diff options
| author | Michele Di Giorgio <michele.digiorgio@arm.com> | 2021-02-02 14:59:09 +0000 |
|---|---|---|
| committer | Michele Di Giorgio <michele.digiorgio@arm.com> | 2021-02-04 17:43:55 +0000 |
| commit | cf87f509fc23d02c56569f794a3fb59e1b8be277 (patch) | |
| tree | 0fe55158f2065dc6a314e82935558b9748165285 /src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_avg_generic_depthfirst/generic.cpp | |
| parent | 89de118ccbebd5a943634137d0c160d4867da49c (diff) | |
| download | ComputeLibrary-cf87f509fc23d02c56569f794a3fb59e1b8be277.tar.gz | |
Tweak scheduling use of SQDMULH in quantized AVG pooling
Resolves COMPMID-4195
Change-Id: Ie5116c1ddddccafba40432fd4b5245bb27890a88
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4997
Reviewed-by: TeresaARM <teresa.charlinreyes@arm.com>
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_avg_generic_depthfirst/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_avg_generic_depthfirst/generic.cpp | 6 |
1 files changed, 3 insertions, 3 deletions
diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_avg_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_avg_generic_depthfirst/generic.cpp index 51d2973ce1..928eb412b5 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_avg_generic_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_avg_generic_depthfirst/generic.cpp @@ -41,10 +41,10 @@ namespace { constexpr RescaleParams rescale_params[8] = { {0x40000000, -0}, // 1/2 - {0x55555555, -1}, // 1/3 + {0x55555556, -1}, // 1/3 {0x40000000, -1}, // 1/4 {0x66666666, -2}, // 1/5 - {0x55555555, -2}, // 1/6 + {0x55555556, -2}, // 1/6 {0x49249249, -2}, // 1/7 {0x40000000, -2}, // 1/8 {0x71c71c72, -3}, // 1/9 @@ -108,7 +108,7 @@ void sve_s8q_nhwc_avg_generic_depthfirst_impl( "mov %w[combined_rescale_value], v18.s[0]\n" : [combined_rescale_value] "=r" (combined_rescale_value) : [per_layer_mul] "r" (qp.per_layer_mul), [rescale_value] "r" (rescale_value) - : "q16", "q17", "q18" + : "v16", "v17", "v18" ); __asm__ __volatile__( |