diff options
Diffstat (limited to 'arm_compute/runtime/CL/functions/CLGEMM.h')
| -rw-r--r-- | arm_compute/runtime/CL/functions/CLGEMM.h | 188 |
1 files changed, 55 insertions, 133 deletions
diff --git a/arm_compute/runtime/CL/functions/CLGEMM.h b/arm_compute/runtime/CL/functions/CLGEMM.h index f5588112ae..f5e6aa1237 100644 --- a/arm_compute/runtime/CL/functions/CLGEMM.h +++ b/arm_compute/runtime/CL/functions/CLGEMM.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2020 ARM Limited. + * Copyright (c) 2016-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -24,11 +24,7 @@ #ifndef ARM_COMPUTE_CLGEMM_H #define ARM_COMPUTE_CLGEMM_H -#include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h" -#include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h" -#include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.h" -#include "arm_compute/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.h" -#include "arm_compute/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h" +#include "arm_compute/function_info/GEMMInfo.h" #include "arm_compute/runtime/CL/CLTensor.h" #include "arm_compute/runtime/CL/CLTypes.h" #include "arm_compute/runtime/IFunction.h" @@ -36,79 +32,16 @@ #include "arm_compute/runtime/IWeightsManager.h" #include "arm_compute/runtime/MemoryGroup.h" +#include <memory> + namespace arm_compute { +// Forward declarations +class CLCompileContext; class ICLTensor; +class ITensorInfo; -namespace weights_transformations -{ -/** Basic function to manage the reshape weights generated from @ref CLGEMMReshapeRHSMatrixKernel */ -class CLGEMMReshapeRHSMatrixKernelManaged : public ITransformWeights -{ -public: - //Inherited method override - void run() override - { - _output.allocator()->allocate(); - CLScheduler::get().enqueue(_kernel, false); - _reshape_run = true; - } - - //Inherited method override - void release() override - { - _output.allocator()->free(); - } - - //Inherited method override - ICLTensor *get_weights() override - { - return &_output; - } - - //Inherited method override - uint32_t uid() override - { - return _uid; - } - - /** Configures the @ref CLGEMMReshapeRHSMatrixKernel kernel - * - * @param[in] input Input tensor. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32 - * @param[in] info RHS matrix information to be used for reshaping. - */ - void configure(const ICLTensor *input, GEMMRHSMatrixInfo info) - { - configure(CLKernelLibrary::get().get_compile_context(), input, info); - } - - /** Configures the @ref CLGEMMReshapeRHSMatrixKernel kernel - * - * @param[in] compile_context The compile context to be used. - * @param[in] input Input tensor. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32 - * @param[in] info RHS matrix information to be used for reshaping. - */ - void configure(const CLCompileContext &compile_context, const ICLTensor *input, GEMMRHSMatrixInfo info) - { - _kernel.configure(compile_context, input, &_output, info); - } - -private: - static constexpr uint32_t _uid = 0x15; - CLTensor _output{}; - CLGEMMReshapeRHSMatrixKernel _kernel{}; -}; -} // namespace weights_transformations - -/** Basic function to execute GEMM on OpenCL. This function calls the following OpenCL kernels: - * - * -# @ref CLGEMMReshapeLHSMatrixKernel (only if the RESHAPED_V1 is selected by the heuristic model) - * -# @ref CLGEMMReshapeRHSMatrixKernel (only if either the RESHAPED_V1 or RESHAPED_ONLY_RHS is selected by the select_gemm_kernel method()) - * -# @ref CLGEMMMatrixMultiplyKernel (only if either the NATIVE or RESHAPED_V1 is selected by the select_gemm_kernel method()) - * -# @ref CLGEMMMatrixMultiplyReshapedKernel (only if RESHAPED_V1 is selected by the select_gemm_kernel method()) - * -# @ref CLGEMMMatrixMultiplyReshapedOnlyRHSKernel (only if RESHAPED_ONLY_RHS is selected by the select_gemm_kernel method()) - * - */ +/** Basic function to execute GEMM on OpenCL */ class CLGEMM : public IFunction { public: @@ -118,34 +51,26 @@ public: * @param[in] weights_manager (Optional) Weights manager. */ CLGEMM(std::shared_ptr<IMemoryManager> memory_manager = nullptr, IWeightsManager *weights_manager = nullptr); + /** Default destructor */ + ~CLGEMM(); /** Prevent instances of this class from being copied (As this class contains pointers) */ CLGEMM(const CLGEMM &) = delete; /** Default move constructor */ - CLGEMM(CLGEMM &&) = default; + CLGEMM(CLGEMM &&); /** Prevent instances of this class from being copied (As this class contains pointers) */ CLGEMM &operator=(const CLGEMM &) = delete; /** Default move assignment operator */ - CLGEMM &operator=(CLGEMM &&) = default; + CLGEMM &operator=(CLGEMM &&); /** Initialise the kernel's inputs and output * - * @note GEMM: General Matrix Multiply - [alpha * A * B + beta * C]. - * - * @note All tensors must have the same data type. + * Valid data layouts: + * - All * - * @note Whilst the first input tensor can be a vector, the second input tensor must be at least a matrix - * - * @param[in] a First input tensor (Matrix or Vector A). Data types supported: F16/F32 - * @param[in] b Second input tensor (Matrix B). Data type supported: same as @p a. - * @param[in] c Third input tensor (Matrix C). It can be a nullptr if just the multiplication between @p a and @p b is needed. Data type supported: same as @p a. - * @param[out] output Output tensor. Data type supported: same as @p a - * @param[in] alpha Weight of the matrix product - * @param[in] beta Weight of matrix C - * @param[in] gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and - * if the reshape of matrix B should happen only for the first run. GEMMInfo also contains information about the reshaping - * in case matrix A and matrix B have been already transformed. - */ - void configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta, const GEMMInfo &gemm_info = GEMMInfo()); - /** Initialise the kernel's inputs and output + * Valid data type configurations: + * |src0 |src1 |src2 |dst | + * |:------------|:-----------|:---------|:--------------| + * |F32 |F32 |F32 |F32 | + * |F16 |F16 |F16 |F16 | * * @note GEMM: General Matrix Multiply - [alpha * A * B + beta * C]. * @@ -153,6 +78,9 @@ public: * * @note Whilst the first input tensor can be a vector, the second input tensor must be at least a matrix * + * @note Batched GEMM only allows RHS tensor's rank to be <= 3 + * @note Batched GEMM only supports broadcasting cases where RHS rank < LHS rank but not the other way around + * * @param[in] compile_context The compile context to be used. * @param[in] a First input tensor (Matrix or Vector A). Data types supported: F16/F32 * @param[in] b Second input tensor (Matrix B). Data type supported: same as @p a. @@ -161,57 +89,51 @@ public: * @param[in] alpha Weight of the matrix product * @param[in] beta Weight of matrix C * @param[in] gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and - * if the reshape of matrix B should happen only for the first run. GEMMInfo also contains information about the reshaping - * in case matrix A and matrix B have been already transformed. + * if the reshape of matrix B should happen only for the first run. GEMMInfo also contains information about the reshaping + * in case matrix A and matrix B have been already transformed. + */ + void configure(const CLCompileContext &compile_context, + const ICLTensor *a, + const ICLTensor *b, + const ICLTensor *c, + ICLTensor *output, + float alpha, + float beta, + const GEMMInfo &gemm_info = GEMMInfo()); + + /** Initialise the kernel's inputs and output + * + * Similar to @ref CLGEMM::configure() */ - void configure(const CLCompileContext &compile_context, const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta, const GEMMInfo &gemm_info = GEMMInfo()); + void configure(const ICLTensor *a, + const ICLTensor *b, + const ICLTensor *c, + ICLTensor *output, + float alpha, + float beta, + const GEMMInfo &gemm_info = GEMMInfo()); + /** Static function to check if given info will lead to a valid configuration of @ref CLGEMM. * - * @param[in] a First input tensor info (Matrix or Vector A). Data types supported: F16/F32 - * @param[in] b Second input tensor info (Matrix B). Data type supported: same as @p a. - * @param[in] c Third input tensor info (Matrix C). It can be a nullptr if just the multiplication between @p a and @p b is needed. Data type supported: same as @p a. - * @param[in] output Output tensor info. Data type supported: same as @p a - * @param[in] alpha Weight of the matrix product - * @param[in] beta Weight of matrix C - * @param[in] gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and - * if the reshape of matrix B should happen only for the first run + * Similar to @ref CLGEMM::configure() * * @return a status */ - static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info = GEMMInfo()); + static Status validate(const ITensorInfo *a, + const ITensorInfo *b, + const ITensorInfo *c, + const ITensorInfo *output, + float alpha, + float beta, + const GEMMInfo &gemm_info = GEMMInfo()); // Inherited methods overridden: void run() override; void prepare() override; private: - static CLGEMMKernelType select_gemm_kernel(unsigned int m, unsigned int n, unsigned int k, DataType data_type, bool reshape_b_only_on_first_run); - - void configure_native_v1(const CLCompileContext &compile_context, const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta, const GEMMInfo &gemm_info); - void configure_reshaped_v1(const CLCompileContext &compile_context, const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta, const GEMMInfo &gemm_info); - void configure_reshaped_v2(const CLCompileContext &compile_context, const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta, const GEMMInfo &gemm_info); - void configure_reshaped_only_rhs(const CLCompileContext &compile_context, const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta, - const GEMMInfo &gemm_info); - - static Status validate_native_v1(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info); - static Status validate_reshaped_v1(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info); - static Status validate_reshaped(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info); - static Status validate_reshaped_only_rhs(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info); - - MemoryGroup _memory_group; - IWeightsManager *_weights_manager; - CLGEMMMatrixMultiplyKernel _mm_kernel; - CLGEMMReshapeLHSMatrixKernel _reshape_lhs_kernel; - CLGEMMReshapeRHSMatrixKernel _reshape_rhs_kernel; - weights_transformations::CLGEMMReshapeRHSMatrixKernelManaged _reshape_rhs_kernel_managed; - CLGEMMMatrixMultiplyReshapedKernel _mm_reshaped_kernel; - CLGEMMMatrixMultiplyReshapedOnlyRHSKernel _mm_reshaped_only_rhs_kernel; - CLTensor _tmp_a; - CLTensor _tmp_b; - const ICLTensor *_original_b; - bool _reshape_b_only_on_first_run; - bool _is_prepared; - CLGEMMKernelType _gemm_kernel_type; + struct Impl; + std::unique_ptr<Impl> _impl; }; } // namespace arm_compute |