projects / mesa.git / blob
? search:


34f8cec43b637524b325d64fe76960fd52e9c95b
[mesa.git] / src / panfrost / pandecode / decode.c
1 /*
2 * Copyright (C) 2017-2019 Alyssa Rosenzweig
3 * Copyright (C) 2017-2019 Connor Abbott
4 * Copyright (C) 2019 Collabora, Ltd.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 * SOFTWARE.
24 */
25
26 #include <panfrost-job.h>
27 #include <stdio.h>
28 #include <stdlib.h>
29 #include <memory.h>
30 #include <stdbool.h>
31 #include <stdarg.h>
32 #include "decode.h"
33 #include "util/u_math.h"
34
35 #include "pan_pretty_print.h"
36 #include "midgard/disassemble.h"
37 #include "bifrost/disassemble.h"
38
39 int pandecode_replay_jc(mali_ptr jc_gpu_va, bool bifrost);
40
41 #define MEMORY_PROP(obj, p) {\
42 if (obj->p) { \
43 char *a = pointer_as_memory_reference(obj->p); \
44 pandecode_prop("%s = %s", #p, a); \
45 free(a); \
46 } \
47 }
48
49 #define DYN_MEMORY_PROP(obj, no, p) { \
50 if (obj->p) \
51 pandecode_prop("%s = %s_%d_p", #p, #p, no); \
52 }
53
54 /* Semantic logging type.
55 *
56 * Raw: for raw messages to be printed as is.
57 * Message: for helpful information to be commented out in replays.
58 * Property: for properties of a struct
59 *
60 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
61 */
62
63 enum pandecode_log_type {
64 PANDECODE_RAW,
65 PANDECODE_MESSAGE,
66 PANDECODE_PROPERTY
67 };
68
69 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
70 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
71 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
72
73 unsigned pandecode_indent = 0;
74
75 static void
76 pandecode_make_indent(void)
77 {
78 for (unsigned i = 0; i < pandecode_indent; ++i)
79 printf(" ");
80 }
81
82 static void
83 pandecode_log_typed(enum pandecode_log_type type, const char *format, ...)
84 {
85 va_list ap;
86
87 pandecode_make_indent();
88
89 if (type == PANDECODE_MESSAGE)
90 printf("// ");
91 else if (type == PANDECODE_PROPERTY)
92 printf(".");
93
94 va_start(ap, format);
95 vprintf(format, ap);
96 va_end(ap);
97
98 if (type == PANDECODE_PROPERTY)
99 printf(",\n");
100 }
101
102 static void
103 pandecode_log_cont(const char *format, ...)
104 {
105 va_list ap;
106
107 va_start(ap, format);
108 vprintf(format, ap);
109 va_end(ap);
110 }
111
112 struct pandecode_flag_info {
113 u64 flag;
114 const char *name;
115 };
116
117 static void
118 pandecode_log_decoded_flags(const struct pandecode_flag_info *flag_info,
119 u64 flags)
120 {
121 bool decodable_flags_found = false;
122
123 for (int i = 0; flag_info[i].name; i++) {
124 if ((flags & flag_info[i].flag) != flag_info[i].flag)
125 continue;
126
127 if (!decodable_flags_found) {
128 decodable_flags_found = true;
129 } else {
130 pandecode_log_cont(" | ");
131 }
132
133 pandecode_log_cont("%s", flag_info[i].name);
134
135 flags &= ~flag_info[i].flag;
136 }
137
138 if (decodable_flags_found) {
139 if (flags)
140 pandecode_log_cont(" | 0x%" PRIx64, flags);
141 } else {
142 pandecode_log_cont("0x%" PRIx64, flags);
143 }
144 }
145
146 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
147 static const struct pandecode_flag_info gl_enable_flag_info[] = {
148 FLAG_INFO(OCCLUSION_QUERY),
149 FLAG_INFO(OCCLUSION_PRECISE),
150 FLAG_INFO(FRONT_CCW_TOP),
151 FLAG_INFO(CULL_FACE_FRONT),
152 FLAG_INFO(CULL_FACE_BACK),
153 {}
154 };
155 #undef FLAG_INFO
156
157 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
158 static const struct pandecode_flag_info clear_flag_info[] = {
159 FLAG_INFO(FAST),
160 FLAG_INFO(SLOW),
161 FLAG_INFO(SLOW_STENCIL),
162 {}
163 };
164 #undef FLAG_INFO
165
166 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
167 static const struct pandecode_flag_info mask_flag_info[] = {
168 FLAG_INFO(R),
169 FLAG_INFO(G),
170 FLAG_INFO(B),
171 FLAG_INFO(A),
172 {}
173 };
174 #undef FLAG_INFO
175
176 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
177 static const struct pandecode_flag_info u3_flag_info[] = {
178 FLAG_INFO(HAS_MSAA),
179 FLAG_INFO(CAN_DISCARD),
180 FLAG_INFO(HAS_BLEND_SHADER),
181 FLAG_INFO(DEPTH_TEST),
182 {}
183 };
184
185 static const struct pandecode_flag_info u4_flag_info[] = {
186 FLAG_INFO(NO_MSAA),
187 FLAG_INFO(NO_DITHER),
188 FLAG_INFO(DEPTH_RANGE_A),
189 FLAG_INFO(DEPTH_RANGE_B),
190 FLAG_INFO(STENCIL_TEST),
191 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER),
192 {}
193 };
194 #undef FLAG_INFO
195
196 #define FLAG_INFO(flag) { MALI_FRAMEBUFFER_##flag, "MALI_FRAMEBUFFER_" #flag }
197 static const struct pandecode_flag_info fb_fmt_flag_info[] = {
198 FLAG_INFO(MSAA_A),
199 FLAG_INFO(MSAA_B),
200 FLAG_INFO(MSAA_8),
201 {}
202 };
203 #undef FLAG_INFO
204
205 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
206 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
207 FLAG_INFO(MSAA),
208 FLAG_INFO(SRGB),
209 {}
210 };
