land http://codereview.appspot.com/6353063/ by Lei

optimizations for D16 using SSE2

skia_bench -config 565 -match bitmap_8888_scale_filter -forceFilter 1 -repeat
30

The result I got on Android platform was below:

w/o this optimization routine: 
D/skia    ( 1868): running bench [640 480]     bitmap_8888_scale_filter
D/skia    ( 1868):    565: cmsecs = 286.50

w/ with optimization:
D/skia    ( 1463): running bench [640 480]     bitmap_8888_scale_filter
D/skia    ( 1463):    565: cmsecs = 186.80

The net gain is 34.80%.



git-svn-id: http://skia.googlecode.com/svn/trunk@4729 2bbb7eff-a529-9590-31e7-b0007b416f81

1 files changed, 133 insertions, 0 deletions

diff --git a/src/opts/SkBitmapProcState_opts_SSE2.cpp b/src/opts/SkBitmapProcState_opts_SSE2.cpp
index 5cb432535d..4ec4ba87dc 100644
--- a/src/opts/SkBitmapProcState_opts_SSE2.cpp
+++ b/src/opts/SkBitmapProcState_opts_SSE2.cpp
@@ -632,3 +632,136 @@ void ClampX_ClampY_nofilter_affine_SSE2(const SkBitmapProcState& s,
         fy += dy;           
     }
 }
+
+/*  SSE version of S32_D16_filter_DX_SSE2
+ *  Definition is in section of "D16 functions for SRC == 8888" in SkBitmapProcState.cpp
+ *  It combines S32_opaque_D32_filter_DX_SSE2 and SkPixel32ToPixel16
+ */
+void S32_D16_filter_DX_SSE2(const SkBitmapProcState& s,
+                                   const uint32_t* xy,
+                                   int count, uint16_t* colors) {
+    SkASSERT(count > 0 && colors != NULL);
+    SkASSERT(s.fDoFilter);
+    SkASSERT(s.fBitmap->config() == SkBitmap::kARGB_8888_Config);
+    SkASSERT(s.fBitmap->isOpaque());
+
+    SkPMColor dstColor, tmpColor;
+    const char* srcAddr = static_cast<const char*>(s.fBitmap->getPixels());
+    unsigned rb = s.fBitmap->rowBytes();
+    uint32_t XY = *xy++;
+    unsigned y0 = XY >> 14;
+    const uint32_t* row0 = reinterpret_cast<const uint32_t*>(srcAddr + (y0 >> 4) * rb);
+    const uint32_t* row1 = reinterpret_cast<const uint32_t*>(srcAddr + (XY & 0x3FFF) * rb);
+    unsigned subY = y0 & 0xF;
+
+    // ( 0,  0,  0,  0,  0,  0,  0, 16)
+    __m128i sixteen = _mm_cvtsi32_si128(16);
+
+    // ( 0,  0,  0,  0, 16, 16, 16, 16)
+    sixteen = _mm_shufflelo_epi16(sixteen, 0);
+
+    // ( 0,  0,  0,  0,  0,  0,  0,  y)
+    __m128i allY = _mm_cvtsi32_si128(subY);
+
+    // ( 0,  0,  0,  0,  y,  y,  y,  y)
+    allY = _mm_shufflelo_epi16(allY, 0);
+
+    // ( 0,  0,  0,  0, 16-y, 16-y, 16-y, 16-y)
+    __m128i negY = _mm_sub_epi16(sixteen, allY);
+
+    // (16-y, 16-y, 16-y, 16-y, y, y, y, y)
+    allY = _mm_unpacklo_epi64(allY, negY);
+
+    // (16, 16, 16, 16, 16, 16, 16, 16 )
+    sixteen = _mm_shuffle_epi32(sixteen, 0);
+
+    // ( 0,  0,  0,  0,  0,  0,  0,  0)
+    __m128i zero = _mm_setzero_si128();
+
+    __m128i _m_shift_5 = _mm_set1_epi32((1<<5)-1);
+    __m128i _m_shift_6 = _mm_set1_epi32((1<<6)-1);
