|
| 1 | +// Copyright (c) Microsoft Corporation. All rights reserved. |
| 2 | +// Licensed under the MIT License. |
| 3 | + |
| 4 | +#include "core/providers/webgpu/tensor/split.h" |
| 5 | +#include "core/providers/webgpu/shader_helper.h" |
| 6 | +#include "core/providers/webgpu/webgpu_supported_types.h" |
| 7 | + |
| 8 | +namespace onnxruntime { |
| 9 | +namespace webgpu { |
| 10 | + |
| 11 | +namespace { |
| 12 | + |
| 13 | +// Helper function to calculate the output index based on the input index and the sizes of the splits. |
| 14 | +void CalculateOutputIndex(std::ostream& os, size_t output_count) { |
| 15 | + os << "fn calculate_output_index(index: u32) -> u32 {\n" |
| 16 | + << " for (var i: u32 = 0u; i < " << output_count << "u; i += 1u ) {\n" |
| 17 | + << " if (index < " << GetElementAt("uniforms.sizes_in_split_axis", "i", output_count) << ") {\n" |
| 18 | + << " return i;\n" |
| 19 | + << " }\n" |
| 20 | + << " }\n" |
| 21 | + << " return " << output_count << "u;\n" |
| 22 | + << "}\n"; |
| 23 | +} |
| 24 | + |
| 25 | +// Helper function to write the buffer data for each output. |
| 26 | +void WriteBufferData(std::ostream& os, const ShaderVariableHelper& input, |
| 27 | + gsl::span<const ShaderVariableHelper*> outputs) { |
| 28 | + os << "fn write_buffer_data(output_number: u32, global_idx: u32, indices: output_0_indices_t) {\n"; |
| 29 | + for (size_t i = 0; i < outputs.size(); ++i) { |
| 30 | + const auto buffer_write = outputs[i]->SetByIndices("indices", input.GetByOffset("global_idx")); |
| 31 | + if (outputs.size() == 1) { |
| 32 | + os << buffer_write; |
| 33 | + } else if (i == 0) { |
| 34 | + os << " if (output_number == 0u) {\n" |
| 35 | + << " " << buffer_write << "\n"; |
| 36 | + } else if (i == outputs.size() - 1) { |
| 37 | + os << " } else {\n" |
| 38 | + << " " << buffer_write << "\n"; |
| 39 | + } else { |
| 40 | + os << " } else if (output_number == " << i << "u) {\n" |
| 41 | + << " " << buffer_write << "\n"; |
| 42 | + } |
| 43 | + } |
| 44 | + os << " }\n" |
| 45 | + << "}\n"; |
| 46 | +} |
| 47 | + |
| 48 | +} // namespace |
| 49 | + |
| 50 | +Status SplitProgram::GenerateShaderCode(ShaderHelper& shader) const { |
| 51 | + const auto& input = shader.AddInput("input", ShaderUsage::UseUniform | ShaderUsage::UseIndicesTypeAlias); |
| 52 | + |
| 53 | + size_t output_count = Outputs().size(); |
| 54 | + std::vector<const ShaderVariableHelper*> outputs; |
| 55 | + outputs.reserve(output_count); |
| 56 | + for (size_t i = 0; i < output_count; ++i) { |
| 57 | + outputs.push_back( |
| 58 | + &shader.AddOutput("output_" + std::to_string(i), ShaderUsage::UseUniform | ShaderUsage::UseIndicesTypeAlias)); |
| 59 | + } |
| 60 | + |
| 61 | + // Add implementation of fn calculate_output_index. |
| 62 | + CalculateOutputIndex(shader.AdditionalImplementation(), output_count); |
| 63 | + // Add implementation of fn write_buffer_data. |
| 64 | + WriteBufferData(shader.AdditionalImplementation(), input, outputs); |
| 65 | + |
| 66 | + shader.MainFunctionBody() << shader.GuardAgainstOutOfBoundsWorkgroupSizes("uniforms.input_size") |
| 67 | + << " var indices = " << input.OffsetToIndices("global_idx") << ";\n" |
| 68 | + << " var index = indices[" << axis_ << "];\n" |
| 69 | + << " let output_number = calculate_output_index(index);\n" |
| 70 | + << " if (output_number != 0u) {\n" |
| 71 | + << " index -= uniforms.sizes_in_split_axis[output_number - 1u];\n" |
| 72 | + << " indices[" << axis_ << "] = index;\n" |
| 73 | + << " }\n" |
| 74 | + << " write_buffer_data(output_number, global_idx, indices);\n"; |
| 75 | + |
| 76 | + return Status::OK(); |
| 77 | +} |
| 78 | + |
| 79 | +Status Split::ComputeInternal(ComputeContext& context) const { |
| 80 | + const Tensor* input = context.Input<Tensor>(0); |
| 81 | + auto& input_shape = input->Shape(); |
| 82 | + auto num_outputs = context.OutputCount(); |
| 83 | + |
| 84 | + int64_t axis = axis_; |
| 85 | + std::vector<int64_t> split_sizes; |
| 86 | + |
| 87 | + split_sizes.assign(split_sizes_.begin(), split_sizes_.end()); |
| 88 | + // Compute split_sizes from the 'split' input tensor. |
