float32", "shape": [1, 768, 64], "dist": "normal", "seed": 422, "scale": 0.05 }, "r": { "dtype": "float32", "shape": [1, 768, 256], "dist": "normal", "seed": 423, "scale": 0.02 }, "b": { "dtype": "float32", "shape": [1, 1536], "dist": "normal", "seed": 424, "scale": 0.05 } }, "outputs": { "y": { "dtype": "float32", "shape": [32, 1, 8, 256] }, "y_h": { "dtype": "float32", "shape": [1, 8, 256] } }, "bench": { "primary": true, "metrics": [ { "type": "gflops", "value": "args.seq * args.batch * 6 * args.hidden * (args.inputSize + args.hidden)" }, { "type": "bandwidth", "name": "recurrent lower-bound BW", "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))" } ] } }, { "name": "h128-in128-b4-seq64-lbr1-vec4", "attrs": { "layout": 0, "direction": "forward", "linear_before_reset": 1, "hidden_size": 128 }, "vars": { "seq": 64, "batch": 4, "inputSize": 128, "hidden": 128 }, "inputs": { "x": { "dtype": "float32", "shape": [64, 4, 128], "dist": "normal", "seed": 431, "scale": 0.3 }, "w": { "dtype": "float32", "shape": [1, 384, 128], "dist": "normal", "seed": 432, "scale": 0.05 }, "r": { "dtype": "float32", "shape": [1, 384, 128], "dist": "normal", "seed": 433, "scale": 0.05 }, "b": { "dtype": "float32", "shape": [1, 768], "dist": "normal", "seed": 434, "scale": 0.05 } }, "outputs": { "y": { "dtype": "float32", "shape": [64, 1, 4, 128] }, "y_h": { "dtype": "float32", "shape": [1, 4, 128] } }, "bench": { "metrics": [ { "type": "gflops", "value": "args.seq * args.batch * 6 * args.hidden * (args.inputSize + args.hidden)" }, { "type": "bandwidth", "name": "recurrent lower-bound BW", "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))" } ] } }, { "name": "large-h512-in64-b2-seq4-vec4", "preset": "model", "attrs": { "layout": 0, "direction": "forward", "linear_before_reset": 0, "hidden_size": 512 }, "vars": { "seq": 4, "batch": 2, "inputSize": 64, "hidden": 512 }, "inputs": { "x": { "dtype": "float32", "shape": [4, 2, 64], "dist": "normal", "seed": 501, "scale": 0.3 }, "w": { "dtype": "float32", "shape": [1, 1536, 64], "dist": "normal", "seed": 502, "scale": 0.05 }, "r": { "dtype": "float32", "shape": [1, 1536, 512], "dist": "normal", "seed": 503, "scale": 0.02 }, "b": { "dtype": "float32", "shape": [1, 3072], "dist": "normal", "seed": 504, "scale": 0.05 } }, "outputs": { "y": { "dtype": "float32", "shape": [4, 1, 2, 512] }, "y_h": { "dtype": "float32", "shape": [1, 2, 512] } }, "bench": { "primary": false, "metrics": [ { "type": "gflops", "value": "args.seq * args.batch * 6 * args.hidden * (args.inputSize + args.hidden)" }, { "type": "bandwidth", "name": "recurrent lower-bound BW", "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))" } ] } }, { "name": "h250-in128-b8-seq32-mixed-vector", "preset": "smoke", "attrs": { "layout": 0, "direction": "forward", "linear_before_reset": 0, "hidden_size": 250 }, "vars": { "seq": 32, "batch": 8, "inputSize": 128, "hidden": 250 }, "inputs": { "x": { "dtype": "float32", "shape": [32, 8, 128], "dist": "normal", "seed": 611, "scale": 0.3 }, "w": { "dtype": "float32", "shape": [1, 750, 128], "dist": "normal", "seed": 612, "scale": 0.05 }, "r": { "dtype": "float32", "shape": [1, 750, 250], "dist": "normal", "seed": 613, "scale": 0.02 }, "b": { "dtype": "float32", "shape": [1, 1500], "dist": "normal", "seed": 614, "scale": 0.05 } }, "outputs": { "y": { "dtype": "float32", "shape": [32, 1, 8, 250] }, "y_h": { "dtype": "float32", "shape": [1, 8, 250] } }, "bench": { "metrics": [ { "type": "gflops", "value": "args.seq * args.batch * 6 * args.hidden * (args.inputSize + args.hidden)" }, { "type": "bandwidth", "name": "recurrent lower-bound BW", "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))" } ] } }, { "name": "h256-in128-b8-seq32-vec4-healthy", "preset": "smoke", "attrs": { "layout": 0, "direction": "forward", "linear_before_reset": 0, "hidden_size": 256 }, "vars": { "seq": 32, "batch": 8, "inputSize": 128, "hidden": 256 }, "inputs": { "x": { "dtype": "float32", "shape": [32, 8, 128], "dist": "normal", "seed": 621, "scale": 0.3 }, "w": { "dtype": "float32", "shape": [1, 768, 128], "dist": "normal", "seed": 622, "scale": 0.05 }, "r": { "dtype": "float32", "shape": [1, 768, 256], "dist": "normal", "seed": 623, "scale": 0.02 }, "b": { "dtype": "float32", "shape": [1, 1536], "dist": "normal", "seed": 624, "scale": 0.05 } }, "outputs": { "y": { "dtype": "float32", "shape": [32, 1, 8, 256] }, "y_h": { "dtype": "float32", "shape": [1, 8, 256] } }, "bench": { "metrics": [ { "type": "gflops", "value": "args.seq * args.batch * 6 * args.hidden * (args.inputSize + args.hidden)" }, { "type": "bandwidth", "name": "recurrent lower-bound BW", "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))" } ] } }, { "name": "h2048-in256-b4-seq16-workgroup-storage-boundary", "preset": "smoke", "attrs": { "layout": 0, "direction": "forward", "linear_before_reset": 0, "hidden_size": 2048 }, "vars": { "seq": 16, "batch": 4, "inputSize": 256, "hidden": 2048 }, "tunableSpace": { "RECURRENT_WORKGROUP_SIZE": [1, 2, 4, 8, 16, 32, 64] }, "inputs": { "x": { "dtype": "float32", "shape": [16, 4, 256], "dist": "normal", "seed": 631, "scale": 0.3 }, "w": { "dtype": "float32", "shape": [1, 6144, 256], "dist": "normal", "seed": 632, "scale": 0.02 }, "r": { "dtype": "float32", "shape": [1, 6144, 2048], "dist": "normal", "seed": 633, "scale": 0.01 }, "b": { "dtype": "float32", "shape": [1, 12288], "dist": "normal", "seed": 634, "scale": 0.05 } }, "outputs": { "y": { "dtype": "float32", "shape": [16, 1, 4, 2048] }, "y_h": { "dtype": "float32", "shape": [1, 4, 2048] } }, "bench": { "metrics": [ { "type": "gflops", "value": "args.seq * args.batch * 6 * args.hidden * (args.inputSize + args.hidden)" }, { "type": "bandwidth", "name": "recurrent lower-bound BW", "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))" } ] } }, { "name": "h512-in256-b4-seq16-vec4-healthy", "preset": "smoke", "attrs": { "layout": 0, "direction": "forward", "linear_before_reset": 0, "hidden_size": 512 }, "vars": { "seq": 16, "batch": 4, "inputSize": 256, "hidden": 512 }, "inputs": { "x": { "dtype": "float32", "shape": [16, 4, 256], "dist": "normal", "seed": 641, "scale": 0.3 }, "w": { "dtype": "float32", "shape": [1, 1536, 256], "dist": "normal", "seed": 642, "scale": 0.02 }, "r": { "dtype": "float32", "shape": [1, 1536, 512], "dist": "normal", "seed": 643, "scale": 0.01 }, "b": { "dtype": "float32", "shape": [1, 3072], "dist": "normal", "seed": 644, "scale": 0.05 } }, "outputs": { "y": { "dtype": "float32", "shape": [16, 1, 4, 512] }, "y_h": { "dtype": "float32", "shape": [1, 4, 512] } }, "bench": { "metrics": [ { "type": "gflops", "value": "args.seq * args.batch * 6 * args.hidden * (args.inputSize + args.hidden)" }, { "type": "bandwidth", "name": "recurrent lower-bound BW", "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))" } ] } }, { "name": "h256-in64-b8-seq1-decode-single-step", "preset": "smoke", "attrs": { "layout": 0, "direction": "forward", "linear_before_reset": 0, "hidden_size": 256 }, "vars": { "seq": 1, "batch": 8, "inputSize": 64, "hidden": 256 }, "inputs": { "x": { "dtype": "float32", "shape": [1, 8, 64], "dist": "normal", "seed": 