aboutsummaryrefslogtreecommitdiffhomepage
path: root/libvo/aclib_template.c
blob: 9ee3c6dfb634f9860fb2f9250fe44a3f50b90138 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
/*
  aclib - advanced C library ;)
  This file contains functions which improve and expand standard C-library
*/

#ifndef HAVE_SSE2
/*
   P3 processor has only one SSE decoder so can execute only 1 sse insn per
   cpu clock, but it has 3 mmx decoders (include load/store unit)
   and executes 3 mmx insns per cpu clock.
   P4 processor has some chances, but after reading:
   http://www.emulators.com/pentium4.htm
   I have doubts. Anyway SSE2 version of this code can be written better.
*/
#undef HAVE_SSE
#endif


/*
 This part of code was taken by me from Linux-2.4.3 and slightly modified
for MMX, MMX2, SSE instruction set. I have done it since linux uses page aligned
blocks but mplayer uses weakly ordered data and original sources can not
speedup them. Only using PREFETCHNTA and MOVNTQ together have effect!

>From IA-32 Intel Architecture Software Developer's Manual Volume 1,

Order Number 245470:
"10.4.6. Cacheability Control, Prefetch, and Memory Ordering Instructions"

Data referenced by a program can be temporal (data will be used again) or
non-temporal (data will be referenced once and not reused in the immediate
future). To make efficient use of the processor's caches, it is generally
desirable to cache temporal data and not cache non-temporal data. Overloading
the processor's caches with non-temporal data is sometimes referred to as
"polluting the caches".
The non-temporal data is written to memory with Write-Combining semantics.

The PREFETCHh instructions permits a program to load data into the processor
at a suggested cache level, so that it is closer to the processors load and
store unit when it is needed. If the data is already present in a level of
the cache hierarchy that is closer to the processor, the PREFETCHh instruction
will not result in any data movement.
But we should you PREFETCHNTA: Non-temporal data fetch data into location
close to the processor, minimizing cache pollution.

The MOVNTQ (store quadword using non-temporal hint) instruction stores
packed integer data from an MMX register to memory, using a non-temporal hint.
The MOVNTPS (store packed single-precision floating-point values using
non-temporal hint) instruction stores packed floating-point data from an
XMM register to memory, using a non-temporal hint.

The SFENCE (Store Fence) instruction controls write ordering by creating a
fence for memory store operations. This instruction guarantees that the results
of every store instruction that precedes the store fence in program order is
globally visible before any store instruction that follows the fence. The
SFENCE instruction provides an efficient way of ensuring ordering between
procedures that produce weakly-ordered data and procedures that consume that
data.

If you have questions please contact with me: Nick Kurshev: nickols_k@mail.ru.
*/

// 3dnow memcpy support from kernel 2.4.2
//  by Pontscho/fresh!mindworkz


#undef HAVE_ONLY_MMX1
#if defined(HAVE_MMX) && !defined(HAVE_MMX2) && !defined(HAVE_3DNOW) && !defined(HAVE_SSE)
/*  means: mmx v.1. Note: Since we added alignment of destinition it speedups
    of memory copying on PentMMX, Celeron-1 and P2 upto 12% versus
    standard (non MMX-optimized) version.
    Note: on K6-2+ it speedups memory copying upto 25% and
          on K7 and P3 about 500% (5 times). */
#define HAVE_ONLY_MMX1
#endif


#undef HAVE_K6_2PLUS
#if !defined( HAVE_MMX2) && defined( HAVE_3DNOW)
#define HAVE_K6_2PLUS
#endif

/* for small memory blocks (<256 bytes) this version is faster */
#define small_memcpy(to,from,n)\
{\
register unsigned long int dummy;\
__asm__ __volatile__(\
	"rep; movsb"\
	:"=&D"(to), "=&S"(from), "=&c"(dummy)\
/* It's most portable way to notify compiler */\
/* that edi, esi and ecx are clobbered in asm block. */\
/* Thanks to A'rpi for hint!!! */\
        :"0" (to), "1" (from),"2" (n)\
	: "memory");\
}

#undef MMREG_SIZE
#ifdef HAVE_SSE
#define MMREG_SIZE 16
#else
#define MMREG_SIZE 64 //8
#endif

#undef PREFETCH
#undef EMMS

#ifdef HAVE_MMX2
#define PREFETCH "prefetchnta"
#elif defined ( HAVE_3DNOW )
#define PREFETCH  "prefetch"
#else
#define PREFETCH "/nop"
#endif

