diff options
author | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2020-07-20 13:31:05 +0100 |
---|---|---|
committer | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2020-07-21 13:12:22 +0000 |
commit | c6eaec3610fa27651582f6c1acad35afffe360f6 (patch) | |
tree | 8664d8e54be6ad1c14cbde61f6b5ac3478d266ab /src/core | |
parent | a449a3638aaaa32414384219a0041b86591a7f41 (diff) | |
download | ComputeLibrary-c6eaec3610fa27651582f6c1acad35afffe360f6.tar.gz |
COMPMID-3326; Update heuristic for GEMMReshaped and GEMMReshapedOnlyRHS
- Update the heuristic for Arm Mali-G77 (F32) in order to use the OpenCL image2d object on GEMM
Change-Id: Ife6736a22ec2a114368bb338908f0c5f239dfad6
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3593
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Diffstat (limited to 'src/core')
-rw-r--r-- | src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationValhall.cpp | 59 |
1 files changed, 49 insertions, 10 deletions
diff --git a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationValhall.cpp b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationValhall.cpp index 11cb90ba19..9f3461e912 100644 --- a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationValhall.cpp +++ b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationValhall.cpp @@ -27,6 +27,9 @@ #include "arm_compute/core/CL/CLKernelLibrary.h" #include "arm_compute/core/CL/gemm/CLGEMMHelpers.h" #include "arm_compute/core/GPUTarget.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" #include <map> #include <utility> @@ -35,6 +38,8 @@ namespace arm_compute { namespace cl_gemm { +using namespace arm_compute::misc::shape_calculator; + CLGEMMReshapedOnlyRHSKernelConfigurationValhall::CLGEMMReshapedOnlyRHSKernelConfigurationValhall(GPUTarget gpu) : ICLGEMMKernelConfiguration(gpu) { @@ -74,32 +79,66 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedOnlyRHSKernelConfi std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedOnlyRHSKernelConfigurationValhall::configure_G77_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b) { ARM_COMPUTE_UNUSED(k); - ARM_COMPUTE_UNUSED(b); + GEMMLHSMatrixInfo lhs_info_buf; + GEMMRHSMatrixInfo rhs_info_buf; + GEMMLHSMatrixInfo lhs_info_img; + GEMMRHSMatrixInfo rhs_info_img; + + // Get lhs_info/rhs_info in case of OpenCL buffer if(m == 1) { - if(n > 2048) + const unsigned int h0 = std::max(n / 4, 1U); + std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 1, 4, 4, 1, h0, false, true, false, true); + } + else + { + if(m > 256) { - return configure_lhs_rhs_info(m, n, 1, 8, 2, 1, 256, false, true, false, true); + const int v0 = std::max(std::min(static_cast<int>(n / 4), static_cast<int>(8)), static_cast<int>(1)); + std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 4, 4, 4, 1, v0, false, true, false, true); } else { - return configure_lhs_rhs_info(m, n, 1, 2, 2, 1, 256, false, true, false, true); + const int v0 = std::max(std::min(static_cast<int>(n / 4), static_cast<int>(8)), static_cast<int>(1)); + std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 2, 4, 4, 1, v0, false, true, false, true); } } + + // Get lhs_info/rhs_info in case of OpenCL image + if(m == 1) + { + std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 1, 4, 4, 1, 8, true, true, false, false, true); + } else { - if(m > 300) + if((m / 4) * (n / 4) > 4096) { - const int v0 = std::max(std::min(static_cast<int>(n / 4), static_cast<int>(256)), static_cast<int>(1)); - return configure_lhs_rhs_info(m, n, 4, 4, 4, 1, v0, false, true, false, true); + const int h0 = std::max(std::min(static_cast<int>(n / 4), static_cast<int>(8)), static_cast<int>(1)); + std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 4, 4, 4, 1, h0, false, true, false, false, true); } else { - const int v0 = std::max(std::min(static_cast<int>(n / 4), static_cast<int>(256)), static_cast<int>(1)); - return configure_lhs_rhs_info(m, n, 2, 4, 4, 1, v0, false, true, false, true); + const int h0 = std::max(std::min(static_cast<int>(n / 4), static_cast<int>(8)), static_cast<int>(1)); + std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 2, 4, 4, 1, h0, false, true, false, false, true); } } + + const TensorInfo tensor_rhs_info(TensorShape(n, k, b), 1, DataType::F32); + const TensorShape shape = compute_rhs_reshaped_shape(tensor_rhs_info, rhs_info_img); + const TensorInfo tensor_reshaped_info(shape, 1, DataType::F32); + + // In case of small workloads, we use the OpenCL buffer rather than the OpenCL image2d + const bool use_cl_image2d = ((m / lhs_info_img.m0) * (n / rhs_info_img.n0)) * b < 1024 ? false : true; + + if(bool(validate_image2d_support_on_rhs(tensor_reshaped_info, rhs_info_img)) && use_cl_image2d) + { + return std::make_pair(lhs_info_img, rhs_info_img); + } + else + { + return std::make_pair(lhs_info_buf, rhs_info_buf); + } } std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedOnlyRHSKernelConfigurationValhall::configure_G77_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b) @@ -120,7 +159,7 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedOnlyRHSKernelConfi return configure_lhs_rhs_info(m, n, 1, 2, 8, 1, h0, false, true, false, true); } } - else if (m < 128) + else if(m < 128) { const int h0 = std::max(std::min(static_cast<int>(n / 4), static_cast<int>(256)), static_cast<int>(1)); if(k >= 512) |