Cosmetics: should not hurt performance, scream if are

Originally committed as revision 5493 to svn://svn.ffmpeg.org/ffmpeg/trunk

Luca Barbato committed 20 years ago

Commit e8772ee1 parent a97c1e92 files · +393 −291Browse files

Showing the changes to libavcodec/ppc (2 of 2 files). Show all

Unified · Split

2 files

libavcodec/ppc/dsputil_altivec.c +97 −61
libavcodec/ppc/dsputil_h264_template_altivec.c +296 −230

@@ -1311,9 +1311,9 @@ POWERPC_PERF_STOP_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1); · 1310 unmodified lines
1311	1311	int hadamard8_diff8x8_altivec(/MpegEncContext/ void s, uint8_t dst, uint8_t *src, int stride, int h){
1312	1312	POWERPC_PERF_DECLARE(altivec_hadamard8_diff8x8_num, 1);
1313	1313	int sum;
1314		−POWERPC_PERF_START_COUNT(altivec_hadamard8_diff8x8_num, 1);
1315	1314	register const_vector unsigned char vzero = (const_vector unsigned char)vec_splat_u8(0);
1316	1315	register vector signed short temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
	1316	+POWERPC_PERF_START_COUNT(altivec_hadamard8_diff8x8_num, 1);
1317	1317	{
1318	1318	register const_vector signed short vprod1 = (const_vector signed short)AVV( 1,-1, 1,-1, 1,-1, 1,-1);
1319	1319	register const_vector signed short vprod2 = (const_vector signed short)AVV( 1, 1,-1,-1, 1, 1,-1,-1);
@@ -1338,6 +1338,8 @@ POWERPC_PERF_START_COUNT(altivec_hadamard8_diff8x8_num, 1); · 18 unmodified lines
1338	1338	{ \
1339	1339	register vector unsigned char src1, src2, srcO; \
1340	1340	register vector unsigned char dst1, dst2, dstO; \
	1341	+ register vector signed short srcV, dstV; \
	1342	+ register vector signed short but0, but1, but2, op1, op2, op3; \
1341	1343	src1 = vec_ld(stride * i, src); \
1342	1344	if ((((stride * i) + (unsigned long)src) & 0x0000000F) > 8) \
1343	1345	src2 = vec_ld((stride * i) + 16, src); \
@@ -1348,17 +1350,19 @@ POWERPC_PERF_START_COUNT(altivec_hadamard8_diff8x8_num, 1); · 4 unmodified lines
1348	1350	dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
1349	1351	/* promote the unsigned chars to signed shorts */ \
1350	1352	/* we're in the 8x8 function, we only care for the first 8 */ \
1351		− register vector signed short srcV = \
1352		− (vector signed short)vec_mergeh((vector signed char)vzero, (vector signed char)srcO); \
1353		− register vector signed short dstV = \
1354		− (vector signed short)vec_mergeh((vector signed char)vzero, (vector signed char)dstO); \
	1353	+ srcV = \
	1354	+ (vector signed short)vec_mergeh((vector signed char)vzero, \
	1355	+ (vector signed char)srcO); \
	1356	+ dstV = \
	1357	+ (vector signed short)vec_mergeh((vector signed char)vzero, \
	1358	+ (vector signed char)dstO); \
1355	1359	/* substractions inside the first butterfly */ \
1356		− register vector signed short but0 = vec_sub(srcV, dstV); \
1357		− register vector signed short op1 = vec_perm(but0, but0, perm1); \
1358		− register vector signed short but1 = vec_mladd(but0, vprod1, op1); \
1359		− register vector signed short op2 = vec_perm(but1, but1, perm2); \
1360		− register vector signed short but2 = vec_mladd(but1, vprod2, op2); \
1361		− register vector signed short op3 = vec_perm(but2, but2, perm3); \
	1360	+ but0 = vec_sub(srcV, dstV); \
	1361	+ op1 = vec_perm(but0, but0, perm1); \
	1362	+ but1 = vec_mladd(but0, vprod1, op1); \
	1363	+ op2 = vec_perm(but1, but1, perm2); \
	1364	+ but2 = vec_mladd(but1, vprod2, op2); \
	1365	+ op3 = vec_perm(but2, but2, perm3); \
1362	1366	res = vec_mladd(but2, vprod3, op3); \
1363	1367	}
1364	1368	ONEITERBUTTERFLY(0, temp0);
@@ -1481,37 +1485,63 @@ static int hadamard8_diff16x8_altivec(/MpegEncContext/ void s, uint8_t dst, · 116 unmodified lines
1481	1485
1482	1486	#define ONEITERBUTTERFLY(i, res1, res2) \
1483	1487	{ \
1484		− register vector unsigned char src1 REG_v(v22), src2 REG_v(v23); \
1485		− register vector unsigned char dst1 REG_v(v24), dst2 REG_v(v25); \
	1488	+ register vector unsigned char src1 REG_v(v22), \
	1489	+ src2 REG_v(v23), \
	1490	+ dst1 REG_v(v24), \
	1491	+ dst2 REG_v(v25), \
	1492	+ srcO REG_v(v22), \
	1493	+ dstO REG_v(v23); \
	1494	+ \
	1495	+ register vector signed short srcV REG_v(v24), \
	1496	+ dstV REG_v(v25), \
	1497	+ srcW REG_v(v26), \
	1498	+ dstW REG_v(v27), \
	1499	+ but0 REG_v(v28), \
	1500	+ but0S REG_v(v29), \
	1501	+ op1 REG_v(v30), \
	1502	+ but1 REG_v(v22), \
	1503	+ op1S REG_v(v23), \
	1504	+ but1S REG_v(v24), \
	1505	+ op2 REG_v(v25), \
	1506	+ but2 REG_v(v26), \
	1507	+ op2S REG_v(v27), \
	1508	+ but2S REG_v(v28), \
	1509	+ op3 REG_v(v29), \
	1510	+ op3S REG_v(v30); \
	1511	+ \
1486	1512	src1 = vec_ld(stride * i, src); \
1487	1513	src2 = vec_ld((stride * i) + 16, src); \
1488		− register vector unsigned char srcO REG_v(v22) = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
	1514	+ srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
1489	1515	dst1 = vec_ld(stride * i, dst); \
1490	1516	dst2 = vec_ld((stride * i) + 16, dst); \
1491		− register vector unsigned char dstO REG_v(v23) = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
	1517	+ dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
1492	1518	/* promote the unsigned chars to signed shorts */ \
1493		− register vector signed short srcV REG_v(v24) = \
1494		− (vector signed short)vec_mergeh((vector signed char)vzero, (vector signed char)srcO); \
1495		− register vector signed short dstV REG_v(v25) = \
1496		− (vector signed short)vec_mergeh((vector signed char)vzero, (vector signed char)dstO); \
1497		− register vector signed short srcW REG_v(v26) = \
1498		− (vector signed short)vec_mergel((vector signed char)vzero, (vector signed char)srcO); \
1499		− register vector signed short dstW REG_v(v27) = \
1500		− (vector signed short)vec_mergel((vector signed char)vzero, (vector signed char)dstO); \
	1519	+ srcV = \
	1520	+ (vector signed short)vec_mergeh((vector signed char)vzero, \
	1521	+ (vector signed char)srcO); \
	1522	+ dstV = \
	1523	+ (vector signed short)vec_mergeh((vector signed char)vzero, \
