michael@0: 
michael@0: /*
michael@0:  * Copyright 2009 The Android Open Source Project
michael@0:  *
michael@0:  * Use of this source code is governed by a BSD-style license that can be
michael@0:  * found in the LICENSE file.
michael@0:  */
michael@0: 
michael@0: 
michael@0: #include <emmintrin.h>
michael@0: #include "SkBitmapProcState_opts_SSE2.h"
michael@0: #include "SkPaint.h"
michael@0: #include "SkUtils.h"
michael@0: 
michael@0: void S32_opaque_D32_filter_DX_SSE2(const SkBitmapProcState& s,
michael@0:                                    const uint32_t* xy,
michael@0:                                    int count, uint32_t* colors) {
michael@0:     SkASSERT(count > 0 && colors != NULL);
michael@0:     SkASSERT(s.fFilterLevel != SkPaint::kNone_FilterLevel);
michael@0:     SkASSERT(s.fBitmap->config() == SkBitmap::kARGB_8888_Config);
michael@0:     SkASSERT(s.fAlphaScale == 256);
michael@0: 
michael@0:     const char* srcAddr = static_cast<const char*>(s.fBitmap->getPixels());
michael@0:     size_t rb = s.fBitmap->rowBytes();
michael@0:     uint32_t XY = *xy++;
michael@0:     unsigned y0 = XY >> 14;
michael@0:     const uint32_t* row0 = reinterpret_cast<const uint32_t*>(srcAddr + (y0 >> 4) * rb);
michael@0:     const uint32_t* row1 = reinterpret_cast<const uint32_t*>(srcAddr + (XY & 0x3FFF) * rb);
michael@0:     unsigned subY = y0 & 0xF;
michael@0: 
michael@0:     // ( 0,  0,  0,  0,  0,  0,  0, 16)
michael@0:     __m128i sixteen = _mm_cvtsi32_si128(16);
michael@0: 
michael@0:     // ( 0,  0,  0,  0, 16, 16, 16, 16)
michael@0:     sixteen = _mm_shufflelo_epi16(sixteen, 0);
michael@0: 
michael@0:     // ( 0,  0,  0,  0,  0,  0,  0,  y)
michael@0:     __m128i allY = _mm_cvtsi32_si128(subY);
michael@0: 
michael@0:     // ( 0,  0,  0,  0,  y,  y,  y,  y)
michael@0:     allY = _mm_shufflelo_epi16(allY, 0);
michael@0: 
michael@0:     // ( 0,  0,  0,  0, 16-y, 16-y, 16-y, 16-y)
michael@0:     __m128i negY = _mm_sub_epi16(sixteen, allY);
michael@0: 
michael@0:     // (16-y, 16-y, 16-y, 16-y, y, y, y, y)
michael@0:     allY = _mm_unpacklo_epi64(allY, negY);
michael@0: 
michael@0:     // (16, 16, 16, 16, 16, 16, 16, 16 )
michael@0:     sixteen = _mm_shuffle_epi32(sixteen, 0);
michael@0: 
michael@0:     // ( 0,  0,  0,  0,  0,  0,  0,  0)
michael@0:     __m128i zero = _mm_setzero_si128();
michael@0:     do {
michael@0:         uint32_t XX = *xy++;    // x0:14 | 4 | x1:14
michael@0:         unsigned x0 = XX >> 18;
michael@0:         unsigned x1 = XX & 0x3FFF;
michael@0: 
michael@0:         // (0, 0, 0, 0, 0, 0, 0, x)
michael@0:         __m128i allX = _mm_cvtsi32_si128((XX >> 14) & 0x0F);
michael@0: 
michael@0:         // (0, 0, 0, 0, x, x, x, x)
michael@0:         allX = _mm_shufflelo_epi16(allX, 0);
michael@0: 
michael@0:         // (x, x, x, x, x, x, x, x)
michael@0:         allX = _mm_shuffle_epi32(allX, 0);
michael@0: 
michael@0:         // (16-x, 16-x, 16-x, 16-x, 16-x, 16-x, 16-x)
michael@0:         __m128i negX = _mm_sub_epi16(sixteen, allX);
michael@0: 
michael@0:         // Load 4 samples (pixels).
michael@0:         __m128i a00 = _mm_cvtsi32_si128(row0[x0]);
michael@0:         __m128i a01 = _mm_cvtsi32_si128(row0[x1]);
michael@0:         __m128i a10 = _mm_cvtsi32_si128(row1[x0]);
michael@0:         __m128i a11 = _mm_cvtsi32_si128(row1[x1]);
michael@0: 
michael@0:         // (0, 0, a00, a10)
michael@0:         __m128i a00a10 = _mm_unpacklo_epi32(a10, a00);
michael@0: 
michael@0:         // Expand to 16 bits per component.
michael@0:         a00a10 = _mm_unpacklo_epi8(a00a10, zero);
michael@0: 
michael@0:         // ((a00 * (16-y)), (a10 * y)).
michael@0:         a00a10 = _mm_mullo_epi16(a00a10, allY);
michael@0: 
michael@0:         // (a00 * (16-y) * (16-x), a10 * y * (16-x)).
michael@0:         a00a10 = _mm_mullo_epi16(a00a10, negX);
michael@0: 
michael@0:         // (0, 0, a01, a10)
michael@0:         __m128i a01a11 = _mm_unpacklo_epi32(a11, a01);
michael@0: 
michael@0:         // Expand to 16 bits per component.
