diff options
Diffstat (limited to 'tests/validation/fixtures/GEMMFixture.h')
| -rw-r--r-- | tests/validation/fixtures/GEMMFixture.h | 500 |
1 files changed, 500 insertions, 0 deletions
diff --git a/tests/validation/fixtures/GEMMFixture.h b/tests/validation/fixtures/GEMMFixture.h index ac8ab2a949..b36bb99246 100644 --- a/tests/validation/fixtures/GEMMFixture.h +++ b/tests/validation/fixtures/GEMMFixture.h @@ -153,6 +153,506 @@ protected: SimpleTensor<T> _reference{}; }; +template <typename TensorType, typename AccessorType, typename T, typename GEMMFunctionType> +class GEMMMatrixMultiplyValidationFixture : public framework::Fixture +{ +public: + template <typename...> + void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, float alpha, float beta, bool broadcast_bias, bool fp16_mixed_precision, const ActivationLayerInfo &act_info, + DataType data_type, GPUTarget gpu_arch) + { + // Set the tensor shapes for LHS and RHS matrices + const TensorShape lhs_shape(k, m, batch_size); + const TensorShape rhs_shape(n, k, batch_size); + const TensorShape bias_shape(n, + broadcast_bias ? 1 : m, + broadcast_bias ? 1 : batch_size); + + _target = compute_target(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, broadcast_bias, fp16_mixed_precision, act_info, gpu_arch); + _reference = compute_reference(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, broadcast_bias, act_info); + } + +protected: + template <typename U> + void fill(U &&tensor, int i) + { + std::uniform_real_distribution<> distribution(-1.0f, 1.0f); + library->fill(tensor, distribution, i); + + // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0) + std::uniform_real_distribution<> distribution_inf(std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity()); + library->fill_borders_with_garbage(tensor, distribution_inf, i); + } + + TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, DataType data_type, float alpha, float beta, bool broadcast_bias, + bool fp16_mixed_precision, const ActivationLayerInfo &act_info, GPUTarget gpu_arch) + { + // Create tensors + TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); + TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); + TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); + TensorType dst; + + const unsigned int m = lhs_shape[1]; + const unsigned int n = rhs_shape[0]; + const unsigned int k = lhs_shape[0]; + GEMMReshapeInfo reshape_info(m, n, k, 1, 1, 0, false, broadcast_bias); + + // The output tensor will be auto-initialized within the function + + // Create and configure function + GEMMFunctionType gemm; + gemm.configure(gpu_arch, &lhs, &rhs, &bias, &dst, alpha, beta, false, reshape_info, fp16_mixed_precision, act_info); + + ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Allocate tensors + lhs.allocator()->allocate(); + rhs.allocator()->allocate(); + bias.allocator()->allocate(); + dst.allocator()->allocate(); + + ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Fill tensors + fill(AccessorType(lhs), 0); + fill(AccessorType(rhs), 1); + fill(AccessorType(bias), 2); + + // Compute GEMM + gemm.run(); + + return dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, DataType data_type, float alpha, float beta, bool broadcast_bias, + const ActivationLayerInfo &act_info) + { + TensorShape dst_shape = lhs_shape; + dst_shape[0] = rhs_shape[0]; + dst_shape[1] = lhs_shape[1]; + + // Create reference + SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; + SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; + SimpleTensor<T> bias{ dst_shape, data_type, 1 }; + + const int n = rhs_shape[0]; + const int m = lhs_shape[1]; + const int batch_size = lhs_shape[2]; + + // Fill reference + fill(lhs, 0); + fill(rhs, 1); + fill(bias, 2); + + if(broadcast_bias) + { + // In case of broadcast, we need simply copy the first into the following "M" ones + for(int i = 1; i < m * batch_size; i++) + { + memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); + } + } + + return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); + } + + TensorType _target{}; + SimpleTensor<T> _reference{}; +}; + +template <typename TensorType, typename AccessorType, typename T, typename GEMMFunctionType> +class GEMMMatrixMultiply3DValidationFixture : public framework::Fixture +{ +public: + template <typename...