| { |
| "domain": "com.microsoft", |
| "name": "CausalConvWithState", |
| "sinceVersion": 1, |
| "inputs": { |
| "inputT": { "onnx": "input", "dtype": "T", "rank": 3 }, |
| "weightT": { "onnx": "weight", "dtype": "T", "rank": 3 }, |
| "biasT": { "onnx": "bias", "dtype": "T", "rank": 1, "optional": true }, |
| "pastStateT": { |
| "onnx": "past_state", |
| "dtype": "T", |
| "rank": "3 if attrs.state_window == 0 else 4", |
| "optional": true |
| } |
| }, |
| "outputs": { |
| "outputT": { "onnx": "output", "dtype": "T", "rank": 3, "shape": "shapes.inputT" }, |
| "presentStateT": { |
| "onnx": "present_state", |
| "dtype": "T", |
| "rank": "3 if attrs.state_window == 0 else 4", |
| "shape": "[dim(shapes.inputT, 0), dim(shapes.inputT, 1), dim(shapes.weightT, 2) - 1] if attrs.state_window == 0 else [attrs.state_window, dim(shapes.inputT, 0), dim(shapes.inputT, 1), dim(shapes.weightT, 2) - 1]" |
| } |
| }, |
| "attributes": { "activation": { "default": "none" }, "ndim": { "default": 1 }, "state_window": { "default": 0 } }, |
| "attributeConstraints": { "activation": { "values": ["none", "silu", "swish"] }, "ndim": { "values": [1] } }, |
| "typeConstraints": { "T": ["float32", "float16"] }, |
| "tunables": { "workgroupSize": { "default": 256 }, "tiledWorkgroupSize": { "default": 128 } }, |
| "derive": { |
| "stateWindow": "attrs.state_window", |
| "windowed": "stateWindow > 0", |
| "stateWindowOk": "stateWindow >= 0 and stateWindow <= 8", |
| "kernelSize": "dim(shapes.weightT, ranks.weightT - 1)", |
| "kernelSizePadded": "ceilDiv(kernelSize, 4) * 4", |
| "weightRankOk": "ranks.weightT == 3 and dim(shapes.weightT, 1) == 1", |
| "stateLength": "kernelSize - 1", |
| "stateSlotStride": "dim(shapes.inputT, 0) * dim(shapes.inputT, 1) * stateLength", |
| "windowedLengthOk": "not windowed or dim(shapes.inputT, 2) > 0", |
| "presentStateOk": "(ranks.presentStateT == 3 and dim(shapes.presentStateT, 0) == dim(shapes.inputT, 0) and dim(shapes.presentStateT, 1) == dim(shapes.inputT, 1) and dim(shapes.presentStateT, 2) == stateLength) if not windowed else (ranks.presentStateT == 4 and dim(shapes.presentStateT, 0) == stateWindow and dim(shapes.presentStateT, 1) == dim(shapes.inputT, 0) and dim(shapes.presentStateT, 2) == dim(shapes.inputT, 1) and dim(shapes.presentStateT, 3) == stateLength)", |
| "pastStateShapeOk": "present.pastStateT and ((ranks.pastStateT == 3 and dim(shapes.pastStateT, 0) == dim(shapes.inputT, 0) and dim(shapes.pastStateT, 1) == dim(shapes.inputT, 1) and dim(shapes.pastStateT, 2) == stateLength) if not windowed else (ranks.pastStateT == 4 and dim(shapes.pastStateT, 0) == stateWindow and dim(shapes.pastStateT, 1) == dim(shapes.inputT, 0) and dim(shapes.pastStateT, 2) == dim(shapes.inputT, 1) and dim(shapes.pastStateT, 3) == stateLength))", |
| "commonContract": "ranks.inputT == 3 and weightRankOk and ranks.outputT == 3 and (tensorDtypes.inputT == \"float32\" or tensorDtypes.inputT == \"float16\") and tensorDtypes.weightT == tensorDtypes.inputT and tensorDtypes.outputT == tensorDtypes.inputT and tensorDtypes.presentStateT == tensorDtypes.inputT and f16Ok(dtypes.T) and dim(shapes.inputT, 1) == dim(shapes.weightT, 0) and dim(shapes.outputT, 0) == dim(shapes.inputT, 0) and dim(shapes.outputT, 1) == dim(shapes.inputT, 1) and dim(shapes.outputT, 2) == dim(shapes.inputT, 2) and stateWindowOk and windowedLengthOk and presentStateOk", |
| "zeroStateContract": "commonContract and not present.pastStateT and not present.biasT", |
| "biasNoStateContract": "commonContract and not present.pastStateT and present.biasT and ranks.biasT == 1 and tensorDtypes.biasT == tensorDtypes.inputT and dim(shapes.biasT, 0) == dim(shapes.inputT, 1)", |