michael@0:         a01a11 = _mm_unpacklo_epi8(a01a11, zero);
michael@0: 
michael@0:         // (a01 * (16-y)), (a11 * y)
michael@0:         a01a11 = _mm_mullo_epi16(a01a11, allY);
michael@0: 
michael@0:         // (a01 * (16-y) * x), (a11 * y * x)
michael@0:         a01a11 = _mm_mullo_epi16(a01a11, allX);
michael@0: 
michael@0:         // (a00*w00 + a01*w01, a10*w10 + a11*w11)
michael@0:         __m128i sum = _mm_add_epi16(a00a10, a01a11);
michael@0: 
michael@0:         // (DC, a00*w00 + a01*w01)
michael@0:         __m128i shifted = _mm_shuffle_epi32(sum, 0xEE);
michael@0: 
michael@0:         // (DC, a00*w00 + a01*w01 + a10*w10 + a11*w11)
michael@0:         sum = _mm_add_epi16(sum, shifted);
michael@0: 
michael@0:         // Divide each 16 bit component by 256.
michael@0:         sum = _mm_srli_epi16(sum, 8);
michael@0: 
michael@0:         // Pack lower 4 16 bit values of sum into lower 4 bytes.
michael@0:         sum = _mm_packus_epi16(sum, zero);
michael@0: 
michael@0:         // Extract low int and store.
michael@0:         *colors++ = _mm_cvtsi128_si32(sum);
michael@0:     } while (--count > 0);
michael@0: }
michael@0: 
michael@0: void S32_alpha_D32_filter_DX_SSE2(const SkBitmapProcState& s,
michael@0:                                   const uint32_t* xy,
michael@0:                                   int count, uint32_t* colors) {
michael@0:     SkASSERT(count > 0 && colors != NULL);
michael@0:     SkASSERT(s.fFilterLevel != SkPaint::kNone_FilterLevel);
michael@0:     SkASSERT(s.fBitmap->config() == SkBitmap::kARGB_8888_Config);
michael@0:     SkASSERT(s.fAlphaScale < 256);
michael@0: 
michael@0:     const char* srcAddr = static_cast<const char*>(s.fBitmap->getPixels());
michael@0:     size_t rb = s.fBitmap->rowBytes();
michael@0:     uint32_t XY = *xy++;
michael@0:     unsigned y0 = XY >> 14;
michael@0:     const uint32_t* row0 = reinterpret_cast<const uint32_t*>(srcAddr + (y0 >> 4) * rb);
michael@0:     const uint32_t* row1 = reinterpret_cast<const uint32_t*>(srcAddr + (XY & 0x3FFF) * rb);
michael@0:     unsigned subY = y0 & 0xF;
michael@0: 
michael@0:     // ( 0,  0,  0,  0,  0,  0,  0, 16)
michael@0:     __m128i sixteen = _mm_cvtsi32_si128(16);
michael@0: 
michael@0:     // ( 0,  0,  0,  0, 16, 16, 16, 16)
michael@0:     sixteen = _mm_shufflelo_epi16(sixteen, 0);
michael@0: 
michael@0:     // ( 0,  0,  0,  0,  0,  0,  0,  y)
michael@0:     __m128i allY = _mm_cvtsi32_si128(subY);
michael@0: 
michael@0:     // ( 0,  0,  0,  0,  y,  y,  y,  y)
michael@0:     allY = _mm_shufflelo_epi16(allY, 0);
michael@0: 
michael@0:     // ( 0,  0,  0,  0, 16-y, 16-y, 16-y, 16-y)
michael@0:     __m128i negY = _mm_sub_epi16(sixteen, allY);
michael@0: 
michael@0:     // (16-y, 16-y, 16-y, 16-y, y, y, y, y)
michael@0:     allY = _mm_unpacklo_epi64(allY, negY);
michael@0: 
michael@0:     // (16, 16, 16, 16, 16, 16, 16, 16 )
michael@0:     sixteen = _mm_shuffle_epi32(sixteen, 0);
michael@0: 
michael@0:     // ( 0,  0,  0,  0,  0,  0,  0,  0)
michael@0:     __m128i zero = _mm_setzero_si128();
michael@0: 
michael@0:     // ( alpha, alpha, alpha, alpha, alpha, alpha, alpha, alpha )
michael@0:     __m128i alpha = _mm_set1_epi16(s.fAlphaScale);
michael@0: 
michael@0:     do {
michael@0:         uint32_t XX = *xy++;    // x0:14 | 4 | x1:14
michael@0:         unsigned x0 = XX >> 18;
michael@0:         unsigned x1 = XX & 0x3FFF;
michael@0: 
michael@0:         // (0, 0, 0, 0, 0, 0, 0, x)
michael@0:         __m128i allX = _mm_cvtsi32_si128((XX >> 14) & 0x0F);
michael@0: 
michael@0:         // (0, 0, 0, 0, x, x, x, x)
michael@0:         allX = _mm_shufflelo_epi16(allX, 0);
michael@0: 
michael@0:         // (x, x, x, x, x, x, x, x)
michael@0:         allX = _mm_shuffle_epi32(allX, 0);
michael@0: 
michael@0:         // (16-x, 16-x, 16-x, 16-x, 16-x, 16-x, 16-x)
michael@0:         __m128i negX = _mm_sub_epi16(sixteen, allX);
michael@0: 
michael@0:         // Load 4 samples (pixels).
michael@0:         __m128i a00 = _mm_cvtsi32_si128(row0[x0]);
michael@0:         __m128i a01 = _mm_cvtsi32_si128(row0[x1]);
michael@0:         __m128i a10 = _mm_cvtsi32_si128(row1[x0]);
michael@0:         __m128i a11 = _mm_cvtsi32_si128(row1[x1]);
michael@0: 
michael@0:         // (0, 0, a00, a10)
michael@0:         __m128i a00a10 = _mm_unpacklo_epi32(a10, a00);
michael@0: 
michael@0:         // Expand to 16 bits per component.
