NEON optimizations for GrayAlpha -> RGBA/BGRA Premul/Unpremul

PNG Decode Time Nexus 6P (for a test set of GrayAlpha encoded PNGs)
Regular   Unpremul 0.91x
Zero Init Unpremul 0.92x
Regular   Premul   0.84x
Zero Init Premul   0.86x

BUG=skia:4767
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1663623002
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot

Review URL: https://codereview.chromium.org/1663623002

diff --git a/src/codec/SkSwizzler.cpp b/src/codec/SkSwizzler.cpp
index ce6dd26..debcd45 100644 (file)
--- a/src/codec/SkSwizzler.cpp
+++ b/src/codec/SkSwizzler.cpp
@@ -309,6 +309,19 @@ static void swizzle_grayalpha_to_n32_unpremul(
     }
 }
 
+static void fast_swizzle_grayalpha_to_n32_unpremul(
+        void* dst, const uint8_t* src, int width, int bpp, int deltaSrc, int offset,
+        const SkPMColor ctable[]) {
+
+    // This function must not be called if we are sampling.  If we are not
+    // sampling, deltaSrc should equal bpp.
+    SkASSERT(deltaSrc == bpp);
+
+    // Note that there is no need to distinguish between RGB and BGR.
+    // Each color channel will get the same value.
+    SkOpts::grayA_to_RGBA((uint32_t*) dst, src + offset, width);
+}
+
 static void swizzle_grayalpha_to_n32_premul(
         void* dst, const uint8_t* src, int width, int bpp, int deltaSrc, int offset,
         const SkPMColor ctable[]) {
@@ -322,6 +335,19 @@ static void swizzle_grayalpha_to_n32_premul(
     }
 }
 
+static void fast_swizzle_grayalpha_to_n32_premul(
+        void* dst, const uint8_t* src, int width, int bpp, int deltaSrc, int offset,
+        const SkPMColor ctable[]) {
+
+    // This function must not be called if we are sampling.  If we are not
+    // sampling, deltaSrc should equal bpp.
+    SkASSERT(deltaSrc == bpp);
+
+    // Note that there is no need to distinguish between rgb and bgr.
+    // Each color channel will get the same value.
+    SkOpts::grayA_to_rgbA((uint32_t*) dst, src + offset, width);
+}
+
 // kBGRX
 
 static void swizzle_bgrx_to_n32(
@@ -718,14 +744,20 @@ SkSwizzler* SkSwizzler::CreateSwizzler(SkSwizzler::SrcConfig sc,
                         if (SkCodec::kYes_ZeroInitialized == zeroInit) {
                             proc = &SkipLeadingGrayAlphaZerosThen
                                     <swizzle_grayalpha_to_n32_unpremul>;
+                            fastProc = &SkipLeadingGrayAlphaZerosThen
+                                    <fast_swizzle_grayalpha_to_n32_unpremul>;
                         } else {
                             proc = &swizzle_grayalpha_to_n32_unpremul;
+                            fastProc = &fast_swizzle_grayalpha_to_n32_unpremul;
                         }
                     } else {
                         if (SkCodec::kYes_ZeroInitialized == zeroInit) {
                             proc = &SkipLeadingGrayAlphaZerosThen<swizzle_grayalpha_to_n32_premul>;
+                            fastProc = &SkipLeadingGrayAlphaZerosThen
+                                    <fast_swizzle_grayalpha_to_n32_premul>;
                         } else {
                             proc = &swizzle_grayalpha_to_n32_premul;
+                            fastProc = &fast_swizzle_grayalpha_to_n32_premul;
                         }
                     }
                     break;

diff --git a/src/core/SkOpts.cpp b/src/core/SkOpts.cpp
index bce6ee1..8065b77 100644 (file)
--- a/src/core/SkOpts.cpp
+++ b/src/core/SkOpts.cpp
@@ -79,12 +79,14 @@ namespace SkOpts {
     decltype(matrix_scale_translate) matrix_scale_translate = sk_default::matrix_scale_translate;
     decltype(matrix_affine)          matrix_affine          = sk_default::matrix_affine;
 
-    decltype(RGBA_to_BGRA) RGBA_to_BGRA = sk_default::RGBA_to_BGRA;
-    decltype(RGBA_to_rgbA) RGBA_to_rgbA = sk_default::RGBA_to_rgbA;
-    decltype(RGBA_to_bgrA) RGBA_to_bgrA = sk_default::RGBA_to_bgrA;
-    decltype(RGB_to_RGB1)  RGB_to_RGB1  = sk_default::RGB_to_RGB1;
-    decltype(RGB_to_BGR1)  RGB_to_BGR1  = sk_default::RGB_to_BGR1;
-    decltype(gray_to_RGB1) gray_to_RGB1 = sk_default::gray_to_RGB1;
+    decltype(RGBA_to_BGRA)  RGBA_to_BGRA  = sk_default::RGBA_to_BGRA;
+    decltype(RGBA_to_rgbA)  RGBA_to_rgbA  = sk_default::RGBA_to_rgbA;
+    decltype(RGBA_to_bgrA)  RGBA_to_bgrA  = sk_default::RGBA_to_bgrA;
+    decltype(RGB_to_RGB1)   RGB_to_RGB1   = sk_default::RGB_to_RGB1;
+    decltype(RGB_to_BGR1)   RGB_to_BGR1   = sk_default::RGB_to_BGR1;
+    decltype(gray_to_RGB1)  gray_to_RGB1  = sk_default::gray_to_RGB1;
+    decltype(grayA_to_RGBA) grayA_to_RGBA = sk_default::grayA_to_RGBA;
+    decltype(grayA_to_rgbA) grayA_to_rgbA = sk_default::grayA_to_rgbA;
 
