[Ffmpeg-cvslog] CVS: ffmpeg/libavcodec/i386 dsputil_mmx.c, 1.114, 1.115 snowdsp_mmx.c, 1.1, 1.2
Corey Hickey CVS
corey
Tue Mar 21 22:51:10 CET 2006
Update of /cvsroot/ffmpeg/ffmpeg/libavcodec/i386
In directory mail:/var2/tmp/cvs-serv7467/libavcodec/i386
Modified Files:
dsputil_mmx.c snowdsp_mmx.c
Log Message:
snow mmx+sse2 optimizations, part 4
Patch by Robert Edele, yartrebo <<at>> earthlink <<dot>> net
Index: dsputil_mmx.c
===================================================================
RCS file: /cvsroot/ffmpeg/ffmpeg/libavcodec/i386/dsputil_mmx.c,v
retrieving revision 1.114
retrieving revision 1.115
diff -u -d -r1.114 -r1.115
--- dsputil_mmx.c 20 Mar 2006 22:27:59 -0000 1.114
+++ dsputil_mmx.c 21 Mar 2006 21:51:07 -0000 1.115
@@ -2564,8 +2564,12 @@
}
#endif
+#ifdef CONFIG_SNOW_ENCODER
+extern void ff_snow_horizontal_compose97i_sse2(DWTELEM *b, int width);
+extern void ff_snow_horizontal_compose97i_mmx(DWTELEM *b, int width);
extern void ff_snow_vertical_compose97i_sse2(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width);
extern void ff_snow_vertical_compose97i_mmx(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width);
+#endif
void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
{
@@ -2956,9 +2960,11 @@
#ifdef CONFIG_SNOW_ENCODER
if(mm_flags & MM_SSE2){
+ c->horizontal_compose97i = ff_snow_horizontal_compose97i_sse2;
c->vertical_compose97i = ff_snow_vertical_compose97i_sse2;
}
else{
+ c->horizontal_compose97i = ff_snow_horizontal_compose97i_mmx;
c->vertical_compose97i = ff_snow_vertical_compose97i_mmx;
}
#endif
Index: snowdsp_mmx.c
===================================================================
RCS file: /cvsroot/ffmpeg/ffmpeg/libavcodec/i386/snowdsp_mmx.c,v
retrieving revision 1.1
retrieving revision 1.2
diff -u -d -r1.1 -r1.2
--- snowdsp_mmx.c 20 Mar 2006 22:27:59 -0000 1.1
+++ snowdsp_mmx.c 21 Mar 2006 21:51:07 -0000 1.2
@@ -21,6 +21,412 @@
#include "../snow.h"
#include "mmx.h"
+static void always_inline snow_interleave_line_header(int * i, int width, DWTELEM * low, DWTELEM * high){
+ (*i) = (width) - 2;
+
+ if (width & 1){
+ low[(*i)+1] = low[((*i)+1)>>1];
+ (*i)--;
+ }
+}
+
+static void always_inline snow_horizontal_compose_lift_lead_out(int i, DWTELEM * dst, DWTELEM * src, DWTELEM * ref, int width, int w, int lift_high, int mul, int add, int shift){
+ for(; i<w; i++){
+ dst[i] = src[i] - ((mul * (ref[i] + ref[i + 1]) + add) >> shift);
+ }
+
+ if((width^lift_high)&1){
+ dst[w] = src[w] - ((mul * 2 * ref[w] + add) >> shift);
+ }
+}
+
+static void always_inline snow_horizontal_compose_liftS_lead_out(int i, DWTELEM * dst, DWTELEM * src, DWTELEM * ref, int width, int w){
+ for(; i<w; i++){
+ dst[i] = src[i] - (((-(ref[i] + ref[(i+1)])+W_BO) - 4 * src[i]) >> W_BS);
+ }
+
+ if(width&1){
+ dst[w] = src[w] - (((-2 * ref[w] + W_BO) - 4 * src[w]) >> W_BS);
+ }
+}
+
+void ff_snow_horizontal_compose97i_sse2(DWTELEM *b, int width){
+ const int w2= (width+1)>>1;
+ // SSE2 code runs faster with pointers aligned on a 32-byte boundary.
+ DWTELEM temp_buf[(width>>1) + 4];
+ DWTELEM * const temp = temp_buf + 4 - (((int)temp_buf & 0xF) >> 2);
+ const int w_l= (width>>1);
+ const int w_r= w2 - 1;
+ int i;
+
+ { // Lift 0
+ DWTELEM * const ref = b + w2 - 1;
+ DWTELEM b_0 = b[0]; //By allowing the first entry in b[0] to be calculated twice
+ // (the first time erroneously), we allow the SSE2 code to run an extra pass.