michael@0:         a00a10 = _mm_unpacklo_epi8(a00a10, zero);
michael@0: 
michael@0:         // ((a00 * (16-y)), (a10 * y)).
michael@0:         a00a10 = _mm_mullo_epi16(a00a10, allY);
michael@0: 
michael@0:         // (a00 * (16-y) * (16-x), a10 * y * (16-x)).
michael@0:         a00a10 = _mm_mullo_epi16(a00a10, negX);
michael@0: 
michael@0:         // (0, 0, a01, a10)
michael@0:         __m128i a01a11 = _mm_unpacklo_epi32(a11, a01);
michael@0: 
michael@0:         // Expand to 16 bits per component.
michael@0:         a01a11 = _mm_unpacklo_epi8(a01a11, zero);
michael@0: 
michael@0:         // (a01 * (16-y)), (a11 * y)
michael@0:         a01a11 = _mm_mullo_epi16(a01a11, allY);
michael@0: 
michael@0:         // (a01 * (16-y) * x), (a11 * y * x)
michael@0:         a01a11 = _mm_mullo_epi16(a01a11, allX);
michael@0: 
michael@0:         // (a00*w00 + a01*w01, a10*w10 + a11*w11)
michael@0:         __m128i sum = _mm_add_epi16(a00a10, a01a11);
michael@0: 
michael@0:         // (DC, a00*w00 + a01*w01)
michael@0:         __m128i shifted = _mm_shuffle_epi32(sum, 0xEE);
michael@0: 
michael@0:         // (DC, a00*w00 + a01*w01 + a10*w10 + a11*w11)
michael@0:         sum = _mm_add_epi16(sum, shifted);
michael@0: 
michael@0:         // Divide each 16 bit component by 256.
michael@0:         sum = _mm_srli_epi16(sum, 8);
michael@0: 
michael@0:         // Multiply by alpha.
michael@0:         sum = _mm_mullo_epi16(sum, alpha);
michael@0: 
michael@0:         // Divide each 16 bit component by 256.
michael@0:         sum = _mm_srli_epi16(sum, 8);
michael@0: 
michael@0:         // Pack lower 4 16 bit values of sum into lower 4 bytes.
michael@0:         sum = _mm_packus_epi16(sum, zero);
michael@0: 
michael@0:         // Extract low int and store.
michael@0:         *colors++ = _mm_cvtsi128_si32(sum);
michael@0:     } while (--count > 0);
michael@0: }
michael@0: 
michael@0: static inline uint32_t ClampX_ClampY_pack_filter(SkFixed f, unsigned max,
michael@0:                                                  SkFixed one) {
michael@0:     unsigned i = SkClampMax(f >> 16, max);
michael@0:     i = (i << 4) | ((f >> 12) & 0xF);
michael@0:     return (i << 14) | SkClampMax((f + one) >> 16, max);
michael@0: }
michael@0: 
michael@0: /*  SSE version of ClampX_ClampY_filter_scale()
michael@0:  *  portable version is in core/SkBitmapProcState_matrix.h
michael@0:  */
michael@0: void ClampX_ClampY_filter_scale_SSE2(const SkBitmapProcState& s, uint32_t xy[],
michael@0:                                      int count, int x, int y) {
michael@0:     SkASSERT((s.fInvType & ~(SkMatrix::kTranslate_Mask |
michael@0:                              SkMatrix::kScale_Mask)) == 0);
michael@0:     SkASSERT(s.fInvKy == 0);
michael@0: 
michael@0:     const unsigned maxX = s.fBitmap->width() - 1;
michael@0:     const SkFixed one = s.fFilterOneX;
michael@0:     const SkFixed dx = s.fInvSx;
michael@0:     SkFixed fx;
michael@0: 
michael@0:     SkPoint pt;
michael@0:     s.fInvProc(s.fInvMatrix, SkIntToScalar(x) + SK_ScalarHalf,
michael@0:                              SkIntToScalar(y) + SK_ScalarHalf, &pt);
michael@0:     const SkFixed fy = SkScalarToFixed(pt.fY) - (s.fFilterOneY >> 1);
michael@0:     const unsigned maxY = s.fBitmap->height() - 1;
michael@0:     // compute our two Y values up front
michael@0:     *xy++ = ClampX_ClampY_pack_filter(fy, maxY, s.fFilterOneY);
michael@0:     // now initialize fx
michael@0:     fx = SkScalarToFixed(pt.fX) - (one >> 1);
michael@0: 
michael@0:     // test if we don't need to apply the tile proc
michael@0:     if (dx > 0 && (unsigned)(fx >> 16) <= maxX &&
michael@0:         (unsigned)((fx + dx * (count - 1)) >> 16) < maxX) {
michael@0:         if (count >= 4) {
michael@0:             // SSE version of decal_filter_scale
michael@0:             while ((size_t(xy) & 0x0F) != 0) {
michael@0:                 SkASSERT((fx >> (16 + 14)) == 0);
michael@0:                 *xy++ = (fx >> 12 << 14) | ((fx >> 16) + 1);
michael@0:                 fx += dx;
michael@0:                 count--;
michael@0:             }
michael@0: 
michael@0:             __m128i wide_1    = _mm_set1_epi32(1);
michael@0:             __m128i wide_dx4  = _mm_set1_epi32(dx * 4);
michael@0:             __m128i wide_fx   = _mm_set_epi32(fx + dx * 3, fx + dx * 2,
michael@0:                                               fx + dx, fx);
michael@0: 
michael@0:             while (count >= 4) {
michael@0:                 __m128i wide_out;
michael@0: 
michael@0:                 wide_out = _mm_slli_epi32(_mm_srai_epi32(wide_fx, 12), 14);
michael@0:                 wide_out = _mm_or_si128(wide_out, _mm_add_epi32(
michael@0:                                         _mm_srai_epi32(wide_fx, 16), wide_1));
michael@0: 
michael@0:                 _mm_store_si128(reinterpret_cast<__m128i*>(xy), wide_out);
michael@0: 
michael@0:                 xy += 4;
michael@0:                 fx += dx * 4;
michael@0:                 wide_fx  = _mm_add_epi32(wide_fx, wide_dx4);
michael@0:                 count -= 4;
michael@0:             } // while count >= 4
michael@0:         } // if count >= 4
michael@0: 
michael@0:         while (count-- > 0) {
michael@0:             SkASSERT((fx >> (16 + 14)) == 0);
michael@0:             *xy++ = (fx >> 12 << 14) | ((fx >> 16) + 1);
michael@0:             fx += dx;
michael@0:         }
michael@0:     } else {
michael@0:         // SSE2 only support 16bit interger max & min, so only process the case
michael@0:         // maxX less than the max 16bit interger. Actually maxX is the bitmap's
michael@0:         // height, there should be rare bitmap whose height will be greater
michael@0:         // than max 16bit interger in the real world.
