michael@0: /*
michael@0:  * Copyright 2013 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 "SkBitmap.h"
michael@0: #include "SkColorPriv.h"
michael@0: #include "SkBlurImage_opts.h"
michael@0: #include "SkRect.h"
michael@0: 
michael@0: #include <arm_neon.h>
michael@0: 
michael@0: namespace {
michael@0: 
michael@0: enum BlurDirection {
michael@0:     kX, kY
michael@0: };
michael@0: 
michael@0: /**
michael@0:  * Helper function to load 2 pixels from diffent rows to a 8x8 NEON register
michael@0:  * and also pre-load pixels for future read
michael@0:  */
michael@0: template<BlurDirection srcDirection>
michael@0: inline uint8x8_t load_2_pixels(const SkPMColor* src, int srcStride) {
michael@0:     if (srcDirection == kX) {
michael@0:         uint32x2_t temp = vdup_n_u32(0);
michael@0:         // 10% faster by adding these 2 prefetches
michael@0:         SK_PREFETCH(src + 16);
michael@0:         SK_PREFETCH(src + srcStride + 16);
michael@0:         return vreinterpret_u8_u32(vld1_lane_u32(src + srcStride, vld1_lane_u32(src, temp, 0), 1));
michael@0:      } else {
michael@0:          return vld1_u8((uint8_t*)src);
michael@0:      }
michael@0: }
michael@0: 
michael@0: /**
michael@0:  * Helper function to store the low 8-bits from a 16x8 NEON register to 2 rows
michael@0:  */
michael@0: template<BlurDirection dstDirection>
michael@0: inline void store_2_pixels(uint16x8_t result16x8, SkPMColor* dst, int dstStride) {
michael@0:     if (dstDirection == kX) {
michael@0:         uint32x2_t temp = vreinterpret_u32_u8(vmovn_u16(result16x8));
michael@0:         vst1_lane_u32(dst, temp, 0);
michael@0:         vst1_lane_u32(dst + dstStride, temp, 1);
michael@0:     } else {
michael@0:         uint8x8_t temp = vmovn_u16(result16x8);
michael@0:         vst1_u8((uint8_t*)dst, temp);
michael@0:     }
michael@0: }
michael@0: 
michael@0: /**
michael@0:  * fast path for kernel size less than 128
michael@0:  */
michael@0: template<BlurDirection srcDirection, BlurDirection dstDirection>
michael@0: void SkDoubleRowBoxBlur_NEON(const SkPMColor** src, int srcStride, SkPMColor** dst, int kernelSize,
michael@0:                         int leftOffset, int rightOffset, int width, int* height)
michael@0: {
michael@0:     const int rightBorder = SkMin32(rightOffset + 1, width);
michael@0:     const int srcStrideX = srcDirection == kX ? 1 : srcStride;
michael@0:     const int dstStrideX = dstDirection == kX ? 1 : *height;
michael@0:     const int srcStrideY = srcDirection == kX ? srcStride : 1;
michael@0:     const int dstStrideY = dstDirection == kX ? width : 1;
michael@0:     const uint16x8_t scale = vdupq_n_u16((1 << 15) / kernelSize);
michael@0: 
michael@0:     for (; *height >= 2; *height -= 2) {
michael@0:         uint16x8_t sum = vdupq_n_u16(0);
michael@0:         const SkPMColor* p = *src;
michael@0:         for (int i = 0; i < rightBorder; i++) {
michael@0:             sum = vaddw_u8(sum,
michael@0:                 load_2_pixels<srcDirection>(p, srcStride));
michael@0:             p += srcStrideX;
michael@0:         }
michael@0: 
michael@0:         const SkPMColor* sptr = *src;
michael@0:         SkPMColor* dptr = *dst;
michael@0:         for (int x = 0; x < width; x++) {
michael@0:             // val = (sum * scale * 2 + 0x8000) >> 16
michael@0:             uint16x8_t resultPixels = vreinterpretq_u16_s16(vqrdmulhq_s16(
michael@0:                 vreinterpretq_s16_u16(sum), vreinterpretq_s16_u16(scale)));
michael@0:             store_2_pixels<dstDirection>(resultPixels, dptr, width);
michael@0: 
michael@0:             if (x >= leftOffset) {
michael@0:                 sum = vsubw_u8(sum,
michael@0:                     load_2_pixels<srcDirection>(sptr - leftOffset * srcStrideX, srcStride));
michael@0:             }
michael@0:             if (x + rightOffset + 1 < width) {
michael@0:                 sum = vaddw_u8(sum,
michael@0:                     load_2_pixels<srcDirection>(sptr + (rightOffset + 1) * srcStrideX, srcStride));
michael@0:             }
michael@0:             sptr += srcStrideX;
michael@0:             dptr += dstStrideX;
michael@0:         }
michael@0:         *src += srcStrideY * 2;
michael@0:         *dst += dstStrideY * 2;
michael@0:     }
michael@0: }
michael@0: 
michael@0: 
michael@0: /**
michael@0:  * Helper function to spread the components of a 32-bit integer into the
michael@0:  * lower 8 bits of each 16-bit element of a NEON register.