+ // The savings in code and time are well worth having to store this value and
+ // calculate b[0] correctly afterwards.
+
+ i = 0;
+ asm volatile(
+ "pcmpeqd %%xmm7, %%xmm7 \n\t"
+ "pslld $31, %%xmm7 \n\t"
+ "psrld $29, %%xmm7 \n\t"
+ ::);
+ for(; i<w_l-7; i+=8){
+ asm volatile(
+ "movdqu (%1), %%xmm1 \n\t"
+ "movdqu 16(%1), %%xmm5 \n\t"
+ "movdqu 4(%1), %%xmm2 \n\t"
+ "movdqu 20(%1), %%xmm6 \n\t"
+ "paddd %%xmm1, %%xmm2 \n\t"
+ "paddd %%xmm5, %%xmm6 \n\t"
+ "movdqa %%xmm2, %%xmm0 \n\t"
+ "movdqa %%xmm6, %%xmm4 \n\t"
+ "paddd %%xmm2, %%xmm2 \n\t"
+ "paddd %%xmm6, %%xmm6 \n\t"
+ "paddd %%xmm0, %%xmm2 \n\t"
+ "paddd %%xmm4, %%xmm6 \n\t"
+ "paddd %%xmm7, %%xmm2 \n\t"
+ "paddd %%xmm7, %%xmm6 \n\t"
+ "psrad $3, %%xmm2 \n\t"
+ "psrad $3, %%xmm6 \n\t"
+ "movdqa (%0), %%xmm0 \n\t"
+ "movdqa 16(%0), %%xmm4 \n\t"
+ "psubd %%xmm2, %%xmm0 \n\t"
+ "psubd %%xmm6, %%xmm4 \n\t"
+ "movdqa %%xmm0, (%0) \n\t"
+ "movdqa %%xmm4, 16(%0) \n\t"
+ :: "r"(&b[i]), "r"(&ref[i])
+ : "memory"
+ );
+ }
+ snow_horizontal_compose_lift_lead_out(i, b, b, ref, width, w_l, 0, W_DM, W_DO, W_DS);
+ b[0] = b_0 - ((W_DM * 2 * ref[1]+W_DO)>>W_DS);
+ }
+
+ { // Lift 1
+ DWTELEM * const dst = b+w2;
+
+ i = 0;
+ for(; (((long)&dst[i]) & 0xF) && i<w_r; i++){
+ dst[i] = dst[i] - (b[i] + b[i + 1]);
+ }
+ for(; i<w_r-7; i+=8){
+ asm volatile(
+ "movdqu (%1), %%xmm1 \n\t"
+ "movdqu 16(%1), %%xmm5 \n\t"
+ "movdqu 4(%1), %%xmm2 \n\t"
+ "movdqu 20(%1), %%xmm6 \n\t"
+ "paddd %%xmm1, %%xmm2 \n\t"
+ "paddd %%xmm5, %%xmm6 \n\t"
+ "movdqa (%0), %%xmm0 \n\t"
+ "movdqa 16(%0), %%xmm4 \n\t"
+ "psubd %%xmm2, %%xmm0 \n\t"
+ "psubd %%xmm6, %%xmm4 \n\t"
+ "movdqa %%xmm0, (%0) \n\t"
+ "movdqa %%xmm4, 16(%0) \n\t"
+ :: "r"(&dst[i]), "r"(&b[i])
+ : "memory"
+ );
+ }
+ snow_horizontal_compose_lift_lead_out(i, dst, dst, b, width, w_r, 1, W_CM, W_CO, W_CS);
+ }
+
+ { // Lift 2
+ DWTELEM * const ref = b+w2 - 1;
+ DWTELEM b_0 = b[0];
+
+ i = 0;
+ asm volatile(
+ "pslld $1, %%xmm7 \n\t" /* xmm7 already holds a '4' from 2 lifts ago. */
+ ::);
+ for(; i<w_l-7; i+=8){
+ asm volatile(
+ "movdqu (%1), %%xmm1 \n\t"
+ "movdqu 16(%1), %%xmm5 \n\t"
+ "movdqu 4(%1), %%xmm0 \n\t"
+ "movdqu 20(%1), %%xmm4 \n\t"
+ "paddd %%xmm1, %%xmm0 \n\t"
+ "paddd %%xmm5, %%xmm4 \n\t"
+ "movdqa %%xmm7, %%xmm1 \n\t"
+ "movdqa %%xmm7, %%xmm5 \n\t"
+ "psubd %%xmm0, %%xmm1 \n\t"
+ "psubd %%xmm4, %%xmm5 \n\t"
