From 4dcc4f8eaa23dc3ff85f6b4eed58d719e50cbe92 Mon Sep 17 00:00:00 2001 From: Vitor Sessak Date: Tue, 6 Jul 2010 16:58:54 +0000 Subject: SSE optimized 32-point DCT Originally committed as revision 24077 to svn://svn.ffmpeg.org/ffmpeg/trunk --- libavcodec/dct.c | 2 + libavcodec/fft.h | 1 + libavcodec/x86/fft.c | 8 ++ libavcodec/x86/fft.h | 1 + libavcodec/x86/fft_sse.c | 266 +++++++++++++++++++++++++++++++++++++++++++++++ 5 files changed, 278 insertions(+) diff --git a/libavcodec/dct.c b/libavcodec/dct.c index 682fe99f80..dab94c30a0 100644 --- a/libavcodec/dct.c +++ b/libavcodec/dct.c @@ -30,6 +30,7 @@ #include #include "libavutil/mathematics.h" #include "fft.h" +#include "x86/fft.h" #define DCT32_FLOAT #include "dct32.c" @@ -213,6 +214,7 @@ av_cold int ff_dct_init(DCTContext *s, int nbits, enum DCTTransformType inverse) s->dct_calc = dct32_func; s->dct32 = dct32; + if (HAVE_MMX) ff_dct_init_mmx(s); return 0; } diff --git a/libavcodec/fft.h b/libavcodec/fft.h index 8930c93ec7..eb6714fe95 100644 --- a/libavcodec/fft.h +++ b/libavcodec/fft.h @@ -112,6 +112,7 @@ void ff_fft_calc_c(FFTContext *s, FFTComplex *z); void ff_fft_init_altivec(FFTContext *s); void ff_fft_init_mmx(FFTContext *s); void ff_fft_init_arm(FFTContext *s); +void ff_dct_init_mmx(DCTContext *s); /** * Do the permutation needed BEFORE calling ff_fft_calc(). diff --git a/libavcodec/x86/fft.c b/libavcodec/x86/fft.c index 2c46c6361b..6f641b6f02 100644 --- a/libavcodec/x86/fft.c +++ b/libavcodec/x86/fft.c @@ -42,3 +42,11 @@ av_cold void ff_fft_init_mmx(FFTContext *s) } #endif } + +av_cold void ff_dct_init_mmx(DCTContext *s) +{ + int has_vectors = mm_support(); + if (has_vectors & FF_MM_SSE && HAVE_SSE) + s->dct32 = ff_dct32_float_sse; +} + diff --git a/libavcodec/x86/fft.h b/libavcodec/x86/fft.h index 7ef5839141..2affacb62b 100644 --- a/libavcodec/x86/fft.h +++ b/libavcodec/x86/fft.h @@ -32,5 +32,6 @@ void ff_imdct_calc_3dn2(FFTContext *s, FFTSample *output, const FFTSample *input void ff_imdct_half_3dn2(FFTContext *s, FFTSample *output, const FFTSample *input); 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_dct32_float_sse(FFTSample *out, const FFTSample *in); #endif diff --git a/libavcodec/x86/fft_sse.c b/libavcodec/x86/fft_sse.c index a4cce69db1..42e5575984 100644 --- a/libavcodec/x86/fft_sse.c +++ b/libavcodec/x86/fft_sse.c @@ -20,6 +20,7 @@ */ #include "libavutil/x86_cpu.h" +#include "libavutil/common.h" #include "libavcodec/dsputil.h" #include "fft.h" @@ -201,3 +202,268 @@ void ff_imdct_calc_sse(FFTContext *s, FFTSample *output, const FFTSample *input) ); } +DECLARE_ALIGNED(16, static const float, b1)[] = { + 0.500603, 0.505471, 0.515447, 0.531043, + 0.553104, 0.582935, 0.622504, 0.674808, + -1.169440, -0.972568, -0.839350, -0.744536, + -10.190008, -3.407609, -2.057781, -1.484165, + 0.502419, 0.522499, 0.566944, 0.646822, + 0.788155, 1.060678, 1.722447, 5.101149, + 0.509796, 0.601345, 0.899976, 2.562916, + 1.000000, 1.000000, 1.306563, 0.541196, + 1.000000, 0.707107, 1.000000, -0.707107 +}; + +DECLARE_ALIGNED(16, static const int32_t, smask)[4] = { + 0, 0, 0x80000000, 0x80000000 +}; + +/* butterfly operator */ +#define BUTTERFLY(a,b,c,tmp) \ + "movaps %%" #a ", %%" #tmp " \n\t" \ + "subps %%" #b ", %%" #a " \n\t" \ + "addps %%" #tmp ", %%" #b " \n\t" \ + "mulps " #c ", %%" #a " \n\t" + +///* Same as BUTTERFLY when vectors a and b overlap */ +#define BUTTERFLY0(val, mask, cos, tmp, shuf) \ + "movaps %%" #val ", %%" #tmp " \n\t" \ + "shufps " #shuf ", %%" #val ",%%" #val " \n\t" \ + "xorps %%" #mask ", %%" #tmp " \n\t" /* flip signs */ \ + "addps %%" #tmp ", %%" #val " \n\t" \ + "mulps %%" #cos ", %%" #val " \n\t" + +#define BUTTERFLY2(val, mask, cos, tmp) BUTTERFLY0(val, mask, cos, tmp, $0x1b) +#define BUTTERFLY3(val, mask, cos, tmp) BUTTERFLY0(val, mask, cos, tmp, $0xb1) + +void ff_dct32_float_sse(FFTSample *out, const FFTSample *in) +{ + int32_t tmp1 = 0; + __asm__ volatile( + /* pass 1 */ + + "movaps (%4), %%xmm0 \n\t" + "movaps 112(%4), %%xmm1 \n\t" + "shufps $0x1b, %%xmm1, %%xmm1 \n\t" + BUTTERFLY(xmm0, xmm1, (%2), xmm3) + + "movaps 64(%4), %%xmm7 \n\t" + "movaps 48(%4), %%xmm4 \n\t" + "shufps $0x1b, %%xmm4, %%xmm4 \n\t" + BUTTERFLY(xmm7, xmm4, 48(%2), xmm3) + + + /* pass 2 */ + "movaps 64(%2), %%xmm2 \n\t" + BUTTERFLY(xmm1, xmm4, %%xmm2, xmm3) + "movaps %%xmm1, 48(%1) \n\t" + "movaps %%xmm4, (%1) \n\t" + + /* pass 1 */ + "movaps 16(%4), %%xmm1 \n\t" + "movaps 96(%4), %%xmm6 \n\t" + "shufps $0x1b, %%xmm6, %%xmm6 \n\t" + BUTTERFLY(xmm1, xmm6, 16(%2), xmm3) + + "movaps 80(%4), %%xmm4 \n\t" + "movaps 32(%4), %%xmm5 \n\t" + "shufps $0x1b, %%xmm5, %%xmm5 \n\t" + BUTTERFLY(xmm4, xmm5, 32(%2), xmm3) + + /* pass 2 */ + BUTTERFLY(xmm0, xmm7, %%xmm2, xmm3) + + "movaps 80(%2), %%xmm2 \n\t" + BUTTERFLY(xmm6, xmm5, %%xmm2, xmm3) + + BUTTERFLY(xmm1, xmm4, %%xmm2, xmm3) + + /* pass 3 */ + "movaps 96(%2), %%xmm2 \n\t" + "shufps $0x1b, %%xmm1, %%xmm1 \n\t" + BUTTERFLY(xmm0, xmm1, %%xmm2, xmm3) + "movaps %%xmm0, 