2 * Copyright (c) 2002 Brian Foley
3 * Copyright (c) 2002 Dieter Shirley
4 * Copyright (c) 2003-2004 Romain Dolbeau <romain@dolbeau.org>
6 * This file is part of FFmpeg.
8 * FFmpeg is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * FFmpeg is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 #include "libavcodec/dsputil.h"
25 #include "dsputil_ppc.h"
27 #include "dsputil_altivec.h"
29 void fdct_altivec(int16_t *block);
30 void gmc1_altivec(uint8_t *dst, uint8_t *src, int stride, int h,
31 int x16, int y16, int rounder);
32 void idct_put_altivec(uint8_t *dest, int line_size, int16_t *block);
33 void idct_add_altivec(uint8_t *dest, int line_size, int16_t *block);
35 void dsputil_h264_init_ppc(DSPContext* c, AVCodecContext *avctx);
37 void dsputil_init_altivec(DSPContext* c, AVCodecContext *avctx);
38 void vc1dsp_init_altivec(DSPContext* c, AVCodecContext *avctx);
39 void snow_init_altivec(DSPContext* c, AVCodecContext *avctx);
40 void float_init_altivec(DSPContext* c, AVCodecContext *avctx);
41 void int_init_altivec(DSPContext* c, AVCodecContext *avctx);
50 result |= FF_MM_ALTIVEC;
56 #if CONFIG_POWERPC_PERF
57 unsigned long long perfdata[POWERPC_NUM_PMC_ENABLED][powerpc_perf_total][powerpc_data_total];
58 /* list below must match enum in dsputil_ppc.h */
59 static unsigned char* perfname[] = {
60 "ff_fft_calc_altivec",
62 "dct_unquantize_h263_altivec",
66 "put_pixels16_altivec",
67 "avg_pixels16_altivec",
68 "avg_pixels8_altivec",
69 "put_pixels8_xy2_altivec",
70 "put_no_rnd_pixels8_xy2_altivec",
71 "put_pixels16_xy2_altivec",
72 "put_no_rnd_pixels16_xy2_altivec",
73 "hadamard8_diff8x8_altivec",
74 "hadamard8_diff16_altivec",
75 "avg_pixels8_xy2_altivec",
76 "clear_blocks_dcbz32_ppc",
77 "clear_blocks_dcbz128_ppc",
78 "put_h264_chroma_mc8_altivec",
79 "avg_h264_chroma_mc8_altivec",
80 "put_h264_qpel16_h_lowpass_altivec",
81 "avg_h264_qpel16_h_lowpass_altivec",
82 "put_h264_qpel16_v_lowpass_altivec",
83 "avg_h264_qpel16_v_lowpass_altivec",
84 "put_h264_qpel16_hv_lowpass_altivec",
85 "avg_h264_qpel16_hv_lowpass_altivec",
91 #if CONFIG_POWERPC_PERF
92 void powerpc_display_perf_report(void)
95 av_log(NULL, AV_LOG_INFO, "PowerPC performance report\n Values are from the PMC registers, and represent whatever the registers are set to record.\n");
96 for(i = 0 ; i < powerpc_perf_total ; i++) {
97 for (j = 0; j < POWERPC_NUM_PMC_ENABLED ; j++) {
98 if (perfdata[j][i][powerpc_data_num] != (unsigned long long)0)
99 av_log(NULL, AV_LOG_INFO,
100 " Function \"%s\" (pmc%d):\n\tmin: %"PRIu64"\n\tmax: %"PRIu64"\n\tavg: %1.2lf (%"PRIu64")\n",
103 perfdata[j][i][powerpc_data_min],
104 perfdata[j][i][powerpc_data_max],
105 (double)perfdata[j][i][powerpc_data_sum] /
106 (double)perfdata[j][i][powerpc_data_num],
107 perfdata[j][i][powerpc_data_num]);
111 #endif /* CONFIG_POWERPC_PERF */
113 /* ***** WARNING ***** WARNING ***** WARNING ***** */
115 clear_blocks_dcbz32_ppc will not work properly on PowerPC processors with a
116 cache line size not equal to 32 bytes.
117 Fortunately all processor used by Apple up to at least the 7450 (aka second
118 generation G4) use 32 bytes cache line.
119 This is due to the use of the 'dcbz' instruction. It simply clear to zero a
120 single cache line, so you need to know the cache line size to use it !
121 It's absurd, but it's fast...
123 update 24/06/2003 : Apple released yesterday the G5, with a PPC970. cache line
124 size: 128 bytes. Oups.
125 The semantic of dcbz was changed, it always clear 32 bytes. so the function
126 below will work, but will be slow. So I fixed check_dcbz_effect to use dcbzl,
127 which is defined to clear a cache line (as dcbz before). So we still can
128 distinguish, and use dcbz (32 bytes) or dcbzl (one cache line) as required.
