[FFmpeg-cvslog] dct32: Add AVX implementation of 32-point DCT
Vitor Sessak
git at videolan.org
Sun May 22 05:14:33 CEST 2011
ffmpeg | branch: master | Vitor Sessak <vitor1001 at gmail.com> | Sat May 14 14:17:15 2011 +0200| [6204feb160c843320f6001d7e2bb2361c82b90ca] | committer: Reinhard Tartler
dct32: Add AVX implementation of 32-point DCT
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=6204feb160c843320f6001d7e2bb2361c82b90ca
---
libavcodec/mpegaudiodec.c | 4 +-
libavcodec/x86/dct32_sse.asm | 334 +++++++++++++++++++++++++++---------------
libavcodec/x86/fft.c | 4 +-
libavcodec/x86/fft.h | 1 +
4 files changed, 224 insertions(+), 119 deletions(-)
diff --git a/libavcodec/mpegaudiodec.c b/libavcodec/mpegaudiodec.c
index 960d13d..ccc93ad 100644
--- a/libavcodec/mpegaudiodec.c
+++ b/libavcodec/mpegaudiodec.c
@@ -69,9 +69,9 @@ typedef struct MPADecodeContext {
uint32_t free_format_next_header;
GetBitContext gb;
GetBitContext in_gb;
- DECLARE_ALIGNED(16, MPA_INT, synth_buf)[MPA_MAX_CHANNELS][512 * 2];
+ DECLARE_ALIGNED(32, MPA_INT, synth_buf)[MPA_MAX_CHANNELS][512 * 2];
int synth_buf_offset[MPA_MAX_CHANNELS];
- DECLARE_ALIGNED(16, INTFLOAT, sb_samples)[MPA_MAX_CHANNELS][36][SBLIMIT];
+ DECLARE_ALIGNED(32, INTFLOAT, sb_samples)[MPA_MAX_CHANNELS][36][SBLIMIT];
INTFLOAT mdct_buf[MPA_MAX_CHANNELS][SBLIMIT * 18]; /* previous samples, for layer 3 MDCT */
GranuleDef granules[2][2]; /* Used in Layer 3 */
#ifdef DEBUG
diff --git a/libavcodec/x86/dct32_sse.asm b/libavcodec/x86/dct32_sse.asm
index fa0a502..2e1176c 100644
--- a/libavcodec/x86/dct32_sse.asm
+++ b/libavcodec/x86/dct32_sse.asm
@@ -20,31 +20,41 @@
;******************************************************************************
%include "x86inc.asm"
+%include "config.asm"
SECTION_RODATA 32
align 32
ps_cos_vec: dd 0.500603, 0.505471, 0.515447, 0.531043
dd 0.553104, 0.582935, 0.622504, 0.674808
- dd -1.169440, -0.972568, -0.839350, -0.744536
dd -10.190008, -3.407609, -2.057781, -1.484165
+ dd -1.169440, -0.972568, -0.839350, -0.744536
dd 0.502419, 0.522499, 0.566944, 0.646822
dd 0.788155, 1.060678, 1.722447, 5.101149
dd 0.509796, 0.601345, 0.899976, 2.562916
+ dd 0.509796, 0.601345, 0.899976, 2.562916
dd 1.000000, 1.000000, 1.306563, 0.541196
+ dd 1.000000, 1.000000, 1.306563, 0.541196
+ dd 1.000000, 0.707107, 1.000000, -0.707107
dd 1.000000, 0.707107, 1.000000, -0.707107
-ps_p1p1m1m1: dd 0, 0, 0x80000000, 0x80000000
+ps_p1p1m1m1: dd 0, 0, 0x80000000, 0x80000000, 0, 0, 0x80000000, 0x80000000
-%macro BUTTERFLY 4
+%macro BUTTERFLY_SSE 4
movaps %4, %1
subps %1, %2
addps %2, %4
mulps %1, %3
%endmacro
-%macro BUTTERFLY0 5
+%macro BUTTERFLY_AVX 4
+ vsubps %4, %1, %2
+ vaddps %2, %2, %1
+ vmulps %1, %4, %3
+%endmacro
+
+%macro BUTTERFLY0_SSE 5