211 #undef FLAG_INFO
212
213 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
214 static const struct pandecode_flag_info mfbd_extra_flag_info[] = {
215 FLAG_INFO(PRESENT),
216 FLAG_INFO(AFBC),
217 FLAG_INFO(ZS),
218 {}
219 };
220 #undef FLAG_INFO
221
222 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
223 static const struct pandecode_flag_info shader_midgard1_flag_info [] = {
224 FLAG_INFO(EARLY_Z),
225 FLAG_INFO(HELPER_INVOCATIONS),
226 FLAG_INFO(READS_TILEBUFFER),
227 FLAG_INFO(READS_ZS),
228 {}
229 };
230 #undef FLAG_INFO
231
232 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
233 static const struct pandecode_flag_info mfbd_flag_info [] = {
234 FLAG_INFO(DEPTH_WRITE),
235 FLAG_INFO(EXTRA),
236 {}
237 };
238 #undef FLAG_INFO
239
240
241 extern char *replace_fragment;
242 extern char *replace_vertex;
243
244 static char *
245 pandecode_job_type_name(enum mali_job_type type)
246 {
247 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
248
249 switch (type) {
250 DEFINE_CASE(NULL);
251 DEFINE_CASE(SET_VALUE);
252 DEFINE_CASE(CACHE_FLUSH);
253 DEFINE_CASE(COMPUTE);
254 DEFINE_CASE(VERTEX);
255 DEFINE_CASE(TILER);
256 DEFINE_CASE(FUSED);
257 DEFINE_CASE(FRAGMENT);
258
259 case JOB_NOT_STARTED:
260 return "NOT_STARTED";
261
262 default:
263 pandecode_log("Warning! Unknown job type %x\n", type);
264 return "!?!?!?";
265 }
266
267 #undef DEFINE_CASE
268 }
269
270 static char *
271 pandecode_draw_mode_name(enum mali_draw_mode mode)
272 {
273 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
274
275 switch (mode) {
276 DEFINE_CASE(DRAW_NONE);
277 DEFINE_CASE(POINTS);
278 DEFINE_CASE(LINES);
279 DEFINE_CASE(TRIANGLES);
280 DEFINE_CASE(TRIANGLE_STRIP);
281 DEFINE_CASE(TRIANGLE_FAN);
282 DEFINE_CASE(LINE_STRIP);
283 DEFINE_CASE(LINE_LOOP);
284 DEFINE_CASE(POLYGON);
285 DEFINE_CASE(QUADS);
286 DEFINE_CASE(QUAD_STRIP);
287
288 default:
289 return "MALI_TRIANGLES /* XXX: Unknown GL mode, check dump */";
290 }
291
292 #undef DEFINE_CASE
293 }
294
295 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
296 static char *
297 pandecode_func_name(enum mali_func mode)
298 {
299 switch (mode) {
300 DEFINE_CASE(NEVER);
301 DEFINE_CASE(LESS);
302 DEFINE_CASE(EQUAL);
303 DEFINE_CASE(LEQUAL);
304 DEFINE_CASE(GREATER);
305 DEFINE_CASE(NOTEQUAL);
306 DEFINE_CASE(GEQUAL);
307 DEFINE_CASE(ALWAYS);
308
309 default:
310 return "MALI_FUNC_NEVER /* XXX: Unknown function, check dump */";
311 }
312 }
313 #undef DEFINE_CASE
314
315 /* Why is this duplicated? Who knows... */
316 #define DEFINE_CASE(name) case MALI_ALT_FUNC_ ## name: return "MALI_ALT_FUNC_" #name
317 static char *
318 pandecode_alt_func_name(enum mali_alt_func mode)
319 {
320 switch (mode) {
321 DEFINE_CASE(NEVER);
322 DEFINE_CASE(LESS);
323 DEFINE_CASE(EQUAL);
324 DEFINE_CASE(LEQUAL);
325 DEFINE_CASE(GREATER);
326 DEFINE_CASE(NOTEQUAL);
327 DEFINE_CASE(GEQUAL);
328 DEFINE_CASE(ALWAYS);
329
330 default:
331 return "MALI_FUNC_NEVER /* XXX: Unknown function, check dump */";
332 }
333 }
334 #undef DEFINE_CASE
335
336 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
337 static char *
338 pandecode_stencil_op_name(enum mali_stencil_op op)
339 {
340 switch (op) {
341 DEFINE_CASE(KEEP);
342 DEFINE_CASE(REPLACE);
343 DEFINE_CASE(ZERO);
344 DEFINE_CASE(INVERT);
345 DEFINE_CASE(INCR_WRAP);
346 DEFINE_CASE(DECR_WRAP);
347 DEFINE_CASE(INCR);
348 DEFINE_CASE(DECR);
349
350 default:
351 return "MALI_STENCIL_KEEP /* XXX: Unknown stencil op, check dump */";
352 }
353 }
354
355 #undef DEFINE_CASE
356
357 #define DEFINE_CASE(name) case MALI_ATTR_ ## name: return "MALI_ATTR_" #name
358 static char *pandecode_attr_mode_name(enum mali_attr_mode mode)
359 {
360 switch(mode) {
361 DEFINE_CASE(UNUSED);
362 DEFINE_CASE(LINEAR);
363 DEFINE_CASE(POT_DIVIDE);
364 DEFINE_CASE(MODULO);
365 DEFINE_CASE(NPOT_DIVIDE);
366 default:
367 return "MALI_ATTR_UNUSED /* XXX: Unknown stencil op, check dump */";
368 }
369 }
370
371 #undef DEFINE_CASE
372
373 #define DEFINE_CASE(name) case MALI_CHANNEL_## name: return "MALI_CHANNEL_" #name
374 static char *
375 pandecode_channel_name(enum mali_channel channel)
376 {
377 switch (channel) {
378 DEFINE_CASE(RED);
379 DEFINE_CASE(GREEN);
380 DEFINE_CASE(BLUE);
381 DEFINE_CASE(ALPHA);
382 DEFINE_CASE(ZERO);
383 DEFINE_CASE(ONE);
384 DEFINE_CASE(RESERVED_0);
385 DEFINE_CASE(RESERVED_1);
386
387 default:
388 return "MALI_CHANNEL_ZERO /* XXX: Unknown channel, check dump */";
389 }
390 }
391 #undef DEFINE_CASE
392
393 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
394 static char *
395 pandecode_wrap_mode_name(enum mali_wrap_mode op)
396 {
397 switch (op) {
398 DEFINE_CASE(REPEAT);
399 DEFINE_CASE(CLAMP_TO_EDGE);
400 DEFINE_CASE(CLAMP_TO_BORDER);
401 DEFINE_CASE(MIRRORED_REPEAT);
402
403 default:
404 return "MALI_WRAP_REPEAT /* XXX: Unknown wrap mode, check dump */";
405 }
406 }
407 #undef DEFINE_CASE
408
409 #define DEFINE_CASE(name) case MALI_TEX_## name: return "MALI_TEX_" #name
410 static char *
411 pandecode_texture_type(enum mali_texture_type type)
412 {
413 switch (type) {
414 DEFINE_CASE(1D);
415 DEFINE_CASE(2D);
416 DEFINE_CASE(3D);
417 DEFINE_CASE(CUBE);
418
419 default:
420 unreachable("Unknown case");
421 }
422 }
423 #undef DEFINE_CASE
424
425 #define DEFINE_CASE(name) case MALI_MFBD_BLOCK_## name: return "MALI_MFBD_BLOCK_" #name
426 static char *
427 pandecode_mfbd_block_format(enum mali_mfbd_block_format fmt)
428 {
429 switch (fmt) {
430 DEFINE_CASE(TILED);
431 DEFINE_CASE(UNKNOWN);
432 DEFINE_CASE(LINEAR);
433 DEFINE_CASE(AFBC);
434
435 default:
436 unreachable("Invalid case");
437 }
438 }
439 #undef DEFINE_CASE
440
441 /* Midgard's tiler descriptor is embedded within the
442 * larger FBD */
443
444 static void
445 pandecode_midgard_tiler_descriptor(const struct midgard_tiler_descriptor *t)
446 {
447 pandecode_log(".tiler = {\n");
448 pandecode_indent++;
449
450 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
451 pandecode_prop("flags = 0x%" PRIx16, t->flags);
452 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
453
454 MEMORY_PROP(t, polygon_list);
455 MEMORY_PROP(t, polygon_list_body);
456
457 MEMORY_PROP(t, heap_start);
458
459 {
460 /* Points to the end of a buffer */
461 char *a = pointer_as_memory_reference(t->heap_end - 1);
462 pandecode_prop("heap_end = %s + 1", a);
463 free(a);
464 }
465
466 bool nonzero_weights = false;
467
468 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
469 nonzero_weights |= t->weights[w] != 0x0;
470 }
471
472 if (nonzero_weights) {
473 pandecode_log(".weights = {");
474
475 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
476 pandecode_log("%d, ", t->weights[w]);
477 }
478
479 pandecode_log("},");
480 }
481
482 pandecode_indent--;
483 pandecode_log("}\n");
484 }
485
486 static void
487 pandecode_replay_sfbd(uint64_t gpu_va, int job_no)
488 {
489 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
490 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
491
492 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
493 pandecode_indent++;
494
495 pandecode_prop("unknown1 = 0x%" PRIx32, s->unknown1);
496 pandecode_prop("unknown2 = 0x%" PRIx32, s->unknown2);
497
498 pandecode_log(".format = ");
499 pandecode_log_decoded_flags(fb_fmt_flag_info, s->format);
500 pandecode_log_cont(",\n");
501
502 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", s->width + 1);
503 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", s->height + 1);
504
505 MEMORY_PROP(s, framebuffer);
506 pandecode_prop("stride = %d", s->stride);
507
508 /* Earlier in the actual commandstream -- right before width -- but we
509 * delay to flow nicer */
510
511 pandecode_log(".clear_flags = ");
512 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
513 pandecode_log_cont(",\n");
514
515 if (s->depth_buffer | s->depth_buffer_enable) {
516 MEMORY_PROP(s, depth_buffer);
517 pandecode_prop("depth_buffer_enable = %s", DS_ENABLE(s->depth_buffer_enable));
518 }
519
520 if (s->stencil_buffer | s->stencil_buffer_enable) {
521 MEMORY_PROP(s, stencil_buffer);
522 pandecode_prop("stencil_buffer_enable = %s", DS_ENABLE(s->stencil_buffer_enable));
523 }
524
525 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
526 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
527 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
528 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
529 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
530 }
531
532 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
533 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
534 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
535 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
536 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
537 }
538
539 if (s->clear_stencil) {
540 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
541 }
542
543 MEMORY_PROP(s, unknown_address_0);
544 const struct midgard_tiler_descriptor t = s->tiler;
545 pandecode_midgard_tiler_descriptor(&t);
546
547 pandecode_indent--;
548 pandecode_log("};\n");
549
550 pandecode_prop("zero0 = 0x%" PRIx64, s->zero0);
551 pandecode_prop("zero1 = 0x%" PRIx64, s->zero1);
552 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
553 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
554
555 printf(".zero3 = {");
556
557 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
558 printf("%X, ", s->zero3[i]);
559
560 printf("},\n");
561
562 printf(".zero6 = {");
563
564 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
565 printf("%X, ", s->zero6[i]);
566
567 printf("},\n");
568 }
569
570 static void
