virtualenv/FFmpeg
1
0
Fork 0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2025-11-23 21:54:53 +02:00
Code Issues Releases Activity

avcodec/x86/mpegvideoencdsp_init: Use xmm registers in SSSE3 functions

Browse Source 
Improves performance and no longer breaks the ABI (by forgetting
to call emms).

Old benchmarks:
add_8x8basis_c:                                         43.6 ( 1.00x)
add_8x8basis_ssse3:                                     12.3 ( 3.55x)

New benchmarks:
add_8x8basis_c:                                         43.0 ( 1.00x)
add_8x8basis_ssse3:                                      6.3 ( 6.79x)

Notice that the output of try_8x8basis_ssse3 changes a bit:
Before this commit, it computes certain values and adds the values
for i,i+1,i+4 and i+5 before right shifting them; now it adds
the values for i,i+1,i+8,i+9. The second pair in these lists
could be avoided (by shifting xmm0 and xmm1 before adding both together
instead of only shifting xmm0 after adding them), but the former
i,i+1 is inherent in using pmaddwd. This is the reason that this
function is not bitexact.

Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
This commit is contained in:
Andreas Rheinhardt
2025-10-12 08:04:11 +02:00
parent cffd029e98
commit f4a87d8ca4
3 changed files with 55 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
libavcodec/mpegvideo_enc.c
View File

@@ -2296,7 +2296,7 @@ static av_always_inline void encode_mb_internal(MPVEncContext *const s,
* and neither of these encoders currently supports 444. */ * and neither of these encoders currently supports 444. */
#define INTERLACED_DCT(s) ((chroma_format == CHROMA_420 || chroma_format == CHROMA_422) && \ #define INTERLACED_DCT(s) ((chroma_format == CHROMA_420 || chroma_format == CHROMA_422) && \
(s)->c.avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) (s)->c.avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT)
int16_t weight[12][64]; DECLARE_ALIGNED(16, int16_t, weight)[12][64];
int16_t orig[12][64]; int16_t orig[12][64];
const int mb_x = s->c.mb_x; const int mb_x = s->c.mb_x;
const int mb_y = s->c.mb_y; const int mb_y = s->c.mb_y;
@@ -4293,7 +4293,7 @@ static int dct_quantize_trellis_c(MPVEncContext *const s,
return last_non_zero; return last_non_zero;
} }
static int16_t basis[64][64]; static DECLARE_ALIGNED(16, int16_t, basis)[64][64];
static void build_basis(uint8_t *perm){ static void build_basis(uint8_t *perm){
int i, j, x, y; int i, j, x, y;
@@ -4317,7 +4317,7 @@ static void build_basis(uint8_t *perm){
static int dct_quantize_refine(MPVEncContext *const s, //FIXME breaks denoise? static int dct_quantize_refine(MPVEncContext *const s, //FIXME breaks denoise?
int16_t *block, int16_t *weight, int16_t *orig, int16_t *block, int16_t *weight, int16_t *orig,
int n, int qscale){ int n, int qscale){
int16_t rem[64]; DECLARE_ALIGNED(16, int16_t, rem)[64];
LOCAL_ALIGNED_16(int16_t, d1, [64]); LOCAL_ALIGNED_16(int16_t, d1, [64]);
const uint8_t *scantable; const uint8_t *scantable;
const uint8_t *perm_scantable; const uint8_t *perm_scantable;

