diff options
Diffstat (limited to 'src/gui/painting/qdrawhelper_sse2.cpp')
| -rw-r--r-- | src/gui/painting/qdrawhelper_sse2.cpp | 138 |
1 files changed, 76 insertions, 62 deletions
diff --git a/src/gui/painting/qdrawhelper_sse2.cpp b/src/gui/painting/qdrawhelper_sse2.cpp index 3212ffdd2d..c82f41ec88 100644 --- a/src/gui/painting/qdrawhelper_sse2.cpp +++ b/src/gui/painting/qdrawhelper_sse2.cpp @@ -233,19 +233,71 @@ void QT_FASTCALL comp_func_Source_sse2(uint *dst, const uint *src, int length, u } } -void qt_memfill32(quint32 *dest, quint32 value, int count) +#ifndef __AVX2__ +static Q_NEVER_INLINE +void Q_DECL_VECTORCALL qt_memfillXX_aligned(void *dest, __m128i value128, quintptr bytecount) { - if (count < 7) { + __m128i *dst128 = reinterpret_cast<__m128i *>(dest); + __m128i *end128 = reinterpret_cast<__m128i *>(static_cast<uchar *>(dest) + bytecount); + + while (dst128 + 4 <= end128) { + _mm_store_si128(dst128 + 0, value128); + _mm_store_si128(dst128 + 1, value128); + _mm_store_si128(dst128 + 2, value128); + _mm_store_si128(dst128 + 3, value128); + dst128 += 4; + } + + bytecount %= 4 * sizeof(__m128i); + switch (bytecount / sizeof(__m128i)) { + case 3: _mm_store_si128(dst128++, value128); Q_FALLTHROUGH(); + case 2: _mm_store_si128(dst128++, value128); Q_FALLTHROUGH(); + case 1: _mm_store_si128(dst128++, value128); + } +} + +void qt_memfill64_sse2(quint64 *dest, quint64 value, qsizetype count) +{ + quintptr misaligned = quintptr(dest) % sizeof(__m128i); + if (misaligned && count) { +#if defined(Q_PROCESSOR_X86_32) + // Before SSE came out, the alignment of the stack used to be only 4 + // bytes and some OS/ABIs (notably, code generated by MSVC) still only + // align to that. In any case, we cannot count on the alignment of + // quint64 to be 8 -- see QtPrivate::AlignOf_WorkaroundForI386Abi in + // qglobal.h. + // + // If the pointer is not aligned to at least 8 bytes, then we'll never + // in turn hit a multiple of 16 for the qt_memfillXX_aligned call + // below. + if (Q_UNLIKELY(misaligned % sizeof(quint64))) + return qt_memfill_template(dest, value, count); +#endif + + *dest++ = value; + --count; + } + + if (count % 2) { + dest[count - 1] = value; + --count; + } + + qt_memfillXX_aligned(dest, _mm_set1_epi64x(value), count * sizeof(quint64)); +} + +void qt_memfill32_sse2(quint32 *dest, quint32 value, qsizetype count) +{ + if (count < 4) { + // this simplifies the code below: the first switch can fall through + // without checking the value of count switch (count) { - case 6: *dest++ = value; Q_FALLTHROUGH(); - case 5: *dest++ = value; Q_FALLTHROUGH(); - case 4: *dest++ = value; Q_FALLTHROUGH(); case 3: *dest++ = value; Q_FALLTHROUGH(); case 2: *dest++ = value; Q_FALLTHROUGH(); case 1: *dest = value; } return; - }; + } const int align = (quintptr)(dest) & 0xf; switch (align) { @@ -263,25 +315,9 @@ void qt_memfill32(quint32 *dest, quint32 value, int count) } } - int count128 = count / 4; - __m128i *dst128 = reinterpret_cast<__m128i*>(dest); - __m128i *end128 = dst128 + count128; - const __m128i value128 = _mm_set_epi32(value, value, value, value); - - while (dst128 + 3 < end128) { - _mm_stream_si128(dst128 + 0, value128); - _mm_stream_si128(dst128 + 1, value128); - _mm_stream_si128(dst128 + 2, value128); - _mm_stream_si128(dst128 + 3, value128); - dst128 += 4; - } - - switch (count128 & 0x3) { - case 3: _mm_stream_si128(dst128++, value128); Q_FALLTHROUGH(); - case 2: _mm_stream_si128(dst128++, value128); Q_FALLTHROUGH(); - case 1: _mm_stream_si128(dst128++, value128); - } + qt_memfillXX_aligned(dest, _mm_set1_epi32(value), count * sizeof(quint32)); } +#endif // !