571 pandecode_u32_slide(unsigned name, const u32 *slide, unsigned count)
572 {
573 pandecode_log(".unknown%d = {", name);
574
575 for (int i = 0; i < count; ++i)
576 printf("%X, ", slide[i]);
577
578 pandecode_log("},\n");
579 }
580
581 #define SHORT_SLIDE(num) \
582 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
583
584 static void
585 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
586 {
587 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
588 const struct mali_compute_fbd *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
589
590 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
591 pandecode_indent++;
592
593 SHORT_SLIDE(1);
594
595 pandecode_indent--;
596 printf("},\n");
597 }
598
599 static void
600 pandecode_replay_swizzle(unsigned swizzle)
601 {
602 pandecode_prop("swizzle = %s | (%s << 3) | (%s << 6) | (%s << 9)",
603 pandecode_channel_name((swizzle >> 0) & 0x7),
604 pandecode_channel_name((swizzle >> 3) & 0x7),
605 pandecode_channel_name((swizzle >> 6) & 0x7),
606 pandecode_channel_name((swizzle >> 9) & 0x7));
607 }
608
609 static void
610 pandecode_rt_format(struct mali_rt_format format)
611 {
612 pandecode_log(".format = {\n");
613 pandecode_indent++;
614
615 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
616 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
617 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
618
619 pandecode_prop("block = %s",
620 pandecode_mfbd_block_format(format.block));
621
622 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
623 MALI_NEGATIVE(format.nr_channels));
624
625 pandecode_log(".flags = ");
626 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
627 pandecode_log_cont(",\n");
628
629 pandecode_replay_swizzle(format.swizzle);
630
631 pandecode_prop("unk4 = 0x%" PRIx32, format.unk4);
632
633 pandecode_indent--;
634 pandecode_log("},\n");
635 }
636
637 static void
638 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct bifrost_framebuffer *fb)
639 {
640 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
641 pandecode_indent++;
642
643 for (int i = 0; i < MALI_NEGATIVE(fb->rt_count_1); i++) {
644 mali_ptr rt_va = gpu_va + i * sizeof(struct bifrost_render_target);
645 struct pandecode_mapped_memory *mem =
646 pandecode_find_mapped_gpu_mem_containing(rt_va);
647 const struct bifrost_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
648
649 pandecode_log("{\n");
650 pandecode_indent++;
651
652 pandecode_rt_format(rt->format);
653
654 if (rt->format.block == MALI_MFBD_BLOCK_AFBC) {
655 pandecode_log(".afbc = {\n");
656 pandecode_indent++;
657
658 char *a = pointer_as_memory_reference(rt->afbc.metadata);
659 pandecode_prop("metadata = %s", a);
660 free(a);
661
662 pandecode_prop("stride = %d", rt->afbc.stride);
663 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
664
665 pandecode_indent--;
666 pandecode_log("},\n");
667 } else {
668 pandecode_log(".chunknown = {\n");
669 pandecode_indent++;
670
671 pandecode_prop("unk = 0x%" PRIx64, rt->chunknown.unk);
672
673 char *a = pointer_as_memory_reference(rt->chunknown.pointer);
674 pandecode_prop("pointer = %s", a);
675 free(a);
676
677 pandecode_indent--;
678 pandecode_log("},\n");
679 }
680
681 MEMORY_PROP(rt, framebuffer);
682 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
683
684 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
685 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
686 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
687 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
688 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
689 }
690
691 if (rt->zero1 || rt->zero2 || rt->zero3) {
692 pandecode_msg("render target zeros tripped\n");
693 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
694 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
695 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
696 }
697
698 pandecode_indent--;
699 pandecode_log("},\n");
700 }
701
702 pandecode_indent--;
703 pandecode_log("};\n");
704 }
705
706 static unsigned
707 pandecode_replay_mfbd_bfr(uint64_t gpu_va, int job_no, bool with_render_targets)
708 {
709 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
710 const struct bifrost_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
711
712 if (fb->sample_locations) {
713 /* The blob stores all possible sample locations in a single buffer
714 * allocated on startup, and just switches the pointer when switching
715 * MSAA state. For now, we just put the data into the cmdstream, but we
716 * should do something like what the blob does with a real driver.
717 *
718 * There seem to be 32 slots for sample locations, followed by another
719 * 16. The second 16 is just the center location followed by 15 zeros
720 * in all the cases I've identified (maybe shader vs. depth/color
721 * samples?).
