Changeset 286

Timestamp:

01/14/07 15:52:33 (2 years ago)

Author:

astrange

Message:

Use non-temporal SSE writes, use cache prefetching in AltiVec?.
Don't use data streams because they're slow on G5 and I don't understand them.

Files:

• trunk/ColorConversions.c (modified) (5 diffs)

trunk/ColorConversions.c

@@ -117 +117 @@
-                
-                for (x = 0,x2 = 0,x4 =0; x < vWidth; x++, x2 += 2, x4 += 4) {
-
+
+
+
+
+
+
-                        ov[x4] = vec_mergeh(chroma, tmp_y);
+                        ov[x4+1] = vec_mergel(chroma, tmp_y);
+                        ov[x4+2] = vec_mergeh(chromal, tmp_y3);
+                        ov[x4+3] = vec_mergel(chromal, tmp_y3);
-                        ov2[x4] = vec_mergeh(chroma, tmp_y2);
-                        ov[x4+1] = vec_mergel(chroma, tmp_y);
-                        ov2[x4+1] = vec_mergel(chroma, tmp_y2);
-
-
-
-
-
-
-
+
+
-                }
-                
@@ -173 +175 @@
-        UInt8          *yc = picture->data[0], *uc = picture->data[1], *vc = picture->data[2];
-        int             rY = picture->linesize[0], rU = picture->linesize[1], rV = picture->linesize[2];
-x2,
+
-        
-        for (y = 0; y < halfheight; y ++) {
@@ -180 +182 @@
-                long long *uv = (long long *)uc, *vv = (long long*)vc;
-                
-,x2 = 0; x < vWidth; x++, x2 += 2
-
- * this avoids reading chroma twice but should we be doing strictly linear writes instead?
- * fun facts: 1. sse2 supports 64-bit as well as 128-bit loads, so we do that for chroma
- * 2: unrolling loops can be very bad. i think we could have done it here, but x86 is so OoO it doesn't really matter */
+; x < vWidth; x++
+ this avoids reading chroma twice
+* sse2 can do 64-bit loads, so we do that. (apple's h264 doesn't seem to, maybe we should copy them?)
+* unrolling loops is very bad on x86 */
+int x2 = x*2;
-                        __builtin_prefetch(&yv[x+1], 0, 0); __builtin_prefetch(&yv2[x+1], 0, 0); // prefetch next y vectors, throw it out of cache immediately after use
-                        __builtin_prefetch(&uv[x+1], 0, 0); __builtin_prefetch(&vv[x+1], 0, 0); // and chroma too
@@ -194 +196 @@
-                                p3 = _mm_unpacklo_epi8(chroma, tmp_y2),
-                                p4 = _mm_unpackhi_epi8(chroma, tmp_y2);
-
-
-
-
+
+
+
+
+
-                }
-                
-                if (__builtin_expect(width % 16, FALSE)) { //spill to scalar for the end if the row isn't a multiple of 16
-                        UInt8 *o2 = o + outRB, *yc2 = yc + rY;
-, x2 = x*2; x < width; x += 2, x2 += 4
- >> 1
+; x < width; x += 2
+>>1, x2 = x*2
-                                o2[x2] = o[x2] = uc[hx];
-                                o[x2 + 1] = yc[x];
@@ -213 +216 @@
-                }
-                
-; o += outRB
-; yc += rY
+*2
+*2
-                uc += rU;
-                vc += rV;
-        }
-empty(); // leave mmx mode
+sfence(); // complete all writes
-}
-#endif