Support Float32 output from SkColorSpaceXform

* Adds Float32 support to SkColorSpaceXform
* Changes API to allows clients to ask for F32, updates clients to
  new API
* Adds Sk4f_load4 and Sk4f_store4 to SkNx
* Make use of new xform in SkGr.cpp

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2339233003
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot

Committed: https://skia.googlesource.com/skia/+/43d6651111374b5d1e4ddd9030dcf079b448ec47
Review-Url: https://codereview.chromium.org/2339233003

3 files changed, 87 insertions, 24 deletions

diff --git a/src/core/SkColorSpaceXform.cpp b/src/core/SkColorSpaceXform.cpp
index 599adc90f9..b1dda34817 100644
--- a/src/core/SkColorSpaceXform.cpp
+++ b/src/core/SkColorSpaceXform.cpp
@@ -908,8 +908,23 @@ static inline void store_f16_1(void* dst, const uint32_t* src,
 }
 
 template <SwapRB kSwapRB>
+static inline void store_f32(void* dst, const uint32_t* src,
+                             Sk4f& dr, Sk4f& dg, Sk4f& db, Sk4f& da,
+                             const uint8_t* const[3]) {
+    Sk4f_store4(dst, dr, dg, db, da);
+}
+
+template <SwapRB kSwapRB>
+static inline void store_f32_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]);
+    rgba.store((float*) dst);
+}
+
+template <SwapRB kSwapRB>
 static inline void store_f16_opaque(void* dst, const uint32_t* src,
-                                    Sk4f& dr, Sk4f& dg, Sk4f& db, Sk4f& da,
+                                    Sk4f& dr, Sk4f& dg, Sk4f& db, Sk4f&,
                                     const uint8_t* const[3]) {
     Sk4h_store4(dst, SkFloatToHalf_finite_ftz(dr),
                      SkFloatToHalf_finite_ftz(dg),
@@ -919,7 +934,7 @@ static inline void store_f16_opaque(void* dst, const uint32_t* src,
 
 template <SwapRB kSwapRB>
 static inline void store_f16_1_opaque(void* dst, const uint32_t* src,
-                                      Sk4f& rgba, const Sk4f& a,
+                                      Sk4f& rgba, const Sk4f&,
                                       const uint8_t* const[3]) {
     uint64_t tmp;
     SkFloatToHalf_finite_ftz(rgba).store(&tmp);
@@ -1089,6 +1104,7 @@ enum DstFormat {
     k8888_2Dot2_DstFormat,
     k8888_Table_DstFormat,
     kF16_Linear_DstFormat,
+    kF32_Linear_DstFormat,
 };
 
 template <SrcFormat kSrc,
@@ -1101,9 +1117,12 @@ static void color_xform_RGBA(void* dst, const uint32_t* src, int len,
                              const uint8_t* const dstTables[3]) {
     LoadFn load;
     Load1Fn load_1;
+    static constexpr bool loadAlpha = (kPremul_SkAlphaType == kAlphaType) ||
+                                      (kF16_Linear_DstFormat == kDst) ||
+                                      (kF32_Linear_DstFormat == kDst);
     switch (kSrc) {
         case kRGBA_8888_Linear_SrcFormat:
-            if (kPremul_SkAlphaType == kAlphaType || kF16_Linear_DstFormat == kDst) {
+            if (loadAlpha) {
                 load = load_rgba_linear;
                 load_1 = load_rgba_linear_1;
             } else {
@@ -1112,7 +1131,7 @@ static void color_xform_RGBA(void* dst, const uint32_t* src, int len,
             }
             break;
         case kRGBA_8888_Table_SrcFormat:
-            if (kPremul_SkAlphaType == kAlphaType || kF16_Linear_DstFormat == kDst) {
+            if (loadAlpha) {
                 load = load_rgba_from_tables;
                 load_1 = load_rgba_from_tables_1;
             } else {
@@ -1153,6 +1172,11 @@ static void color_xform_RGBA(void* dst, const uint32_t* src, int len,
                                                             store_f16_1<kSwapRB>;
             sizeOfDstPixel = 8;
             break;
+        case kF32_Linear_DstFormat:
+            store   = store_f32<kSwapRB>;
+            store_1 = store_f32_1<kSwapRB>;
+            sizeOfDstPixel = 16;
+            break;
     }
 
