Revert "explore always using SkRasterPipeline for color space transforms"

This reverts commit 58564425e5acb5911cc3719f9c1e39190cc829b8.

Reason for revert: iOS

Original change's description:
> explore always using SkRasterPipeline for color space transforms
> 
> On my trashcan, bench times change: mostly slowdowns, some speedups.
> On the other hand, this cuts about 270K of code out of Skia.
> 
> Change-Id: Ib1069792508ced361c11ea809b0b8187f5e28a5c
> Reviewed-on: https://skia-review.googlesource.com/19744
> Reviewed-by: Matt Sarett <msarett@google.com>
> Commit-Queue: Mike Klein <mtklein@chromium.org>

TBR=mtklein@chromium.org,msarett@google.com

Change-Id: I26469e174f0281597ac5720ec790686ef3fbb4e0
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/20036
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>

1 files changed, 698 insertions, 21 deletions

diff --git a/src/core/SkColorSpaceXform.cpp b/src/core/SkColorSpaceXform.cpp
index 78d8812869..8887bc279f 100644
--- a/src/core/SkColorSpaceXform.cpp
+++ b/src/core/SkColorSpaceXform.cpp
@@ -349,7 +349,570 @@ std::unique_ptr<SkColorSpaceXform> SkColorSpaceXform_Base::New(SkColorSpace* src
 
 ///////////////////////////////////////////////////////////////////////////////////////////////////
 
-static inline int num_tables(SkColorSpace_XYZ* space) {
+#define AI SK_ALWAYS_INLINE
+
+static AI void load_matrix(const float matrix[13],
+                           Sk4f& rXgXbX, Sk4f& rYgYbY, Sk4f& rZgZbZ, Sk4f& rTgTbT) {
+    rXgXbX = Sk4f::Load(matrix + 0);
+    rYgYbY = Sk4f::Load(matrix + 3);
+    rZgZbZ = Sk4f::Load(matrix + 6);
+    rTgTbT = Sk4f::Load(matrix + 9);
+}
+
+enum Order {
+    kRGBA_Order,
+    kBGRA_Order,
+};
+
+static AI void set_rb_shifts(Order kOrder, int* kRShift, int* kBShift) {
+    if (kRGBA_Order == kOrder) {
+        *kRShift = 0;
+        *kBShift = 16;
+    } else {
+        *kRShift = 16;
+        *kBShift = 0;
+    }
+}
+
+template <Order kOrder>
+static AI void load_rgb_from_tables(const uint32_t* src,
+                                    Sk4f& r, Sk4f& g, Sk4f& b, Sk4f& a,
+                                    const float* const srcTables[3]) {
+    int kRShift, kGShift = 8, kBShift;
+    set_rb_shifts(kOrder, &kRShift, &kBShift);
+    r = { srcTables[0][(src[0] >> kRShift) & 0xFF],
+          srcTables[0][(src[1] >> kRShift) & 0xFF],
+          srcTables[0][(src[2] >> kRShift) & 0xFF],
+          srcTables[0][(src[3] >> kRShift) & 0xFF], };
+    g = { srcTables[1][(src[0] >> kGShift) & 0xFF],
+          srcTables[1][(src[1] >> kGShift) & 0xFF],
+          srcTables[1][(src[2] >> kGShift) & 0xFF],
+          srcTables[1][(src[3] >> kGShift) & 0xFF], };
+    b = { srcTables[2][(src[0] >> kBShift) & 0xFF],
+          srcTables[2][(src[1] >> kBShift) & 0xFF],
+          srcTables[2][(src[2] >> kBShift) & 0xFF],
+          srcTables[2][(src[3] >> kBShift) & 0xFF], };
+    a = 0.0f; // Don't let the compiler complain that |a| is uninitialized.