	1524	+ (vector signed char)dstO); \
	1525	+ srcW = \
	1526	+ (vector signed short)vec_mergel((vector signed char)vzero, \
	1527	+ (vector signed char)srcO); \
	1528	+ dstW = \
	1529	+ (vector signed short)vec_mergel((vector signed char)vzero, \
	1530	+ (vector signed char)dstO); \
1501	1531	/* substractions inside the first butterfly */ \
1502		− register vector signed short but0 REG_v(v28) = vec_sub(srcV, dstV); \
1503		− register vector signed short but0S REG_v(v29) = vec_sub(srcW, dstW); \
1504		− register vector signed short op1 REG_v(v30) = vec_perm(but0, but0, perm1); \
1505		− register vector signed short but1 REG_v(v22) = vec_mladd(but0, vprod1, op1); \
1506		− register vector signed short op1S REG_v(v23) = vec_perm(but0S, but0S, perm1); \
1507		− register vector signed short but1S REG_v(v24) = vec_mladd(but0S, vprod1, op1S); \
1508		− register vector signed short op2 REG_v(v25) = vec_perm(but1, but1, perm2); \
1509		− register vector signed short but2 REG_v(v26) = vec_mladd(but1, vprod2, op2); \
1510		− register vector signed short op2S REG_v(v27) = vec_perm(but1S, but1S, perm2); \
1511		− register vector signed short but2S REG_v(v28) = vec_mladd(but1S, vprod2, op2S); \
1512		− register vector signed short op3 REG_v(v29) = vec_perm(but2, but2, perm3); \
	1532	+ but0 = vec_sub(srcV, dstV); \
	1533	+ but0S = vec_sub(srcW, dstW); \
	1534	+ op1 = vec_perm(but0, but0, perm1); \
	1535	+ but1 = vec_mladd(but0, vprod1, op1); \
	1536	+ op1S = vec_perm(but0S, but0S, perm1); \
	1537	+ but1S = vec_mladd(but0S, vprod1, op1S); \
	1538	+ op2 = vec_perm(but1, but1, perm2); \
	1539	+ but2 = vec_mladd(but1, vprod2, op2); \
	1540	+ op2S = vec_perm(but1S, but1S, perm2); \
	1541	+ but2S = vec_mladd(but1S, vprod2, op2S); \
	1542	+ op3 = vec_perm(but2, but2, perm3); \
1513	1543	res1 = vec_mladd(but2, vprod3, op3); \
1514		− register vector signed short op3S REG_v(v30) = vec_perm(but2S, but2S, perm3); \
	1544	+ op3S = vec_perm(but2S, but2S, perm3); \
1515	1545	res2 = vec_mladd(but2S, vprod3, op3S); \
1516	1546	}
1517	1547	ONEITERBUTTERFLY(0, temp0, temp0S);
@@ -1526,6 +1556,12 @@ static int hadamard8_diff16x8_altivec(/MpegEncContext/ void s, uint8_t dst, · 8 unmodified lines
1526	1556	#undef ONEITERBUTTERFLY
1527	1557	{
1528	1558	register vector signed int vsum;
	1559	+ register vector signed short line0S, line1S, line2S, line3S, line4S,
	1560	+ line5S, line6S, line7S, line0BS,line2BS,
	1561	+ line1BS,line3BS,line4BS,line6BS,line5BS,
	1562	+ line7BS,line0CS,line4CS,line1CS,line5CS,
	1563	+ line2CS,line6CS,line3CS,line7CS;
	1564	+
1529	1565	register vector signed short line0 = vec_add(temp0, temp1);
1530	1566	register vector signed short line1 = vec_sub(temp0, temp1);
1531	1567	register vector signed short line2 = vec_add(temp2, temp3);
@@ -1562,32 +1598,32 @@ static int hadamard8_diff16x8_altivec(/MpegEncContext/ void s, uint8_t dst, · 30 unmodified lines
1562	1598	vsum = vec_sum4s(vec_abs(line6C), vsum);
1563	1599	vsum = vec_sum4s(vec_abs(line7C), vsum);
1564	1600
1565		− register vector signed short line0S = vec_add(temp0S, temp1S);
1566		− register vector signed short line1S = vec_sub(temp0S, temp1S);
1567		− register vector signed short line2S = vec_add(temp2S, temp3S);
1568		− register vector signed short line3S = vec_sub(temp2S, temp3S);
1569		− register vector signed short line4S = vec_add(temp4S, temp5S);
1570		− register vector signed short line5S = vec_sub(temp4S, temp5S);
1571		− register vector signed short line6S = vec_add(temp6S, temp7S);
1572		− register vector signed short line7S = vec_sub(temp6S, temp7S);
1573		−
1574		− register vector signed short line0BS = vec_add(line0S, line2S);
1575		− register vector signed short line2BS = vec_sub(line0S, line2S);
1576		− register vector signed short line1BS = vec_add(line1S, line3S);
1577		− register vector signed short line3BS = vec_sub(line1S, line3S);
1578		− register vector signed short line4BS = vec_add(line4S, line6S);
1579		− register vector signed short line6BS = vec_sub(line4S, line6S);
1580		− register vector signed short line5BS = vec_add(line5S, line7S);
1581		− register vector signed short line7BS = vec_sub(line5S, line7S);
1582		−
1583		− register vector signed short line0CS = vec_add(line0BS, line4BS);
1584		− register vector signed short line4CS = vec_sub(line0BS, line4BS);
1585		− register vector signed short line1CS = vec_add(line1BS, line5BS);
1586		− register vector signed short line5CS = vec_sub(line1BS, line5BS);
1587		− register vector signed short line2CS = vec_add(line2BS, line6BS);
1588		− register vector signed short line6CS = vec_sub(line2BS, line6BS);
1589		− register vector signed short line3CS = vec_add(line3BS, line7BS);
1590		− register vector signed short line7CS = vec_sub(line3BS, line7BS);
	1601	+ line0S = vec_add(temp0S, temp1S);
	1602	+ line1S = vec_sub(temp0S, temp1S);
	1603	+ line2S = vec_add(temp2S, temp3S);
	1604	+ line3S = vec_sub(temp2S, temp3S);
	1605	+ line4S = vec_add(temp4S, temp5S);
	1606	+ line5S = vec_sub(temp4S, temp5S);
	1607	+ line6S = vec_add(temp6S, temp7S);
	1608	+ line7S = vec_sub(temp6S, temp7S);
	1609	+
	1610	+ line0BS = vec_add(line0S, line2S);
	1611	+ line2BS = vec_sub(line0S, line2S);
	1612	+ line1BS = vec_add(line1S, line3S);
	1613	+ line3BS = vec_sub(line1S, line3S);
	1614	+ line4BS = vec_add(line4S, line6S);
	1615	+ line6BS = vec_sub(line4S, line6S);
	1616	+ line5BS = vec_add(line5S, line7S);
	1617	+ line7BS = vec_sub(line5S, line7S);
	1618	+
	1619	+ line0CS = vec_add(line0BS, line4BS);
	1620	+ line4CS = vec_sub(line0BS, line4BS);
	1621	+ line1CS = vec_add(line1BS, line5BS);
	1622	+ line5CS = vec_sub(line1BS, line5BS);
	1623	+ line2CS = vec_add(line2BS, line6BS);
	1624	+ line6CS = vec_sub(line2BS, line6BS);
	1625	+ line3CS = vec_add(line3BS, line7BS);
	1626	+ line7CS = vec_sub(line3BS, line7BS);
1591	1627
1592	1628	vsum = vec_sum4s(vec_abs(line0CS), vsum);
1593	1629	vsum = vec_sum4s(vec_abs(line1CS), vsum);

