sketch hooking into PNG_FILTER_OPTIMIZATIONS

Local timing says this 4-byte Paeth function takes about 0.3x the time the serial libpng code does, dropping from ~10 cycles per byte to ~2.9.

bpp=4 is mainly an easy demo.  This approach can work for any bpp up to 16, 1 pixel at a time, at roughly the same cost per pixel.  Doing more than 1 pixel at a time is a tricky math problem I have yet to attempt to solve.

Everything here can be trivially downgraded to MMX, supporting bpp up to 8.  It seems to be a little slower (~3.5 cycles per byte), but it would make the code compatible with every x86 that can still power on.

I've tried four approaches:
  - this way;
  - doing things naively in 16-bit;
  - a 16-bit version that requires division by 3 (i.e. mulhi_epu16(..., 0x5580) );
  - a mostly 8-bit version of the same.

They're all fine, but this one is consistently the fastest I've measured.
I'd be happy to settle on the naive 16-bit version too, which would have a very clear implementation that's only minorly slower than this version.  The other two are way more complicated, and would require us to draw some serious ASCII diagrams to explain.

I have learned that the .skp serialization tests (serialize-8888) have a nice side effect of testing the correctness of these filters!

(Since writing the description above, I've bumped things up to {Paeth,Sub,Avg} x { 3 bpp, 4 bpp }.)

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1573943002

Review URL: https://codereview.chromium.org/1573943002

3 files changed, 220 insertions, 8 deletions

diff --git a/src/codec/SkPngCodec.cpp b/src/codec/SkPngCodec.cpp
index 66215a6c73..9691f12d4d 100644
--- a/src/codec/SkPngCodec.cpp
+++ b/src/codec/SkPngCodec.cpp
@@ -11,11 +11,29 @@
 #include "SkBitmap.h"
 #include "SkMath.h"
 #include "SkPngCodec.h"
+#include "SkPngFilters.h"
 #include "SkSize.h"
 #include "SkStream.h"
 #include "SkSwizzler.h"
 #include "SkTemplates.h"
 
+#if defined(__SSE2__)
+    #include "pngstruct.h"
+
+    extern "C" void sk_png_init_filter_functions_sse2(png_structp png, unsigned int bpp) {
+        if (bpp == 3) {
+            png->read_filter[PNG_FILTER_VALUE_SUB  -1] =   sk_sub3_sse2;
+            png->read_filter[PNG_FILTER_VALUE_AVG  -1] =   sk_avg3_sse2;
+            png->read_filter[PNG_FILTER_VALUE_PAETH-1] = sk_paeth3_sse2;
+        }
+        if (bpp == 4) {
+            png->read_filter[PNG_FILTER_VALUE_SUB  -1] =   sk_sub4_sse2;
+            png->read_filter[PNG_FILTER_VALUE_AVG  -1] =   sk_avg4_sse2;
+            png->read_filter[PNG_FILTER_VALUE_PAETH-1] = sk_paeth4_sse2;
+        }
+    }
+#endif
+
 ///////////////////////////////////////////////////////////////////////////////
 // Helper macros
 ///////////////////////////////////////////////////////////////////////////////
@@ -331,8 +349,8 @@ static bool read_header(SkStream* stream, SkPngChunkReader* chunkReader,
             }
             break;
         case PNG_COLOR_TYPE_GRAY_ALPHA:
-            //FIXME: support gray with alpha as a color type 
-            //convert to RGBA 
+            //FIXME: support gray with alpha as a color type
+            //convert to RGBA
             png_set_gray_to_rgb(png_ptr);
             skColorType = kN32_SkColorType;
             skAlphaType = kUnpremul_SkAlphaType;
@@ -406,7 +424,7 @@ SkCodec::Result SkPngCodec::initializeSwizzler(const SkImageInfo& requestedInfo,
         SkCodecPrintf("setjmp long jump!\n");
         return kInvalidInput;
     }
-    png_read_update_info(fPng_ptr, fInfo_ptr);  
+    png_read_update_info(fPng_ptr, fInfo_ptr);
 
     //srcColorType was determined in read_header() which determined png color type
     const SkColorType srcColorType = this->getInfo().colorType();
@@ -422,7 +440,7 @@ SkCodec::Result SkPngCodec::initializeSwizzler(const SkImageInfo& requestedInfo,
             break;
         case kGray_8_SkColorType:
             fSrcConfig = SkSwizzler::kGray;
-            break; 
+            break;
         case kN32_SkColorType:
             if (this->getInfo().alphaType() == kOpaque_SkAlphaType) {
                     fSrcConfig = SkSwizzler::kRGB;
@@ -433,7 +451,7 @@ SkCodec::Result SkPngCodec::initializeSwizzler(const SkImageInfo& requestedInfo,
         default:
             //would have exited before now if the colorType was supported by png
             SkASSERT(false);
-    }  
+    }
 
