COMPMID-3326: Update heuristic for GEMMReshaped and GEMMReshapedOnlyRHS
- Update the heuristic for Arm Mali-G76 (F32) in order to use the OpenCL image2d object
on GEMM
- Create utility function to validate the support for image2d
Change-Id: I0913ac5f27fd07992b0ac188af753a2abeb034ca
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3559
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/CL/gemm/CLGEMMHelpers.cpp b/src/core/CL/gemm/CLGEMMHelpers.cpp
index 1165e70..5734c93 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 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 @@
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