Changeset 846 for trunk/src/gui/painting/qdrawhelper_sse2.cpp

Timestamp:

May 5, 2011, 5:36:53 AM (14 years ago)

Author:

Dmitry A. Kuminov

Message:

trunk: Merged in qt 4.7.2 sources from branches/vendor/nokia/qt.

Location:

trunk

Files:

• . (modified) (1 prop)
• src/gui/painting/qdrawhelper_sse2.cpp (modified) (10 diffs)

trunk/src/gui/painting/qdrawhelper_sse2.cpp

@@ -1 +1 @@
 /****************************************************************************
 **
-** Copyright (C) 2010 Nokia Corporation and/or its subsidiary(-ies).
+** Copyright (C) 2011 Nokia Corporation and/or its subsidiary(-ies).
 ** All rights reserved.
 ** Contact: Nokia Corporation (qt-info@nokia.com)
@@ -44 +44 @@
 #ifdef QT_HAVE_SSE2
 
+#include <private/qdrawingprimitive_sse2_p.h>
 #include <private/qpaintengine_raster_p.h>
 
-#ifdef QT_LINUXBASE
-// this is an evil hack - the posix_memalign declaration in LSB
-// is wrong - see http://bugs.linuxbase.org/show_bug.cgi?id=2431
-#  define posix_memalign _lsb_hack_posix_memalign
-#  include <emmintrin.h>
-#  undef posix_memalign
-#else
-#  include <emmintrin.h>
-#endif
-
 QT_BEGIN_NAMESPACE
-
-/*
- * Multiply the components of pixelVector by alphaChannel
- * Each 32bits components of alphaChannel must be in the form 0x00AA00AA
- * colorMask must have 0x00ff00ff on each 32 bits component
- * half must have the value 128 (0x80) for each 32 bits compnent
- */
-#define BYTE_MUL_SSE2(result, pixelVector, alphaChannel, colorMask, half) \
-{ \
-    /* 1. separate the colors in 2 vectors so each color is on 16 bits \
-       (in order to be multiplied by the alpha \
-       each 32 bit of dstVectorAG are in the form 0x00AA00GG \
-       each 32 bit of dstVectorRB are in the form 0x00RR00BB */\
-    __m128i pixelVectorAG = _mm_srli_epi16(pixelVector, 8); \
-    __m128i pixelVectorRB = _mm_and_si128(pixelVector, colorMask); \
- \
-    /* 2. multiply the vectors by the alpha channel */\
-    pixelVectorAG = _mm_mullo_epi16(pixelVectorAG, alphaChannel); \
-    pixelVectorRB = _mm_mullo_epi16(pixelVectorRB, alphaChannel); \
- \
-    /* 3. devide by 255, that's the tricky part. \
-       we do it like for BYTE_MUL(), with bit shift: X/255 ~= (X + X/256 + rounding)/256 */ \
-    /** so first (X + X/256 + rounding) */\
-    pixelVectorRB = _mm_add_epi16(pixelVectorRB, _mm_srli_epi16(pixelVectorRB, 8)); \
-    pixelVectorRB = _mm_add_epi16(pixelVectorRB, half); \
-    pixelVectorAG = _mm_add_epi16(pixelVectorAG, _mm_srli_epi16(pixelVectorAG, 8)); \
-    pixelVectorAG = _mm_add_epi16(pixelVectorAG, half); \
- \
-    /** second devide by 256 */\
-    pixelVectorRB = _mm_srli_epi16(pixelVectorRB, 8); \
-    /** for AG, we could >> 8 to divide followed by << 8 to put the \
-        bytes in the correct position. By masking instead, we execute \
-        only one instruction */\
-    pixelVectorAG = _mm_andnot_si128(colorMask, pixelVectorAG); \
- \
-    /* 4. combine the 2 pairs of colors */ \
-    result = _mm_or_si128(pixelVectorAG, pixelVectorRB); \
-}
-
-/*
- * Each 32bits components of alphaChannel must be in the form 0x00AA00AA
- * oneMinusAlphaChannel must be 255 - alpha for each 32 bits component
- * colorMask must have 0x00ff00ff on each 32 bits component
- * half must have the value 128 (0x80) for each 32 bits compnent
- */
-#define INTERPOLATE_PIXEL_255_SSE2(result, srcVector, dstVector, alphaChannel, oneMinusAlphaChannel, colorMask, half) { \
-    /* interpolate AG */\
-    __m128i srcVectorAG = _mm_srli_epi16(srcVector, 8); \
-    __m128i dstVectorAG = _mm_srli_epi16(dstVector, 8); \
-    __m128i srcVectorAGalpha = _mm_mullo_epi16(srcVectorAG, alphaChannel); \
-    __m128i dstVectorAGoneMinusAlphalpha = _mm_mullo_epi16(dstVectorAG, oneMinusAlphaChannel); \
-    __m128i finalAG = _mm_add_epi16(srcVectorAGalpha, dstVectorAGoneMinusAlphalpha); \
-    finalAG = _mm_add_epi16(finalAG, _mm_srli_epi16(finalAG, 8)); \
-    finalAG = _mm_add_epi16(finalAG, half); \
-    finalAG = _mm_andnot_si128(colorMask, finalAG); \
- \
-    /* interpolate RB */\
-    __m128i srcVectorRB = _mm_and_si128(srcVector, colorMask); \
-    __m128i dstVectorRB = _mm_and_si128(dstVector, colorMask); \
-    __m128i srcVectorRBalpha = _mm_mullo_epi16(srcVectorRB, alphaChannel); \
-    __m128i dstVectorRBoneMinusAlphalpha = _mm_mullo_epi16(dstVectorRB, oneMinusAlphaChannel); \
-    __m128i finalRB = _mm_add_epi16(srcVectorRBalpha, dstVectorRBoneMinusAlphalpha); \
-    finalRB = _mm_add_epi16(finalRB, _mm_srli_epi16(finalRB, 8)); \
-    finalRB = _mm_add_epi16(finalRB, half); \
-    finalRB = _mm_srli_epi16(finalRB, 8); \
- \
-    /* combine */\
-    result = _mm_or_si128(finalAG, finalRB); \
-}
 
