Commit b2d17d7c4ecd31cba36ef79909e2e819217a0971

zecke
Mon Nov 09 07:28:09 GMT 2009

Tree SHA1: 0136660
Parent SHA1: 50b67cf (Detect architecture and features automatically on GNU)
raw diff | raw patch

Use the PLD instruction for CC_RCVT and GNU when it is available

Separate the CC_RCVT pragma to generate arm assembly from the preload
code. Use the preload code when PLD is available, add GNU inline assembly.

before:
Read paint buffer with 217 frames
/qtgraphics-cycler.trace, iterations: 3, frames: 217, min(ms): 12417, median(ms): 12418, stddev: 0.021134 %, max(fps): 17.476041

after:
Read paint buffer with 217 frames
/qtgraphics-cycler.trace, iterations: 3, frames: 217, min(ms): 11963, median(ms): 11963, stddev: 0.003940 %, max(fps): 18.139263

Diff rendering mode:
inline
side by side


1214	1214	return b;
1215	1215	}
1216	1216
1217		#if defined(Q_CC_RVCT)
1218	1217	// Force ARM code generation for comp_func_* -methods
	1218	#if defined(Q_CC_RVCT)
1219	1219	# pragma push
1220	1220	# pragma arm
1221		# if defined(QT_HAVE_ARMV6)
	1221	#endif
	1222
	1223
	1224	// Use PLD on ARM core that have them in two assembly flavors
	1225	#ifdef QT_ARM_HAS_PLD
	1226
	1227	// assembly flavors
	1228	# if defined (Q_CC_RVCT)
1222	1229	static __forceinline void preload(const uint *start)
1223	1230	{
1224	1231	asm( "pld [start]" );
1225	1232	}
	1233	# elif defined (Q_CC_GNU)
	1234	static __attribute__((always_inline)) void preload(const uint *start)
	1235	{
	1236	asm __volatile__("pld [%[addr]]\n"
	1237	:
	1238	: [addr] "r"(start));
	1239	}
	1240	# endif
	1241
1226	1242	static const uint L2CacheLineLength = 32;
1227	1243	static const uint L2CacheLineLengthInInts = L2CacheLineLength/sizeof(uint);
1228	1244	# define PRELOAD_INIT(x) preload(x);
…	…
1247	1247	// Two consecutive preloads stall, so space them out a bit by using different modulus.
1248	1248	# define PRELOAD_COND2(x,y) if (((uint)&x[i])%L2CacheLineLength == 0) preload(&x[i] + L2CacheLineLengthInInts); \
1249	1249	if (((uint)&y[i])%L2CacheLineLength == 16) preload(&y[i] + L2CacheLineLengthInInts);
1250		# endif // QT_HAVE_ARMV6
1251		#endif // Q_CC_RVCT
1252	1250
1253		#if !defined(Q_CC_RVCT) \|\| !defined(QT_HAVE_ARMV6)
	1251	#else
1254	1252	# define PRELOAD_INIT(x)
1255	1253	# define PRELOAD_INIT2(x,y)
1256	1254	# define PRELOAD_COND(x)
1257	1255	# define PRELOAD_COND2(x,y)
1258		#endif
	1256	#endif // QT_ARM_HAS_PLD
1259	1257
1260	1258	/* The constant alpha factor describes an alpha factor that gets applied
1261	1259	to the result of the composition operation combining it with the destination.

Commit b2d17d7c4ecd31cba36ef79909e2e819217a0971

Comments

Add a new comment:

Add your comment