| "stateNoBiasContract": "commonContract and present.pastStateT and not present.biasT and tensorDtypes.pastStateT == tensorDtypes.inputT and pastStateShapeOk", |
| "stateBiasContract": "commonContract and present.pastStateT and present.biasT and ranks.biasT == 1 and tensorDtypes.pastStateT == tensorDtypes.inputT and tensorDtypes.biasT == tensorDtypes.inputT and pastStateShapeOk and dim(shapes.biasT, 0) == dim(shapes.inputT, 1)" |
| }, |
| "bindings": { |
| "input": { "arg": "inputT", "buffer": "read-only-storage", "elementType": "$inputVec4" }, |
| "weight": { "arg": "weightT", "buffer": "read-only-storage", "elementType": "$weightElem" }, |
| "output": { "arg": "outputT", "buffer": "storage", "elementType": "$outputVec4" }, |
| "present_state": { "arg": "presentStateT", "buffer": "storage", "elementType": "$outputScalar" }, |
| "params": { |
| "buffer": "uniform", |
| "struct": [ |
| { "name": "batchSize", "type": "u32", "value": "dim(shapes.inputT, 0)" }, |
| { "name": "channels", "type": "u32", "value": "dim(shapes.inputT, 1)" }, |
| { "name": "length", "type": "u32", "value": "dim(shapes.inputT, 2)" }, |
| { "name": "stateWindow", "type": "u32", "value": "stateWindow" }, |
| { "name": "stateSlotStride", "type": "u32", "value": "stateSlotStride" } |
| ] |
| }, |
| "bias": { "arg": "biasT", "buffer": "read-only-storage", "elementType": "$inputScalar" }, |
| "past_state": { "arg": "pastStateT", "buffer": "read-only-storage", "elementType": "$inputScalar" }, |
| "input_2": { "arg": "inputT", "name": "input", "buffer": "read-only-storage", "elementType": "$inputScalar" }, |
| "weight_2": { "arg": "weightT", "name": "weight", "buffer": "read-only-storage", "elementType": "$inputScalar" }, |
| "output_2": { "arg": "outputT", "name": "output", "buffer": "storage", "elementType": "$outputScalar" }, |
| "params_2": { |
| "name": "params", |
| "buffer": "uniform", |
| "struct": [ |
| { "name": "batchSize", "type": "u32", "value": "dim(shapes.inputT, 0)" }, |
| { "name": "channels", "type": "u32", "value": "dim(shapes.inputT, 1)" }, |
| { "name": "length", "type": "u32", "value": "dim(shapes.inputT, 2)" }, |
| { "name": "kernelSize", "type": "u32", "value": "kernelSize" }, |
| { "name": "stateWindow", "type": "u32", "value": "stateWindow" }, |
| { "name": "stateSlotStride", "type": "u32", "value": "stateSlotStride" } |
| ] |
| } |
| }, |
| "variants": [ |
| { |
| "id": "zero_state_vec4", |
| "priority": 20, |
| "when": ["zeroStateContract", "kernelSize >= 2", "kernelSize <= 4", "dim(shapes.inputT, 2) >= 4", "dim(shapes.inputT, 2) % 4 == 0"], |
| "derive": { |
| "useSilu": "attrs.activation == \"silu\" or attrs.activation == \"swish\"", |
| "workgroupSize": 256, |
| "hasStateWindow": "windowed", |
| "usesF16": "tensorDtypes.inputT == \"float16\"", |
| "inputScalar": "dtypes.T", |
| "outputScalar": "dtypes.T", |
| "inputVec4": "\"vec4<f16>\" if tensorDtypes.inputT == \"float16\" else \"vec4<f32>\"", |
| "outputVec4": "\"vec4<f16>\" if tensorDtypes.inputT == \"float16\" else \"vec4<f32>\"", |
| "hasBias": false, |
| "hasState": false, |
| "weightElem": "(\"vec4<f16>\" if tensorDtypes.inputT == \"float16\" else \"vec4<f32>\") if kernelSize == 4 else dtypes.T" |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "CausalConvWithState.Vec4", |
| "shader": "causal-conv-with-state-vec4.wgsl.jinja", |
| "derive": { "materializeConvBeforeActivation": false }, |
| "bindings": ["input", "weight", "output", "present_state", "params"], |
| "dispatch": { |
| "x": "min(ceilDiv((dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * (dim(shapes.outputT, 2) / 4)), (workgroupSize)), 65535)", |
| "y": "ceilDiv(ceilDiv((dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * (dim(shapes.outputT, 2) / 4)), (workgroupSize)), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "state_bias_vec4", |
