NEON fast path for box blur

Calculate 8 channels in parallel by using 16-bits to store each channel. Due to the limitation of VQRDMULH, (int16 * int16 * 2 + 0x8000) >> 16, the fast path can only support kernelSize < 128.
8 significant bits are kept at least in each stage, the final error should less-equal than 1.

Pre-fetching memory for X-direction read. In fact pre-fetching memory doesn't help much for Y direction read, since it is a waste to load a cache line for only read 8 bytes.(I left it there to keep the symmetry. pre-fetch is cheap :) )

bench data on Nexus 10
before:
running bench [640 480]      blur_image_filter_large_10.00_10.00   8888:  cmsecs =  25081.48
running bench [640 480]      blur_image_filter_small_10.00_10.00   8888:  cmsecs =  25038.04
running bench [640 480]        blur_image_filter_large_1.00_1.00   8888:  cmsecs =  25209.04
running bench [640 480]        blur_image_filter_small_1.00_1.00   8888:  cmsecs =  24928.01
running bench [640 480]        blur_image_filter_large_0.00_1.00   8888:  cmsecs =  17160.98
running bench [640 480]       blur_image_filter_large_0.00_10.00   8888:  cmsecs =  17924.11
running bench [640 480]        blur_image_filter_large_1.00_0.00   8888:  cmsecs =  14609.19
running bench [640 480]       blur_image_filter_large_10.00_0.00   8888:  cmsecs =  14625.91

after:
running bench [640 480]      blur_image_filter_large_10.00_10.00   8888:  cmsecs =  14848.42
running bench [640 480]      blur_image_filter_small_10.00_10.00   8888:  cmsecs =  16037.29
running bench [640 480]        blur_image_filter_large_1.00_1.00   8888:  cmsecs =  14819.55
running bench [640 480]        blur_image_filter_small_1.00_1.00   8888:  cmsecs =  14563.69
running bench [640 480]        blur_image_filter_large_0.00_1.00   8888:  cmsecs =  11905.34
running bench [640 480]       blur_image_filter_large_0.00_10.00   8888:  cmsecs =  11883.85
running bench [640 480]        blur_image_filter_large_1.00_0.00   8888:  cmsecs =   9576.51
running bench [640 480]       blur_image_filter_large_10.00_0.00   8888:  cmsecs =   9793.84

BUG=
R=senorblanco@chromium.org, mtklein@google.com, reed@google.com, kevin.petit@arm.com, kevin.petit.arm@gmail.com

Author: zheng.xu@arm.com

Review URL: https://codereview.chromium.org/105893003

git-svn-id: http://skia.googlecode.com/svn/trunk@13036 2bbb7eff-a529-9590-31e7-b0007b416f81

1 files changed, 90 insertions, 3 deletions

diff --git a/src/opts/SkBlurImage_opts_neon.cpp b/src/opts/SkBlurImage_opts_neon.cpp
index 4e33d72d46..ebe17f05b4 100644
--- a/src/opts/SkBlurImage_opts_neon.cpp
+++ b/src/opts/SkBlurImage_opts_neon.cpp
@@ -20,6 +20,86 @@ enum BlurDirection {
 };
 
