diff --git a/include/tim/vx/ops/matmul.h b/include/tim/vx/ops/matmul.h new file mode 100644 index 0000000..af47a58 --- /dev/null +++ b/include/tim/vx/ops/matmul.h @@ -0,0 +1,59 @@ +/**************************************************************************** +* +* Copyright (c) 2021 Vivante Corporation +* +* Permission is hereby granted, free of charge, to any person obtaining a +* copy of this software and associated documentation files (the "Software"), +* to deal in the Software without restriction, including without limitation +* the rights to use, copy, modify, merge, publish, distribute, sublicense, +* and/or sell copies of the Software, and to permit persons to whom the +* Software is furnished to do so, subject to the following conditions: +* +* The above copyright notice and this permission notice shall be included in +* all copies or substantial portions of the Software. +* +* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER +* DEALINGS IN THE SOFTWARE. +* +*****************************************************************************/ +#ifndef TIM_VX_OPS_MATMUL_H_ +#define TIM_VX_OPS_MATMUL_H_ +#include "tim/vx/operation.h" + +namespace tim { +namespace vx { +namespace ops { + +/** + * ## Matmul + * + * Multiplies matrix a by matrix b, producing a * b. + * + * - transpose_a: If True, a is transposed before multiplication. + * - transpose_b: If True, b is transposed before multiplication. + * - adjoint_a: If True, a is conjugated and transposed before multiplication. + * - adjoint_b: If True, b is conjugated and transposed before multiplication. + */ + +class Matmul : public Operation { + public: + Matmul(Graph* graph, bool transpose_a = false, bool transpose_b = false, + bool adjoint_a = false, bool adjoint_b = false); + + protected: + bool transpose_a_; + bool transpose_b_; + bool adjoint_a_; + bool adjoint_b_; +}; + +} // namespace ops +} // namespace vx +} // namespace tim + +#endif /* TIM_VX_OPS_MATMUL_H_ */ \ No newline at end of file diff --git a/src/tim/vx/ops/README.md b/src/tim/vx/ops/README.md index 82dc49a..6a3357a 100644 --- a/src/tim/vx/ops/README.md +++ b/src/tim/vx/ops/README.md @@ -42,6 +42,7 @@ Elu|ELU|Mapped|[tf.nn.elu](https://tensorflow.google.cn/api_docs/python/tf/nn/el Batch2Space|BATCH2SPACE|Mapped|[tf.batch_to_space](https://tensorflow.google.cn/api_docs/python/tf/batch_to_space) Space2Batch|SPACE2BATCH|Mapped|[tf.space_to_batch](https://tensorflow.google.cn/api_docs/python/tf/space_to_batch) Pad|PAD|Mapped|[tf.pad](https://tensorflow.google.cn/api_docs/python/tf/pad) +Matmul|MATRIXMUL|Mapped|[tf.linalg.matmul](https://www.tensorflow.org/api_docs/python/tf/linalg/matmul) LayerNormalization|LAYER_NORM|Mapped|[tf.keras.layers.LayerNormalization](https://tensorflow.google.cn/api_docs/python/tf/keras/layers/LayerNormalization) ReduceMin|REDUCE_MIN|Mapped|[tf.math.reduce_min](https://tensorflow.google.cn/api_docs/python/tf/math/reduce_min) ReduceMax|REDUCE_MAX|Mapped|[tf.math.reduce_max](https://tensorflow.google.cn/api_docs/python/tf/math/reduce_max) diff --git a/src/tim/vx/ops/matmul.cc b/src/tim/vx/ops/matmul.cc new file mode 100644 index 0000000..b062e64 --- /dev/null +++ b/src/tim/vx/ops/matmul.cc @@ -0,0 +1,46 @@ +/**************************************************************************** +* +* Copyright (c) 2021 Vivante Corporation +* +* Permission is hereby granted, free of charge, to any person obtaining a +* copy of this software and associated documentation files (the "Software"), +* to deal in the Software without restriction, including without limitation +* the rights to use, copy, modify, merge, publish, distribute, sublicense, +* and/or sell copies