Merge commit '82ee14d2cedd7867920529b408ed6c7ec2f13ff1'

* commit '82ee14d2cedd7867920529b408ed6c7ec2f13ff1': ppc: dsputil: comment formatting and wording/grammar improvements Conflicts: libavcodec/ppc/gmc_altivec.c libavcodec/ppc/idct_altivec.c

Michael Niedermayer committed 12 years ago

Commit 047cf462 parents 75cc29a 82ee14d7 files · +180 −199Browse files

Showing the changes to libavcodec/ppc (7 of 7 files). Show all

Unified · Split

7 files

libavcodec/ppc/dsputil_altivec.c +114 −119
libavcodec/ppc/dsputil_ppc.c +29 −31
libavcodec/ppc/fdct_altivec.c +3 −4
libavcodec/ppc/fft_altivec.c +6 −6
libavcodec/ppc/gmc_altivec.c +17 −23
libavcodec/ppc/idct_altivec.c +4 −9
libavcodec/ppc/int_altivec.c +7 −7

libavcodec/ppc/dsputil_altivec.c+114−119

@@ -47,27 +47,27 @@ static int sad16_x2_altivec(void v, uint8_t pix1, uint8_t *pix2, int line_size · 46 unmodified lines
47	47	sad = (vector unsigned int)vec_splat_u32(0);
48	48	for (i = 0; i < h; i++) {
49	49	/* Read unaligned pixels into our vectors. The vectors are as follows:
50		− pix1v: pix1[0]-pix1[15]
51		− pix2v: pix2[0]-pix2[15] pix2iv: pix2[1]-pix2[16] */
	50	+ * pix1v: pix1[0] - pix1[15]
	51	+ * pix2v: pix2[0] - pix2[15] pix2iv: pix2[1] - pix2[16] */
52	52	pix1v = vec_ld( 0, pix1);
53	53	pix2l = vec_ld( 0, pix2);
54	54	pix2r = vec_ld(16, pix2);
55	55	pix2v = vec_perm(pix2l, pix2r, perm1);
56	56	pix2iv = vec_perm(pix2l, pix2r, perm2);
57	57
58		− /* Calculate the average vector */
	58	+ /* Calculate the average vector. */
59	59	avgv = vec_avg(pix2v, pix2iv);
60	60
61		− /* Calculate a sum of abs differences vector */
	61	+ /* Calculate a sum of abs differences vector. */
62	62	t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
63	63
64		− /* Add each 4 pixel group together and put 4 results into sad */
	64	+ /* Add each 4 pixel group together and put 4 results into sad. */
65	65	sad = vec_sum4s(t5, sad);
66	66
67	67	pix1 += line_size;
68	68	pix2 += line_size;
69	69	}
70		− /* Sum up the four partial sums, and put the result into s */
	70	+ /* Sum up the four partial sums, and put the result into s. */
71	71	sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
72	72	sumdiffs = vec_splat(sumdiffs, 3);
73	73	vec_ste(sumdiffs, 0, &s);
@@ -91,33 +91,33 @@ static int sad16_y2_altivec(void v, uint8_t pix1, uint8_t *pix2, int line_size · 17 unmodified lines
91	91	sad = (vector unsigned int)vec_splat_u32(0);
92	92
93	93	/* Due to the fact that pix3 = pix2 + line_size, the pix3 of one
94		− iteration becomes pix2 in the next iteration. We can use this
95		− fact to avoid a potentially expensive unaligned read, each
96		− time around the loop.
97		− Read unaligned pixels into our vectors. The vectors are as follows:
98		− pix2v: pix2[0]-pix2[15]
99		− Split the pixel vectors into shorts */
	94	+ * iteration becomes pix2 in the next iteration. We can use this
	95	+ * fact to avoid a potentially expensive unaligned read, each
	96	+ * time around the loop.
	97	+ * Read unaligned pixels into our vectors. The vectors are as follows:
	98	+ * pix2v: pix2[0] - pix2[15]
	99	+ * Split the pixel vectors into shorts. */
100	100	pix2l = vec_ld( 0, pix2);
101	101	pix2r = vec_ld(15, pix2);
102	102	pix2v = vec_perm(pix2l, pix2r, perm);
103	103
104	104	for (i = 0; i < h; i++) {
105	105	/* Read unaligned pixels into our vectors. The vectors are as follows:
106		− pix1v: pix1[0]-pix1[15]
107		− pix3v: pix3[0]-pix3[15] */
	106	+ * pix1v: pix1[0] - pix1[15]
	107	+ * pix3v: pix3[0] - pix3[15] */
108	108	pix1v = vec_ld(0, pix1);
109	109
110	110	pix2l = vec_ld( 0, pix3);
111	111	pix2r = vec_ld(15, pix3);
112	112	pix3v = vec_perm(pix2l, pix2r, perm);
113	113
114		− /* Calculate the average vector */
	114	+ /* Calculate the average vector. */
115	115	avgv = vec_avg(pix2v, pix3v);
116	116
117		− /* Calculate a sum of abs differences vector */
	117	+ /* Calculate a sum of abs differences vector. */
118	118	t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
119	119
120		− /* Add each 4 pixel group together and put 4 results into sad */
	120	+ /* Add each 4 pixel group together and put 4 results into sad. */
121	121	sad = vec_sum4s(t5, sad);
122	122
123	123	pix1 += line_size;
@@ -126,7 +126,7 @@ static int sad16_y2_altivec(void v, uint8_t pix1, uint8_t *pix2, int line_size · 2 unmodified lines
126	126
127	127	}
128	128
129		− /* Sum up the four partial sums, and put the result into s */
	129	+ /* Sum up the four partial sums, and put the result into s. */
130	130	sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
131	131	sumdiffs = vec_splat(sumdiffs, 3);
132	132	vec_ste(sumdiffs, 0, &s);
@@ -157,12 +157,12 @@ static int sad16_xy2_altivec(void v, uint8_t pix1, uint8_t *pix2, int line_siz · 24 unmodified lines
157	157	s = 0;
158	158
159	159	/* Due to the fact that pix3 = pix2 + line_size, the pix3 of one
160		− iteration becomes pix2 in the next iteration. We can use this
161		− fact to avoid a potentially expensive unaligned read, as well
162		− as some splitting, and vector addition each time around the loop.
163		− Read unaligned pixels into our vectors. The vectors are as follows:
164		− pix2v: pix2[0]-pix2[15] pix2iv: pix2[1]-pix2[16]
165		− Split the pixel vectors into shorts */
	160	+ * iteration becomes pix2 in the next iteration. We can use this