651, "scale": 0.3 }, "w": { "dtype": "float32", "shape": [1, 768, 64], "dist": "normal", "seed": 652, "scale": 0.05 }, "r": { "dtype": "float32", "shape": [1, 768, 256], "dist": "normal", "seed": 653, "scale": 0.02 }, "b": { "dtype": "float32", "shape": [1, 1536], "dist": "normal", "seed": 654, "scale": 0.05 } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 8, 256] }, "y_h": { "dtype": "float32", "shape": [1, 8, 256] } }, "bench": { "metrics": [ { "type": "gflops", "value": "args.seq * args.batch * 6 * args.hidden * (args.inputSize + args.hidden)" }, { "type": "bandwidth", "name": "recurrent lower-bound BW", "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))" } ] } }, { "name": "h512-in64-b1-seq64-single-batch-recurrence-starved", "preset": "stress", "attrs": { "layout": 0, "direction": "forward", "linear_before_reset": 0, "hidden_size": 512 }, "vars": { "seq": 64, "batch": 1, "inputSize": 64, "hidden": 512 }, "inputs": { "x": { "dtype": "float32", "shape": [64, 1, 64], "dist": "normal", "seed": 711, "scale": 0.3 }, "w": { "dtype": "float32", "shape": [1, 1536, 64], "dist": "normal", "seed": 712, "scale": 0.05 }, "r": { "dtype": "float32", "shape": [1, 1536, 512], "dist": "normal", "seed": 713, "scale": 0.02 }, "b": { "dtype": "float32", "shape": [1, 3072], "dist": "normal", "seed": 714, "scale": 0.05 } }, "outputs": { "y": { "dtype": "float32", "shape": [64, 1, 1, 512] }, "y_h": { "dtype": "float32", "shape": [1, 1, 512] } }, "bench": { "metrics": [ { "type": "gflops", "value": "args.seq * args.batch * 6 * args.hidden * (args.inputSize + args.hidden)" }, { "type": "bandwidth", "name": "recurrent lower-bound BW", "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))" } ] } }, { "name": "h2048-in256-b1-seq8-workgroup-storage-single-workgroup", "preset": "stress", "attrs": { "layout": 0, "direction": "forward", "linear_before_reset": 0, "hidden_size": 2048 }, "vars": { "seq": 8, "batch": 1, "inputSize": 256, "hidden": 2048 }, "tunableSpace": { "RECURRENT_WORKGROUP_SIZE": [1, 2, 4, 8, 16, 32, 64] }, "inputs": { "x": { "dtype": "float32", "shape": [8, 1, 256], "dist": "normal", "seed": 721, "scale": 0.3 }, "w": { "dtype": "float32", "shape": [1, 6144, 256], "dist": "normal", "seed": 722, "scale": 0.02 }, "r": { "dtype": "float32", "shape": [1, 6144, 2048], "dist": "normal", "seed": 723, "scale": 0.01 }, "b": { "dtype": "float32", "shape": [1, 12288], "dist": "normal", "seed": 724, "scale": 0.05 } }, "outputs": { "y": { "dtype": "float32", "shape": [8, 1, 1, 2048] }, "y_h": { "dtype": "float32", "shape": [1, 1, 2048] } }, "bench": { "metrics": [ { "type": "gflops", "value": "args.seq * args.batch * 6 * args.hidden * (args.inputSize + args.hidden)" }, { "type": "bandwidth", "name": "recurrent lower-bound BW", "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))" } ] } }, { "name": "reverse-h128-in128-b8-seq64-general-scalar", "preset": "stress", "attrs": { "layout": 0, "direction": "reverse", "linear_before_reset": 0, "hidden_size": 128 }, "vars": { "seq": 64, "batch": 8, "inputSize": 128, "hidden": 128 }, "inputs": { "x": { "dtype": "float32", "shape": [64, 8, 128], "dist": "normal", "seed": 731, "scale": 0.3 }, "w": { "dtype": "float32", "shape": [1, 384, 128], "dist": "normal", "seed": 732, "scale": 0.05 }, "r": { "dtype": "float32", "shape": [1, 384, 128], "dist": "normal", "seed": 733, "scale": 0.02 }, "b": { "dtype": "float32", "shape": [1, 768], "dist": "normal", "seed": 734, "scale": 0.05 } }, "outputs": { "y": { "dtype": "float32", "shape": [64, 1, 8, 128] }, "y_h": { "dtype": "float32", "shape": [1, 