> + void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, float alpha, float beta, bool broadcast_bias, bool fp16_mixed_precision, + const ActivationLayerInfo &act_info, DataType data_type, GPUTarget gpu_arch) + { + // In case of GEMM3D, m is the product between m_w and m_h + const unsigned int m = m_w * m_h; + + // Set the tensor shapes for LHS and RHS matrices + const TensorShape lhs_shape(k, m, batch_size); + const TensorShape rhs_shape(n, k, batch_size); + const TensorShape bias_shape(n, 1, 1); + + _target = compute_target(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, m_h, fp16_mixed_precision, act_info, gpu_arch); + _reference = compute_reference(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, m_h, act_info); + } + +protected: + template <typename U> + void fill(U &&tensor, int i) + { + std::uniform_real_distribution<> distribution(-1.0f, 1.0f); + library->fill(tensor, distribution, i); + } + + TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, DataType data_type, float alpha, float beta, unsigned int m_h, + bool fp16_mixed_precision, const ActivationLayerInfo &act_info, GPUTarget gpu_arch) + { + // Create tensors + TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); + TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); + TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); + TensorType dst; + + const unsigned int m = lhs_shape[1]; + const unsigned int n = rhs_shape[0]; + const unsigned int k = lhs_shape[0]; + GEMMReshapeInfo reshape_info(m, n, k, 1, 1, m_h, false, true); + + // The output tensor will be auto-initialized within the function + + // Create and configure function + GEMMFunctionType gemm; + gemm.configure(gpu_arch, &lhs, &rhs, &bias, &dst, alpha, beta, false, reshape_info, fp16_mixed_precision, act_info); + + ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Allocate tensors + lhs.allocator()->allocate(); + rhs.allocator()->allocate(); + bias.allocator()->allocate(); + dst.allocator()->allocate(); + + ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Fill tensors + fill(AccessorType(lhs), 0); + fill(AccessorType(rhs), 1); + fill(AccessorType(bias), 2); + + // Compute GEMM + gemm.run(); + + return dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, DataType data_type, float alpha, float beta, unsigned int m_h, + const ActivationLayerInfo &act_info) + { + TensorShape dst_shape = lhs_shape; + dst_shape.set(0, rhs_shape[0]); + dst_shape.set(1, lhs_shape[1] / m_h); + dst_shape.set(2, m_h); + dst_shape.set(3, lhs_shape[2]); + + // Create reference + SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; + SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; + SimpleTensor<T> bias{ dst_shape, data_type, 1 }; + + const int n = rhs_shape[0]; + const int m = lhs_shape[1]; + const int batch_size = lhs_shape[2]; + + // Fill reference + fill(lhs, 0); + fill(rhs, 1); + fill(bias, 2); + + // In case of broadcast, we need simply copy the first into the following "M" ones + for(int i = 1; i < m * batch_size; i++) + { + memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); + } + + return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); + } + + TensorType _target{}; + SimpleTensor<T> _reference{}; +}; + +template <typename TensorType, typename AccessorType, typename T, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType> +class GEMMMatrixMultiplyInterleavedTransposedValidationFixture : public framework::Fixture +{ +public: + template <typename...> + void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, float alpha, float beta, unsigned int v0, unsigned int h0, bool broadcast_bias, bool fp16_mixed_precision, + const ActivationLayerInfo &act_info, DataType data_type, GPUTarget gpu_arch) + { + GEMMLHSMatrixInfo lhs_info; + lhs_info.m0 = 4; + lhs_info.k0 = 4; + lhs_info.v0 = v0; + lhs_info.interleave = true; + lhs_info.transpose = true; + + GEMMRHSMatrixInfo rhs_info; + rhs_info.n0 = 16 / sizeof(T); + rhs_info.k0 = 1; + rhs_info.h0 = h0; + rhs_info.interleave = false; + rhs_info.transpose = false; + + // Set the tensor shapes for LHS and RHS matrices + const TensorShape lhs_shape(k, m, batch_size); + const TensorShape rhs_shape(n, k, batch_size); + const TensorShape bias_shape(n, + broadcast_bias ? 