 void qt_blend_argb32_on_argb32_sse2(uchar *destPixels, int dbpl,
@@ -133 +55 @@
 {
     const quint32 *src = (const quint32 *) srcPixels;
-    quint32 *dst = (uint *) destPixels;
+    quint32 *dst = (quint32 *) destPixels;
     if (const_alpha == 256) {
         const __m128i alphaMask = _mm_set1_epi32(0xff000000);
@@ -141 +63 @@
         const __m128i colorMask = _mm_set1_epi32(0x00ff00ff);
         for (int y = 0; y < h; ++y) {
-            int x = 0;
-            for (; x < w-3; x += 4) {
-                const __m128i srcVector = _mm_loadu_si128((__m128i *)&src[x]);
-                const __m128i srcVectorAlpha = _mm_and_si128(srcVector, alphaMask);
-                if (_mm_movemask_epi8(_mm_cmpeq_epi32(srcVectorAlpha, alphaMask)) == 0xffff) {
-                    // all opaque
-                    _mm_storeu_si128((__m128i *)&dst[x], srcVector);
-                } else if (_mm_movemask_epi8(_mm_cmpeq_epi32(srcVectorAlpha, nullVector)) != 0xffff) {
-                    // not fully transparent
-                    // result = s + d * (1-alpha)
-
-                    // extract the alpha channel on 2 x 16 bits
-                    // so we have room for the multiplication
-                    // each 32 bits will be in the form 0x00AA00AA
-                    // with A being the 1 - alpha
-                    __m128i alphaChannel = _mm_srli_epi32(srcVector, 24);
-                    alphaChannel = _mm_or_si128(alphaChannel, _mm_slli_epi32(alphaChannel, 16));
-                    alphaChannel = _mm_sub_epi16(one, alphaChannel);
-
-                    const __m128i dstVector = _mm_loadu_si128((__m128i *)&dst[x]);
-                    __m128i destMultipliedByOneMinusAlpha;
-                    BYTE_MUL_SSE2(destMultipliedByOneMinusAlpha, dstVector, alphaChannel, colorMask, half);
-
-                    // result = s + d * (1-alpha)
-                    const __m128i result = _mm_add_epi8(srcVector, destMultipliedByOneMinusAlpha);
-                    _mm_storeu_si128((__m128i *)&dst[x], result);
-                }
-            }
-            for (; x<w; ++x) {
-                uint s = src[x];
-                if (s >= 0xff000000)
-                    dst[x] = s;
-                else if (s != 0)
-                    dst[x] = s + BYTE_MUL(dst[x], qAlpha(~s));
-            }
+            BLEND_SOURCE_OVER_ARGB32_SSE2(dst, src, w, nullVector, half, one, colorMask, alphaMask);
             dst = (quint32 *)(((uchar *) dst) + dbpl);
             src = (const quint32 *)(((const uchar *) src) + sbpl);
@@ -190 +78 @@
         const __m128i constAlphaVector = _mm_set1_epi16(const_alpha);
         for (int y = 0; y < h; ++y) {
-            int x = 0;
-            for (; x < w-3; x += 4) {
-                __m128i srcVector = _mm_loadu_si128((__m128i *)&src[x]);
