Optimize ARGB32->RGBA64PM better

This conversion is critical for ARGB32 painting, and no compiler
optimized the premultiplication efficiently.

Change-Id: Iee137c2f7020246478d09e880a7a1bf2ed3c6fd4
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>

3 files changed, 199 insertions, 20 deletions

diff --git a/src/gui/painting/qdrawhelper.cpp b/src/gui/painting/qdrawhelper.cpp
index af9375fcec..cdf5826672 100644
--- a/src/gui/painting/qdrawhelper.cpp
+++ b/src/gui/painting/qdrawhelper.cpp
@@ -1086,18 +1086,8 @@ static const QRgba64 *QT_FASTCALL fetchRGB32ToRGB64(QRgba64 *buffer, const uchar
 static const QRgba64 *QT_FASTCALL convertARGB32ToRGBA64PM(QRgba64 *buffer, const uint *src, int count,
                                                           const QVector<QRgb> *, QDitherInfo *)
 {
-#ifdef __SSE2__
-    qConvertARGB32PMToRGBA64PM_sse2<false, false>(buffer, src, count);
-    for (int i = 0; i < count; ++i)
-        buffer[i] = buffer[i].premultiplied();
-#elif defined(__ARM_NEON__)
-    qConvertARGB32PMToRGBA64PM_neon<false, false>(buffer, src, count);
-    for (int i = 0; i < count; ++i)
-        buffer[i] = buffer[i].premultiplied();
-#else
     for (int i = 0; i < count; ++i)
         buffer[i] = QRgba64::fromArgb32(src[i]).premultiplied();
-#endif
     return buffer;
 }
 
@@ -1149,18 +1139,8 @@ static const QRgba64 *QT_FASTCALL fetchRGBA64ToRGBA64PM(QRgba64 *buffer, const u
 static const QRgba64 *QT_FASTCALL convertRGBA8888ToRGBA64PM(QRgba64 *buffer, const uint *src, int count,
                                                             const QVector<QRgb> *, QDitherInfo *)
 {
-#ifdef __SSE2__
-    qConvertARGB32PMToRGBA64PM_sse2<true, false>(buffer, src, count);
-    for (int i = 0; i < count; ++i)
-        buffer[i] = buffer[i].premultiplied();
-#elif defined(__ARM_NEON__)
-    qConvertARGB32PMToRGBA64PM_neon<true, false>(buffer, src, count);
-    for (int i = 0; i < count; ++i)
-        buffer[i] = buffer[i].premultiplied();
-#else
     for (int i = 0; i < count; ++i)
         buffer[i] = QRgba64::fromArgb32(RGBA2ARGB(src[i])).premultiplied();
-#endif
     return buffer;
 }
 
