snow mmx+sse2 optimizations, part 4
authorRobert Edele <yartrebo@earthlink.net>
Tue, 21 Mar 2006 21:51:07 +0000 (21:51 +0000)
committerCorey Hickey <bugfood-ml@fatooh.org>
Tue, 21 Mar 2006 21:51:07 +0000 (21:51 +0000)
Patch by Robert Edele, yartrebo <<at>> earthlink <<dot>> net

Originally committed as revision 5191 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/i386/dsputil_mmx.c
libavcodec/i386/snowdsp_mmx.c

index a5e9675..72e3bd1 100644 (file)
@@ -2564,8 +2564,12 @@ static void ff_idct_xvid_mmx2_add(uint8_t *dest, int line_size, DCTELEM *block)
 }
 #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 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
 
 #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 9f5f067..a66b9fb 100644 (file)
 #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"\