Commit a69e152dcc704eff6b7e409ad1ed43f732e6be9c

zecke
Mon Nov 09 07:28:12 GMT 2009

Tree SHA1: c2b75fe
Parent SHA1: c51c85d (Port the qt_memfill_armv6 to GNU syntax)
raw diff | raw patch

Create a specialized SourceOver routine in plain C

The following tweaks were done to boost performance:
     - Handle 0xff and 0x00 alpha special in the const_alpha=255 case
     - Unroll the above block eight times
     - Place and tweak the pld to give optimal performance on Cortex-A8

Read paint buffer with 217 frames
/qtgraphics-cycler.trace, iterations: 3, frames: 217, min(ms): 7981, median(ms): 7997, stddev: 0.842720 %, max(fps): 27.189575

Diff rendering mode:
inline
side by side


7913	7913	#endif
7914	7914	}
7915	7915
7916		#if defined(Q_CC_RVCT) && defined(QT_HAVE_ARMV6)
	7916	#if (defined(Q_CC_RVCT) && defined(QT_HAVE_ARMV6)) \|\| defined (Q_CC_GNU)
7917	7917	// Move these to qdrawhelper_arm.c when all
7918	7918	// functions are implemented using arm assembly.
7919		static CompositionFunctionSolid qt_functionForModeSolid_ARMv6[numCompositionFunctions] = {
	7919	static CompositionFunctionSolid qt_functionForModeSolid_ARM[numCompositionFunctions] = {
7920	7920	comp_func_solid_SourceOver,
7921	7921	comp_func_solid_DestinationOver,
7922	7922	comp_func_solid_Clear,
…	…
7952	7952	rasterop_solid_SourceAndNotDestination
7953	7953	};
7954	7954
7955		static CompositionFunction qt_functionForMode_ARMv6[numCompositionFunctions] = {
	7955	static CompositionFunction qt_functionForMode_ARM[numCompositionFunctions] = {
	7956	#if QT_HAVE_ARMV6
7956	7957	comp_func_SourceOver_armv6,
	7958	#else
	7959	comp_func_SourceOver_arm,
	7960	#endif
7957	7961	comp_func_DestinationOver,
7958	7962	comp_func_Clear,
	7963	#if QT_HAVE_ARMV6
7959	7964	comp_func_Source_armv6,
	7965	#else
	7966	comp_func_Source,
	7967	#endif
7960	7968	comp_func_Destination,
7961	7969	comp_func_SourceIn,
7962	7970	comp_func_DestinationIn,
…	…
7996	7996	rasterop_SourceAndNotDestination
7997	7997	};
7998	7998
	7999	#if defined(Q_CC_RVCT)
7999	8000	static void qt_blend_color_argb_armv6(int count, const QSpan spans, void userData)
8000	8001	{
8001	8002	QSpanData data = reinterpret_cast<QSpanData >(userData);
8002	8003
8003		CompositionFunctionSolid func = qt_functionForModeSolid_ARMv6[data->rasterBuffer->compositionMode];
	8004	CompositionFunctionSolid func = qt_functionForModeSolid_ARM[data->rasterBuffer->compositionMode];
8004	8005	while (count--) {
8005	8006	uint target = ((uint )data->rasterBuffer->scanLine(spans->y)) + spans->x;
8006	8007	func(target, spans->len, data->solid.color, spans->coverage);
8007	8008	++spans;
8008	8009	}
8009	8010	}
	8011	#endif
8010	8012
8011	8013	#endif // Q_CC_RVCT && QT_HAVE_ARMV6
8012	8014
…	…
8128	8128
8129	8129	#endif // QT_NO_DEBUG
8130	8130
	8131	// ARM optimized drawhing helpers. Currently this is used for GCC and RVCT
8131	8132	#if defined(QT_ARCH_ARM) \|\| defined(QT_ARCH_ARMV6)
8132	8133
8133	8134	# if defined (Q_CC_RVCT) \|\| defined(Q_CC_GNU)
8134	8135	qt_memfill32 = qt_memfill32_arm;
	8136
	8137	functionForModeAsm = qt_functionForMode_ARM;
	8138	functionForModeSolidAsm = qt_functionForModeSolid_ARM;
8135	8139	# endif
8136	8140
8137		# if defined(Q_CC_RVCT) && defined(QT_HAVE_ARMV6)
8138		functionForModeAsm = qt_functionForMode_ARMv6;
8139		functionForModeSolidAsm = qt_functionForModeSolid_ARMv6;
8140	8141
8141
	8142	# if defined(Q_CC_RVCT) && defined(QT_HAVE_ARMV6)
8142	8143	qDrawHelper[QImage::Format_ARGB32_Premultiplied].blendColor = qt_blend_color_argb_armv6;
8143	8144
8144	8145	qBlendFunctions[QImage::Format_RGB32][QImage::Format_RGB32] = qt_blend_rgb32_on_rgb32_armv6;


