From 856f66e6c61b77d03f754cd0fa8439891f0e4aca Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Thu, 22 Apr 2021 21:13:21 +0100 Subject: Port CLGEMM to memory injecting interface Moves the following kernels: - CLGEMMMatrixMultiplyKernel - CLGEMMMatrixMultiplyNativeKernel - CLGEMMMatrixMultipluReshapedKernel - CLGEMMMatrixMultiplyReshapedOnlyRHSKernel Moves the following functions - CLGEMM Introduces facilities to easy handling of auxiliary temporary buffers under then new run interface. Such are: - CLAuxTensorHandler: That allows wrapping of workspace buffers memory to CLBuffer objects - Ability to inject TensorInfo to allocator without transferring ownership. This reduce the copy overhead if needed. Resolves: COMPMID-4188 Signed-off-by: Georgios Pinitas Change-Id: I7055435d831b05b749b26302082e4ac45f26dfb0 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5498 Tested-by: Arm Jenkins Reviewed-by: Michalis Spyrou Comments-Addressed: Arm Jenkins --- arm_compute/runtime/CL/functions/CLGEMM.h | 145 ++++-------------------------- 1 file changed, 18 insertions(+), 127 deletions(-) (limited to 'arm_compute/runtime/CL/functions/CLGEMM.h') diff --git a/arm_compute/runtime/CL/functions/CLGEMM.h b/arm_compute/runtime/CL/functions/CLGEMM.h index 1e2ae7be64..38a07ef9fb 100644 --- a/arm_compute/runtime/CL/functions/CLGEMM.h +++ b/arm_compute/runtime/CL/functions/CLGEMM.h @@ -35,76 +35,12 @@ namespace arm_compute { +// Forward declarations class CLCompileContext; -class CLGEMMReshapeRHSMatrixKernel; -class CLGEMMMatrixMultiplyKernel; -class CLGEMMMatrixMultiplyReshapedKernel; -class CLGEMMMatrixMultiplyReshapedOnlyRHSKernel; -class CLGEMMReshapeLHSMatrixKernel; class ICLTensor; class ITensorInfo; -namespace weights_transformations -{ -/** Basic function to manage the reshape weights generated from @ref CLGEMMReshapeRHSMatrixKernel */ -class CLGEMMReshapeRHSMatrixKernelManaged : public ITransformWeights -{ -public: - /** Default constructor */ - CLGEMMReshapeRHSMatrixKernelManaged(); - /** Prevent instances of this class from being copied (As this class contains pointers) */ - CLGEMMReshapeRHSMatrixKernelManaged(const CLGEMMReshapeRHSMatrixKernelManaged &) = delete; - /** Default move constructor */ - CLGEMMReshapeRHSMatrixKernelManaged(CLGEMMReshapeRHSMatrixKernelManaged &&) = default; - /** Prevent instances of this class from being copied (As this class contains pointers) */ - CLGEMMReshapeRHSMatrixKernelManaged &operator=(const CLGEMMReshapeRHSMatrixKernelManaged &) = delete; - /** Default move assignment operator */ - CLGEMMReshapeRHSMatrixKernelManaged &operator=(CLGEMMReshapeRHSMatrixKernelManaged &&) = default; - /** Default desctructor */ - ~CLGEMMReshapeRHSMatrixKernelManaged(); - //Inherited method override - void run() override; - - //Inherited method override - void release() override; - - //Inherited method override - ICLTensor *get_weights() override; - - //Inherited method override - uint32_t uid() override; - - /** Configures the @ref CLGEMMReshapeRHSMatrixKernel kernel - * - * @param[in] input Input tensor. Data types supported: All - * @param[in] info RHS matrix information to be used for reshaping. - */ - void configure(const ICLTensor *input, GEMMRHSMatrixInfo info); - - /** Configures the @ref CLGEMMReshapeRHSMatrixKernel kernel - * - * @param[in] compile_context The compile context to be used. - * @param[in] input Input tensor. Data types supported: All - * @param[in] info RHS matrix information to be used for reshaping. - */ - void configure(const CLCompileContext &compile_context, const ICLTensor *input, GEMMRHSMatrixInfo info); - -private: - static constexpr uint32_t _uid{ 0x15 }; - CLTensor _output{}; - std::unique_ptr _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: @@ -114,16 +50,16 @@ public: * @param[in] weights_manager (Optional) Weights manager. */ CLGEMM(std::shared_ptr 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; - /** Default destructor */ - ~CLGEMM(); + CLGEMM &operator=(CLGEMM &&); /** Initialise the kernel's inputs and output * * Valid data layouts: @@ -134,25 +70,6 @@ public: * |:------------|:-----------|:---------|:--------------| * |F32 |F32 |F32 |F32 | * |F16 |F16 |F16 |F16 | - * - * @note GEMM: General Matrix Multiply - [alpha * A * B + beta * C]. - * - * @note All tensors must have the same data type. - * - * @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 * * @note GEMM: General Matrix Multiply - [alpha * A * B + beta * C]. * @@ -168,20 +85,20 @@ 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 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 */ @@ -192,34 +109,8 @@ public: void prepare() override; private: - 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; - std::unique_ptr _mm_kernel; - std::unique_ptr _reshape_lhs_kernel; - std::unique_ptr _reshape_rhs_kernel; - std::unique_ptr _reshape_rhs_kernel_managed; - std::unique_ptr _mm_reshaped_kernel; - std::unique_ptr _mm_reshaped_only_rhs_kernel; - std::unique_ptr _mm_reshaped_only_rhs_fallback_kernel; - CLTensor _tmp_a; - CLTensor _tmp_b; - const ICLTensor *_original_b; - const ICLTensor *_lhs; - ICLTensor *_dst; - bool _reshape_b_only_on_first_run; - bool _is_prepared; - CLGEMMKernelType _gemm_kernel_type; + struct Impl; + std::unique_ptr _impl; }; } // namespace arm_compute -- cgit v1.2.1