libavcodec/ppc/dsputil_h264_template_altivec.c+296−230

@@ -19,13 +19,13 @@ · 18 unmodified lines
19	19	/* this code assume that stride % 16 == 0 */
20	20	void PREFIX_h264_chroma_mc8_altivec(uint8_t * dst, uint8_t * src, int stride, int h, int x, int y) {
21	21	POWERPC_PERF_DECLARE(PREFIX_h264_chroma_mc8_num, 1);
22		− POWERPC_PERF_START_COUNT(PREFIX_h264_chroma_mc8_num, 1);
23		− signed int ABCD[4] __attribute__((aligned(16)));
	22	+ signed int ABCD[4] __attribute__((aligned(16))) =
	23	+ {((8 - x) * (8 - y)),
	24	+ ((x) * (8 - y)),
	25	+ ((8 - x) * (y)),
	26	+ ((x) * (y))};
24	27	register int i;
25		− ABCD[0] = ((8 - x) * (8 - y));
26		− ABCD[1] = ((x) * (8 - y));
27		− ABCD[2] = ((8 - x) * (y));
28		− ABCD[3] = ((x) * (y));
	28	+ vector unsigned char fperm;
29	29	const vector signed int vABCD = vec_ld(0, ABCD);
30	30	const vector signed short vA = vec_splat((vector signed short)vABCD, 1);
31	31	const vector signed short vB = vec_splat((vector signed short)vABCD, 3);
@@ -34,55 +34,61 @@ void PREFIX_h264_chroma_mc8_altivec(uint8_t * dst, uint8_t * src, int stride, in · 2 unmodified lines
34	34	const vector signed int vzero = vec_splat_s32(0);
35	35	const vector signed short v32ss = vec_sl(vec_splat_s16(1),vec_splat_u16(5));
36	36	const vector unsigned short v6us = vec_splat_u16(6);
	37	+ register int loadSecond = (((unsigned long)src) % 16) <= 7 ? 0 : 1;
	38	+ register int reallyBadAlign = (((unsigned long)src) % 16) == 15 ? 1 : 0;
37	39
38		− vector unsigned char fperm;
	40	+ vector unsigned char vsrcAuc, vsrcBuc, vsrcperm0, vsrcperm1;
	41	+ vector unsigned char vsrc0uc, vsrc1uc;
	42	+ vector signed short vsrc0ssH, vsrc1ssH;
	43	+ vector unsigned char vsrcCuc, vsrc2uc, vsrc3uc;
	44	+ vector signed short vsrc2ssH, vsrc3ssH, psum;
	45	+ vector unsigned char vdst, ppsum, vfdst, fsum;
	46	+
	47	+ POWERPC_PERF_START_COUNT(PREFIX_h264_chroma_mc8_num, 1);
39	48
40	49	if (((unsigned long)dst) % 16 == 0) {
41		− fperm = (vector unsigned char)AVV(0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17,
42		− 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F);
	50	+ fperm = (vector unsigned char)AVV(0x10, 0x11, 0x12, 0x13,
	51	+ 0x14, 0x15, 0x16, 0x17,
	52	+ 0x08, 0x09, 0x0A, 0x0B,
	53	+ 0x0C, 0x0D, 0x0E, 0x0F);
43	54	} else {
44		− fperm = (vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
45		− 0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F);
	55	+ fperm = (vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03,
	56	+ 0x04, 0x05, 0x06, 0x07,
	57	+ 0x18, 0x19, 0x1A, 0x1B,
	58	+ 0x1C, 0x1D, 0x1E, 0x1F);
46	59	}
47	60
48		− register int loadSecond = (((unsigned long)src) % 16) <= 7 ? 0 : 1;
49		− register int reallyBadAlign = (((unsigned long)src) % 16) == 15 ? 1 : 0;
50		−
51		− vector unsigned char vsrcAuc;
52		− vector unsigned char vsrcBuc;
53		− vector unsigned char vsrcperm0;
54		− vector unsigned char vsrcperm1;
55	61	vsrcAuc = vec_ld(0, src);
	62	+
56	63	if (loadSecond)
57	64	vsrcBuc = vec_ld(16, src);
58	65	vsrcperm0 = vec_lvsl(0, src);
59	66	vsrcperm1 = vec_lvsl(1, src);
60	67
61		− vector unsigned char vsrc0uc;
62		− vector unsigned char vsrc1uc;
63	68	vsrc0uc = vec_perm(vsrcAuc, vsrcBuc, vsrcperm0);
64	69	if (reallyBadAlign)
65	70	vsrc1uc = vsrcBuc;
66	71	else
67	72	vsrc1uc = vec_perm(vsrcAuc, vsrcBuc, vsrcperm1);
68	73
69		− vector signed short vsrc0ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, (vector unsigned char)vsrc0uc);
70		− vector signed short vsrc1ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, (vector unsigned char)vsrc1uc);
	74	+ vsrc0ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero,
	75	+ (vector unsigned char)vsrc0uc);
	76	+ vsrc1ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero,
	77	+ (vector unsigned char)vsrc1uc);
71	78
72	79	if (!loadSecond) {// -> !reallyBadAlign
73	80	for (i = 0 ; i < h ; i++) {
74		− vector unsigned char vsrcCuc;
	81	+
	82	+
75	83	vsrcCuc = vec_ld(stride + 0, src);
76	84
77		− vector unsigned char vsrc2uc;
78		− vector unsigned char vsrc3uc;
79	85	vsrc2uc = vec_perm(vsrcCuc, vsrcCuc, vsrcperm0);
80	86	vsrc3uc = vec_perm(vsrcCuc, vsrcCuc, vsrcperm1);
81	87
82		− vector signed short vsrc2ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, (vector unsigned char)vsrc2uc);
83		− vector signed short vsrc3ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, (vector unsigned char)vsrc3uc);
84		−
85		− vector signed short psum;
	88	+ vsrc2ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero,
	89	+ (vector unsigned char)vsrc2uc);
	90	+ vsrc3ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero,
	91	+ (vector unsigned char)vsrc3uc);
86	92
87	93	psum = vec_mladd(vA, vsrc0ssH, vec_splat_s16(0));
88	94	psum = vec_mladd(vB, vsrc1ssH, psum);
@@ -91,11 +97,9 @@ void PREFIX_h264_chroma_mc8_altivec(uint8_t * dst, uint8_t * src, int stride, in · 2 unmodified lines
91	97	psum = vec_add(v32ss, psum);
92	98	psum = vec_sra(psum, v6us);
93	99
94		− vector unsigned char vdst = vec_ld(0, dst);
95		− vector unsigned char ppsum = (vector unsigned char)vec_packsu(psum, psum);
96		−
97		− vector unsigned char vfdst = vec_perm(vdst, ppsum, fperm);
98		− vector unsigned char fsum;
	100	+ vdst = vec_ld(0, dst);
	101	+ ppsum = (vector unsigned char)vec_packsu(psum, psum);
	102	+ vfdst = vec_perm(vdst, ppsum, fperm);
99	103
100	104	OP_U8_ALTIVEC(fsum, vfdst, vdst);
101	105
@@ -108,24 +112,21 @@ void PREFIX_h264_chroma_mc8_altivec(uint8_t * dst, uint8_t * src, int stride, in · 6 unmodified lines
108	112	src += stride;
109	113	}
110	114	} else {
111		− for (i = 0 ; i < h ; i++) {
112		− vector unsigned char vsrcCuc;
113	115	vector unsigned char vsrcDuc;
	116	+ for (i = 0 ; i < h ; i++) {
114	117	vsrcCuc = vec_ld(stride + 0, src);
115	118	vsrcDuc = vec_ld(stride + 16, src);
116	119
117		− vector unsigned char vsrc2uc;
118		− vector unsigned char vsrc3uc;
119	120	vsrc2uc = vec_perm(vsrcCuc, vsrcDuc, vsrcperm0);
120	121	if (reallyBadAlign)
121	122	vsrc3uc = vsrcDuc;
122	123	else
123	124	vsrc3uc = vec_perm(vsrcCuc, vsrcDuc, vsrcperm1);
124	125