michael@0:         if ((count >= 4) && (maxX <= 0xFFFF)) {
michael@0:             while (((size_t)xy & 0x0F) != 0) {
michael@0:                 *xy++ = ClampX_ClampY_pack_filter(fx, maxX, one);
michael@0:                 fx += dx;
michael@0:                 count--;
michael@0:             }
michael@0: 
michael@0:             __m128i wide_fx   = _mm_set_epi32(fx + dx * 3, fx + dx * 2,
michael@0:                                               fx + dx, fx);
michael@0:             __m128i wide_dx4  = _mm_set1_epi32(dx * 4);
michael@0:             __m128i wide_one  = _mm_set1_epi32(one);
michael@0:             __m128i wide_maxX = _mm_set1_epi32(maxX);
michael@0:             __m128i wide_mask = _mm_set1_epi32(0xF);
michael@0: 
michael@0:              while (count >= 4) {
michael@0:                 __m128i wide_i;
michael@0:                 __m128i wide_lo;
michael@0:                 __m128i wide_fx1;
michael@0: 
michael@0:                 // i = SkClampMax(f>>16,maxX)
michael@0:                 wide_i = _mm_max_epi16(_mm_srli_epi32(wide_fx, 16),
michael@0:                                        _mm_setzero_si128());
michael@0:                 wide_i = _mm_min_epi16(wide_i, wide_maxX);
michael@0: 
michael@0:                 // i<<4 | TILEX_LOW_BITS(fx)
michael@0:                 wide_lo = _mm_srli_epi32(wide_fx, 12);
michael@0:                 wide_lo = _mm_and_si128(wide_lo, wide_mask);
michael@0:                 wide_i  = _mm_slli_epi32(wide_i, 4);
michael@0:                 wide_i  = _mm_or_si128(wide_i, wide_lo);
michael@0: 
michael@0:                 // i<<14
michael@0:                 wide_i = _mm_slli_epi32(wide_i, 14);
michael@0: 
michael@0:                 // SkClampMax(((f+one))>>16,max)
michael@0:                 wide_fx1 = _mm_add_epi32(wide_fx, wide_one);
michael@0:                 wide_fx1 = _mm_max_epi16(_mm_srli_epi32(wide_fx1, 16),
michael@0:                                                         _mm_setzero_si128());
michael@0:                 wide_fx1 = _mm_min_epi16(wide_fx1, wide_maxX);
michael@0: 
michael@0:                 // final combination
michael@0:                 wide_i = _mm_or_si128(wide_i, wide_fx1);
michael@0:                 _mm_store_si128(reinterpret_cast<__m128i*>(xy), wide_i);
michael@0: 
michael@0:                 wide_fx = _mm_add_epi32(wide_fx, wide_dx4);
michael@0:                 fx += dx * 4;
michael@0:                 xy += 4;
michael@0:                 count -= 4;
michael@0:             } // while count >= 4
michael@0:         } // if count >= 4
michael@0: 
michael@0:         while (count-- > 0) {
michael@0:             *xy++ = ClampX_ClampY_pack_filter(fx, maxX, one);
michael@0:             fx += dx;
michael@0:         }
michael@0:     }
michael@0: }
michael@0: 
michael@0: /*  SSE version of ClampX_ClampY_nofilter_scale()
michael@0:  *  portable version is in core/SkBitmapProcState_matrix.h
michael@0:  */
michael@0: void ClampX_ClampY_nofilter_scale_SSE2(const SkBitmapProcState& s,
michael@0:                                     uint32_t xy[], int count, int x, int y) {
michael@0:     SkASSERT((s.fInvType & ~(SkMatrix::kTranslate_Mask |
michael@0:                              SkMatrix::kScale_Mask)) == 0);
michael@0: 
michael@0:     // we store y, x, x, x, x, x
michael@0:     const unsigned maxX = s.fBitmap->width() - 1;
michael@0:     SkFixed fx;
michael@0:     SkPoint pt;
michael@0:     s.fInvProc(s.fInvMatrix, SkIntToScalar(x) + SK_ScalarHalf,
michael@0:                              SkIntToScalar(y) + SK_ScalarHalf, &pt);
michael@0:     fx = SkScalarToFixed(pt.fY);
michael@0:     const unsigned maxY = s.fBitmap->height() - 1;
michael@0:     *xy++ = SkClampMax(fx >> 16, maxY);
michael@0:     fx = SkScalarToFixed(pt.fX);
michael@0: 
michael@0:     if (0 == maxX) {
michael@0:         // all of the following X values must be 0
michael@0:         memset(xy, 0, count * sizeof(uint16_t));
michael@0:         return;
michael@0:     }
michael@0: 
michael@0:     const SkFixed dx = s.fInvSx;
michael@0: 
michael@0:     // test if we don't need to apply the tile proc
michael@0:     if ((unsigned)(fx >> 16) <= maxX &&
michael@0:         (unsigned)((fx + dx * (count - 1)) >> 16) <= maxX) {
michael@0:         // SSE version of decal_nofilter_scale
michael@0:         if (count >= 8) {
michael@0:             while (((size_t)xy & 0x0F) != 0) {