/* On K6 femms is faster of emms. On K7 femms is directly mapped on emms. */
#ifdef HAVE_3DNOW
#define EMMS     "femms"
#else
#define EMMS     "emms"
#endif

#undef MOVNTQ
#ifdef HAVE_MMX2
#define MOVNTQ "movntq"
#else
#define MOVNTQ "movq"
#endif

#undef MIN_LEN
#ifdef HAVE_ONLY_MMX1
#define MIN_LEN 0x800  /* 2K blocks */
#else
#define MIN_LEN 0x40  /* 64-byte blocks */
#endif

static void * RENAME(fast_memcpy)(void * to, const void * from, size_t len)
{
	void *retval;
	size_t i;
	retval = to;
#ifdef STATISTICS
	{
		static int freq[33];
		static int t=0;
		int i;
		for(i=0; len>(1<<i); i++);
		freq[i]++;
		t++;
		if(1024*1024*1024 % t == 0)
			for(i=0; i<32; i++)
				printf("freq < %8d %4d\n", 1<<i, freq[i]);
	}
#endif
#ifndef HAVE_ONLY_MMX1
        /* PREFETCH has effect even for MOVSB instruction ;) */
	__asm__ __volatile__ (
	        PREFETCH" (%0)\n"
	        PREFETCH" 64(%0)\n"
	        PREFETCH" 128(%0)\n"
        	PREFETCH" 192(%0)\n"
        	PREFETCH" 256(%0)\n"
		: : "r" (from) );
#endif
        if(len >= MIN_LEN)
	{
	  register unsigned long int delta;
          /* Align destinition to MMREG_SIZE -boundary */
          delta = ((unsigned long int)to)&(MMREG_SIZE-1);
          if(delta)
	  {
	    delta=MMREG_SIZE-delta;
	    len -= delta;
	    small_memcpy(to, from, delta);
	  }
	  i = len >> 6; /* len/64 */
	  len&=63;
        /*
           This algorithm is top effective when the code consequently
           reads and writes blocks which have size of cache line.
           Size of cache line is processor-dependent.
           It will, however, be a minimum of 32 bytes on any processors.
           It would be better to have a number of instructions which
           perform reading and writing to be multiple to a number of
           processor's decoders, but it's not always possible.
        */
#ifdef HAVE_SSE /* Only P3 (may be Cyrix3) */
	if(((unsigned long)from) & 15)
	/* if SRC is misaligned */
	for(; i>0; i--)
	{
		__asm__ __volatile__ (
		PREFETCH" 320(%0)\n"
		"movups (%0), %%xmm0\n"
		"movups 16(%0), %%xmm1\n"
		"movups 32(%0), %%xmm2\n"
		"movups 48(%0), %%xmm3\n"
		"movntps %%xmm0, (%1)\n"
		"movntps %%xmm1, 16(%1)\n"
		"movntps %%xmm2, 32(%1)\n"
		"movntps %%xmm3, 48(%1)\n"
		:: "r" (from), "r" (to) : "memory");
		from=((const unsigned char *) from)+64;
		to=((unsigned char *)to)+64;
	}
	else
	/*
	   Only if SRC is aligned on 16-byte boundary.
	   It allows to use movaps instead of movups, which required data
	   to be aligned or a general-protection exception (#GP) is generated.
	*/
	for(; i>0; i--)
	{
		__asm__ __volatile__ (
		PREFETCH" 320(%0)\n"
		"movaps (%0), %%xmm0\n"
		"movaps 16(%0), %%xmm1\n"
		"movaps 32(%0), %%xmm2\n"
		"movaps 48(%0), %%xmm3\n"
		"movntps %%xmm0, (%1)\n"
		"movntps %%xmm1, 16(%1)\n"
		"movntps %%xmm2, 32(%1)\n"
		"movntps %%xmm3, 48(%1)\n"
		:: "r" (from), "r" (to) : "memory");
		from=((const unsigned char *)from)+64;
		to=((unsigned char *)to)+64;
	}
#else
	// Align destination at BLOCK_SIZE boundary
	for(; ((int)to & (BLOCK_SIZE-1)) && i>0; i--)
	{
		__asm__ __volatile__ (
#ifndef HAVE_ONLY_MMX1
        	PREFETCH" 320(%0)\n"
#endif
		"movq (%0), %%mm0\n"
		"movq 8(%0), %%mm1\n"
		"movq 16(%0), %%mm2\n"
		"movq 24(%0), %%mm3\n"
		"movq 32(%0), %%mm4\n"
		"movq 40(%0), %%mm5\n"
		"movq 48(%0), %%mm6\n"
		"movq 56(%0), %%mm7\n"
		MOVNTQ" %%mm0, (%1)\n"
		MOVNTQ" %%mm1, 8(%1)\n"
		MOVNTQ" %%mm2, 16(%1)\n"
		MOVNTQ" %%mm3, 24(%1)\n"
		MOVNTQ" %%mm4, 32(%1)\n"
		MOVNTQ" %%mm5, 40(%1)\n"
		MOVNTQ" %%mm6, 48(%1)\n"
		MOVNTQ" %%mm7, 56(%1)\n"
		:: "r" (from), "r" (to) : "memory");
		from=((const unsigned char *)from)+64;
		to=((unsigned char *)to)+64;
	}