125		− vector signed short vsrc2ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, (vector unsigned char)vsrc2uc);
126		− vector signed short vsrc3ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, (vector unsigned char)vsrc3uc);
127		−
128		− vector signed short psum;
	126	+ vsrc2ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero,
	127	+ (vector unsigned char)vsrc2uc);
	128	+ vsrc3ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero,
	129	+ (vector unsigned char)vsrc3uc);
129	130
130	131	psum = vec_mladd(vA, vsrc0ssH, vec_splat_s16(0));
131	132	psum = vec_mladd(vB, vsrc1ssH, psum);
@@ -134,11 +135,9 @@ void PREFIX_h264_chroma_mc8_altivec(uint8_t * dst, uint8_t * src, int stride, in · 2 unmodified lines
134	135	psum = vec_add(v32ss, psum);
135	136	psum = vec_sr(psum, v6us);
136	137
137		− vector unsigned char vdst = vec_ld(0, dst);
138		− vector unsigned char ppsum = (vector unsigned char)vec_pack(psum, psum);
139		−
140		− vector unsigned char vfdst = vec_perm(vdst, ppsum, fperm);
141		− vector unsigned char fsum;
	138	+ vdst = vec_ld(0, dst);
	139	+ ppsum = (vector unsigned char)vec_pack(psum, psum);
	140	+ vfdst = vec_perm(vdst, ppsum, fperm);
142	141
143	142	OP_U8_ALTIVEC(fsum, vfdst, vdst);
144	143
@@ -157,7 +156,6 @@ void PREFIX_h264_chroma_mc8_altivec(uint8_t * dst, uint8_t * src, int stride, in · 12 unmodified lines
157	156	/* this code assume stride % 16 == 0 */
158	157	static void PREFIX_h264_qpel16_h_lowpass_altivec(uint8_t * dst, uint8_t * src, int dstStride, int srcStride) {
159	158	POWERPC_PERF_DECLARE(PREFIX_h264_qpel16_h_lowpass_num, 1);
160		− POWERPC_PERF_START_COUNT(PREFIX_h264_qpel16_h_lowpass_num, 1);
161	159	register int i;
162	160
163	161	const vector signed int vzero = vec_splat_s32(0);
@@ -172,13 +170,30 @@ static void PREFIX_h264_qpel16_h_lowpass_altivec(uint8_t * dst, uint8_t * src, i · 8 unmodified lines
172	170	const vector signed short v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
173	171	const vector signed short v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4));
174	172	const vector unsigned char dstperm = vec_lvsr(0, dst);
175		− const vector unsigned char neg1 = (const vector unsigned char)vec_splat_s8(-1);
176		− const vector unsigned char dstmask = vec_perm((const vector unsigned char)vzero, neg1, dstperm);
	173	+ const vector unsigned char neg1 =
	174	+ (const vector unsigned char) vec_splat_s8(-1);
	175	+
	176	+ const vector unsigned char dstmask =
	177	+ vec_perm((const vector unsigned char)vzero,
	178	+ neg1, dstperm);
	179	+
	180	+ vector unsigned char srcM2, srcM1, srcP0, srcP1, srcP2, srcP3;
177	181
178	182	register int align = ((((unsigned long)src) - 2) % 16);
179	183
	184	+ vector signed short srcP0A, srcP0B, srcP1A, srcP1B,
	185	+ srcP2A, srcP2B, srcP3A, srcP3B,
	186	+ srcM1A, srcM1B, srcM2A, srcM2B,
	187	+ sum1A, sum1B, sum2A, sum2B, sum3A, sum3B,
	188	+ pp1A, pp1B, pp2A, pp2B, pp3A, pp3B,
	189	+ psumA, psumB, sumA, sumB;
	190	+
	191	+ vector unsigned char sum, dst1, dst2, vdst, fsum,
	192	+ rsum, fdst1, fdst2;
	193	+
	194	+ POWERPC_PERF_START_COUNT(PREFIX_h264_qpel16_h_lowpass_num, 1);
	195	+
180	196	for (i = 0 ; i < 16 ; i ++) {
181		− vector unsigned char srcM2, srcM1, srcP0, srcP1, srcP2, srcP3;
182	197	vector unsigned char srcR1 = vec_ld(-2, src);
183	198	vector unsigned char srcR2 = vec_ld(14, src);
184	199
@@ -237,55 +252,54 @@ static void PREFIX_h264_qpel16_h_lowpass_altivec(uint8_t * dst, uint8_t * src, i · 52 unmodified lines
237	252	} break;
238	253	}
239	254
240		− const vector signed short srcP0A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP0);
241		− const vector signed short srcP0B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP0);
242		− const vector signed short srcP1A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP1);
243		− const vector signed short srcP1B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP1);
	255	+ srcP0A = vec_mergeh((vector unsigned char)vzero, srcP0);
	256	+ srcP0B = vec_mergel((vector unsigned char)vzero, srcP0);
	257	+ srcP1A = vec_mergeh((vector unsigned char)vzero, srcP1);
	258	+ srcP1B = vec_mergel((vector unsigned char)vzero, srcP1);
244	259
245		− const vector signed short srcP2A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP2);
246		− const vector signed short srcP2B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP2);
247		− const vector signed short srcP3A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP3);
248		− const vector signed short srcP3B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP3);
	260	+ srcP2A = vec_mergeh((vector unsigned char)vzero, srcP2);
	261	+ srcP2B = vec_mergel((vector unsigned char)vzero, srcP2);
	262	+ srcP3A = vec_mergeh((vector unsigned char)vzero, srcP3);
	263	+ srcP3B = vec_mergel((vector unsigned char)vzero, srcP3);
249	264
250		− const vector signed short srcM1A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcM1);
251		− const vector signed short srcM1B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcM1);
252		− const vector signed short srcM2A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcM2);
253		− const vector signed short srcM2B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcM2);
	265	+ srcM1A = vec_mergeh((vector unsigned char)vzero, srcM1);
	266	+ srcM1B = vec_mergel((vector unsigned char)vzero, srcM1);
	267	+ srcM2A = vec_mergeh((vector unsigned char)vzero, srcM2);
	268	+ srcM2B = vec_mergel((vector unsigned char)vzero, srcM2);
254	269
255		− const vector signed short sum1A = vec_adds(srcP0A, srcP1A);
256		− const vector signed short sum1B = vec_adds(srcP0B, srcP1B);
257		− const vector signed short sum2A = vec_adds(srcM1A, srcP2A);
258		− const vector signed short sum2B = vec_adds(srcM1B, srcP2B);
259		− const vector signed short sum3A = vec_adds(srcM2A, srcP3A);
260		− const vector signed short sum3B = vec_adds(srcM2B, srcP3B);
	270	+ sum1A = vec_adds(srcP0A, srcP1A);
	271	+ sum1B = vec_adds(srcP0B, srcP1B);
	272	+ sum2A = vec_adds(srcM1A, srcP2A);
	273	+ sum2B = vec_adds(srcM1B, srcP2B);
	274	+ sum3A = vec_adds(srcM2A, srcP3A);
	275	+ sum3B = vec_adds(srcM2B, srcP3B);