99
libavcodec/x86/mpegvideoencdsp_init.c
View File

@@ -35,13 +35,6 @@ int ff_pix_norm1_sse2(const uint8_t *pix, ptrdiff_t line_size);
#if HAVE_SSSE3_INLINE #if HAVE_SSSE3_INLINE
#define SCALE_OFFSET -1 #define SCALE_OFFSET -1
/*
* pmulhrsw: dst[0 - 15] = (src[0 - 15] * dst[0 - 15] + 0x4000)[15 - 30]
*/
#define PMULHRW(x, y, s, o) \
"pmulhrsw " #s ", " #x " \n\t" \
"pmulhrsw " #s ", " #y " \n\t"
#define MAX_ABS 512 #define MAX_ABS 512
static int try_8x8basis_ssse3(const int16_t rem[64], const int16_t weight[64], const int16_t basis[64], int scale) static int try_8x8basis_ssse3(const int16_t rem[64], const int16_t weight[64], const int16_t basis[64], int scale)
@@ -52,36 +45,39 @@ static int try_8x8basis_ssse3(const int16_t rem[64], const int16_t weight[64], c
scale <<= 16 + SCALE_OFFSET - BASIS_SHIFT + RECON_SHIFT; scale <<= 16 + SCALE_OFFSET - BASIS_SHIFT + RECON_SHIFT;
__asm__ volatile( __asm__ volatile(
"pxor %%mm7, %%mm7 \n\t" "pxor %%xmm2, %%xmm2 \n\t"
"movd %4, %%mm5 \n\t" "movd %4, %%xmm3 \n\t"
"punpcklwd %%mm5, %%mm5 \n\t" "punpcklwd %%xmm3, %%xmm3 \n\t"
"punpcklwd %%mm5, %%mm5 \n\t" "pshufd $0, %%xmm3, %%xmm3 \n\t"
".p2align 4 \n\t" ".p2align 4 \n\t"
"1: \n\t" "1: \n\t"
"movq (%1, %0), %%mm0 \n\t" "movdqa (%1, %0), %%xmm0 \n\t"
"movq 8(%1, %0), %%mm1 \n\t" "movdqa 16(%1, %0), %%xmm1 \n\t"
PMULHRW(%%mm0, %%mm1, %%mm5, %%mm6) "pmulhrsw %%xmm3, %%xmm0 \n\t"
"paddw (%2, %0), %%mm0 \n\t" "pmulhrsw %%xmm3, %%xmm1 \n\t"
"paddw 8(%2, %0), %%mm1 \n\t" "paddw (%2, %0), %%xmm0 \n\t"
"psraw $6, %%mm0 \n\t" "paddw 16(%2, %0), %%xmm1 \n\t"
"psraw $6, %%mm1 \n\t" "psraw $6, %%xmm0 \n\t"
"pmullw (%3, %0), %%mm0 \n\t" "psraw $6, %%xmm1 \n\t"
"pmullw 8(%3, %0), %%mm1 \n\t" "pmullw (%3, %0), %%xmm0 \n\t"
"pmaddwd %%mm0, %%mm0 \n\t" "pmullw 16(%3, %0), %%xmm1 \n\t"
"pmaddwd %%mm1, %%mm1 \n\t" "pmaddwd %%xmm0, %%xmm0 \n\t"
"paddd %%mm1, %%mm0 \n\t" "pmaddwd %%xmm1, %%xmm1 \n\t"
"psrld $4, %%mm0 \n\t" "paddd %%xmm1, %%xmm0 \n\t"
"paddd %%mm0, %%mm7 \n\t" "psrld $4, %%xmm0 \n\t"
"add $16, %0 \n\t" "paddd %%xmm0, %%xmm2 \n\t"
"cmp $128, %0 \n\t" //FIXME optimize & bench "add $32, %0 \n\t"
" jb 1b \n\t" "cmp $128, %0 \n\t" //FIXME optimize & bench
"pshufw $0x0E, %%mm7, %%mm6 \n\t" " jb 1b \n\t"
"paddd %%mm6, %%mm7 \n\t" // faster than phaddd on core2 "pshufd $0x0E, %%xmm2, %%xmm0 \n\t"
"psrld $2, %%mm7 \n\t" "paddd %%xmm0, %%xmm2 \n\t"
"movd %%mm7, %0 \n\t" "pshufd $0x01, %%xmm2, %%xmm0 \n\t"
"paddd %%xmm0, %%xmm2 \n\t"
"psrld $2, %%xmm2 \n\t"
"movd %%xmm2, %0 \n\t"
: "+r" (i) : "+r" (i)
: "r"(basis), "r"(rem), "r"(weight), "g"(scale) : "r"(basis), "r"(rem), "r"(weight), "g"(scale)
XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3")
); );
return i; return i;
} }
@@ -93,24 +89,25 @@ static void add_8x8basis_ssse3(int16_t rem[64], const int16_t basis[64], int sca
if (FFABS(scale) < 1024) { if (FFABS(scale) < 1024) {
scale <<= 16 + SCALE_OFFSET - BASIS_SHIFT + RECON_SHIFT; scale <<= 16 + SCALE_OFFSET - BASIS_SHIFT + RECON_SHIFT;
__asm__ volatile( __asm__ volatile(
"movd %3, %%mm5 \n\t" "movd %3, %%xmm2 \n\t"
"punpcklwd %%mm5, %%mm5 \n\t" "punpcklwd %%xmm2, %%xmm2 \n\t"
"punpcklwd %%mm5, %%mm5 \n\t" "pshufd $0, %%xmm2, %%xmm2 \n\t"
".p2align 4 \n\t" ".p2align 4 \n\t"
"1: \n\t" "1: \n\t"
"movq (%1, %0), %%mm0 \n\t" "movdqa (%1, %0), %%xmm0 \n\t"
"movq 8(%1, %0), %%mm1 \n\t" "movdqa 16(%1, %0), %%xmm1 \n\t"
PMULHRW(%%mm0, %%mm1, %%mm5, %%mm6) "pmulhrsw %%xmm2, %%xmm0 \n\t"
"paddw (%2, %0), %%mm0 \n\t" "pmulhrsw %%xmm2, %%xmm1 \n\t"
"paddw 8(%2, %0), %%mm1 \n\t" "paddw (%2, %0), %%xmm0 \n\t"
"movq %%mm0, (%2, %0) \n\t" "paddw 16(%2, %0), %%xmm1 \n\t"
"movq %%mm1, 8(%2, %0) \n\t" "movdqa %%xmm0, (%2, %0) \n\t"
"add $16, %0 \n\t" "movdqa %%xmm1, 16(%2, %0) \n\t"
"cmp $128, %0 \n\t" // FIXME optimize & bench "add $32, %0 \n\t"
" jb 1b \n\t" "cmp $128, %0 \n\t" // FIXME optimize & bench
" jb 1b \n\t"
: "+r" (i) : "+r" (i)
: "r"(basis), "r"(rem), "g"(scale) : "r"(basis), "r"(rem), "g"(scale)
XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2")
); );
} else { } else {
for (i=0; i<8*8; i++) { for (i=0; i<8*8; i++) {