@@ -6514,6 +6494,14 @@ static void qInitDrawhelperFunctions()
                                                                   const QVector<QRgb> *, QDitherInfo *);
         extern const uint *QT_FASTCALL fetchRGBA8888ToARGB32PM_sse4(uint *buffer, const uchar *src, int index, int count,
                                                                     const QVector<QRgb> *, QDitherInfo *);
+        extern const QRgba64 * QT_FASTCALL convertARGB32ToRGBA64PM_sse4(QRgba64 *buffer, const uint *src, int count,
+                                                                        const QVector<QRgb> *, QDitherInfo *);
+        extern const QRgba64 * QT_FASTCALL convertRGBA8888ToRGBA64PM_sse4(QRgba64 *buffer, const uint *src, int count,
+                                                                          const QVector<QRgb> *, QDitherInfo *);
+        extern const QRgba64 *QT_FASTCALL fetchARGB32ToRGBA64PM_sse4(QRgba64 *buffer, const uchar *src, int index, int count,
+                                                                     const QVector<QRgb> *, QDitherInfo *);
+        extern const QRgba64 *QT_FASTCALL fetchRGBA8888ToRGBA64PM_sse4(QRgba64 *buffer, const uchar *src, int index, int count,
+                                                                       const QVector<QRgb> *, QDitherInfo *);
         extern void QT_FASTCALL storeARGB32FromARGB32PM_sse4(uchar *dest, const uint *src, int index, int count,
                                                                       const QVector<QRgb> *, QDitherInfo *);
         extern void QT_FASTCALL storeRGBA8888FromARGB32PM_sse4(uchar *dest, const uint *src, int index, int count,
@@ -6530,8 +6518,14 @@ static void qInitDrawhelperFunctions()
         qPixelLayouts[QImage::Format_ARGB32].convertToARGB32PM = convertARGB32ToARGB32PM_sse4;
         qPixelLayouts[QImage::Format_RGBA8888].fetchToARGB32PM = fetchRGBA8888ToARGB32PM_sse4;
         qPixelLayouts[QImage::Format_RGBA8888].convertToARGB32PM = convertRGBA8888ToARGB32PM_sse4;
+        qPixelLayouts[QImage::Format_ARGB32].fetchToRGBA64PM = fetchARGB32ToRGBA64PM_sse4;
+        qPixelLayouts[QImage::Format_ARGB32].convertToRGBA64PM = convertARGB32ToRGBA64PM_sse4;
         qPixelLayouts[QImage::Format_ARGB32].storeFromARGB32PM = storeARGB32FromARGB32PM_sse4;
+        qPixelLayouts[QImage::Format_RGBA8888].fetchToRGBA64PM = fetchRGBA8888ToRGBA64PM_sse4;
+        qPixelLayouts[QImage::Format_RGBA8888].convertToRGBA64PM = convertRGBA8888ToRGBA64PM_sse4;
         qPixelLayouts[QImage::Format_RGBA8888].storeFromARGB32PM = storeRGBA8888FromARGB32PM_sse4;
+        qPixelLayouts[QImage::Format_RGBX8888].fetchToRGBA64PM = fetchRGBA8888ToRGBA64PM_sse4;
+        qPixelLayouts[QImage::Format_RGBX8888].convertToRGBA64PM = convertRGBA8888ToRGBA64PM_sse4;
         qPixelLayouts[QImage::Format_RGBX8888].storeFromARGB32PM = storeRGBXFromARGB32PM_sse4;
         qPixelLayouts[QImage::Format_A2BGR30_Premultiplied].storeFromARGB32PM = storeA2RGB30PMFromARGB32PM_sse4<PixelOrderBGR>;
         qPixelLayouts[QImage::Format_A2RGB30_Premultiplied].storeFromARGB32PM = storeA2RGB30PMFromARGB32PM_sse4<PixelOrderRGB>;
@@ -6620,6 +6614,14 @@ static void qInitDrawhelperFunctions()
                                                               const QVector<QRgb> *, QDitherInfo *);
     extern const uint *QT_FASTCALL fetchRGBA8888ToARGB32PM_neon(uint *buffer, const uchar *src, int index, int count,
                                                                 const QVector<QRgb> *, QDitherInfo *);
+   extern const QRgba64 * QT_FASTCALL convertARGB32ToRGBA64PM_neon(QRgba64 *buffer, const uint *src, int count,
+                                                                   const QVector<QRgb> *, QDitherInfo *);
+   extern const QRgba64 * QT_FASTCALL convertRGBA8888ToRGBA64PM_neon(QRgba64 *buffer, const uint *src, int count,
+                                                                     const QVector<QRgb> *, QDitherInfo *);
+   extern const QRgba64 *QT_FASTCALL fetchARGB32ToRGBA64PM_neon(QRgba64 *buffer, const uchar *src, int index, int count,
+                                                                const QVector<QRgb> *, QDitherInfo *);
+   extern const QRgba64 *QT_FASTCALL fetchRGBA8888ToRGBA64PM_neon(QRgba64 *buffer, const uchar *src, int index, int count,
+                                                                  const QVector<QRgb> *, QDitherInfo *);
     extern void QT_FASTCALL storeARGB32FromARGB32PM_neon(uchar *dest, const uint *src, int index, int count,
                                                          const QVector<QRgb> *, QDitherInfo *);
     extern void QT_FASTCALL storeRGBA8888FromARGB32PM_neon(uchar *dest, const uint *src, int index, int count,
@@ -6629,10 +6631,16 @@ static void qInitDrawhelperFunctions()
     qPixelLayouts[QImage::Format_ARGB32].fetchToARGB32PM = fetchARGB32ToARGB32PM_neon;
     qPixelLayouts[QImage::Format_ARGB32].convertToARGB32PM = convertARGB32ToARGB32PM_neon;
     qPixelLayouts[QImage::Format_ARGB32].storeFromARGB32PM = storeARGB32FromARGB32PM_neon;
+    qPixelLayouts[QImage::Format_ARGB32].fetchToRGBA64PM = fetchARGB32ToRGBA64PM_neon;
+    qPixelLayouts[QImage::Format_ARGB32].convertToRGBA64PM = convertARGB32ToRGBA64PM_neon;
     qPixelLayouts[QImage::Format_RGBA8888].fetchToARGB32PM = fetchRGBA8888ToARGB32PM_neon;
     qPixelLayouts[QImage::Format_RGBA8888].convertToARGB32PM = convertRGBA8888ToARGB32PM_neon;
     qPixelLayouts[QImage::Format_RGBA8888].storeFromARGB32PM = storeRGBA8888FromARGB32PM_neon;
+    qPixelLayouts[QImage::Format_RGBA8888].fetchToRGBA64PM = fetchRGBA8888ToRGBA64PM_neon;
+    qPixelLayouts[QImage::Format_RGBA8888].convertToRGBA64PM = convertRGBA8888ToRGBA64PM_neon;
     qPixelLayouts[QImage::Format_RGBX8888].storeFromARGB32PM = storeRGBXFromARGB32PM_neon;
+    qPixelLayouts[QImage::Format_RGBX8888].fetchToRGBA64PM = fetchRGBA8888ToRGBA64PM_neon;
+    qPixelLayouts[QImage::Format_RGBX8888].convertToRGBA64PM = convertRGBA8888ToRGBA64PM_neon;
 #endif
 
