diff options
| author | Michalis Spyrou <michalis.spyrou@arm.com> | 2020-10-09 11:52:10 +0100 |
|---|---|---|
| committer | Sheri Zhang <sheri.zhang@arm.com> | 2020-10-09 12:43:56 +0000 |
| commit | c226853f80d53619a2f49e646635e04ee0885c3b (patch) | |
| tree | 252620ed7db38810c1590d7caab62625ec07ad6f /src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.cpp | |
| parent | 4adaddbaa633a4025f29f2e0a63c7126d9d7c530 (diff) | |
| download | ComputeLibrary-c226853f80d53619a2f49e646635e04ee0885c3b.tar.gz | |
COMPMID-3794: Fix window loops causing performance regression
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Change-Id: Id4d95c6ce5fed91bb079b8bfe1abceedefd20c97
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4117
Reviewed-by: Sheri Zhang <sheri.zhang@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.cpp')
| -rw-r--r-- | src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.cpp | 6 |
1 files changed, 3 insertions, 3 deletions
diff --git a/src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.cpp index 5710897329..4ac33d1e29 100644 --- a/src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.cpp +++ b/src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.cpp @@ -139,7 +139,7 @@ void run_offset_contribution(const Window &window, const int32x4_t b_offset_term_s32_vec = vdupq_n_s32(b_offset_term_s32); int x = window_start_x; - for(; x < (window_end_x - window_step_x); x += window_step_x) + for(; x <= (window_end_x - window_step_x); x += window_step_x) { // Compute the leftover term due to a_offset. int32x4x4_t a_offset_term_s32 = @@ -237,7 +237,7 @@ void run_offset_contribution(const Window &window, const int32x4_t b_offset_term_s32_vec = vdupq_n_s32(b_offset_term_s32); int x = window_start_x; - for(; x < (window_end_x - window_step_x); x += window_step_x) + for(; x <= (window_end_x - window_step_x); x += window_step_x) { int32x4x4_t in_s32 = { @@ -291,7 +291,7 @@ void run_offset_contribution(const Window &window, auto mm_result_ptr = reinterpret_cast<int32_t *>(mm_result_it.ptr()); int x = window_start_x; - for(; x < (window_end_x - window_step_x); x += window_step_x) + for(; x <= (window_end_x - window_step_x); x += window_step_x) { // Compute the leftover term due to a_offset. int32x4x4_t a_offset_term_s32 = |