movaps %4, %1
shufps %1, %1, %5
xorps %4, %2
@@ -52,6 +62,13 @@ ps_p1p1m1m1: dd 0, 0, 0x80000000, 0x80000000
mulps %1, %3
%endmacro
+%macro BUTTERFLY0_AVX 5
+ vshufps %4, %1, %1, %5
+ vxorps %1, %1, %2
+ vaddps %4, %4, %1
+ vmulps %1, %4, %3
+%endmacro
+
%macro BUTTERFLY2 4
BUTTERFLY0 %1, %2, %3, %4, 0x1b
%endmacro
@@ -60,8 +77,199 @@ ps_p1p1m1m1: dd 0, 0, 0x80000000, 0x80000000
BUTTERFLY0 %1, %2, %3, %4, 0xb1
%endmacro
-INIT_XMM
+%macro PASS6_AND_PERMUTE 0
+ mov tmpd, [outq+4]
+ movss m7, [outq+72]
+ addss m7, [outq+76]
+ movss m3, [outq+56]
+ addss m3, [outq+60]
+ addss m4, m3
+ movss m2, [outq+52]
+ addss m2, m3
+ movss m3, [outq+104]
+ addss m3, [outq+108]
+ addss m1, m3
+ addss m5, m4
+ movss [outq+ 16], m1
+ movss m1, [outq+100]
+ addss m1, m3
+ movss m3, [outq+40]
+ movss [outq+ 48], m1
+ addss m3, [outq+44]
+ movss m1, [outq+100]
+ addss m4, m3
+ addss m3, m2
+ addss m1, [outq+108]
+ movss [outq+ 40], m3
+ addss m2, [outq+36]
+ movss m3, [outq+8]
+ movss [outq+ 56], m2
+ addss m3, [outq+12]
+ movss [outq+ 32], m3
+ movss m3, [outq+80]
+ movss [outq+ 8], m5
+ movss [outq+ 80], m1
+ movss m2, [outq+52]
+ movss m5, [outq+120]
+ addss m5, [outq+124]
+ movss m1, [outq+64]
+ addss m2, [outq+60]
+ addss m0, m5
+ addss m5, [outq+116]
+ mov [outq+64], tmpd
+ addss m6, m0
+ addss m1, m6
+ mov tmpd, [outq+12]
+ mov [outq+ 96], tmpd
+ movss [outq+ 4], m1
+ movss m1, [outq+24]
+ movss [outq+ 24], m4
+ movss m4, [outq+88]
+ addss m4, [outq+92]
+ addss m3, m4
+ addss m4, [outq+84]
+ mov tmpd, [outq+108]
+ addss m1, [outq+28]
+ addss m0, m1
+ addss m1, m5
+ addss m6, m3
+ addss m3, m0
+ addss m0, m7
+ addss m5, [outq+20]
+ addss m7, m1
+ movss [outq+ 12], m6
+ mov [outq+112], tmpd
+ movss m6, [outq+28]
+ movss [outq+ 28], m0
+ movss m0, [outq+36]
+ movss [outq+ 36], m7
+ addss m1, m4
+ movss m7, [outq+116]
+ addss m0, m2
+ addss m7, [outq+124]
+ movss [outq+ 72], m0
+ movss m0, [outq+44]
+ addss m2, m0
+ movss [outq+ 44], m1
+ movss [outq+ 88], m2
+ addss m0, [outq+60]
+ mov tmpd, [outq+60]
+ mov [outq+120], tmpd
+ movss [outq+104], m0
+ addss m4, m5
+ addss m5, [outq+68]
+ movss [outq+52], m4
+ movss [outq+60], m5
+ movss m4, [outq+68]
+ movss m5, [outq+20]
+ movss [outq+ 20], m3
+ addss m5, m7
+ addss m7, m6
+ addss m4, m5
+ movss m2, [outq+84]
+ addss m2, [outq+92]
+ addss m5, m2
+ movss [outq+ 68], m4
+ addss m2, m7
+ movss m4, [outq+76]
+ movss [outq+ 84], m2
+ movss [outq+ 76], m5
+ addss m7, m4
+ addss m6, [outq+124]
+ addss m4, m6
+ addss m6, [outq+92]
+ movss [outq+100], m4
+ movss [outq+108], m6
+ movss m6, [outq+92]
+ movss [outq+92], m7
+ addss m6, [outq+124]
+ movss [outq+116], m6
+%endmacro
+
+%define BUTTERFLY BUTTERFLY_AVX
+%define BUTTERFLY0 BUTTERFLY0_AVX