 #if defined(ENABLE_PIXMAN_DRAWHELPERS)
diff --git a/src/gui/painting/qdrawhelper_neon.cpp b/src/gui/painting/qdrawhelper_neon.cpp
index e33af3b784..3fcbcfd053 100644
--- a/src/gui/painting/qdrawhelper_neon.cpp
+++ b/src/gui/painting/qdrawhelper_neon.cpp
@@ -1149,6 +1149,72 @@ static inline void convertARGBToARGB32PM_neon(uint *buffer, const uint *src, int
     }
 }
 
+template<bool RGBA>
+static inline void convertARGB32ToRGBA64PM_neon(QRgba64 *buffer, const uint *src, int count)
+{
+    if (count <= 0)
+        return;
+
+    const uint8x8_t shuffleMask = { 3, 3, 3, 3, 7, 7, 7, 7};
+    const uint64x2_t blendMask = vdupq_n_u64(Q_UINT64_C(0xffff000000000000));
+
+    int i = 0;
+    for (; i < count-3; i += 4) {
+        uint32x4_t vs32 = vld1q_u32(src + i);
+        uint32x4_t alphaVector = vshrq_n_u32(vs32, 24);
+#if defined(Q_PROCESSOR_ARM_64)
+        uint32_t alphaSum = vaddvq_u32(alphaVector);
+#else
+        // no vaddvq_u32
+        uint32x2_t tmp = vpadd_u32(vget_low_u32(alphaVector), vget_high_u32(alphaVector));
+        uint32_t alphaSum = vget_lane_u32(vpadd_u32(tmp, tmp), 0);
+#endif
+        if (alphaSum) {
+            if (!RGBA)
+                vs32 = vrgba2argb(vs32);
+            const uint8x16_t vs8 = vreinterpretq_u8_u32(vs32);
+            const uint8x16x2_t v = vzipq_u8(vs8, vs8);
+            if (alphaSum != 255 * 4) {
+                const uint8x8_t s1 = vreinterpret_u8_u32(vget_low_u32(vs32));
+                const uint8x8_t s2 = vreinterpret_u8_u32(vget_high_u32(vs32));
+                const uint8x8_t alpha1 = vtbl1_u8(s1, shuffleMask);
+                const uint8x8_t alpha2 = vtbl1_u8(s2, shuffleMask);
+                uint16x8_t src1 = vmull_u8(s1, alpha1);
+                uint16x8_t src2 = vmull_u8(s2, alpha2);
+                // convert from 0->(255x255) to 0->(255x257)
+                src1 = vsraq_n_u16(src1, src1, 7);
+                src2 = vsraq_n_u16(src2, src2, 7);
+
+                // now restore alpha from the trivial conversion
+                const uint64x2_t d1 = vbslq_u64(blendMask, vreinterpretq_u64_u8(v.val[0]), vreinterpretq_u64_u16(src1));
+                const uint64x2_t d2 = vbslq_u64(blendMask, vreinterpretq_u64_u8(v.val[1]), vreinterpretq_u64_u16(src2));
+
+                vst1q_u16((uint16_t *)buffer, vreinterpretq_u16_u64(d1));
+                buffer += 2;
+                vst1q_u16((uint16_t *)buffer, vreinterpretq_u16_u64(d2));
+                buffer += 2;
+            } else {
+                vst1q_u16((uint16_t *)buffer, vreinterpretq_u16_u8(v.val[0]));
+                buffer += 2;
+                vst1q_u16((uint16_t *)buffer, vreinterpretq_u16_u8(v.val[1]));
+                buffer += 2;
+            }
+        } else {
+            vst1q_u16((uint16_t *)buffer, vdupq_n_u16(0));
+            buffer += 2;
+            vst1q_u16((uint16_t *)buffer, vdupq_n_u16(0));
+            buffer += 2;
+        }
+    }
+
+    SIMD_EPILOGUE(i, count, 3) {
+        uint s = src[i];
+        if (RGBA)
+            s = RGBA2ARGB(s);
+        *buffer++ = QRgba64::fromArgb32(s).premultiplied();
+    }
+}
+
 static inline float32x4_t reciprocal_mul_ps(float32x4_t a, float mul)
 {
     float32x4_t ia = vrecpeq_f32(a); // estimate 1/a
@@ -1258,6 +1324,34 @@ const uint *QT_FASTCALL fetchRGBA8888ToARGB32PM_neon(uint *buffer, const uchar *
     return buffer;
 }
 