| "priority": 20, |
| "when": ["stateBiasContract", "kernelSize >= 2", "kernelSize <= 4", "dim(shapes.inputT, 2) >= 4", "dim(shapes.inputT, 2) % 4 == 0"], |
| "derive": { |
| "useSilu": "attrs.activation == \"silu\" or attrs.activation == \"swish\"", |
| "workgroupSize": 256, |
| "hasStateWindow": "windowed", |
| "usesF16": "tensorDtypes.inputT == \"float16\"", |
| "inputScalar": "dtypes.T", |
| "outputScalar": "dtypes.T", |
| "inputVec4": "\"vec4<f16>\" if tensorDtypes.inputT == \"float16\" else \"vec4<f32>\"", |
| "outputVec4": "\"vec4<f16>\" if tensorDtypes.inputT == \"float16\" else \"vec4<f32>\"", |
| "hasBias": true, |
| "hasState": true, |
| "weightElem": "(\"vec4<f16>\" if tensorDtypes.inputT == \"float16\" else \"vec4<f32>\") if kernelSize == 4 else dtypes.T" |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "CausalConvWithState.Vec4", |
| "shader": "causal-conv-with-state-vec4.wgsl.jinja", |
| "derive": { "materializeConvBeforeActivation": false }, |
| "bindings": ["input", "weight", "bias", "past_state", "output", "present_state", "params"], |
| "dispatch": { |
| "x": "min(ceilDiv((dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * (dim(shapes.outputT, 2) / 4)), (workgroupSize)), 65535)", |
| "y": "ceilDiv(ceilDiv((dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * (dim(shapes.outputT, 2) / 4)), (workgroupSize)), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "bias_no_state_vec4", |
| "priority": 20, |
| "when": ["biasNoStateContract", "kernelSize >= 2", "kernelSize <= 4", "dim(shapes.inputT, 2) >= 4", "dim(shapes.inputT, 2) % 4 == 0"], |
| "derive": { |
| "useSilu": "attrs.activation == \"silu\" or attrs.activation == \"swish\"", |
| "workgroupSize": 256, |
| "hasStateWindow": "windowed", |
| "usesF16": "tensorDtypes.inputT == \"float16\"", |
| "inputScalar": "dtypes.T", |
| "outputScalar": "dtypes.T", |
| "inputVec4": "\"vec4<f16>\" if tensorDtypes.inputT == \"float16\" else \"vec4<f32>\"", |
| "outputVec4": "\"vec4<f16>\" if tensorDtypes.inputT == \"float16\" else \"vec4<f32>\"", |
| "hasBias": true, |
| "hasState": false, |
| "weightElem": "(\"vec4<f16>\" if tensorDtypes.inputT == \"float16\" else \"vec4<f32>\") if kernelSize == 4 else dtypes.T" |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "CausalConvWithState.Vec4", |
| "shader": "causal-conv-with-state-vec4.wgsl.jinja", |
| "derive": { "materializeConvBeforeActivation": false }, |
| "bindings": ["input", "weight", "bias", "output", "present_state", "params"], |
| "dispatch": { |
| "x": "min(ceilDiv((dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * (dim(shapes.outputT, 2) / 4)), (workgroupSize)), 65535)", |
| "y": "ceilDiv(ceilDiv((dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * (dim(shapes.outputT, 2) / 4)), (workgroupSize)), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "state_no_bias_vec4", |
| "priority": 20, |
| "when": ["stateNoBiasContract", "kernelSize >= 2", "kernelSize <= 4", "dim(shapes.inputT, 2) >= 4", "dim(shapes.inputT, 2) % 4 == 0"], |
| "derive": { |
| "useSilu": "attrs.activation == \"silu\" or attrs.activation == \"swish\"", |
| "workgroupSize": 256, |
| "hasStateWindow": "windowed", |
| "usesF16": "tensorDtypes.inputT == \"float16\"", |
| "inputScalar": "dtypes.T", |
| "outputScalar": "dtypes.T", |
| "inputVec4": "\"vec4<f16>\" if tensorDtypes.inputT == \"float16\" else \"vec4<f32>\"", |
| "outputVec4": "\"vec4<f16>\" if tensorDtypes.inputT == \"float16\" else \"vec4<f32>\"", |
| "hasBias": false, |
| "hasState": true, |
| "weightElem": "(\"vec4<f16>\" if tensorDtypes.inputT == \"float16\" else \"vec4<f32>\") if kernelSize == 4 else dtypes.T" |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "CausalConvWithState.Vec4", |