     do_color_xform<kAlphaType, kCSM>
@@ -1245,7 +1269,7 @@ static inline void apply_set_src(void* dst, const uint32_t* src, int len, SkAlph
 
 template <SrcGamma kSrc, DstGamma kDst, ColorSpaceMatch kCSM>
 void SkColorSpaceXform_Base<kSrc, kDst, kCSM>
-::apply(void* dst, const uint32_t* src, int len, SkColorType dstColorType, SkAlphaType alphaType)
+::apply(void* dst, const uint32_t* src, int len, ColorFormat dstColorFormat, SkAlphaType alphaType)
 const
 {
     if (kFull_ColorSpaceMatch == kCSM) {
@@ -1255,13 +1279,14 @@ const
                 // linear space.
                 break;
             default:
-                switch (dstColorType) {
-                    case kRGBA_8888_SkColorType:
+                switch (dstColorFormat) {
+                    case kRGBA_8888_ColorFormat:
                         return (void) memcpy(dst, src, len * sizeof(uint32_t));
-                    case kBGRA_8888_SkColorType:
+                    case kBGRA_8888_ColorFormat:
                         return SkOpts::RGBA_to_BGRA((uint32_t*) dst, src, len);
-                    case kRGBA_F16_SkColorType:
-                        // There's still work to do to xform to linear F16.
+                    case kRGBA_F16_ColorFormat:
+                    case kRGBA_F32_ColorFormat:
+                        // There's still work to do to xform to linear floats.
                         break;
                     default:
                         SkASSERT(false);
@@ -1283,8 +1308,8 @@ const
         src = (const uint32_t*) storage.get();
     }
 
-    switch (dstColorType) {
-        case kRGBA_8888_SkColorType:
+    switch (dstColorFormat) {
+        case kRGBA_8888_ColorFormat:
             switch (kDst) {
                 case kLinear_DstGamma:
                     return apply_set_src<kSrc, k8888_Linear_DstFormat, kCSM, kNo_SwapRB>
@@ -1299,7 +1324,7 @@ const
                     return apply_set_src<kSrc, k8888_Table_DstFormat, kCSM, kNo_SwapRB>
                             (dst, src, len, alphaType, fSrcGammaTables, fSrcToDst, fDstGammaTables);
             }
-        case kBGRA_8888_SkColorType:
+        case kBGRA_8888_ColorFormat:
             switch (kDst) {
                 case kLinear_DstGamma:
                     return apply_set_src<kSrc, k8888_Linear_DstFormat, kCSM, kYes_SwapRB>
@@ -1314,7 +1339,7 @@ const
                     return apply_set_src<kSrc, k8888_Table_DstFormat, kCSM, kYes_SwapRB>
                             (dst, src, len, alphaType, fSrcGammaTables, fSrcToDst, fDstGammaTables);
             }
-        case kRGBA_F16_SkColorType:
+        case kRGBA_F16_ColorFormat:
             switch (kDst) {
                 case kLinear_DstGamma:
                     return apply_set_src<kSrc, kF16_Linear_DstFormat, kCSM, kNo_SwapRB>
@@ -1323,6 +1348,15 @@ const
                     SkASSERT(false);
                     return;
             }
+        case kRGBA_F32_ColorFormat:
+            switch (kDst) {
+                case kLinear_DstGamma:
+                    return apply_set_src<kSrc, kF32_Linear_DstFormat, kCSM, kNo_SwapRB>
+                            (dst, src, len, alphaType, fSrcGammaTables, fSrcToDst, nullptr);
+                default:
+                    SkASSERT(false);
+                    return;
+            }
         default:
             SkASSERT(false);
             return;
diff --git a/src/core/SkColorSpaceXform.h b/src/core/SkColorSpaceXform.h
index c80cd15dca..6c996a9c09 100644
--- a/src/core/SkColorSpaceXform.h
+++ b/src/core/SkColorSpaceXform.h
@@ -25,20 +25,27 @@ public:
     static std::unique_ptr<SkColorSpaceXform> New(const sk_sp<SkColorSpace>& srcSpace,
                                                   const sk_sp<SkColorSpace>& dstSpace);
 