//	printf(" %d %d\n", (int)from&1023, (int)to&1023);
	// Pure Assembly cuz gcc is a bit unpredictable ;)
	if(i>=BLOCK_SIZE/64)
		asm volatile(
			"xor %%"REG_a", %%"REG_a"	\n\t"
			".balign 16		\n\t"
			"1:			\n\t"
				"movl (%0, %%"REG_a"), %%ebx 	\n\t"
				"movl 32(%0, %%"REG_a"), %%ebx 	\n\t"
				"movl 64(%0, %%"REG_a"), %%ebx 	\n\t"
				"movl 96(%0, %%"REG_a"), %%ebx 	\n\t"
				"add $128, %%"REG_a"		\n\t"
				"cmp %3, %%"REG_a"		\n\t"
				" jb 1b				\n\t"

			"xor %%"REG_a", %%"REG_a"	\n\t"

				".balign 16		\n\t"
				"2:			\n\t"
				"movq (%0, %%"REG_a"), %%mm0\n"
				"movq 8(%0, %%"REG_a"), %%mm1\n"
				"movq 16(%0, %%"REG_a"), %%mm2\n"
				"movq 24(%0, %%"REG_a"), %%mm3\n"
				"movq 32(%0, %%"REG_a"), %%mm4\n"
				"movq 40(%0, %%"REG_a"), %%mm5\n"
				"movq 48(%0, %%"REG_a"), %%mm6\n"
				"movq 56(%0, %%"REG_a"), %%mm7\n"
				MOVNTQ" %%mm0, (%1, %%"REG_a")\n"
				MOVNTQ" %%mm1, 8(%1, %%"REG_a")\n"
				MOVNTQ" %%mm2, 16(%1, %%"REG_a")\n"
				MOVNTQ" %%mm3, 24(%1, %%"REG_a")\n"
				MOVNTQ" %%mm4, 32(%1, %%"REG_a")\n"
				MOVNTQ" %%mm5, 40(%1, %%"REG_a")\n"
				MOVNTQ" %%mm6, 48(%1, %%"REG_a")\n"
				MOVNTQ" %%mm7, 56(%1, %%"REG_a")\n"
				"add $64, %%"REG_a"		\n\t"
				"cmp %3, %%"REG_a"		\n\t"
				"jb 2b				\n\t"

#if CONFUSION_FACTOR > 0
	// a few percent speedup on out of order executing CPUs
			"mov %5, %%"REG_a"		\n\t"
				"2:			\n\t"
				"movl (%0), %%ebx	\n\t"
				"movl (%0), %%ebx	\n\t"
				"movl (%0), %%ebx	\n\t"
				"movl (%0), %%ebx	\n\t"
				"dec %%"REG_a"		\n\t"
				" jnz 2b		\n\t"
#endif

			"xor %%"REG_a", %%"REG_a"	\n\t"
			"add %3, %0		\n\t"
			"add %3, %1		\n\t"
			"sub %4, %2		\n\t"
			"cmp %4, %2		\n\t"
			" jae 1b		\n\t"
				: "+r" (from), "+r" (to), "+r" (i)
				: "r" ((long)BLOCK_SIZE), "i" (BLOCK_SIZE/64), "i" ((long)CONFUSION_FACTOR)
				: "%"REG_a, "%ebx"
		);