+}
+
+template <Order kOrder>
+static AI void load_rgba_from_tables(const uint32_t* src,
+                                     Sk4f& r, Sk4f& g, Sk4f& b, Sk4f& a,
+                                     const float* const srcTables[3]) {
+    int kRShift, kGShift = 8, kBShift;
+    set_rb_shifts(kOrder, &kRShift, &kBShift);
+    r = { srcTables[0][(src[0] >> kRShift) & 0xFF],
+          srcTables[0][(src[1] >> kRShift) & 0xFF],
+          srcTables[0][(src[2] >> kRShift) & 0xFF],
+          srcTables[0][(src[3] >> kRShift) & 0xFF], };
+    g = { srcTables[1][(src[0] >> kGShift) & 0xFF],
+          srcTables[1][(src[1] >> kGShift) & 0xFF],
+          srcTables[1][(src[2] >> kGShift) & 0xFF],
+          srcTables[1][(src[3] >> kGShift) & 0xFF], };
+    b = { srcTables[2][(src[0] >> kBShift) & 0xFF],
+          srcTables[2][(src[1] >> kBShift) & 0xFF],
+          srcTables[2][(src[2] >> kBShift) & 0xFF],
+          srcTables[2][(src[3] >> kBShift) & 0xFF], };
+    a = (1.0f / 255.0f) * SkNx_cast<float>(Sk4u::Load(src) >> 24);
+}
+
+template <Order kOrder>
+static AI void load_rgb_linear(const uint32_t* src, Sk4f& r, Sk4f& g, Sk4f& b, Sk4f& a,
+                               const float* const[3]) {
+    int kRShift, kGShift = 8, kBShift;
+    set_rb_shifts(kOrder, &kRShift, &kBShift);
+    r = (1.0f / 255.0f) * SkNx_cast<float>((Sk4u::Load(src) >> kRShift) & 0xFF);
+    g = (1.0f / 255.0f) * SkNx_cast<float>((Sk4u::Load(src) >> kGShift) & 0xFF);
+    b = (1.0f / 255.0f) * SkNx_cast<float>((Sk4u::Load(src) >> kBShift) & 0xFF);
+    a = 0.0f; // Don't let the compiler complain that |a| is uninitialized.
+}
+
+template <Order kOrder>
+static AI void load_rgba_linear(const uint32_t* src, Sk4f& r, Sk4f& g, Sk4f& b, Sk4f& a,
+                                const float* const[3]) {
+    int kRShift, kGShift = 8, kBShift;
+    set_rb_shifts(kOrder, &kRShift, &kBShift);
+    r = (1.0f / 255.0f) * SkNx_cast<float>((Sk4u::Load(src) >> kRShift) & 0xFF);
+    g = (1.0f / 255.0f) * SkNx_cast<float>((Sk4u::Load(src) >> kGShift) & 0xFF);
+    b = (1.0f / 255.0f) * SkNx_cast<float>((Sk4u::Load(src) >> kBShift) & 0xFF);
+    a = (1.0f / 255.0f) * SkNx_cast<float>((Sk4u::Load(src) >> 24));
+}
+
+template <Order kOrder>
+static AI void load_rgb_from_tables_1(const uint32_t* src,
+                                      Sk4f& r, Sk4f& g, Sk4f& b, Sk4f& a,
+                                      const float* const srcTables[3]) {
+    int kRShift, kGShift = 8, kBShift;
+    set_rb_shifts(kOrder, &kRShift, &kBShift);
+    r = Sk4f(srcTables[0][(*src >> kRShift) & 0xFF]);
+    g = Sk4f(srcTables[1][(*src >> kGShift) & 0xFF]);
+    b = Sk4f(srcTables[2][(*src >> kBShift) & 0xFF]);
+    a = 0.0f; // Don't let MSAN complain that |a| is uninitialized.
+}
+
+template <Order kOrder>
+static AI void load_rgba_from_tables_1(const uint32_t* src,
+                                       Sk4f& r, Sk4f& g, Sk4f& b, Sk4f& a,
+                                       const float* const srcTables[3]) {
+    int kRShift, kGShift = 8, kBShift;
+    set_rb_shifts(kOrder, &kRShift, &kBShift);
+    r = Sk4f(srcTables[0][(*src >> kRShift) & 0xFF]);
+    g = Sk4f(srcTables[1][(*src >> kGShift) & 0xFF]);
+    b = Sk4f(srcTables[2][(*src >> kBShift) & 0xFF]);
+    a = (1.0f / 255.0f) * Sk4f(*src >> 24);
+}
+
+template <Order kOrder>
+static AI void load_rgb_linear_1(const uint32_t* src,
+                                 Sk4f& r, Sk4f& g, Sk4f& b, Sk4f& a,
+                                 const float* const srcTables[3]) {
+    int kRShift, kGShift = 8, kBShift;
+    set_rb_shifts(kOrder, &kRShift, &kBShift);
+    r = Sk4f((1.0f / 255.0f) * ((*src >> kRShift) & 0xFF));
+    g = Sk4f((1.0f / 255.0f) * ((*src >> kGShift) & 0xFF));
+    b = Sk4f((1.0f / 255.0f) * ((*src >> kBShift) & 0xFF));
+    a = 0.0f; // Don't let MSAN complain that |a| is uninitialized.