	161	+ * fact to avoid a potentially expensive unaligned read, as well
	162	+ * as some splitting, and vector addition each time around the loop.
	163	+ * Read unaligned pixels into our vectors. The vectors are as follows:
	164	+ * pix2v: pix2[0] - pix2[15] pix2iv: pix2[1] - pix2[16]
	165	+ * Split the pixel vectors into shorts. */
166	166	pix2l = vec_ld( 0, pix2);
167	167	pix2r = vec_ld(16, pix2);
168	168	pix2v = vec_perm(pix2l, pix2r, perm1);
@@ -177,8 +177,8 @@ static int sad16_xy2_altivec(void v, uint8_t pix1, uint8_t *pix2, int line_siz · 8 unmodified lines
177	177
178	178	for (i = 0; i < h; i++) {
179	179	/* Read unaligned pixels into our vectors. The vectors are as follows:
180		− pix1v: pix1[0]-pix1[15]
181		− pix3v: pix3[0]-pix3[15] pix3iv: pix3[1]-pix3[16] */
	180	+ * pix1v: pix1[0] - pix1[15]
	181	+ * pix3v: pix3[0] - pix3[15] pix3iv: pix3[1] - pix3[16] */
182	182	pix1v = vec_ld(0, pix1);
183	183
184	184	pix2l = vec_ld( 0, pix3);
@@ -187,40 +187,40 @@ static int sad16_xy2_altivec(void v, uint8_t pix1, uint8_t *pix2, int line_siz · 2 unmodified lines
187	187	pix3iv = vec_perm(pix2l, pix2r, perm2);
188	188
189	189	/* Note that AltiVec does have vec_avg, but this works on vector pairs
190		− and rounds up. We could do avg(avg(a,b),avg(c,d)), but the rounding
191		− would mean that, for example, avg(3,0,0,1) = 2, when it should be 1.
192		− Instead, we have to split the pixel vectors into vectors of shorts,
193		− and do the averaging by hand. */
	190	+ * and rounds up. We could do avg(avg(a, b), avg(c, d)), but the
	191	+ * rounding would mean that, for example, avg(3, 0, 0, 1) = 2, when
	192	+ * it should be 1. Instead, we have to split the pixel vectors into
	193	+ * vectors of shorts and do the averaging by hand. */
194	194
195		− /* Split the pixel vectors into shorts */
	195	+ /* Split the pixel vectors into shorts. */
196	196	pix3hv = (vector unsigned short) vec_mergeh(zero, pix3v);
197	197	pix3lv = (vector unsigned short) vec_mergel(zero, pix3v);
198	198	pix3ihv = (vector unsigned short) vec_mergeh(zero, pix3iv);
199	199	pix3ilv = (vector unsigned short) vec_mergel(zero, pix3iv);
200	200
201		− /* Do the averaging on them */
	201	+ /* Do the averaging on them. */
202	202	t3 = vec_add(pix3hv, pix3ihv);
203	203	t4 = vec_add(pix3lv, pix3ilv);
204	204
205	205	avghv = vec_sr(vec_add(vec_add(t1, t3), two), two);
206	206	avglv = vec_sr(vec_add(vec_add(t2, t4), two), two);
207	207
208		− /* Pack the shorts back into a result */
	208	+ /* Pack the shorts back into a result. */
209	209	avgv = vec_pack(avghv, avglv);
210	210
211		− /* Calculate a sum of abs differences vector */
	211	+ /* Calculate a sum of abs differences vector. */
212	212	t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
213	213
214		− /* Add each 4 pixel group together and put 4 results into sad */
	214	+ /* Add each 4 pixel group together and put 4 results into sad. */
215	215	sad = vec_sum4s(t5, sad);
216	216
217	217	pix1 += line_size;
218	218	pix3 += line_size;
219		− /* Transfer the calculated values for pix3 into pix2 */
	219	+ /* Transfer the calculated values for pix3 into pix2. */
220	220	t1 = t3;
221	221	t2 = t4;
222	222	}
223		− /* Sum up the four partial sums, and put the result into s */
	223	+ /* Sum up the four partial sums, and put the result into s. */
224	224	sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
225	225	sumdiffs = vec_splat(sumdiffs, 3);
226	226	vec_ste(sumdiffs, 0, &s);
@@ -242,25 +242,25 @@ static int sad16_altivec(void v, uint8_t pix1, uint8_t *pix2, int line_size, i · 15 unmodified lines
242	242
243	243
244	244	for (i = 0; i < h; i++) {
245		− /* Read potentially unaligned pixels into t1 and t2 */
	245	+ /* Read potentially unaligned pixels into t1 and t2. */
246	246	vector unsigned char pix2l = vec_ld( 0, pix2);
247	247	vector unsigned char pix2r = vec_ld(15, pix2);
248	248	t1 = vec_ld(0, pix1);
249	249	t2 = vec_perm(pix2l, pix2r, perm);
250	250
251		− /* Calculate a sum of abs differences vector */
	251	+ /* Calculate a sum of abs differences vector. */
252	252	t3 = vec_max(t1, t2);
253	253	t4 = vec_min(t1, t2);
254	254	t5 = vec_sub(t3, t4);
255	255
256		− /* Add each 4 pixel group together and put 4 results into sad */
	256	+ /* Add each 4 pixel group together and put 4 results into sad. */
257	257	sad = vec_sum4s(t5, sad);
258	258
259	259	pix1 += line_size;
260	260	pix2 += line_size;
261	261	}
262	262
263		− /* Sum up the four partial sums, and put the result into s */
	263	+ /* Sum up the four partial sums, and put the result into s. */
264	264	sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
265	265	sumdiffs = vec_splat(sumdiffs, 3);
266	266	vec_ste(sumdiffs, 0, &s);
@@ -283,9 +283,9 @@ static int sad8_altivec(void v, uint8_t pix1, uint8_t *pix2, int line_size, in · 16 unmodified lines
283	283	sad = (vector unsigned int)vec_splat_u32(0);
284	284
285	285	for (i = 0; i < h; i++) {
286		− /* Read potentially unaligned pixels into t1 and t2
287		− Since we're reading 16 pixels, and actually only want 8,
288		− mask out the last 8 pixels. The 0s don't change the sum. */
	286	+ /* Read potentially unaligned pixels into t1 and t2.
	287	+ * Since we're reading 16 pixels, and actually only want 8,
