(svn r26255) -Codechange: improve performance of brightness adjustment (MJP)

4 files changed, 21 insertions, 22 deletions

diff --git a/src/blitter/32bpp_sse2.cpp b/src/blitter/32bpp_sse2.cpp
index fa43eb4c5..0b3eb1899 100644
--- a/src/blitter/32bpp_sse2.cpp
+++ b/src/blitter/32bpp_sse2.cpp
@@ -294,7 +294,7 @@ inline Colour Blitter_32bppSSE2::AdjustBrightness(Colour colour, uint8 brightnes
 }
 
 IGNORE_UNINITIALIZED_WARNING_START
-/* static */ Colour Blitter_32bppSSE2::ReallyAdjustBrightness(Colour colour, uint8 brightness)
+Colour Blitter_32bppSSE2::ReallyAdjustBrightness(Colour colour, uint8 brightness)
 {
 	uint64 c16 = colour.b | (uint64) colour.g << 16 | (uint64) colour.r << 32;
 	c16 *= brightness;
@@ -304,24 +304,14 @@ IGNORE_UNINITIALIZED_WARNING_START
 
 	/* Sum overbright (maximum for each rgb is 508, 9 bits, -255 is changed in -256 so we just have to take the 8 lower bits into account). */
 	c16_ob = (((c16_ob >> (8 + 7)) & 0x0100010001) * 0xFF) & c16;
-	uint64 ob = (uint16) c16_ob + (uint16) (c16_ob >> 16) + (uint16) (c16_ob >> 32);
+	const uint ob = ((uint16) c16_ob + (uint16) (c16_ob >> 16) + (uint16) (c16_ob >> 32)) / 2;
 
 	const uint32 alpha32 = colour.data & 0xFF000000;
 	__m128i ret;
-#ifdef _SQ64
-	ret = _mm_cvtsi64_si128(c16);
-#else
-	INSR64(c16, ret, 0);
-#endif
+	LOAD64(c16, ret);
 	if (ob != 0) {
-		/* Reduce overbright strength. */
-		ob /= 2;
-		__m128i ob128;
-#ifdef _SQ64
-		ob128 = _mm_cvtsi64_si128(ob | ob << 16 | ob << 32);
-#else
-		INSR64(ob | ob << 16 | ob << 32, ob128, 0);
-#endif
+		__m128i ob128 = _mm_cvtsi32_si128(ob);
+		ob128 = _mm_shufflelo_epi16(ob128, 0xC0);
 		__m128i white = OVERBRIGHT_VALUE_MASK;
 		__m128i c128 = ret;
 		ret = _mm_subs_epu16(white, c128); /* PSUBUSW,   (255 - rgb) */
diff --git a/src/blitter/32bpp_sse2.hpp b/src/blitter/32bpp_sse2.hpp
index 386b5d3e3..1c3307c70 100644
--- a/src/blitter/32bpp_sse2.hpp
+++ b/src/blitter/32bpp_sse2.hpp
@@ -54,6 +54,12 @@ typedef union ALIGN(16) um128i {
 }
 #define INSR64(m_val, m_into, m_rank) (*(um128i*) &m_into).m128i_u64[m_rank] = (m_val)
 
+#ifdef _SQ64
+	#define LOAD64(m_val, m_into) m_into = _mm_cvtsi64_si128(m_val);
+#else
+	#define LOAD64(m_val, m_into) INSR64(m_val, m_into, 0)
+#endif
+
 /* PUT_ALPHA_IN_FRONT_OF_RGB is redefined in 32bpp_ssse3.hpp. */
 #define PUT_ALPHA_IN_FRONT_OF_RGB(m_from, m_into) \
 	m_into = _mm_shufflelo_epi16(m_from, 0x3F); /* PSHUFLW, put alpha1 in front of each rgb1 */ \
diff --git a/src/blitter/32bpp_sse4.cpp b/src/blitter/32bpp_sse4.cpp
index a0ddf2b63..4feebc221 100644
--- a/src/blitter/32bpp_sse4.cpp
+++ b/src/blitter/32bpp_sse4.cpp
@@ -232,7 +232,7 @@ inline Colour Blitter_32bppSSE4::AdjustBrightness(Colour colour, uint8 brightnes
 }
 
 IGNORE_UNINITIALIZED_WARNING_START
-/* static */ Colour Blitter_32bppSSE4::ReallyAdjustBrightness(Colour colour, uint8 brightness)
+Colour Blitter_32bppSSE4::ReallyAdjustBrightness(Colour colour, uint8 brightness)
 {
 	uint64 c16 = colour.b | (uint64) colour.g << 16 | (uint64) colour.r << 32;
 	c16 *= brightness;
@@ -242,16 +242,14 @@ IGNORE_UNINITIALIZED_WARNING_START
 
 	/* Sum overbright (maximum for each rgb is 508, 9 bits, -255 is changed in -256 so we just have to take the 8 lower bits into account). */
 	c16_ob = (((c16_ob >> (8 + 7)) & 0x0100010001) * 0xFF) & c16;
-	uint64 ob = (uint16) c16_ob + (uint16) (c16_ob >> 16) + (uint16) (c16_ob >> 32);
+	const uint ob = ((uint16) c16_ob + (uint16) (c16_ob >> 16) + (uint16) (c16_ob >> 32)) / 2;
 
 	const uint32 alpha32 = colour.data & 0xFF000000;
 	__m128i ret;
-	INSR64(c16, ret, 0);
+	LOAD64(c16, ret);
 	if (ob != 0) {
-		/* Reduce overbright strength. */
-		ob /= 2;
-		__m128i ob128;
-		INSR64(ob | ob << 16 | ob << 32, ob128, 0);
+		__m128i ob128 = _mm_cvtsi32_si128(ob);
+		ob128 = _mm_shufflelo_epi16(ob128, 0xC0);
 		__m128i white = OVERBRIGHT_VALUE_MASK;
 		__m128i c128 = ret;
 		ret = _mm_subs_epu16(white, c128); /* PSUBUSW,   (255 - rgb) */
diff --git a/src/blitter/32bpp_sse4.hpp b/src/blitter/32bpp_sse4.hpp
index d21b5ff33..f8a563b85 100644
--- a/src/blitter/32bpp_sse4.hpp
+++ b/src/blitter/32bpp_sse4.hpp
@@ -35,6 +35,11 @@ IGNORE_UNINITIALIZED_WARNING_START
 		(*(um128i*) &m_into).m128i = _mm_insert_epi32((*(um128i*) &m_into).m128i, v.u32.low, (m_rank)*2); \
 		(*(um128i*) &m_into).m128i = _mm_insert_epi32((*(um128i*) &m_into).m128i, v.u32.high, (m_rank)*2 + 1); \
 	}
+
+	#undef LOAD64
+	#define LOAD64(m_val, m_into) \
+		m_into = _mm_cvtsi32_si128(m_val); \
+		INSR32((m_val) >> 32, m_into, 1);
 #endif
 IGNORE_UNINITIALIZED_WARNING_STOP