+}
+
+template <Order kOrder>
+static AI void load_rgba_linear_1(const uint32_t* src,
+                                  Sk4f& r, Sk4f& g, Sk4f& b, Sk4f& a,
+                                  const float* const srcTables[3]) {
+    int kRShift, kGShift = 8, kBShift;
+    set_rb_shifts(kOrder, &kRShift, &kBShift);
+    r = Sk4f((1.0f / 255.0f) * ((*src >> kRShift) & 0xFF));
+    g = Sk4f((1.0f / 255.0f) * ((*src >> kGShift) & 0xFF));
+    b = Sk4f((1.0f / 255.0f) * ((*src >> kBShift) & 0xFF));
+    a = Sk4f((1.0f / 255.0f) * ((*src >> 24)));
+}
+
+static AI void transform_gamut(const Sk4f& r, const Sk4f& g, const Sk4f& b, const Sk4f& a,
+                               const Sk4f& rXgXbX, const Sk4f& rYgYbY, const Sk4f& rZgZbZ,
+                               Sk4f& dr, Sk4f& dg, Sk4f& db, Sk4f& da) {
+    dr = rXgXbX[0]*r + rYgYbY[0]*g + rZgZbZ[0]*b;
+    dg = rXgXbX[1]*r + rYgYbY[1]*g + rZgZbZ[1]*b;
+    db = rXgXbX[2]*r + rYgYbY[2]*g + rZgZbZ[2]*b;
+    da = a;
+}
+
+static AI void transform_gamut_1(const Sk4f& r, const Sk4f& g, const Sk4f& b,
+                                 const Sk4f& rXgXbX, const Sk4f& rYgYbY, const Sk4f& rZgZbZ,
+                                 Sk4f& rgba) {
+    rgba = rXgXbX*r + rYgYbY*g + rZgZbZ*b;
+}
+
+static AI void translate_gamut(const Sk4f& rTgTbT, Sk4f& dr, Sk4f& dg, Sk4f& db) {
+    dr = dr + rTgTbT[0];
+    dg = dg + rTgTbT[1];
+    db = db + rTgTbT[2];
+}
+
+static AI void translate_gamut_1(const Sk4f& rTgTbT, Sk4f& rgba) {
+    rgba = rgba + rTgTbT;
+}
+
+template <Order kOrder>
+static AI void store_srgb(void* dst, const uint32_t* src, Sk4f& dr, Sk4f& dg, Sk4f& db, Sk4f&,
+                          const uint8_t* const[3]) {
+    int kRShift, kGShift = 8, kBShift;
+    set_rb_shifts(kOrder, &kRShift, &kBShift);
+    dr = sk_linear_to_srgb_needs_trunc(dr);
+    dg = sk_linear_to_srgb_needs_trunc(dg);
+    db = sk_linear_to_srgb_needs_trunc(db);
+
+    dr = sk_clamp_0_255(dr);
+    dg = sk_clamp_0_255(dg);
+    db = sk_clamp_0_255(db);
+
+    Sk4i da = Sk4i::Load(src) & 0xFF000000;
+
+    Sk4i rgba = (SkNx_cast<int>(dr) << kRShift)
+              | (SkNx_cast<int>(dg) << kGShift)
+              | (SkNx_cast<int>(db) << kBShift)
+              | (da                           );
+    rgba.store(dst);
+}
+
+template <Order kOrder>
+static AI void store_srgb_1(void* dst, const uint32_t* src,
+                            Sk4f& rgba, const Sk4f&,
+                            const uint8_t* const[3]) {
+    rgba = sk_clamp_0_255(sk_linear_to_srgb_needs_trunc(rgba));
+
+    uint32_t tmp;
+    SkNx_cast<uint8_t>(SkNx_cast<int32_t>(rgba)).store(&tmp);
+    tmp = (*src & 0xFF000000) | (tmp & 0x00FFFFFF);
+    if (kBGRA_Order == kOrder) {
+        tmp = SkSwizzle_RB(tmp);
+    }
+
+    *(uint32_t*)dst = tmp;
+}
+
+static AI Sk4f linear_to_2dot2(const Sk4f& x) {
+    // x^(29/64) is a very good approximation of the true value, x^(1/2.2).