+    enum ColorFormat : uint8_t {
+        kRGBA_8888_ColorFormat,
+        kBGRA_8888_ColorFormat,
+        kRGBA_F16_ColorFormat,
+        kRGBA_F32_ColorFormat,
+    };
+
     /**
      *  Apply the color conversion to a |src| buffer, storing the output in the |dst| buffer.
      *
-     *  @param dst          Stored in the format described by |dstColorType| and |dstAlphaType|
-     *  @param src          Stored as RGBA_8888, kUnpremul (note kOpaque is a form of kUnpremul)
-     *  @param len          Number of pixels in the buffers
-     *  @param dstColorType Describes color type of |dst|
-     *  @param dstAlphaType Describes alpha type of |dst|
-     *                      kUnpremul preserves input alpha values
-     *                      kPremul   performs a premultiplication and also preserves alpha values
-     *                      kOpaque   optimization hint, |dst| alphas set to 1
+     *  @param dst            Stored in the format described by |dstColorType| and |dstAlphaType|
+     *  @param src            Stored as RGBA_8888, kUnpremul (note kOpaque is a form of kUnpremul)
+     *  @param len            Number of pixels in the buffers
+     *  @param dstColorFormat Describes color format of |dst|
+     *  @param dstAlphaType   Describes alpha type of |dst|
+     *                        kUnpremul preserves input alpha values
+     *                        kPremul   performs a premultiplication and also preserves alpha values
+     *                        kOpaque   optimization hint, |dst| alphas set to 1
      *
      */
-    virtual void apply(void* dst, const uint32_t* src, int len, SkColorType dstColorType,
+    virtual void apply(void* dst, const uint32_t* src, int len, ColorFormat dstColorFormat,
                        SkAlphaType dstAlphaType) const = 0;
 
     virtual ~SkColorSpaceXform() {}
@@ -66,7 +73,7 @@ template <SrcGamma kSrc, DstGamma kDst, ColorSpaceMatch kCSM>
 class SkColorSpaceXform_Base : public SkColorSpaceXform {
 public:
 
-    void apply(void* dst, const uint32_t* src, int len, SkColorType dstColorType,
+    void apply(void* dst, const uint32_t* src, int len, ColorFormat dstColorFormat,
                SkAlphaType dstAlphaType) const override;
 
     static constexpr int      kDstGammaTableSize = 1024;
diff --git a/src/core/SkNx.h b/src/core/SkNx.h
index 881a475ce0..6bca856d8b 100644
--- a/src/core/SkNx.h
+++ b/src/core/SkNx.h
@@ -332,6 +332,28 @@ SI void Sk4h_store4(void* dst, const Sk4h& r, const Sk4h& g, const Sk4h& b, cons
     Sk4h(r[3], g[3], b[3], a[3]).store(dst64 + 3);
 }
 
+// Load 4 Sk4f and transpose them (512 bits total).
+SI void Sk4f_load4(const void* vptr, Sk4f* r, Sk4f* g, Sk4f* b, Sk4f* a) {
+    const float* ptr = (const float*) vptr;
+    auto p0 = Sk4f::Load(ptr +  0),
+         p1 = Sk4f::Load(ptr +  4),
+         p2 = Sk4f::Load(ptr +  8),
+         p3 = Sk4f::Load(ptr + 12);
+    *r = { p0[0], p1[0], p2[0], p3[0] };
+    *g = { p0[1], p1[1], p2[1], p3[1] };
+    *b = { p0[2], p1[2], p2[2], p3[2] };
+    *a = { p0[3], p1[3], p2[3], p3[3] };
+}
+
+// Transpose 4 Sk4f and store (512 bits total).
+SI void Sk4f_store4(void* vdst, const Sk4f& r, const Sk4f& g, const Sk4f& b, const Sk4f& a) {
+    float* dst = (float*) vdst;
+    Sk4f(r[0], g[0], b[0], a[0]).store(dst +  0);
+    Sk4f(r[1], g[1], b[1], a[1]).store(dst +  4);
+    Sk4f(r[2], g[2], b[2], a[2]).store(dst +  8);
+    Sk4f(r[3], g[3], b[3], a[3]).store(dst + 12);
+}
+
 #endif
 
 SI void Sk4f_ToBytes(uint8_t p[16], const Sk4f& a, const Sk4f& b, const Sk4f& c, const Sk4f& d) {