From 3321950c109b416e63eda59c76e6365abc2072b8 Mon Sep 17 00:00:00 2001 From: Sebastian Pop Date: Thu, 2 Jul 2020 16:57:58 +0000 Subject: [PATCH] [aarch64] improve hscale by 50% with multi-threading hscale is bound by the number of multiply-adds available on a given core. The patch doubles the number of multiply-adds by distributing half the load to a helper thread. The performance improves up to 50% on Graviton2 Arm Neoverse-N1 processors. $ ./ffmpeg_g -nostats -f lavfi -i testsrc2=4k:d=2 -vf bench=start,scale=1024x1024,bench=stop -f null - before: [bench @ 0xaaaad62c3d30] t:0.013293 avg:0.013315 max:0.013697 min:0.013293 after: [bench @ 0xaaaae9346d30] t:0.009637 avg:0.009691 max:0.010005 min:0.009637 38% improvement scale=1280x720 49% improvement before: [bench @ 0xaaaadba88d30] t:0.015973 avg:0.016321 max:0.016917 min:0.015973 after: [bench @ 0xaaaabc78dd30] t:0.010823 avg:0.010869 max:0.011552 min:0.010708 scale=852x480 45% improvement before: [bench @ 0xaaaaeeed0d30] t:0.013731 avg:0.013727 max:0.013773 min:0.013279 after: [bench @ 0xaaaaf5f5dd30] t:0.009279 avg:0.009296 max:0.009328 min:0.009187 scale=640x360 45% improvement before: [bench @ 0xaaaacee25d30] t:0.012010 avg:0.012006 max:0.012053 min:0.011653 after: [bench @ 0xaaaaea2b5d30] t:0.008077 avg:0.008084 max:0.008409 min:0.008057 scale=284x160 36% improvement before: [bench @ 0xaaaadbb9ed30] t:0.008384 avg:0.008367 max:0.008421 min:0.008193 after: [bench @ 0xaaaafb1d6d30] t:0.006099 avg:0.006100 max:0.006120 min:0.006026 --- libswscale/aarch64/swscale.c | 44 ++++++++++++++++++++++++++++++++++- libswscale/swscale_internal.h | 15 ++++++++++++ libswscale/utils.c | 14 +++++++++++ 3 files changed, 72 insertions(+), 1 deletion(-) diff --git a/libswscale/aarch64/swscale.c b/libswscale/aarch64/swscale.c index eecbea88ca..56191ab612 100644 --- a/libswscale/aarch64/swscale.c +++ b/libswscale/aarch64/swscale.c @@ -20,6 +20,7 @@ #include "libswscale/swscale.h" #include "libswscale/swscale_internal.h" #include "libavutil/aarch64/cpu.h" +#include "pthread.h" void ff_hscale_8_to_15_neon(SwsContext *c, int16_t *dst, int dstW, const uint8_t *src, const int16_t *filter, @@ -29,6 +30,47 @@ void ff_yuv2planeX_8_neon(const int16_t *filter, int filterSize, const int16_t **src, uint8_t *dest, int dstW, const uint8_t *dither, int offset); +void *aarch64_thread_fun(void *ctx) { + SwsContext *c = (SwsContext *)ctx; + + /* Run until canceled. */ + do { + if (__atomic_load_n(&c->has_work, __ATOMIC_ACQUIRE)) { + ff_hscale_8_to_15_neon(c, c->dst_, c->dstW_, c->src_, c->filter_, c->filterPos_, c->filterSize_); + __atomic_store_n(&c->has_work, 0, __ATOMIC_RELEASE); + } + } while (1); +} + +static void ff_hscale_8_to_15_neon_1(SwsContext *c, int16_t *dst, int dstW, + const uint8_t *src, const int16_t *filter, + const int32_t *filterPos, int filterSize) +{ + int rem; + if (dstW <= 0 || filterSize <= 0) + return; + + if (!c) { + ff_hscale_8_to_15_neon(c, dst, dstW, src, filter, filterPos, filterSize); + return; + } + /* Half the outer loop load. */ + rem = dstW % 2; + dstW /= 2; + c->dst_ = dst; + c->dstW_ = dstW; + c->src_ = src; + c->filter_ = filter; + c->filterPos_ = filterPos; + c->filterSize_ = filterSize; + __atomic_store_n(&c->has_work, 1, __ATOMIC_RELEASE); + + ff_hscale_8_to_15_neon(c, dst + dstW, dstW + rem, src, filter + dstW * filterSize, filterPos + dstW, filterSize); + + /* Wait for the helper thread to finish work. */ + do {} while (__atomic_load_n(&c->has_work, __ATOMIC_ACQUIRE)); +} + av_cold void ff_sws_init_swscale_aarch64(SwsContext *c) { int cpu_flags = av_get_cpu_flags(); @@ -38,7 +80,7 @@ av_cold void ff_sws_init_swscale_aarch64(SwsContext *c) (c->hLumFilterSize % 8) == 0 && (c->hChrFilterSize % 8) == 0) { - c->hyScale = c->hcScale = ff_hscale_8_to_15_neon; + c->hyScale = c->hcScale = ff_hscale_8_to_15_neon_1; } if (c->dstBpc == 8) { c->yuv2planeX = ff_yuv2planeX_8_neon; diff --git a/libswscale/swscale_internal.h b/libswscale/swscale_internal.h index 1a1b6f0dee..fcfb7f0ef0 100644 --- a/libswscale/swscale_internal.h +++ b/libswscale/swscale_internal.h @@ -625,9 +625,24 @@ typedef struct SwsContext { SwsDither dither; SwsAlphaBlend alphablend; + +#ifdef __aarch64__ + pthread_t thread; + int has_work; + int16_t *dst_; + int dstW_; + const uint8_t *src_; + const int16_t *filter_; + const int32_t *filterPos_; + int filterSize_; +#endif } SwsContext; //FIXME check init (where 0) +#ifdef __aarch64__ +void *aarch64_thread_fun(void *ctx); +#endif + SwsFunc ff_yuv2rgb_get_func_ptr(SwsContext *c); int ff_yuv2rgb_c_init_tables(SwsContext *c, const int inv_table[4], int fullRange, int brightness, diff --git a/libswscale/utils.c b/libswscale/utils.c index dcd1dbaa76..75d40c11c5 100644 --- a/libswscale/utils.c +++ b/libswscale/utils.c @@ -38,6 +38,10 @@ #include #endif +#ifdef __aarch64__ +#include "pthread.h" +#endif + #include "libavutil/attributes.h" #include "libavutil/avassert.h" #include "libavutil/avutil.h" @@ -1846,6 +1850,12 @@ av_cold int sws_init_context(SwsContext *c, SwsFilter *srcFilter, } } +#ifdef __aarch64__ + c->has_work = 0; + if (0 != pthread_create (&c->thread, NULL, aarch64_thread_fun, (void *)c)) + goto fail; +#endif + c->swscale = ff_getSwsFunc(c); return ff_init_filters(c); nomem: @@ -2385,6 +2395,10 @@ void sws_freeContext(SwsContext *c) ff_free_filters(c); +#ifdef __aarch64__ + pthread_cancel(c->thread); +#endif + av_free(c); } -- 2.25.1