722 */
723
724 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->sample_locations);
725
726 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->sample_locations);
727
728 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
729 pandecode_indent++;
730
731 for (int i = 0; i < 32 + 16; i++) {
732 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
733 }
734
735 pandecode_indent--;
736 pandecode_log("};\n");
737 }
738
739 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
740 pandecode_indent++;
741
742 pandecode_prop("unk0 = 0x%x", fb->unk0);
743
744 if (fb->sample_locations)
745 pandecode_prop("sample_locations = sample_locations_%d", job_no);
746
747 /* Assume that unknown1 was emitted in the last job for
748 * now */
749 MEMORY_PROP(fb, unknown1);
750
751 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
752 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
753 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
754 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
755
756 pandecode_prop("unk1 = 0x%x", fb->unk1);
757 pandecode_prop("unk2 = 0x%x", fb->unk2);
758 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
759 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
760
761 pandecode_log(".mfbd_flags = ");
762 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
763 pandecode_log_cont(",\n");
764
765 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
766 pandecode_prop("clear_depth = %f", fb->clear_depth);
767
768 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
769 MEMORY_PROP(fb, scratchpad);
770 const struct midgard_tiler_descriptor t = fb->tiler;
771 pandecode_midgard_tiler_descriptor(&t);
772
773 if (fb->zero3 || fb->zero4) {
774 pandecode_msg("framebuffer zeros tripped\n");
775 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
776 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
777 }
778
779 pandecode_indent--;
780 pandecode_log("};\n");
781
782 gpu_va += sizeof(struct bifrost_framebuffer);
783
784 if ((fb->mfbd_flags & MALI_MFBD_EXTRA) && with_render_targets) {
785 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
786 const struct bifrost_fb_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
787
788 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
789 pandecode_indent++;
790
791 MEMORY_PROP(fbx, checksum);
792
793 if (fbx->checksum_stride)
794 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
795
796 pandecode_log(".flags = ");
797 pandecode_log_decoded_flags(mfbd_extra_flag_info, fbx->flags);
798 pandecode_log_cont(",\n");
799
800 if (fbx->flags & MALI_EXTRA_AFBC_ZS) {
801 pandecode_log(".ds_afbc = {\n");
802 pandecode_indent++;
803
804 MEMORY_PROP((&fbx->ds_afbc), depth_stencil_afbc_metadata);
805 pandecode_prop("depth_stencil_afbc_stride = %d",
806 fbx->ds_afbc.depth_stencil_afbc_stride);
807 MEMORY_PROP((&fbx->ds_afbc), depth_stencil);
808
809 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
810 pandecode_msg("Depth/stencil AFBC zeros tripped\n");
811 pandecode_prop("zero1 = 0x%" PRIx32,
812 fbx->ds_afbc.zero1);
813 pandecode_prop("padding = 0x%" PRIx64,
814 fbx->ds_afbc.padding);
815 }
816
817 pandecode_indent--;
818 pandecode_log("},\n");
819 } else {
820 pandecode_log(".ds_linear = {\n");
821 pandecode_indent++;
822
823 if (fbx->ds_linear.depth) {
824 MEMORY_PROP((&fbx->ds_linear), depth);
825 pandecode_prop("depth_stride = %d",
826 fbx->ds_linear.depth_stride);
827 }
828
829 if (fbx->ds_linear.stencil) {
830 MEMORY_PROP((&fbx->ds_linear), stencil);
831 pandecode_prop("stencil_stride = %d",
832 fbx->ds_linear.stencil_stride);
833 }
834
835 if (fbx->ds_linear.depth_stride_zero ||
836 fbx->ds_linear.stencil_stride_zero ||
837 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
838 pandecode_msg("Depth/stencil zeros tripped\n");
839 pandecode_prop("depth_stride_zero = 0x%x",
840 fbx->ds_linear.depth_stride_zero);
841 pandecode_prop("stencil_stride_zero = 0x%x",
842 fbx->ds_linear.stencil_stride_zero);
843 pandecode_prop("zero1 = 0x%" PRIx32,
844 fbx->ds_linear.zero1);
845 pandecode_prop("zero2 = 0x%" PRIx32,
846 fbx->ds_linear.zero2);
847 }
848
849 pandecode_indent--;
850 pandecode_log("},\n");
851 }
852
853 if (fbx->zero3 || fbx->zero4) {
854 pandecode_msg("fb_extra zeros tripped\n");
855 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
856 pandecode_prop("zero4 = 0x%" PRIx64, fbx->zero4);
857 }
858
859 pandecode_indent--;
860 pandecode_log("};\n");
861
862 gpu_va += sizeof(struct bifrost_fb_extra);
863 }
864
865 if (with_render_targets)
866 pandecode_render_target(gpu_va, job_no, fb);
867
868 /* Passback the render target count */
869 return MALI_NEGATIVE(fb->rt_count_1);
870 }
871
872 /* Just add a comment decoding the shift/odd fields forming the padded vertices
873 * count */
874
875 static void
876 pandecode_padded_vertices(unsigned shift, unsigned k)
877 {
878 unsigned odd = 2*k + 1;
879 unsigned pot = 1 << shift;
880 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
881 }
882
883 /* Given a magic divisor, recover what we were trying to divide by.
884 *
885 * Let m represent the magic divisor. By definition, m is an element on Z, whre
886 * 0 <= m < 2^N, for N bits in m.
887 *
888 * Let q represent the number we would like to divide by.
889 *
890 * By definition of a magic divisor for N-bit unsigned integers (a number you
891 * multiply by to magically get division), m is a number such that:
892 *
893 * (m * x) & (2^N - 1) = floor(x/q).
894 * for all x on Z where 0 <= x < 2^N
895 *
896 * Ignore the case where any of the above values equals zero; it is irrelevant
897 * for our purposes (instanced arrays).
898 *
899 * Choose x = q. Then:
900 *
901 * (m * x) & (2^N - 1) = floor(x/q).
902 * (m * q) & (2^N - 1) = floor(q/q).
903 *
904 * floor(q/q) = floor(1) = 1, therefore:
905 *
906 * (m * q) & (2^N - 1) = 1
907 *
908 * Recall the identity that the bitwise AND of one less than a power-of-two
909 * equals the modulo with that power of two, i.e. for all x:
910 *
911 * x & (2^N - 1) = x % N
912 *
913 * Therefore:
914 *
915 * mq % (2^N) = 1
916 *
917 * By definition, a modular multiplicative inverse of a number m is the number
918 * q such that with respect to a modulos M:
919 *
920 * mq % M = 1
921 *
922 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
923 *
924 */
925
926 static void
927 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
928 {
929 /* Compute the modular inverse of `magic` with respect to 2^(32 -
930 * shift) the most lame way possible... just repeatedly add.
