commit 93b9bdb49f0f1e715c7ad251b6886c1a49945b5a
author Anthony Barbier <anthony.barbier@arm.com> Tue Dec 12 11:27:55 2017 +0000
committer Anthony Barbier <anthony.barbier@arm.com> Fri Nov 02 16:42:33 2018 +0000
tree c31436a3c8f634fb22bd30a9ee25f8788fbdb69c
parent 40626803bc2df7cda8a86e947636e9b950726384

COMPMID-743 Fixed AccessWindow in GEMMLowpMatrixMultiply

Change-Id: I8c9ae389756aa8ca346ad1ebfa1feac9a47964a5
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/112863
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Anthony Barbier <anthony.barbier@arm.com>

src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp

diff --git a/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp
index d22773d..9104f0b 100644
--- a/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp
@@ -743,6 +743,7 @@
 
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(in0_shape[2] != out_shape[2], "Output tensor must have the same number of batches of input0 tensor");
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(in1_shape[2] != 1 && in0_shape[2] != in1_shape[2], "Input1 tensor must have the same number of batches of input0 or the number of batches must be set to 1");
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(in1_shape[0] % 16, "Input1's width must be a multiple of 16");
     }
 
     return Status{};
@@ -777,8 +778,10 @@
     {
         win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
 
-        AccessWindowStatic     in0_access(input0, 0, 0, ceil_to_multiple(input0->dimension(0), 8), input0->dimension(1));
-        AccessWindowHorizontal in1_access(input1, 0, num_elems_processed_per_iteration_x);
+        unsigned int num_k_iterations = ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x) / 16;
+        // For each iteration of "k" we increment the input pointer by 4, and we load 8 elements a the time:
+        AccessWindowStatic     in0_access(input0, 0, 0, (num_k_iterations - 1) * 4 + 8, input0->dimension(1));
+        AccessWindowHorizontal in1_access(input1, 0, input1->dimension(0));
         AccessWindowRectangle  output_access(output, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
 
         window_changed = update_window_and_padding(win, in0_access, in1_access, output_access);