+    auto x2  = x.rsqrt(),                            // x^(-1/2)
+         x32 = x2.rsqrt().rsqrt().rsqrt().rsqrt(),   // x^(-1/32)
+         x64 = x32.rsqrt();                          // x^(+1/64)
+
+    // 29 = 32 - 2 - 1
+    return 255.0f * x2.invert() * x32 * x64.invert();
+}
+
+template <Order kOrder>
+static AI void store_2dot2(void* dst, const uint32_t* src, Sk4f& dr, Sk4f& dg, Sk4f& db, Sk4f&,
+                           const uint8_t* const[3]) {
+    int kRShift, kGShift = 8, kBShift;
+    set_rb_shifts(kOrder, &kRShift, &kBShift);
+    dr = linear_to_2dot2(dr);
+    dg = linear_to_2dot2(dg);
+    db = linear_to_2dot2(db);
+
+    dr = sk_clamp_0_255(dr);
+    dg = sk_clamp_0_255(dg);
+    db = sk_clamp_0_255(db);
+
+    Sk4i da = Sk4i::Load(src) & 0xFF000000;
+
+    Sk4i rgba = (Sk4f_round(dr) << kRShift)
+              | (Sk4f_round(dg) << kGShift)
+              | (Sk4f_round(db) << kBShift)
+              | (da                       );
+    rgba.store(dst);
+}
+
+template <Order kOrder>
+static AI void store_2dot2_1(void* dst, const uint32_t* src,
+                             Sk4f& rgba, const Sk4f&,
+                             const uint8_t* const[3]) {
+    rgba = sk_clamp_0_255(linear_to_2dot2(rgba));
+
+    uint32_t tmp;
+    SkNx_cast<uint8_t>(Sk4f_round(rgba)).store(&tmp);
+    tmp = (*src & 0xFF000000) | (tmp & 0x00FFFFFF);
+    if (kBGRA_Order == kOrder) {
+        tmp = SkSwizzle_RB(tmp);
+    }
+
+    *(uint32_t*)dst = tmp;
+}
+
+template <Order kOrder>
+static AI void store_linear(void* dst, const uint32_t* src, Sk4f& dr, Sk4f& dg, Sk4f& db, Sk4f&,
+                            const uint8_t* const[3]) {
+    int kRShift, kGShift = 8, kBShift;
+    set_rb_shifts(kOrder, &kRShift, &kBShift);
+    dr = sk_clamp_0_255(255.0f * dr);
+    dg = sk_clamp_0_255(255.0f * dg);
+    db = sk_clamp_0_255(255.0f * db);
+
+    Sk4i da = Sk4i::Load(src) & 0xFF000000;
+
+    Sk4i rgba = (Sk4f_round(dr) << kRShift)
+              | (Sk4f_round(dg) << kGShift)
+              | (Sk4f_round(db) << kBShift)
+              | (da                       );
+    rgba.store(dst);
+}
+
+template <Order kOrder>
+static AI void store_linear_1(void* dst, const uint32_t* src,
+                              Sk4f& rgba, const Sk4f&,
+                              const uint8_t* const[3]) {
+    rgba = sk_clamp_0_255(255.0f * rgba);
+
+    uint32_t tmp;
+    SkNx_cast<uint8_t>(Sk4f_round(rgba)).store(&tmp);
+    tmp = (*src & 0xFF000000) | (tmp & 0x00FFFFFF);
+    if (kBGRA_Order == kOrder) {
+        tmp = SkSwizzle_RB(tmp);
+    }
+
+    *(uint32_t*)dst = tmp;
+}
+
+template <Order kOrder>
+static AI void store_f16(void* dst, const uint32_t* src, Sk4f& dr, Sk4f& dg, Sk4f& db, Sk4f& da,
+                         const uint8_t* const[3]) {
+    Sk4h::Store4(dst, SkFloatToHalf_finite_ftz(dr),
+                      SkFloatToHalf_finite_ftz(dg),
+                      SkFloatToHalf_finite_ftz(db),
+                      SkFloatToHalf_finite_ftz(da));
+}
+
+template <Order kOrder>
+static AI void store_f16_1(void* dst, const uint32_t* src,
+                           Sk4f& rgba, const Sk4f& a,
+                           const uint8_t* const[3]) {
+    rgba = Sk4f(rgba[0], rgba[1], rgba[2], a[3]);
+    SkFloatToHalf_finite_ftz(rgba).store((uint64_t*) dst);
+}
+
+template <Order kOrder>
+static AI void store_f16_opaque(void* dst, const uint32_t* src, Sk4f& dr, Sk4f& dg, Sk4f& db,
+                                Sk4f&, const uint8_t* const[3]) {
+    Sk4h::Store4(dst, SkFloatToHalf_finite_ftz(dr),
+                      SkFloatToHalf_finite_ftz(dg),
+                      SkFloatToHalf_finite_ftz(db),
+                      SK_Half1);
+}
+
+template <Order kOrder>