+ "movdqa (%0), %%xmm0 \n\t"
+ "movdqa 16(%0), %%xmm4 \n\t"
+ "pslld $2, %%xmm0 \n\t"
+ "pslld $2, %%xmm4 \n\t"
+ "psubd %%xmm0, %%xmm1 \n\t"
+ "psubd %%xmm4, %%xmm5 \n\t"
+ "psrad $4, %%xmm1 \n\t"
+ "psrad $4, %%xmm5 \n\t"
+ "movdqa (%0), %%xmm0 \n\t"
+ "movdqa 16(%0), %%xmm4 \n\t"
+ "psubd %%xmm1, %%xmm0 \n\t"
+ "psubd %%xmm5, %%xmm4 \n\t"
+ "movdqa %%xmm0, (%0) \n\t"
+ "movdqa %%xmm4, 16(%0) \n\t"
+ :: "r"(&b[i]), "r"(&ref[i])
+ : "memory"
+ );
+ }
+ snow_horizontal_compose_liftS_lead_out(i, b, b, ref, width, w_l);
+ b[0] = b_0 - (((-2 * ref[1] + W_BO) - 4 * b_0) >> W_BS);
+ }
+
+ { // Lift 3
+ DWTELEM * const src = b+w2;
+
+ i = 0;
+ for(; (((long)&temp[i]) & 0xF) && i<w_r; i++){
+ temp[i] = src[i] - ((-W_AM*(b[i] + b[i+1]))>>W_AS);
+ }
+ for(; i<w_r-7; i+=8){
+ asm volatile(
+ "movdqu 4(%1), %%xmm2 \n\t"
+ "movdqu 20(%1), %%xmm6 \n\t"
+ "paddd (%1), %%xmm2 \n\t"
+ "paddd 16(%1), %%xmm6 \n\t"
+ "movdqa %%xmm2, %%xmm0 \n\t"
+ "movdqa %%xmm6, %%xmm4 \n\t"
+ "pslld $2, %%xmm2 \n\t"
+ "pslld $2, %%xmm6 \n\t"
+ "psubd %%xmm2, %%xmm0 \n\t"
+ "psubd %%xmm6, %%xmm4 \n\t"
+ "psrad $1, %%xmm0 \n\t"
+ "psrad $1, %%xmm4 \n\t"
+ "movdqu (%0), %%xmm2 \n\t"
+ "movdqu 16(%0), %%xmm6 \n\t"
+ "psubd %%xmm0, %%xmm2 \n\t"
+ "psubd %%xmm4, %%xmm6 \n\t"
+ "movdqa %%xmm2, (%2) \n\t"
+ "movdqa %%xmm6, 16(%2) \n\t"
+ :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i])
+ : "memory"
+ );
+ }
+ snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO, W_AS);
+ }
+
+ {
+ snow_interleave_line_header(&i, width, b, temp);
+
+ for (; (i & 0x1E) != 0x1E; i-=2){
+ b[i+1] = temp[i>>1];
+ b[i] = b[i>>1];
+ }
+ for (i-=30; i>=0; i-=32){
+ asm volatile(
+ "movdqa (%1), %%xmm0 \n\t"
+ "movdqa 16(%1), %%xmm2 \n\t"
+ "movdqa 32(%1), %%xmm4 \n\t"
+ "movdqa 48(%1), %%xmm6 \n\t"
+ "movdqa (%1), %%xmm1 \n\t"
+ "movdqa 16(%1), %%xmm3 \n\t"
+ "movdqa 32(%1), %%xmm5 \n\t"
+ "movdqa 48(%1), %%xmm7 \n\t"
+ "punpckldq (%2), %%xmm0 \n\t"
+ "punpckldq 16(%2), %%xmm2 \n\t"
+ "punpckldq 32(%2), %%xmm4 \n\t"
+ "punpckldq 48(%2), %%xmm6 \n\t"
+ "movdqa %%xmm0, (%0) \n\t"
+ "movdqa %%xmm2, 32(%0) \n\t"
+ "movdqa %%xmm4, 64(%0) \n\t"
+ "movdqa %%xmm6, 96(%0) \n\t"
+ "punpckhdq (%2), %%xmm1 \n\t"
+ "punpckhdq 16(%2), %%xmm3 \n\t"
+ "punpckhdq 32(%2), %%xmm5 \n\t"
+ "punpckhdq 48(%2), %%xmm7 \n\t"
+ "movdqa %%xmm1, 16(%0) \n\t"