	288	+ * mask out the last 8 pixels. The 0s don't change the sum. */
289	289	vector unsigned char pix1l = vec_ld(0, pix1);
290	290	vector unsigned char pix1r = vec_ld(7, pix1);
291	291	vector unsigned char pix2l = vec_ld(0, pix2);
@@ -293,19 +293,19 @@ static int sad8_altivec(void v, uint8_t pix1, uint8_t *pix2, int line_size, in · 1 unmodified line
293	293	t1 = vec_and(vec_perm(pix1l, pix1r, perm1), permclear);
294	294	t2 = vec_and(vec_perm(pix2l, pix2r, perm2), permclear);
295	295
296		− /* Calculate a sum of abs differences vector */
	296	+ /* Calculate a sum of abs differences vector. */
297	297	t3 = vec_max(t1, t2);
298	298	t4 = vec_min(t1, t2);
299	299	t5 = vec_sub(t3, t4);
300	300
301		− /* Add each 4 pixel group together and put 4 results into sad */
	301	+ /* Add each 4 pixel group together and put 4 results into sad. */
302	302	sad = vec_sum4s(t5, sad);
303	303
304	304	pix1 += line_size;
305	305	pix2 += line_size;
306	306	}
307	307
308		− /* Sum up the four partial sums, and put the result into s */
	308	+ /* Sum up the four partial sums, and put the result into s. */
309	309	sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
310	310	sumdiffs = vec_splat(sumdiffs, 3);
311	311	vec_ste(sumdiffs, 0, &s);
@@ -327,17 +327,17 @@ static int pix_norm1_altivec(uint8_t *pix, int line_size) · 15 unmodified lines
327	327
328	328	s = 0;
329	329	for (i = 0; i < 16; i++) {
330		− /* Read in the potentially unaligned pixels */
	330	+ /* Read the potentially unaligned pixels. */
331	331	vector unsigned char pixl = vec_ld( 0, pix);
332	332	vector unsigned char pixr = vec_ld(15, pix);
333	333	pixv = vec_perm(pixl, pixr, perm);
334	334
335		− /* Square the values, and add them to our sum */
	335	+ /* Square the values, and add them to our sum. */
336	336	sv = vec_msum(pixv, pixv, sv);
337	337
338	338	pix += line_size;
339	339	}
340		− /* Sum up the four partial sums, and put the result into s */
	340	+ /* Sum up the four partial sums, and put the result into s. */
341	341	sum = vec_sums((vector signed int) sv, (vector signed int) zero);
342	342	sum = vec_splat(sum, 3);
343	343	vec_ste(sum, 0, &s);
@@ -345,11 +345,8 @@ static int pix_norm1_altivec(uint8_t *pix, int line_size) · 1 unmodified line
345	345	return s;
346	346	}
347	347
348		−/**
349		− * Sum of Squared Errors for a 8x8 block.
350		− * AltiVec-enhanced.
351		− * It's the sad8_altivec code above w/ squaring added.
352		− */
	348	+/* Sum of Squared Errors for an 8x8 block, AltiVec-enhanced.
	349	+ * It's the sad8_altivec code above w/ squaring added. */
353	350	static int sse8_altivec(void v, uint8_t pix1, uint8_t *pix2, int line_size, int h)
354	351	{
355	352	int i;
@@ -365,9 +362,9 @@ static int sse8_altivec(void v, uint8_t pix1, uint8_t *pix2, int line_size, in · 9 unmodified lines
365	362	sum = (vector unsigned int)vec_splat_u32(0);
366	363
367	364	for (i = 0; i < h; i++) {
368		− /* Read potentially unaligned pixels into t1 and t2
369		− Since we're reading 16 pixels, and actually only want 8,
370		− mask out the last 8 pixels. The 0s don't change the sum. */
	365	+ /* Read potentially unaligned pixels into t1 and t2.
	366	+ * Since we're reading 16 pixels, and actually only want 8,
	367	+ * mask out the last 8 pixels. The 0s don't change the sum. */
371	368	vector unsigned char pix1l = vec_ld(0, pix1);
372	369	vector unsigned char pix1r = vec_ld(7, pix1);
373	370	vector unsigned char pix2l = vec_ld(0, pix2);
@@ -376,21 +373,21 @@ static int sse8_altivec(void v, uint8_t pix1, uint8_t *pix2, int line_size, in · 2 unmodified lines
376	373	t2 = vec_and(vec_perm(pix2l, pix2r, perm2), permclear);
377	374
378	375	/* Since we want to use unsigned chars, we can take advantage
379		− of the fact that abs(a-b)^2 = (a-b)^2. */
	376	+ * of the fact that abs(a - b) ^ 2 = (a - b) ^ 2. */
380	377
381		− /* Calculate abs differences vector */
	378	+ /* Calculate abs differences vector. */
382	379	t3 = vec_max(t1, t2);
383	380	t4 = vec_min(t1, t2);
384	381	t5 = vec_sub(t3, t4);
385	382
386		− /* Square the values and add them to our sum */
	383	+ /* Square the values and add them to our sum. */
387	384	sum = vec_msum(t5, t5, sum);
388	385
389	386	pix1 += line_size;
390	387	pix2 += line_size;
391	388	}
392	389
393		− /* Sum up the four partial sums, and put the result into s */
	390	+ /* Sum up the four partial sums, and put the result into s. */
394	391	sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
395	392	sumsqr = vec_splat(sumsqr, 3);
396	393	vec_ste(sumsqr, 0, &s);
@@ -398,11 +395,8 @@ static int sse8_altivec(void v, uint8_t pix1, uint8_t *pix2, int line_size, in · 1 unmodified line
398	395	return s;
399	396	}
400	397
401		−/**
402		− * Sum of Squared Errors for a 16x16 block.
403		− * AltiVec-enhanced.
404		− * It's the sad16_altivec code above w/ squaring added.
405		− */
	398	+/* Sum of Squared Errors for a 16x16 block, AltiVec-enhanced.
	399	+ * It's the sad16_altivec code above w/ squaring added. */
406	400	static int sse16_altivec(void v, uint8_t pix1, uint8_t *pix2, int line_size, int h)
407	401	{
408	402	int i;
@@ -416,28 +410,28 @@ static int sse16_altivec(void v, uint8_t pix1, uint8_t *pix2, int line_size, i · 7 unmodified lines
416	410	sum = (vector unsigned int)vec_splat_u32(0);
417	411
418	412	for (i = 0; i < h; i++) {