+static AI void store_f16_1_opaque(void* dst, const uint32_t* src,
+                                  Sk4f& rgba, const Sk4f&,
+                                  const uint8_t* const[3]) {
+    uint64_t tmp;
+    SkFloatToHalf_finite_ftz(rgba).store(&tmp);
+    tmp &= 0x0000FFFFFFFFFFFF;
+    tmp |= static_cast<uint64_t>(SK_Half1) << 48;
+    *((uint64_t*) dst) = tmp;
+}
+
+template <Order kOrder>
+static AI void store_generic(void* dst, const uint32_t* src, Sk4f& dr, Sk4f& dg, Sk4f& db, Sk4f&,
+                             const uint8_t* const dstTables[3]) {
+    int kRShift, kGShift = 8, kBShift;
+    set_rb_shifts(kOrder, &kRShift, &kBShift);
+    dr = Sk4f::Min(Sk4f::Max(1023.0f * dr, 0.0f), 1023.0f);
+    dg = Sk4f::Min(Sk4f::Max(1023.0f * dg, 0.0f), 1023.0f);
+    db = Sk4f::Min(Sk4f::Max(1023.0f * db, 0.0f), 1023.0f);
+
+    Sk4i ir = Sk4f_round(dr);
+    Sk4i ig = Sk4f_round(dg);
+    Sk4i ib = Sk4f_round(db);
+
+    Sk4i da = Sk4i::Load(src) & 0xFF000000;
+
+    uint32_t* dst32 = (uint32_t*) dst;
+    dst32[0] = dstTables[0][ir[0]] << kRShift
+             | dstTables[1][ig[0]] << kGShift
+             | dstTables[2][ib[0]] << kBShift
+             | da[0];
+    dst32[1] = dstTables[0][ir[1]] << kRShift
+             | dstTables[1][ig[1]] << kGShift
+             | dstTables[2][ib[1]] << kBShift
+             | da[1];
+    dst32[2] = dstTables[0][ir[2]] << kRShift
+             | dstTables[1][ig[2]] << kGShift
+             | dstTables[2][ib[2]] << kBShift
+             | da[2];
+    dst32[3] = dstTables[0][ir[3]] << kRShift
+             | dstTables[1][ig[3]] << kGShift
+             | dstTables[2][ib[3]] << kBShift
+             | da[3];
+}
+
+template <Order kOrder>
+static AI void store_generic_1(void* dst, const uint32_t* src,
+                               Sk4f& rgba, const Sk4f&,
+                               const uint8_t* const dstTables[3]) {
+    int kRShift, kGShift = 8, kBShift;
+    set_rb_shifts(kOrder, &kRShift, &kBShift);
+    rgba = Sk4f::Min(Sk4f::Max(1023.0f * rgba, 0.0f), 1023.0f);
+
+    Sk4i indices = Sk4f_round(rgba);
+
+    *((uint32_t*) dst) = dstTables[0][indices[0]] << kRShift
+                       | dstTables[1][indices[1]] << kGShift
+                       | dstTables[2][indices[2]] << kBShift
+                       | (*src & 0xFF000000);
+}
+
+typedef decltype(load_rgb_from_tables<kRGBA_Order>  )* LoadFn;
+typedef decltype(load_rgb_from_tables_1<kRGBA_Order>)* Load1Fn;
+typedef decltype(store_generic<kRGBA_Order>         )* StoreFn;
+typedef decltype(store_generic_1<kRGBA_Order>       )* Store1Fn;
+
+enum SrcFormat {
+    kRGBA_8888_Linear_SrcFormat,
+    kRGBA_8888_Table_SrcFormat,
+    kBGRA_8888_Linear_SrcFormat,
+    kBGRA_8888_Table_SrcFormat,
+};
+
+enum DstFormat {
+    kRGBA_8888_Linear_DstFormat,
+    kRGBA_8888_SRGB_DstFormat,
+    kRGBA_8888_2Dot2_DstFormat,
+    kRGBA_8888_Table_DstFormat,
+    kBGRA_8888_Linear_DstFormat,
+    kBGRA_8888_SRGB_DstFormat,
+    kBGRA_8888_2Dot2_DstFormat,
+    kBGRA_8888_Table_DstFormat,
+    kF16_Linear_DstFormat,
+};
+
+template <SrcFormat kSrc,
+          DstFormat kDst,
+          SkAlphaType kAlphaType,
+          ColorSpaceMatch kCSM>
+static void color_xform_RGBA(void* dst, const void* vsrc, int len,
+                             const float* const srcTables[3], const float matrix[13],
+                             const uint8_t* const dstTables[3]) {
+    LoadFn load;
+    Load1Fn load_1;
+    const bool kLoadAlpha = kF16_Linear_DstFormat == kDst && kOpaque_SkAlphaType != kAlphaType;
+    switch (kSrc) {
+        case kRGBA_8888_Linear_SrcFormat:
+            if (kLoadAlpha) {
+                load = load_rgba_linear<kRGBA_Order>;
+                load_1 = load_rgba_linear_1<kRGBA_Order>;
+            } else {