-                if (_mm_movemask_epi8(_mm_cmpeq_epi32(srcVector, nullVector)) != 0xffff) {
-                    BYTE_MUL_SSE2(srcVector, srcVector, constAlphaVector, colorMask, half);
-
-                    __m128i alphaChannel = _mm_srli_epi32(srcVector, 24);
-                    alphaChannel = _mm_or_si128(alphaChannel, _mm_slli_epi32(alphaChannel, 16));
-                    alphaChannel = _mm_sub_epi16(one, alphaChannel);
-
-                    const __m128i dstVector = _mm_loadu_si128((__m128i *)&dst[x]);
-                    __m128i destMultipliedByOneMinusAlpha;
-                    BYTE_MUL_SSE2(destMultipliedByOneMinusAlpha, dstVector, alphaChannel, colorMask, half);
-
-                    const __m128i result = _mm_add_epi8(srcVector, destMultipliedByOneMinusAlpha);
-                    _mm_storeu_si128((__m128i *)&dst[x], result);
-                }
-            }
-            for (; x<w; ++x) {
-                quint32 s = src[x];
-                if (s != 0) {
-                    s = BYTE_MUL(s, const_alpha);
-                    dst[x] = s + BYTE_MUL(dst[x], qAlpha(~s));
-                }
-            }
+            BLEND_SOURCE_OVER_ARGB32_WITH_CONST_ALPHA_SSE2(dst, src, w, nullVector, half, one, colorMask, constAlphaVector)
             dst = (quint32 *)(((uchar *) dst) + dbpl);
             src = (const quint32 *)(((const uchar *) src) + sbpl);
@@ -233 +97 @@
 {
     const quint32 *src = (const quint32 *) srcPixels;
-    quint32 *dst = (uint *) destPixels;
+    quint32 *dst = (quint32 *) destPixels;
     if (const_alpha != 256) {
         if (const_alpha != 0) {
@@ -246 +110 @@
             for (int y = 0; y < h; ++y) {
                 int x = 0;
+
+                // First, align dest to 16 bytes:
+                ALIGNMENT_PROLOGUE_16BYTES(dst, x, w) {
+                    quint32 s = src[x];
+                    s = BYTE_MUL(s, const_alpha);
+                    dst[x] = INTERPOLATE_PIXEL_255(src[x], const_alpha, dst[x], one_minus_const_alpha);
+                }
+
                 for (; x < w-3; x += 4) {
                     __m128i srcVector = _mm_loadu_si128((__m128i *)&src[x]);
                     if (_mm_movemask_epi8(_mm_cmpeq_epi32(srcVector, nullVector)) != 0xffff) {
-                        const __m128i dstVector = _mm_loadu_si128((__m128i *)&dst[x]);
+                        const __m128i dstVector = _mm_load_si128((__m128i *)&dst[x]);
                         __m128i result;
                         INTERPOLATE_PIXEL_255_SSE2(result, srcVector, dstVector, constAlphaVector, oneMinusConstAlpha, colorMask, half);
-                        _mm_storeu_si128((__m128i *)&dst[x], result);
+                        _mm_store_si128((__m128i *)&dst[x], result);
                     }
                 }
@@ -269 +141 @@
 }
 
