1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/gfx/skia/trunk/src/opts/SkBlurImage_opts_neon.cpp Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,188 @@
1.4 +/*
1.5 + * Copyright 2013 The Android Open Source Project
1.6 + *
1.7 + * Use of this source code is governed by a BSD-style license that can be
1.8 + * found in the LICENSE file.
1.9 + */
1.10 +
1.11 +
1.12 +#include "SkBitmap.h"
1.13 +#include "SkColorPriv.h"
1.14 +#include "SkBlurImage_opts.h"
1.15 +#include "SkRect.h"
1.16 +
1.17 +#include <arm_neon.h>
1.18 +
1.19 +namespace {
1.20 +
1.21 +enum BlurDirection {
1.22 + kX, kY
1.23 +};
1.24 +
1.25 +/**
1.26 + * Helper function to load 2 pixels from diffent rows to a 8x8 NEON register
1.27 + * and also pre-load pixels for future read
1.28 + */
1.29 +template<BlurDirection srcDirection>
1.30 +inline uint8x8_t load_2_pixels(const SkPMColor* src, int srcStride) {
1.31 + if (srcDirection == kX) {
1.32 + uint32x2_t temp = vdup_n_u32(0);
1.33 + // 10% faster by adding these 2 prefetches
1.34 + SK_PREFETCH(src + 16);
1.35 + SK_PREFETCH(src + srcStride + 16);
1.36 + return vreinterpret_u8_u32(vld1_lane_u32(src + srcStride, vld1_lane_u32(src, temp, 0), 1));
1.37 + } else {
1.38 + return vld1_u8((uint8_t*)src);
1.39 + }
1.40 +}
1.41 +
1.42 +/**
1.43 + * Helper function to store the low 8-bits from a 16x8 NEON register to 2 rows
1.44 + */
1.45 +template<BlurDirection dstDirection>
1.46 +inline void store_2_pixels(uint16x8_t result16x8, SkPMColor* dst, int dstStride) {
1.47 + if (dstDirection == kX) {
1.48 + uint32x2_t temp = vreinterpret_u32_u8(vmovn_u16(result16x8));
1.49 + vst1_lane_u32(dst, temp, 0);
1.50 + vst1_lane_u32(dst + dstStride, temp, 1);
1.51 + } else {
1.52 + uint8x8_t temp = vmovn_u16(result16x8);
1.53 + vst1_u8((uint8_t*)dst, temp);
1.54 + }
1.55 +}
1.56 +
1.57 +/**
1.58 + * fast path for kernel size less than 128
1.59 + */
1.60 +template<BlurDirection srcDirection, BlurDirection dstDirection>
1.61 +void SkDoubleRowBoxBlur_NEON(const SkPMColor** src, int srcStride, SkPMColor** dst, int kernelSize,
1.62 + int leftOffset, int rightOffset, int width, int* height)
1.63 +{
1.64 + const int rightBorder = SkMin32(rightOffset + 1, width);
1.65 + const int srcStrideX = srcDirection == kX ? 1 : srcStride;
1.66 + const int dstStrideX = dstDirection == kX ? 1 : *height;
1.67 + const int srcStrideY = srcDirection == kX ? srcStride : 1;
1.68 + const int dstStrideY = dstDirection == kX ? width : 1;
1.69 + const uint16x8_t scale = vdupq_n_u16((1 << 15) / kernelSize);
1.70 +
1.71 + for (; *height >= 2; *height -= 2) {
1.72 + uint16x8_t sum = vdupq_n_u16(0);
1.73 + const SkPMColor* p = *src;
1.74 + for (int i = 0; i < rightBorder; i++) {
1.75 + sum = vaddw_u8(sum,
1.76 + load_2_pixels<srcDirection>(p, srcStride));
1.77 + p += srcStrideX;
1.78 + }
1.79 +
1.80 + const SkPMColor* sptr = *src;
1.81 + SkPMColor* dptr = *dst;
1.82 + for (int x = 0; x < width; x++) {
1.83 + // val = (sum * scale * 2 + 0x8000) >> 16
1.84 + uint16x8_t resultPixels = vreinterpretq_u16_s16(vqrdmulhq_s16(
1.85 + vreinterpretq_s16_u16(sum), vreinterpretq_s16_u16(scale)));
1.86 + store_2_pixels<dstDirection>(resultPixels, dptr, width);
1.87 +
1.88 + if (x >= leftOffset) {
1.89 + sum = vsubw_u8(sum,
1.90 + load_2_pixels<srcDirection>(sptr - leftOffset * srcStrideX, srcStride));
1.91 + }
1.92 + if (x + rightOffset + 1 < width) {
1.93 + sum = vaddw_u8(sum,
1.94 + load_2_pixels<srcDirection>(sptr + (rightOffset + 1) * srcStrideX, srcStride));
1.95 + }
1.96 + sptr += srcStrideX;
1.97 + dptr += dstStrideX;
1.98 + }
1.99 + *src += srcStrideY * 2;
1.100 + *dst += dstStrideY * 2;
1.101 + }
1.102 +}
1.103 +
1.104 +
1.105 +/**
1.106 + * Helper function to spread the components of a 32-bit integer into the
1.107 + * lower 8 bits of each 16-bit element of a NEON register.
