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: Port add_8x8basis_ssse3() to ASM

Browse Source 
Both GCC and Clang completely unroll the unlikely loop at -O3,
leading to codesize bloat; their code is also suboptimal, as they
don't make use of pmulhrsw (even with -mssse3). This commit
therefore ports the whole function to external assembly. The new
function occupies 176B here vs 1406B for GCC.

Benchmarks for a testcase with huge qscale (notice that the C version
is unrolled just like the unlikely loop in the SSSE3 version):
add_8x8basis_c:                                         43.4 ( 1.00x)
add_8x8basis_ssse3 (old):                               43.6 ( 1.00x)
add_8x8basis_ssse3 (new):                               11.9 ( 3.63x)

Reviewed-by: Kieran Kunhya <kieran@kunhya.com>
Reviewed-by: Lynne <dev@lynne.ee>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
This commit is contained in:
Andreas Rheinhardt
2025-11-15 19:44:02 +01:00
parent 0d3a88e55f
commit aeb138679a
2 changed files with 59 additions and 39 deletions
Download Patch File Download Diff File Expand all files Collapse all files

52
libavcodec/x86/mpegvideoencdsp.asm
View File

@@ -25,6 +25,58 @@
SECTION .text SECTION .text
; void ff_add_8x8basis_ssse3(int16_t rem[64], const int16_t basis[64], int scale)
INIT_XMM ssse3
cglobal add_8x8basis, 3, 3+ARCH_X86_64, 4, rem, basis, scale
movd m0, scaled
add scaled, 1024
add basisq, 128
add remq, 128
%if ARCH_X86_64
%define OFF r3q
mov r3q, -128
cmp scaled, 2047
%else
%define OFF r2q
cmp scaled, 2047
mov r2q, -128
%endif
ja .huge_scale
punpcklwd m0, m0
pshufd m0, m0, 0x0
psllw m0, 5
.loop1:
mova m1, [basisq+OFF]
mova m2, [basisq+OFF+16]
pmulhrsw m1, m0
pmulhrsw m2, m0
paddw m1, [remq+OFF]
paddw m2, [remq+OFF+16]
mova [remq+OFF], m1
mova [remq+OFF+16], m2
add OFF, 32
js .loop1
RET
.huge_scale:
pslld m0, 6
punpcklwd m0, m0
pshufd m1, m0, 0x55
psrlw m0, 1
pshufd m0, m0, 0x0
.loop2:
mova m2, [basisq+OFF]
pmulhrsw m3, m2, m0
pmullw m2, m1
paddw m2, m3
paddw m2, [remq+OFF]
mova [remq+OFF], m2
add OFF, 16
js .loop2
RET
INIT_XMM sse2 INIT_XMM sse2
cglobal mpv_denoise_dct, 3, 4, 7, block, sum, offset cglobal mpv_denoise_dct, 3, 4, 7, block, sum, offset
pxor m6, m6 pxor m6, m6

46
libavcodec/x86/mpegvideoencdsp_init.c
View File

@@ -32,6 +32,7 @@ void ff_mpv_denoise_dct_sse2(int16_t block[64], int dct_error_sum[64],
int ff_pix_sum16_sse2(const uint8_t *pix, ptrdiff_t line_size); int ff_pix_sum16_sse2(const uint8_t *pix, ptrdiff_t line_size);
int ff_pix_sum16_xop(const uint8_t *pix, ptrdiff_t line_size); int ff_pix_sum16_xop(const uint8_t *pix, ptrdiff_t line_size);
int ff_pix_norm1_sse2(const uint8_t *pix, ptrdiff_t line_size); int ff_pix_norm1_sse2(const uint8_t *pix, ptrdiff_t line_size);
void ff_add_8x8basis_ssse3(int16_t rem[64], const int16_t basis[64], int scale);
#if HAVE_INLINE_ASM #if HAVE_INLINE_ASM
#if HAVE_SSSE3_INLINE #if HAVE_SSSE3_INLINE
@@ -83,41 +84,6 @@ static int try_8x8basis_ssse3(const int16_t rem[64], const int16_t weight[64], c
); );
return i; return i;
} }
static void add_8x8basis_ssse3(int16_t rem[64], const int16_t basis[64], int scale)
{
x86_reg i=0;
if (FFABS(scale) < 1024) {
scale *= 1 << (16 + SCALE_OFFSET - BASIS_SHIFT + RECON_SHIFT);
__asm__ volatile(
"movd %3, %%xmm2 \n\t"
"punpcklwd %%xmm2, %%xmm2 \n\t"
"pshufd $0, %%xmm2, %%xmm2 \n\t"
".p2align 4 \n\t"
"1: \n\t"
"movdqa (%1, %0), %%xmm0 \n\t"
"movdqa 16(%1, %0), %%xmm1 \n\t"
"pmulhrsw %%xmm2, %%xmm0 \n\t"
"pmulhrsw %%xmm2, %%xmm1 \n\t"
"paddw (%2, %0), %%xmm0 \n\t"
"paddw 16(%2, %0), %%xmm1 \n\t"
"movdqa %%xmm0, (%2, %0) \n\t"
"movdqa %%xmm1, 16(%2, %0) \n\t"
"add $32, %0 \n\t"
"cmp $128, %0 \n\t" // FIXME optimize & bench
" jb 1b \n\t"
: "+r" (i)
: "r"(basis), "r"(rem), "g"(scale)
XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2")
);
} else {
for (i=0; i<8*8; i++) {
rem[i] += (basis[i]*scale + (1<<(BASIS_SHIFT - RECON_SHIFT-1)))>>(BASIS_SHIFT - RECON_SHIFT);
}
}
}
#endif /* HAVE_SSSE3_INLINE */ #endif /* HAVE_SSSE3_INLINE */
/* Draw the edges of width 'w' of an image of size width, height */ /* Draw the edges of width 'w' of an image of size width, height */
@@ -227,15 +193,17 @@ av_cold void ff_mpegvideoencdsp_init_x86(MpegvideoEncDSPContext *c,
c->draw_edges = draw_edges_mmx; c->draw_edges = draw_edges_mmx;
} }
} }
#endif /* HAVE_INLINE_ASM */
if (X86_SSSE3(cpu_flags)) {
#if HAVE_SSSE3_INLINE #if HAVE_SSSE3_INLINE
if (INLINE_SSSE3(cpu_flags)) {
if (!(avctx->flags & AV_CODEC_FLAG_BITEXACT)) { if (!(avctx->flags & AV_CODEC_FLAG_BITEXACT)) {
c->try_8x8basis = try_8x8basis_ssse3; c->try_8x8basis = try_8x8basis_ssse3;
} }
c->add_8x8basis = add_8x8basis_ssse3;
}
#endif /* HAVE_SSSE3_INLINE */ #endif /* HAVE_SSSE3_INLINE */
#if HAVE_SSSE3_EXTERNAL
c->add_8x8basis = ff_add_8x8basis_ssse3;
#endif
}
#endif /* HAVE_INLINE_ASM */
} }