112(%1) \n\t" + "movaps %%xmm1, 96(%1) \n\t" + + "movaps 0(%1), %%xmm0 \n\t" + "shufps $0x1b, %%xmm5, %%xmm5 \n\t" + BUTTERFLY(xmm0, xmm5, %%xmm2, xmm3) + + "movaps 48(%1), %%xmm1 \n\t" + "shufps $0x1b, %%xmm6, %%xmm6 \n\t" + BUTTERFLY(xmm1, xmm6, %%xmm2, xmm3) + "movaps %%xmm1, 48(%1) \n\t" + + "shufps $0x1b, %%xmm4, %%xmm4 \n\t" + BUTTERFLY(xmm7, xmm4, %%xmm2, xmm3) + + /* pass 4 */ + "movaps (%3), %%xmm3 \n\t" + "movaps 112(%2), %%xmm2 \n\t" + + BUTTERFLY2(xmm5, xmm3, xmm2, xmm1) + + BUTTERFLY2(xmm0, xmm3, xmm2, xmm1) + "movaps %%xmm0, 16(%1) \n\t" + + BUTTERFLY2(xmm6, xmm3, xmm2, xmm1) + "movaps %%xmm6, 32(%1) \n\t" + + "movaps 48(%1), %%xmm0 \n\t" + BUTTERFLY2(xmm0, xmm3, xmm2, xmm1) + "movaps %%xmm0, 48(%1) \n\t" + + BUTTERFLY2(xmm4, xmm3, xmm2, xmm1) + + BUTTERFLY2(xmm7, xmm3, xmm2, xmm1) + + "movaps 96(%1), %%xmm6 \n\t" + BUTTERFLY2(xmm6, xmm3, xmm2, xmm1) + + "movaps 112(%1), %%xmm0 \n\t" + BUTTERFLY2(xmm0, xmm3, xmm2, xmm1) + + /* pass 5 */ + "movaps 128(%2), %%xmm2 \n\t" + "shufps $0xCC, %%xmm3,%%xmm3 \n\t" + + BUTTERFLY3(xmm5, xmm3, xmm2, xmm1) + "movaps %%xmm5, (%1) \n\t" + + "movaps 16(%1), %%xmm1 \n\t" + BUTTERFLY3(xmm1, xmm3, xmm2, xmm5) + "movaps %%xmm1, 16(%1) \n\t" + + BUTTERFLY3(xmm4, xmm3, xmm2, xmm5) + "movaps %%xmm4, 64(%1) \n\t" + + BUTTERFLY3(xmm7, xmm3, xmm2, xmm5) + "movaps %%xmm7, 80(%1) \n\t" + + "movaps 32(%1), %%xmm5 \n\t" + BUTTERFLY3(xmm5, xmm3, xmm2, xmm7) + "movaps %%xmm5, 32(%1) \n\t" + + "movaps 48(%1), %%xmm4 \n\t" + BUTTERFLY3(xmm4, xmm3, xmm2, xmm7) + "movaps %%xmm4, 48(%1) \n\t" + + BUTTERFLY3(xmm6, xmm3, xmm2, xmm7) + "movaps %%xmm6, 96(%1) \n\t" + + BUTTERFLY3(xmm0, xmm3, xmm2, xmm7) + "movaps %%xmm0, 112(%1) \n\t" + + + /* pass 6, no SIMD... */ + "movss 56(%1), %%xmm3 \n\t" + "movl 4(%1), %0 \n\t" + "addss 60(%1), %%xmm3 \n\t" + "movss 72(%1), %%xmm7 \n\t" + "addss %%xmm3, %%xmm4 \n\t" + "movss 52(%1), %%xmm2 \n\t" + "addss %%xmm3, %%xmm2 \n\t" + "movss 24(%1), %%xmm3 \n\t" + "addss 28(%1), %%xmm3 \n\t" + "addss 76(%1), %%xmm7 \n\t" + "addss %%xmm3, %%xmm1 \n\t" + "addss %%xmm4, %%xmm5 \n\t" + "movss %%xmm1, 16(%1) \n\t" + "movss 20(%1), %%xmm1 \n\t" + "addss %%xmm3, %%xmm1 \n\t" + "movss 40(%1), %%xmm3 \n\t" + "movss %%xmm1, 48(%1) \n\t" + "addss 44(%1), %%xmm3 \n\t" + "movss 20(%1), %%xmm1 \n\t" + "addss %%xmm3, %%xmm4 \n\t" + "addss %%xmm2, %%xmm3 \n\t" + "addss 28(%1), %%xmm1 \n\t" + "movss %%xmm3, 40(%1) \n\t" + "addss 36(%1), %%xmm2 \n\t" + "movss 8(%1), %%xmm3 \n\t" + "movss %%xmm2, 