diff options
| author | Allan Sandfeld Jensen <allan.jensen@theqtcompany.com> | 2016-08-10 11:58:24 +0200 |
|---|---|---|
| committer | Allan Sandfeld Jensen <allan.jensen@qt.io> | 2016-08-12 22:27:17 +0000 |
| commit | fcd3c11945c20c1b314fee57f7bbe457412844bc (patch) | |
| tree | 1d71ae3e0d75b203a77f70623abf4d7233063ccb /src | |
| parent | 4a1bafcc4ee5b7d968620808e155c1617aa6f273 (diff) | |
Improve vectorized bound calculations
Adds a proper bound calculation to vectorized rotated sampling to avoid
checking inside the loop, and fixes the existing bounds check to use
the incrementor fdx instead the float value it was calculated from
which may cause different rounding.
Change-Id: I5226926a142573c026db5504414204b6ee8dd8a7
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
Diffstat (limited to 'src')
| -rw-r--r-- | src/gui/painting/qdrawhelper.cpp | 33 |
1 files changed, 19 insertions, 14 deletions
diff --git a/src/gui/painting/qdrawhelper.cpp b/src/gui/painting/qdrawhelper.cpp index 4812cee9bb..affbc43afe 100644 --- a/src/gui/painting/qdrawhelper.cpp +++ b/src/gui/painting/qdrawhelper.cpp @@ -2180,6 +2180,8 @@ static const uint * QT_FASTCALL fetchTransformedBilinearARGB32PM(uint *buffer, c if (blendType != BlendTransformedBilinearTiled) { #define BILINEAR_DOWNSCALE_BOUNDS_PROLOG \ + const qint64 min_fx = qint64(image_x1) * fixed_scale; \ + const qint64 max_fx = qint64(image_x2) * fixed_scale; \ while (b < end) { \ int x1 = (fx >> 16); \ int x2; \ @@ -2195,11 +2197,11 @@ static const uint * QT_FASTCALL fetchTransformedBilinearARGB32PM(uint *buffer, c fx += fdx; \ ++b; \ } \ - uint *boundedEnd; \ + uint *boundedEnd = end; \ if (fdx > 0) \ - boundedEnd = qMin(end, buffer + uint((image_x2 - (fx >> 16)) / data->m11)); \ - else \ - boundedEnd = qMin(end, buffer + uint((image_x1 - (fx >> 16)) / data->m11)); \ + boundedEnd = qMin(boundedEnd, b + (max_fx - fx) / fdx); \ + else if (fdx < 0) \ + boundedEnd = qMin(boundedEnd, b + (min_fx - fx) / fdx); \ boundedEnd -= 3; #if defined(__SSE2__) @@ -2333,6 +2335,10 @@ static const uint * QT_FASTCALL fetchTransformedBilinearARGB32PM(uint *buffer, c if (blendType != BlendTransformedBilinearTiled) { #define BILINEAR_ROTATE_BOUNDS_PROLOG \ + const qint64 min_fx = qint64(image_x1) * fixed_scale; \ + const qint64 max_fx = qint64(image_x2) * fixed_scale; \ + const qint64 min_fy = qint64(image_y1) * fixed_scale; \ + const qint64 max_fy = qint64(image_y2) * fixed_scale; \ while (b < end) { \ int x1 = (fx >> 16); \ int x2; \ @@ -2355,7 +2361,15 @@ static const uint * QT_FASTCALL fetchTransformedBilinearARGB32PM(uint *buffer, c fy += fdy; \ ++b; \ } \ - uint *boundedEnd = end - 3; \ + uint *boundedEnd = end; \ + if (fdx > 0) \ + boundedEnd = qMin(boundedEnd, b + (max_fx - fx) / fdx); \ + else if (fdx < 0) \ + boundedEnd = qMin(boundedEnd, b + (min_fx - fx) / fdx); \ + if (fdy > 0) \ + boundedEnd = qMin(boundedEnd, b + (max_fy - fy) / fdy); \ + else if (fdy < 0) \ + boundedEnd = qMin(boundedEnd, b + (min_fy - fy) / fdy); \ boundedEnd -= 3; #if defined(__SSE2__) @@ -2374,15 +2388,6 @@ static const uint * QT_FASTCALL fetchTransformedBilinearARGB32PM(uint *buffer, c const __m128i vbpl = _mm_shufflelo_epi16(_mm_cvtsi32_si128(bytesPerLine/4), _MM_SHUFFLE(0, 0, 0, 0)); while (b < boundedEnd) { - if (fdx > 0 && (short)_mm_extract_epi16(v_fx, 7) >= image_x2) - break; - if (fdx < 0 && (short)_mm_extract_epi16(v_fx, 7) < image_x1) - break; - if (fdy > 0 && (short)_mm_extract_epi16(v_fy, 7) >= image_y2) - break; - if (fdy < 0 && (short)_mm_extract_epi16(v_fy, 7) < image_y1) - break; - const __m128i vy = _mm_packs_epi32(_mm_srli_epi32(v_fy, 16), _mm_setzero_si128()); // 4x16bit * 4x16bit -> 4x32bit __m128i offset = _mm_unpacklo_epi16(_mm_mullo_epi16(vy, vbpl), _mm_mulhi_epi16(vy, vbpl)); |