419		− /* Read potentially unaligned pixels into t1 and t2 */
	413	+ /* Read potentially unaligned pixels into t1 and t2. */
420	414	vector unsigned char pix2l = vec_ld( 0, pix2);
421	415	vector unsigned char pix2r = vec_ld(15, pix2);
422	416	t1 = vec_ld(0, pix1);
423	417	t2 = vec_perm(pix2l, pix2r, perm);
424	418
425	419	/* Since we want to use unsigned chars, we can take advantage
426		− of the fact that abs(a-b)^2 = (a-b)^2. */
	420	+ * of the fact that abs(a - b) ^ 2 = (a - b) ^ 2. */
427	421
428		− /* Calculate abs differences vector */
	422	+ /* Calculate abs differences vector. */
429	423	t3 = vec_max(t1, t2);
430	424	t4 = vec_min(t1, t2);
431	425	t5 = vec_sub(t3, t4);
432	426
433		− /* Square the values and add them to our sum */
	427	+ /* Square the values and add them to our sum. */
434	428	sum = vec_msum(t5, t5, sum);
435	429
436	430	pix1 += line_size;
437	431	pix2 += line_size;
438	432	}
439	433
440		− /* Sum up the four partial sums, and put the result into s */
	434	+ /* Sum up the four partial sums, and put the result into s. */
441	435	sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
442	436	sumsqr = vec_splat(sumsqr, 3);
443	437	vec_ste(sumsqr, 0, &s);
@@ -459,18 +453,18 @@ static int pix_sum_altivec(uint8_t * pix, int line_size) · 15 unmodified lines
459	453	sad = (vector unsigned int)vec_splat_u32(0);
460	454
461	455	for (i = 0; i < 16; i++) {
462		− /* Read the potentially unaligned 16 pixels into t1 */
	456	+ /* Read the potentially unaligned 16 pixels into t1. */
463	457	vector unsigned char pixl = vec_ld( 0, pix);
464	458	vector unsigned char pixr = vec_ld(15, pix);
465	459	t1 = vec_perm(pixl, pixr, perm);
466	460
467		− /* Add each 4 pixel group together and put 4 results into sad */
	461	+ /* Add each 4 pixel group together and put 4 results into sad. */
468	462	sad = vec_sum4s(t1, sad);
469	463
470	464	pix += line_size;
471	465	}
472	466
473		− /* Sum up the four partial sums, and put the result into s */
	467	+ /* Sum up the four partial sums, and put the result into s. */
474	468	sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
475	469	sumdiffs = vec_splat(sumdiffs, 3);
476	470	vec_ste(sumdiffs, 0, &s);
@@ -487,6 +481,9 @@ static void get_pixels_altivec(int16_t restrict block, const uint8_t pixels, i · 10 unmodified lines
487	481	vector signed short shorts;
488	482
489	483	for (i = 0; i < 8; i++) {
	484	+ /* Read potentially unaligned pixels.
	485	+ * We're reading 16 pixels, and actually only want 8,
	486	+ * but we simply ignore the extras. */
490	487	// Read potentially unaligned pixels.
491	488	// We're reading 16 pixels, and actually only want 8,
492	489	// but we simply ignore the extras.
@@ -494,10 +491,10 @@ static void get_pixels_altivec(int16_t restrict block, const uint8_t pixels, i · 1 unmodified line
494	491	vector unsigned char pixr = vec_ld(7, pixels);
495	492	bytes = vec_perm(pixl, pixr, perm);
496	493
497		− // convert the bytes into shorts
	494	+ // Convert the bytes into shorts.
498	495	shorts = (vector signed short)vec_mergeh(zero, bytes);
499	496
500		− // save the data to the block, we assume the block is 16-byte aligned
	497	+ // Save the data to the block, we assume the block is 16-byte aligned.
501	498	vec_st(shorts, i16, (vector signed short)block);
502	499
503	500	pixels += line_size;
@@ -515,60 +512,59 @@ static void diff_pixels_altivec(int16_t restrict block, const uint8_t s1, · 11 unmodified lines
515	512	vector signed short shorts1, shorts2;
516	513
517	514	for (i = 0; i < 4; i++) {
518		− // Read potentially unaligned pixels
519		− // We're reading 16 pixels, and actually only want 8,
520		− // but we simply ignore the extras.
	515	+ /* Read potentially unaligned pixels.
	516	+ * We're reading 16 pixels, and actually only want 8,
	517	+ * but we simply ignore the extras. */
521	518	pixl = vec_ld( 0, s1);
522	519	pixr = vec_ld(15, s1);
523	520	bytes = vec_perm(pixl, pixr, perm1);
524	521
525		− // convert the bytes into shorts
	522	+ // Convert the bytes into shorts.
526	523	shorts1 = (vector signed short)vec_mergeh(zero, bytes);
527	524
528		− // Do the same for the second block of pixels
	525	+ // Do the same for the second block of pixels.
529	526	pixl = vec_ld( 0, s2);
530	527	pixr = vec_ld(15, s2);
531	528	bytes = vec_perm(pixl, pixr, perm2);
532	529
533		− // convert the bytes into shorts
	530	+ // Convert the bytes into shorts.
534	531	shorts2 = (vector signed short)vec_mergeh(zero, bytes);
535	532
536		− // Do the subtraction
	533	+ // Do the subtraction.
537	534	shorts1 = vec_sub(shorts1, shorts2);
538	535
539		− // save the data to the block, we assume the block is 16-byte aligned
	536	+ // Save the data to the block, we assume the block is 16-byte aligned.
540	537	vec_st(shorts1, 0, (vector signed short*)block);
541	538
542	539	s1 += stride;
543	540	s2 += stride;
544	541	block += 8;
545	542
	543	+ /* The code below is a copy of the code above...
	544	+ * This is a manual unroll. */
546	545
547		− // The code below is a copy of the code above... This is a manual
548		− // unroll.
549		−
550		− // Read potentially unaligned pixels
551		− // We're reading 16 pixels, and actually only want 8,
552		− // but we simply ignore the extras.
	546	+ /* Read potentially unaligned pixels.