     // Each Init_foo() is defined in src/opts/SkOpts_foo.cpp.
     void Init_ssse3();

diff --git a/src/core/SkOpts.h b/src/core/SkOpts.h
index b5286e4..69da362 100644 (file)
--- a/src/core/SkOpts.h
+++ b/src/core/SkOpts.h
@@ -62,7 +62,9 @@ namespace SkOpts {
                         RGBA_to_bgrA,  // i.e. swap RB and premultiply
                         RGB_to_RGB1,   // i.e. insert an opaque alpha
                         RGB_to_BGR1,   // i.e. swap RB and insert an opaque alpha
-                        gray_to_RGB1;  // i.e. set color channels to same value + an opaque alpha
+                        gray_to_RGB1,  // i.e. expand to color channels + an opaque alpha
+                        grayA_to_RGBA, // i.e. expand to color channels
+                        grayA_to_rgbA; // i.e. expand to color channels and premultiply
 }
 
 #endif//SkOpts_DEFINED

diff --git a/src/opts/SkOpts_neon.cpp b/src/opts/SkOpts_neon.cpp
index 79d3140..01f43ef 100644 (file)
--- a/src/opts/SkOpts_neon.cpp
+++ b/src/opts/SkOpts_neon.cpp
@@ -47,11 +47,13 @@ namespace SkOpts {
         matrix_scale_translate = sk_neon::matrix_scale_translate;
         matrix_affine          = sk_neon::matrix_affine;
 
-        RGBA_to_BGRA = sk_neon::RGBA_to_BGRA;
-        RGBA_to_rgbA = sk_neon::RGBA_to_rgbA;
-        RGBA_to_bgrA = sk_neon::RGBA_to_bgrA;
-        RGB_to_RGB1  = sk_neon::RGB_to_RGB1;
-        RGB_to_BGR1  = sk_neon::RGB_to_BGR1;
-        gray_to_RGB1 = sk_neon::gray_to_RGB1;
+        RGBA_to_BGRA  = sk_neon::RGBA_to_BGRA;
+        RGBA_to_rgbA  = sk_neon::RGBA_to_rgbA;
+        RGBA_to_bgrA  = sk_neon::RGBA_to_bgrA;
+        RGB_to_RGB1   = sk_neon::RGB_to_RGB1;
+        RGB_to_BGR1   = sk_neon::RGB_to_BGR1;
+        gray_to_RGB1  = sk_neon::gray_to_RGB1;
+        grayA_to_RGBA = sk_neon::grayA_to_RGBA;
+        grayA_to_rgbA = sk_neon::grayA_to_rgbA;
     }
 }

diff --git a/src/opts/SkOpts_ssse3.cpp b/src/opts/SkOpts_ssse3.cpp
index 22eda58..13c8fbf 100644 (file)
--- a/src/opts/SkOpts_ssse3.cpp
+++ b/src/opts/SkOpts_ssse3.cpp
@@ -18,11 +18,13 @@ namespace SkOpts {
         blit_mask_d32_a8 = sk_ssse3::blit_mask_d32_a8;
         color_cube_filter_span = sk_ssse3::color_cube_filter_span;
 
-        RGBA_to_BGRA = sk_ssse3::RGBA_to_BGRA;
-        RGBA_to_rgbA = sk_ssse3::RGBA_to_rgbA;
-        RGBA_to_bgrA = sk_ssse3::RGBA_to_bgrA;
-        RGB_to_RGB1  = sk_ssse3::RGB_to_RGB1;
-        RGB_to_BGR1  = sk_ssse3::RGB_to_BGR1;
-        gray_to_RGB1 = sk_ssse3::gray_to_RGB1;
+        RGBA_to_BGRA  = sk_ssse3::RGBA_to_BGRA;
+        RGBA_to_rgbA  = sk_ssse3::RGBA_to_rgbA;
+        RGBA_to_bgrA  = sk_ssse3::RGBA_to_bgrA;
+        RGB_to_RGB1   = sk_ssse3::RGB_to_RGB1;
+        RGB_to_BGR1   = sk_ssse3::RGB_to_BGR1;
+        gray_to_RGB1  = sk_ssse3::gray_to_RGB1;
+        grayA_to_RGBA = sk_ssse3::grayA_to_RGBA;
+        grayA_to_rgbA = sk_ssse3::grayA_to_rgbA;
     }
 }

diff --git a/src/opts/SkSwizzler_opts.h b/src/opts/SkSwizzler_opts.h
index 27416ea..f0492e7 100644 (file)
--- a/src/opts/SkSwizzler_opts.h
+++ b/src/opts/SkSwizzler_opts.h
@@ -98,6 +98,33 @@ static void gray_to_RGB1_portable(uint32_t dst[], const void* vsrc, int count) {
     }
 }
 