+void QT_FASTCALL comp_func_SourceOver_sse2(uint *destPixels, const uint *srcPixels, int length, uint const_alpha)
+{
+    Q_ASSERT(const_alpha < 256);
+
+    const quint32 *src = (const quint32 *) srcPixels;
+    quint32 *dst = (quint32 *) destPixels;
+
+    const __m128i nullVector = _mm_set1_epi32(0);
+    const __m128i half = _mm_set1_epi16(0x80);
+    const __m128i one = _mm_set1_epi16(0xff);
+    const __m128i colorMask = _mm_set1_epi32(0x00ff00ff);
+    if (const_alpha == 255) {
+        const __m128i alphaMask = _mm_set1_epi32(0xff000000);
+        BLEND_SOURCE_OVER_ARGB32_SSE2(dst, src, length, nullVector, half, one, colorMask, alphaMask);
+    } else {
+        const __m128i constAlphaVector = _mm_set1_epi16(const_alpha);
+        BLEND_SOURCE_OVER_ARGB32_WITH_CONST_ALPHA_SSE2(dst, src, length, nullVector, half, one, colorMask, constAlphaVector);
+    }
+}
+
+void QT_FASTCALL comp_func_Plus_sse2(uint *dst, const uint *src, int length, uint const_alpha)
+{
+    int x = 0;
+
+    if (const_alpha == 255) {
+        // 1) Prologue: align destination on 16 bytes
+        ALIGNMENT_PROLOGUE_16BYTES(dst, x, length)
+            dst[x] = comp_func_Plus_one_pixel(dst[x], src[x]);
+
+        // 2) composition with SSE2
+        for (; x < length - 3; x += 4) {
+            const __m128i srcVector = _mm_loadu_si128((__m128i *)&src[x]);
+            const __m128i dstVector = _mm_load_si128((__m128i *)&dst[x]);
+
+            const __m128i result = _mm_adds_epu8(srcVector, dstVector);
+            _mm_store_si128((__m128i *)&dst[x], result);
+        }
+
+        // 3) Epilogue:
+        for (; x < length; ++x)
+            dst[x] = comp_func_Plus_one_pixel(dst[x], src[x]);
+    } else {
+        const int one_minus_const_alpha = 255 - const_alpha;
+        const __m128i constAlphaVector = _mm_set1_epi16(const_alpha);
+        const __m128i oneMinusConstAlpha =  _mm_set1_epi16(one_minus_const_alpha);
+
+        // 1) Prologue: align destination on 16 bytes
+        ALIGNMENT_PROLOGUE_16BYTES(dst, x, length)
+            dst[x] = comp_func_Plus_one_pixel_const_alpha(dst[x], src[x], const_alpha, one_minus_const_alpha);
+
+        const __m128i half = _mm_set1_epi16(0x80);
+        const __m128i colorMask = _mm_set1_epi32(0x00ff00ff);
+        // 2) composition with SSE2
+        for (; x < length - 3; x += 4) {
+            const __m128i srcVector = _mm_loadu_si128((__m128i *)&src[x]);
+            const __m128i dstVector = _mm_load_si128((__m128i *)&dst[x]);
+
+            __m128i result = _mm_adds_epu8(srcVector, dstVector);
+            INTERPOLATE_PIXEL_255_SSE2(result, result, dstVector, constAlphaVector, oneMinusConstAlpha, colorMask, half)
+            _mm_store_si128((__m128i *)&dst[x], result);
+        }
+
+        // 3) Epilogue:
+        for (; x < length; ++x)