__AVX2__ void QT_FASTCALL comp_func_solid_SourceOver_sse2(uint *destPixels, int length, uint color, uint const_alpha) { @@ -314,28 +350,6 @@ void QT_FASTCALL comp_func_solid_SourceOver_sse2(uint *destPixels, int length, u } } -void qt_memfill16(quint16 *dest, quint16 value, int count) -{ - if (count < 3) { - switch (count) { - case 2: *dest++ = value; Q_FALLTHROUGH(); - case 1: *dest = value; - } - return; - } - - const int align = (quintptr)(dest) & 0x3; - switch (align) { - case 2: *dest++ = value; --count; - } - - const quint32 value32 = (value << 16) | value; - qt_memfill32(reinterpret_cast<quint32*>(dest), value32, count / 2); - - if (count & 0x1) - dest[count - 1] = value; -} - void qt_bitmapblit32_sse2_base(QRasterBuffer *rasterBuffer, int x, int y, quint32 color, const uchar *src, int width, int height, int stride) @@ -440,30 +454,30 @@ public: union Vect_buffer_i { Int32x4 v; int i[4]; }; union Vect_buffer_f { Float32x4 v; float f[4]; }; - static inline Float32x4 v_dup(float x) { return _mm_set1_ps(x); } - static inline Float32x4 v_dup(double x) { return _mm_set1_ps(x); } - static inline Int32x4 v_dup(int x) { return _mm_set1_epi32(x); } - static inline Int32x4 v_dup(uint x) { return _mm_set1_epi32(x); } + static inline Float32x4 Q_DECL_VECTORCALL v_dup(float x) { return _mm_set1_ps(x); } + static inline Float32x4 Q_DECL_VECTORCALL v_dup(double x) { return _mm_set1_ps(x); } + static inline Int32x4 Q_DECL_VECTORCALL v_dup(int x) { return _mm_set1_epi32(x); } + static inline Int32x4 Q_DECL_VECTORCALL v_dup(uint x) { return _mm_set1_epi32(x); } - static inline Float32x4 v_add(Float32x4 a, Float32x4 b) { return _mm_add_ps(a, b); } - static inline Int32x4 v_add(Int32x4 a, Int32x4 b) { return _mm_add_epi32(a, b); } + static inline Float32x4 Q_DECL_VECTORCALL v_add(Float32x4 a, Float32x4 b) { return _mm_add_ps(a, b); } + static inline Int32x4 Q_DECL_VECTORCALL v_add(Int32x4 a, Int32x4 b) { return _mm_add_epi32(a, b); } - static inline Float32x4 v_max(Float32x4 a, Float32x4 b) { return _mm_max_ps(a, b); } - static inline Float32x4 v_min(Float32x4 a, Float32x4 b) { return _mm_min_ps(a, b); } - static inline Int32x4 v_min_16(Int32x4 a, Int32x4 b) { return _mm_min_epi16(a, b); } + static inline Float32x4 Q_DECL_VECTORCALL v_max(Float32x4 a, Float32x4 b) { return _mm_max_ps(a, b); } + static inline Float32x4 Q_DECL_VECTORCALL v_min(Float32x4 a, Float32x4 b) { return _mm_min_ps(a, b); } + static inline Int32x4 Q_DECL_VECTORCALL v_min_16(Int32x4 a, Int32x4 b) { return _mm_min_epi16(a, b); } - static inline Int32x4 v_and(Int32x4 a, Int32x4 b) { return _mm_and_si128(a, b); } + static inline Int32x4 Q_DECL_VECTORCALL v_and(Int32x4 a, Int32x4 b) { return _mm_and_si128(a, b); } - static inline Float32x4 v_sub(Float32x4 a, Float32x4 b) { return _mm_sub_ps(a, b); } - static inline Int32x4 v_sub(Int32x4 a, Int32x4 b) { return _mm_sub_epi32(a, b); } + static inline Float32x4 Q_DECL_VECTORCALL v_sub(Float32x4 a, Float32x4 b) { return _mm_sub_ps(a, b); } + static inline Int32x4 Q_DECL_VECTORCALL v_sub(Int32x4 a, Int32x4 b) { return _mm_sub_epi32(a, b); } - static inline Float32x4 v_mul(Float32x4 a, Float32x4 b) { return _mm_mul_ps(a, b); } + static inline Float32x4 Q_DECL_VECTORCALL v_mul(Float32x4 a, Float32x4 b) { return _mm_mul_ps(a, b); } - static inline Float32x4 v_sqrt(Float32x4 x) { return _mm_sqrt_ps(x); } + static inline Float32x4 Q_DECL_VECTORCALL v_sqrt(Float32x4 x) { return _mm_sqrt_ps(x); } - static inline Int32x4 v_toInt(Float32x4 x) { return _mm_cvttps_epi32(x); } + static inline Int32x4 Q_DECL_VECTORCALL v_toInt(Float32x4 x) { return _mm_cvttps_epi32(x); } - static inline Int32x4 v_greaterOrEqual(Float32x4 a, Float32x4 b) { return _mm_castps_si128(_mm_cmpgt_ps(a, b)); } + static inline Int32x4 Q_DECL_VECTORCALL v_greaterOrEqual(Float32x4 a, Float32x4 b) { return _mm_castps_si128(_mm_cmpgt_ps(a, b)); } }; const uint * QT_FASTCALL qt_fetch_radial_gradient_sse2(uint *buffer, const Operator *op, const QSpanData *data, |