+                load = load_rgb_linear<kRGBA_Order>;
+                load_1 = load_rgb_linear_1<kRGBA_Order>;
+            }
+            break;
+        case kRGBA_8888_Table_SrcFormat:
+            if (kLoadAlpha) {
+                load = load_rgba_from_tables<kRGBA_Order>;
+                load_1 = load_rgba_from_tables_1<kRGBA_Order>;
+            } else {
+                load = load_rgb_from_tables<kRGBA_Order>;
+                load_1 = load_rgb_from_tables_1<kRGBA_Order>;
+            }
+            break;
+        case kBGRA_8888_Linear_SrcFormat:
+            if (kLoadAlpha) {
+                load = load_rgba_linear<kBGRA_Order>;
+                load_1 = load_rgba_linear_1<kBGRA_Order>;
+            } else {
+                load = load_rgb_linear<kBGRA_Order>;
+                load_1 = load_rgb_linear_1<kBGRA_Order>;
+            }
+            break;
+        case kBGRA_8888_Table_SrcFormat:
+            if (kLoadAlpha) {
+                load = load_rgba_from_tables<kBGRA_Order>;
+                load_1 = load_rgba_from_tables_1<kBGRA_Order>;
+            } else {
+                load = load_rgb_from_tables<kBGRA_Order>;
+                load_1 = load_rgb_from_tables_1<kBGRA_Order>;
+            }
+            break;
+    }
+
+    StoreFn store;
+    Store1Fn store_1;
+    size_t sizeOfDstPixel;
+    switch (kDst) {
+        case kRGBA_8888_Linear_DstFormat:
+            store   = store_linear<kRGBA_Order>;
+            store_1 = store_linear_1<kRGBA_Order>;
+            sizeOfDstPixel = 4;
+            break;
+        case kRGBA_8888_SRGB_DstFormat:
+            store   = store_srgb<kRGBA_Order>;
+            store_1 = store_srgb_1<kRGBA_Order>;
+            sizeOfDstPixel = 4;
+            break;
+        case kRGBA_8888_2Dot2_DstFormat:
+            store   = store_2dot2<kRGBA_Order>;
+            store_1 = store_2dot2_1<kRGBA_Order>;
+            sizeOfDstPixel = 4;
+            break;
+        case kRGBA_8888_Table_DstFormat:
+            store   = store_generic<kRGBA_Order>;
+            store_1 = store_generic_1<kRGBA_Order>;
+            sizeOfDstPixel = 4;
+            break;
+        case kBGRA_8888_Linear_DstFormat:
+            store   = store_linear<kBGRA_Order>;
+            store_1 = store_linear_1<kBGRA_Order>;
+            sizeOfDstPixel = 4;
+            break;
+        case kBGRA_8888_SRGB_DstFormat:
+            store   = store_srgb<kBGRA_Order>;
+            store_1 = store_srgb_1<kBGRA_Order>;
+            sizeOfDstPixel = 4;
+            break;
+        case kBGRA_8888_2Dot2_DstFormat:
+            store   = store_2dot2<kBGRA_Order>;
+            store_1 = store_2dot2_1<kBGRA_Order>;
+            sizeOfDstPixel = 4;
+            break;
+        case kBGRA_8888_Table_DstFormat:
+            store   = store_generic<kBGRA_Order>;
+            store_1 = store_generic_1<kBGRA_Order>;
+            sizeOfDstPixel = 4;
+            break;
+        case kF16_Linear_DstFormat:
+            store   = (kOpaque_SkAlphaType == kAlphaType) ? store_f16_opaque<kRGBA_Order> :
+                                                            store_f16<kRGBA_Order>;
+            store_1 = (kOpaque_SkAlphaType == kAlphaType) ? store_f16_1_opaque<kRGBA_Order> :
+                                                            store_f16_1<kRGBA_Order>;
+            sizeOfDstPixel = 8;
+            break;
+    }
+
+    const uint32_t* src = (const uint32_t*) vsrc;
+    Sk4f rXgXbX, rYgYbY, rZgZbZ, rTgTbT;
+    load_matrix(matrix, rXgXbX, rYgYbY, rZgZbZ, rTgTbT);
+
+    if (len >= 4) {
+        // Naively this would be a loop of load-transform-store, but we found it faster to
+        // move the N+1th load ahead of the Nth store.  We don't bother doing this for N<4.