michael@0:                 *xy++ = pack_two_shorts(fx >> 16, (fx + dx) >> 16);
michael@0:                 fx += 2 * dx;
michael@0:                 count -= 2;
michael@0:             }
michael@0: 
michael@0:             __m128i wide_dx4 = _mm_set1_epi32(dx * 4);
michael@0:             __m128i wide_dx8 = _mm_add_epi32(wide_dx4, wide_dx4);
michael@0: 
michael@0:             __m128i wide_low = _mm_set_epi32(fx + dx * 3, fx + dx * 2,
michael@0:                                              fx + dx, fx);
michael@0:             __m128i wide_high = _mm_add_epi32(wide_low, wide_dx4);
michael@0: 
michael@0:             while (count >= 8) {
michael@0:                 __m128i wide_out_low = _mm_srli_epi32(wide_low, 16);
michael@0:                 __m128i wide_out_high = _mm_srli_epi32(wide_high, 16);
michael@0: 
michael@0:                 __m128i wide_result = _mm_packs_epi32(wide_out_low,
michael@0:                                                       wide_out_high);
michael@0:                 _mm_store_si128(reinterpret_cast<__m128i*>(xy), wide_result);
michael@0: 
michael@0:                 wide_low = _mm_add_epi32(wide_low, wide_dx8);
michael@0:                 wide_high = _mm_add_epi32(wide_high, wide_dx8);
michael@0: 
michael@0:                 xy += 4;
michael@0:                 fx += dx * 8;
michael@0:                 count -= 8;
michael@0:             }
michael@0:         } // if count >= 8
michael@0: 
michael@0:         uint16_t* xx = reinterpret_cast<uint16_t*>(xy);
michael@0:         while (count-- > 0) {
michael@0:             *xx++ = SkToU16(fx >> 16);
michael@0:             fx += dx;
michael@0:         }
michael@0:     } else {
michael@0:         // SSE2 only support 16bit interger max & min, so only process the case
michael@0:         // maxX less than the max 16bit interger. Actually maxX is the bitmap's
michael@0:         // height, there should be rare bitmap whose height will be greater
michael@0:         // than max 16bit interger in the real world.
michael@0:         if ((count >= 8) && (maxX <= 0xFFFF)) {
michael@0:             while (((size_t)xy & 0x0F) != 0) {
michael@0:                 *xy++ = pack_two_shorts(SkClampMax((fx + dx) >> 16, maxX),
michael@0:                                         SkClampMax(fx >> 16, maxX));
michael@0:                 fx += 2 * dx;
michael@0:                 count -= 2;
michael@0:             }
michael@0: 
michael@0:             __m128i wide_dx4 = _mm_set1_epi32(dx * 4);
michael@0:             __m128i wide_dx8 = _mm_add_epi32(wide_dx4, wide_dx4);
michael@0: 
michael@0:             __m128i wide_low = _mm_set_epi32(fx + dx * 3, fx + dx * 2,
michael@0:                                              fx + dx, fx);
michael@0:             __m128i wide_high = _mm_add_epi32(wide_low, wide_dx4);
michael@0:             __m128i wide_maxX = _mm_set1_epi32(maxX);
michael@0: 
michael@0:             while (count >= 8) {
michael@0:                 __m128i wide_out_low = _mm_srli_epi32(wide_low, 16);
michael@0:                 __m128i wide_out_high = _mm_srli_epi32(wide_high, 16);
michael@0: 
michael@0:                 wide_out_low  = _mm_max_epi16(wide_out_low,
michael@0:                                               _mm_setzero_si128());
michael@0:                 wide_out_low  = _mm_min_epi16(wide_out_low, wide_maxX);
michael@0:                 wide_out_high = _mm_max_epi16(wide_out_high,
michael@0:                                               _mm_setzero_si128());
michael@0:                 wide_out_high = _mm_min_epi16(wide_out_high, wide_maxX);
michael@0: 
michael@0:                 __m128i wide_result = _mm_packs_epi32(wide_out_low,
michael@0:                                                       wide_out_high);
michael@0:                 _mm_store_si128(reinterpret_cast<__m128i*>(xy), wide_result);
michael@0: 
michael@0:                 wide_low  = _mm_add_epi32(wide_low, wide_dx8);
michael@0:                 wide_high = _mm_add_epi32(wide_high, wide_dx8);
michael@0: 
michael@0:                 xy += 4;
michael@0:                 fx += dx * 8;
michael@0:                 count -= 8;
michael@0:             }
michael@0:         } // if count >= 8
michael@0: 
michael@0:         uint16_t* xx = reinterpret_cast<uint16_t*>(xy);
michael@0:         while (count-- > 0) {
michael@0:             *xx++ = SkClampMax(fx >> 16, maxX);
