[Ffmpeg-devel] [PATCH] SSE counterpart of ff_imdct_calc_3dn2
Zuxy Meng
zuxy.meng
Wed Aug 23 12:07:06 CEST 2006
Hi,
2006/8/23, Michael Niedermayer <michaelni at gmx.at>:
> Hi
>
> i think that can be done faster with:
>
> "movaps %0, %%xmm0 \n\t" // xmm0 = 0 1 2 3
> "movaps %1, %%xmm1 \n\t" // xmm1 = 4 5 6 7
> ::"m"(z[k]), "m"(z[-2 - k])
> );
> asm (
> "xorps %%xmm7, %%xmm0 \n\t" // xmm1 =-0-1-2-3
> "movaps %%xmm0, %%xmm2 \n\t" // xmm2 =-0-1-2-3
> "shufps $141,%%xmm1, %%xmm0 \n\t" // xmm0 =-1-3 4 6
> "shufps $216,%%xmm1, %%xmm2 \n\t" // xmm2 =-0-2 5 7
> "shufps $156,%%xmm0, %%xmm0 \n\t" // xmm0 =-1 6-3 4 !
> "shufps $156,%%xmm2, %%xmm2 \n\t" // xmm2 =-0 7-2 5 !
> "movaps %%xmm0, %0 \n\t"
> "movaps %%xmm2, %2 \n\t"
> "shufps $27, %%xmm0, %%xmm0 \n\t" // xmm1 = 4 -3 6 -1
> "xorps %%xmm7, %%xmm0 \n\t" // xmm1 = -4 3 -6 1 !
> "shufps $27, %%xmm2, %%xmm2 \n\t" // xmm3 = 5 -2 7 -0 !
> "movaps %%xmm0, %1 \n\t"
> "movaps %%xmm2, %3 \n\t"
>
> or
>
> "movlps %0, %%xmm0 \n\t" // xmm0 = 0 1 X X
> "movlps 8+%0, %%xmm1 \n\t" // xmm1 = 2 3 X X
> "movhps 8+%1, %%xmm0 \n\t" // xmm0 = 0 1 6 7
> "movhps %1, %%xmm1 \n\t" // xmm1 = 2 3 4 5
> ::"m"(z[k]), "m"(z[-2 - k])
> );
> asm (
> "xorps %%xmm7, %%xmm0 \n\t" // xmm1 =-0-1 6 7
> "xorps %%xmm7, %%xmm1 \n\t" // xmm1 =-2-3 4 5
> "movaps %%xmm0, %%xmm2 \n\t" // xmm2 =-0-1 6 7
> "shufps $??,%%xmm1, %%xmm0 \n\t" // xmm0 =-1 6-3 4 !
> "shufps $??,%%xmm1, %%xmm2 \n\t" // xmm2 =-0 7-2 5 !
> "movaps %%xmm0, %0 \n\t"
> "movaps %%xmm2, %2 \n\t"
> "shufps $27, %%xmm0, %%xmm0 \n\t" // xmm1 = 4 -3 6 -1
> "xorps %%xmm6, %%xmm0 \n\t" // xmm1 = -4 3 -6 1 !
> "shufps $27, %%xmm2, %%xmm2 \n\t" // xmm3 = 5 -2 7 -0 !
> "movaps %%xmm0, %1 \n\t"
> "movaps %%xmm2, %3 \n\t"
>
> both untested, so maybe ive missed something silly
>
> and of course should the 2 asms be merged, and the whole loop be done in
> asm
See if this is better, with fewer instructions. BTW I also checked the
code against gcc2.95.3-8 and seems that the single asm block for the
last loop works all right.
--
Zuxy
Beauty is truth,
While truth is beauty.
PGP KeyID: E8555ED6
-------------- next part --------------
Index: libavcodec/i386/fft_sse.c
===================================================================
--- libavcodec/i386/fft_sse.c ???????? 6051??
+++ libavcodec/i386/fft_sse.c ????????????