130 see <http://developer.apple.com/technotes/tn/tn2087.html>
131 and <http://developer.apple.com/technotes/tn/tn2086.html>
133 void clear_blocks_dcbz32_ppc(DCTELEM *blocks)
135 POWERPC_PERF_DECLARE(powerpc_clear_blocks_dcbz32, 1);
136 register int misal = ((unsigned long)blocks & 0x00000010);
138 POWERPC_PERF_START_COUNT(powerpc_clear_blocks_dcbz32, 1);
141 ((unsigned long*)blocks)[0] = 0L;
142 ((unsigned long*)blocks)[1] = 0L;
143 ((unsigned long*)blocks)[2] = 0L;
144 ((unsigned long*)blocks)[3] = 0L;
147 for ( ; i < sizeof(DCTELEM)*6*64-31 ; i += 32) {
148 __asm__ volatile("dcbz %0,%1" : : "b" (blocks), "r" (i) : "memory");
151 ((unsigned long*)blocks)[188] = 0L;
152 ((unsigned long*)blocks)[189] = 0L;
153 ((unsigned long*)blocks)[190] = 0L;
154 ((unsigned long*)blocks)[191] = 0L;
158 memset(blocks, 0, sizeof(DCTELEM)*6*64);
160 POWERPC_PERF_STOP_COUNT(powerpc_clear_blocks_dcbz32, 1);
163 /* same as above, when dcbzl clear a whole 128B cache line
164 i.e. the PPC970 aka G5 */
166 void clear_blocks_dcbz128_ppc(DCTELEM *blocks)
168 POWERPC_PERF_DECLARE(powerpc_clear_blocks_dcbz128, 1);
169 register int misal = ((unsigned long)blocks & 0x0000007f);
171 POWERPC_PERF_START_COUNT(powerpc_clear_blocks_dcbz128, 1);
174 // we could probably also optimize this case,
175 // but there's not much point as the machines
176 // aren't available yet (2003-06-26)
177 memset(blocks, 0, sizeof(DCTELEM)*6*64);
180 for ( ; i < sizeof(DCTELEM)*6*64 ; i += 128) {
181 __asm__ volatile("dcbzl %0,%1" : : "b" (blocks), "r" (i) : "memory");
184 memset(blocks, 0, sizeof(DCTELEM)*6*64);
186 POWERPC_PERF_STOP_COUNT(powerpc_clear_blocks_dcbz128, 1);
189 void clear_blocks_dcbz128_ppc(DCTELEM *blocks)
191 memset(blocks, 0, sizeof(DCTELEM)*6*64);
196 /* check dcbz report how many bytes are set to 0 by dcbz */
197 /* update 24/06/2003 : replace dcbz by dcbzl to get
198 the intended effect (Apple "fixed" dcbz)
199 unfortunately this cannot be used unless the assembler
200 knows about dcbzl ... */
201 long check_dcbzl_effect(void)
203 register char *fakedata = av_malloc(1024);
204 register char *fakedata_middle;
205 register long zero = 0;
213 fakedata_middle = (fakedata + 512);
215 memset(fakedata, 0xFF, 1024);
217 /* below the constraint "b" seems to mean "Address base register"
218 in gcc-3.3 / RS/6000 speaks. seems to avoid using r0, so.... */
219 __asm__ volatile("dcbzl %0, %1" : : "b" (fakedata_middle), "r" (zero));
221 for (i = 0; i < 1024 ; i ++) {
222 if (fakedata[i] == (char)0)
231 long check_dcbzl_effect(void)
237 static void prefetch_ppc(void *mem, int stride, int h)
239 register const uint8_t *p = mem;
241 __asm__ volatile ("dcbt 0,%0" : : "r" (p));
246 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx)
248 // Common optimizations whether AltiVec is available or not
249 c->prefetch = prefetch_ppc;
250 switch (check_dcbzl_effect()) {
252 c->clear_blocks = clear_blocks_dcbz32_ppc;
255 c->clear_blocks = clear_blocks_dcbz128_ppc;
262 if(CONFIG_H264_DECODER) dsputil_h264_init_ppc(c, avctx);
265 mm_flags |= FF_MM_ALTIVEC;
267 dsputil_init_altivec(c, avctx);
268 if(CONFIG_SNOW_DECODER) snow_init_altivec(c, avctx);
269 if(CONFIG_VC1_DECODER || CONFIG_WMV3_DECODER)
270 vc1dsp_init_altivec(c, avctx);
271 float_init_altivec(c, avctx);
272 int_init_altivec(c, avctx);
273 c->gmc1 = gmc1_altivec;
276 if (avctx->dct_algo == FF_DCT_AUTO ||
277 avctx->dct_algo == FF_DCT_ALTIVEC) {
278 c->fdct = fdct_altivec;
280 #endif //CONFIG_ENCODERS
282 if (avctx->lowres==0) {
283 if ((avctx->idct_algo == FF_IDCT_AUTO) ||
284 (avctx->idct_algo == FF_IDCT_ALTIVEC)) {
285 c->idct_put = idct_put_altivec;
286 c->idct_add = idct_add_altivec;
287 c->idct_permutation_type = FF_TRANSPOSE_IDCT_PERM;
291 #if CONFIG_POWERPC_PERF
294 for (i = 0 ; i < powerpc_perf_total ; i++) {
295 for (j = 0; j < POWERPC_NUM_PMC_ENABLED ; j++) {
296 perfdata[j][i][powerpc_data_min] = 0xFFFFFFFFFFFFFFFFULL;
297 perfdata[j][i][powerpc_data_max] = 0x0000000000000000ULL;
298 perfdata[j][i][powerpc_data_sum] = 0x0000000000000000ULL;
299 perfdata[j][i][powerpc_data_num] = 0x0000000000000000ULL;
303 #endif /* CONFIG_POWERPC_PERF */
305 #endif /* HAVE_ALTIVEC */