261	276
262		− const vector signed short pp1A = vec_mladd(sum1A, v20ss, v16ss);
263		− const vector signed short pp1B = vec_mladd(sum1B, v20ss, v16ss);
	277	+ pp1A = vec_mladd(sum1A, v20ss, v16ss);
	278	+ pp1B = vec_mladd(sum1B, v20ss, v16ss);
264	279
265		− const vector signed short pp2A = vec_mladd(sum2A, v5ss, (vector signed short)vzero);
266		− const vector signed short pp2B = vec_mladd(sum2B, v5ss, (vector signed short)vzero);
	280	+ pp2A = vec_mladd(sum2A, v5ss, (vector signed short)vzero);
	281	+ pp2B = vec_mladd(sum2B, v5ss, (vector signed short)vzero);
267	282
268		− const vector signed short pp3A = vec_add(sum3A, pp1A);
269		− const vector signed short pp3B = vec_add(sum3B, pp1B);
	283	+ pp3A = vec_add(sum3A, pp1A);
	284	+ pp3B = vec_add(sum3B, pp1B);
270	285
271		− const vector signed short psumA = vec_sub(pp3A, pp2A);
272		− const vector signed short psumB = vec_sub(pp3B, pp2B);
	286	+ psumA = vec_sub(pp3A, pp2A);
	287	+ psumB = vec_sub(pp3B, pp2B);
273	288
274		− const vector signed short sumA = vec_sra(psumA, v5us);
275		− const vector signed short sumB = vec_sra(psumB, v5us);
	289	+ sumA = vec_sra(psumA, v5us);
	290	+ sumB = vec_sra(psumB, v5us);
276	291
277		− const vector unsigned char sum = vec_packsu(sumA, sumB);
	292	+ sum = vec_packsu(sumA, sumB);
278	293
279		− const vector unsigned char dst1 = vec_ld(0, dst);
280		− const vector unsigned char dst2 = vec_ld(16, dst);
281		− const vector unsigned char vdst = vec_perm(dst1, dst2, vec_lvsl(0, dst));
	294	+ dst1 = vec_ld(0, dst);
	295	+ dst2 = vec_ld(16, dst);
	296	+ vdst = vec_perm(dst1, dst2, vec_lvsl(0, dst));
282	297
283		− vector unsigned char fsum;
284	298	OP_U8_ALTIVEC(fsum, sum, vdst);
285	299
286		− const vector unsigned char rsum = vec_perm(fsum, fsum, dstperm);
287		− const vector unsigned char fdst1 = vec_sel(dst1, rsum, dstmask);
288		− const vector unsigned char fdst2 = vec_sel(rsum, dst2, dstmask);
	300	+ rsum = vec_perm(fsum, fsum, dstperm);
	301	+ fdst1 = vec_sel(dst1, rsum, dstmask);
	302	+ fdst2 = vec_sel(rsum, dst2, dstmask);
289	303
290	304	vec_st(fdst1, 0, dst);
291	305	vec_st(fdst2, 16, dst);
@@ -299,7 +313,6 @@ POWERPC_PERF_STOP_COUNT(PREFIX_h264_qpel16_h_lowpass_num, 1); · 7 unmodified lines
299	313	/* this code assume stride % 16 == 0 */
300	314	static void PREFIX_h264_qpel16_v_lowpass_altivec(uint8_t * dst, uint8_t * src, int dstStride, int srcStride) {
301	315	POWERPC_PERF_DECLARE(PREFIX_h264_qpel16_v_lowpass_num, 1);
302		− POWERPC_PERF_START_COUNT(PREFIX_h264_qpel16_v_lowpass_num, 1);
303	316
304	317	register int i;
305	318
@@ -318,49 +331,71 @@ static void PREFIX_h264_qpel16_v_lowpass_altivec(uint8_t * dst, uint8_t * src, i · 12 unmodified lines
318	331	const vector unsigned char srcM2a = vec_ld(0, srcbis);
319	332	const vector unsigned char srcM2b = vec_ld(16, srcbis);
320	333	const vector unsigned char srcM2 = vec_perm(srcM2a, srcM2b, perm);
321		− srcbis += srcStride;
322		− const vector unsigned char srcM1a = vec_ld(0, srcbis);
	334	+// srcbis += srcStride;
	335	+ const vector unsigned char srcM1a = vec_ld(0, srcbis += srcStride);
323	336	const vector unsigned char srcM1b = vec_ld(16, srcbis);
324	337	const vector unsigned char srcM1 = vec_perm(srcM1a, srcM1b, perm);
325		− srcbis += srcStride;
326		− const vector unsigned char srcP0a = vec_ld(0, srcbis);
	338	+// srcbis += srcStride;
	339	+ const vector unsigned char srcP0a = vec_ld(0, srcbis += srcStride);
327	340	const vector unsigned char srcP0b = vec_ld(16, srcbis);
328	341	const vector unsigned char srcP0 = vec_perm(srcP0a, srcP0b, perm);
329		− srcbis += srcStride;
330		− const vector unsigned char srcP1a = vec_ld(0, srcbis);
	342	+// srcbis += srcStride;
	343	+ const vector unsigned char srcP1a = vec_ld(0, srcbis += srcStride);
331	344	const vector unsigned char srcP1b = vec_ld(16, srcbis);
332	345	const vector unsigned char srcP1 = vec_perm(srcP1a, srcP1b, perm);
333		− srcbis += srcStride;
334		− const vector unsigned char srcP2a = vec_ld(0, srcbis);
	346	+// srcbis += srcStride;
	347	+ const vector unsigned char srcP2a = vec_ld(0, srcbis += srcStride);
335	348	const vector unsigned char srcP2b = vec_ld(16, srcbis);
336	349	const vector unsigned char srcP2 = vec_perm(srcP2a, srcP2b, perm);
337		− srcbis += srcStride;
338		−
339		− vector signed short srcM2ssA = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcM2);
340		− vector signed short srcM2ssB = (vector signed short)vec_mergel((vector unsigned char)vzero, srcM2);
341		− vector signed short srcM1ssA = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcM1);
342		− vector signed short srcM1ssB = (vector signed short)vec_mergel((vector unsigned char)vzero, srcM1);
343		− vector signed short srcP0ssA = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP0);
344		− vector signed short srcP0ssB = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP0);
345		− vector signed short srcP1ssA = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP1);
346		− vector signed short srcP1ssB = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP1);
347		− vector signed short srcP2ssA = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP2);
348		− vector signed short srcP2ssB = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP2);
	350	+// srcbis += srcStride;
	351	+
	352	+ vector signed short srcM2ssA = (vector signed short)
	353	+ vec_mergeh((vector unsigned char)vzero, srcM2);
	354	+ vector signed short srcM2ssB = (vector signed short)
	355	+ vec_mergel((vector unsigned char)vzero, srcM2);
	356	+ vector signed short srcM1ssA = (vector signed short)
	357	+ vec_mergeh((vector unsigned char)vzero, srcM1);
	358	+ vector signed short srcM1ssB = (vector signed short)
	359	+ vec_mergel((vector unsigned char)vzero, srcM1);
	360	+ vector signed short srcP0ssA = (vector signed short)
	361	+ vec_mergeh((vector unsigned char)vzero, srcP0);
	362	+ vector signed short srcP0ssB = (vector signed short)
	363	+ vec_mergel((vector unsigned char)vzero, srcP0);
	364	+ vector signed short srcP1ssA = (vector signed short)
	365	+ vec_mergeh((vector unsigned char)vzero, srcP1);
