add bgra as 1st class format

This is a start to eliminating swap_rb as a stage.

I've just hit the main hot spots here.  Going to look into
the ~dozen other spots to see how they should work next.

Change-Id: I26fb46a042facf7bd6fff3b47c9fcee86d7142fd
Reviewed-on: https://skia-review.googlesource.com/20982
Commit-Queue: Mike Klein <mtklein@chromium.org>
Reviewed-by: Mike Reed <reed@google.com>

src/core/SkRasterPipeline.h

@@ -64,16 +64,18 @@ struct SkJumper_constants;
     M(move_src_dst) M(move_dst_src)                              \
     M(clamp_0) M(clamp_1) M(clamp_a) M(clamp_a_dst)              \
     M(unpremul) M(premul)                                        \
-    M(set_rgb) M(swap_rb) M(swap_rb_dst)                         \
+    M(set_rgb) M(swap_rb)                                        \
     M(from_srgb) M(from_srgb_dst) M(to_srgb)                     \
     M(constant_color) M(seed_shader) M(dither)                   \
-    M(load_a8)   M(load_a8_dst)   M(store_a8)                    \
+                                                M(gather_i8)     \
-    M(load_g8)   M(load_g8_dst)                                  \
+    M(load_a8)   M(load_a8_dst)   M(store_a8)   M(gather_a8)     \
-    M(load_565)  M(load_565_dst)  M(store_565)                   \
+    M(load_g8)   M(load_g8_dst)                 M(gather_g8)     \
-    M(load_4444) M(load_4444_dst) M(store_4444)                  \
+    M(load_565)  M(load_565_dst)  M(store_565)  M(gather_565)    \
-    M(load_f16)  M(load_f16_dst)  M(store_f16)                   \
+    M(load_4444) M(load_4444_dst) M(store_4444) M(gather_4444)   \
+    M(load_f16)  M(load_f16_dst)  M(store_f16)  M(gather_f16)    \
     M(load_f32)  M(load_f32_dst)  M(store_f32)                   \
-    M(load_8888) M(load_8888_dst) M(store_8888)                  \
+    M(load_8888) M(load_8888_dst) M(store_8888) M(gather_8888)   \
+    M(load_bgra) M(load_bgra_dst) M(store_bgra) M(gather_bgra)   \
     M(load_u16_be) M(load_rgb_u16_be) M(store_u16_be)            \
     M(load_tables_u16_be) M(load_tables_rgb_u16_be)              \
     M(load_tables) M(load_rgba) M(store_rgba)                    \
@@ -96,8 +98,6 @@ struct SkJumper_constants;
     M(clamp_x)   M(mirror_x)   M(repeat_x)                       \
     M(clamp_y)   M(mirror_y)   M(repeat_y)                       \
     M(clamp_x_1) M(mirror_x_1) M(repeat_x_1)                     \
-    M(gather_a8) M(gather_g8) M(gather_i8)                       \
-    M(gather_565) M(gather_4444) M(gather_8888) M(gather_f16)    \
     M(bilinear_nx) M(bilinear_px) M(bilinear_ny) M(bilinear_py)  \
     M(bicubic_n3x) M(bicubic_n1x) M(bicubic_p1x) M(bicubic_p3x)  \
     M(bicubic_n3y) M(bicubic_n1y) M(bicubic_p1y) M(bicubic_p3y)  \

src/core/SkRasterPipelineBlitter.cpp

@@ -227,13 +227,10 @@ void SkRasterPipelineBlitter::append_load_d(SkRasterPipeline* p) const {
         case kAlpha_8_SkColorType:   p->append(SkRasterPipeline::load_a8_dst,   &fDstPtr); break;
         case kRGB_565_SkColorType:   p->append(SkRasterPipeline::load_565_dst,  &fDstPtr); break;
         case kARGB_4444_SkColorType: p->append(SkRasterPipeline::load_4444_dst, &fDstPtr); break;
-        case kBGRA_8888_SkColorType:
+        case kBGRA_8888_SkColorType: p->append(SkRasterPipeline::load_bgra_dst, &fDstPtr); break;
         case kRGBA_8888_SkColorType: p->append(SkRasterPipeline::load_8888_dst, &fDstPtr); break;
         case kRGBA_F16_SkColorType:  p->append(SkRasterPipeline::load_f16_dst,  &fDstPtr); break;
-        default: break;
+        default:                                                                           break;
-    }
-    if (fDst.info().colorType() == kBGRA_8888_SkColorType) {
-        p->append(SkRasterPipeline::swap_rb_dst);
     }
     if (fDst.info().gammaCloseToSRGB()) {
         p->append_from_srgb_dst(fDst.info().alphaType());
@@ -250,15 +247,12 @@ void SkRasterPipelineBlitter::append_store(SkRasterPipeline* p) const {
         p->append(SkRasterPipeline::dither, &fDitherRate);
     }
 