+const QRgba64 * QT_FASTCALL convertARGB32ToRGBA64PM_neon(QRgba64 *buffer, const uint *src, int count,
+                                                         const QVector<QRgb> *, QDitherInfo *)
+{
+    convertARGB32ToRGBA64PM_neon<false>(buffer, src, count);
+    return buffer;
+}
+
+const QRgba64 * QT_FASTCALL convertRGBA8888ToRGBA64PM_neon(QRgba64 *buffer, const uint *src, int count,
+                                                           const QVector<QRgb> *, QDitherInfo *)
+{
+    convertARGB32ToRGBA64PM_neon<true>(buffer, src, count);
+    return buffer;
+}
+
+const QRgba64 *QT_FASTCALL fetchARGB32ToRGBA64PM_neon(QRgba64 *buffer, const uchar *src, int index, int count,
+                                                      const QVector<QRgb> *, QDitherInfo *)
+{
+    convertARGB32ToRGBA64PM_neon<false>(buffer, reinterpret_cast<const uint *>(src) + index, count);
+    return buffer;
+}
+
+const QRgba64 *QT_FASTCALL fetchRGBA8888ToRGBA64PM_neon(QRgba64 *buffer, const uchar *src, int index, int count,
+                                                        const QVector<QRgb> *, QDitherInfo *)
+{
+    convertARGB32ToRGBA64PM_neon<true>(buffer, reinterpret_cast<const uint *>(src) + index, count);
+    return buffer;
+}
+
 void QT_FASTCALL storeRGB32FromARGB32PM_neon(uchar *dest, const uint *src, int index, int count,
                                              const QVector<QRgb> *, QDitherInfo *)
 {
diff --git a/src/gui/painting/qdrawhelper_sse4.cpp b/src/gui/painting/qdrawhelper_sse4.cpp
index 06bfd3465e..9e3e6682bb 100644
--- a/src/gui/painting/qdrawhelper_sse4.cpp
+++ b/src/gui/painting/qdrawhelper_sse4.cpp
@@ -94,6 +94,55 @@ static void convertARGBToARGB32PM_sse4(uint *buffer, const uint *src, int count)
     }
 }
 