@@ -27,6 +27,12 @@
static const int p1p1m1m1[4] __attribute__((aligned(16))) =
{ 0, 0, 1 << 31, 1 << 31 };
+static const int p1m1p1m1[4] __attribute__((aligned(16))) =
+ { 0, 1 << 31, 0, 1 << 31 };
+
+static const int m1m1m1m1[4] __attribute__((aligned(16))) =
+ { 1 << 31, 1 << 31, 1 << 31, 1 << 31 };
+
#if 0
static void print_v4sf(const char *str, __m128 a)
{
@@ -118,3 +124,119 @@
} while (nblocks != 0);
}
+void ff_imdct_calc_sse(MDCTContext *s, FFTSample *output,
+ const FFTSample *input, FFTSample *tmp)
+{
+ long k, n8, n4, n2, n;
+ const uint16_t *revtab = s->fft.revtab;
+ const FFTSample *tcos = s->tcos;
+ const FFTSample *tsin = s->tsin;
+ const FFTSample *in1, *in2;
+ FFTComplex *z = (FFTComplex *)tmp;
+
+ n = 1 << s->nbits;
+ n2 = n >> 1;
+ n4 = n >> 2;
+ n8 = n >> 3;
+
+ asm volatile ("movaps %0, %%xmm7\n\t"::"m"(*p1m1p1m1));
+
+ /* pre rotation */
+ in1 = input;
+ in2 = input + n2 - 4;
+
+ /* Complex multiplication
+ Two complex products per iteration, we could have 4 with 8 xmm
+ registers, 8 with 16 xmm registers.
+ Maybe we should unroll more.
+ */
+ for (k = 0; k < n4; k += 2) {
+ asm volatile (
+ "movaps %0, %%xmm0 \n\t" // xmm0 = r0 X r1 X : in2
+ "movaps %1, %%xmm3 \n\t" // xmm3 = X i1 X i0: in1
+ "movlps %2, %%xmm1 \n\t" // xmm1 = X X R1 R0: tcos
+ "movlps %3, %%xmm2 \n\t" // xmm2 = X X I1 I0: tsin
+ "shufps $95, %%xmm0, %%xmm0 \n\t" // xmm0 = r1 r1 r0 r0
+ "shufps $160,%%xmm3, %%xmm3 \n\t" // xmm3 = i1 i1 i0 i0
+ "unpcklps %%xmm2, %%xmm1 \n\t" // xmm1 = I1 R1 I0 R0
+ "movaps %%xmm1, %%xmm2 \n\t" // xmm2 = I1 R1 I0 R0
+ "xorps %%xmm7, %%xmm2 \n\t" // xmm2 = -I1 R1 -I0 R0
+ "mulps %%xmm1, %%xmm0 \n\t" // xmm0 = rI rR rI rR
+ "shufps $177,%%xmm2, %%xmm2 \n\t" // xmm2 = R1 -I1 R0 -I0
+ "mulps %%xmm2, %%xmm3 \n\t" // xmm3 = Ri -Ii Ri -Ii
+ "addps %%xmm3, %%xmm0 \n\t" // xmm0 = result
+ ::"m"(in2[-2*k]), "m"(in1[2*k]),
+ "m"(tcos[k]), "m"(tsin[k])
+ );
+ /* Should be in the same block, hack for gcc2.95 & gcc3 */
+ asm (
+ "movlps %%xmm0, %0 \n\t"
+ "movhps %%xmm0, %1 \n\t"
+ :"=m"(z[revtab[k]]), "=m"(z[revtab[k + 1]])
+ );
+ }
+
+ ff_fft_calc_sse(&s->fft, z);
+
+ /* Not currently needed, added for safety */
+ asm volatile ("movaps %0, %%xmm7\n\t"::"m"(*p1m1p1m1));
+
+ /* post rotation + reordering */
+ for (k = 0; k < n4; k += 2) {
+ asm (
+ "movaps %0, %%xmm0 \n\t" // xmm0 = i1 r1 i0 r0: z
+ "movlps %1, %%xmm1 \n\t" // xmm1 = X X R1 R0: tcos
+ "movaps %%xmm0, %%xmm3 \n\t" // xmm3 = i1 r1 i0 r0
+ "movlps %2, %%xmm2 \n\t" // xmm2 = X X I1 I0: tsin
+ "shufps $160,%%xmm0, %%xmm0 \n\t" // xmm0 = r1 r1 r0 r0
+ "shufps $245,%%xmm3, %%xmm3 \n\t" // xmm3 = i1 i1 i0 i0
+ "unpcklps %%xmm2, %%xmm1 \n\t" // xmm1 = I1 R1 I0 R0
+ "movaps %%xmm1, %%xmm2 \n\t" // xmm2 = I1 R1 I0 R0
+ "xorps %%xmm7, %%xmm2 \n\t" // xmm2 = -I1 R1 -I0 R0
+ "mulps %%xmm1, %%xmm0 \n\t" // xmm0 = rI rR rI rR