+        Sk4f r, g, b, a;
+        load(src, r, g, b, a, srcTables);
+        src += 4;
+        len -= 4;
+
+        Sk4f dr, dg, db, da;
+        while (len >= 4) {
+            if (kNone_ColorSpaceMatch == kCSM) {
+                transform_gamut(r, g, b, a, rXgXbX, rYgYbY, rZgZbZ, dr, dg, db, da);
+                translate_gamut(rTgTbT, dr, dg, db);
+            } else {
+                dr = r;
+                dg = g;
+                db = b;
+                da = a;
+            }
+
+            load(src, r, g, b, a, srcTables);
+
+            store(dst, src - 4, dr, dg, db, da, dstTables);
+            dst = SkTAddOffset<void>(dst, 4 * sizeOfDstPixel);
+            src += 4;
+            len -= 4;
+        }
+
+        if (kNone_ColorSpaceMatch == kCSM) {
+            transform_gamut(r, g, b, a, rXgXbX, rYgYbY, rZgZbZ, dr, dg, db, da);
+            translate_gamut(rTgTbT, dr, dg, db);
+        } else {
+            dr = r;
+            dg = g;
+            db = b;
+            da = a;
+        }
+
+        store(dst, src - 4, dr, dg, db, da, dstTables);
+        dst = SkTAddOffset<void>(dst, 4 * sizeOfDstPixel);
+    }
+
+    while (len > 0) {
+        Sk4f r, g, b, a;
+        load_1(src, r, g, b, a, srcTables);
+
+        Sk4f rgba;
+        if (kNone_ColorSpaceMatch == kCSM) {
+            transform_gamut_1(r, g, b, rXgXbX, rYgYbY, rZgZbZ, rgba);
+            translate_gamut_1(rTgTbT, rgba);
+        } else {
+            rgba = Sk4f(r[0], g[0], b[0], a[0]);
+        }
+
+        store_1(dst, src, rgba, a, dstTables);
+
+        src += 1;
+        len -= 1;
+        dst = SkTAddOffset<void>(dst, sizeOfDstPixel);
+    }
+}
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+static AI int num_tables(SkColorSpace_XYZ* space) {
     switch (space->gammaNamed()) {
         case kSRGB_SkGammaNamed:
         case k2Dot2Curve_SkGammaNamed:
@@ -427,31 +990,145 @@ SkColorSpaceXform_XYZ<kCSM>
 
 ///////////////////////////////////////////////////////////////////////////////////////////////////
 
-template <ColorSpaceMatch kCSM>
-bool SkColorSpaceXform_XYZ<kCSM>::onApply(ColorFormat dstColorFormat, void* dst,
-                                          ColorFormat srcColorFormat, const void* src,
-                                          int len, SkAlphaType alphaType) const {
-    if (kFull_ColorSpaceMatch == kCSM && kPremul_SkAlphaType != alphaType) {
-        if ((kRGBA_8888_ColorFormat == dstColorFormat &&
-             kRGBA_8888_ColorFormat == srcColorFormat) ||
-            (kBGRA_8888_ColorFormat == dstColorFormat &&
-             kBGRA_8888_ColorFormat == srcColorFormat))
-        {
-            memcpy(dst, src, len * sizeof(uint32_t));
+template <SrcFormat kSrc, DstFormat kDst, ColorSpaceMatch kCSM>
+static AI bool apply_set_alpha(void* dst, const void* src, int len, SkAlphaType alphaType,
+                               const float* const srcTables[3], const float matrix[13],
+                               const uint8_t* const dstTables[3]) {
+    switch (alphaType) {
+        case kOpaque_SkAlphaType:
+            color_xform_RGBA<kSrc, kDst, kOpaque_SkAlphaType, kCSM>
+                    (dst, src, len, srcTables, matrix, dstTables);
             return true;
-        }
-
-        if ((kRGBA_8888_ColorFormat == dstColorFormat &&
-             kBGRA_8888_ColorFormat == srcColorFormat) ||
-            (kBGRA_8888_ColorFormat == dstColorFormat &&
-             kRGBA_8888_ColorFormat == srcColorFormat))
-        {
-            SkOpts::RGBA_to_BGRA((uint32_t*)dst, src, len);
+        case kUnpremul_SkAlphaType:
+            color_xform_RGBA<kSrc, kDst, kUnpremul_SkAlphaType, kCSM>
+                    (dst, src, len, srcTables, matrix, dstTables);
             return true;
+        default:
+            return false;
+    }
+}
+
+template <DstFormat kDst, ColorSpaceMatch kCSM>
+static AI bool apply_set_src(void* dst, const void* src, int len, SkAlphaType alphaType,
+                             const float* const srcTables[3], const float matrix[13],
+                             const uint8_t* const dstTables[3],
+                             SkColorSpaceXform::ColorFormat srcColorFormat,
+                             SrcGamma srcGamma) {
+    switch (srcColorFormat) {
+        case SkColorSpaceXform::kRGBA_8888_ColorFormat:
+            switch (srcGamma) {
+                case kLinear_SrcGamma:
+                    return apply_set_alpha<kRGBA_8888_Linear_SrcFormat, kDst, kCSM>
+                            (dst, src, len, alphaType, nullptr, matrix, dstTables);
+                default:
+                    return apply_set_alpha<kRGBA_8888_Table_SrcFormat, kDst, kCSM>
+                            (dst, src, len, alphaType, srcTables, matrix, dstTables);
+            }
+        case SkColorSpaceXform::kBGRA_8888_ColorFormat:
+            switch (srcGamma) {
+                case kLinear_SrcGamma:
+                    return apply_set_alpha<kBGRA_8888_Linear_SrcFormat, kDst, kCSM>