	547	+ * We're reading 16 pixels, and actually only want 8,
	548	+ * but we simply ignore the extras. */
553	549	pixl = vec_ld( 0, s1);
554	550	pixr = vec_ld(15, s1);
555	551	bytes = vec_perm(pixl, pixr, perm1);
556	552
557		− // convert the bytes into shorts
	553	+ // Convert the bytes into shorts.
558	554	shorts1 = (vector signed short)vec_mergeh(zero, bytes);
559	555
560		− // Do the same for the second block of pixels
	556	+ // Do the same for the second block of pixels.
561	557	pixl = vec_ld( 0, s2);
562	558	pixr = vec_ld(15, s2);
563	559	bytes = vec_perm(pixl, pixr, perm2);
564	560
565		− // convert the bytes into shorts
	561	+ // Convert the bytes into shorts.
566	562	shorts2 = (vector signed short)vec_mergeh(zero, bytes);
567	563
568		− // Do the subtraction
	564	+ // Do the subtraction.
569	565	shorts1 = vec_sub(shorts1, shorts2);
570	566
571		− // save the data to the block, we assume the block is 16-byte aligned
	567	+ // Save the data to the block, we assume the block is 16-byte aligned.
572	568	vec_st(shorts1, 0, (vector signed short*)block);
573	569
574	570	s1 += stride;
@@ -595,14 +591,14 @@ static void add_bytes_altivec(uint8_t dst, uint8_t src, int w) { · 20 unmodified lines
595	591	register int i;
596	592	register vector unsigned char vdst, vsrc;
597	593
598		− /* dst and src are 16 bytes-aligned (guaranteed) */
	594	+ /* dst and src are 16 bytes-aligned (guaranteed). */
599	595	for (i = 0 ; (i + 15) < w ; i+=16) {
600	596	vdst = vec_ld(i, (unsigned char*)dst);
601	597	vsrc = vec_ld(i, (unsigned char*)src);
602	598	vdst = vec_add(vsrc, vdst);
603	599	vec_st(vdst, i, (unsigned char*)dst);
604	600	}
605		− /* if w is not a multiple of 16 */
	601	+ /* If w is not a multiple of 16. */
606	602	for (; (i < w) ; i++) {
607	603	dst[i] = src[i];
608	604	}
@@ -643,8 +639,8 @@ static int hadamard8_diff8x8_altivec(/MpegEncContext/ void s, uint8_t dst, u · 34 unmodified lines
643	639	dst1 = vec_ld(stride * i, dst); \
644	640	dst2 = vec_ld((stride * i) + 15, dst); \
645	641	dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
646		− /* promote the unsigned chars to signed shorts */ \
647		− /* we're in the 8x8 function, we only care for the first 8 */ \
	642	+ /* Promote the unsigned chars to signed shorts. */ \
	643	+ /* We're in the 8x8 function, we only care for the first 8. */ \
648	644	srcV = (vector signed short)vec_mergeh((vector signed char)vzero, \
649	645	(vector signed char)srcO); \
650	646	dstV = (vector signed short)vec_mergeh((vector signed char)vzero, \
@@ -713,24 +709,23 @@ static int hadamard8_diff8x8_altivec(/MpegEncContext/ void s, uint8_t dst, u · 62 unmodified lines
713	709	}
714	710
715	711	/*
716		−16x8 works with 16 elements; it allows to avoid replicating loads, and
717		−give the compiler more rooms for scheduling. It's only used from
718		−inside hadamard8_diff16_altivec.
719		−
720		−Unfortunately, it seems gcc-3.3 is a bit dumb, and the compiled code has a LOT
721		−of spill code, it seems gcc (unlike xlc) cannot keep everything in registers
722		−by itself. The following code include hand-made registers allocation. It's not
723		−clean, but on a 7450 the resulting code is much faster (best case fall from
724		−700+ cycles to 550).
725		−
726		−xlc doesn't add spill code, but it doesn't know how to schedule for the 7450,
727		−and its code isn't much faster than gcc-3.3 on the 7450 (but uses 25% less
728		−instructions...)
729		−
730		−On the 970, the hand-made RA is still a win (around 690 vs. around 780), but
731		−xlc goes to around 660 on the regular C code...
732		−*/
733		−
	712	+ * 16x8 works with 16 elements; it allows to avoid replicating loads, and
	713	+ * gives the compiler more room for scheduling. It's only used from
	714	+ * inside hadamard8_diff16_altivec.
	715	+ *
	716	+ * Unfortunately, it seems gcc-3.3 is a bit dumb, and the compiled code has
	717	+ * a LOT of spill code, it seems gcc (unlike xlc) cannot keep everything in
	718	+ * registers by itself. The following code includes hand-made register
	719	+ * allocation. It's not clean, but on a 7450 the resulting code is much faster
	720	+ * (best case falls from 700+ cycles to 550).
	721	+ *
	722	+ * xlc doesn't add spill code, but it doesn't know how to schedule for the
	723	+ * 7450, and its code isn't much faster than gcc-3.3 on the 7450 (but uses
	724	+ * 25% fewer instructions...)
	725	+ *
	726	+ * On the 970, the hand-made RA is still a win (around 690 vs. around 780),
	727	+ * but xlc goes to around 660 on the regular C code...
	728	+ */
734	729	static int hadamard8_diff16x8_altivec(/MpegEncContext/ void s, uint8_t dst, uint8_t *src, int stride, int h) {
735	730	int sum;
736	731	register vector signed short
@@ -805,7 +800,7 @@ static int hadamard8_diff16x8_altivec(/MpegEncContext/ void s, uint8_t dst, · 68 unmodified lines
805	800	dst1 = vec_ld(stride * i, dst); \
806	801	dst2 = vec_ld((stride * i) + 16, dst); \
807	802	dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
808		− /* promote the unsigned chars to signed shorts */ \
	803	+ /* Promote the unsigned chars to signed shorts. */ \
809	804	srcV = (vector signed short)vec_mergeh((vector signed char)vzero, \
810	805	(vector signed char)srcO); \
811	806	dstV = (vector signed short)vec_mergeh((vector signed char)vzero, \