	366	+ vector signed short srcP1ssB = (vector signed short)
	367	+ vec_mergel((vector unsigned char)vzero, srcP1);
	368	+ vector signed short srcP2ssA = (vector signed short)
	369	+ vec_mergeh((vector unsigned char)vzero, srcP2);
	370	+ vector signed short srcP2ssB = (vector signed short)
	371	+ vec_mergel((vector unsigned char)vzero, srcP2);
	372	+
	373	+ vector signed short pp1A, pp1B, pp2A, pp2B, pp3A, pp3B,
	374	+ psumA, psumB, sumA, sumB,
	375	+ srcP3ssA, srcP3ssB,
	376	+ sum1A, sum1B, sum2A, sum2B, sum3A, sum3B;
	377	+
	378	+ vector unsigned char sum, dst1, dst2, vdst, fsum, rsum, fdst1, fdst2,
	379	+ srcP3a, srcP3b, srcP3;
	380	+
	381	+ POWERPC_PERF_START_COUNT(PREFIX_h264_qpel16_v_lowpass_num, 1);
349	382
350	383	for (i = 0 ; i < 16 ; i++) {
351		− const vector unsigned char srcP3a = vec_ld(0, srcbis);
352		− const vector unsigned char srcP3b = vec_ld(16, srcbis);
353		− const vector unsigned char srcP3 = vec_perm(srcP3a, srcP3b, perm);
354		− const vector signed short srcP3ssA = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP3);
355		− const vector signed short srcP3ssB = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP3);
356		− srcbis += srcStride;
357		−
358		− const vector signed short sum1A = vec_adds(srcP0ssA, srcP1ssA);
359		− const vector signed short sum1B = vec_adds(srcP0ssB, srcP1ssB);
360		− const vector signed short sum2A = vec_adds(srcM1ssA, srcP2ssA);
361		− const vector signed short sum2B = vec_adds(srcM1ssB, srcP2ssB);
362		− const vector signed short sum3A = vec_adds(srcM2ssA, srcP3ssA);
363		− const vector signed short sum3B = vec_adds(srcM2ssB, srcP3ssB);
	384	+ srcP3a = vec_ld(0, srcbis += srcStride);
	385	+ srcP3b = vec_ld(16, srcbis);
	386	+ srcP3 = vec_perm(srcP3a, srcP3b, perm);
	387	+ srcP3ssA = (vector signed short)
	388	+ vec_mergeh((vector unsigned char)vzero, srcP3);
	389	+ srcP3ssB = (vector signed short)
	390	+ vec_mergel((vector unsigned char)vzero, srcP3);
	391	+// srcbis += srcStride;
	392	+
	393	+ sum1A = vec_adds(srcP0ssA, srcP1ssA);
	394	+ sum1B = vec_adds(srcP0ssB, srcP1ssB);
	395	+ sum2A = vec_adds(srcM1ssA, srcP2ssA);
	396	+ sum2B = vec_adds(srcM1ssB, srcP2ssB);
	397	+ sum3A = vec_adds(srcM2ssA, srcP3ssA);
	398	+ sum3B = vec_adds(srcM2ssB, srcP3ssB);
364	399
365	400	srcM2ssA = srcM1ssA;
366	401	srcM2ssB = srcM1ssB;
@@ -373,33 +408,32 @@ static void PREFIX_h264_qpel16_v_lowpass_altivec(uint8_t * dst, uint8_t * src, i · 6 unmodified lines
373	408	srcP2ssA = srcP3ssA;
374	409	srcP2ssB = srcP3ssB;
375	410
376		− const vector signed short pp1A = vec_mladd(sum1A, v20ss, v16ss);
377		− const vector signed short pp1B = vec_mladd(sum1B, v20ss, v16ss);
	411	+ pp1A = vec_mladd(sum1A, v20ss, v16ss);
	412	+ pp1B = vec_mladd(sum1B, v20ss, v16ss);
378	413
379		− const vector signed short pp2A = vec_mladd(sum2A, v5ss, (vector signed short)vzero);
380		− const vector signed short pp2B = vec_mladd(sum2B, v5ss, (vector signed short)vzero);
	414	+ pp2A = vec_mladd(sum2A, v5ss, (vector signed short)vzero);
	415	+ pp2B = vec_mladd(sum2B, v5ss, (vector signed short)vzero);
381	416
382		− const vector signed short pp3A = vec_add(sum3A, pp1A);
383		− const vector signed short pp3B = vec_add(sum3B, pp1B);
	417	+ pp3A = vec_add(sum3A, pp1A);
	418	+ pp3B = vec_add(sum3B, pp1B);
384	419
385		− const vector signed short psumA = vec_sub(pp3A, pp2A);
386		− const vector signed short psumB = vec_sub(pp3B, pp2B);
	420	+ psumA = vec_sub(pp3A, pp2A);
	421	+ psumB = vec_sub(pp3B, pp2B);
387	422
388		− const vector signed short sumA = vec_sra(psumA, v5us);
389		− const vector signed short sumB = vec_sra(psumB, v5us);
	423	+ sumA = vec_sra(psumA, v5us);
	424	+ sumB = vec_sra(psumB, v5us);
390	425
391		− const vector unsigned char sum = vec_packsu(sumA, sumB);
	426	+ sum = vec_packsu(sumA, sumB);
392	427
393		− const vector unsigned char dst1 = vec_ld(0, dst);
394		− const vector unsigned char dst2 = vec_ld(16, dst);
395		− const vector unsigned char vdst = vec_perm(dst1, dst2, vec_lvsl(0, dst));
	428	+ dst1 = vec_ld(0, dst);
	429	+ dst2 = vec_ld(16, dst);
	430	+ vdst = vec_perm(dst1, dst2, vec_lvsl(0, dst));
396	431
397		− vector unsigned char fsum;
398	432	OP_U8_ALTIVEC(fsum, sum, vdst);
399	433
400		− const vector unsigned char rsum = vec_perm(fsum, fsum, dstperm);
401		− const vector unsigned char fdst1 = vec_sel(dst1, rsum, dstmask);
402		− const vector unsigned char fdst2 = vec_sel(rsum, dst2, dstmask);
	434	+ rsum = vec_perm(fsum, fsum, dstperm);
	435	+ fdst1 = vec_sel(dst1, rsum, dstmask);
	436	+ fdst2 = vec_sel(rsum, dst2, dstmask);
403	437
404	438	vec_st(fdst1, 0, dst);
405	439	vec_st(fdst2, 16, dst);
@@ -412,7 +446,6 @@ static void PREFIX_h264_qpel16_v_lowpass_altivec(uint8_t * dst, uint8_t * src, i · 6 unmodified lines
412	446	/* this code assume stride % 16 == 0 and tmp is properly aligned */
413	447	static void PREFIX_h264_qpel16_hv_lowpass_altivec(uint8_t * dst, int16_t * tmp, uint8_t * src, int dstStride, int tmpStride, int srcStride) {
414	448	POWERPC_PERF_DECLARE(PREFIX_h264_qpel16_hv_lowpass_num, 1);
415		− POWERPC_PERF_START_COUNT(PREFIX_h264_qpel16_hv_lowpass_num, 1);
416	449	register int i;
417	450	const vector signed int vzero = vec_splat_s32(0);
418	451	const vector unsigned char permM2 = vec_lvsl(-2, src);
@@ -430,8 +463,38 @@ static void PREFIX_h264_qpel16_hv_lowpass_altivec(uint8_t * dst, int16_t * tmp, · 11 unmodified lines
430	463
431	464	register int align = ((((unsigned long)src) - 2) % 16);
432	465
433		− src -= (2 * srcStride);
	466	+ const vector unsigned char neg1 = (const vector unsigned char)
	467	+ vec_splat_s8(-1);
	468	+
	469	+ vector signed short srcP0A, srcP0B, srcP1A, srcP1B,
	470	+ srcP2A, srcP2B, srcP3A, srcP3B,
	471	+ srcM1A, srcM1B, srcM2A, srcM2B,
	472	+ sum1A, sum1B, sum2A, sum2B, sum3A, sum3B,
	473	+ pp1A, pp1B, pp2A, pp2B, psumA, psumB;
	474	+
	475	+ const vector unsigned char dstperm = vec_lvsr(0, dst);
	476	+
	477	+ const vector unsigned char dstmask = vec_perm((const vector unsigned char)vzero, neg1, dstperm);
	478	+
	479	+ const vector unsigned char mperm = (const vector unsigned char)