1 : m, + broadcast_bias ? 1 : batch_size); + + _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, fp16_mixed_precision, act_info, gpu_arch); + _reference = compute_reference(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, broadcast_bias, act_info); + } + +protected: + template <typename U> + void fill(U &&tensor, int i) + { + std::uniform_real_distribution<> distribution(-1.0f, 1.0f); + library->fill(tensor, distribution, i); + + // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0) + std::uniform_real_distribution<> distribution_inf(std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity()); + library->fill_borders_with_garbage(tensor, distribution_inf, i); + } + + TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, + DataType data_type, float alpha, float beta, bool broadcast_bias, bool fp16_mixed_precision, const ActivationLayerInfo &act_info, GPUTarget gpu_arch) + { + // Create tensors + TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); + TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); + TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); + TensorType lhs_reshaped; + TensorType rhs_reshaped; + TensorType dst; + + const unsigned int m = lhs_shape[1]; + const unsigned int n = rhs_shape[0]; + const unsigned int k = lhs_shape[0]; + GEMMReshapeInfo reshape_info(m, n, k, rhs_info.h0, lhs_info.v0, 0, false, broadcast_bias); + + // The output tensor will be auto-initialized within the function + + // Create and configure function + ReshapeLHSFunctionType reshape_lhs; + ReshapeRHSFunctionType reshape_rhs; + GEMMFunctionType gemm; + reshape_lhs.configure(&lhs, &lhs_reshaped, lhs_info); + reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info); + gemm.configure(gpu_arch, &lhs_reshaped, &rhs_reshaped, &bias, &dst, alpha, beta, true, reshape_info, fp16_mixed_precision, act_info); + + ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Allocate tensors + lhs.allocator()->allocate(); + rhs.allocator()->allocate(); + lhs_reshaped.allocator()->allocate(); + rhs_reshaped.allocator()->allocate(); + bias.allocator()->allocate(); + dst.allocator()->allocate(); + + ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!lhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Fill tensors + fill(AccessorType(lhs), 0); + fill(AccessorType(rhs), 1); + fill(AccessorType(bias), 2); + + // Compute GEMM + reshape_lhs.run(); + reshape_rhs.run(); + gemm.run(); + + return dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, DataType data_type, float alpha, float beta, bool broadcast_bias, + const ActivationLayerInfo &act_info) + { + TensorShape dst_shape = lhs_shape; + dst_shape[0] = rhs_shape[0]; + dst_shape[1] = lhs_shape[1]; + + // Create reference + SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; + SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; + SimpleTensor<T> bias{ dst_shape, data_type, 1 }; + + const int n = rhs_shape[0]; + const int m = lhs_shape[1]; + const int batch_size = lhs_shape[2]; + + // Fill reference + fill(lhs, 0); + fill(rhs, 1); + fill(bias, 2); + + if(broadcast_bias) + { + // In case of broadcast, we need simply copy the first into the following "M" ones + for(int i = 1; i < m * batch_size; i++) + { + memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); + } + } + + return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); + } + + TensorType _target{}; + SimpleTensor<T> _reference{}; +}; + +template <typename TensorType, typename AccessorType, typename T, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType> +class GEMMMatrixMultiplyInterleavedTransposed3DValidationFixture : public framework::Fixture +{ +public: + template <typename...