michael@0:             fx += dx;
michael@0:         }
michael@0:     }
michael@0: }
michael@0: 
michael@0: /*  SSE version of ClampX_ClampY_filter_affine()
michael@0:  *  portable version is in core/SkBitmapProcState_matrix.h
michael@0:  */
michael@0: void ClampX_ClampY_filter_affine_SSE2(const SkBitmapProcState& s,
michael@0:                                       uint32_t xy[], int count, int x, int y) {
michael@0:     SkPoint srcPt;
michael@0:     s.fInvProc(s.fInvMatrix,
michael@0:                SkIntToScalar(x) + SK_ScalarHalf,
michael@0:                SkIntToScalar(y) + SK_ScalarHalf, &srcPt);
michael@0: 
michael@0:     SkFixed oneX = s.fFilterOneX;
michael@0:     SkFixed oneY = s.fFilterOneY;
michael@0:     SkFixed fx = SkScalarToFixed(srcPt.fX) - (oneX >> 1);
michael@0:     SkFixed fy = SkScalarToFixed(srcPt.fY) - (oneY >> 1);
michael@0:     SkFixed dx = s.fInvSx;
michael@0:     SkFixed dy = s.fInvKy;
michael@0:     unsigned maxX = s.fBitmap->width() - 1;
michael@0:     unsigned maxY = s.fBitmap->height() - 1;
michael@0: 
michael@0:     if (count >= 2 && (maxX <= 0xFFFF)) {
michael@0:         SkFixed dx2 = dx + dx;
michael@0:         SkFixed dy2 = dy + dy;
michael@0: 
michael@0:         __m128i wide_f = _mm_set_epi32(fx + dx, fy + dy, fx, fy);
michael@0:         __m128i wide_d2  = _mm_set_epi32(dx2, dy2, dx2, dy2);
michael@0:         __m128i wide_one  = _mm_set_epi32(oneX, oneY, oneX, oneY);
michael@0:         __m128i wide_max = _mm_set_epi32(maxX, maxY, maxX, maxY);
michael@0:         __m128i wide_mask = _mm_set1_epi32(0xF);
michael@0: 
michael@0:         while (count >= 2) {
michael@0:             // i = SkClampMax(f>>16,maxX)
michael@0:             __m128i wide_i = _mm_max_epi16(_mm_srli_epi32(wide_f, 16),
michael@0:                                            _mm_setzero_si128());
michael@0:             wide_i = _mm_min_epi16(wide_i, wide_max);
michael@0: 
michael@0:             // i<<4 | TILEX_LOW_BITS(f)
michael@0:             __m128i wide_lo = _mm_srli_epi32(wide_f, 12);
michael@0:             wide_lo = _mm_and_si128(wide_lo, wide_mask);
michael@0:             wide_i  = _mm_slli_epi32(wide_i, 4);
michael@0:             wide_i  = _mm_or_si128(wide_i, wide_lo);
michael@0: 
michael@0:             // i<<14
michael@0:             wide_i = _mm_slli_epi32(wide_i, 14);
michael@0: 
michael@0:             // SkClampMax(((f+one))>>16,max)
michael@0:             __m128i wide_f1 = _mm_add_epi32(wide_f, wide_one);
michael@0:             wide_f1 = _mm_max_epi16(_mm_srli_epi32(wide_f1, 16),
michael@0:                                                    _mm_setzero_si128());
michael@0:             wide_f1 = _mm_min_epi16(wide_f1, wide_max);
michael@0: 
michael@0:             // final combination
michael@0:             wide_i = _mm_or_si128(wide_i, wide_f1);
michael@0:             _mm_storeu_si128(reinterpret_cast<__m128i*>(xy), wide_i);
michael@0: 
michael@0:             wide_f = _mm_add_epi32(wide_f, wide_d2);
michael@0: 
michael@0:             fx += dx2;
michael@0:             fy += dy2;
michael@0:             xy += 4;
michael@0:             count -= 2;
michael@0:         } // while count >= 2
michael@0:     } // if count >= 2
michael@0: 
michael@0:     while (count-- > 0) {
michael@0:         *xy++ = ClampX_ClampY_pack_filter(fy, maxY, oneY);
michael@0:         fy += dy;
michael@0:         *xy++ = ClampX_ClampY_pack_filter(fx, maxX, oneX);
michael@0:         fx += dx;
michael@0:     }
michael@0: }
michael@0: 
michael@0: /*  SSE version of ClampX_ClampY_nofilter_affine()
michael@0:  *  portable version is in core/SkBitmapProcState_matrix.h
michael@0:  */
michael@0: void ClampX_ClampY_nofilter_affine_SSE2(const SkBitmapProcState& s,
michael@0:                                       uint32_t xy[], int count, int x, int y) {
michael@0:     SkASSERT(s.fInvType & SkMatrix::kAffine_Mask);
michael@0:     SkASSERT((s.fInvType & ~(SkMatrix::kTranslate_Mask |
michael@0:                              SkMatrix::kScale_Mask |
michael@0:                              SkMatrix::kAffine_Mask)) == 0);
michael@0: 
michael@0:     SkPoint srcPt;
michael@0:     s.fInvProc(s.fInvMatrix,
michael@0:                SkIntToScalar(x) + SK_ScalarHalf,
michael@0:                SkIntToScalar(y) + SK_ScalarHalf, &srcPt);
michael@0: 
michael@0:     SkFixed fx = SkScalarToFixed(srcPt.fX);