+
+INIT_YMM
section .text align=16
+%ifdef HAVE_AVX
+; void ff_dct32_float_avx(FFTSample *out, const FFTSample *in)
+cglobal dct32_float_avx, 2,3,8, out, in, tmp
+ ; pass 1
+ vmovaps m4, [inq+0]
+ vinsertf128 m5, m5, [inq+96], 1
+ vinsertf128 m5, m5, [inq+112], 0
+ vshufps m5, m5, m5, 0x1b
+ BUTTERFLY m4, m5, [ps_cos_vec], m6
+
+ vmovaps m2, [inq+64]
+ vinsertf128 m6, m6, [inq+32], 1
+ vinsertf128 m6, m6, [inq+48], 0
+ vshufps m6, m6, m6, 0x1b
+ BUTTERFLY m2, m6, [ps_cos_vec+32], m0
+
+ ; pass 2
+
+ BUTTERFLY m5, m6, [ps_cos_vec+64], m0
+ BUTTERFLY m4, m2, [ps_cos_vec+64], m7
+
+
+ ; pass 3
+ vperm2f128 m3, m6, m4, 0x31
+ vperm2f128 m1, m6, m4, 0x20
+ vshufps m3, m3, m3, 0x1b
+
+ BUTTERFLY m1, m3, [ps_cos_vec+96], m6
+
+
+ vperm2f128 m4, m5, m2, 0x20
+ vperm2f128 m5, m5, m2, 0x31
+ vshufps m5, m5, m5, 0x1b
+
+ BUTTERFLY m4, m5, [ps_cos_vec+96], m6
+
+ ; pass 4
+ vmovaps m6, [ps_p1p1m1m1+0]
+ vmovaps m2, [ps_cos_vec+128]
+
+ BUTTERFLY2 m5, m6, m2, m7
+ BUTTERFLY2 m4, m6, m2, m7
+ BUTTERFLY2 m1, m6, m2, m7
+ BUTTERFLY2 m3, m6, m2, m7
+
+
+ ; pass 5
+ vshufps m6, m6, m6, 0xcc
+ vmovaps m2, [ps_cos_vec+160]
+
+ BUTTERFLY3 m5, m6, m2, m7
+ BUTTERFLY3 m4, m6, m2, m7
+ BUTTERFLY3 m1, m6, m2, m7
+ BUTTERFLY3 m3, m6, m2, m7
+
+ vperm2f128 m6, m3, m3, 0x31
+ vmovaps [outq], m3
+
+ vextractf128 [outq+64], m5, 1
+ vextractf128 [outq+32], m5, 0
+
+ vextractf128 [outq+80], m4, 1
+ vextractf128 [outq+48], m4, 0
+
+ vperm2f128 m0, m1, m1, 0x31
+ vmovaps [outq+96], m1
+
+ vzeroupper
+
+ ; pass 6, no SIMD...
+INIT_XMM
+ PASS6_AND_PERMUTE
+ RET
+%endif
+
+%define BUTTERFLY BUTTERFLY_SSE
+%define BUTTERFLY0 BUTTERFLY0_SSE
+
+INIT_XMM
; void ff_dct32_float_sse(FFTSample *out, const FFTSample *in)
cglobal dct32_float_sse, 2,3,8, out, in, tmp
; pass 1
@@ -74,8 +282,7 @@ cglobal dct32_float_sse, 2,3,8, out, in, tmp
movaps m7, [inq+64]
movaps m4, [inq+48]
shufps m4, m4, 0x1b
- BUTTERFLY m7, m4, [ps_cos_vec+48], m3
-
+ BUTTERFLY m7, m4, [ps_cos_vec+32], m3
; pass 2
movaps m2, [ps_cos_vec+64]
@@ -92,7 +299,7 @@ cglobal dct32_float_sse, 2,3,8, out, in, tmp
movaps m4, [inq+80]
movaps m5, [inq+32]
shufps m5, m5, 0x1b
- BUTTERFLY m4, m5, [ps_cos_vec+32], m3
+ BUTTERFLY m4, m5, [ps_cos_vec+48], m3
; pass 2
BUTTERFLY m0, m7, m2, m3
@@ -123,7 +330,7 @@ cglobal dct32_float_sse, 2,3,8, out, in, tmp
; pass 4
movaps m3, [ps_p1p1m1m1+0]
- movaps m2, [ps_cos_vec+112]
+ movaps m2, [ps_cos_vec+128]
BUTTERFLY2 m5, m3, m2, m1
@@ -148,7 +355,7 @@ cglobal dct32_float_sse, 2,3,8, out, in, tmp
BUTTERFLY2 m0, m3, m2, m1
; pass 5
- movaps m2, [ps_cos_vec+128]
+ movaps m2, [ps_cos_vec+160]
shufps m3, m3, 0xcc
BUTTERFLY3 m5, m3, m2, m1
@@ -180,110 +387,5 @@ cglobal dct32_float_sse, 2,3,8, out, in, tmp
; pass 6, no SIMD...