1.108 + */
1.109 +
1.110 +static inline uint16x4_t expand(uint32_t a) {
1.111 + // ( ARGB ) -> ( ARGB ARGB ) -> ( A R G B A R G B )
1.112 + uint8x8_t v8 = vreinterpret_u8_u32(vdup_n_u32(a));
1.113 + // ( A R G B A R G B ) -> ( 0A 0R 0G 0B 0A 0R 0G 0B ) -> ( 0A 0R 0G 0B )
1.114 + return vget_low_u16(vmovl_u8(v8));
1.115 +}
1.116 +
1.117 +template<BlurDirection srcDirection, BlurDirection dstDirection>
1.118 +void SkBoxBlur_NEON(const SkPMColor* src, int srcStride, SkPMColor* dst, int kernelSize,
1.119 + int leftOffset, int rightOffset, int width, int height)
1.120 +{
1.121 + const int rightBorder = SkMin32(rightOffset + 1, width);
1.122 + const int srcStrideX = srcDirection == kX ? 1 : srcStride;
1.123 + const int dstStrideX = dstDirection == kX ? 1 : height;
1.124 + const int srcStrideY = srcDirection == kX ? srcStride : 1;
1.125 + const int dstStrideY = dstDirection == kX ? width : 1;
1.126 + const uint32x4_t scale = vdupq_n_u32((1 << 24) / kernelSize);
1.127 + const uint32x4_t half = vdupq_n_u32(1 << 23);
1.128 +
1.129 + if (kernelSize < 128)
1.130 + {
1.131 + SkDoubleRowBoxBlur_NEON<srcDirection, dstDirection>(&src, srcStride, &dst, kernelSize,
1.132 + leftOffset, rightOffset, width, &height);
1.133 + }
1.134 +
1.135 + for (; height > 0; height--) {
1.136 + uint32x4_t sum = vdupq_n_u32(0);
1.137 + const SkPMColor* p = src;
1.138 + for (int i = 0; i < rightBorder; ++i) {
1.139 + sum = vaddw_u16(sum, expand(*p));
1.140 + p += srcStrideX;
1.141 + }
1.142 +
1.143 + const SkPMColor* sptr = src;
1.144 + SkPMColor* dptr = dst;
1.145 + for (int x = 0; x < width; ++x) {
1.146 + // ( half+sumA*scale half+sumR*scale half+sumG*scale half+sumB*scale )
1.147 + uint32x4_t result = vmlaq_u32(half, sum, scale);
1.148 +
1.149 + // Saturated conversion to 16-bit.
1.150 + // ( AAAA RRRR GGGG BBBB ) -> ( 0A 0R 0G 0B )
1.151 + uint16x4_t result16 = vqshrn_n_u32(result, 16);
1.152 +
1.153 + // Saturated conversion to 8-bit.
1.154 + // ( 0A 0R 0G 0B ) -> ( 0A 0R 0G 0B 0A 0R 0G 0B ) -> ( A R G B A R G B )
1.155 + uint8x8_t result8 = vqshrn_n_u16(vcombine_u16(result16, result16), 8);
1.156 +
1.157 + // ( A R G B A R G B ) -> ( ARGB ARGB ) -> ( ARGB )
1.158 + // Store low 32 bits to destination.
1.159 + vst1_lane_u32(dptr, vreinterpret_u32_u8(result8), 0);
1.160 +
1.161 + if (x >= leftOffset) {
1.162 + const SkPMColor* l = sptr - leftOffset * srcStrideX;
1.163 + sum = vsubw_u16(sum, expand(*l));
1.164 + }
1.165 + if (x + rightOffset + 1 < width) {
1.166 + const SkPMColor* r = sptr + (rightOffset + 1) * srcStrideX;
1.167 + sum = vaddw_u16(sum, expand(*r));
1.168 + }
1.169 + sptr += srcStrideX;
1.170 + if (srcDirection == kX) {
1.171 + SK_PREFETCH(sptr + (rightOffset + 16) * srcStrideX);
1.172 + }
1.173 + dptr += dstStrideX;
1.174 + }
1.175 + src += srcStrideY;
1.176 + dst += dstStrideY;
1.177 + }
1.178 +}
1.179 +
1.180 +} // namespace
1.181 +
1.182 +bool SkBoxBlurGetPlatformProcs_NEON(SkBoxBlurProc* boxBlurX,
1.183 + SkBoxBlurProc* boxBlurY,
1.184 + SkBoxBlurProc* boxBlurXY,
1.185 + SkBoxBlurProc* boxBlurYX) {
1.186 + *boxBlurX = SkBoxBlur_NEON<kX, kX>;
1.187 + *boxBlurY = SkBoxBlur_NEON<kY, kY>;
1.188 + *boxBlurXY = SkBoxBlur_NEON<kX, kY>;
1.189 + *boxBlurYX = SkBoxBlur_NEON<kY, kX>;
1.190 + return true;
1.191 +}
