altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)

Originally committed as revision 1283 to svn://svn.ffmpeg.org/ffmpeg/trunk

3 files changed, 143 insertions, 1 deletions

diff --git a/libavcodec/dsputil.h b/libavcodec/dsputil.h
index 0b1d34f908..29aca1ac22 100644
--- a/libavcodec/dsputil.h
+++ b/libavcodec/dsputil.h
@@ -190,6 +190,10 @@ void dsputil_init_alpha(DSPContext* c, unsigned mask);
 
 #elif defined(ARCH_POWERPC)
 
+#define MM_ALTIVEC    0x0001 /* standard AltiVec */
+
+extern int mm_flags;
+
 #define __align8 __attribute__ ((aligned (16)))
 
 void dsputil_init_ppc(DSPContext* c, unsigned mask);
diff --git a/libavcodec/imgresample.c b/libavcodec/imgresample.c
index 1197f858bc..630f2eb2c6 100644
--- a/libavcodec/imgresample.c
+++ b/libavcodec/imgresample.c
@@ -22,7 +22,7 @@
 #ifdef USE_FASTMEMCPY
 #include "fastmemcpy.h"
 #endif
-
+extern int mm_flags;
 
 #define NB_COMPONENTS 3
 
@@ -264,6 +264,133 @@ static void v_resample4_mmx(UINT8 *dst, int dst_width, UINT8 *src, int wrap,
 }
 #endif
 
