skcpu: sse4.1 floor, f16c f16<->f32

-  floor with roundps is about 4.5x faster when available
  -  f16 srcover_n is similar to but a little faster than the version in https://codereview.chromium.org/1884683002.  This new one fuses the dst load/stores into the f16<->f32 conversions:

+0x180	    movups              (%r15), %xmm1
+0x184	    vcvtph2ps           (%rbx), %xmm2
+0x189	    movaps              %xmm1, %xmm3
+0x18c	    shufps              $255, %xmm3, %xmm3
+0x190	    movaps              %xmm0, %xmm4
+0x193	    subps               %xmm3, %xmm4
+0x196	    mulps               %xmm2, %xmm4
+0x199	    addps               %xmm1, %xmm4
+0x19c	    vcvtps2ph           $0, %xmm4, (%rbx)
+0x1a2	    addq                $16, %r15
+0x1a6	    addq                $8, %rbx
+0x1aa	    decl                %r14d
+0x1ad	    jne                 +0x180

If we decide to land this it'd be a good idea to convert most or all users of SkFloatToHalf_01 and SkHalfToFloat_01 over to the pointer-based versions.

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1891513002
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot

Committed: https://skia.googlesource.com/skia/+/cbe3c1af987d622ea67ef560d855b41bb14a0ce9

Committed: https://skia.googlesource.com/skia/+/3faf74b8364491ca806f523fbb1d8a97be592663

Review URL: https://codereview.chromium.org/1891513002

src/core/SkHalf.h

@@ -8,6 +8,7 @@
 #ifndef SkHalf_DEFINED
 #define SkHalf_DEFINED
 
+#include "SkCpu.h"
 #include "SkNx.h"
 #include "SkTypes.h"
 
@@ -122,3 +123,32 @@ static inline uint64_t SkFloatToHalf_01(const Sk4f& fs) {
 }
 
 #endif
+
+static inline Sk4f SkHalfToFloat_01(const uint64_t* hs) {
+#if !defined(SKNX_NO_SIMD) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
+    if (SkCpu::Supports(SkCpu::F16C)) {
+        __m128 fs;
+    #if defined(__GNUC__) || defined(__clang__)
+        asm("vcvtph2ps %[hs], %[fs]" : [fs]"=x"(fs) : [hs]"m"(*hs));
+    #else
+        fs = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i*)hs));
+    #endif
+        return fs;
+    }
+#endif
+    return SkHalfToFloat_01(*hs);
+}
+
+static inline void SkFloatToHalf_01(const Sk4f& fs, uint64_t* hs) {
+#if !defined(SKNX_NO_SIMD) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
+    if (SkCpu::Supports(SkCpu::F16C)) {
+    #if defined(__GNUC__) || defined(__clang__)
+        asm("vcvtps2ph $0, %[fs], %[hs]" : [hs]"=m"(*hs) : [fs]"x"(fs.fVec));
+    #else
+        _mm_storel_epi64((__m128i*)hs, _mm_cvtps_ph(fs.fVec, 0));
+    #endif
+        return;
+    }
+#endif
+    *hs = SkFloatToHalf_01(fs);
+}

src/core/SkXfermodeF16.cpp

@@ -134,15 +134,13 @@ static void srcover_1(const SkXfermode*, uint64_t dst[], const SkPM4f* src, int
 static void srcover_n(const SkXfermode*, uint64_t dst[], const SkPM4f src[], int count,
                       const SkAlpha aa[]) {
     for (int i = 0; i < count; ++i) {
-        const Sk4f s4 = Sk4f::Load(src[i].fVec);
-        const Sk4f dst_scale = Sk4f(1 - get_alpha(s4));
-        const Sk4f d4 = SkHalfToFloat_01(dst[i]);
-        const Sk4f r4 = s4 + d4 * dst_scale;
+        Sk4f s = Sk4f::Load(src+i),
+             d = SkHalfToFloat_01(dst+i),
+             r = s + d*(1.0f - SkNx_shuffle<3,3,3,3>(s));
         if (aa) {
-            dst[i] = SkFloatToHalf_01(lerp_by_coverage(r4, d4, aa[i]));
-        } else {
-            dst[i] = SkFloatToHalf_01(r4);
+            r = lerp_by_coverage(r, d, aa[i]);
         }
+        SkFloatToHalf_01(r, dst+i);
     }
 }
 

src/opts/SkNx_sse.h

@@ -8,21 +8,13 @@
 #ifndef SkNx_sse_DEFINED
 #define SkNx_sse_DEFINED
 
+#include "SkCpu.h"
+
 // This file may assume <= SSE2, but must check SK_CPU_SSE_LEVEL for anything more recent.
 // If you do, make sure this is in a static inline function... anywhere else risks violating ODR.
 
 #define SKNX_IS_FAST
 
-// SSE 4.1 has _mm_floor_ps to floor 4 floats.  We emulate it:
-//   - roundtrip through integers via truncation
-//   - subtract 1 if that's too big (possible for negative values).
-// This restricts the domain of our inputs to a maximum somehwere around 2^31.  Seems plenty big.
-static inline __m128 sse2_mm_floor_ps(__m128 v) {
-    __m128 roundtrip = _mm_cvtepi32_ps(_mm_cvttps_epi32(v));
-    __m128 too_big = _mm_cmpgt_ps(roundtrip, v);
-    return _mm_sub_ps(roundtrip, _mm_and_ps(too_big, _mm_set1_ps(1.0f)));
-}
-
 template <>
 class SkNx<2, float> {
 public:
@@ -97,7 +89,25 @@ public:
     static SkNx Max(const SkNx& l, const SkNx& r) { return _mm_max_ps(l.fVec, r.fVec); }
 
     SkNx abs() const { return _mm_andnot_ps(_mm_set1_ps(-0.0f), fVec); }
-    SkNx floor() const { return sse2_mm_floor_ps(fVec); }
+    SkNx floor() const {
+        if (SkCpu::Supports(SkCpu::SSE41)) {
+            __m128 r;
+        #if defined(__GNUC__) || defined(__clang__)
+            asm("roundps $0x1, %[fVec], %[r]" : [r]"=x"(r) : [fVec]"x"(fVec));
+        #else
+            r = _mm_floor_ps(fVec);
+        #endif
+            return r;
+        }
+        // Emulate _mm_floor_ps() with SSE2:
+        //   - roundtrip through integers via truncation
+        //   - subtract 1 if that's too big (possible for negative values).
+        // This restricts the domain of our inputs to a maximum somehwere around 2^31.
+        // Seems plenty big.
+        __m128 roundtrip = _mm_cvtepi32_ps(_mm_cvttps_epi32(fVec));
+        __m128 too_big = _mm_cmpgt_ps(roundtrip, fVec);
+        return _mm_sub_ps(roundtrip, _mm_and_ps(too_big, _mm_set1_ps(1.0f)));
+    }
 
     SkNx   sqrt() const { return _mm_sqrt_ps (fVec);  }
     SkNx  rsqrt() const { return _mm_rsqrt_ps(fVec); }