dave/openttd
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge branch 'master' into jgrpp

Browse Source 
# Conflicts:
#	src/airport_gui.cpp
#	src/blitter/32bpp_anim_sse4.cpp
#	src/console_cmds.cpp
#	src/linkgraph/linkgraph_gui.cpp
#	src/newgrf_object.h
#	src/road_gui.cpp
#	src/widgets/road_widget.h
#	src/window.cpp
This commit is contained in:
Jonathan G Rennison
2022-10-16 20:14:13 +01:00
parent 7f782a1bc3 3b071f47e4
commit 0a9c44d1a2
82 changed files with 778 additions and 341 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
src/blitter/32bpp_sse_func.hpp
View File

@@ -65,27 +65,31 @@ static inline __m128i DistributeAlpha(const __m128i from, const __m128i &mask)
{
#if (SSE_VERSION == 2)
__m128i alphaAB = _mm_shufflelo_epi16(from, 0x3F); // PSHUFLW, put alpha1 in front of each rgb1
return _mm_shufflehi_epi16(alphaAB, 0x3F); // PSHUFHW, put alpha2 in front of each rgb2
alphaAB = _mm_shufflehi_epi16(alphaAB, 0x3F); // PSHUFHW, put alpha2 in front of each rgb2
return _mm_andnot_si128(mask, alphaAB); // PANDN, set alpha fields to 0
#else
return _mm_shuffle_epi8(from, mask);
#endif
}
GNU_TARGET(SSE_TARGET)
static inline __m128i AlphaBlendTwoPixels(__m128i src, __m128i dst, const __m128i &distribution_mask, const __m128i &pack_mask)
static inline __m128i AlphaBlendTwoPixels(__m128i src, __m128i dst, const __m128i &distribution_mask, const __m128i &pack_mask, const __m128i &alpha_mask)
{
__m128i srcAB = _mm_unpacklo_epi8(src, _mm_setzero_si128()); // PUNPCKLBW, expand each uint8 into uint16
__m128i dstAB = _mm_unpacklo_epi8(dst, _mm_setzero_si128());
__m128i alphaAB = _mm_cmpgt_epi16(srcAB, _mm_setzero_si128()); // PCMPGTW, if (alpha > 0) a++;
alphaAB = _mm_srli_epi16(alphaAB, 15);
alphaAB = _mm_add_epi16(alphaAB, srcAB);
__m128i alphaMaskAB = _mm_cmpgt_epi16(srcAB, _mm_setzero_si128()); // PCMPGTW (alpha > 0) ? 0xFFFF : 0
__m128i alphaAB = _mm_sub_epi16(srcAB, alphaMaskAB); // if (alpha > 0) a++;
alphaAB = DistributeAlpha(alphaAB, distribution_mask);
srcAB = _mm_sub_epi16(srcAB, dstAB); // PSUBW, (r - Cr)
srcAB = _mm_mullo_epi16(srcAB, alphaAB); // PMULLW, a*(r - Cr)
srcAB = _mm_srli_epi16(srcAB, 8); // PSRLW, a*(r - Cr)/256
srcAB = _mm_add_epi16(srcAB, dstAB); // PADDW, a*(r - Cr)/256 + Cr
alphaMaskAB = _mm_and_si128(alphaMaskAB, alpha_mask); // PAND, set non alpha fields to 0
srcAB = _mm_or_si128(srcAB, alphaMaskAB); // POR, set alpha fields to 0xFFFF is src alpha was > 0
return PackUnsaturated(srcAB, pack_mask);
}
@@ -233,9 +237,11 @@ inline void Blitter_32bppSSE4::Draw(const Blitter::BlitterParams *bp, ZoomLevel
const MapValue *src_mv = src_mv_line;
/* Load these variables into register before loop. */
const __m128i alpha_and = ALPHA_AND_MASK;
#define ALPHA_BLEND_PARAM_3 alpha_and
#if (SSE_VERSION == 2)
const __m128i clear_hi = CLEAR_HIGH_BYTE_MASK;
#define ALPHA_BLEND_PARAM_1 clear_hi
#define ALPHA_BLEND_PARAM_1 alpha_and
#define ALPHA_BLEND_PARAM_2 clear_hi
#define DARKEN_PARAM_1 tr_nom_base
#define DARKEN_PARAM_2 tr_nom_base
@@ -281,7 +287,7 @@ inline void Blitter_32bppSSE4::Draw(const Blitter::BlitterParams *bp, ZoomLevel
for (uint x = (uint) effective_width / 2; x > 0; x--) {
