From d018425c324704949c7f65230def9586e71f07f5 Mon Sep 17 00:00:00 2001 From: Alexander Monakov Date: Wed, 14 Feb 2024 10:34:24 -1000 Subject: [PATCH] util/bufferiszero: Remove AVX512 variant Thanks to early checks in the inline buffer_is_zero wrapper, the SIMD routines are invoked much more rarely in normal use when most buffers are non-zero. This makes use of AVX512 unprofitable, as it incurs extra frequency and voltage transition periods during which the CPU operates at reduced performance, as described in https://travisdowns.github.io/blog/2020/01/17/avxfreq1.html Signed-off-by: Mikhail Romanov Signed-off-by: Alexander Monakov Reviewed-by: Richard Henderson Message-Id: <20240206204809.9859-4-amonakov@ispras.ru> Signed-off-by: Richard Henderson --- util/bufferiszero.c | 38 +++----------------------------------- 1 file changed, 3 insertions(+), 35 deletions(-) diff --git a/util/bufferiszero.c b/util/bufferiszero.c index f5a3634f9a..641d5f9b9e 100644 --- a/util/bufferiszero.c +++ b/util/bufferiszero.c @@ -64,7 +64,7 @@ buffer_zero_int(const void *buf, size_t len) } } -#if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT) || defined(__SSE2__) +#if defined(CONFIG_AVX2_OPT) || defined(__SSE2__) #include /* Note that each of these vectorized functions require len >= 64. */ @@ -128,41 +128,12 @@ buffer_zero_avx2(const void *buf, size_t len) } #endif /* CONFIG_AVX2_OPT */ -#ifdef CONFIG_AVX512F_OPT -static bool __attribute__((target("avx512f"))) -buffer_zero_avx512(const void *buf, size_t len) -{ - /* Begin with an unaligned head of 64 bytes. */ - __m512i t = _mm512_loadu_si512(buf); - __m512i *p = (__m512i *)(((uintptr_t)buf + 5 * 64) & -64); - __m512i *e = (__m512i *)(((uintptr_t)buf + len) & -64); - - /* Loop over 64-byte aligned blocks of 256. */ - while (p <= e) { - __builtin_prefetch(p); - if (unlikely(_mm512_test_epi64_mask(t, t))) { - return false; - } - t = p[-4] | p[-3] | p[-2] | p[-1]; - p += 4; - } - - t |= _mm512_loadu_si512(buf + len - 4 * 64); - t |= _mm512_loadu_si512(buf + len - 3 * 64); - t |= _mm512_loadu_si512(buf + len - 2 * 64); - t |= _mm512_loadu_si512(buf + len - 1 * 64); - - return !_mm512_test_epi64_mask(t, t); - -} -#endif /* CONFIG_AVX512F_OPT */ - /* * Make sure that these variables are appropriately initialized when * SSE2 is enabled on the compiler command-line, but the compiler is * too old to support CONFIG_AVX2_OPT. */ -#if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT) +#if defined(CONFIG_AVX2_OPT) # define INIT_USED 0 # define INIT_LENGTH 0 # define INIT_ACCEL buffer_zero_int @@ -188,9 +159,6 @@ select_accel_cpuinfo(unsigned info) unsigned len; bool (*fn)(const void *, size_t); } all[] = { -#ifdef CONFIG_AVX512F_OPT - { CPUINFO_AVX512F, 256, buffer_zero_avx512 }, -#endif #ifdef CONFIG_AVX2_OPT { CPUINFO_AVX2, 128, buffer_zero_avx2 }, #endif @@ -208,7 +176,7 @@ select_accel_cpuinfo(unsigned info) return 0; } -#if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT) +#if defined(CONFIG_AVX2_OPT) static void __attribute__((constructor)) init_accel(void) { used_accel = select_accel_cpuinfo(cpuinfo_init()); -- 2.11.4.GIT