michael@0:     SkFixed fy = SkScalarToFixed(srcPt.fY);
michael@0:     SkFixed dx = s.fInvSx;
michael@0:     SkFixed dy = s.fInvKy;
michael@0:     int maxX = s.fBitmap->width() - 1;
michael@0:     int maxY = s.fBitmap->height() - 1;
michael@0: 
michael@0:     if (count >= 4 && (maxX <= 0xFFFF)) {
michael@0:         while (((size_t)xy & 0x0F) != 0) {
michael@0:             *xy++ = (SkClampMax(fy >> 16, maxY) << 16) |
michael@0:                                   SkClampMax(fx >> 16, maxX);
michael@0:             fx += dx;
michael@0:             fy += dy;
michael@0:             count--;
michael@0:         }
michael@0: 
michael@0:         SkFixed dx4 = dx * 4;
michael@0:         SkFixed dy4 = dy * 4;
michael@0: 
michael@0:         __m128i wide_fx   = _mm_set_epi32(fx + dx * 3, fx + dx * 2,
michael@0:                                           fx + dx, fx);
michael@0:         __m128i wide_fy   = _mm_set_epi32(fy + dy * 3, fy + dy * 2,
michael@0:                                           fy + dy, fy);
michael@0:         __m128i wide_dx4  = _mm_set1_epi32(dx4);
michael@0:         __m128i wide_dy4  = _mm_set1_epi32(dy4);
michael@0: 
michael@0:         __m128i wide_maxX = _mm_set1_epi32(maxX);
michael@0:         __m128i wide_maxY = _mm_set1_epi32(maxY);
michael@0: 
michael@0:         while (count >= 4) {
michael@0:             // SkClampMax(fx>>16,maxX)
michael@0:             __m128i wide_lo = _mm_max_epi16(_mm_srli_epi32(wide_fx, 16),
michael@0:                                             _mm_setzero_si128());
michael@0:             wide_lo = _mm_min_epi16(wide_lo, wide_maxX);
michael@0: 
michael@0:             // SkClampMax(fy>>16,maxY)
michael@0:             __m128i wide_hi = _mm_max_epi16(_mm_srli_epi32(wide_fy, 16),
michael@0:                                             _mm_setzero_si128());
michael@0:             wide_hi = _mm_min_epi16(wide_hi, wide_maxY);
michael@0: 
michael@0:             // final combination
michael@0:             __m128i wide_i = _mm_or_si128(_mm_slli_epi32(wide_hi, 16),
michael@0:                                           wide_lo);
michael@0:             _mm_store_si128(reinterpret_cast<__m128i*>(xy), wide_i);
michael@0: 
michael@0:             wide_fx = _mm_add_epi32(wide_fx, wide_dx4);
michael@0:             wide_fy = _mm_add_epi32(wide_fy, wide_dy4);
michael@0: 
michael@0:             fx += dx4;
michael@0:             fy += dy4;
michael@0:             xy += 4;
michael@0:             count -= 4;
michael@0:         } // while count >= 4
michael@0:     } // if count >= 4
michael@0: 
michael@0:     while (count-- > 0) {
michael@0:         *xy++ = (SkClampMax(fy >> 16, maxY) << 16) |
michael@0:                               SkClampMax(fx >> 16, maxX);
michael@0:         fx += dx;
michael@0:         fy += dy;
michael@0:     }
michael@0: }
michael@0: 
michael@0: /*  SSE version of S32_D16_filter_DX_SSE2
michael@0:  *  Definition is in section of "D16 functions for SRC == 8888" in SkBitmapProcState.cpp
michael@0:  *  It combines S32_opaque_D32_filter_DX_SSE2 and SkPixel32ToPixel16
michael@0:  */
michael@0: void S32_D16_filter_DX_SSE2(const SkBitmapProcState& s,
michael@0:                                    const uint32_t* xy,
michael@0:                                    int count, uint16_t* colors) {
michael@0:     SkASSERT(count > 0 && colors != NULL);
michael@0:     SkASSERT(s.fFilterLevel != SkPaint::kNone_FilterLevel);
michael@0:     SkASSERT(s.fBitmap->config() == SkBitmap::kARGB_8888_Config);
michael@0:     SkASSERT(s.fBitmap->isOpaque());
michael@0: 
michael@0:     SkPMColor dstColor;
michael@0:     const char* srcAddr = static_cast<const char*>(s.fBitmap->getPixels());
michael@0:     size_t rb = s.fBitmap->rowBytes();
michael@0:     uint32_t XY = *xy++;
michael@0:     unsigned y0 = XY >> 14;
michael@0:     const uint32_t* row0 = reinterpret_cast<const uint32_t*>(srcAddr + (y0 >> 4) * rb);
michael@0:     const uint32_t* row1 = reinterpret_cast<const uint32_t*>(srcAddr + (XY & 0x3FFF) * rb);
michael@0:     unsigned subY = y0 & 0xF;
michael@0: 
michael@0:     // ( 0,  0,  0,  0,  0,  0,  0, 16)
michael@0:     __m128i sixteen = _mm_cvtsi32_si128(16);
michael@0: 
michael@0:     // ( 0,  0,  0,  0, 16, 16, 16, 16)
michael@0:     sixteen = _mm_shufflelo_epi16(sixteen, 0);
michael@0: 
michael@0:     // ( 0,  0,  0,  0,  0,  0,  0,  y)
