From 2393313673442256d70e8860f2a4e9df496c1bee Mon Sep 17 00:00:00 2001 From: Martin Vignali Date: Thu, 5 Oct 2017 16:31:01 +0200 Subject: [PATCH 2/4] libavcodec/blockdsp : add clear_block_prores using a loop and call clear_block, can be slower than memset this new func, clear even block number and is faster than a memset --- libavcodec/blockdsp.c | 6 ++++++ libavcodec/blockdsp.h | 1 + libavcodec/x86/blockdsp.asm | 34 ++++++++++++++++++++++++++++++++++ libavcodec/x86/blockdsp_init.c | 4 ++++ tests/checkasm/blockdsp.c | 19 +++++++++++++++++++ 5 files changed, 64 insertions(+) diff --git a/libavcodec/blockdsp.c b/libavcodec/blockdsp.c index c7efe7e77b..564149243f 100644 --- a/libavcodec/blockdsp.c +++ b/libavcodec/blockdsp.c @@ -35,6 +35,11 @@ static void clear_blocks_c(int16_t *blocks) memset(blocks, 0, sizeof(int16_t) * 6 * 64); } +static void clear_blocks_prores_c(int16_t * blocks, ptrdiff_t block_count) +{ + memset(blocks, 0, sizeof(int16_t) * 64 * block_count); +} + static void fill_block16_c(uint8_t *block, uint8_t value, ptrdiff_t line_size, int h) { @@ -61,6 +66,7 @@ av_cold void ff_blockdsp_init(BlockDSPContext *c, AVCodecContext *avctx) { c->clear_block = clear_block_c; c->clear_blocks = clear_blocks_c; + c->clear_blocks_prores = clear_blocks_prores_c; c->fill_block_tab[0] = fill_block16_c; c->fill_block_tab[1] = fill_block8_c; diff --git a/libavcodec/blockdsp.h b/libavcodec/blockdsp.h index 26fc2ea13b..56881b38c9 100644 --- a/libavcodec/blockdsp.h +++ b/libavcodec/blockdsp.h @@ -35,6 +35,7 @@ typedef void (*op_fill_func)(uint8_t *block /* align width (8 or 16) */, typedef struct BlockDSPContext { void (*clear_block)(int16_t *block /* align 32 */); void (*clear_blocks)(int16_t *blocks /* align 32 */); + void (*clear_blocks_prores)(int16_t * blocks, ptrdiff_t block_count); /* align 32, block_count even */ op_fill_func fill_block_tab[2]; } BlockDSPContext; diff --git a/libavcodec/x86/blockdsp.asm b/libavcodec/x86/blockdsp.asm index 9d203df8f5..e99425272a 100644 --- a/libavcodec/x86/blockdsp.asm +++ b/libavcodec/x86/blockdsp.asm @@ -86,3 +86,37 @@ INIT_XMM sse CLEAR_BLOCKS 1 INIT_YMM avx CLEAR_BLOCKS 1 + + +;---------------------------------------------------------------- +; void ff_clear_blocks_prores(int16_t *blocks, ptrdiff_t block_count); +;---------------------------------------------------------------- +; %1 = number of xmm registers used +%macro CLEAR_BLOCKS_PRORES 1 +cglobal clear_blocks_prores, 2, 2, %1, blocks, block_count + shl block_countq, 7 ; 64*sizeof(int16_t) * block_count + add blocksq, block_countq + neg block_countq + ZERO m0, m0, m0 +.loop: + mova [blocksq+ block_countq +mmsize*0], m0 + mova [blocksq+ block_countq +mmsize*1], m0 + mova [blocksq+ block_countq +mmsize*2], m0 + mova [blocksq+ block_countq +mmsize*3], m0 + mova [blocksq+ block_countq +mmsize*4], m0 + mova [blocksq+ block_countq +mmsize*5], m0 + mova [blocksq+ block_countq +mmsize*6], m0 + mova [blocksq+ block_countq +mmsize*7], m0 + add block_countq, mmsize*8 + js .loop + RET +%endmacro + +INIT_MMX mmx +%define ZERO pxor +CLEAR_BLOCKS_PRORES 0 +INIT_XMM sse +%define ZERO xorps +CLEAR_BLOCKS_PRORES 1 +INIT_YMM avx +CLEAR_BLOCKS_PRORES 1 diff --git a/libavcodec/x86/blockdsp_init.c b/libavcodec/x86/blockdsp_init.c index 8b01a447cd..80988598e2 100644 --- a/libavcodec/x86/blockdsp_init.c +++ b/libavcodec/x86/blockdsp_init.c @@ -32,6 +32,8 @@ void ff_clear_block_avx(int16_t *block); void ff_clear_blocks_mmx(int16_t *blocks); void ff_clear_blocks_sse(int16_t *blocks); void ff_clear_blocks_avx(int16_t *blocks); +void ff_clear_blocks_prores_sse(int16_t * blocks, ptrdiff_t block_count); +void ff_clear_blocks_prores_avx(int16_t * blocks, ptrdiff_t block_count); av_cold void ff_blockdsp_init_x86(BlockDSPContext *c, AVCodecContext *avctx) @@ -51,10 +53,12 @@ av_cold void ff_blockdsp_init_x86(BlockDSPContext *c, if (EXTERNAL_SSE(cpu_flags)) { c->clear_block = ff_clear_block_sse; c->clear_blocks = ff_clear_blocks_sse; + c->clear_blocks_prores = ff_clear_blocks_prores_sse; } if (EXTERNAL_AVX_FAST(cpu_flags)) { c->clear_block = ff_clear_block_avx; c->clear_blocks = ff_clear_blocks_avx; + c->clear_blocks_prores = ff_clear_blocks_prores_avx; } #endif /* HAVE_X86ASM */ } diff --git a/tests/checkasm/blockdsp.c b/tests/checkasm/blockdsp.c index c753506b3c..5136aa80f5 100644 --- a/tests/checkasm/blockdsp.c +++ b/tests/checkasm/blockdsp.c @@ -28,6 +28,8 @@ #include "libavutil/internal.h" #include "libavutil/intreadwrite.h" +#define PRORES_MAX_BUF_SIZE 2048//32*8*8 + #define randomize_buffers(size) \ do { \ int i; \ @@ -51,6 +53,20 @@ do { \ } \ } while (0) +static void check_clear_blocks_prores(ptrdiff_t block_count) { + LOCAL_ALIGNED_32(uint16_t, buf0, [PRORES_MAX_BUF_SIZE]); + LOCAL_ALIGNED_32(uint16_t, buf1, [PRORES_MAX_BUF_SIZE]); + + declare_func(void, uint16_t * block, ptrdiff_t size); + + randomize_buffers(PRORES_MAX_BUF_SIZE/2); + call_ref(buf0, block_count); + call_new(buf1, block_count); + if (memcmp(buf0, buf1, PRORES_MAX_BUF_SIZE)) + fail(); + bench_new(buf0, block_count); +} + void checkasm_check_blockdsp(void) { LOCAL_ALIGNED_32(uint16_t, buf0, [6 * 8 * 8]); @@ -64,5 +80,8 @@ void checkasm_check_blockdsp(void) check_clear(clear_block, 8 * 8); check_clear(clear_blocks, 8 * 8 * 6); + if (check_func(h.clear_blocks_prores, "blockdsp.clear_blocks_prores")) + check_clear_blocks_prores(32); + report("blockdsp"); } -- 2.11.0 (Apple Git-81)