8
tests/checkasm/mpegvideoencdsp.c
View File

@@ -39,13 +39,13 @@
static void check_add_8x8basis(MpegvideoEncDSPContext *c) static void check_add_8x8basis(MpegvideoEncDSPContext *c)
{ {
declare_func_emms(AV_CPU_FLAG_SSSE3, void, int16_t rem[64], const int16_t basis[64], int scale); declare_func(void, int16_t rem[64], const int16_t basis[64], int scale);
if (check_func(c->add_8x8basis, "add_8x8basis")) { if (check_func(c->add_8x8basis, "add_8x8basis")) {
// FIXME: What are the actual ranges for these values? // FIXME: What are the actual ranges for these values?
int scale = sign_extend(rnd(), 12); int scale = sign_extend(rnd(), 12);
int16_t rem1[64]; DECLARE_ALIGNED(16, int16_t, rem1)[64];
int16_t rem2[64]; DECLARE_ALIGNED(16, int16_t, rem2)[64];
int16_t basis[64]; DECLARE_ALIGNED(16, int16_t, basis)[64];
randomize_buffer_clipped(basis, -15760, 15760); randomize_buffer_clipped(basis, -15760, 15760);
randomize_buffers(rem1, sizeof(rem1)); randomize_buffers(rem1, sizeof(rem1));