> + void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, float alpha, float beta, unsigned int v0, unsigned int h0, bool broadcast_bias, + bool fp16_mixed_precision, const ActivationLayerInfo &act_info, DataType data_type, GPUTarget gpu_arch) + { + GEMMLHSMatrixInfo lhs_info; + lhs_info.m0 = 4; + lhs_info.k0 = 4; + lhs_info.v0 = v0; + lhs_info.interleave = true; + lhs_info.transpose = true; + + GEMMRHSMatrixInfo rhs_info; + rhs_info.n0 = 16 / sizeof(T); + rhs_info.k0 = 1; + rhs_info.h0 = h0; + rhs_info.interleave = false; + rhs_info.transpose = false; + + // In case of GEMM3D, m is the product between m_w and m_h + const unsigned int m = m_w * m_h; + + // Set the tensor shapes for LHS and RHS matrices + const TensorShape lhs_shape(k, m, batch_size); + const TensorShape rhs_shape(n, k, batch_size); + const TensorShape bias_shape(n, 1, 1); + + _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, m_h, fp16_mixed_precision, act_info, gpu_arch); + _reference = compute_reference(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, m_h, act_info); + } + +protected: + template <typename U> + void fill(U &&tensor, int i) + { + std::uniform_real_distribution<> distribution(-1.0f, 1.0f); + library->fill(tensor, distribution, i); + } + + TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, + DataType data_type, float alpha, float beta, unsigned int m_h, bool fp16_mixed_precision, const ActivationLayerInfo &act_info, GPUTarget gpu_arch) + { + // Create tensors + TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); + TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); + TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); + TensorType lhs_reshaped; + TensorType rhs_reshaped; + TensorType dst; + + const unsigned int m = lhs_shape[1]; + const unsigned int n = rhs_shape[0]; + const unsigned int k = lhs_shape[0]; + GEMMReshapeInfo reshape_info(m, n, k, rhs_info.h0, lhs_info.v0, m_h, false, true); + + // The output tensor will be auto-initialized within the function + + // Create and configure function + ReshapeLHSFunctionType reshape_lhs; + ReshapeRHSFunctionType reshape_rhs; + GEMMFunctionType gemm; + reshape_lhs.configure(&lhs, &lhs_reshaped, lhs_info); + reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info); + gemm.configure(gpu_arch, &lhs_reshaped, &rhs_reshaped, &bias, &dst, alpha, beta, true, reshape_info, fp16_mixed_precision, act_info); + + ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Allocate tensors + lhs.allocator()->allocate(); + rhs.allocator()->allocate(); + lhs_reshaped.allocator()->allocate(); + rhs_reshaped.allocator()->allocate(); + bias.allocator()->allocate(); + dst.allocator()->allocate(); + + ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!lhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Fill tensors + fill(AccessorType(lhs), 0); + fill(AccessorType(rhs), 1); + fill(AccessorType(bias), 2); + + // Compute GEMM + reshape_lhs.run(); + reshape_rhs.run(); + gemm.run(); + + return dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, DataType data_type, float alpha, float beta, unsigned int m_h, + const ActivationLayerInfo &act_info) + { + TensorShape dst_shape = lhs_shape; + dst_shape.set(0, rhs_shape[0]); + dst_shape.set(1, lhs_shape[1] / m_h); + dst_shape.set(2, m_h); + dst_shape.set(3, lhs_shape[2]); + + // Create reference + SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; + SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; + SimpleTensor<T> bias{ dst_shape, data_type, 1 }; + + const int n = rhs_shape[0]; + const int m = lhs_shape[1]; + const int batch_size = lhs_shape[2]; + + // Fill reference + fill(lhs, 0); + fill(rhs, 1); + fill(bias, 2); + + // In case of broadcast, we need simply copy the first into the following "M" ones + for(int i = 1; i < m * batch_size; i++) + { + memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); + } + + return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); + } + + TensorType _target{}; + SimpleTensor<T> _reference{}; +}; + template <typename TensorType, typename AccessorType, typename T, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType> class GEMMMatrixMultiplyReshapedValidationFixture : public framework::Fixture { |