| "shader": "causal-conv-with-state-vec4.wgsl.jinja", |
| "derive": { "materializeConvBeforeActivation": false }, |
| "bindings": ["input", "weight", "past_state", "output", "present_state", "params"], |
| "dispatch": { |
| "x": "min(ceilDiv((dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * (dim(shapes.outputT, 2) / 4)), (workgroupSize)), 65535)", |
| "y": "ceilDiv(ceilDiv((dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * (dim(shapes.outputT, 2) / 4)), (workgroupSize)), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "zero_state_tiled_large_kernel", |
| "priority": 10, |
| "when": ["zeroStateContract", "kernelSize >= 32", "dim(shapes.inputT, 2) >= 256", "dim(shapes.inputT, 2) % 8 == 0", "tunables.tiledWorkgroupSize >= 1", "floor(tunables.tiledWorkgroupSize) == tunables.tiledWorkgroupSize", "tunables.tiledWorkgroupSize <= device.limits.maxComputeInvocationsPerWorkgroup", "tunables.tiledWorkgroupSize <= device.limits.maxComputeWorkgroupSizeX", "(tunables.tiledWorkgroupSize * 8 + 2 * kernelSizePadded - 1) * 4 <= device.limits.maxComputeWorkgroupStorageSize"], |
| "derive": { |
| "hasBias": false, |
| "hasState": false, |
| "useSilu": "attrs.activation == \"silu\" or attrs.activation == \"swish\"", |
| "inputScalar": "dtypes.T", |
| "outputScalar": "dtypes.T", |
| "workgroupSize": "tunables.tiledWorkgroupSize", |
| "tileSize": "tunables.tiledWorkgroupSize * 8", |
| "inputTileSize": "tunables.tiledWorkgroupSize * 8 + kernelSizePadded - 1", |
| "hasStateWindow": "windowed", |
| "usesF16": "tensorDtypes.inputT == \"float16\"" |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "CausalConvWithState.TiledLargeKernel", |
| "shader": "causal-conv-with-state-tiled.wgsl.jinja", |
| "derive": { "materializeConvBeforeActivation": false }, |
| "bindings": ["input_2", "weight_2", "output_2", "present_state", "params"], |
| "dispatch": { |
| "x": "min(dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * ceilDiv(dim(shapes.outputT, 2), tileSize), 65535)", |
| "y": "ceilDiv(dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * ceilDiv(dim(shapes.outputT, 2), tileSize), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "state_bias_tiled_large_kernel", |
| "priority": 10, |
| "when": ["stateBiasContract", "kernelSize >= 32", "dim(shapes.inputT, 2) >= 256", "dim(shapes.inputT, 2) % 8 == 0", "tunables.tiledWorkgroupSize >= 1", "floor(tunables.tiledWorkgroupSize) == tunables.tiledWorkgroupSize", "tunables.tiledWorkgroupSize <= device.limits.maxComputeInvocationsPerWorkgroup", "tunables.tiledWorkgroupSize <= device.limits.maxComputeWorkgroupSizeX", "(tunables.tiledWorkgroupSize * 8 + 2 * kernelSizePadded - 1) * 4 <= device.limits.maxComputeWorkgroupStorageSize"], |
| "derive": { |
| "hasBias": true, |
| "hasState": true, |
| "useSilu": "attrs.activation == \"silu\" or attrs.activation == \"swish\"", |
| "inputScalar": "dtypes.T", |
| "outputScalar": "dtypes.T", |
| "workgroupSize": "tunables.tiledWorkgroupSize", |
| "tileSize": "tunables.tiledWorkgroupSize * 8", |
| "inputTileSize": "tunables.tiledWorkgroupSize * 8 + kernelSizePadded - 1", |
| "hasStateWindow": "windowed", |
| "usesF16": "tensorDtypes.inputT == \"float16\"" |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "CausalConvWithState.TiledLargeKernel", |
| "shader": "causal-conv-with-state-tiled.wgsl.jinja", |
| "derive": { "materializeConvBeforeActivation": false }, |
| "bindings": ["input_2", "weight_2", "bias", "past_state", "output_2", "present_state", "params"], |
| "dispatch": { |
| "x": "min(dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * ceilDiv(dim(shapes.outputT, 2), tileSize), 65535)", |