- mov tmpd, [outq+4]
- movss m7, [outq+72]
- addss m7, [outq+76]
- movss m3, [outq+56]
- addss m3, [outq+60]
- addss m4, m3
- movss m2, [outq+52]
- addss m2, m3
- movss m3, [outq+104]
- addss m3, [outq+108]
- addss m1, m3
- addss m5, m4
- movss [outq+ 16], m1
- movss m1, [outq+100]
- addss m1, m3
- movss m3, [outq+40]
- movss [outq+ 48], m1
- addss m3, [outq+44]
- movss m1, [outq+100]
- addss m4, m3
- addss m3, m2
- addss m1, [outq+108]
- movss [outq+ 40], m3
- addss m2, [outq+36]
- movss m3, [outq+8]
- movss [outq+ 56], m2
- addss m3, [outq+12]
- movss [outq+ 32], m3
- movss m3, [outq+80]
- movss [outq+ 8], m5
- movss [outq+ 80], m1
- movss m2, [outq+52]
- movss m5, [outq+120]
- addss m5, [outq+124]
- movss m1, [outq+64]
- addss m2, [outq+60]
- addss m0, m5
- addss m5, [outq+116]
- mov [outq+64], tmpd
- addss m6, m0
- addss m1, m6
- mov tmpd, [outq+12]
- mov [outq+ 96], tmpd
- movss [outq+ 4], m1
- movss m1, [outq+24]
- movss [outq+ 24], m4
- movss m4, [outq+88]
- addss m4, [outq+92]
- addss m3, m4
- addss m4, [outq+84]
- mov tmpd, [outq+108]
- addss m1, [outq+28]
- addss m0, m1
- addss m1, m5
- addss m6, m3
- addss m3, m0
- addss m0, m7
- addss m5, [outq+20]
- addss m7, m1
- movss [outq+ 12], m6
- mov [outq+112], tmpd
- movss m6, [outq+28]
- movss [outq+ 28], m0
- movss m0, [outq+36]
- movss [outq+ 36], m7
- addss m1, m4
- movss m7, [outq+116]
- addss m0, m2
- addss m7, [outq+124]
- movss [outq+ 72], m0
- movss m0, [outq+44]
- addss m2, m0
- movss [outq+ 44], m1
- movss [outq+ 88], m2
- addss m0, [outq+60]
- mov tmpd, [outq+60]
- mov [outq+120], tmpd
- movss [outq+104], m0
- addss m4, m5
- addss m5, [outq+68]
- movss [outq+52], m4
- movss [outq+60], m5
- movss m4, [outq+68]
- movss m5, [outq+20]
- movss [outq+ 20], m3
- addss m5, m7
- addss m7, m6
- addss m4, m5
- movss m2, [outq+84]
- addss m2, [outq+92]
- addss m5, m2
- movss [outq+ 68], m4
- addss m2, m7
- movss m4, [outq+76]
- movss [outq+ 84], m2
- movss [outq+ 76], m5
- addss m7, m4
- addss m6, [outq+124]
- addss m4, m6
- addss m6, [outq+92]
- movss [outq+100], m4
- movss [outq+108], m6
- movss m6, [outq+92]
- movss [outq+92], m7
- addss m6, [outq+124]
- movss [outq+116], m6
+ PASS6_AND_PERMUTE
RET
diff --git a/libavcodec/x86/fft.c b/libavcodec/x86/fft.c
index b29412c..8eef421 100644
--- a/libavcodec/x86/fft.c
+++ b/libavcodec/x86/fft.c
@@ -57,7 +57,9 @@ av_cold void ff_fft_init_mmx(FFTContext *s)
av_cold void ff_dct_init_mmx(DCTContext *s)
{
int has_vectors = av_get_cpu_flags();
- if (has_vectors & AV_CPU_FLAG_SSE && HAVE_SSE)
+ if (has_vectors & AV_CPU_FLAG_AVX && HAVE_AVX)
+ s->dct32 = ff_dct32_float_avx;
+ else if (has_vectors & AV_CPU_FLAG_SSE && HAVE_SSE)
s->dct32 = ff_dct32_float_sse;
}
#endif
diff --git a/libavcodec/x86/fft.h b/libavcodec/x86/fft.h
index c637905..0ade2b2 100644
--- a/libavcodec/x86/fft.h
+++ b/libavcodec/x86/fft.h
@@ -35,5 +35,6 @@ void ff_imdct_calc_sse(FFTContext *s, FFTSample *output, const FFTSample *input)
void ff_imdct_half_sse(FFTContext *s, FFTSample *output, const FFTSample *input);
void ff_imdct_half_avx(FFTContext *s, FFTSample *output, const FFTSample *input);
void ff_dct32_float_sse(FFTSample *out, const FFTSample *in);
+void ff_dct32_float_avx(FFTSample *out, const FFTSample *in);
#endif /* AVCODEC_X86_FFT_H */
More information about the ffmpeg-cvslog
mailing list