+                            (dst, src, len, alphaType, nullptr, matrix, dstTables);
+                default:
+                    return apply_set_alpha<kBGRA_8888_Table_SrcFormat, kDst, kCSM>
+                            (dst, src, len, alphaType, srcTables, matrix, dstTables);
+            }
+        default:
+            return false;
+    }
+}
+
+#undef AI
+
+template <ColorSpaceMatch kCSM>
+bool SkColorSpaceXform_XYZ<kCSM>
+::onApply(ColorFormat dstColorFormat, void* dst, ColorFormat srcColorFormat, const void* src,
+          int len, SkAlphaType alphaType) const
+{
+    if (kFull_ColorSpaceMatch == kCSM) {
+        if (kPremul_SkAlphaType != alphaType) {
+            if ((kRGBA_8888_ColorFormat == dstColorFormat &&
+                 kRGBA_8888_ColorFormat == srcColorFormat) ||
+                (kBGRA_8888_ColorFormat == dstColorFormat &&
+                 kBGRA_8888_ColorFormat == srcColorFormat))
+            {
+                memcpy(dst, src, len * sizeof(uint32_t));
+                return true;
+            }
+            if ((kRGBA_8888_ColorFormat == dstColorFormat &&
+                 kBGRA_8888_ColorFormat == srcColorFormat) ||
+                (kBGRA_8888_ColorFormat == dstColorFormat &&
+                 kRGBA_8888_ColorFormat == srcColorFormat))
+            {
+                SkOpts::RGBA_to_BGRA((uint32_t*) dst, src, len);
+                return true;
+            }
         }
     }
 
-    return this->applyPipeline(dstColorFormat, dst, srcColorFormat, src, len, alphaType);
+    if (kRGBA_F32_ColorFormat == dstColorFormat ||
+        kBGR_565_ColorFormat == dstColorFormat ||
+        kRGBA_F32_ColorFormat == srcColorFormat ||
+        kRGBA_F16_ColorFormat == srcColorFormat ||
+        kRGBA_U16_BE_ColorFormat == srcColorFormat ||
+        kRGB_U16_BE_ColorFormat == srcColorFormat ||
+        kPremul_SkAlphaType == alphaType)
+    {
+        return this->applyPipeline(dstColorFormat, dst, srcColorFormat, src, len, alphaType);
+    }
+
+    switch (dstColorFormat) {
+        case kRGBA_8888_ColorFormat:
+            switch (fDstGamma) {
+                case kLinear_DstGamma:
+                    return apply_set_src<kRGBA_8888_Linear_DstFormat, kCSM>
+                            (dst, src, len, alphaType, fSrcGammaTables, fSrcToDst, nullptr,
+                             srcColorFormat, fSrcGamma);
+                case kSRGB_DstGamma:
+                    return apply_set_src<kRGBA_8888_SRGB_DstFormat, kCSM>
+                            (dst, src, len, alphaType, fSrcGammaTables, fSrcToDst, nullptr,
+                             srcColorFormat, fSrcGamma);
+                case k2Dot2_DstGamma:
+                    return apply_set_src<kRGBA_8888_2Dot2_DstFormat, kCSM>
+                            (dst, src, len, alphaType, fSrcGammaTables, fSrcToDst, nullptr,
+                             srcColorFormat, fSrcGamma);
+                case kTable_DstGamma:
+                    return apply_set_src<kRGBA_8888_Table_DstFormat, kCSM>
+                            (dst, src, len, alphaType, fSrcGammaTables, fSrcToDst, fDstGammaTables,
+                             srcColorFormat, fSrcGamma);
+            }
+        case kBGRA_8888_ColorFormat:
+            switch (fDstGamma) {
+                case kLinear_DstGamma:
+                    return apply_set_src<kBGRA_8888_Linear_DstFormat, kCSM>
+                            (dst, src, len, alphaType, fSrcGammaTables, fSrcToDst, nullptr,
+                             srcColorFormat, fSrcGamma);
+                case kSRGB_DstGamma:
+                    return apply_set_src<kBGRA_8888_SRGB_DstFormat, kCSM>
+                            (dst, src, len, alphaType, fSrcGammaTables, fSrcToDst, nullptr,
+                             srcColorFormat, fSrcGamma);
+                case k2Dot2_DstGamma:
+                    return apply_set_src<kBGRA_8888_2Dot2_DstFormat, kCSM>
+                            (dst, src, len, alphaType, fSrcGammaTables, fSrcToDst, nullptr,
+                             srcColorFormat, fSrcGamma);
+                case kTable_DstGamma:
+                    return apply_set_src<kBGRA_8888_Table_DstFormat, kCSM>
+                            (dst, src, len, alphaType, fSrcGammaTables, fSrcToDst, fDstGammaTables,
+                             srcColorFormat, fSrcGamma);
+            }
+        case kRGBA_F16_ColorFormat:
+            switch (fDstGamma) {
+                case kLinear_DstGamma:
+                    return apply_set_src<kF16_Linear_DstFormat, kCSM>
+                            (dst, src, len, alphaType, fSrcGammaTables, fSrcToDst, nullptr,
+                             srcColorFormat, fSrcGamma);
+                default:
+                    return false;
+            }
+        default:
+            SkASSERT(false);
+            return false;
+    }
 }
 
 bool SkColorSpaceXform::apply(ColorFormat dstColorFormat, void* dst, ColorFormat srcColorFormat,