SkNx: add Sk4u

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
 
@@ skipping 112 matching lines @@
     #else
         return _mm_or_ps(_mm_and_ps   (fVec, t.fVec),
                          _mm_andnot_ps(fVec, e.fVec));
     #endif
     }
 
     __m128 fVec;
 };
 
 template <>
-class SkNx<4, int> {
+class SkNx<4, int32_t> {
 public:
     SkNx(const __m128i& vec) : fVec(vec) {}
 
     SkNx() {}
-    SkNx(int val) : fVec(_mm_set1_epi32(val)) {}
+    SkNx(int32_t val) : fVec(_mm_set1_epi32(val)) {}
     static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); }
-    SkNx(int a, int b, int c, int d) : fVec(_mm_setr_epi32(a,b,c,d)) {}
+    SkNx(int32_t a, int32_t b, int32_t c, int32_t d) : fVec(_mm_setr_epi32(a,b,c,d)) {}
 
     void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); }
 
     SkNx operator + (const SkNx& o) const { return _mm_add_epi32(fVec, o.fVec); }
     SkNx operator - (const SkNx& o) const { return _mm_sub_epi32(fVec, o.fVec); }
     SkNx operator * (const SkNx& o) const {
         __m128i mul20 = _mm_mul_epu32(fVec, o.fVec),
                 mul31 = _mm_mul_epu32(_mm_srli_si128(fVec, 4), _mm_srli_si128(o.fVec, 4));
         return _mm_unpacklo_epi32(_mm_shuffle_epi32(mul20, _MM_SHUFFLE(0,0,2,0)),
                                   _mm_shuffle_epi32(mul31, _MM_SHUFFLE(0,0,2,0)));
     }
 
     SkNx operator & (const SkNx& o) const { return _mm_and_si128(fVec, o.fVec); }
     SkNx operator | (const SkNx& o) const { return _mm_or_si128(fVec, o.fVec); }
     SkNx operator ^ (const SkNx& o) const { return _mm_xor_si128(fVec, o.fVec); }
 
     SkNx operator << (int bits) const { return _mm_slli_epi32(fVec, bits); }
     SkNx operator >> (int bits) const { return _mm_srai_epi32(fVec, bits); }
 
     SkNx operator == (const SkNx& o) const { return _mm_cmpeq_epi32 (fVec, o.fVec); }
     SkNx operator  < (const SkNx& o) const { return _mm_cmplt_epi32 (fVec, o.fVec); }
     SkNx operator  > (const SkNx& o) const { return _mm_cmpgt_epi32 (fVec, o.fVec); }
 
-    int operator[](int k) const {
+    int32_t operator[](int k) const {
         SkASSERT(0 <= k && k < 4);
-        union { __m128i v; int is[4]; } pun = {fVec};
+        union { __m128i v; int32_t is[4]; } pun = {fVec};
         return pun.is[k&3];
     }
 
     SkNx thenElse(const SkNx& t, const SkNx& e) const {
     #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
         return _mm_blendv_epi8(e.fVec, t.fVec, fVec);
     #else
         return _mm_or_si128(_mm_and_si128   (fVec, t.fVec),
                             _mm_andnot_si128(fVec, e.fVec));
     #endif
     }
 
     __m128i fVec;
 };
 
 template <>
+class SkNx<4, uint32_t> {
+public:
+    SkNx(const __m128i& vec) : fVec(vec) {}
+
+    SkNx() {}
+    SkNx(uint32_t val) : fVec(_mm_set1_epi32(val)) {}
+    static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); }
+    SkNx(uint32_t a, uint32_t b, uint32_t c, uint32_t d) : fVec(_mm_setr_epi32(a,b,c,d)) {}
+
+    void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); }
+
+    SkNx operator + (const SkNx& o) const { return _mm_add_epi32(fVec, o.fVec); }
+    SkNx operator - (const SkNx& o) const { return _mm_sub_epi32(fVec, o.fVec); }
+    // Not quite sure how to best do operator * in SSE2.  We probably don't use it.
+
+    SkNx operator & (const SkNx& o) const { return _mm_and_si128(fVec, o.fVec); }
+    SkNx operator | (const SkNx& o) const { return _mm_or_si128(fVec, o.fVec); }
+    SkNx operator ^ (const SkNx& o) const { return _mm_xor_si128(fVec, o.fVec); }
+
+    SkNx operator << (int bits) const { return _mm_slli_epi32(fVec, bits); }
+    SkNx operator >> (int bits) const { return _mm_srli_epi32(fVec, bits); }
+
+    SkNx operator == (const SkNx& o) const { return _mm_cmpeq_epi32 (fVec, o.fVec); }
+    // operator < and > take a little extra fiddling to make work for unsigned ints.
+
+    uint32_t operator[](int k) const {
+        SkASSERT(0 <= k && k < 4);
+        union { __m128i v; uint32_t us[4]; } pun = {fVec};
+        return pun.us[k&3];
+    }
+
+    SkNx thenElse(const SkNx& t, const SkNx& e) const {
+    #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
+        return _mm_blendv_epi8(e.fVec, t.fVec, fVec);
+    #else
+        return _mm_or_si128(_mm_and_si128   (fVec, t.fVec),
+                            _mm_andnot_si128(fVec, e.fVec));
+    #endif
+    }
+
+    __m128i fVec;
+};
+
+
+template <>
 class SkNx<4, uint16_t> {
 public:
     SkNx(const __m128i& vec) : fVec(vec) {}
 
