diff options
author | Gunes Bayir <gunes.bayir@arm.com> | 2022-09-11 15:59:19 +0100 |
---|---|---|
committer | Gunes Bayir <gunes.bayir@arm.com> | 2022-09-16 12:38:59 +0000 |
commit | c4f2743951473f8d97f5a43767fdbb31a4df967c (patch) | |
tree | 26c49b1af8113bb169931f3af5e502904d455a25 /src/core/NEON/wrapper | |
parent | 0d05b6690fe69c57f63ca43d59b551f074613062 (diff) | |
download | ComputeLibrary-c4f2743951473f8d97f5a43767fdbb31a4df967c.tar.gz |
Optimize Quantized/Integer Bilinear Scale for Neon™
This patch introduces several performance optimizations regarding the Bilinear Scale operator with REPLICATE Border mode. Changes apply only to NHWC.
This patch
- Reduces the memory footprint by disabling precomputation of indices and weights when they're not used
- Rewrites the kernels for QASYMM8/QASYMM8_SIGNED/U8(Uint8)
- Adds S8(Int8) Bilinear Scale for Border mode REPLICATE
- Removes Bilinear Scale SVE kernels for Quantized and Integer types and adjust the heuristics to choose the Neon™ implementation
- Adds new test cases where the input and output of the Bilinear Scale operator have different quantization scale and offset
Resolves: COMPMID-5453, COMPMID-5454
Change-Id: I3d251e76e0c6978fd5a0a1795ec62ab536bec93c
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8250
Reviewed-by: SiCong Li <sicong.li@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/wrapper')
-rw-r--r-- | src/core/NEON/wrapper/intrinsics/cvt.h | 20 |
1 files changed, 18 insertions, 2 deletions
diff --git a/src/core/NEON/wrapper/intrinsics/cvt.h b/src/core/NEON/wrapper/intrinsics/cvt.h index e52e3dd0c4..baad1319b2 100644 --- a/src/core/NEON/wrapper/intrinsics/cvt.h +++ b/src/core/NEON/wrapper/intrinsics/cvt.h @@ -59,19 +59,35 @@ VCVT_TO_F16_IMPL(float16x4_t, float32x4_t, vcvt, f16, f32) #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC template <typename T> -inline typename std::enable_if<std::is_same<T, uint8_t>::value, uint32x4_t>::type +inline typename std::enable_if < std::is_same<T, uint8_t>::value || std::is_same<T, uint32_t>::value, uint32x4_t >::type vcvt(const float32x4_t &a) { return vcvtq_u32_f32(a); } template <typename T> -inline typename std::enable_if<std::is_same<T, int8_t>::value, int32x4_t>::type +inline typename std::enable_if < std::is_same<T, int8_t>::value || std::is_same<T, int32_t>::value, int32x4_t >::type vcvt(const float32x4_t &a) { return vcvtq_s32_f32(a); } +#ifdef __aarch64__ +template <typename T> +inline typename std::enable_if<std::is_same<T, uint32_t>::value, uint32x4_t>::type +vcvta(const float32x4_t &a) +{ + return vcvtaq_u32_f32(a); +} + +template <typename T> +inline typename std::enable_if<std::is_same<T, int32_t>::value, int32x4_t>::type +vcvta(const float32x4_t &a) +{ + return vcvtaq_s32_f32(a); +} +#endif //__aarch64__ + #if defined(ARM_COMPUTE_ENABLE_BF16) /** Convert 2x128-bit floating point vectors into 1x128-bit bfloat16 vector * |