From 95eed2b75f1fa4fe3bf3dc9ccabbb99549f03234 Mon Sep 17 00:00:00 2001 From: George Steed Date: Thu, 25 Apr 2024 09:28:31 +0100 Subject: [PATCH] [AArch64] Add Neon dot-product implementation of HammingDistance We can use the Neon dot-product instructions as a slightly faster widening accumulation. This also has the advantage of widening to 32 bits so avoids the risk of overflow present in the original Neon code. Reduction in runtimes observed for HammingDistance compared to the existing Neon code: Cortex-A55: -4.4% Cortex-A510: -26.5% Cortex-A76: -8.1% Cortex-A720: -15.5% Cortex-X1: -4.1% Cortex-X2: -5.1% Bug: libyuv:977 Change-Id: I9e5e10d228c339d905cb2e668a9811ff0a6af5de Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/5490049 Reviewed-by: Frank Barchard Commit-Queue: Frank Barchard --- include/libyuv/compare_row.h | 6 +++++- source/compare.cc | 5 +++++ source/compare_neon64.cc | 32 ++++++++++++++++++++++++++++++++ 3 files changed, 42 insertions(+), 1 deletion(-) diff --git a/include/libyuv/compare_row.h b/include/libyuv/compare_row.h index e78d742be..4ab4b49ad 100644 --- a/include/libyuv/compare_row.h +++ b/include/libyuv/compare_row.h @@ -78,12 +78,13 @@ extern "C" { // The following are available for Neon: #if !defined(LIBYUV_DISABLE_NEON) && \ (defined(__ARM_NEON__) || defined(LIBYUV_NEON) || defined(__aarch64__)) -#define HAS_SUMSQUAREERROR_NEON #define HAS_HAMMINGDISTANCE_NEON +#define HAS_SUMSQUAREERROR_NEON #endif // The following are available for AArch64 Neon: #if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__) +#define HAS_HAMMINGDISTANCE_NEON_DOTPROD #define HAS_SUMSQUAREERROR_NEON_DOTPROD #endif @@ -107,6 +108,9 @@ uint32_t HammingDistance_AVX2(const uint8_t* src_a, uint32_t HammingDistance_NEON(const uint8_t* src_a, const uint8_t* src_b, int count); +uint32_t HammingDistance_NEON_DotProd(const uint8_t* src_a, + const uint8_t* src_b, + int count); uint32_t HammingDistance_MSA(const uint8_t* src_a, const uint8_t* src_b, int count); diff --git a/source/compare.cc b/source/compare.cc index e128dfabc..3783baea2 100644 --- a/source/compare.cc +++ b/source/compare.cc @@ -134,6 +134,11 @@ uint64_t ComputeHammingDistance(const uint8_t* src_a, HammingDistance = HammingDistance_NEON; } #endif +#if defined(HAS_HAMMINGDISTANCE_NEON_DOTPROD) + if (TestCpuFlag(kCpuHasNeonDotProd)) { + HammingDistance = HammingDistance_NEON_DotProd; + } +#endif #if defined(HAS_HAMMINGDISTANCE_SSSE3) if (TestCpuFlag(kCpuHasSSSE3)) { HammingDistance = HammingDistance_SSSE3; diff --git a/source/compare_neon64.cc b/source/compare_neon64.cc index 0a4d52393..0b378c632 100644 --- a/source/compare_neon64.cc +++ b/source/compare_neon64.cc @@ -86,6 +86,38 @@ uint32_t SumSquareError_NEON(const uint8_t* src_a, return sse; } +uint32_t HammingDistance_NEON_DotProd(const uint8_t* src_a, + const uint8_t* src_b, + int count) { + uint32_t diff; + asm volatile( + "movi v4.4s, #0 \n" + "movi v5.4s, #0 \n" + "movi v6.16b, #1 \n" + + "1: \n" + "ldp q0, q1, [%0], #32 \n" + "ldp q2, q3, [%1], #32 \n" + "eor v0.16b, v0.16b, v2.16b \n" + "prfm pldl1keep, [%0, 448] \n" // prefetch 7 lines ahead + "eor v1.16b, v1.16b, v3.16b \n" + "cnt v0.16b, v0.16b \n" + "prfm pldl1keep, [%1, 448] \n" + "cnt v1.16b, v1.16b \n" + "subs %w2, %w2, #32 \n" + "udot v4.4s, v0.16b, v6.16b \n" + "udot v5.4s, v1.16b, v6.16b \n" + "b.gt 1b \n" + + "add v0.4s, v4.4s, v5.4s \n" + "addv s0, v0.4s \n" + "fmov %w3, s0 \n" + : "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(diff) + : + : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6"); + return diff; +} + uint32_t SumSquareError_NEON_DotProd(const uint8_t* src_a, const uint8_t* src_b, int count) {