     // Copy the color table to the client if they request kIndex8 mode
     copy_color_table(requestedInfo, fColorTable, ctable, ctableCount);
@@ -624,8 +642,8 @@ public:
             SkCodecPrintf("setjmp long jump!\n");
             return false;
         }
-        //there is a potential tradeoff of memory vs speed created by putting this in a loop. 
-        //calling png_read_rows in a loop is insignificantly slower than calling it once with count 
+        //there is a potential tradeoff of memory vs speed created by putting this in a loop.
+        //calling png_read_rows in a loop is insignificantly slower than calling it once with count
         //as png_read_rows has it's own loop which calls png_read_row count times.
         for (int row = 0; row < count; row++) {
             png_read_rows(this->png_ptr(), &fSrcRow, png_bytepp_NULL, 1);
@@ -656,7 +674,7 @@ public:
     Result onStartScanlineDecode(const SkImageInfo& dstInfo, const Options& options,
             SkPMColor ctable[], int* ctableCount) override {
         if (!conversion_possible(dstInfo, this->getInfo())) {
-            return kInvalidConversion;    
+            return kInvalidConversion;
         }
 
         const Result result = this->initializeSwizzler(dstInfo, options, ctable,
diff --git a/src/codec/SkPngFilters.cpp b/src/codec/SkPngFilters.cpp
new file mode 100644
index 0000000000..115810ea7c
--- /dev/null
+++ b/src/codec/SkPngFilters.cpp
@@ -0,0 +1,168 @@
+/*
+ * Copyright 2016 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#include "SkPngFilters.h"
+#include "SkTypes.h"
+
+// Functions in this file look at most 3 pixels (a,b,c) to predict the fourth (d).
+// They're positioned like this:
+//     prev:  c b
+//     row:   a d
+// The Sub filter predicts d=a, Avg d=(a+b)/2,  and Paeth predicts d to be whichever
+// of a, b, or c is closest to p=a+b-c.  (Up also exists, predicting d=b.)
+
+#if defined(__SSE2__)
+
+    template <int bpp>
+    static __m128i load(const void* p) {
+        static_assert(bpp <= 4, "");
+
+        uint32_t packed;
+        memcpy(&packed, p, bpp);
+        return _mm_cvtsi32_si128(packed);
+    }
+
+    template <int bpp>
+    static void store(void* p, __m128i v) {
+        static_assert(bpp <= 4, "");
+
+        uint32_t packed = _mm_cvtsi128_si32(v);
+        memcpy(p, &packed, bpp);
+    }
+
+    template <int bpp>
+    static void sk_sub_sse2(png_row_infop row_info, png_bytep row, png_const_bytep) {
+        // The Sub filter predicts each pixel as the previous pixel, a.
+        // There is no pixel to the left of the first pixel.  It's encoded directly.
+        // That works with our main loop if we just say that left pixel was zero.
+        __m128i a, d = _mm_setzero_si128();
+
+        int rb = row_info->rowbytes;
+        while (rb > 0) {
+            a = d; d = load<bpp>(row);
+            d = _mm_add_epi8(d, a);
+            store<bpp>(row, d);
+
+            row += bpp;
+            rb  -= bpp;
+        }
+    }
+
+    template <int bpp>
+    void sk_avg_sse2(png_row_infop row_info, png_bytep row, png_const_bytep prev) {
+        // The Avg filter predicts each pixel as the (truncated) average of a and b.
+        // There's no pixel to the left of the first pixel.  Luckily, it's
+        // predicted to be half of the pixel above it.  So again, this works
+        // perfectly with our loop if we make sure a starts at zero.
+        const __m128i zero = _mm_setzero_si128();
+        __m128i    b;
+        __m128i a, d = zero;
+
+        int rb = row_info->rowbytes;
+        while (rb > 0) {
+                   b = load<bpp>(prev);
+            a = d; d = load<bpp>(row );
+
+            // PNG requires a truncating average here, so sadly we can't just use _mm_avg_epu8...
+            __m128i avg = _mm_avg_epu8(a,b);
+            // ...but we can fix it up by subtracting off 1 if it rounded up.
+            avg = _mm_sub_epi8(avg, _mm_and_si128(_mm_xor_si128(a,b), _mm_set1_epi8(1)));
+
+            d = _mm_add_epi8(d, avg);
+            store<bpp>(row, d);
+
+            prev += bpp;
+            row  += bpp;
+            rb   -= bpp;
+        }
+    }
+
+    // Returns bytewise |x-y|.
+    static __m128i absdiff_u8(__m128i x, __m128i y) {
+        // One of these two saturated subtractions will be the answer, the other zero.
+        return _mm_or_si128(_mm_subs_epu8(x,y), _mm_subs_epu8(y,x));
+    }
+
+    // Bytewise c ? t : e.
+    static __m128i if_then_else(__m128i c, __m128i t, __m128i e) {
