Home | History | Annotate | Download | only in opts 
      1 /*
      2  * Copyright 2011 Google Inc.
      3  *
      4  * Use of this source code is governed by a BSD-style license that can be
      5  * found in the LICENSE file.
      6  */
      7 
      8 #include "SkBlitRect_opts_SSE2.h"
      9 #include "SkBlitRow.h"
     10 #include "SkColorPriv.h"
     11 
     12 #include <emmintrin.h>
     13 
     14 /** Simple blitting of opaque rectangles less than 31 pixels wide:
     15     inlines and merges sections of Color32_SSE2 and sk_memset32_SSE2.
     16 */
     17 static void BlitRect32_OpaqueNarrow_SSE2(SkPMColor* SK_RESTRICT destination,
     18                                   int width, int height,
     19                                   size_t rowBytes, uint32_t color) {
     20     SkASSERT(255 == SkGetPackedA32(color));
     21     SkASSERT(width > 0);
     22     SkASSERT(width < 31);
     23 
     24     while (--height >= 0) {
     25         SkPMColor* dst = destination;
     26         int count = width;
     27 
     28         while (count > 4) {
     29             *dst++ = color;
     30             *dst++ = color;
     31             *dst++ = color;
     32             *dst++ = color;
     33             count -= 4;
     34         }
     35 
     36         while (count > 0) {
     37             *dst++ = color;
     38             --count;
     39         }
     40 
     41         destination = (uint32_t*)((char*)destination + rowBytes);
     42     }
     43 }
     44 
     45 /**
     46   Fast blitting of opaque rectangles at least 31 pixels wide:
     47   inlines and merges sections of Color32_SSE2 and sk_memset32_SSE2.
     48   A 31 pixel rectangle is guaranteed to have at least one
     49   16-pixel aligned span that can take advantage of mm_store.
     50 */
     51 static void BlitRect32_OpaqueWide_SSE2(SkPMColor* SK_RESTRICT destination,
     52                                 int width, int height,
     53                                 size_t rowBytes, uint32_t color) {
     54     SkASSERT(255 == SkGetPackedA32(color));
     55     SkASSERT(width >= 31);
     56 
     57     __m128i color_wide = _mm_set1_epi32(color);
     58     while (--height >= 0) {
     59         // Prefetching one row ahead to L1 cache can equal hardware
     60         // performance for large/tall rects, but never *beats*
     61         // hardware performance.
     62         SkPMColor* dst = destination;
     63         int count = width;
     64 
     65         while (((size_t)dst) & 0x0F) {
     66             *dst++ = color;
     67             --count;
     68         }
     69         __m128i *d = reinterpret_cast<__m128i*>(dst);
     70 
     71         // Googling suggests _mm_stream is only going to beat _mm_store
     72         // for things that wouldn't fit in L2 cache anyway, typically
     73         // >500kB, and precisely fill cache lines.  For us, with
     74         // arrays > 100k elements _mm_stream is still 100%+ slower than
     75         // mm_store.
     76 
     77         // Unrolling to count >= 64 is a break-even for most
     78         // input patterns; we seem to be saturating the bus and having
     79         // low enough overhead at 32.
     80 
     81         while (count >= 32) {
     82             _mm_store_si128(d++, color_wide);
     83             _mm_store_si128(d++, color_wide);
     84             _mm_store_si128(d++, color_wide);
     85             _mm_store_si128(d++, color_wide);
     86             _mm_store_si128(d++, color_wide);
     87             _mm_store_si128(d++, color_wide);
     88             _mm_store_si128(d++, color_wide);
     89             _mm_store_si128(d++, color_wide);
     90             count -= 32;
     91         }
     92         if (count >= 16) {
     93             _mm_store_si128(d++, color_wide);
     94             _mm_store_si128(d++, color_wide);
     95             _mm_store_si128(d++, color_wide);
     96             _mm_store_si128(d++, color_wide);
     97             count -= 16;
     98         }
     99         dst = reinterpret_cast<uint32_t*>(d);
    100 
    101         // Unrolling the loop in the Narrow code is a significant performance
    102         // gain, but unrolling this loop appears to make no difference in
    103         // benchmarks with either mm_store_si128 or individual sets.
    104 
    105         while (count > 0) {
    106             *dst++ = color;
    107             --count;
    108         }
    109 
    110         destination = (uint32_t*)((char*)destination + rowBytes);
    111     }
    112 }
    113 
    114 void ColorRect32_SSE2(SkPMColor* destination,
    115                       int width, int height,
    116                       size_t rowBytes, uint32_t color) {
    117     if (0 == height || 0 == width || 0 == color) {
    118         return;
    119     }
    120     unsigned colorA = SkGetPackedA32(color);
    121     if (false && 255 == colorA) { // disabled but compilable to suppress warning
    122         if (width < 31) {
    123             BlitRect32_OpaqueNarrow_SSE2(destination, width, height,
    124                                          rowBytes, color);
    125         } else {
    126             BlitRect32_OpaqueWide_SSE2(destination, width, height,
    127                                        rowBytes, color);
    128         }
    129     } else {
    130         SkBlitRow::ColorRect32(destination, width, height, rowBytes, color);
    131     }
    132 }
    133