	480	+ AVV(0x00, 0x08, 0x01, 0x09, 0x02, 0x0A, 0x03, 0x0B,
	481	+ 0x04, 0x0C, 0x05, 0x0D, 0x06, 0x0E, 0x07, 0x0F);
	482	+ int16_t *tmpbis = tmp;
	483	+
	484	+ vector signed short tmpM1ssA, tmpM1ssB, tmpM2ssA, tmpM2ssB,
	485	+ tmpP0ssA, tmpP0ssB, tmpP1ssA, tmpP1ssB,
	486	+ tmpP2ssA, tmpP2ssB;
434	487
	488	+ vector signed int pp1Ae, pp1Ao, pp1Be, pp1Bo, pp2Ae, pp2Ao, pp2Be, pp2Bo,
	489	+ pp3Ae, pp3Ao, pp3Be, pp3Bo, pp1cAe, pp1cAo, pp1cBe, pp1cBo,
	490	+ pp32Ae, pp32Ao, pp32Be, pp32Bo, sumAe, sumAo, sumBe, sumBo,
	491	+ ssumAe, ssumAo, ssumBe, ssumBo;
	492	+ vector unsigned char fsum, sumv, sum, dst1, dst2, vdst,
	493	+ rsum, fdst1, fdst2;
	494	+ vector signed short ssume, ssumo;
	495	+
	496	+ POWERPC_PERF_START_COUNT(PREFIX_h264_qpel16_hv_lowpass_num, 1);
	497	+ src -= (2 * srcStride);
435	498	for (i = 0 ; i < 21 ; i ++) {
436	499	vector unsigned char srcM2, srcM1, srcP0, srcP1, srcP2, srcP3;
437	500	vector unsigned char srcR1 = vec_ld(-2, src);
@@ -492,36 +555,48 @@ static void PREFIX_h264_qpel16_hv_lowpass_altivec(uint8_t * dst, int16_t * tmp, · 54 unmodified lines
492	555	} break;
493	556	}
494	557
495		− const vector signed short srcP0A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP0);
496		− const vector signed short srcP0B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP0);
497		− const vector signed short srcP1A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP1);
498		− const vector signed short srcP1B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP1);
499		−
500		− const vector signed short srcP2A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP2);
501		− const vector signed short srcP2B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP2);
502		− const vector signed short srcP3A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP3);
503		− const vector signed short srcP3B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP3);
504		−
505		− const vector signed short srcM1A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcM1);
506		− const vector signed short srcM1B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcM1);
507		− const vector signed short srcM2A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcM2);
508		− const vector signed short srcM2B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcM2);
509		−
510		− const vector signed short sum1A = vec_adds(srcP0A, srcP1A);
511		− const vector signed short sum1B = vec_adds(srcP0B, srcP1B);
512		− const vector signed short sum2A = vec_adds(srcM1A, srcP2A);
513		− const vector signed short sum2B = vec_adds(srcM1B, srcP2B);
514		− const vector signed short sum3A = vec_adds(srcM2A, srcP3A);
515		− const vector signed short sum3B = vec_adds(srcM2B, srcP3B);
516		−
517		− const vector signed short pp1A = vec_mladd(sum1A, v20ss, sum3A);
518		− const vector signed short pp1B = vec_mladd(sum1B, v20ss, sum3B);
519		−
520		− const vector signed short pp2A = vec_mladd(sum2A, v5ss, (vector signed short)vzero);
521		− const vector signed short pp2B = vec_mladd(sum2B, v5ss, (vector signed short)vzero);
522		−
523		− const vector signed short psumA = vec_sub(pp1A, pp2A);
524		− const vector signed short psumB = vec_sub(pp1B, pp2B);
	558	+ srcP0A = (vector signed short)
	559	+ vec_mergeh((vector unsigned char)vzero, srcP0);
	560	+ srcP0B = (vector signed short)
	561	+ vec_mergel((vector unsigned char)vzero, srcP0);
	562	+ srcP1A = (vector signed short)
	563	+ vec_mergeh((vector unsigned char)vzero, srcP1);
	564	+ srcP1B = (vector signed short)
	565	+ vec_mergel((vector unsigned char)vzero, srcP1);
	566	+
	567	+ srcP2A = (vector signed short)
	568	+ vec_mergeh((vector unsigned char)vzero, srcP2);
	569	+ srcP2B = (vector signed short)
	570	+ vec_mergel((vector unsigned char)vzero, srcP2);
	571	+ srcP3A = (vector signed short)
	572	+ vec_mergeh((vector unsigned char)vzero, srcP3);
	573	+ srcP3B = (vector signed short)
	574	+ vec_mergel((vector unsigned char)vzero, srcP3);
	575	+
	576	+ srcM1A = (vector signed short)
	577	+ vec_mergeh((vector unsigned char)vzero, srcM1);
	578	+ srcM1B = (vector signed short)
	579	+ vec_mergel((vector unsigned char)vzero, srcM1);
	580	+ srcM2A = (vector signed short)
	581	+ vec_mergeh((vector unsigned char)vzero, srcM2);
	582	+ srcM2B = (vector signed short)
	583	+ vec_mergel((vector unsigned char)vzero, srcM2);
	584	+
	585	+ sum1A = vec_adds(srcP0A, srcP1A);
	586	+ sum1B = vec_adds(srcP0B, srcP1B);
	587	+ sum2A = vec_adds(srcM1A, srcP2A);
	588	+ sum2B = vec_adds(srcM1B, srcP2B);
	589	+ sum3A = vec_adds(srcM2A, srcP3A);
	590	+ sum3B = vec_adds(srcM2B, srcP3B);
	591	+
	592	+ pp1A = vec_mladd(sum1A, v20ss, sum3A);
	593	+ pp1B = vec_mladd(sum1B, v20ss, sum3B);
	594	+
	595	+ pp2A = vec_mladd(sum2A, v5ss, (vector signed short)vzero);
	596	+ pp2B = vec_mladd(sum2B, v5ss, (vector signed short)vzero);
	597	+
	598	+ psumA = vec_sub(pp1A, pp2A);
	599	+ psumB = vec_sub(pp1B, pp2B);
525	600
526	601	vec_st(psumA, 0, tmp);
527	602	vec_st(psumB, 16, tmp);
@@ -530,35 +605,25 @@ static void PREFIX_h264_qpel16_hv_lowpass_altivec(uint8_t * dst, int16_t * tmp, · 2 unmodified lines
530	605	tmp += tmpStride; /* int16_t, and stride is 16, so it's OK here /
531	606	}
532	607
533		− const vector unsigned char dstperm = vec_lvsr(0, dst);
534		− const vector unsigned char neg1 = (const vector unsigned char)vec_splat_s8(-1);
535		− const vector unsigned char dstmask = vec_perm((const vector unsigned char)vzero, neg1, dstperm);
536		− const vector unsigned char mperm = (const vector unsigned char)
537		− AVV(0x00, 0x08, 0x01, 0x09, 0x02, 0x0A, 0x03, 0x0B,
538		− 0x04, 0x0C, 0x05, 0x0D, 0x06, 0x0E, 0x07, 0x0F);
539		−
540		− int16_t tmpbis = tmp - (tmpStride 21);
541		−
542		− vector signed short tmpM2ssA = vec_ld(0, tmpbis);
543		− vector signed short tmpM2ssB = vec_ld(16, tmpbis);
	608	+ tmpM2ssA = vec_ld(0, tmpbis);
	609	+ tmpM2ssB = vec_ld(16, tmpbis);
544	610	tmpbis += tmpStride;
545		− vector signed short tmpM1ssA = vec_ld(0, tmpbis);
546		− vector signed short tmpM1ssB = vec_ld(16, tmpbis);
	611	+ tmpM1ssA = vec_ld(0, tmpbis);