of the Software, and to permit persons to whom the +* Software is furnished to do so, subject to the following conditions: +* +* The above copyright notice and this permission notice shall be included in +* all copies or substantial portions of the Software. +* +* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER +* DEALINGS IN THE SOFTWARE. +* +*****************************************************************************/ +#include "tim/vx/ops/matmul.h" + +#include "operation_private.h" +#include "vsi_nn_pub.h" +#include "type_utils.h" + +namespace tim { +namespace vx { +namespace ops { + +Matmul::Matmul(Graph* graph, bool transpose_a, bool transpose_b, + bool adjoint_a, bool adjoint_b) + : Operation(graph, VSI_NN_OP_MATRIXMUL), transpose_a_(transpose_a), + transpose_b_(transpose_b), adjoint_a_(adjoint_a), adjoint_b_(adjoint_b) { + this->impl()->node()->nn_param.matrixmul.transpose[0] = ToVxBool(transpose_a_); + this->impl()->node()->nn_param.matrixmul.transpose[1] = ToVxBool(transpose_b_); + this->impl()->node()->nn_param.matrixmul.adjoint[0] = ToVxBool(adjoint_a_); + this->impl()->node()->nn_param.matrixmul.adjoint[1] = ToVxBool(adjoint_b_); +} + +} // namespace ops +} // namespace vx +} // namespace tim \ No newline at end of file diff --git a/src/tim/vx/ops/matmul_test.cc b/src/tim/vx/ops/matmul_test.cc new file mode 100644 index 0000000..47ad77f --- /dev/null +++ b/src/tim/vx/ops/matmul_test.cc @@ -0,0 +1,204 @@ +/**************************************************************************** +* +* Copyright (c) 2021 Vivante Corporation +* +* Permission is hereby granted, free of charge, to any person obtaining a +* copy of this software and associated documentation files (the "Software"), +* to deal in the Software without restriction, including without limitation +* the rights to use, copy, modify, merge, publish, distribute, sublicense, +* and/or sell copies of the Software, and to permit persons to whom the +* Software is furnished to do so, subject to the following conditions: +* +* The above copyright notice and this permission notice shall be included in +* all copies or substantial portions of the Software. +* +* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER +* DEALINGS IN THE SOFTWARE. +* +*****************************************************************************/ +#include "tim/vx/context.h" +#include "tim/vx/graph.h" +#include "tim/vx/ops/matmul.h" + +#include "gtest/gtest.h" + +namespace { +template +::testing::AssertionResult ArraysMatch(const std::vector& expected, + const std::vector& actual, + T abs_error){ + for (size_t i = 0; i < expected.size(); ++i){ + EXPECT_NEAR(expected[i], actual[i], abs_error) << "at index:" << i; + } + + return ::testing::AssertionSuccess(); +} +} + +TEST(Matmul, shape_2_6_shape_6_2_float) { + auto ctx = tim::vx::Context::Create(); + auto graph = ctx->CreateGraph(); + + tim::vx::ShapeType a_shape({6, 2}); + tim::vx::ShapeType b_shape({2, 6}); + tim::vx::ShapeType out_shape({2, 2}); + tim::vx::TensorSpec a_spec(tim::vx::DataType::FLOAT32, + a_shape, tim::vx::TensorAttribute::INPUT); + tim::vx::TensorSpec b_spec(tim::vx::DataType::FLOAT32, + b_shape, tim::vx::TensorAttribute::INPUT); + tim::vx::TensorSpec out_spec(tim::vx::DataType::FLOAT32, + out_shape, tim::vx::TensorAttribute::OUTPUT); + + auto a_tensor = graph->CreateTensor(a_spec); + auto b_tensor = graph->CreateTensor(b_spec); + auto out_tensor = graph->CreateTensor(out_spec); + + std::vector a_data = { + 1, 2, 3, 4, 5, 6, + -1, -2, -3, -4, -5, -6 + }; + std::vector b_data = { + 6, 5, + 4, 3, + 2, 1, + -6, -5, + -4, -3, + -2, -1 + }; + std::vector golden = { + -36, -27, + 36, 27 + }; + + EXPECT_TRUE(a_tensor->CopyDataToTensor(a_data.data(), a_data.size() * sizeof(float))); + EXPECT_TRUE(b_tensor->CopyDataToTensor(b_data.data(), b_data.size() * sizeof(float))); + + auto op = graph->CreateOperation(); + (*op).BindInputs({a_tensor, b_tensor}).BindOutputs({out_tensor}); + + EXPECT_TRUE(graph->Compile()); + EXPECT_TRUE(graph->Run()); + + std::vector output(golden.size()); + EXPECT_TRUE(out_tensor->CopyDataFromTensor(output.data())); + EXPECT_TRUE(ArraysMatch(golden, output, 1e-5f)); +} + +TEST(Matmul, shape_2_3_2_shape_2_3_2_float_transpose_b) { + auto ctx = tim::vx::Context::Create(); + auto graph = ctx->CreateGraph(); + + tim::vx::ShapeType a_shape({2, 3, 2}); + tim::vx::ShapeType b_shape({2, 3, 2}); + tim::vx::ShapeType out_shape({3, 3, 2}); + tim::vx::TensorSpec a_spec(tim::vx::DataType::FLOAT32, + a_shape, tim::vx::TensorAttribute::INPUT); + tim::vx::TensorSpec b_spec(tim::vx::DataType::FLOAT32, + b_shape, tim::vx::TensorAttribute::INPUT); + tim::vx::TensorSpec out_spec(tim::vx::DataType::FLOAT32, + out_shape, tim::vx::TensorAttribute::OUTPUT); + + auto a_tensor = graph->CreateTensor(a_spec); + auto b_tensor = graph->CreateTensor(b_spec); + auto out_tensor = graph->CreateTensor(out_spec); + + std::vector a_data = { + 1, 2, + 3, 4, + 5, 6, + -1, -2, + -3, -4, + -5, -6 + }; + std::vector b_data = { + 6, 5, + 4, 3, + 2, 1, + -6, -5, + -4, -3, + -2, -1 + }; + std::vector golden = { + 16, 10, 4, + 38, 24, 10, + 60, 38, 16, + 16, 10, 4, + 38, 24, 10, + 60, 38, 16, + }; + + EXPECT_TRUE(a_tensor->CopyDataToTensor(a_data.data(), a_data.size() * sizeof(float))); + EXPECT_TRUE(b_tensor->CopyDataToTensor(b_data.data(), b_data.size() * sizeof(float))); + + auto op = graph->CreateOperation(false, true); + (*op).BindInputs({a_tensor, b_tensor}).BindOutputs({out_tensor}); + + EXPECT_TRUE(graph->Compile()); + EXPECT_TRUE(graph->Run()); + + std::vector output(golden.size() * sizeof(float)); + EXPECT_TRUE(out_tensor->CopyDataFromTensor(output.data())); + EXPECT_TRUE(ArraysMatch(golden, output, 1e-5f)); +} + +TEST(Matmul, shape_2_3_2_shape_2_3_2_uint8_transpose_a) { + auto ctx = tim::vx::Context::Create(); + auto graph = ctx->CreateGraph(); + + tim::vx::ShapeType a_shape({2, 3, 2}); + tim::vx::ShapeType b_shape({2, 3, 2}); + tim::vx::ShapeType out_shape({2, 2, 2}); + tim::vx::Quantization input_quant(tim::vx::QuantType::ASYMMETRIC, 1, 6); + tim::vx::Quantization output_quant(tim::vx::QuantType::ASYMMETRIC, 1, 0); + tim::vx::TensorSpec a_spec(tim::vx::DataType::UINT8, + a_shape, tim::vx::TensorAttribute::INPUT, input_quant); + tim::vx::TensorSpec b_spec(tim::vx::DataType::UINT8, + b_shape, tim::vx::TensorAttribute::INPUT, input_quant); + tim::vx::TensorSpec out_spec(tim::vx::DataType::UINT8, + out_shape, tim::vx::TensorAttribute::OUTPUT, output_quant); + + auto a_tensor = graph->CreateTensor(a_spec); + auto b_tensor = graph->CreateTensor(b_spec); + auto out_tensor = graph->CreateTensor(out_spec); + + std::vector a_data = { + 7, 8, + 9, 10, + 11, 12, + 5, 4, + 3, 2, + 1, 0, + }; + std::vector b_data = { + 12, 11, + 10, 9, + 8, 7, + 0, 1, + 2, 3, + 4, 5, + }; + std::vector golden = { + 28, 19, + 40, 28, + 28, 19, + 40, 28, + }; + + EXPECT_TRUE(a_tensor->CopyDataToTensor(a_data.data(), a_data.size())); + EXPECT_TRUE(b_tensor->CopyDataToTensor(b_data.data(), b_data.size())); + + auto op = graph->CreateOperation(true); + (*op).BindInputs({a_tensor, b_tensor}).BindOutputs({out_tensor}); + + EXPECT_TRUE(graph->Compile()); + EXPECT_TRUE(graph->Run()); + + std::vector output(golden.size()); + EXPECT_TRUE(out_tensor->CopyDataFromTensor(output.data())); + EXPECT_EQ(golden, output); +} \ No newline at end of file