__m128i srcABCD = _mm_loadl_epi64((const __m128i*) src);
__m128i dstABCD = _mm_loadl_epi64((__m128i*) dst);
_mm_storel_epi64((__m128i*) dst, AlphaBlendTwoPixels(srcABCD, dstABCD, ALPHA_BLEND_PARAM_1, ALPHA_BLEND_PARAM_2));
_mm_storel_epi64((__m128i*) dst, AlphaBlendTwoPixels(srcABCD, dstABCD, ALPHA_BLEND_PARAM_1, ALPHA_BLEND_PARAM_2, ALPHA_BLEND_PARAM_3));
src += 2;
dst += 2;
}
@@ -289,7 +295,7 @@ inline void Blitter_32bppSSE4::Draw(const Blitter::BlitterParams *bp, ZoomLevel
if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) {
__m128i srcABCD = _mm_cvtsi32_si128(src->data);
__m128i dstABCD = _mm_cvtsi32_si128(dst->data);
dst->data = _mm_cvtsi128_si32(AlphaBlendTwoPixels(srcABCD, dstABCD, ALPHA_BLEND_PARAM_1, ALPHA_BLEND_PARAM_2));
dst->data = _mm_cvtsi128_si32(AlphaBlendTwoPixels(srcABCD, dstABCD, ALPHA_BLEND_PARAM_1, ALPHA_BLEND_PARAM_2, ALPHA_BLEND_PARAM_3));
}
break;
@@ -334,7 +340,7 @@ inline void Blitter_32bppSSE4::Draw(const Blitter::BlitterParams *bp, ZoomLevel
}
/* Blend colours. */
_mm_storel_epi64((__m128i *) dst, AlphaBlendTwoPixels(srcABCD, dstABCD, ALPHA_BLEND_PARAM_1, ALPHA_BLEND_PARAM_2));
_mm_storel_epi64((__m128i *) dst, AlphaBlendTwoPixels(srcABCD, dstABCD, ALPHA_BLEND_PARAM_1, ALPHA_BLEND_PARAM_2, ALPHA_BLEND_PARAM_3));
dst += 2;
src += 2;
src_mv += 2;
@@ -363,7 +369,7 @@ inline void Blitter_32bppSSE4::Draw(const Blitter::BlitterParams *bp, ZoomLevel
if (src->a < 255) {
bmcr_alpha_blend_single:
__m128i dstABCD = _mm_cvtsi32_si128(dst->data);
srcABCD = AlphaBlendTwoPixels(srcABCD, dstABCD, ALPHA_BLEND_PARAM_1, ALPHA_BLEND_PARAM_2);
srcABCD = AlphaBlendTwoPixels(srcABCD, dstABCD, ALPHA_BLEND_PARAM_1, ALPHA_BLEND_PARAM_2, ALPHA_BLEND_PARAM_3);
}
dst->data = _mm_cvtsi128_si32(srcABCD);
}
@@ -429,7 +435,7 @@ bmcr_alpha_blend_single:
__m128i srcABCD = _mm_setr_epi32(AdjustBrightneSSE(src->data, DEFAULT_BRIGHTNESS + bp->brightness_adjust).data, AdjustBrightneSSE((src + 1)->data, DEFAULT_BRIGHTNESS + bp->brightness_adjust).data, 0, 0);
#endif
__m128i dstABCD = _mm_loadl_epi64((__m128i*) dst);
_mm_storel_epi64((__m128i*) dst, AlphaBlendTwoPixels(srcABCD, dstABCD, ALPHA_BLEND_PARAM_1, ALPHA_BLEND_PARAM_2));
_mm_storel_epi64((__m128i*) dst, AlphaBlendTwoPixels(srcABCD, dstABCD, ALPHA_BLEND_PARAM_1, ALPHA_BLEND_PARAM_2, ALPHA_BLEND_PARAM_3));
src += 2;
dst += 2;
}
@@ -437,7 +443,7 @@ bmcr_alpha_blend_single:
if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) {
__m128i srcABCD = _mm_cvtsi32_si128(AdjustBrightneSSE(src->data, DEFAULT_BRIGHTNESS + bp->brightness_adjust).data);
__m128i dstABCD = _mm_cvtsi32_si128(dst->data);
dst->data = _mm_cvtsi128_si32(AlphaBlendTwoPixels(srcABCD, dstABCD, ALPHA_BLEND_PARAM_1, ALPHA_BLEND_PARAM_2));
dst->data = _mm_cvtsi128_si32(AlphaBlendTwoPixels(srcABCD, dstABCD, ALPHA_BLEND_PARAM_1, ALPHA_BLEND_PARAM_2, ALPHA_BLEND_PARAM_3));
}
break;
@@ -465,7 +471,7 @@ bmcr_alpha_blend_single:
if (src->a < 255) {
bmcr_alpha_blend_single_brightness:
__m128i dstABCD = _mm_cvtsi32_si128(dst->data);
srcABCD = AlphaBlendTwoPixels(srcABCD, dstABCD, ALPHA_BLEND_PARAM_1, ALPHA_BLEND_PARAM_2);
srcABCD = AlphaBlendTwoPixels(srcABCD, dstABCD, ALPHA_BLEND_PARAM_1, ALPHA_BLEND_PARAM_2, ALPHA_BLEND_PARAM_3);
}
dst->data = _mm_cvtsi128_si32(srcABCD);
}