+#ifdef HAVE_ALTIVEC
+typedef	union {
+    vector unsigned char v;
+    unsigned char c[16];
+} vec_uc_t;
+
+typedef	union {
+    vector signed short v;
+    signed short s[8];
+} vec_ss_t;
+
+void v_resample16_altivec(UINT8 *dst, int dst_width, UINT8 *src, int wrap,
+                            INT16 *filter)
+{
+    int sum, i;
+    uint8_t *s;
+    vector unsigned char *tv, tmp, dstv, zero;
+    vec_ss_t srchv[4], srclv[4], fv[4];
+    vector signed short zeros, sumhv, sumlv;    
+    s = src;
+
+    for(i=0;i<4;i++)
+    {
+        /*
+           The vec_madds later on does an implicit >>15 on the result.
+           Since FILTER_BITS is 8, and we have 15 bits of magnitude in
+           a signed short, we have just enough bits to pre-shift our
+           filter constants <<7 to compensate for vec_madds.
+        */
+        fv[i].s[0] = filter[i] << (15-FILTER_BITS);
+        fv[i].v = vec_splat(fv[i].v, 0);
+    }
+    
+    zero = vec_splat_u8(0);
+    zeros = vec_splat_s16(0);
+
+
+    /*
+       When we're resampling, we'd ideally like both our input buffers,
+       and output buffers to be 16-byte aligned, so we can do both aligned
+       reads and writes. Sadly we can't always have this at the moment, so
+       we opt for aligned writes, as unaligned writes have a huge overhead.
+       To do this, do enough scalar resamples to get dst 16-byte aligned.
+    */
+    i = (16-((int)dst) & 0xf) & 0xf;
+    while(i>0) {
+        sum = s[0 * wrap] * filter[0] +
+        s[1 * wrap] * filter[1] +
+        s[2 * wrap] * filter[2] +
+        s[3 * wrap] * filter[3];
+        sum = sum >> FILTER_BITS;
+        if (sum<0) sum = 0; else if (sum>255) sum=255;
+        dst[0] = sum;
+        dst++;
+        s++;
+        dst_width--;
+        i--;
+    }
+    
+    /* Do our altivec resampling on 16 pixels at once. */
+    while(dst_width>=16) {
+        /*
+           Read 16 (potentially unaligned) bytes from each of
+           4 lines into 4 vectors, and split them into shorts.
+           Interleave the multipy/accumulate for the resample
+           filter with the loads to hide the 3 cycle latency
+           the vec_madds have.
+        */
+        tv = (vector unsigned char *) &s[0 * wrap];
+        tmp = vec_perm(tv[0], tv[1], vec_lvsl(0, &s[i * wrap]));
+        srchv[0].v = (vector signed short) vec_mergeh(zero, tmp);
+        srclv[0].v = (vector signed short) vec_mergel(zero, tmp);
+        sumhv = vec_madds(srchv[0].v, fv[0].v, zeros);
+        sumlv = vec_madds(srclv[0].v, fv[0].v, zeros);
+
+        tv = (vector unsigned char *) &s[1 * wrap];
+        tmp = vec_perm(tv[0], tv[1], vec_lvsl(0, &s[1 * wrap]));
+        srchv[1].v = (vector signed short) vec_mergeh(zero, tmp);
+        srclv[1].v = (vector signed short) vec_mergel(zero, tmp);
+        sumhv = vec_madds(srchv[1].v, fv[1].v, sumhv);
+        sumlv = vec_madds(srclv[1].v, fv[1].v, sumlv);
+
+        tv = (vector unsigned char *) &s[2 * wrap];
+        tmp = vec_perm(tv[0], tv[1], vec_lvsl(0, &s[2 * wrap]));
+        srchv[2].v = (vector signed short) vec_mergeh(zero, tmp);
+        srclv[2].v = (vector signed short) vec_mergel(zero, tmp);
+        sumhv = vec_madds(srchv[2].v, fv[2].v, sumhv);
+        sumlv = vec_madds(srclv[2].v, fv[2].v, sumlv);
+
+        tv = (vector unsigned char *) &s[3 * wrap];
+        tmp = vec_perm(tv[0], tv[1], vec_lvsl(0, &s[3 * wrap]));
+        srchv[3].v = (vector signed short) vec_mergeh(zero, tmp);
+        srclv[3].v = (vector signed short) vec_mergel(zero, tmp);
+        sumhv = vec_madds(srchv[3].v, fv[3].v, sumhv);
+        sumlv = vec_madds(srclv[3].v, fv[3].v, sumlv);
+    
+        /*
+           Pack the results into our destination vector,
+           and do an aligned write of that back to memory.
+        */
+        dstv = vec_packsu(sumhv, sumlv) ;
+        vec_st(dstv, 0, (vector unsigned char *) dst);
+        
+        dst+=16;
+        s+=16;
+        dst_width-=16;
+    }
+
+    /*
+       If there are any leftover pixels, resample them
+       with the slow scalar method.
+    */
+    while(dst_width>0) {
+        sum = s[0 * wrap] * filter[0] +
+        s[1 * wrap] * filter[1] +
+        s[2 * wrap] * filter[2] +
+        s[3 * wrap] * filter[3];
+        sum = sum >> FILTER_BITS;
+        if (sum<0) sum = 0; else if (sum>255) sum=255;
+        dst[0] = sum;
+        dst++;
+        s++;
+        dst_width--;
+    }
+}
+#endif
+
 /* slow version to handle limit cases. Does not need optimisation */
 static void h_resample_slow(UINT8 *dst, int dst_width, UINT8 *src, int src_width,
                             int src_start, int src_incr, INT16 *filters)
@@ -384,6 +511,13 @@ static void component_resample(ImgReSampleContext *s,
                             &s->v_filters[phase_y][0]);
         else
 #endif
+#ifdef HAVE_ALTIVEC
+            if ((mm_flags & MM_ALTIVEC) && NB_TAPS == 4 && FILTER_BITS == 8)
+                v_resample16_altivec(output, owidth,
+                                s->line_buf + (ring_y - NB_TAPS + 1) * owidth, owidth,
+                                &s->v_filters[phase_y][0]);
+        else
+#endif
             v_resample(output, owidth, 
                        s->line_buf + (ring_y - NB_TAPS + 1) * owidth, owidth, 
                        &s->v_filters[phase_y][0]);
diff --git a/libavcodec/ppc/dsputil_ppc.c b/libavcodec/ppc/dsputil_ppc.c
index c4dae6cb3b..ffe3ce0633 100644
--- a/libavcodec/ppc/dsputil_ppc.c
+++ b/libavcodec/ppc/dsputil_ppc.c
@@ -23,6 +23,8 @@
 #include "dsputil_altivec.h"
 #endif
 
+int mm_flags = 0;
+
 void dsputil_init_ppc(DSPContext* c, unsigned mask)
 {
     // Common optimisations whether Altivec or not
@@ -31,6 +33,8 @@ void dsputil_init_ppc(DSPContext* c, unsigned mask)
 
 #if HAVE_ALTIVEC
     if (has_altivec()) {
+        mm_flags |= MM_ALTIVEC;
+        
         // Altivec specific optimisations
         c->pix_abs16x16_x2 = pix_abs16x16_x2_altivec;
         c->pix_abs16x16_y2 = pix_abs16x16_y2_altivec;