| 89 | + if (split_sizes_.size() == 0 && context.InputCount() > 1) { |
| 90 | + const Tensor* split_tensor = context.Input<Tensor>(1); |
| 91 | + // Check if split_tensor is valid. |
| 92 | + if (split_tensor != nullptr) { |
| 93 | + ORT_ENFORCE(split_tensor->Shape().NumDimensions() == 1, "The split tensor must be a vector tensor."); |
| 94 | + // Get split_sizes from the input tensor. |
| 95 | + auto nDims = static_cast<size_t>(split_tensor->Shape()[0]); |
| 96 | + const auto* data = split_tensor->Data<int64_t>(); |
| 97 | + split_sizes.assign(data, data + nDims); |
| 98 | + } |
| 99 | + } |
| 100 | + |
| 101 | + // The variables below are not actually used in the current implementation. |
| 102 | + int before_dims = 0; |
| 103 | + int after_dims_including_split_axis = 0; |
| 104 | + int after_dims_excluding_split = 0; |
| 105 | + // This handles the case where the axis is negative. It also splits outputs evenly according to num_ouputs if |
| 106 | + // split_sizes is empty. |
| 107 | + ORT_RETURN_IF_ERROR(PrepareForCompute(input_shape, num_outputs, axis, before_dims, after_dims_including_split_axis, |
| 108 | + after_dims_excluding_split, split_sizes)); |
| 109 | + |
| 110 | + SplitProgram program{gsl::narrow_cast<uint32_t>(axis)}; |
| 111 | + program.AddInput({input, ProgramTensorMetadataDependency::TypeAndRank}); |
| 112 | + |
| 113 | + auto output_dimensions = input_shape.AsShapeVector(); |
| 114 | + for (int i = 0; i < num_outputs; ++i) { |
| 115 | + // Update the size of dimension for axis we're splitting on. |
| 116 | + auto split_size = narrow<int>(split_sizes[i]); |
| 117 | + output_dimensions[narrow<size_t>(axis)] = split_size; |
| 118 | + |
| 119 | + Tensor* output = context.Output(i, TensorShape{output_dimensions}); |
| 120 | + program.AddOutput({output, ProgramTensorMetadataDependency::Rank}); |
| 121 | + } |
| 122 | + |
| 123 | + uint32_t input_size = gsl::narrow<uint32_t>(input_shape.Size()); |
| 124 | + // Early return if the input tensor is empty. |
| 125 | + if (input_size == 0) { |
| 126 | + return Status::OK(); |
| 127 | + } |
| 128 | + |
| 129 | + uint32_t previous_sum = 0; |
| 130 | + std::vector<uint32_t> sizes_in_split_axis; |
| 131 | + // sizes_in_split_axis are the cumulative sizes of the splits in the split axis. |
| 132 | + for (auto split_size : split_sizes) { |
| 133 | + previous_sum += gsl::narrow<uint32_t>(split_size); |
| 134 | + sizes_in_split_axis.push_back(previous_sum); |
| 135 | + } |
| 136 | + |
| 137 | + program |
| 138 | + .SetDispatchGroupSize((input_size + WORKGROUP_SIZE - 1) / WORKGROUP_SIZE) |
| 139 | + .CacheHint(std::to_string(axis)) |
| 140 | + .AddUniformVariables( |
| 141 | + {input_size, gsl::span<const uint32_t>(sizes_in_split_axis.data(), sizes_in_split_axis.size())}); |
| 142 | + return context.RunProgram(program); |
| 143 | +} |
| 144 | + |
| 145 | +#define WEBGPU_SPLIT_KERNEL(OP_TYPE, VERSION, KERNEL_CLASS, TYPE) \ |
| 146 | + ONNX_OPERATOR_KERNEL_EX(OP_TYPE, kOnnxDomain, VERSION, kWebGpuExecutionProvider, \ |
| 147 | + KernelDefBuilder().TypeConstraint("T", TYPE).InputMemoryType(OrtMemTypeCPU, 1), \ |
| 148 | + KERNEL_CLASS); |
| 149 | + |
| 150 | +#define WEBGPU_SPLIT_VERSIONED_KERNEL(OP_TYPE, VERSION_FROM, VERSION_TO, KERNEL_CLASS, TYPE) \ |
| 151 | + ONNX_OPERATOR_VERSIONED_KERNEL_EX(OP_TYPE, kOnnxDomain, VERSION_FROM, VERSION_TO, kWebGpuExecutionProvider, \ |
| 152 | + KernelDefBuilder().TypeConstraint("T", TYPE).InputMemoryType(OrtMemTypeCPU, 1), \ |
| 153 | + KERNEL_CLASS); |
| 154 | + |
| 155 | +WEBGPU_SPLIT_VERSIONED_KERNEL(Split, 1, 1, Split_1, WebGpuSupportedNumberTypes()) |
| 156 | +WEBGPU_SPLIT_VERSIONED_KERNEL(Split, 2, 10, Split_2_10, WebGpuSupportedNumberTypes()) |
| 157 | +WEBGPU_SPLIT_VERSIONED_KERNEL(Split, 11, 12, Split_11_12, WebGpuSupportedNumberTypes()) |
| 158 | +WEBGPU_SPLIT_VERSIONED_KERNEL(Split, 13, 17, Split_13_17, WebGpuSupportedNumberTypes()) |
| 159 | +WEBGPU_SPLIT_KERNEL(Split, 18, Split_18, WebGpuSupportedNumberTypes()); |
| 160 | + |
| 161 | +} // namespace webgpu |
| 162 | +} // namespace onnxruntime |
0 commit comments