| "y": "ceilDiv(dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * ceilDiv(dim(shapes.outputT, 2), tileSize), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "bias_no_state_tiled_large_kernel", |
| "priority": 10, |
| "when": ["biasNoStateContract", "kernelSize >= 32", "dim(shapes.inputT, 2) >= 256", "dim(shapes.inputT, 2) % 8 == 0", "tunables.tiledWorkgroupSize >= 1", "floor(tunables.tiledWorkgroupSize) == tunables.tiledWorkgroupSize", "tunables.tiledWorkgroupSize <= device.limits.maxComputeInvocationsPerWorkgroup", "tunables.tiledWorkgroupSize <= device.limits.maxComputeWorkgroupSizeX", "(tunables.tiledWorkgroupSize * 8 + 2 * kernelSizePadded - 1) * 4 <= device.limits.maxComputeWorkgroupStorageSize"], |
| "derive": { |
| "hasBias": true, |
| "hasState": false, |
| "useSilu": "attrs.activation == \"silu\" or attrs.activation == \"swish\"", |
| "inputScalar": "dtypes.T", |
| "outputScalar": "dtypes.T", |
| "workgroupSize": "tunables.tiledWorkgroupSize", |
| "tileSize": "tunables.tiledWorkgroupSize * 8", |
| "inputTileSize": "tunables.tiledWorkgroupSize * 8 + kernelSizePadded - 1", |
| "hasStateWindow": "windowed", |
| "usesF16": "tensorDtypes.inputT == \"float16\"" |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "CausalConvWithState.TiledLargeKernel", |
| "shader": "causal-conv-with-state-tiled.wgsl.jinja", |
| "derive": { "materializeConvBeforeActivation": false }, |
| "bindings": ["input_2", "weight_2", "bias", "output_2", "present_state", "params"], |
| "dispatch": { |
| "x": "min(dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * ceilDiv(dim(shapes.outputT, 2), tileSize), 65535)", |
| "y": "ceilDiv(dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * ceilDiv(dim(shapes.outputT, 2), tileSize), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "state_no_bias_tiled_large_kernel", |
| "priority": 10, |
| "when": ["stateNoBiasContract", "kernelSize >= 32", "dim(shapes.inputT, 2) >= 256", "dim(shapes.inputT, 2) % 8 == 0", "tunables.tiledWorkgroupSize >= 1", "floor(tunables.tiledWorkgroupSize) == tunables.tiledWorkgroupSize", "tunables.tiledWorkgroupSize <= device.limits.maxComputeInvocationsPerWorkgroup", "tunables.tiledWorkgroupSize <= device.limits.maxComputeWorkgroupSizeX", "(tunables.tiledWorkgroupSize * 8 + 2 * kernelSizePadded - 1) * 4 <= device.limits.maxComputeWorkgroupStorageSize"], |
| "derive": { |
| "hasBias": false, |
| "hasState": true, |
| "useSilu": "attrs.activation == \"silu\" or attrs.activation == \"swish\"", |
| "inputScalar": "dtypes.T", |
| "outputScalar": "dtypes.T", |
| "workgroupSize": "tunables.tiledWorkgroupSize", |
| "tileSize": "tunables.tiledWorkgroupSize * 8", |
| "inputTileSize": "tunables.tiledWorkgroupSize * 8 + kernelSizePadded - 1", |
| "hasStateWindow": "windowed", |
| "usesF16": "tensorDtypes.inputT == \"float16\"" |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "CausalConvWithState.TiledLargeKernel", |
| "shader": "causal-conv-with-state-tiled.wgsl.jinja", |
| "derive": { "materializeConvBeforeActivation": false }, |
| "bindings": ["input_2", "weight_2", "past_state", "output_2", "present_state", "params"], |
| "dispatch": { |
| "x": "min(dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * ceilDiv(dim(shapes.outputT, 2), tileSize), 65535)", |
| "y": "ceilDiv(dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * ceilDiv(dim(shapes.outputT, 2), tileSize), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "zero_state", |
| "priority": 0, |
| "when": ["zeroStateContract"], |
| "derive": { |
| "hasBias": false, |
| "hasState": false, |
| "useSilu": "attrs.activation == \"silu\" or attrs.activation == \"swish\"", |
| "inputScalar": "dtypes.T", |
| "outputScalar": "dtypes.T", |
| "workgroupSize": "tunables.workgroupSize", |
| "hasStateWindow": "windowed", |