87	87	);
88	88	}
89	89
	90	static __attribute__((always_inline)) void preload(const uint *start)
	91	{
	92	#if defined(QT_ARM_HAS_PLD)
	93	asm __volatile__("pld [%[addr]]\n"
	94	:
	95	: [addr] "r"(start));
	96	#endif
	97	}
	98
	99	#define UNROLL_4_TIMES(block) block block block block
	100	#define UNROLL_8_TIMES(block) UNROLL_4_TIMES(block) UNROLL_4_TIMES(block)
	101
	102	#define UNROLL_LOOP_WITH_PLD(block) \
	103	int i = 0; \
	104	int end = length & ~7; \
	105	while (i < end) { \
	106	preload(&src[i] + 32); \
	107	preload(&dest[i]+ 32); \
	108	UNROLL_8_TIMES(block) \
	109	} \
	110	preload(&src[i] + 32); \
	111	preload(&dest[i]+ 32); \
	112	while (i < length) { \
	113	block \
	114	}
	115
	116
	117	// TODO: on Cortex-A8 this is 64.
	118	static const uint L2CacheLineLength = 32;
	119	static const uint L2CacheLineLengthInInts = L2CacheLineLength/sizeof(uint);
	120	# define PRELOAD_INIT2(x,y) preload(x); preload(y);
	121	// Two consecutive preloads stall, so space them out a bit by using different modulus.
	122	// -- I could not verify this on a Cortex-A8 with the above UNROLL_LOOP_WITH_PLD
	123	# define PRELOAD_COND2(x,y) if (((uint)&x[i])%L2CacheLineLength == 0) preload(&x[i] + L2CacheLineLengthInInts); \
	124	if (((uint)&y[i])%L2CacheLineLength == 16) preload(&y[i] + L2CacheLineLengthInInts);
	125	void comp_func_SourceOver_arm(uint dest, const uint src, int length, uint const_alpha)
	126	{
	127	PRELOAD_INIT2(dest, src)
	128	if (const_alpha == 255) {
	129	UNROLL_LOOP_WITH_PLD(
	130	{
	131	if ((src[i] & 0xff000000) == 0x00000000) {
	132	; // nothing
	133	} else if ((src[i] & 0xff000000) == 0xff000000) {
	134	dest[i] = src[i];
	135	} else {
	136	uint s = src[i];
	137	dest[i] = s + BYTE_MUL(dest[i], qAlpha(~s));
	138	}
	139
	140	++i;
	141	});
	142	} else {
	143	int i = 0;
	144	while (i < length) {
	145	uint s = BYTE_MUL(src[i], const_alpha);
	146	dest[i] = s + BYTE_MUL(dest[i], qAlpha(~s));
	147	PRELOAD_COND2(dest, src)
	148	++i;
	149	}
	150	}
	151
	152	return;
	153	}
	154
90	155	#endif


99	99
100	100	#if defined(Q_CC_RVCT) \|\| defined(Q_CC_GNU)
101	101	extern "C" void qt_memfill32_arm(quint32 *dest, quint32 value, int count);
	102	extern "C" void comp_func_SourceOver_arm(uint dest, const uint src, int length, uint const_alpha);
102	103	#endif
103	104
104	105	#if defined(Q_CC_RVCT) && defined(QT_HAVE_ARMV6)

Commit a69e152dcc704eff6b7e409ad1ed43f732e6be9c

Comments

Add a new comment:

Add your comment