skrpb: evaluate color filters for constant shaders once.

The simplest thing to do here is just run shader+color filter pipeline at
construction time to create a new constant color shader (replacing the paint
color).

This reduces a pipeline like:
  - constant_color  (paint color)
  - matrix_4x5
  - clamp_a
  - load_d_foo, xfermode, lerp, store_foo
to
  - constant_color  (paint color -> matrix_4x5 -> clamp_a)
  - load_d_foo, xfermode, lerp, store_foo

To implement this all, we add a new store_f32 stage that writes SkPM4f, and
finally get around to implementing Sk8f::Store4() (store while reinterlacing).
Sk4f::Store4() already exists for both SSE and NEON.

Next step: reduce simple constant_color -> store pipelines (src mode, full
coverage) into non-pipeline memsets.
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2480823002

Review-Url: https://codereview.chromium.org/2480823002

2 files changed, 52 insertions, 3 deletions

diff --git a/src/opts/SkNx_sse.h b/src/opts/SkNx_sse.h
index a4594115e0..a4783c6302 100644
--- a/src/opts/SkNx_sse.h
+++ b/src/opts/SkNx_sse.h
@@ -544,6 +544,14 @@ public:
         __m256i fVec;
     };
 
+    // _mm256_unpack{lo,hi}_pd() auto-casting to and from __m256d.
+    AI static __m256 unpacklo_pd(__m256 x, __m256 y) {
+        return _mm256_castpd_ps(_mm256_unpacklo_pd(_mm256_castps_pd(x), _mm256_castps_pd(y)));
+    }
+    AI static __m256 unpackhi_pd(__m256 x, __m256 y) {
+        return _mm256_castpd_ps(_mm256_unpackhi_pd(_mm256_castps_pd(x), _mm256_castps_pd(y)));
+    }
+
     template <>
     class SkNx<8, float> {
     public:
@@ -560,6 +568,29 @@ public:
         AI static SkNx Load(const void* ptr) { return _mm256_loadu_ps((const float*)ptr); }
         AI void store(void* ptr) const { _mm256_storeu_ps((float*)ptr, fVec); }
 
+        AI static void Store4(void* ptr,
+                              const SkNx& r, const SkNx& g, const SkNx& b, const SkNx& a) {
+            __m256 rg0145 = _mm256_unpacklo_ps(r.fVec, g.fVec),  // r0 g0 r1 g1 | r4 g4 r5 g5
+                   rg2367 = _mm256_unpackhi_ps(r.fVec, g.fVec),  // r2 ...      | r6 ...
+                   ba0145 = _mm256_unpacklo_ps(b.fVec, a.fVec),  // b0 a0 b1 a1 | b4 a4 b5 a5
+                   ba2367 = _mm256_unpackhi_ps(b.fVec, a.fVec);  // b2 ...      | b6 ...
+
+            __m256 _04 = unpacklo_pd(rg0145, ba0145),  // r0 g0 b0 a0 | r4 g4 b4 a4
+                   _15 = unpackhi_pd(rg0145, ba0145),  // r1 ...      | r5 ...
+                   _26 = unpacklo_pd(rg2367, ba2367),  // r2 ...      | r6 ...
+                   _37 = unpackhi_pd(rg2367, ba2367);  // r3 ...      | r7 ...
+
+            __m256 _01 = _mm256_permute2f128_ps(_04, _15, 16),  // 16 == 010 000 == lo, lo
+                   _23 = _mm256_permute2f128_ps(_26, _37, 16),
+                   _45 = _mm256_permute2f128_ps(_04, _15, 25),  // 25 == 011 001 == hi, hi
+                   _67 = _mm256_permute2f128_ps(_26, _37, 25);
+
+            _mm256_storeu_ps((float*)ptr + 0*8, _01);
+            _mm256_storeu_ps((float*)ptr + 1*8, _23);
+            _mm256_storeu_ps((float*)ptr + 2*8, _45);
+            _mm256_storeu_ps((float*)ptr + 3*8, _67);
+        }
+
         AI SkNx operator+(const SkNx& o) const { return _mm256_add_ps(fVec, o.fVec); }
         AI SkNx operator-(const SkNx& o) const { return _mm256_sub_ps(fVec, o.fVec); }
         AI SkNx operator*(const SkNx& o) const { return _mm256_mul_ps(fVec, o.fVec); }
diff --git a/src/opts/SkRasterPipeline_opts.h b/src/opts/SkRasterPipeline_opts.h
index bd42632b4b..155558e776 100644
--- a/src/opts/SkRasterPipeline_opts.h
+++ b/src/opts/SkRasterPipeline_opts.h
@@ -23,9 +23,9 @@ namespace {
     static constexpr int N = 4;
 #endif
 
-using SkNf = SkNx<N, float>;
-using SkNi = SkNx<N, int>;
-using SkNh = SkNx<N, uint16_t>;
+    using SkNf = SkNx<N, float>;
+    using SkNi = SkNx<N, int>;
+    using SkNh = SkNx<N, uint16_t>;
 
     struct BodyStage;
     struct TailStage;
@@ -379,6 +379,24 @@ STAGE(store_f16, false) {
     }
 }
 
+STAGE(store_f32, false) {
+    auto ptr = *(SkPM4f**)ctx + x;
+
+    SkPM4f buf[8];
+    SkNf::Store4(kIsTail ? buf : ptr, r,g,b,a);
+    if (kIsTail) {
+        switch (tail & (N-1)) {
+            case 7: ptr[6] = buf[6];
+            case 6: ptr[5] = buf[5];
+            case 5: ptr[4] = buf[4];
+            case 4: ptr[3] = buf[3];
+            case 3: ptr[2] = buf[2];
+            case 2: ptr[1] = buf[1];
+        }
+        ptr[0] = buf[0];
+    }
+}
+
 
 // Load 8-bit SkPMColor-order sRGB.
 STAGE(load_d_srgb, true) {