summaryrefslogtreecommitdiff
path: root/src/blitter
diff options
context:
space:
mode:
authorrubidium <rubidium@openttd.org>2014-01-03 18:03:14 +0000
committerrubidium <rubidium@openttd.org>2014-01-03 18:03:14 +0000
commit115c712f4ea2ee09c90ee3c0f04f4bfd70d00944 (patch)
tree503b63b8738173becd919c24f650f92868b0e872 /src/blitter
parent613fb9c6ef6888568a47e5e08f462839bcfe7758 (diff)
downloadopenttd-115c712f4ea2ee09c90ee3c0f04f4bfd70d00944.tar.xz
(svn r26222) -Codechange: use SSE instructions for transparent colour "remap" in the SSE2/SSSE3 blitters as well
Diffstat (limited to 'src/blitter')
-rw-r--r--src/blitter/32bpp_sse2.cpp49
-rw-r--r--src/blitter/32bpp_ssse3.cpp45
2 files changed, 73 insertions, 21 deletions
diff --git a/src/blitter/32bpp_sse2.cpp b/src/blitter/32bpp_sse2.cpp
index 52f37d690..d515d428d 100644
--- a/src/blitter/32bpp_sse2.cpp
+++ b/src/blitter/32bpp_sse2.cpp
@@ -44,7 +44,8 @@ inline void Blitter_32bppSSE2::Draw(const Blitter::BlitterParams *bp, ZoomLevel
}
/* Load these variables into register before loop. */
- const __m128i clear_hi = CLEAR_HIGH_BYTE_MASK;
+ const __m128i clear_hi = CLEAR_HIGH_BYTE_MASK;
+ const __m128i tr_nom_base = TRANSPARENT_NOM_BASE;
for (int y = bp->height; y != 0; y--) {
Colour *dst = dst_line;
@@ -143,18 +144,44 @@ inline void Blitter_32bppSSE2::Draw(const Blitter::BlitterParams *bp, ZoomLevel
src_mv_line += si->sprite_width;
break;
}
- case BM_TRANSPARENT:
- /* Make the current colour a bit more black, so it looks like this image is transparent */
- for (int x = bp->width; x != 0; x--) {
- if (src->a == 255) {
- *dst = MakeTransparent(*dst, 3, 4);
- } else {
- *dst = MakeTransparent(*dst, (256 * 4 - src->a), 256 * 4);
- }
- dst++;
- src++;
+ case BM_TRANSPARENT: {
+ /* Make the current colour a bit more black, so it looks like this image is transparent.
+ * rgb = rgb * ((256/4) * 4 - (alpha/4)) / ((256/4) * 4)
+ */
+ __m128i srcABCD = _mm_loadu_si128((const __m128i*) src);
+ __m128i dstABCD = _mm_loadu_si128((__m128i*) dst);
+ for (uint x = (uint) bp->width / 2; x > 0; x--) {
+ __m128i srcAB = _mm_unpacklo_epi8(srcABCD, _mm_setzero_si128());
+ __m128i dstAB = _mm_unpacklo_epi8(dstABCD, _mm_setzero_si128());
+ __m128i dstCD = _mm_unpackhi_epi8(dstABCD, _mm_setzero_si128());
+ __m128i alphaAB = _mm_shufflelo_epi16(srcAB, 0x3F);
+ alphaAB = _mm_shufflehi_epi16(alphaAB, 0x3F);
+ alphaAB = _mm_srli_epi16(alphaAB, 2); // Reduce to 64 levels of shades so the max value fits in 16 bits.