+template<bool RGBA>
+static void convertARGBToRGBA64PM_sse4(QRgba64 *buffer, const uint *src, int count)
+{
+    int i = 0;
+    const __m128i alphaMask = _mm_set1_epi32(0xff000000);
+    const __m128i rgbaMask = _mm_setr_epi8(2, 1, 0, 3, 6, 5, 4, 7, 10, 9, 8, 11, 14, 13, 12, 15);
+    const __m128i shuffleMask = _mm_setr_epi8(6, 7, 6, 7, 6, 7, 6, 7, 14, 15, 14, 15, 14, 15, 14, 15);
+    const __m128i zero = _mm_setzero_si128();
+
+    for (; i < count - 3; i += 4) {
+        __m128i srcVector = _mm_loadu_si128((const __m128i *)&src[i]);
+        if (!_mm_testz_si128(srcVector, alphaMask)) {
+            if (!_mm_testc_si128(srcVector, alphaMask)) {
+                if (!RGBA)
+                    srcVector = _mm_shuffle_epi8(srcVector, rgbaMask);
+                __m128i src1 = _mm_unpacklo_epi8(srcVector, zero);
+                __m128i src2 = _mm_unpackhi_epi8(srcVector, zero);
+                __m128i alpha1 = _mm_shuffle_epi8(src1, shuffleMask);
+                __m128i alpha2 = _mm_shuffle_epi8(src2, shuffleMask);
+                src1 = _mm_mullo_epi16(src1, alpha1);
+                src2 = _mm_mullo_epi16(src2, alpha2);
+                alpha1 = _mm_unpacklo_epi8(srcVector, srcVector);
+                alpha2 = _mm_unpackhi_epi8(srcVector, srcVector);
+                src1 = _mm_add_epi16(src1, _mm_srli_epi16(src1, 7));
+                src2 = _mm_add_epi16(src2, _mm_srli_epi16(src2, 7));
+                src1 = _mm_blend_epi16(src1, alpha1, 0x88);
+                src2 = _mm_blend_epi16(src2, alpha2, 0x88);
+                _mm_storeu_si128((__m128i *)&buffer[i], src1);
+                _mm_storeu_si128((__m128i *)&buffer[i + 2], src2);
+            } else {
+                if (!RGBA)
+                    srcVector = _mm_shuffle_epi8(srcVector, rgbaMask);
+                const __m128i src1 = _mm_unpacklo_epi8(srcVector, srcVector);
+                const __m128i src2 = _mm_unpackhi_epi8(srcVector, srcVector);
+                _mm_storeu_si128((__m128i *)&buffer[i], src1);
+                _mm_storeu_si128((__m128i *)&buffer[i + 2], src2);
+            }
+        } else {
+            _mm_storeu_si128((__m128i *)&buffer[i], zero);
+            _mm_storeu_si128((__m128i *)&buffer[i + 2], zero);
+        }
+    }
+
+    SIMD_EPILOGUE(i, count, 3) {
+        const uint s = RGBA ? RGBA2ARGB(src[i]) : src[i];
+        buffer[i] = QRgba64::fromArgb32(s).premultiplied();
+    }
+}
+
 static inline __m128 Q_DECL_VECTORCALL reciprocal_mul_ps(__m128 a, float mul)
 {
     __m128 ia = _mm_rcp_ps(a); // Approximate 1/a
@@ -269,6 +318,20 @@ void QT_FASTCALL convertRGBA8888ToARGB32PM_sse4(uint *buffer, int count, const Q
     convertARGBToARGB32PM_sse4<true>(buffer, buffer, count);
 }
 
+const QRgba64 * QT_FASTCALL convertARGB32ToRGBA64PM_sse4(QRgba64 *buffer, const uint *src, int count,
+                                                         const QVector<QRgb> *, QDitherInfo *)
+{
+    convertARGBToRGBA64PM_sse4<false>(buffer, src, count);
+    return buffer;
+}
+
+const QRgba64 * QT_FASTCALL convertRGBA8888ToRGBA64PM_sse4(QRgba64 *buffer, const uint *src, int count,
+                                                           const QVector<QRgb> *, QDitherInfo *)
+{
+    convertARGBToRGBA64PM_sse4<true>(buffer, src, count);
+    return buffer;
+}
+
 const uint *QT_FASTCALL fetchARGB32ToARGB32PM_sse4(uint *buffer, const uchar *src, int index, int count,
                                                   const QVector<QRgb> *, QDitherInfo *)
 {
@@ -283,6 +346,20 @@ const uint *QT_FASTCALL fetchRGBA8888ToARGB32PM_sse4(uint *buffer, const uchar *
     return buffer;
 }
 
+const QRgba64 *QT_FASTCALL fetchARGB32ToRGBA64PM_sse4(QRgba64 *buffer, const uchar *src, int index, int count,
+                                                      const QVector<QRgb> *, QDitherInfo *)
+{
+    convertARGBToRGBA64PM_sse4<false>(buffer, reinterpret_cast<const uint *>(src) + index, count);
+    return buffer;
+}
+
+const QRgba64 *QT_FASTCALL fetchRGBA8888ToRGBA64PM_sse4(QRgba64 *buffer, const uchar *src, int index, int count,
+                                                        const QVector<QRgb> *, QDitherInfo *)
+{
+    convertARGBToRGBA64PM_sse4<true>(buffer, reinterpret_cast<const uint *>(src) + index, count);
+    return buffer;
+}
+
 void QT_FASTCALL storeRGB32FromARGB32PM_sse4(uchar *dest, const uint *src, int index, int count,
                                              const QVector<QRgb> *, QDitherInfo *)
 {