+ "shufps $177,%%xmm2, %%xmm2 \n\t" // xmm2 = R1 -I1 R0 -I0
+ "mulps %%xmm2, %%xmm3 \n\t" // xmm3 = Ri -Ii Ri -Ii
+ "addps %%xmm3, %%xmm0 \n\t" // xmm0 = result
+ "movaps %%xmm0, %0 \n\t"
+ :"+m"(z[k])
+ :"m"(tcos[k]), "m"(tsin[k])
+ );
+ }
+
+ z += n8;
+ asm volatile("movaps %0, %%xmm7 \n\t"::"m"(*m1m1m1m1));
+ for(k = 0; k < n8; k += 2) {
+ /*
+ Mnemonics:
+ 0 = z[k].re
+ 1 = z[k].im
+ 2 = z[k + 1].re
+ 3 = z[k + 1].im
+ 4 = z[-k - 2].re
+ 5 = z[-k - 2].im
+ 6 = z[-k - 1].re
+ 7 = z[-k - 1].im
+ */
+ asm (
+ "movaps %4, %%xmm0 \n\t" // xmm0 = 0 1 2 3
+ "movaps %5, %%xmm1 \n\t" // xmm1 = 4 5 6 7
+ "xorps %%xmm7, %%xmm0 \n\t" // xmm0 = -0 -1 -2 -3
+ "movaps %%xmm0, %%xmm2 \n\t" // xmm2 = -0 -1 -2 -3
+ "shufps $141,%%xmm1, %%xmm0 \n\t" // xmm0 = -1 -3 4 6
+ "shufps $216,%%xmm1, %%xmm2 \n\t" // xmm2 = -0 -2 5 7
+ "shufps $156,%%xmm0, %%xmm0 \n\t" // xmm0 = -1 6 -3 4 !
+ "shufps $156,%%xmm2, %%xmm2 \n\t" // xmm2 = -0 7 -2 5 !
+ "movaps %%xmm0, %0 \n\t"
+ "movaps %%xmm2, %2 \n\t"
+ "shufps $27, %%xmm0, %%xmm0 \n\t" // xmm0 = 4 -3 6 -1
+ "xorps %%xmm7, %%xmm0 \n\t" // xmm0 = -4 3 -6 1 !
+ "shufps $27, %%xmm2, %%xmm2 \n\t" // xmm2 = 5 -2 7 -0 !
+ "movaps %%xmm0, %1 \n\t"
+ "movaps %%xmm2, %3 \n\t"
+ :"=m"(output[2*k]), "=m"(output[n2 - 4 - 2*k]),
+ "=m"(output[n2 + 2*k]), "=m"(output[n - 4 - 2*k])
+ :"m"(z[k]), "m"(z[-2 - k])
+ );
+ }
+}
+
Index: libavcodec/dsputil.h
===================================================================
--- libavcodec/dsputil.h ???????? 6051??
+++ libavcodec/dsputil.h ????????????
@@ -655,6 +655,8 @@
const FFTSample *input, FFTSample *tmp);
void ff_imdct_calc_3dn2(MDCTContext *s, FFTSample *output,
const FFTSample *input, FFTSample *tmp);
+void ff_imdct_calc_sse(MDCTContext *s, FFTSample *output,
+ const FFTSample *input, FFTSample *tmp);
void ff_mdct_calc(MDCTContext *s, FFTSample *out,
const FFTSample *input, FFTSample *tmp);
void ff_mdct_end(MDCTContext *s);
Index: libavcodec/fft.c
===================================================================
--- libavcodec/fft.c ???????? 6051??
+++ libavcodec/fft.c ????????????
@@ -65,17 +65,18 @@
if (has_vectors) {
#if defined(HAVE_MMX)
- if (has_vectors & MM_3DNOWEXT)
+ if (has_vectors & MM_3DNOWEXT) {
+ /* 3DNowEx for K7/K8 */
s->imdct_calc = ff_imdct_calc_3dn2;
- if (has_vectors & MM_3DNOWEXT)
- /* 3DNowEx for K7/K8 */
s->fft_calc = ff_fft_calc_3dn2;
- else if (has_vectors & MM_3DNOW)
+ } else if (has_vectors & MM_3DNOW)
/* 3DNow! for K6-2/3 */
s->fft_calc = ff_fft_calc_3dn;
- else if (has_vectors & MM_SSE)
+ else if (has_vectors & MM_SSE) {
/* SSE for P3/P4 */
s->fft_calc = ff_fft_calc_sse;
+ s->imdct_calc = ff_imdct_calc_sse;
+ }
#else /* HAVE_MMX */
if (has_vectors & MM_ALTIVEC)
s->fft_calc = ff_fft_calc_altivec;
More information about the ffmpeg-devel
mailing list