From 926afe1c8ad6ba6a7bada62a4027fcb79d727104 Mon Sep 17 00:00:00 2001 From: Gian Marco Iodice Date: Tue, 19 Mar 2019 11:44:13 +0000 Subject: COMPMID-2097: Implement a heuristic to dispatch CLGEMMReshapedOnlyRHS kernel from CLGEMM Change-Id: I4170a80647b02501aa669e2c0347ddc39888ee76 Signed-off-by: Gian Marco Iodice Reviewed-on: https://review.mlplatform.org/c/928 Reviewed-by: Giuseppe Rossini Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins --- arm_compute/core/utils/misc/ShapeCalculator.h | 15 ++++++--------- 1 file changed, 6 insertions(+), 9 deletions(-) (limited to 'arm_compute/core/utils') diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h index 0d07266403..384bd460a0 100644 --- a/arm_compute/core/utils/misc/ShapeCalculator.h +++ b/arm_compute/core/utils/misc/ShapeCalculator.h @@ -887,23 +887,20 @@ inline TensorShape compute_mm_shape(const ITensorInfo &input0, const ITensorInfo { ARM_COMPUTE_ERROR_ON_MSG(input0.num_dimensions() > 4, "The number of dimensions for the matrix A must be <= 4"); + const bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); const bool reinterpret_output_as_3d = gemm_info.depth_output_gemm3d() != 0; const int depth_output_gemm3d = reinterpret_output_as_3d ? gemm_info.depth_output_gemm3d() : 1; // If the output of GEMM has to be reinterpreted as 3D, the number of input0 rows (M) is obtained collapsing the second and third // dimension of the output tensor - const int dim0 = gemm_info.n(); - const int dim1 = gemm_info.m() / depth_output_gemm3d; - const int dim2 = input0.tensor_shape()[2]; - const int dim3 = input0.tensor_shape()[3]; + const int batch_size = reinterpret_input_as_3d ? input0.tensor_shape()[3] : input0.tensor_shape()[2]; TensorShape output_shape{ input0.tensor_shape() }; - output_shape.set(0, dim0); - output_shape.set(1, dim1); - output_shape.set(2, reinterpret_output_as_3d ? depth_output_gemm3d : dim2); - output_shape.set(3, reinterpret_output_as_3d ? dim2 : dim3); - output_shape.set(4, reinterpret_output_as_3d ? dim3 : 1); + output_shape.set(0, gemm_info.n()); + output_shape.set(1, gemm_info.m() / depth_output_gemm3d); + output_shape.set(2, reinterpret_output_as_3d ? depth_output_gemm3d : batch_size); + output_shape.set(3, reinterpret_output_as_3d ? batch_size : 1); return output_shape; } -- cgit v1.2.1