931 * Asymptoptically slow but nobody cares in practice, unless you have
932 * massive numbers of vertices or high divisors. */
933
934 unsigned inverse = 0;
935
936 /* Magic implicitly has the highest bit set */
937 magic |= (1 << 31);
938
939 /* Depending on rounding direction */
940 if (extra)
941 magic++;
942
943 for (;;) {
944 uint32_t product = magic * inverse;
945
946 if (shift) {
947 product >>= shift;
948 }
949
950 if (product == 1)
951 break;
952
953 ++inverse;
954 }
955
956 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
957
958 /* Recall we're supposed to divide by (gl_level_divisor *
959 * padded_num_vertices) */
960
961 unsigned padded_num_vertices = inverse / orig_divisor;
962
963 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
964 }
965
966 static void
967 pandecode_replay_attributes(const struct pandecode_mapped_memory *mem,
968 mali_ptr addr, int job_no, char *suffix,
969 int count, bool varying)
970 {
971 char *prefix = varying ? "varyings" : "attributes";
972
973 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
974
975 char base[128];
976 snprintf(base, sizeof(base), "%s_data_%d%s", prefix, job_no, suffix);
977
978 for (int i = 0; i < count; ++i) {
979 enum mali_attr_mode mode = attr[i].elements & 7;
980
981 if (mode == MALI_ATTR_UNUSED)
982 continue;
983
984 mali_ptr raw_elements = attr[i].elements & ~7;
985
986 /* TODO: Do we maybe want to dump the attribute values
987 * themselves given the specified format? Or is that too hard?
988 * */
989
990 char *a = pointer_as_memory_reference(raw_elements);
991 pandecode_log("mali_ptr %s_%d_p = %s;\n", base, i, a);
992 free(a);
993 }
994
995 pandecode_log("union mali_attr %s_%d[] = {\n", prefix, job_no);
996 pandecode_indent++;
997
998 for (int i = 0; i < count; ++i) {
999 pandecode_log("{\n");
1000 pandecode_indent++;
1001
1002 unsigned mode = attr[i].elements & 7;
1003 pandecode_prop("elements = (%s_%d_p) | %s", base, i, pandecode_attr_mode_name(mode));
1004 pandecode_prop("shift = %d", attr[i].shift);
1005 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1006 pandecode_prop("stride = 0x%" PRIx32, attr[i].stride);
1007 pandecode_prop("size = 0x%" PRIx32, attr[i].size);
1008
1009 /* Decode further where possible */
1010
1011 if (mode == MALI_ATTR_MODULO) {
1012 pandecode_padded_vertices(
1013 attr[i].shift,
1014 attr[i].extra_flags);
1015 }
1016
1017 pandecode_indent--;
1018 pandecode_log("}, \n");
1019
1020 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1021 i++;
1022 pandecode_log("{\n");
1023 pandecode_indent++;
1024 pandecode_prop("unk = 0x%x", attr[i].unk);
1025 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1026 if (attr[i].zero != 0)
1027 pandecode_prop("zero = 0x%x /* XXX zero tripped */", attr[i].zero);
1028 pandecode_prop("divisor = %d", attr[i].divisor);
1029 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1030 pandecode_indent--;
1031 pandecode_log("}, \n");
1032 }
1033
1034 }
1035
1036 pandecode_indent--;
1037 pandecode_log("};\n");
1038 }
1039
1040 static mali_ptr
1041 pandecode_replay_shader_address(const char *name, mali_ptr ptr)
1042 {
1043 /* TODO: Decode flags */
1044 mali_ptr shader_ptr = ptr & ~15;
1045
1046 char *a = pointer_as_memory_reference(shader_ptr);
1047 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1048 free(a);
1049
1050 return shader_ptr;
1051 }
1052
1053 static bool
1054 all_zero(unsigned *buffer, unsigned count)
1055 {
1056 for (unsigned i = 0; i < count; ++i) {
1057 if (buffer[i])
1058 return false;
1059 }
1060
1061 return true;
1062 }
1063
1064 static void
1065 pandecode_replay_stencil(const char *name, const struct mali_stencil_test *stencil)
1066 {
1067 if (all_zero((unsigned *) stencil, sizeof(stencil) / sizeof(unsigned)))
1068 return;
1069
1070 const char *func = pandecode_func_name(stencil->func);
1071 const char *sfail = pandecode_stencil_op_name(stencil->sfail);
1072 const char *dpfail = pandecode_stencil_op_name(stencil->dpfail);
1073 const char *dppass = pandecode_stencil_op_name(stencil->dppass);
1074
1075 if (stencil->zero)
1076 pandecode_msg("Stencil zero tripped: %X\n", stencil->zero);
1077
1078 pandecode_log(".stencil_%s = {\n", name);
1079 pandecode_indent++;
1080 pandecode_prop("ref = %d", stencil->ref);
1081 pandecode_prop("mask = 0x%02X", stencil->mask);
1082 pandecode_prop("func = %s", func);
1083 pandecode_prop("sfail = %s", sfail);
1084 pandecode_prop("dpfail = %s", dpfail);
1085 pandecode_prop("dppass = %s", dppass);
1086 pandecode_indent--;
1087 pandecode_log("},\n");
1088 }
1089
1090 static void
1091 pandecode_replay_blend_equation(const struct mali_blend_equation *blend)
1092 {
1093 if (blend->zero1)
1094 pandecode_msg("Blend zero tripped: %X\n", blend->zero1);
1095
1096 pandecode_log(".equation = {\n");
1097 pandecode_indent++;
1098
1099 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1100 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1101
1102 pandecode_log(".color_mask = ");
1103 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1104 pandecode_log_cont(",\n");
1105
1106 pandecode_indent--;
1107 pandecode_log("},\n");
1108 }
1109
1110 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1111
1112 static unsigned
1113 decode_bifrost_constant(u16 constant)
1114 {
