1 /* 2 * Copyright 2013 The Android Open Source Project 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 9 #include "SkBitmap.h" 10 #include "SkColorPriv.h" 11 #include "SkBlurImage_opts.h" 12 #include "SkRect.h" 13 14 #include <arm_neon.h> 15 16 namespace { 17 18 enum BlurDirection { 19 kX, kY 20 }; 21 22 /** 23 * Helper function to spread the components of a 32-bit integer into the 24 * lower 8 bits of each 16-bit element of a NEON register. 25 */ 26 27 static inline uint16x4_t expand(uint32_t a) { 28 // ( ARGB ) -> ( ARGB ARGB ) -> ( A R G B A R G B ) 29 uint8x8_t v8 = vreinterpret_u8_u32(vdup_n_u32(a)); 30 // ( A R G B A R G B ) -> ( 0A 0R 0G 0B 0A 0R 0G 0B ) -> ( 0A 0R 0G 0B ) 31 return vget_low_u16(vmovl_u8(v8)); 32 } 33 34 template<BlurDirection srcDirection, BlurDirection dstDirection> 35 void SkBoxBlur_NEON(const SkPMColor* src, int srcStride, SkPMColor* dst, int kernelSize, 36 int leftOffset, int rightOffset, int width, int height) 37 { 38 const int rightBorder = SkMin32(rightOffset + 1, width); 39 const int srcStrideX = srcDirection == kX ? 1 : srcStride; 40 const int dstStrideX = dstDirection == kX ? 1 : height; 41 const int srcStrideY = srcDirection == kX ? srcStride : 1; 42 const int dstStrideY = dstDirection == kX ? width : 1; 43 const uint32x4_t scale = vdupq_n_u32((1 << 24) / kernelSize); 44 const uint32x4_t half = vdupq_n_u32(1 << 23); 45 for (int y = 0; y < height; ++y) { 46 uint32x4_t sum = vdupq_n_u32(0); 47 const SkPMColor* p = src; 48 for (int i = 0; i < rightBorder; ++i) { 49 sum = vaddw_u16(sum, expand(*p)); 50 p += srcStrideX; 51 } 52 53 const SkPMColor* sptr = src; 54 SkPMColor* dptr = dst; 55 for (int x = 0; x < width; ++x) { 56 // ( half+sumA*scale half+sumR*scale half+sumG*scale half+sumB*scale ) 57 uint32x4_t result = vmlaq_u32(half, sum, scale); 58 59 // Saturated conversion to 16-bit. 60 // ( AAAA RRRR GGGG BBBB ) -> ( 0A 0R 0G 0B ) 61 uint16x4_t result16 = vqshrn_n_u32(result, 16); 62 63 // Saturated conversion to 8-bit. 64 // ( 0A 0R 0G 0B ) -> ( 0A 0R 0G 0B 0A 0R 0G 0B ) -> ( A R G B A R G B ) 65 uint8x8_t result8 = vqshrn_n_u16(vcombine_u16(result16, result16), 8); 66 67 // ( A R G B A R G B ) -> ( ARGB ARGB ) -> ( ARGB ) 68 // Store low 32 bits to destination. 69 vst1_lane_u32(dptr, vreinterpret_u32_u8(result8), 0); 70 71 if (x >= leftOffset) { 72 const SkPMColor* l = sptr - leftOffset * srcStrideX; 73 sum = vsubw_u16(sum, expand(*l)); 74 } 75 if (x + rightOffset + 1 < width) { 76 const SkPMColor* r = sptr + (rightOffset + 1) * srcStrideX; 77 sum = vaddw_u16(sum, expand(*r)); 78 } 79 sptr += srcStrideX; 80 if (srcDirection == kY) { 81 SK_PREFETCH(sptr + (rightOffset + 1) * srcStrideX); 82 } 83 dptr += dstStrideX; 84 } 85 src += srcStrideY; 86 dst += dstStrideY; 87 } 88 } 89 90 } // namespace 91 92 bool SkBoxBlurGetPlatformProcs_NEON(SkBoxBlurProc* boxBlurX, 93 SkBoxBlurProc* boxBlurY, 94 SkBoxBlurProc* boxBlurXY, 95 SkBoxBlurProc* boxBlurYX) { 96 *boxBlurX = SkBoxBlur_NEON<kX, kX>; 97 *boxBlurY = SkBoxBlur_NEON<kY, kY>; 98 *boxBlurXY = SkBoxBlur_NEON<kX, kY>; 99 *boxBlurYX = SkBoxBlur_NEON<kY, kX>; 100 return true; 101 } 102