restore SkOpts::blit_row_color32

We think there are some seemingly minor codegen improvements (pmovzxbw
instead of punpcklbw) when compiling for SSE 4.1 that might actually be
a bigger deal speed-wise than they'd seem.

Also rewrite using SkVx in a way that should scale well up to AVX2.

Change-Id: Ie7c0194dc4fe9fe81c1c932187c0bb00da69190b
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/207260
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Lee Salzman <lsalzman@mozilla.com>

src/core/SkBlitRow_D32.cpp

@@ -5,7 +5,6 @@
  * found in the LICENSE file.
  */
 
-#include "Sk4px.h"
 #include "SkBlitRow.h"
 #include "SkColorData.h"
 #include "SkOpts.h"
@@ -311,15 +310,5 @@ void SkBlitRow::Color32(SkPMColor dst[], const SkPMColor src[], int count, SkPMC
         case   0: memmove(dst, src, count * sizeof(SkPMColor)); return;
         case 255: sk_memset32(dst, color, count);               return;
     }
-
-    unsigned invA = 255 - SkGetPackedA32(color);
-    invA += invA >> 7;
-    SkASSERT(invA < 256);  // We've should have already handled alpha == 0 externally.
-
-    Sk16h colorHighAndRound = (Sk4px::DupPMColor(color).widen() << 8) + Sk16h(128);
-    Sk16b invA_16x(invA);
-
-    Sk4px::MapSrc(count, dst, src, [&](const Sk4px& src4) -> Sk4px {
-        return (src4 * invA_16x).addNarrowHi(colorHighAndRound);
-    });
+    return SkOpts::blit_row_color32(dst, src, count, color);
 }

src/core/SkOpts.cpp

@@ -55,6 +55,7 @@ namespace SkOpts {
 
     DEFINE_DEFAULT(blit_mask_d32_a8);
 
+    DEFINE_DEFAULT(blit_row_color32);
     DEFINE_DEFAULT(blit_row_s32a_opaque);
 
     DEFINE_DEFAULT(RGBA_to_BGRA);

src/core/SkOpts.h

@@ -26,6 +26,7 @@ namespace SkOpts {
     extern SkXfermode* (*create_xfermode)(SkBlendMode);
 
     extern void (*blit_mask_d32_a8)(SkPMColor*, size_t, const SkAlpha*, size_t, SkColor, int, int);
+    extern void (*blit_row_color32)(SkPMColor*, const SkPMColor*, int, SkPMColor);
     extern void (*blit_row_s32a_opaque)(SkPMColor*, const SkPMColor*, int, U8CPU);
 
     // Swizzle input into some sort of 8888 pixel, {premul,unpremul} x {rgba,bgra}.

src/opts/SkBlitRow_opts.h

@@ -8,6 +8,7 @@
 #ifndef SkBlitRow_opts_DEFINED
 #define SkBlitRow_opts_DEFINED
 
+#include "SkVx.h"
 #include "SkColorData.h"
 #include "SkMSAN.h"
 
@@ -40,6 +41,38 @@
 
 namespace SK_OPTS_NS {
 
+// Blend constant color over count src pixels, writing into dst.
+inline void blit_row_color32(SkPMColor* dst, const SkPMColor* src, int count, SkPMColor color) {
+    constexpr int N = 8;  // 4, 16 also reasonable choices
+    using U32 = skvx::Vec<  N, uint32_t>;
+    using U16 = skvx::Vec<4*N, uint16_t>;
+    using U8  = skvx::Vec<4*N, uint8_t>;
+
+    auto kernel = [color](U32 src) {
+        unsigned invA = 255 - SkGetPackedA32(color);
+        invA += invA >> 7;
+        SkASSERT(0 < invA && invA < 256);  // We handle alpha == 0 or alpha == 255 specially.
+
+        // (src * invA + (color << 8) + 128) >> 8
+        // Should all fit in 16 bits.
+        // TODO(mtklein): can we do src * invA with umull on ARM?
+        U16 s = skvx::cast<uint16_t>(skvx::bit_pun<U8>(src)),
+            c = skvx::cast<uint16_t>(skvx::bit_pun<U8>(U32(color))),
+            d = (s * invA + (c << 8) + 128)>>8;
+        return skvx::bit_pun<U32>(skvx::cast<uint8_t>(d));
+    };
+
+    while (count >= N) {
+        kernel(U32::Load(src)).store(dst);
+        src   += N;
+        dst   += N;
+        count -= N;
+    }
+    while (count --> 0) {
+        *dst++ = kernel(U32{*src++})[0];
+    }
+}
+
 #if defined(SK_ARM_HAS_NEON)
 
 // Return a uint8x8_t value, r, computed as r[i] = SkMulDiv255Round(x[i], y[i]), where r[i], x[i],

src/opts/SkOpts_hsw.cpp

@@ -8,11 +8,15 @@
 #include "SkOpts.h"
 
 #define SK_OPTS_NS hsw
+#include "SkBlitRow_opts.h"
 #include "SkRasterPipeline_opts.h"
 #include "SkUtils_opts.h"
 
 namespace SkOpts {
     void Init_hsw() {
+        blit_row_color32     = hsw::blit_row_color32;
+        blit_row_s32a_opaque = hsw::blit_row_s32a_opaque;
+
     #define M(st) stages_highp[SkRasterPipeline::st] = (StageFn)SK_OPTS_NS::st;
         SK_RASTER_PIPELINE_STAGES(M)
         just_return_highp = (StageFn)SK_OPTS_NS::just_return;

src/opts/SkOpts_sse41.cpp

@@ -8,11 +8,12 @@
 #include "SkOpts.h"
 
 #define SK_OPTS_NS sse41
-#include "SkRasterPipeline_opts.h"
 #include "SkBlitRow_opts.h"
+#include "SkRasterPipeline_opts.h"
 
 namespace SkOpts {
     void Init_sse41() {
+        blit_row_color32     = sse41::blit_row_color32;
         blit_row_s32a_opaque = sse41::blit_row_s32a_opaque;
 
     #define M(st) stages_highp[SkRasterPipeline::st] = (StageFn)SK_OPTS_NS::st;