56(%1) \n\t" + "addss 12(%1), %%xmm3 \n\t" + "movss %%xmm5, 8(%1) \n\t" + "movss %%xmm3, 32(%1) \n\t" + "movss 52(%1), %%xmm2 \n\t" + "movss 80(%1), %%xmm3 \n\t" + "movss 120(%1), %%xmm5 \n\t" + "movss %%xmm1, 80(%1) \n\t" + "movss %%xmm4, 24(%1) \n\t" + "addss 124(%1), %%xmm5 \n\t" + "movss 64(%1), %%xmm1 \n\t" + "addss 60(%1), %%xmm2 \n\t" + "addss %%xmm5, %%xmm0 \n\t" + "addss 116(%1), %%xmm5 \n\t" + "movl %0, 64(%1) \n\t" + "addss %%xmm0, %%xmm6 \n\t" + "addss %%xmm6, %%xmm1 \n\t" + "movl 12(%1), %0 \n\t" + "movss %%xmm1, 4(%1) \n\t" + "movss 88(%1), %%xmm1 \n\t" + "movl %0, 96(%1) \n\t" + "addss 92(%1), %%xmm1 \n\t" + "movss 104(%1), %%xmm4 \n\t" + "movl 28(%1), %0 \n\t" + "addss 108(%1), %%xmm4 \n\t" + "addss %%xmm4, %%xmm0 \n\t" + "addss %%xmm1, %%xmm3 \n\t" + "addss 84(%1), %%xmm1 \n\t" + "addss %%xmm5, %%xmm4 \n\t" + "addss %%xmm3, %%xmm6 \n\t" + "addss %%xmm0, %%xmm3 \n\t" + "addss %%xmm7, %%xmm0 \n\t" + "addss 100(%1), %%xmm5 \n\t" + "addss %%xmm4, %%xmm7 \n\t" + "movl %0, 112(%1) \n\t" + "movss %%xmm0, 28(%1) \n\t" + "movss 36(%1), %%xmm0 \n\t" + "movss %%xmm7, 36(%1) \n\t" + "addss %%xmm1, %%xmm4 \n\t" + "movss 116(%1), %%xmm7 \n\t" + "addss %%xmm2, %%xmm0 \n\t" + "addss 124(%1), %%xmm7 \n\t" + "movss %%xmm0, 72(%1) \n\t" + "movss 44(%1), %%xmm0 \n\t" + "movss %%xmm6, 12(%1) \n\t" + "movss %%xmm3, 20(%1) \n\t" + "addss %%xmm0, %%xmm2 \n\t" + "movss %%xmm4, 44(%1) \n\t" + "movss %%xmm2, 88(%1) \n\t" + "addss 60(%1), %%xmm0 \n\t" + "movl 60(%1), %0 \n\t" + "movl %0, 120(%1) \n\t" + "movss %%xmm0, 104(%1) \n\t" + "addss %%xmm5, %%xmm1 \n\t" + "addss 68(%1), %%xmm5 \n\t" + "movss %%xmm1, 52(%1) \n\t" + "movss %%xmm5, 60(%1) \n\t" + "movss 68(%1), %%xmm1 \n\t" + "movss 100(%1), %%xmm5 \n\t" + "addss %%xmm7, %%xmm5 \n\t" + "addss 108(%1), %%xmm7 \n\t" + "addss %%xmm5, %%xmm1 \n\t" + "movss 84(%1), %%xmm2 \n\t" + "addss 92(%1), %%xmm2 \n\t" + "addss %%xmm2, %%xmm5 \n\t" + "movss %%xmm1, 68(%1) \n\t" + "addss %%xmm7, %%xmm2 \n\t" + "movss 76(%1), %%xmm1 \n\t" + "movss %%xmm2, 84(%1) \n\t" + "movss %%xmm5, 76(%1) \n\t" + "movss 108(%1), %%xmm2 \n\t" + "addss %%xmm1, %%xmm7 \n\t" + "addss 124(%1), %%xmm2 \n\t" + "addss %%xmm2, %%xmm1 \n\t" + "addss 92(%1), %%xmm2 \n\t" + "movss %%xmm1, 100(%1) \n\t" + "movss %%xmm2, 108(%1) \n\t" + "movss 92(%1), %%xmm2 \n\t" + "movss %%xmm7, 92(%1) \n\t" + "addss 124(%1), %%xmm2 \n\t" + "movss %%xmm2, 116(%1) \n\t" + :"+&r"(tmp1) + :"r"(out), "r"(b1), "r"(smask), "r"(in) + :"memory" + ); +} -- cgit v1.2.3