 /**
+ * Helper function to load 2 pixels from diffent rows to a 8x8 NEON register
+ * and also pre-load pixels for future read
+ */
+template<BlurDirection srcDirection>
+inline uint8x8_t load_2_pixels(const SkPMColor* src, int srcStride) {
+    if (srcDirection == kX) {
+        uint32x2_t temp = vdup_n_u32(0);
+        // 10% faster by adding these 2 prefetches
+        SK_PREFETCH(src + 16);
+        SK_PREFETCH(src + srcStride + 16);
+        return vreinterpret_u8_u32(vld1_lane_u32(src + srcStride, vld1_lane_u32(src, temp, 0), 1));
+     } else {
+         return vld1_u8((uint8_t*)src);
+     }
+}
+
+/**
+ * Helper function to store the low 8-bits from a 16x8 NEON register to 2 rows
+ */
+template<BlurDirection dstDirection>
+inline void store_2_pixels(uint16x8_t result16x8, SkPMColor* dst, int dstStride) {
+    if (dstDirection == kX) {
+        uint32x2_t temp = vreinterpret_u32_u8(vmovn_u16(result16x8));
+        vst1_lane_u32(dst, temp, 0);
+        vst1_lane_u32(dst + dstStride, temp, 1);
+    } else {
+        uint8x8_t temp = vmovn_u16(result16x8);
+        vst1_u8((uint8_t*)dst, temp);
+    }
+}
+
+/**
+ * fast path for kernel size less than 128
+ */
+template<BlurDirection srcDirection, BlurDirection dstDirection>
+void SkDoubleRowBoxBlur_NEON(const SkPMColor** src, int srcStride, SkPMColor** dst, int kernelSize,
+                        int leftOffset, int rightOffset, int width, int* height)
+{
+    const int rightBorder = SkMin32(rightOffset + 1, width);
+    const int srcStrideX = srcDirection == kX ? 1 : srcStride;
+    const int dstStrideX = dstDirection == kX ? 1 : *height;
+    const int srcStrideY = srcDirection == kX ? srcStride : 1;
+    const int dstStrideY = dstDirection == kX ? width : 1;
+    const uint16x8_t scale = vdupq_n_u16((1 << 15) / kernelSize);
+
+    for (; *height >= 2; *height -= 2) {
+        uint16x8_t sum = vdupq_n_u16(0);
+        const SkPMColor* p = *src;
+        for (int i = 0; i < rightBorder; i++) {
+            sum = vaddw_u8(sum,
+                load_2_pixels<srcDirection>(p, srcStride));
+            p += srcStrideX;
+        }
+
+        const SkPMColor* sptr = *src;
+        SkPMColor* dptr = *dst;
+        for (int x = 0; x < width; x++) {
+            // val = (sum * scale * 2 + 0x8000) >> 16
+            uint16x8_t resultPixels = vreinterpretq_u16_s16(vqrdmulhq_s16(
+                vreinterpretq_s16_u16(sum), vreinterpretq_s16_u16(scale)));
+            store_2_pixels<dstDirection>(resultPixels, dptr, width);
+
+            if (x >= leftOffset) {
+                sum = vsubw_u8(sum, 
+                    load_2_pixels<srcDirection>(sptr - leftOffset * srcStrideX, srcStride));
+            }
+            if (x + rightOffset + 1 < width) {
+                sum = vaddw_u8(sum, 
+                    load_2_pixels<srcDirection>(sptr + (rightOffset + 1) * srcStrideX, srcStride));
+            }
+            sptr += srcStrideX;
+            dptr += dstStrideX;
+        }
+        *src += srcStrideY * 2;
+        *dst += dstStrideY * 2;
+    }
+}
+
+
+/**
  * Helper function to spread the components of a 32-bit integer into the
  * lower 8 bits of each 16-bit element of a NEON register.
  */
@@ -42,7 +122,14 @@ void SkBoxBlur_NEON(const SkPMColor* src, int srcStride, SkPMColor* dst, int ker
     const int dstStrideY = dstDirection == kX ? width : 1;
     const uint32x4_t scale = vdupq_n_u32((1 << 24) / kernelSize);
     const uint32x4_t half = vdupq_n_u32(1 << 23);
-    for (int y = 0; y < height; ++y) {
+
+    if (kernelSize < 128)
+    {
+        SkDoubleRowBoxBlur_NEON<srcDirection, dstDirection>(&src, srcStride, &dst, kernelSize,
+            leftOffset, rightOffset, width, &height);
+    }
+
+    for (; height > 0; height--) {
         uint32x4_t sum = vdupq_n_u32(0);
         const SkPMColor* p = src;
         for (int i = 0; i < rightBorder; ++i) {
@@ -77,8 +164,8 @@ void SkBoxBlur_NEON(const SkPMColor* src, int srcStride, SkPMColor* dst, int ker
                 sum = vaddw_u16(sum, expand(*r));
             }
             sptr += srcStrideX;
-            if (srcDirection == kY) {
-                SK_PREFETCH(sptr + (rightOffset + 1) * srcStrideX);
+            if (srcDirection == kX) {
+                SK_PREFETCH(sptr + (rightOffset + 16) * srcStrideX);
             }
             dptr += dstStrideX;
         }