michael@0:     __m128i allY = _mm_cvtsi32_si128(subY);
michael@0: 
michael@0:     // ( 0,  0,  0,  0,  y,  y,  y,  y)
michael@0:     allY = _mm_shufflelo_epi16(allY, 0);
michael@0: 
michael@0:     // ( 0,  0,  0,  0, 16-y, 16-y, 16-y, 16-y)
michael@0:     __m128i negY = _mm_sub_epi16(sixteen, allY);
michael@0: 
michael@0:     // (16-y, 16-y, 16-y, 16-y, y, y, y, y)
michael@0:     allY = _mm_unpacklo_epi64(allY, negY);
michael@0: 
michael@0:     // (16, 16, 16, 16, 16, 16, 16, 16 )
michael@0:     sixteen = _mm_shuffle_epi32(sixteen, 0);
michael@0: 
michael@0:     // ( 0,  0,  0,  0,  0,  0,  0,  0)
michael@0:     __m128i zero = _mm_setzero_si128();
michael@0: 
michael@0:     do {
michael@0:         uint32_t XX = *xy++;    // x0:14 | 4 | x1:14
michael@0:         unsigned x0 = XX >> 18;
michael@0:         unsigned x1 = XX & 0x3FFF;
michael@0: 
michael@0:         // (0, 0, 0, 0, 0, 0, 0, x)
michael@0:         __m128i allX = _mm_cvtsi32_si128((XX >> 14) & 0x0F);
michael@0: 
michael@0:         // (0, 0, 0, 0, x, x, x, x)
michael@0:         allX = _mm_shufflelo_epi16(allX, 0);
michael@0: 
michael@0:         // (x, x, x, x, x, x, x, x)
michael@0:         allX = _mm_shuffle_epi32(allX, 0);
michael@0: 
michael@0:         // (16-x, 16-x, 16-x, 16-x, 16-x, 16-x, 16-x)
michael@0:         __m128i negX = _mm_sub_epi16(sixteen, allX);
michael@0: 
michael@0:         // Load 4 samples (pixels).
michael@0:         __m128i a00 = _mm_cvtsi32_si128(row0[x0]);
michael@0:         __m128i a01 = _mm_cvtsi32_si128(row0[x1]);
michael@0:         __m128i a10 = _mm_cvtsi32_si128(row1[x0]);
michael@0:         __m128i a11 = _mm_cvtsi32_si128(row1[x1]);
michael@0: 
michael@0:         // (0, 0, a00, a10)
michael@0:         __m128i a00a10 = _mm_unpacklo_epi32(a10, a00);
michael@0: 
michael@0:         // Expand to 16 bits per component.
michael@0:         a00a10 = _mm_unpacklo_epi8(a00a10, zero);
michael@0: 
michael@0:         // ((a00 * (16-y)), (a10 * y)).
michael@0:         a00a10 = _mm_mullo_epi16(a00a10, allY);
michael@0: 
michael@0:         // (a00 * (16-y) * (16-x), a10 * y * (16-x)).
michael@0:         a00a10 = _mm_mullo_epi16(a00a10, negX);
michael@0: 
michael@0:         // (0, 0, a01, a10)
michael@0:         __m128i a01a11 = _mm_unpacklo_epi32(a11, a01);
michael@0: 
michael@0:         // Expand to 16 bits per component.
michael@0:         a01a11 = _mm_unpacklo_epi8(a01a11, zero);
michael@0: 
michael@0:         // (a01 * (16-y)), (a11 * y)
michael@0:         a01a11 = _mm_mullo_epi16(a01a11, allY);
michael@0: 
michael@0:         // (a01 * (16-y) * x), (a11 * y * x)
michael@0:         a01a11 = _mm_mullo_epi16(a01a11, allX);
michael@0: 
michael@0:         // (a00*w00 + a01*w01, a10*w10 + a11*w11)
michael@0:         __m128i sum = _mm_add_epi16(a00a10, a01a11);
michael@0: 
michael@0:         // (DC, a00*w00 + a01*w01)
michael@0:         __m128i shifted = _mm_shuffle_epi32(sum, 0xEE);
michael@0: 
michael@0:         // (DC, a00*w00 + a01*w01 + a10*w10 + a11*w11)
michael@0:         sum = _mm_add_epi16(sum, shifted);
michael@0: 
michael@0:         // Divide each 16 bit component by 256.
michael@0:         sum = _mm_srli_epi16(sum, 8);
michael@0: 
michael@0:         // Pack lower 4 16 bit values of sum into lower 4 bytes.
michael@0:         sum = _mm_packus_epi16(sum, zero);
michael@0: 
michael@0:         // Extract low int and store.
michael@0:         dstColor = _mm_cvtsi128_si32(sum);
michael@0: 
michael@0:         //*colors++ = SkPixel32ToPixel16(dstColor);
michael@0:         // below is much faster than the above. It's tested for Android benchmark--Softweg
michael@0:         __m128i _m_temp1 = _mm_set1_epi32(dstColor);
michael@0:         __m128i _m_temp2 = _mm_srli_epi32(_m_temp1, 3);
michael@0: 
michael@0:         unsigned int r32 = _mm_cvtsi128_si32(_m_temp2);
michael@0:         unsigned r = (r32 & ((1<<5) -1)) << 11;
michael@0: 
michael@0:         _m_temp2 = _mm_srli_epi32(_m_temp2, 7);
michael@0:         unsigned int g32 = _mm_cvtsi128_si32(_m_temp2);
michael@0:         unsigned g = (g32 & ((1<<6) -1)) << 5;
michael@0: 
michael@0:         _m_temp2 = _mm_srli_epi32(_m_temp2, 9);
michael@0:         unsigned int b32 = _mm_cvtsi128_si32(_m_temp2);
michael@0:         unsigned b = (b32 & ((1<<5) -1));
michael@0: 
michael@0:         *colors++ = r | g | b;
michael@0: 
michael@0:     } while (--count > 0);
michael@0: }