+    do {
+        uint32_t XX = *xy++;    // x0:14 | 4 | x1:14
+        unsigned x0 = XX >> 18;
+        unsigned x1 = XX & 0x3FFF;
+
+        // (0, 0, 0, 0, 0, 0, 0, x)
+        __m128i allX = _mm_cvtsi32_si128((XX >> 14) & 0x0F);
+
+        // (0, 0, 0, 0, x, x, x, x)
+        allX = _mm_shufflelo_epi16(allX, 0);
+
+        // (x, x, x, x, x, x, x, x)
+        allX = _mm_shuffle_epi32(allX, 0);
+
+        // (16-x, 16-x, 16-x, 16-x, 16-x, 16-x, 16-x)
+        __m128i negX = _mm_sub_epi16(sixteen, allX);
+
+        // Load 4 samples (pixels).
+        __m128i a00 = _mm_cvtsi32_si128(row0[x0]);
+        __m128i a01 = _mm_cvtsi32_si128(row0[x1]);
+        __m128i a10 = _mm_cvtsi32_si128(row1[x0]);
+        __m128i a11 = _mm_cvtsi32_si128(row1[x1]);
+
+        // (0, 0, a00, a10)
+        __m128i a00a10 = _mm_unpacklo_epi32(a10, a00);
+
+        // Expand to 16 bits per component.
+        a00a10 = _mm_unpacklo_epi8(a00a10, zero);
+
+        // ((a00 * (16-y)), (a10 * y)).
+        a00a10 = _mm_mullo_epi16(a00a10, allY);
+
+        // (a00 * (16-y) * (16-x), a10 * y * (16-x)).
+        a00a10 = _mm_mullo_epi16(a00a10, negX);
+
+        // (0, 0, a01, a10)
+        __m128i a01a11 = _mm_unpacklo_epi32(a11, a01);
+
+        // Expand to 16 bits per component.
+        a01a11 = _mm_unpacklo_epi8(a01a11, zero);
+
+        // (a01 * (16-y)), (a11 * y)
+        a01a11 = _mm_mullo_epi16(a01a11, allY);
+
+        // (a01 * (16-y) * x), (a11 * y * x)
+        a01a11 = _mm_mullo_epi16(a01a11, allX);
+
+        // (a00*w00 + a01*w01, a10*w10 + a11*w11)
+        __m128i sum = _mm_add_epi16(a00a10, a01a11);
+
+        // (DC, a00*w00 + a01*w01)
+        __m128i shifted = _mm_shuffle_epi32(sum, 0xEE);
+
+        // (DC, a00*w00 + a01*w01 + a10*w10 + a11*w11)
+        sum = _mm_add_epi16(sum, shifted);
+
+        // Divide each 16 bit component by 256.
+        sum = _mm_srli_epi16(sum, 8);
+
+        // Pack lower 4 16 bit values of sum into lower 4 bytes.
+        sum = _mm_packus_epi16(sum, zero);
+
+        // Extract low int and store.
+        dstColor = _mm_cvtsi128_si32(sum);
+
+        //*colors++ = SkPixel32ToPixel16(dstColor);
+        // below is much faster than the above. It's tested for Android benchmark--Softweg
+        __m128i _m_temp1 = _mm_set1_epi32(dstColor);
+        __m128i _m_temp2 = _mm_srli_epi32(_m_temp1, 3);
+
+        unsigned int r32 = _mm_cvtsi128_si32(_m_temp2);
+        unsigned r = (r32 & ((1<<5) -1)) << 11;
+
+        _m_temp2 = _mm_srli_epi32(_m_temp2, 7);
+        unsigned int g32 = _mm_cvtsi128_si32(_m_temp2);
+        unsigned g = (g32 & ((1<<6) -1)) << 5;
+
+        _m_temp2 = _mm_srli_epi32(_m_temp2, 9);
+        unsigned int b32 = _mm_cvtsi128_si32(_m_temp2);
+        unsigned b = (b32 & ((1<<5) -1));
+
+        *colors++ = r | g | b;
+
+    } while (--count > 0);
+}