1115 float lo = (float) (constant & 0xFF);
1116 float hi = (float) (constant >> 8);
1117
1118 return (hi / 255.0) + (lo / 65535.0);
1119 }
1120
1121 static mali_ptr
1122 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1123 {
1124 struct bifrost_blend_rt *b =
1125 ((struct bifrost_blend_rt *) descs) + rt_no;
1126
1127 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1128 pandecode_indent++;
1129
1130 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1131 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1132 b->constant, decode_bifrost_constant(b->constant));
1133
1134 /* TODO figure out blend shader enable bit */
1135 pandecode_replay_blend_equation(&b->equation);
1136 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1137 pandecode_prop("index = 0x%" PRIx16, b->index);
1138 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1139
1140 pandecode_indent--;
1141 pandecode_log("},\n");
1142
1143 return 0;
1144 }
1145
1146 static mali_ptr
1147 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1148 {
1149 if (all_zero((unsigned *) blend, sizeof(blend) / sizeof(unsigned)))
1150 return 0;
1151
1152 pandecode_log(".blend = {\n");
1153 pandecode_indent++;
1154
1155 if (is_shader) {
1156 pandecode_replay_shader_address("shader", blend->shader);
1157 } else {
1158 pandecode_replay_blend_equation(&blend->equation);
1159 pandecode_prop("constant = %f", blend->constant);
1160 }
1161
1162 pandecode_indent--;
1163 pandecode_log("},\n");
1164
1165 /* Return blend shader to disassemble if present */
1166 return is_shader ? (blend->shader & ~0xF) : 0;
1167 }
1168
1169 static mali_ptr
1170 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1171 {
1172 struct midgard_blend_rt *b =
1173 ((struct midgard_blend_rt *) descs) + rt_no;
1174
1175 /* Flags determine presence of blend shader */
1176 bool is_shader = (b->flags & 0xF) >= 0x2;
1177
1178 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1179 pandecode_indent++;
1180
1181 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1182
1183 union midgard_blend blend = b->blend;
1184 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1185
1186 pandecode_indent--;
1187 pandecode_log("};\n");
1188
1189 return shader;
1190 }
1191
1192 static int
1193 pandecode_replay_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1194 {
1195 char base[128];
1196 char *prefix = varying ? "varying" : "attribute";
1197 unsigned max_index = 0;
1198 snprintf(base, sizeof(base), "%s_meta", prefix);
1199
1200 pandecode_log("struct mali_attr_meta %s_%d%s[] = {\n", base, job_no, suffix);
1201 pandecode_indent++;
1202
1203 struct mali_attr_meta *attr_meta;
1204 mali_ptr p = varying ? (v->varying_meta & ~0xF) : v->attribute_meta;
1205
1206 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1207
1208 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1209 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1210 sizeof(*attr_mem));
1211
1212 pandecode_log("{\n");
1213 pandecode_indent++;
1214 pandecode_prop("index = %d", attr_meta->index);
1215
1216 if (attr_meta->index > max_index)
1217 max_index = attr_meta->index;
1218 pandecode_replay_swizzle(attr_meta->swizzle);
1219 pandecode_prop("format = %s", pandecode_format_name(attr_meta->format));
1220
1221 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1222 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1223 pandecode_prop("src_offset = %d", attr_meta->src_offset);
1224 pandecode_indent--;
1225 pandecode_log("},\n");
1226
1227 }
1228
1229 pandecode_indent--;
1230 pandecode_log("};\n");
1231
1232 return max_index;
1233 }
1234
1235 static void
1236 pandecode_replay_indices(uintptr_t pindices, uint32_t index_count, int job_no)
1237 {
1238 struct pandecode_mapped_memory *imem = pandecode_find_mapped_gpu_mem_containing(pindices);
1239
1240 if (imem) {
1241 /* Indices are literally just a u32 array :) */
1242
1243 uint32_t *PANDECODE_PTR_VAR(indices, imem, pindices);
1244
1245 pandecode_log("uint32_t indices_%d[] = {\n", job_no);
1246 pandecode_indent++;
1247
1248 for (unsigned i = 0; i < (index_count + 1); i += 3)
1249 pandecode_log("%d, %d, %d,\n",
1250 indices[i],
1251 indices[i + 1],
1252 indices[i + 2]);
1253
1254 pandecode_indent--;
1255 pandecode_log("};\n");
1256 }
1257 }
1258
1259 /* return bits [lo, hi) of word */
1260 static u32
1261 bits(u32 word, u32 lo, u32 hi)
1262 {
1263 if (hi - lo >= 32)
1264 return word; // avoid undefined behavior with the shift
1265
1266 return (word >> lo) & ((1 << (hi - lo)) - 1);
1267 }
1268
1269 static void
1270 pandecode_replay_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no)
1271 {
1272 pandecode_log_cont("{\n");
1273 pandecode_indent++;
1274
1275 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1276 pandecode_prop("size_y_shift = %d", p->size_y_shift);
1277 pandecode_prop("size_z_shift = %d", p->size_z_shift);
1278 pandecode_prop("workgroups_x_shift = %d", p->workgroups_x_shift);
1279 pandecode_prop("workgroups_y_shift = %d", p->workgroups_y_shift);
1280 pandecode_prop("workgroups_z_shift = %d", p->workgroups_z_shift);
1281 pandecode_prop("workgroups_x_shift_2 = 0x%" PRIx32, p->workgroups_x_shift_2);
1282
1283 /* Decode invocation_count. See the comment before the definition of
1284 * invocation_count for an explanation.