libavcodec/ppc/dsputil_ppc.c+29−31

@@ -32,24 +32,23 @@ · 31 unmodified lines
32	32
33	33	/* *** WARNING * WARNING * WARNING *** */
34	34	/*
35		−clear_blocks_dcbz32_ppc will not work properly on PowerPC processors with a
36		−cache line size not equal to 32 bytes.
37		−Fortunately all processor used by Apple up to at least the 7450 (aka second
38		−generation G4) use 32 bytes cache line.
39		−This is due to the use of the 'dcbz' instruction. It simply clear to zero a
40		−single cache line, so you need to know the cache line size to use it !
41		−It's absurd, but it's fast...
42		−
43		−update 24/06/2003 : Apple released yesterday the G5, with a PPC970. cache line
44		−size: 128 bytes. Oups.
45		−The semantic of dcbz was changed, it always clear 32 bytes. so the function
46		−below will work, but will be slow. So I fixed check_dcbz_effect to use dcbzl,
47		−which is defined to clear a cache line (as dcbz before). So we still can
48		−distinguish, and use dcbz (32 bytes) or dcbzl (one cache line) as required.
49		−
50		−see <http://developer.apple.com/technotes/tn/tn2087.html>
51		−and <http://developer.apple.com/technotes/tn/tn2086.html>
52		−*/
	35	+ * clear_blocks_dcbz32_ppc will not work properly on PowerPC processors with
	36	+ * a cache line size not equal to 32 bytes. Fortunately all processors used
	37	+ * by Apple up to at least the 7450 (AKA second generation G4) use 32-byte
	38	+ * cache lines. This is due to the use of the 'dcbz' instruction. It simply
	39	+ * clears a single cache line to zero, so you need to know the cache line
	40	+ * size to use it! It's absurd, but it's fast...
	41	+ *
	42	+ * update 24/06/2003: Apple released the G5 yesterday, with a PPC970.
	43	+ * cache line size: 128 bytes. Oups.
	44	+ * The semantics of dcbz was changed, it always clears 32 bytes. So the function
	45	+ * below will work, but will be slow. So I fixed check_dcbz_effect to use dcbzl,
	46	+ * which is defined to clear a cache line (as dcbz before). So we can still
	47	+ * distinguish, and use dcbz (32 bytes) or dcbzl (one cache line) as required.
	48	+ *
	49	+ * see <http://developer.apple.com/technotes/tn/tn2087.html>
	50	+ * and <http://developer.apple.com/technotes/tn/tn2086.html>
	51	+ */
53	52	static void clear_blocks_dcbz32_ppc(int16_t *blocks)
54	53	{
55	54	register int misal = ((unsigned long)blocks & 0x00000010);
@@ -73,17 +72,17 @@ static void clear_blocks_dcbz32_ppc(int16_t *blocks) · 17 unmodified lines
73	72	}
74	73	}
75	74
76		−/* same as above, when dcbzl clear a whole 128B cache line
77		− i.e. the PPC970 aka G5 */
	75	+/* Same as above, when dcbzl clears a whole 128 bytes cache line
	76	+ * i.e. the PPC970 AKA G5. */
78	77	#if HAVE_DCBZL
79	78	static void clear_blocks_dcbz128_ppc(int16_t *blocks)
80	79	{
81	80	register int misal = ((unsigned long)blocks & 0x0000007f);
82	81	register int i = 0;
83	82	if (misal) {
84		− // we could probably also optimize this case,
85		− // but there's not much point as the machines
86		− // aren't available yet (2003-06-26)
	83	+ /* We could probably also optimize this case,
	84	+ * but there's not much point as the machines
	85	+ * aren't available yet (2003-06-26). */
87	86	memset(blocks, 0, sizeof(int16_t)664);
88	87	}
89	88	else
@@ -99,11 +98,10 @@ static void clear_blocks_dcbz128_ppc(int16_t *blocks) · 9 unmodified lines
99	98	#endif
100	99
101	100	#if HAVE_DCBZL
102		−/* check dcbz report how many bytes are set to 0 by dcbz */
103		−/* update 24/06/2003 : replace dcbz by dcbzl to get
104		− the intended effect (Apple "fixed" dcbz)
105		− unfortunately this cannot be used unless the assembler
106		− knows about dcbzl ... */
	101	+/* Check dcbz report how many bytes are set to 0 by dcbz. */
	102	+/* update 24/06/2003: Replace dcbz by dcbzl to get the intended effect
	103	+ * (Apple "fixed" dcbz). Unfortunately this cannot be used unless the
	104	+ * assembler knows about dcbzl ... */
107	105	static long check_dcbzl_effect(void)
108	106	{
109	107	register char *fakedata = av_malloc(1024);
@@ -120,8 +118,8 @@ static long check_dcbzl_effect(void) · 10 unmodified lines
120	118
121	119	memset(fakedata, 0xFF, 1024);
122	120
123		− /* below the constraint "b" seems to mean "Address base register"
124		− in gcc-3.3 / RS/6000 speaks. seems to avoid using r0, so.... */
	121	+ /* Below the constraint "b" seems to mean "address base register"
	122	+ * in gcc-3.3 / RS/6000 speaks. Seems to avoid using r0, so.... */
125	123	__asm__ volatile("dcbzl %0, %1" : : "b" (fakedata_middle), "r" (zero));
126	124
127	125	for (i = 0; i < 1024 ; i ++) {
@@ -145,7 +143,7 @@ av_cold void ff_dsputil_init_ppc(DSPContext c, AVCodecContext avctx) · 17 unmodified lines
145	143	const int high_bit_depth = avctx->bits_per_raw_sample > 8;
146	144	int mm_flags = av_get_cpu_flags();
147	145
148		− // Common optimizations whether AltiVec is available or not
	146	+ // common optimizations whether AltiVec is available or not
149	147	if (!high_bit_depth) {
150	148	switch (check_dcbzl_effect()) {
151	149	case 32:

libavcodec/ppc/fdct_altivec.c+3−4

@@ -259,11 +259,10 @@ void ff_fdct_altivec(int16_t *block) · 258 unmodified lines
259	259	#undef MERGE_S16
260	260	/* }}} */
261	261
	262	+ /* Some of the initial calculations can be done as vector short
	263	+ * before conversion to vector float. The following code section
	264	+ * takes advantage of this. */
262	265
263		−/* Some of the initial calculations can be done as vector short before
264		− * conversion to vector float. The following code section takes advantage
265		− * of this.
266		− */
267	266	/* fdct rows {{{ */
268	267	x0 = ((vector float)vec_add(vs16(b00), vs16(b70)));
269	268	x7 = ((vector float)vec_sub(vs16(b00), vs16(b70)));

libavcodec/ppc/fft_altivec.c+6−6

@@ -27,12 +27,12 @@ · 26 unmodified lines
27	27	#include "libavcodec/fft.h"
28	28
29	29	/**
30		− * Do a complex FFT with the parameters defined in ff_fft_init(). The
31		− * input data must be permuted before with s->revtab table. No
32		− * 1.0/sqrt(n) normalization is done.
33		− * AltiVec-enabled
34		− * This code assumes that the 'z' pointer is 16 bytes-aligned
35		− * It also assumes all FFTComplex are 8 bytes-aligned pair of float
	30	+ * Do a complex FFT with the parameters defined in ff_fft_init().
	31	+ * The input data must be permuted before with s->revtab table.
	32	+ * No 1.0 / sqrt(n) normalization is done.
	33	+ * AltiVec-enabled:
	34	+ * This code assumes that the 'z' pointer is 16 bytes-aligned.
	35	+ * It also assumes all FFTComplex are 8 bytes-aligned pairs of floats.
36	36	*/
37	37
38	38	void ff_fft_calc_altivec(FFTContext s, FFTComplex z);