+ "movdqa %%xmm3, 48(%0) \n\t"
+ "movdqa %%xmm5, 80(%0) \n\t"
+ "movdqa %%xmm7, 112(%0) \n\t"
+ :: "r"(&(b)[i]), "r"(&(b)[i>>1]), "r"(&(temp)[i>>1])
+ : "memory"
+ );
+ }
+ }
+}
+
+void ff_snow_horizontal_compose97i_mmx(DWTELEM *b, int width){
+ const int w2= (width+1)>>1;
+ DWTELEM temp[width >> 1];
+ const int w_l= (width>>1);
+ const int w_r= w2 - 1;
+ int i;
+
+ { // Lift 0
+ DWTELEM * const ref = b + w2 - 1;
+
+ i = 1;
+ b[0] = b[0] - ((W_DM * 2 * ref[1]+W_DO)>>W_DS);
+ asm volatile(
+ "pcmpeqd %%mm7, %%mm7 \n\t"
+ "pslld $31, %%mm7 \n\t"
+ "psrld $29, %%mm7 \n\t"
+ ::);
+ for(; i<w_l-3; i+=4){
+ asm volatile(
+ "movq (%1), %%mm2 \n\t"
+ "movq 8(%1), %%mm6 \n\t"
+ "paddd 4(%1), %%mm2 \n\t"
+ "paddd 12(%1), %%mm6 \n\t"
+ "movq %%mm2, %%mm0 \n\t"
+ "movq %%mm6, %%mm4 \n\t"
+ "paddd %%mm2, %%mm2 \n\t"
+ "paddd %%mm6, %%mm6 \n\t"
+ "paddd %%mm0, %%mm2 \n\t"
+ "paddd %%mm4, %%mm6 \n\t"
+ "paddd %%mm7, %%mm2 \n\t"
+ "paddd %%mm7, %%mm6 \n\t"
+ "psrad $3, %%mm2 \n\t"
+ "psrad $3, %%mm6 \n\t"
+ "movq (%0), %%mm0 \n\t"
+ "movq 8(%0), %%mm4 \n\t"
+ "psubd %%mm2, %%mm0 \n\t"
+ "psubd %%mm6, %%mm4 \n\t"
+ "movq %%mm0, (%0) \n\t"
+ "movq %%mm4, 8(%0) \n\t"
+ :: "r"(&b[i]), "r"(&ref[i])
+ : "memory"
+ );
+ }
+ snow_horizontal_compose_lift_lead_out(i, b, b, ref, width, w_l, 0, W_DM, W_DO, W_DS);
+ }
+
+ { // Lift 1
+ DWTELEM * const dst = b+w2;
+
+ i = 0;
+ for(; i<w_r-3; i+=4){
+ asm volatile(
+ "movq (%1), %%mm2 \n\t"
+ "movq 8(%1), %%mm6 \n\t"
+ "paddd 4(%1), %%mm2 \n\t"
+ "paddd 12(%1), %%mm6 \n\t"
+ "movq (%0), %%mm0 \n\t"
+ "movq 8(%0), %%mm4 \n\t"
+ "psubd %%mm2, %%mm0 \n\t"
+ "psubd %%mm6, %%mm4 \n\t"
+ "movq %%mm0, (%0) \n\t"
+ "movq %%mm4, 8(%0) \n\t"
+ :: "r"(&dst[i]), "r"(&b[i])
+ : "memory"
+ );
+ }
+ snow_horizontal_compose_lift_lead_out(i, dst, dst, b, width, w_r, 1, W_CM, W_CO, W_CS);
+ }
+
+ { // Lift 2
+ DWTELEM * const ref = b+w2 - 1;
+
+ i = 1;
+ b[0] = b[0] - (((-2 * ref[1] + W_BO) - 4 * b[0]) >> W_BS);
+ asm volatile(
+ "pslld $1, %%mm7 \n\t" /* xmm7 already holds a '4' from 2 lifts ago. */
+ ::);
+ for(; i<w_l-3; i+=4){
+ asm volatile(
+ "movq (%1), %%mm0 \n\t"
+ "movq 8(%1), %%mm4 \n\t"
+ "paddd 4(%1), %%mm0 \n\t"
+ "paddd 12(%1), %%mm4 \n\t"
+ "movq %%mm7, %%mm1 \n\t"
+ "movq %%mm7, %%mm5 \n\t"
+ "psubd %%mm0, %%mm1 \n\t"
+ "psubd %%mm4, %%mm5 \n\t"
+ "movq (%0), %%mm0 \n\t"
+ "movq 8(%0), %%mm4 \n\t"