michael@0:  */
michael@0: 
michael@0: static inline uint16x4_t expand(uint32_t a) {
michael@0:     // ( ARGB ) -> ( ARGB ARGB ) -> ( A R G B A R G B )
michael@0:     uint8x8_t v8 = vreinterpret_u8_u32(vdup_n_u32(a));
michael@0:     // ( A R G B A R G B ) -> ( 0A 0R 0G 0B 0A 0R 0G 0B ) -> ( 0A 0R 0G 0B )
michael@0:     return vget_low_u16(vmovl_u8(v8));
michael@0: }
michael@0: 
michael@0: template<BlurDirection srcDirection, BlurDirection dstDirection>
michael@0: void SkBoxBlur_NEON(const SkPMColor* src, int srcStride, SkPMColor* dst, int kernelSize,
michael@0:                     int leftOffset, int rightOffset, int width, int height)
michael@0: {
michael@0:     const int rightBorder = SkMin32(rightOffset + 1, width);
michael@0:     const int srcStrideX = srcDirection == kX ? 1 : srcStride;
michael@0:     const int dstStrideX = dstDirection == kX ? 1 : height;
michael@0:     const int srcStrideY = srcDirection == kX ? srcStride : 1;
michael@0:     const int dstStrideY = dstDirection == kX ? width : 1;
michael@0:     const uint32x4_t scale = vdupq_n_u32((1 << 24) / kernelSize);
michael@0:     const uint32x4_t half = vdupq_n_u32(1 << 23);
michael@0: 
michael@0:     if (kernelSize < 128)
michael@0:     {
michael@0:         SkDoubleRowBoxBlur_NEON<srcDirection, dstDirection>(&src, srcStride, &dst, kernelSize,
michael@0:             leftOffset, rightOffset, width, &height);
michael@0:     }
michael@0: 
michael@0:     for (; height > 0; height--) {
michael@0:         uint32x4_t sum = vdupq_n_u32(0);
michael@0:         const SkPMColor* p = src;
michael@0:         for (int i = 0; i < rightBorder; ++i) {
michael@0:             sum = vaddw_u16(sum, expand(*p));
michael@0:             p += srcStrideX;
michael@0:         }
michael@0: 
michael@0:         const SkPMColor* sptr = src;
michael@0:         SkPMColor* dptr = dst;
michael@0:         for (int x = 0; x < width; ++x) {
michael@0:             // ( half+sumA*scale half+sumR*scale half+sumG*scale half+sumB*scale )
michael@0:             uint32x4_t result = vmlaq_u32(half, sum, scale);
michael@0: 
michael@0:             // Saturated conversion to 16-bit.
michael@0:             // ( AAAA RRRR GGGG BBBB ) -> ( 0A 0R 0G 0B )
michael@0:             uint16x4_t result16 = vqshrn_n_u32(result, 16);
michael@0: 
michael@0:             // Saturated conversion to 8-bit.
michael@0:             // ( 0A 0R 0G 0B ) -> ( 0A 0R 0G 0B 0A 0R 0G 0B ) -> ( A R G B A R G B )
michael@0:             uint8x8_t result8 = vqshrn_n_u16(vcombine_u16(result16, result16), 8);
michael@0: 
michael@0:             // ( A R G B A R G B ) -> ( ARGB ARGB ) -> ( ARGB )
michael@0:             // Store low 32 bits to destination.
michael@0:             vst1_lane_u32(dptr, vreinterpret_u32_u8(result8), 0);
michael@0: 
michael@0:             if (x >= leftOffset) {
michael@0:                 const SkPMColor* l = sptr - leftOffset * srcStrideX;
michael@0:                 sum = vsubw_u16(sum, expand(*l));
michael@0:             }
michael@0:             if (x + rightOffset + 1 < width) {
michael@0:                 const SkPMColor* r = sptr + (rightOffset + 1) * srcStrideX;
michael@0:                 sum = vaddw_u16(sum, expand(*r));
michael@0:             }
michael@0:             sptr += srcStrideX;
michael@0:             if (srcDirection == kX) {
michael@0:                 SK_PREFETCH(sptr + (rightOffset + 16) * srcStrideX);
michael@0:             }
michael@0:             dptr += dstStrideX;
michael@0:         }
michael@0:         src += srcStrideY;
michael@0:         dst += dstStrideY;
michael@0:     }
michael@0: }
michael@0: 
michael@0: } // namespace
michael@0: 
michael@0: bool SkBoxBlurGetPlatformProcs_NEON(SkBoxBlurProc* boxBlurX,
michael@0:                                     SkBoxBlurProc* boxBlurY,
michael@0:                                     SkBoxBlurProc* boxBlurXY,
michael@0:                                     SkBoxBlurProc* boxBlurYX) {
michael@0:     *boxBlurX = SkBoxBlur_NEON<kX, kX>;
michael@0:     *boxBlurY = SkBoxBlur_NEON<kY, kY>;
michael@0:     *boxBlurXY = SkBoxBlur_NEON<kX, kY>;
michael@0:     *boxBlurYX = SkBoxBlur_NEON<kY, kX>;
michael@0:     return true;
michael@0: }