+            dst[x] = comp_func_Plus_one_pixel_const_alpha(dst[x], src[x], const_alpha, one_minus_const_alpha);
+    }
+}
+
+void QT_FASTCALL comp_func_Source_sse2(uint *dst, const uint *src, int length, uint const_alpha)
+{
+    if (const_alpha == 255) {
+        ::memcpy(dst, src, length * sizeof(uint));
+    } else {
+        const int ialpha = 255 - const_alpha;
+
+        int x = 0;
+
+        // 1) prologue, align on 16 bytes
+        ALIGNMENT_PROLOGUE_16BYTES(dst, x, length)
+            dst[x] = INTERPOLATE_PIXEL_255(src[x], const_alpha, dst[x], ialpha);
+
+        // 2) interpolate pixels with SSE2
+        const __m128i half = _mm_set1_epi16(0x80);
+        const __m128i colorMask = _mm_set1_epi32(0x00ff00ff);
+        const __m128i constAlphaVector = _mm_set1_epi16(const_alpha);
+        const __m128i oneMinusConstAlpha =  _mm_set1_epi16(ialpha);
+        for (; x < length - 3; x += 4) {
+            const __m128i srcVector = _mm_loadu_si128((__m128i *)&src[x]);
+            __m128i dstVector = _mm_load_si128((__m128i *)&dst[x]);
+            INTERPOLATE_PIXEL_255_SSE2(dstVector, srcVector, dstVector, constAlphaVector, oneMinusConstAlpha, colorMask, half)
+            _mm_store_si128((__m128i *)&dst[x], dstVector);
+        }
+
+        // 3) Epilogue
+        for (; x < length; ++x)
+            dst[x] = INTERPOLATE_PIXEL_255(src[x], const_alpha, dst[x], ialpha);
+    }
+}
+
 void qt_memfill32_sse2(quint32 *dest, quint32 value, int count)
 {
@@ -296 +267 @@
     int n = (count128 + 3) / 4;
     switch (count128 & 0x3) {
-    case 0: do { _mm_store_si128(dst128++, value128);
-    case 3:      _mm_store_si128(dst128++, value128);
-    case 2:      _mm_store_si128(dst128++, value128);
-    case 1:      _mm_store_si128(dst128++, value128);
+    case 0: do { _mm_stream_si128(dst128++, value128);
+    case 3:      _mm_stream_si128(dst128++, value128);
+    case 2:      _mm_stream_si128(dst128++, value128);
+    case 1:      _mm_stream_si128(dst128++, value128);
     } while (--n > 0);
     }
@@ -312 +283 @@
     }
 }
+
+void QT_FASTCALL comp_func_solid_SourceOver_sse2(uint *destPixels, int length, uint color, uint const_alpha)
+{
+    if ((const_alpha & qAlpha(color)) == 255) {
+        qt_memfill32_sse2(destPixels, color, length);
+    } else {
+        if (const_alpha != 255)
+            color = BYTE_MUL(color, const_alpha);
+
+        const quint32 minusAlphaOfColor = qAlpha(~color);
+        int x = 0;
+
+        quint32 *dst = (quint32 *) destPixels;
+        const __m128i colorVector = _mm_set1_epi32(color);
+        const __m128i colorMask = _mm_set1_epi32(0x00ff00ff);
+        const __m128i half = _mm_set1_epi16(0x80);
+        const __m128i minusAlphaOfColorVector = _mm_set1_epi16(minusAlphaOfColor);
+
+        ALIGNMENT_PROLOGUE_16BYTES(dst, x, length)
+            destPixels[x] = color + BYTE_MUL(destPixels[x], minusAlphaOfColor);
