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

src/opts/SkNx_sse.h

 /*
  * Copyright 2015 Google Inc.
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #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:
     SkNx(const __m128& vec) : fVec(vec) {}
 
     SkNx() {}
     SkNx(float val) : fVec(_mm_set1_ps(val)) {}
     static SkNx Load(const void* ptr) {
         return _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*)ptr));
     }
@@ skipping 54 matching lines @@
     SkNx operator != (const SkNx& o) const { return _mm_cmpneq_ps(fVec, o.fVec); }
     SkNx operator  < (const SkNx& o) const { return _mm_cmplt_ps (fVec, o.fVec); }
     SkNx operator  > (const SkNx& o) const { return _mm_cmpgt_ps (fVec, o.fVec); }
     SkNx operator <= (const SkNx& o) const { return _mm_cmple_ps (fVec, o.fVec); }
     SkNx operator >= (const SkNx& o) const { return _mm_cmpge_ps (fVec, o.fVec); }
 
     static SkNx Min(const SkNx& l, const SkNx& r) { return _mm_min_ps(l.fVec, r.fVec); }
     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); }
     SkNx invert() const { return _mm_rcp_ps(fVec); }
 
     float operator[](int k) const {
         SkASSERT(0 <= k && k < 4);
         union { __m128 v; float fs[4]; } pun = {fVec};
         return pun.fs[k&3];
     }
@@ skipping 243 matching lines @@
 
 template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, uint8_t>(const Sk4b& src) {
     return _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128());
 }
 
 template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src) {
     return _mm_packus_epi16(src.fVec, src.fVec);
 }
 
 #endif//SkNx_sse_DEFINED