+        // SSE 4.1+ would be: return _mm_blendv_epi8(e,t,c);
+        return _mm_or_si128(_mm_and_si128(c, t), _mm_andnot_si128(c, e));
+    }
+
+    template <int bpp>
+    void sk_paeth_sse2(png_row_infop row_info, png_bytep row, png_const_bytep prev) {
+        // Paeth tries to predict pixel d using the pixel to the left of it, a,
+        // and two pixels from the previous row, b and c:
+        //   prev: c b
+        //   row:  a d
+        // The Paeth function predicts d to be whichever of a, b, or c is nearest to p=a+b-c.
+
+        // The first pixel has no left context, and so uses an Up filter, p = b.
+        // This works naturally with our main loop's p = a+b-c if we force a and c to zero.
+        // Here we zero b and d, which become c and a respectively at the start of the loop.
+        __m128i c, b = _mm_setzero_si128(),
+                a, d = _mm_setzero_si128();
+
+        int rb = row_info->rowbytes;
+        while (rb > 0) {
+            c = b; b = load<bpp>(prev);
+            a = d; d = load<bpp>(row );
+
+            // We can't express p in 8 bits, but luckily we can use this faux p instead.
+            // (I have no deep insight here... I just proved this with brute force.)
+            __m128i min = _mm_min_epu8(a,b),
+                    max = _mm_max_epu8(a,b),
+                    faux_p = _mm_adds_epu8(min, _mm_subs_epu8(max, c));
+
+            // We could use faux_p for calculating all three of pa, pb, and pc,
+            // but it's a little quicker to calculate the correct pa and pb directly,
+            // and the predictor remains the same.  (Again, brute force.)
+            __m128i pa = absdiff_u8(b,c),             // |a+b-c - a| == |b-c|
+                    pb = absdiff_u8(a,c),             // |a+b-c - b| == |a-c|
+               faux_pc = absdiff_u8(faux_p, c);
+
+            // From here, things are straightforward.  Find the smallest distance to p...
+            __m128i smallest = _mm_min_epu8(_mm_min_epu8(pa, pb), faux_pc);
+
+            // ... then the predictor is the input corresponding to that smallest distance,
+            // breaking ties in favor of a over b over c.
+            __m128i nearest = if_then_else(_mm_cmpeq_epi8(smallest, pa), a,
+                              if_then_else(_mm_cmpeq_epi8(smallest, pb), b,
+                                                                         c));
+
+            // We've reconstructed d!  Leave it for next round to become a, and write it out.
+            d = _mm_add_epi8(d, nearest);
+            store<bpp>(row, d);
+
+            prev += bpp;
+            row  += bpp;
+            rb   -= bpp;
+        }
+    }
+
+    void sk_sub3_sse2(png_row_infop row_info, png_bytep row, png_const_bytep prev) {
+        sk_sub_sse2<3>(row_info, row, prev);
+    }
+    void sk_sub4_sse2(png_row_infop row_info, png_bytep row, png_const_bytep prev) {
+        sk_sub_sse2<4>(row_info, row, prev);
+    }
+
+    void sk_avg3_sse2(png_row_infop row_info, png_bytep row, png_const_bytep prev) {
+        sk_avg_sse2<3>(row_info, row, prev);
+    }
+    void sk_avg4_sse2(png_row_infop row_info, png_bytep row, png_const_bytep prev) {
+        sk_avg_sse2<4>(row_info, row, prev);
+    }
+
+    void sk_paeth3_sse2(png_row_infop row_info, png_bytep row, png_const_bytep prev) {
+        sk_paeth_sse2<3>(row_info, row, prev);
+    }
+    void sk_paeth4_sse2(png_row_infop row_info, png_bytep row, png_const_bytep prev) {
+        sk_paeth_sse2<4>(row_info, row, prev);
+    }
+
+#endif
diff --git a/src/codec/SkPngFilters.h b/src/codec/SkPngFilters.h
new file mode 100644
index 0000000000..5fc446881c
--- /dev/null
+++ b/src/codec/SkPngFilters.h
@@ -0,0 +1,26 @@
+/*
+ * Copyright 2016 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#ifndef SkPngFilters_DEFINED
+#define SkPngFilters_DEFINED
+
+#include "png.h"
+
+// We don't bother specializing Up...
+// it's so simple it's usually already perfectly autovectorized.
+
+// These all require bpp=3 (i.e. RGB).
+void   sk_sub3_sse2(png_row_infop, png_bytep, png_const_bytep);
+void   sk_avg3_sse2(png_row_infop, png_bytep, png_const_bytep);
+void sk_paeth3_sse2(png_row_infop, png_bytep, png_const_bytep);
+
+// These all require bpp=4 (i.e. RGBA).
+void   sk_sub4_sse2(png_row_infop, png_bytep, png_const_bytep);
+void   sk_avg4_sse2(png_row_infop, png_bytep, png_const_bytep);
+void sk_paeth4_sse2(png_row_infop, png_bytep, png_const_bytep);
+
+#endif//SkPngFilterOpts_DEFINED