| "usesF16": "tensorDtypes.inputT == \"float16\"" |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "CausalConvWithState", |
| "shader": "causal-conv-with-state.wgsl.jinja", |
| "derive": { "materializeConvBeforeActivation": false }, |
| "bindings": ["input_2", "weight_2", "output_2", "present_state", "params_2"], |
| "dispatch": { |
| "x": "min(ceilDiv((dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * max(1, dim(shapes.outputT, 2))), (workgroupSize)), 65535)", |
| "y": "ceilDiv(ceilDiv((dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * max(1, dim(shapes.outputT, 2))), (workgroupSize)), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "state_bias", |
| "priority": 0, |
| "when": ["stateBiasContract"], |
| "derive": { |
| "hasBias": true, |
| "hasState": true, |
| "useSilu": "attrs.activation == \"silu\" or attrs.activation == \"swish\"", |
| "inputScalar": "dtypes.T", |
| "outputScalar": "dtypes.T", |
| "workgroupSize": "tunables.workgroupSize", |
| "hasStateWindow": "windowed", |
| "usesF16": "tensorDtypes.inputT == \"float16\"" |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "CausalConvWithState", |
| "shader": "causal-conv-with-state.wgsl.jinja", |
| "derive": { "materializeConvBeforeActivation": false }, |
| "bindings": ["input_2", "weight_2", "bias", "past_state", "output_2", "present_state", "params_2"], |
| "dispatch": { |
| "x": "min(ceilDiv((dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * max(1, dim(shapes.outputT, 2))), (workgroupSize)), 65535)", |
| "y": "ceilDiv(ceilDiv((dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * max(1, dim(shapes.outputT, 2))), (workgroupSize)), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "bias_no_state", |
| "priority": 0, |
| "when": ["biasNoStateContract"], |
| "derive": { |
| "hasBias": true, |
| "hasState": false, |
| "useSilu": "attrs.activation == \"silu\" or attrs.activation == \"swish\"", |
| "inputScalar": "dtypes.T", |
| "outputScalar": "dtypes.T", |
| "workgroupSize": "tunables.workgroupSize", |
| "hasStateWindow": "windowed", |
| "usesF16": "tensorDtypes.inputT == \"float16\"" |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "CausalConvWithState", |
| "shader": "causal-conv-with-state.wgsl.jinja", |
| "derive": { "materializeConvBeforeActivation": false }, |
| "bindings": ["input_2", "weight_2", "bias", "output_2", "present_state", "params_2"], |
| "dispatch": { |
| "x": "min(ceilDiv((dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * max(1, dim(shapes.outputT, 2))), (workgroupSize)), 65535)", |
| "y": "ceilDiv(ceilDiv((dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * max(1, dim(shapes.outputT, 2))), (workgroupSize)), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| }, |
| { |
| "id": "state_no_bias", |
| "priority": 0, |
| "when": ["stateNoBiasContract"], |
| "derive": { |
| "hasBias": false, |
| "hasState": true, |
| "useSilu": "attrs.activation == \"silu\" or attrs.activation == \"swish\"", |
| "inputScalar": "dtypes.T", |
| "outputScalar": "dtypes.T", |
| "workgroupSize": "tunables.workgroupSize", |
| "hasStateWindow": "windowed", |
| "usesF16": "tensorDtypes.inputT == \"float16\"" |
| }, |
| "passes": [ |
| { |
| "id": "main", |
| "name": "CausalConvWithState", |
| "shader": "causal-conv-with-state.wgsl.jinja", |
| "derive": { "materializeConvBeforeActivation": false }, |
| "bindings": ["input_2", "weight_2", "past_state", "output_2", "present_state", "params_2"], |
| "dispatch": { |
| "x": "min(ceilDiv((dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * max(1, dim(shapes.outputT, 2))), (workgroupSize)), 65535)", |
| "y": "ceilDiv(ceilDiv((dim(shapes.outputT, 0) * dim(shapes.outputT, 1) * max(1, dim(shapes.outputT, 2))), (workgroupSize)), 65535)", |
| "z": 1 |
| } |
| } |
| ] |
| } |
| ] |
| } |
|
|