	612	+ tmpM1ssB = vec_ld(16, tmpbis);
547	613	tmpbis += tmpStride;
548		− vector signed short tmpP0ssA = vec_ld(0, tmpbis);
549		− vector signed short tmpP0ssB = vec_ld(16, tmpbis);
	614	+ tmpP0ssA = vec_ld(0, tmpbis);
	615	+ tmpP0ssB = vec_ld(16, tmpbis);
550	616	tmpbis += tmpStride;
551		− vector signed short tmpP1ssA = vec_ld(0, tmpbis);
552		− vector signed short tmpP1ssB = vec_ld(16, tmpbis);
	617	+ tmpP1ssA = vec_ld(0, tmpbis);
	618	+ tmpP1ssB = vec_ld(16, tmpbis);
553	619	tmpbis += tmpStride;
554		− vector signed short tmpP2ssA = vec_ld(0, tmpbis);
555		− vector signed short tmpP2ssB = vec_ld(16, tmpbis);
	620	+ tmpP2ssA = vec_ld(0, tmpbis);
	621	+ tmpP2ssB = vec_ld(16, tmpbis);
556	622	tmpbis += tmpStride;
557	623
558	624	for (i = 0 ; i < 16 ; i++) {
559	625	const vector signed short tmpP3ssA = vec_ld(0, tmpbis);
560	626	const vector signed short tmpP3ssB = vec_ld(16, tmpbis);
561		− tmpbis += tmpStride;
562	627
563	628	const vector signed short sum1A = vec_adds(tmpP0ssA, tmpP1ssA);
564	629	const vector signed short sum1B = vec_adds(tmpP0ssB, tmpP1ssB);
@@ -567,6 +632,8 @@ static void PREFIX_h264_qpel16_hv_lowpass_altivec(uint8_t * dst, int16_t * tmp, · 2 unmodified lines
567	632	const vector signed short sum3A = vec_adds(tmpM2ssA, tmpP3ssA);
568	633	const vector signed short sum3B = vec_adds(tmpM2ssB, tmpP3ssB);
569	634
	635	+ tmpbis += tmpStride;
	636	+
570	637	tmpM2ssA = tmpM1ssA;
571	638	tmpM2ssB = tmpM1ssB;
572	639	tmpM1ssA = tmpP0ssA;
@@ -578,57 +645,56 @@ static void PREFIX_h264_qpel16_hv_lowpass_altivec(uint8_t * dst, int16_t * tmp, · 5 unmodified lines
578	645	tmpP2ssA = tmpP3ssA;
579	646	tmpP2ssB = tmpP3ssB;
580	647
581		− const vector signed int pp1Ae = vec_mule(sum1A, v20ss);
582		− const vector signed int pp1Ao = vec_mulo(sum1A, v20ss);
583		− const vector signed int pp1Be = vec_mule(sum1B, v20ss);
584		− const vector signed int pp1Bo = vec_mulo(sum1B, v20ss);
	648	+ pp1Ae = vec_mule(sum1A, v20ss);
	649	+ pp1Ao = vec_mulo(sum1A, v20ss);
	650	+ pp1Be = vec_mule(sum1B, v20ss);
	651	+ pp1Bo = vec_mulo(sum1B, v20ss);
585	652
586		− const vector signed int pp2Ae = vec_mule(sum2A, v5ss);
587		− const vector signed int pp2Ao = vec_mulo(sum2A, v5ss);
588		− const vector signed int pp2Be = vec_mule(sum2B, v5ss);
589		− const vector signed int pp2Bo = vec_mulo(sum2B, v5ss);
	653	+ pp2Ae = vec_mule(sum2A, v5ss);
	654	+ pp2Ao = vec_mulo(sum2A, v5ss);
	655	+ pp2Be = vec_mule(sum2B, v5ss);
	656	+ pp2Bo = vec_mulo(sum2B, v5ss);
590	657
591		− const vector signed int pp3Ae = vec_sra((vector signed int)sum3A, v16ui);
592		− const vector signed int pp3Ao = vec_mulo(sum3A, v1ss);
593		− const vector signed int pp3Be = vec_sra((vector signed int)sum3B, v16ui);
594		− const vector signed int pp3Bo = vec_mulo(sum3B, v1ss);
	658	+ pp3Ae = vec_sra((vector signed int)sum3A, v16ui);
	659	+ pp3Ao = vec_mulo(sum3A, v1ss);
	660	+ pp3Be = vec_sra((vector signed int)sum3B, v16ui);
	661	+ pp3Bo = vec_mulo(sum3B, v1ss);
595	662
596		− const vector signed int pp1cAe = vec_add(pp1Ae, v512si);
597		− const vector signed int pp1cAo = vec_add(pp1Ao, v512si);
598		− const vector signed int pp1cBe = vec_add(pp1Be, v512si);
599		− const vector signed int pp1cBo = vec_add(pp1Bo, v512si);
	663	+ pp1cAe = vec_add(pp1Ae, v512si);
	664	+ pp1cAo = vec_add(pp1Ao, v512si);
	665	+ pp1cBe = vec_add(pp1Be, v512si);
	666	+ pp1cBo = vec_add(pp1Bo, v512si);
600	667
601		− const vector signed int pp32Ae = vec_sub(pp3Ae, pp2Ae);
602		− const vector signed int pp32Ao = vec_sub(pp3Ao, pp2Ao);
603		− const vector signed int pp32Be = vec_sub(pp3Be, pp2Be);
604		− const vector signed int pp32Bo = vec_sub(pp3Bo, pp2Bo);
	668	+ pp32Ae = vec_sub(pp3Ae, pp2Ae);
	669	+ pp32Ao = vec_sub(pp3Ao, pp2Ao);
	670	+ pp32Be = vec_sub(pp3Be, pp2Be);
	671	+ pp32Bo = vec_sub(pp3Bo, pp2Bo);
605	672
606		− const vector signed int sumAe = vec_add(pp1cAe, pp32Ae);
607		− const vector signed int sumAo = vec_add(pp1cAo, pp32Ao);
608		− const vector signed int sumBe = vec_add(pp1cBe, pp32Be);
609		− const vector signed int sumBo = vec_add(pp1cBo, pp32Bo);
	673	+ sumAe = vec_add(pp1cAe, pp32Ae);
	674	+ sumAo = vec_add(pp1cAo, pp32Ao);
	675	+ sumBe = vec_add(pp1cBe, pp32Be);
	676	+ sumBo = vec_add(pp1cBo, pp32Bo);
610	677
611		− const vector signed int ssumAe = vec_sra(sumAe, v10ui);
612		− const vector signed int ssumAo = vec_sra(sumAo, v10ui);
613		− const vector signed int ssumBe = vec_sra(sumBe, v10ui);
614		− const vector signed int ssumBo = vec_sra(sumBo, v10ui);
	678	+ ssumAe = vec_sra(sumAe, v10ui);
	679	+ ssumAo = vec_sra(sumAo, v10ui);
	680	+ ssumBe = vec_sra(sumBe, v10ui);
	681	+ ssumBo = vec_sra(sumBo, v10ui);
615	682
616		− const vector signed short ssume = vec_packs(ssumAe, ssumBe);
617		− const vector signed short ssumo = vec_packs(ssumAo, ssumBo);
	683	+ ssume = vec_packs(ssumAe, ssumBe);
	684	+ ssumo = vec_packs(ssumAo, ssumBo);
618	685
619		− const vector unsigned char sumv = vec_packsu(ssume, ssumo);
620		− const vector unsigned char sum = vec_perm(sumv, sumv, mperm);
	686	+ sumv = vec_packsu(ssume, ssumo);
	687	+ sum = vec_perm(sumv, sumv, mperm);
621	688
622		− const vector unsigned char dst1 = vec_ld(0, dst);
623		− const vector unsigned char dst2 = vec_ld(16, dst);
624		− const vector unsigned char vdst = vec_perm(dst1, dst2, vec_lvsl(0, dst));
	689	+ dst1 = vec_ld(0, dst);
	690	+ dst2 = vec_ld(16, dst);
	691	+ vdst = vec_perm(dst1, dst2, vec_lvsl(0, dst));
625	692
626		− vector unsigned char fsum;
627	693	OP_U8_ALTIVEC(fsum, sum, vdst);
628	694
629		− const vector unsigned char rsum = vec_perm(fsum, fsum, dstperm);
630		− const vector unsigned char fdst1 = vec_sel(dst1, rsum, dstmask);
631		− const vector unsigned char fdst2 = vec_sel(rsum, dst2, dstmask);
	695	+ rsum = vec_perm(fsum, fsum, dstperm);
	696	+ fdst1 = vec_sel(dst1, rsum, dstmask);
	697	+ fdst2 = vec_sel(rsum, dst2, dstmask);
632	698
633	699	vec_st(fdst1, 0, dst);
634	700	vec_st(fdst2, 16, dst);

Cosmetics: should not hurt performance, scream if are

Keyboard shortcuts

General

Issues and pull requests

Issues, pull requests and Inbox

Repository code