libavcodec/ppc/gmc_altivec.c+17−23

@@ -1,6 +1,6 @@
1	1	/*
2		− * GMC (Global Motion Compensation)
3		− * AltiVec-enabled
	2	+ * GMC (Global Motion Compensation), AltiVec-enabled
	3	+ *
4	4	* Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org>
5	5	*
6	6	* This file is part of FFmpeg.
@@ -25,10 +25,8 @@ · 18 unmodified lines
25	25	#include "libavutil/ppc/util_altivec.h"
26	26	#include "dsputil_altivec.h"
27	27
28		−/*
29		− altivec-enhanced gmc1. ATM this code assume stride is a multiple of 8,
30		− to preserve proper dst alignment.
31		−*/
	28	+/* AltiVec-enhanced gmc1. ATM this code assumes stride is a multiple of 8
	29	+ * to preserve proper dst alignment. */
32	30	void ff_gmc1_altivec(uint8_t dst / align 8 /, uint8_t src /* align1 */, int stride, int h, int x16, int y16, int rounder)
33	31	{
34	32	const DECLARE_ALIGNED(16, unsigned short, rounder_a) = rounder;
@@ -56,18 +54,16 @@ void ff_gmc1_altivec(uint8_t dst / align 8 /, uint8_t src /* align1 */, int · 21 unmodified lines
56	54
57	55	rounderV = vec_splat((vec_u16)vec_lde(0, &rounder_a), 0);
58	56
59		− // we'll be able to pick-up our 9 char elements
60		− // at src from those 32 bytes
61		− // we load the first batch here, as inside the loop
62		− // we can re-use 'src+stride' from one iteration
63		− // as the 'src' of the next.
	57	+ /* we'll be able to pick-up our 9 char elements at src from those
	58	+ * 32 bytes we load the first batch here, as inside the loop we can
	59	+ * reuse 'src + stride' from one iteration as the 'src' of the next. */
64	60	src_0 = vec_ld(0, src);
65	61	src_1 = vec_ld(16, src);
66	62	srcvA = vec_perm(src_0, src_1, vec_lvsl(0, src));
67	63
68	64	if (src_really_odd != 0x0000000F) {
69		− // if (src & 0xF) == 0xF, then (src+1) is properly aligned
70		− // on the second vector.
	65	+ /* If (src & 0xF) == 0xF, then (src + 1) is properly aligned
	66	+ * on the second vector. */
71	67	srcvB = vec_perm(src_0, src_1, vec_lvsl(1, src));
72	68	} else {
73	69	srcvB = src_1;
@@ -81,17 +77,16 @@ void ff_gmc1_altivec(uint8_t dst / align 8 /, uint8_t src /* align1 */, int · 7 unmodified lines
81	77
82	78	dstv = vec_ld(0, dst);
83	79
84		− // we we'll be able to pick-up our 9 char elements
85		− // at src + stride from those 32 bytes
86		− // then reuse the resulting 2 vectors srvcC and srcvD
87		− // as the next srcvA and srcvB
	80	+ /* We'll be able to pick-up our 9 char elements at src + stride from
	81	+ * those 32 bytes then reuse the resulting 2 vectors srvcC and srcvD
	82	+ * as the next srcvA and srcvB. */
88	83	src_0 = vec_ld(stride + 0, src);
89	84	src_1 = vec_ld(stride + 16, src);
90	85	srcvC = vec_perm(src_0, src_1, vec_lvsl(stride + 0, src));
91	86
92	87	if (src_really_odd != 0x0000000F) {
93		− // if (src & 0xF) == 0xF, then (src+1) is properly aligned
94		− // on the second vector.
	88	+ /* If (src & 0xF) == 0xF, then (src + 1) is properly aligned
	89	+ * on the second vector. */
95	90	srcvD = vec_perm(src_0, src_1, vec_lvsl(stride + 1, src));
96	91	} else {
97	92	srcvD = src_1;
@@ -100,10 +95,9 @@ void ff_gmc1_altivec(uint8_t dst / align 8 /, uint8_t src /* align1 */, int · 2 unmodified lines
100	95	srcvC = vec_mergeh(vczero, srcvC);
101	96	srcvD = vec_mergeh(vczero, srcvD);
102	97
103		−
104		− // OK, now we (finally) do the math :-)
105		− // those four instructions replaces 32 int muls & 32 int adds.
106		− // isn't AltiVec nice ?
	98	+ /* OK, now we (finally) do the math :-)
	99	+ * Those four instructions replace 32 int muls & 32 int adds.
	100	+ * Isn't AltiVec nice? */
107	101	tempA = vec_mladd((vector unsigned short)srcvA, Av, rounderV);
108	102	tempB = vec_mladd((vector unsigned short)srcvB, Bv, tempA);
109	103	tempC = vec_mladd((vector unsigned short)srcvC, Cv, tempB);

libavcodec/ppc/idct_altivec.c+4−9

@@ -18,24 +18,19 @@ · 17 unmodified lines
18	18	* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19	19	*/
20	20
21		−/*
22		− * NOTE: This code is based on GPL code from the libmpeg2 project. The
	21	+/* NOTE: This code is based on GPL code from the libmpeg2 project. The
23	22	* author, Michel Lespinasses, has given explicit permission to release
24	23	* under LGPL as part of FFmpeg.
25		− */
26		−
27		−/*
	24	+ *
28	25	* FFmpeg integration by Dieter Shirley
29	26	*
30	27	* This file is a direct copy of the AltiVec IDCT module from the libmpeg2
31	28	* project. I've deleted all of the libmpeg2-specific code, renamed the
32	29	* functions and reordered the function parameters. The only change to the
33	30	* IDCT function itself was to factor out the partial transposition, and to
34		− * perform a full transpose at the end of the function.
35		− */
36		−
	31	+ * perform a full transpose at the end of the function. */
37	32
38		−#include <stdlib.h> /* malloc(), free() */
	33	+#include <stdlib.h>
39	34	#include <string.h>
40	35	#include "config.h"
41	36	#if HAVE_ALTIVEC_H

libavcodec/ppc/int_altivec.c+7−7

@@ -19,9 +19,9 @@ · 18 unmodified lines
19	19	*/
20	20
21	21	/**
22		− ** @file
23		− ** integer misc ops.
24		− **/
	22	+ * @file
	23	+ * miscellaneous integer operations
	24	+ */
25	25
26	26	#include "config.h"
27	27	#if HAVE_ALTIVEC_H
@@ -43,8 +43,8 @@ static int ssd_int8_vs_int16_altivec(const int8_t pix1, const int16_t pix2, · 15 unmodified lines
43	43	int32_t score[4];
44	44	} u;
45	45	u.vscore = vec_splat_s32(0);
46		−//
47		−//XXX lazy way, fix it later
	46	+
	47	+// XXX lazy way, fix it later
48	48
49	49	#define vec_unaligned_load(b) \
50	50	vec_perm(vec_ld(0,b),vec_ld(15,b),vec_lvsl(0, b));
@@ -52,12 +52,12 @@ static int ssd_int8_vs_int16_altivec(const int8_t pix1, const int16_t pix2, · 1 unmodified line
52	52	size16 = size >> 4;
53	53	while(size16) {
54	54	// score += (pix1[i]-pix2[i])*(pix1[i]-pix2[i]);
55		− //load pix1 and the first batch of pix2
	55	+ // load pix1 and the first batch of pix2
56	56
57	57	vpix1 = vec_unaligned_load(pix1);
58	58	vpix2 = vec_unaligned_load(pix2);
59	59	pix2 += 8;
60		− //unpack
	60	+ // unpack
61	61	vpix1h = vec_unpackh(vpix1);
62	62	vdiff = vec_sub(vpix1h, vpix2);
63	63	vpix1l = vec_unpackl(vpix1);

Merge commit '82ee14d2cedd7867920529b408ed6c7ec2f13ff1'

Keyboard shortcuts

General

Issues and pull requests

Issues, pull requests and Inbox

Repository code