	for(; i>0; i--)
	{
		__asm__ __volatile__ (
#ifndef HAVE_ONLY_MMX1
        	PREFETCH" 320(%0)\n"
#endif
		"movq (%0), %%mm0\n"
		"movq 8(%0), %%mm1\n"
		"movq 16(%0), %%mm2\n"
		"movq 24(%0), %%mm3\n"
		"movq 32(%0), %%mm4\n"
		"movq 40(%0), %%mm5\n"
		"movq 48(%0), %%mm6\n"
		"movq 56(%0), %%mm7\n"
		MOVNTQ" %%mm0, (%1)\n"
		MOVNTQ" %%mm1, 8(%1)\n"
		MOVNTQ" %%mm2, 16(%1)\n"
		MOVNTQ" %%mm3, 24(%1)\n"
		MOVNTQ" %%mm4, 32(%1)\n"
		MOVNTQ" %%mm5, 40(%1)\n"
		MOVNTQ" %%mm6, 48(%1)\n"
		MOVNTQ" %%mm7, 56(%1)\n"
		:: "r" (from), "r" (to) : "memory");
		from=((const unsigned char *)from)+64;
		to=((unsigned char *)to)+64;
	}

#endif /* Have SSE */
#ifdef HAVE_MMX2
                /* since movntq is weakly-ordered, a "sfence"
		 * is needed to become ordered again. */
		__asm__ __volatile__ ("sfence":::"memory");
#endif
#ifndef HAVE_SSE
		/* enables to use FPU */
		__asm__ __volatile__ (EMMS:::"memory");
#endif
	}
	/*
	 *	Now do the tail of the block
	 */
	if(len) small_memcpy(to, from, len);
	return retval;
}

/**
 * special copy routine for mem -> agp/pci copy (based upon fast_memcpy)
 */
static void * RENAME(mem2agpcpy)(void * to, const void * from, size_t len)
{
	void *retval;
	size_t i;
	retval = to;
#ifdef STATISTICS
	{
		static int freq[33];
		static int t=0;
		int i;
		for(i=0; len>(1<<i); i++);
		freq[i]++;
		t++;
		if(1024*1024*1024 % t == 0)
			for(i=0; i<32; i++)
				printf("mem2agp freq < %8d %4d\n", 1<<i, freq[i]);
	}
#endif
        if(len >= MIN_LEN)
	{
	  register unsigned long int delta;
          /* Align destinition to MMREG_SIZE -boundary */
          delta = ((unsigned long int)to)&7;
          if(delta)
	  {
	    delta=8-delta;
	    len -= delta;
	    small_memcpy(to, from, delta);
	  }
	  i = len >> 6; /* len/64 */
	  len &= 63;
        /*
           This algorithm is top effective when the code consequently
           reads and writes blocks which have size of cache line.
           Size of cache line is processor-dependent.
           It will, however, be a minimum of 32 bytes on any processors.
           It would be better to have a number of instructions which
           perform reading and writing to be multiple to a number of
           processor's decoders, but it's not always possible.
        */
	for(; i>0; i--)
	{
		__asm__ __volatile__ (
        	PREFETCH" 320(%0)\n"
		"movq (%0), %%mm0\n"
		"movq 8(%0), %%mm1\n"
		"movq 16(%0), %%mm2\n"
		"movq 24(%0), %%mm3\n"
		"movq 32(%0), %%mm4\n"
		"movq 40(%0), %%mm5\n"
		"movq 48(%0), %%mm6\n"
		"movq 56(%0), %%mm7\n"
		MOVNTQ" %%mm0, (%1)\n"
		MOVNTQ" %%mm1, 8(%1)\n"
		MOVNTQ" %%mm2, 16(%1)\n"
		MOVNTQ" %%mm3, 24(%1)\n"
		MOVNTQ" %%mm4, 32(%1)\n"
		MOVNTQ" %%mm5, 40(%1)\n"
		MOVNTQ" %%mm6, 48(%1)\n"
		MOVNTQ" %%mm7, 56(%1)\n"
		:: "r" (from), "r" (to) : "memory");
		from=((const unsigned char *)from)+64;
		to=((unsigned char *)to)+64;
	}
#ifdef HAVE_MMX2
                /* since movntq is weakly-ordered, a "sfence"
		 * is needed to become ordered again. */
		__asm__ __volatile__ ("sfence":::"memory");
#endif
		/* enables to use FPU */
		__asm__ __volatile__ (EMMS:::"memory");
	}
	/*
	 *	Now do the tail of the block
	 */
	if(len) small_memcpy(to, from, len);
	return retval;
}