aboutsummaryrefslogtreecommitdiff
path: root/src/core/CL/gemm
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/CL/gemm')
-rw-r--r--src/core/CL/gemm/CLGEMMHelpers.cpp48
-rw-r--r--src/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationBifrost.cpp44
-rw-r--r--src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp42
3 files changed, 113 insertions, 21 deletions
diff --git a/src/core/CL/gemm/CLGEMMHelpers.cpp b/src/core/CL/gemm/CLGEMMHelpers.cpp
index 1165e70273..5734c93021 100644
--- a/src/core/CL/gemm/CLGEMMHelpers.cpp
+++ b/src/core/CL/gemm/CLGEMMHelpers.cpp
@@ -23,7 +23,10 @@
*/
#include "arm_compute/core/CL/gemm/CLGEMMHelpers.h"
+#include "arm_compute/core/CL/CLHelpers.h"
#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/OpenCL.h"
+#include "arm_compute/core/ITensorInfo.h"
#include <utility>
@@ -32,24 +35,13 @@ namespace arm_compute
namespace cl_gemm
{
std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_lhs_rhs_info(unsigned int m, unsigned int n, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0,
- bool lhs_interleave, bool rhs_interleave, bool lhs_transpose, bool rhs_transpose)
+ bool lhs_interleave, bool rhs_interleave, bool lhs_transpose, bool rhs_transpose, bool export_to_cl_image)
{
- GEMMLHSMatrixInfo lhs_info;
- GEMMRHSMatrixInfo rhs_info;
+ v0 = ((m / (m0 * v0)) == 0) ? 1 : v0;
+ h0 = ((n / (n0 * h0)) == 0) ? 1 : h0;
- // Configure GEMMLHSMatrixInfo
- lhs_info.m0 = m0;
- lhs_info.k0 = k0;
- lhs_info.v0 = ((m / (lhs_info.m0 * v0)) == 0) ? 1 : v0;
- lhs_info.interleave = lhs_interleave;
- lhs_info.transpose = lhs_transpose;
-
- // Configure GEMMRHSMatrixInfo
- rhs_info.n0 = n0;
- rhs_info.k0 = lhs_info.k0;
- rhs_info.h0 = ((n / (rhs_info.n0 * h0)) == 0) ? 1 : h0;
- rhs_info.interleave = rhs_interleave;
- rhs_info.transpose = rhs_transpose;
+ const GEMMLHSMatrixInfo lhs_info(m0, k0, v0, lhs_transpose, lhs_interleave);
+ const GEMMRHSMatrixInfo rhs_info(n0, k0, h0, rhs_transpose, rhs_interleave, export_to_cl_image);
return std::make_pair(lhs_info, rhs_info);
}
@@ -66,5 +58,27 @@ void update_padding_for_cl_image(ITensorInfo *tensor)
tensor->extend_padding(PaddingSize(0, padding, 0, 0));
}
+
+Status validate_image2d_support_on_rhs(const ITensorInfo &tensor_reshaped_info, const GEMMRHSMatrixInfo &rhs_info)
+{
+ if(rhs_info.export_to_cl_image)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG((rhs_info.n0 == 2) || (rhs_info.n0 == 3), "Export to cl_image only supported with n0 = 4, 8 or 16");
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG((rhs_info.k0 == 2) || (rhs_info.k0 == 3), "Export to cl_image only supported with k0 = 4, 8 or 16");
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(tensor_reshaped_info.data_type() != DataType::F32, "Export to cl_image only supported with F32 data type");
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(!image2d_from_buffer_supported(CLKernelLibrary::get().get_device()), "The extension cl_khr_image2d_from_buffer is not supported on the target platform");
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(get_cl_image_pitch_alignment(CLKernelLibrary::get().get_device()) == 0, "Impossible to retrieve the cl_image pitch alignment");
+
+ // Check the width and height of the output tensor.
+ // Since we cannot create a 3d image from a buffer, the third dimension is collapsed on the second dimension
+ const size_t max_image_w = CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_IMAGE2D_MAX_WIDTH>();
+ const size_t max_image_h = CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_IMAGE2D_MAX_HEIGHT>();
+
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(tensor_reshaped_info.tensor_shape()[0] > max_image_w * 4, "Not supported width for cl_image");
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(tensor_reshaped_info.tensor_shape()[1] * tensor_reshaped_info.tensor_shape()[2] > max_image_h, "Not supported height for cl_image");
+ }
+
+ return Status{};
+}
} // namespace cl_gemm
-} // namespace arm_compute \ No newline at end of file
+} // namespace arm_compute
diff --git a/src/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationBifrost.cpp b/src/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationBifrost.cpp
index f2954be7d2..a533f14d02 100644
--- a/src/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationBifrost.cpp
+++ b/src/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationBifrost.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;
+
CLGEMMReshapedKernelConfigurationBifrost::CLGEMMReshapedKernelConfigurationBifrost(GPUTarget gpu)
: ICLGEMMKernelConfiguration(gpu)
{
@@ -153,13 +158,48 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedKernelConfiguratio
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(n <= 4)
{
- return configure_lhs_rhs_info(m, n, 4, 2, 8, 16, 16, true, false, false, true);
+ std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 4, 2, 8, 16, 16, true, false, false, true);
+ }
+ else
+ {
+ std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 4, 4, 2, 8, 16, false, false, false, true);
+ }
+
+ // Get lhs_info/rhs_info in case of OpenCL image
+ // Condition on the GPU workload
+ if((m / 4) * (n / 4) >= 2560)
+ {
+ // Big workload
+ std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 8, true, true, true, false, true);
+ }
+ else
+ {
+ // Small workload
+ std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 2, 4, 4, 1, 1, true, true, true, 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 vector by matrix with few work-items, we use the OpenCL buffer rather than the OpenCL image2d
+ const bool use_cl_image2d = (n <= 4) ? 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 configure_lhs_rhs_info(m, n, 4, 4, 2, 8, 16, false, false, false, true);
+ return std::make_pair(lhs_info_buf, rhs_info_buf);
}
}
diff --git a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp
index f662089c77..581c2d2199 100644
--- a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp
+++ b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.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;
+
CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::CLGEMMReshapedOnlyRHSKernelConfigurationBifrost(GPUTarget gpu)
: ICLGEMMKernelConfiguration(gpu)
{
@@ -139,14 +144,47 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedOnlyRHSKernelConfi
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)
{
const unsigned int h0 = std::max(n / 2, 1U);
- return configure_lhs_rhs_info(m, n, 1, 2, 8, 1, h0, false, true, false, true);
+ std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 1, 2, 8, 1, h0, false, true, false, true);
}
else
{
- return configure_lhs_rhs_info(m, n, 4, 4, 4, 1, 2, false, true, false, true);
+ std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 4, 4, 4, 1, 2, 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, 4, false, true, false, false, true);
+ }
+ else
+ {
+ const int h0 = std::max(std::min(static_cast<int>(n / 4), static_cast<int>(16)), 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);
+ }
+
+ 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 vector by matrix with few work-items, we use the OpenCL buffer rather than the OpenCL image2d
+ const bool use_cl_image2d = (m == 1 && n <= 4096) ? 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);
}
}
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-18 12:08:48 +0000