commit 1c9efebf4344e8db97e6d9282b2bf48b52090b58
author giuros01 <giuseppe.rossini@arm.com> Fri Jan 11 14:04:43 2019 +0000
committer Giuseppe Rossini <giuseppe.rossini@arm.com> Mon Jan 14 12:42:42 2019 +0000
tree 40c8737a23ff289cb176618d7c16fd4b2d811689
parent 685c181d5ac36948f8df22f627ef9b3a74b5f87e

Issue COMPMID-1835: Remove CLGEMMInterleave4x4Kernel and replace with CLGEMMReshapeLHSMatrixKernel

Change-Id: Id6a1bd78f9b1698b64a004e4adebc41002b15745
Reviewed-on: https://review.mlplatform.org/496
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp

diff --git a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp
index 66fafe4..2c072a8 100644
--- a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -72,16 +72,24 @@
     else
     {
         GEMMRHSMatrixInfo rhs_info;
+        GEMMLHSMatrixInfo lhs_info;
         const int         m                         = reshape_info.m();
         const int         n                         = reshape_info.n();
         const int         k                         = reshape_info.k();
         const int         mult_transpose1xW_width   = reshape_info.mult_transpose1xW_width();
         const int         mult_interleave4x4_height = reshape_info.mult_interleave4x4_height();
-        rhs_info.n0                                 = 16 / input1->element_size();
-        rhs_info.k0                                 = 1;
-        rhs_info.h0                                 = mult_transpose1xW_width;
-        rhs_info.interleave                         = false;
-        rhs_info.transpose                          = false;
+        const bool        unroll_block              = dot8_supported(CLKernelLibrary::get().get_device());
+
+        rhs_info.n0         = 16 / input1->element_size();
+        rhs_info.k0         = 1;
+        rhs_info.h0         = mult_transpose1xW_width;
+        rhs_info.interleave = false;
+        rhs_info.transpose  = false;
+        lhs_info.m0         = 4;
+        lhs_info.k0         = 4;
+        lhs_info.v0         = mult_interleave4x4_height;
+        lhs_info.interleave = true;
+        lhs_info.transpose  = unroll_block;
 
         TensorShape tensor_shape0{ input0->tensor_shape() };
         tensor_shape0.set(0, k);
@@ -94,7 +102,7 @@
         const TensorInfo tensor_info0 = input0->clone()->set_tensor_shape(tensor_shape0);
         const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1);
 
-        const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(compute_interleaved_shape(tensor_info0, mult_interleave4x4_height));
+        const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(compute_lhs_reshaped_shape(tensor_info0, lhs_info));
         const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(compute_rhs_reshaped_shape(tensor_info1, rhs_info));
 
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input0, &tensor_info_reshaped0);