-    if (fDst.info().colorType() == kBGRA_8888_SkColorType) {
-        p->append(SkRasterPipeline::swap_rb);
-    }
     switch (fDst.info().colorType()) {
         case kGray_8_SkColorType:    p->append(SkRasterPipeline::luminance_to_alpha); // fallthru
         case kAlpha_8_SkColorType:   p->append(SkRasterPipeline::store_a8,   &fDstPtr); break;
         case kRGB_565_SkColorType:   p->append(SkRasterPipeline::store_565,  &fDstPtr); break;
         case kARGB_4444_SkColorType: p->append(SkRasterPipeline::store_4444, &fDstPtr); break;
-        case kBGRA_8888_SkColorType:
+        case kBGRA_8888_SkColorType: p->append(SkRasterPipeline::store_bgra, &fDstPtr); break;
         case kRGBA_8888_SkColorType: p->append(SkRasterPipeline::store_8888, &fDstPtr); break;
         case kRGBA_F16_SkColorType:  p->append(SkRasterPipeline::store_f16,  &fDstPtr); break;
         default: break;

src/jumper/SkJumper.cpp

@@ -75,9 +75,10 @@ using StartPipelineFn = void(size_t,size_t,size_t,void**,K*);
     M(set_rgb)           \
     M(premul)            \
     M(load_8888) M(load_8888_dst) M(store_8888) \
+    M(load_bgra) M(load_bgra_dst) M(store_bgra) \
     M(load_a8)   M(load_a8_dst)   M(store_a8)   \
     M(load_g8)   M(load_g8_dst)                 \
-    M(swap_rb)   M(swap_rb_dst)                 \
+    M(swap_rb)           \
     M(srcover_rgba_8888) \
     M(lerp_1_float)      \
     M(lerp_u8)           \

src/jumper/SkJumper_stages.cpp

@@ -520,12 +520,6 @@ STAGE(swap_rb) {
     b = tmp;
 }
 
-STAGE(swap_rb_dst) {
-    auto tmp = dr;
-    dr = db;
-    db = tmp;
-}
-
 STAGE(move_src_dst) {
     dr = r;
     dg = g;
@@ -906,6 +900,29 @@ STAGE(store_8888) {
     store(ptr, px, tail);
 }
 
+STAGE(load_bgra) {
+    auto ptr = *(const uint32_t**)ctx + x;
+    from_8888(load<U32>(ptr, tail), &b,&g,&r,&a);
+}
+STAGE(load_bgra_dst) {
+    auto ptr = *(const uint32_t**)ctx + x;
+    from_8888(load<U32>(ptr, tail), &db,&dg,&dr,&da);
+}
+STAGE(gather_bgra) {
+    const uint32_t* ptr;
+    U32 ix = ix_and_ptr(&ptr, ctx, r,g);
+    from_8888(gather(ptr, ix), &b,&g,&r,&a);
+}
+STAGE(store_bgra) {
+    auto ptr = *(uint32_t**)ctx + x;
+
+    U32 px = round(b, 255.0f)
+           | round(g, 255.0f) <<  8
+           | round(r, 255.0f) << 16
+           | round(a, 255.0f) << 24;
+    store(ptr, px, tail);
+}
+
 STAGE(load_f16) {
     auto ptr = *(const uint64_t**)ctx + x;
 

src/jumper/SkJumper_stages_lowp.cpp

@@ -249,6 +249,19 @@ STAGE(store_8888) {
     store(ptr, to_8888(r,g,b,a), tail);
 }
 
+STAGE(load_bgra) {
+    auto ptr = *(const uint32_t**)ctx + x;
+    from_8888(load<U32>(ptr, tail), &b,&g,&r,&a);
+}
+STAGE(load_bgra_dst) {
+    auto ptr = *(const uint32_t**)ctx + x;
+    from_8888(load<U32>(ptr, tail), &db,&dg,&dr,&da);
+}
+STAGE(store_bgra) {
+    auto ptr = *(uint32_t**)ctx + x;
+    store(ptr, to_8888(b,g,r,a), tail);
+}
+
 STAGE(load_a8) {
     auto ptr = *(const uint8_t**)ctx + x;
     r = g = b = 0.0f;
@@ -334,12 +347,6 @@ STAGE(swap_rb) {
     r = b;
     b = tmp;
 }
-STAGE(swap_rb_dst) {
-    auto tmp = dr;
-    dr = db;
-    db = tmp;
-}
-
 STAGE(move_src_dst) {
     dr = r;
     dg = g;

src/shaders/SkImageShader.cpp

@@ -336,8 +336,8 @@ bool SkImageShader::onAppendStages(SkRasterPipeline* p, SkColorSpace* dstCS, SkA
             case kGray_8_SkColorType:    p->append(SkRasterPipeline::gather_g8,   gather); break;
             case kRGB_565_SkColorType:   p->append(SkRasterPipeline::gather_565,  gather); break;
             case kARGB_4444_SkColorType: p->append(SkRasterPipeline::gather_4444, gather); break;
-            case kRGBA_8888_SkColorType:
+            case kBGRA_8888_SkColorType: p->append(SkRasterPipeline::gather_bgra, gather); break;
-            case kBGRA_8888_SkColorType: p->append(SkRasterPipeline::gather_8888, gather); break;
+            case kRGBA_8888_SkColorType: p->append(SkRasterPipeline::gather_8888, gather); break;
             case kRGBA_F16_SkColorType:  p->append(SkRasterPipeline::gather_f16,  gather); break;
             default: SkASSERT(false);
         }
@@ -396,11 +396,7 @@ bool SkImageShader::onAppendStages(SkRasterPipeline* p, SkColorSpace* dstCS, SkA
         p->append(SkRasterPipeline::move_dst_src);
     }
 
-    auto effective_color_type = [](SkColorType ct) {
+    if (info.colorType() == kIndex_8_SkColorType && kN32_SkColorType == kBGRA_8888_SkColorType) {
-        return ct == kIndex_8_SkColorType ? kN32_SkColorType : ct;
-    };
-
-    if (effective_color_type(info.colorType()) == kBGRA_8888_SkColorType) {
         p->append(SkRasterPipeline::swap_rb);
     }
     if (info.colorType() == kAlpha_8_SkColorType) {