+ __m128i nom = _mm_sub_epi16(tr_nom_base, alphaAB);
+ dstAB = _mm_mullo_epi16(dstAB, nom);
+ dstAB = _mm_srli_epi16(dstAB, 8);
+ dstAB = _mm_packus_epi16(dstAB, dstCD);
+ Colour *old_dst = dst;
+ src += 2;
+ dst += 2;
+ dstABCD = _mm_loadu_si128((__m128i*) dst);
+ _mm_storeu_si128((__m128i *) old_dst, dstAB);
+ srcABCD = _mm_loadu_si128((const __m128i*) src);
+ }
+ if (bp->width & 1) {
+ __m128i srcAB = _mm_unpacklo_epi8(srcABCD, _mm_setzero_si128());
+ __m128i dstAB = _mm_unpacklo_epi8(dstABCD, _mm_setzero_si128());
+ __m128i alphaAB = _mm_shufflelo_epi16(srcAB, 0x3F);
+ alphaAB = _mm_shufflehi_epi16(alphaAB, 0x3F);
+ alphaAB = _mm_srli_epi16(alphaAB, 2);
+ __m128i nom = _mm_sub_epi16(tr_nom_base, alphaAB);
+ dstAB = _mm_mullo_epi16(dstAB, nom);
+ dstAB = _mm_srli_epi16(dstAB, 8);
+ dstAB = _mm_packus_epi16(dstAB, dstAB);
+ (*dst).data = EXTR32(dstAB, 0);
}
break;
+ }
}
src_rgba_line = (const Colour*) ((const byte*) src_rgba_line + si->sprite_line_size);
diff --git a/src/blitter/32bpp_ssse3.cpp b/src/blitter/32bpp_ssse3.cpp
index eac4e8c11..531262162 100644
--- a/src/blitter/32bpp_ssse3.cpp
+++ b/src/blitter/32bpp_ssse3.cpp
@@ -56,6 +56,7 @@ inline void Blitter_32bppSSSE3::Draw(const Blitter::BlitterParams *bp, ZoomLevel
const __m128i ob_check = OVERBRIGHT_PRESENCE_MASK;
const __m128i ob_mask = OVERBRIGHT_VALUE_MASK;
const __m128i ob_cm = OVERBRIGHT_CONTROL_MASK;
+ const __m128i tr_nom_base = TRANSPARENT_NOM_BASE;
for (int y = bp->height; y != 0; y--) {
Colour *dst = dst_line;
@@ -199,18 +200,42 @@ inline void Blitter_32bppSSSE3::Draw(const Blitter::BlitterParams *bp, ZoomLevel
src_mv_line += si->sprite_width;
break;
}
- case BM_TRANSPARENT:
- /* Make the current colour a bit more black, so it looks like this image is transparent */
- for (int x = bp->width; x != 0; x--) {
- if (src->a == 255) {
- *dst = MakeTransparent(*dst, 3, 4);
- } else {
- *dst = MakeTransparent(*dst, (256 * 4 - src->a), 256 * 4);
- }
- dst++;
- src++;
+ case BM_TRANSPARENT: {
+ /* Make the current colour a bit more black, so it looks like this image is transparent.
+ * rgb = rgb * ((256/4) * 4 - (alpha/4)) / ((256/4) * 4)
+ */
+ __m128i srcABCD = _mm_loadu_si128((const __m128i*) src);
+ __m128i dstABCD = _mm_loadu_si128((__m128i*) dst);
+ for (uint x = (uint) bp->width / 2; x > 0; x--) {
+ __m128i srcAB = _mm_unpacklo_epi8(srcABCD, _mm_setzero_si128());
+ __m128i dstAB = _mm_unpacklo_epi8(dstABCD, _mm_setzero_si128());
+ __m128i dstCD = _mm_unpackhi_epi8(dstABCD, _mm_setzero_si128());
+ __m128i alphaAB = _mm_shuffle_epi8(srcAB, a_cm);
+ alphaAB = _mm_srli_epi16(alphaAB, 2); // Reduce to 64 levels of shades so the max value fits in 16 bits.
+ __m128i nom = _mm_sub_epi16(tr_nom_base, alphaAB);
+ dstAB = _mm_mullo_epi16(dstAB, nom);
+ dstAB = _mm_srli_epi16(dstAB, 8);
+ dstAB = _mm_packus_epi16(dstAB, dstCD);
+ Colour *old_dst = dst;
+ src += 2;
+ dst += 2;
+ dstABCD = _mm_loadu_si128((__m128i*) dst);
+ _mm_storeu_si128((__m128i *) old_dst, dstAB);
+ srcABCD = _mm_loadu_si128((const __m128i*) src);
+ }
+ if (bp->width & 1) {
+ __m128i srcAB = _mm_unpacklo_epi8(srcABCD, _mm_setzero_si128());
+ __m128i dstAB = _mm_unpacklo_epi8(dstABCD, _mm_setzero_si128());
+ __m128i alphaAB = _mm_shuffle_epi8(srcAB, a_cm);
+ alphaAB = _mm_srli_epi16(alphaAB, 2);
+ __m128i nom = _mm_sub_epi16(tr_nom_base, alphaAB);
+ dstAB = _mm_mullo_epi16(dstAB, nom);
+ dstAB = _mm_srli_epi16(dstAB, 8);
+ dstAB = _mm_packus_epi16(dstAB, dstAB);
+ (*dst).data = EXTR32(dstAB, 0);
}
break;
+ }
}
src_rgba_line = (const Colour*) ((const byte*) src_rgba_line + si->sprite_line_size);