+static void grayA_to_RGBA_portable(uint32_t dst[], const void* vsrc, int count) {
+    const uint8_t* src = (const uint8_t*)vsrc;
+    for (int i = 0; i < count; i++) {
+        uint8_t g = src[0],
+                a = src[1];
+        src += 2;
+        dst[i] = (uint32_t)a << 24
+               | (uint32_t)g << 16
+               | (uint32_t)g <<  8
+               | (uint32_t)g <<  0;
+    }
+}
+
+static void grayA_to_rgbA_portable(uint32_t dst[], const void* vsrc, int count) {
+    const uint8_t* src = (const uint8_t*)vsrc;
+    for (int i = 0; i < count; i++) {
+        uint8_t g = src[0],
+                a = src[1];
+        src += 2;
+        g = (g*a+127)/255;
+        dst[i] = (uint32_t)a << 24
+               | (uint32_t)g << 16
+               | (uint32_t)g <<  8
+               | (uint32_t)g <<  0;
+    }
+}
+
 #if defined(SK_ARM_HAS_NEON)
 
 // Rounded divide by 255, (x + 127) / 255
@@ -311,6 +338,69 @@ static void gray_to_RGB1(uint32_t dst[], const void* vsrc, int count) {
     gray_to_RGB1_portable(dst, src, count);
 }
 
+template <bool kPremul>
+static void expand_grayA(uint32_t dst[], const void* vsrc, int count) {
+    const uint8_t* src = (const uint8_t*) vsrc;
+    while (count >= 16) {
+        // Load 16 pixels.
+        uint8x16x2_t ga = vld2q_u8(src);
+
+        // Premultiply if requested.
+        if (kPremul) {
+            ga.val[0] = vcombine_u8(
+                    scale(vget_low_u8(ga.val[0]),  vget_low_u8(ga.val[1])),
+                    scale(vget_high_u8(ga.val[0]), vget_high_u8(ga.val[1])));
+        }
+
+        // Set each of the color channels.
+        uint8x16x4_t rgba;
+        rgba.val[0] = ga.val[0];
+        rgba.val[1] = ga.val[0];
+        rgba.val[2] = ga.val[0];
+        rgba.val[3] = ga.val[1];
+
+        // Store 16 pixels.
+        vst4q_u8((uint8_t*) dst, rgba);
+        src += 16*2;
+        dst += 16;
+        count -= 16;
+    }
+
+    if (count >= 8) {
+        // Load 8 pixels.
+        uint8x8x2_t ga = vld2_u8(src);
+
+        // Premultiply if requested.
+        if (kPremul) {
+            ga.val[0] = scale(ga.val[0], ga.val[1]);
+        }
+
+        // Set each of the color channels.
+        uint8x8x4_t rgba;
+        rgba.val[0] = ga.val[0];
+        rgba.val[1] = ga.val[0];
+        rgba.val[2] = ga.val[0];
+        rgba.val[3] = ga.val[1];
+
+        // Store 8 pixels.
+        vst4_u8((uint8_t*) dst, rgba);
+        src += 8*2;
+        dst += 8;
+        count -= 8;
+    }
+
+    auto proc = kPremul ? grayA_to_rgbA_portable : grayA_to_RGBA_portable;
+    proc(dst, src, count);
+}
+
+static void grayA_to_RGBA(uint32_t dst[], const void* src, int count) {
+    expand_grayA<false>(dst, src, count);
+}
+
+static void grayA_to_rgbA(uint32_t dst[], const void* src, int count) {
+    expand_grayA<true>(dst, src, count);
+}
+
 #elif SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
 
 template <bool kSwapRB>
@@ -482,6 +572,14 @@ static void gray_to_RGB1(uint32_t dst[], const void* vsrc, int count) {
     gray_to_RGB1_portable(dst, src, count);
 }
 
+static void grayA_to_RGBA(uint32_t dst[], const void* src, int count) {
+    grayA_to_RGBA_portable(dst, src, count);
+}
+
+static void grayA_to_rgbA(uint32_t dst[], const void* src, int count) {
+    grayA_to_rgbA_portable(dst, src, count);
+}
+
 #else
 
 static void RGBA_to_rgbA(uint32_t* dst, const void* src, int count) {
@@ -508,6 +606,14 @@ static void gray_to_RGB1(uint32_t dst[], const void* src, int count) {
     gray_to_RGB1_portable(dst, src, count);
 }
 
+static void grayA_to_RGBA(uint32_t dst[], const void* src, int count) {
+    grayA_to_RGBA_portable(dst, src, count);
+}
+
+static void grayA_to_rgbA(uint32_t dst[], const void* src, int count) {
+    grayA_to_rgbA_portable(dst, src, count);
+}
+
 #endif
 
 }