8, 128] } }, "bench": { "metrics": [ { "type": "gflops", "value": "args.seq * args.batch * 6 * args.hidden * (args.inputSize + args.hidden)" }, { "type": "bandwidth", "name": "recurrent lower-bound BW", "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))" } ] } }, { "name": "manifest-gap-sequence-lens-b8-h256-in256-seq64-general-path", "preset": "stress", "provenance": { "source": "synthetic benchmark", "notes": "Full-length sequence_lens case that verifies the shared post-mask route keeps the projected parallel GRU path eligible." }, "attrs": { "layout": 0, "direction": "forward", "linear_before_reset": 0, "hidden_size": 256 }, "vars": { "seq": 64, "batch": 8, "inputSize": 256, "hidden": 256 }, "inputs": { "x": { "dtype": "float32", "shape": [64, 8, 256], "dist": "normal", "seed": 9191, "scale": 0.3 }, "w": { "dtype": "float32", "shape": [1, 768, 256], "dist": "normal", "seed": 9192, "scale": 0.05 }, "r": { "dtype": "float32", "shape": [1, 768, 256], "dist": "normal", "seed": 9193, "scale": 0.02 }, "b": { "dtype": "float32", "shape": [1, 1536], "dist": "normal", "seed": 9194, "scale": 0.05 }, "sequence_lens": { "dtype": "int32", "shape": [8], "data": { "kind": "values", "values": [64, 64, 64, 64, 64, 64, 64, 64] } } }, "outputs": { "y": { "dtype": "float32", "shape": [64, 1, 8, 256], "dist": "empty" }, "y_h": { "dtype": "float32", "shape": [1, 8, 256], "dist": "empty" } }, "bench": { "metrics": [ { "type": "gflops", "value": "args.seq * args.batch * 6 * args.hidden * (args.inputSize + args.hidden)" }, { "type": "bandwidth", "name": "recurrent lower-bound BW", "value": "4 * (numel(shapes.x) + numel(shapes.w) + 2 * dim(shapes.x, 0) * dim(shapes.x, 1) * dim(shapes.w, 0) * dim(shapes.w, 1) + max(dim(shapes.x, 0) - 1, 0) * numel(shapes.r) + 2 * numel(shapes.y))" } ] } }, { "name": "gru-layout1-batch-major-h256-in256-b8-seq64", "preset": "edge", "attrs": { "layout": 1, "direction": "forward", "linear_before_reset": 0, "hidden_size": 256 }, "inputs": { "x": { "dtype": "float32", "shape": [8, 64, 256], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02 } }, "w": { "dtype": "float32", "shape": [1, 768, 256], "data": { "kind": "fillFloat32", "sinStep": 0.03, "cosStep": 0.04, "scale": 0.05 } }, "r": { "dtype": "float32", "shape": [1, 768, 256], "data": { "kind": "fillFloat32", "sinStep": 0.05, "cosStep": 0.06, "scale": 0.05 } }, "b": { "dtype": "float32", "shape": [1, 1536], "data": { "kind": "constant", "value": 0.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [8, 64, 1, 256] }, "y_h": { "dtype": "float32", "shape": [8, 1, 256] } }, "bench": { "metrics": [{ "type": "gflops", "value": "2 * 64 * 8 * 3 * 256 * (256 + 256)" }] } }, { "name": "gru-layout0-time-major-h256-in256-b8-seq64-control", "preset": "edge", "attrs": { "layout": 0, "direction": "forward", "linear_before_reset": 0, "hidden_size": 256 }, "inputs": { "x": { "dtype": "float32", "shape": [64, 8, 256], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02 } }, "w": { "dtype": "float32", "shape": [1, 768, 256], "data": { "kind": "fillFloat32", "sinStep": 0.03, "cosStep": 0.04, "scale": 0.05 } }, "r": { "dtype": "float32", "shape": [1, 768, 256], "data": { "kind": "fillFloat32", "sinStep": 0.05, "cosStep": 0.06, "scale": 0.05 } }, "b": { "dtype": "float32", "shape": [1, 1536], "data": { "kind": "constant", "value": 0.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [64, 1, 8, 256] }, "y_h": { "dtype": "float32", "shape": [1, 8, 256] } }, "bench": { "metrics": [{ "type": "gflops", "value": "2 * 64 * 8 * 3 * 256 * (256 + 256)" }] } } ] }