+ "pslld $2, %%mm0 \n\t"
+ "pslld $2, %%mm4 \n\t"
+ "psubd %%mm0, %%mm1 \n\t"
+ "psubd %%mm4, %%mm5 \n\t"
+ "psrad $4, %%mm1 \n\t"
+ "psrad $4, %%mm5 \n\t"
+ "movq (%0), %%mm0 \n\t"
+ "movq 8(%0), %%mm4 \n\t"
+ "psubd %%mm1, %%mm0 \n\t"
+ "psubd %%mm5, %%mm4 \n\t"
+ "movq %%mm0, (%0) \n\t"
+ "movq %%mm4, 8(%0) \n\t"
+ :: "r"(&b[i]), "r"(&ref[i])
+ : "memory"
+ );
+ }
+ snow_horizontal_compose_liftS_lead_out(i, b, b, ref, width, w_l);
+ }
+
+ { // Lift 3
+ DWTELEM * const src = b+w2;
+ i = 0;
+
+ for(; i<w_r-3; i+=4){
+ asm volatile(
+ "movq 4(%1), %%mm2 \n\t"
+ "movq 12(%1), %%mm6 \n\t"
+ "paddd (%1), %%mm2 \n\t"
+ "paddd 8(%1), %%mm6 \n\t"
+ "movq %%mm2, %%mm0 \n\t"
+ "movq %%mm6, %%mm4 \n\t"
+ "pslld $2, %%mm2 \n\t"
+ "pslld $2, %%mm6 \n\t"
+ "psubd %%mm2, %%mm0 \n\t"
+ "psubd %%mm6, %%mm4 \n\t"
+ "psrad $1, %%mm0 \n\t"
+ "psrad $1, %%mm4 \n\t"
+ "movq (%0), %%mm2 \n\t"
+ "movq 8(%0), %%mm6 \n\t"
+ "psubd %%mm0, %%mm2 \n\t"
+ "psubd %%mm4, %%mm6 \n\t"
+ "movq %%mm2, (%2) \n\t"
+ "movq %%mm6, 8(%2) \n\t"
+ :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i])
+ : "memory"
+ );
+ }
+ snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO, W_AS);
+ }
+
+ {
+ snow_interleave_line_header(&i, width, b, temp);
+
+ for (; (i & 0xE) != 0xE; i-=2){
+ b[i+1] = temp[i>>1];
+ b[i] = b[i>>1];
+ }
+ for (i-=14; i>=0; i-=16){
+ asm volatile(
+ "movq (%1), %%mm0 \n\t"
+ "movq 8(%1), %%mm2 \n\t"
+ "movq 16(%1), %%mm4 \n\t"
+ "movq 24(%1), %%mm6 \n\t"
+ "movq (%1), %%mm1 \n\t"
+ "movq 8(%1), %%mm3 \n\t"
+ "movq 16(%1), %%mm5 \n\t"
+ "movq 24(%1), %%mm7 \n\t"
+ "punpckldq (%2), %%mm0 \n\t"
+ "punpckldq 8(%2), %%mm2 \n\t"
+ "punpckldq 16(%2), %%mm4 \n\t"
+ "punpckldq 24(%2), %%mm6 \n\t"
+ "movq %%mm0, (%0) \n\t"
+ "movq %%mm2, 16(%0) \n\t"
+ "movq %%mm4, 32(%0) \n\t"
+ "movq %%mm6, 48(%0) \n\t"
+ "punpckhdq (%2), %%mm1 \n\t"
+ "punpckhdq 8(%2), %%mm3 \n\t"
+ "punpckhdq 16(%2), %%mm5 \n\t"
+ "punpckhdq 24(%2), %%mm7 \n\t"
+ "movq %%mm1, 8(%0) \n\t"
+ "movq %%mm3, 24(%0) \n\t"
+ "movq %%mm5, 40(%0) \n\t"
+ "movq %%mm7, 56(%0) \n\t"
+ :: "r"(&b[i]), "r"(&b[i>>1]), "r"(&temp[i>>1])
+ : "memory"
+ );
+ }
+ }
+}
+
#define snow_vertical_compose_sse2_load_add(op,r,t0,t1,t2,t3)\
""op" (%%"r",%%"REG_d",4), %%"t0" \n\t"\
""op" 16(%%"r",%%"REG_d",4), %%"t1" \n\t"\
More information about the ffmpeg-cvslog
mailing list