     SkNx() {}
     SkNx(uint16_t val) : fVec(_mm_set1_epi16(val)) {}
     static SkNx Load(const void* ptr) { return _mm_loadl_epi64((const __m128i*)ptr); }
     SkNx(uint16_t a, uint16_t b, uint16_t c, uint16_t d) : fVec(_mm_setr_epi16(a,b,c,d,0,0,0,0)) {}
 
     void store(void* ptr) const { _mm_storel_epi64((__m128i*)ptr, fVec); }
@@ skipping 115 matching lines @@
     }
 
     SkNx thenElse(const SkNx& t, const SkNx& e) const {
         return _mm_or_si128(_mm_and_si128   (fVec, t.fVec),
                             _mm_andnot_si128(fVec, e.fVec));
     }
 
     __m128i fVec;
 };
 
-template<> /*static*/ inline Sk4f SkNx_cast<float, int>(const Sk4i& src) {
+template<> /*static*/ inline Sk4f SkNx_cast<float, int32_t>(const Sk4i& src) {
     return _mm_cvtepi32_ps(src.fVec);
 }
+template<> /*static*/ inline Sk4f SkNx_cast<float, uint32_t>(const Sk4u& src) {
+    return SkNx_cast<float>(Sk4i::Load(&src));
+}
 
-template <> /*static*/ inline Sk4i SkNx_cast<int, float>(const Sk4f& src) {
+template <> /*static*/ inline Sk4i SkNx_cast<int32_t, float>(const Sk4f& src) {
     return _mm_cvttps_epi32(src.fVec);
 }
 
-template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, int>(const Sk4i& src) {
+template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, int32_t>(const Sk4i& src) {
 #if 0 && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
     // TODO: This seems to be causing code generation problems.   Investigate?
     return _mm_packus_epi32(src.fVec);
 #elif SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
     // With SSSE3, we can just shuffle the low 2 bytes from each lane right into place.
     const int _ = ~0;
     return _mm_shuffle_epi8(src.fVec, _mm_setr_epi8(0,1, 4,5, 8,9, 12,13, _,_,_,_,_,_,_,_));
 #else
     // With SSE2, we have to sign extend our input, making _mm_packs_epi32 do the pack we want.
     __m128i x = _mm_srai_epi32(_mm_slli_epi32(src.fVec, 16), 16);
     return _mm_packs_epi32(x,x);
 #endif
 }
 
 template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, float>(const Sk4f& src) {
-    return SkNx_cast<uint16_t>(SkNx_cast<int>(src));
+    return SkNx_cast<uint16_t>(SkNx_cast<int32_t>(src));
 }
 
 template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, float>(const Sk4f& src) {
     auto _32 = _mm_cvttps_epi32(src.fVec);
 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
     const int _ = ~0;
     return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,4,8,12, _,_,_,_, _,_,_,_, _,_,_,_));
 #else
     auto _16 = _mm_packus_epi16(_32, _32);
     return     _mm_packus_epi16(_16, _16);
@@ skipping 31 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);
 }
 
-template<> /*static*/ inline Sk4i SkNx_cast<int, uint16_t>(const Sk4h& src) {
+template<> /*static*/ inline Sk4i SkNx_cast<int32_t, uint16_t>(const Sk4h& src) {
     return _mm_unpacklo_epi16(src.fVec, _mm_setzero_si128());
 }
 
-template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, int>(const Sk4i& src) {
+template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, int32_t>(const Sk4i& src) {
     return _mm_packus_epi16(_mm_packus_epi16(src.fVec, src.fVec), src.fVec);
 }
 
 static inline Sk4i Sk4f_round(const Sk4f& x) {
     return _mm_cvtps_epi32(x.fVec);
 }
 
 static inline void Sk4h_load4(const void* ptr, Sk4h* r, Sk4h* g, Sk4h* b, Sk4h* a) {
     __m128i lo = _mm_loadu_si128(((__m128i*)ptr) + 0),
             hi = _mm_loadu_si128(((__m128i*)ptr) + 1);
@@ skipping 11 matching lines @@
                                const Sk4h& a) {
     __m128i rg = _mm_unpacklo_epi16(r.fVec, g.fVec);
     __m128i ba = _mm_unpacklo_epi16(b.fVec, a.fVec);
     __m128i lo = _mm_unpacklo_epi32(rg, ba);
     __m128i hi = _mm_unpackhi_epi32(rg, ba);
     _mm_storeu_si128(((__m128i*) dst) + 0, lo);
     _mm_storeu_si128(((__m128i*) dst) + 1, hi);
 }
 
 #endif//SkNx_sse_DEFINED