+
+        for (; x < length-3; x += 4) {
+            __m128i dstVector = _mm_load_si128((__m128i *)&dst[x]);
+            BYTE_MUL_SSE2(dstVector, dstVector, minusAlphaOfColorVector, colorMask, half);
+            dstVector = _mm_add_epi8(colorVector, dstVector);
+            _mm_store_si128((__m128i *)&dst[x], dstVector);
+        }
+        for (;x < length; ++x)
+            destPixels[x] = color + BYTE_MUL(destPixels[x], minusAlphaOfColor);
+    }
+}
+
+CompositionFunctionSolid qt_functionForModeSolid_onlySSE2[numCompositionFunctions] = {
+    comp_func_solid_SourceOver_sse2,
+    comp_func_solid_DestinationOver,
+    comp_func_solid_Clear,
+    comp_func_solid_Source,
+    comp_func_solid_Destination,
+    comp_func_solid_SourceIn,
+    comp_func_solid_DestinationIn,
+    comp_func_solid_SourceOut,
+    comp_func_solid_DestinationOut,
+    comp_func_solid_SourceAtop,
+    comp_func_solid_DestinationAtop,
+    comp_func_solid_XOR,
+    comp_func_solid_Plus,
+    comp_func_solid_Multiply,
+    comp_func_solid_Screen,
+    comp_func_solid_Overlay,
+    comp_func_solid_Darken,
+    comp_func_solid_Lighten,
+    comp_func_solid_ColorDodge,
+    comp_func_solid_ColorBurn,
+    comp_func_solid_HardLight,
+    comp_func_solid_SoftLight,
+    comp_func_solid_Difference,
+    comp_func_solid_Exclusion,
+    rasterop_solid_SourceOrDestination,
+    rasterop_solid_SourceAndDestination,
+    rasterop_solid_SourceXorDestination,
+    rasterop_solid_NotSourceAndNotDestination,
+    rasterop_solid_NotSourceOrNotDestination,
+    rasterop_solid_NotSourceXorDestination,
+    rasterop_solid_NotSource,
+    rasterop_solid_NotSourceAndDestination,
+    rasterop_solid_SourceAndNotDestination
+};
+
+CompositionFunction qt_functionForMode_onlySSE2[numCompositionFunctions] = {
+    comp_func_SourceOver_sse2,
+    comp_func_DestinationOver,
+    comp_func_Clear,
+    comp_func_Source_sse2,
+    comp_func_Destination,
+    comp_func_SourceIn,
+    comp_func_DestinationIn,
+    comp_func_SourceOut,
+    comp_func_DestinationOut,
+    comp_func_SourceAtop,
+    comp_func_DestinationAtop,
+    comp_func_XOR,
+    comp_func_Plus_sse2,
+    comp_func_Multiply,
+    comp_func_Screen,
+    comp_func_Overlay,
+    comp_func_Darken,
+    comp_func_Lighten,
+    comp_func_ColorDodge,
+    comp_func_ColorBurn,
+    comp_func_HardLight,
+    comp_func_SoftLight,
+    comp_func_Difference,
+    comp_func_Exclusion,
+    rasterop_SourceOrDestination,
+    rasterop_SourceAndDestination,
+    rasterop_SourceXorDestination,
+    rasterop_NotSourceAndNotDestination,
+    rasterop_NotSourceOrNotDestination,
+    rasterop_NotSourceXorDestination,
+    rasterop_NotSource,
+    rasterop_NotSourceAndDestination,
+    rasterop_SourceAndNotDestination
+};
 
 void qt_memfill16_sse2(quint16 *dest, quint16 value, int count)