1285 */
1286 pandecode_msg("size: (%d, %d, %d)\n",
1287 bits(p->invocation_count, 0, p->size_y_shift) + 1,
1288 bits(p->invocation_count, p->size_y_shift, p->size_z_shift) + 1,
1289 bits(p->invocation_count, p->size_z_shift,
1290 p->workgroups_x_shift) + 1);
1291 pandecode_msg("workgroups: (%d, %d, %d)\n",
1292 bits(p->invocation_count, p->workgroups_x_shift,
1293 p->workgroups_y_shift) + 1,
1294 bits(p->invocation_count, p->workgroups_y_shift,
1295 p->workgroups_z_shift) + 1,
1296 bits(p->invocation_count, p->workgroups_z_shift,
1297 32) + 1);
1298
1299 /* TODO: Decode */
1300 if (p->unknown_draw)
1301 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1302
1303 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1304
1305 pandecode_prop("draw_mode = %s", pandecode_draw_mode_name(p->draw_mode));
1306
1307 /* Index count only exists for tiler jobs anyway */
1308
1309 if (p->index_count)
1310 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1311
1312 if (p->negative_start)
1313 pandecode_prop("negative_start = %d", p->negative_start);
1314
1315 DYN_MEMORY_PROP(p, job_no, indices);
1316
1317 if (p->zero1) {
1318 pandecode_msg("Zero tripped\n");
1319 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1320 }
1321
1322 pandecode_indent--;
1323 pandecode_log("},\n");
1324 }
1325
1326 static void
1327 pandecode_replay_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1328 {
1329 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1330
1331 struct mali_uniform_buffer_meta *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1332
1333 for (int i = 0; i < ubufs_count; i++) {
1334 mali_ptr ptr = ubufs[i].ptr << 2;
1335 struct pandecode_mapped_memory *umem2 = pandecode_find_mapped_gpu_mem_containing(ptr);
1336 uint32_t *PANDECODE_PTR_VAR(ubuf, umem2, ptr);
1337 char name[50];
1338 snprintf(name, sizeof(name), "ubuf_%d", i);
1339 /* The blob uses ubuf 0 to upload internal stuff and
1340 * uniforms that won't fit/are accessed indirectly, so
1341 * it puts it in the batchbuffer.
1342 */
1343 pandecode_log("uint32_t %s_%d[] = {\n", name, job_no);
1344 pandecode_indent++;
1345
1346 for (int j = 0; j <= ubufs[i].size; j++) {
1347 for (int k = 0; k < 4; k++) {
1348 if (k == 0)
1349 pandecode_log("0x%"PRIx32", ", ubuf[4 * j + k]);
1350 else
1351 pandecode_log_cont("0x%"PRIx32", ", ubuf[4 * j + k]);
1352
1353 }
1354
1355 pandecode_log_cont("\n");
1356 }
1357
1358 pandecode_indent--;
1359 pandecode_log("};\n");
1360 }
1361
1362 pandecode_log("struct mali_uniform_buffer_meta uniform_buffers_%d[] = {\n",
1363 job_no);
1364 pandecode_indent++;
1365
1366 for (int i = 0; i < ubufs_count; i++) {
1367 pandecode_log("{\n");
1368 pandecode_indent++;
1369 pandecode_prop("size = MALI_POSITIVE(%d)", ubufs[i].size + 1);
1370 pandecode_prop("ptr = ubuf_%d_%d_p >> 2", i, job_no);
1371 pandecode_indent--;
1372 pandecode_log("},\n");
1373 }
1374
1375 pandecode_indent--;
1376 pandecode_log("};\n");
1377 }
1378
1379 static void
1380 pandecode_replay_scratchpad(uintptr_t pscratchpad, int job_no, char *suffix)
1381 {
1382
1383 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(pscratchpad);
1384
1385 struct bifrost_scratchpad *PANDECODE_PTR_VAR(scratchpad, mem, pscratchpad);
1386
1387 if (scratchpad->zero)
1388 pandecode_msg("XXX scratchpad zero tripped");
1389
1390 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64"_%d%s = {\n", pscratchpad, job_no, suffix);
1391 pandecode_indent++;
1392
1393 pandecode_prop("flags = 0x%x", scratchpad->flags);
1394 MEMORY_PROP(scratchpad, gpu_scratchpad);
1395
1396 pandecode_indent--;
1397 pandecode_log("};\n");
1398 }
1399
1400 static void
1401 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1402 bool is_bifrost)
1403 {
1404 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1405 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1406
1407 /* Compute maximum possible size */
1408 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1409
1410 /* Print some boilerplate to clearly denote the assembly (which doesn't
1411 * obey indentation rules), and actually do the disassembly! */
1412
1413 printf("\n\n");
1414
1415 if (is_bifrost) {
1416 disassemble_bifrost(code, sz, false);
1417 } else {
1418 disassemble_midgard(code, sz);
1419 }
1420
1421 printf("\n\n");
1422 }
1423
1424 static void
1425 pandecode_replay_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix *p,
1426 int job_no, enum mali_job_type job_type,
1427 char *suffix, bool is_bifrost)
1428 {
1429 mali_ptr shader_meta_ptr = (u64) (uintptr_t) (p->_shader_upper << 4);
1430 struct pandecode_mapped_memory *attr_mem;
1431
1432 unsigned rt_count = 1;
1433
1434 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1435 * are the only things actually needed from the FBD, vertex/tiler jobs
1436 * no longer reference the FBD -- instead, this field points to some
1437 * info about the scratchpad.
1438 */
1439 if (is_bifrost)
1440 pandecode_replay_scratchpad(p->framebuffer & ~FBD_TYPE, job_no, suffix);
1441 else if (p->framebuffer & MALI_MFBD)
1442 rt_count = pandecode_replay_mfbd_bfr((u64) ((uintptr_t) p->framebuffer) & FBD_MASK, job_no, false);
1443 else if (job_type == JOB_TYPE_COMPUTE)
1444 pandecode_compute_fbd((u64) (uintptr_t) p->framebuffer, job_no);
1445 else
1446 pandecode_replay_sfbd((u64) (uintptr_t) p->framebuffer, job_no);
1447
1448 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
1449 int texture_count = 0, sampler_count = 0;
1450
1451 if (shader_meta_ptr) {
1452 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(shader_meta_ptr);
1453 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, shader_meta_ptr);
1454
1455 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", shader_meta_ptr, job_no, suffix);
1456 pandecode_indent++;
1457
1458 /* Save for dumps */
1459 attribute_count = s->attribute_count;
1460 varying_count = s->varying_count;
1461 texture_count = s->texture_count;
1462 sampler_count = s->sampler_count;
1463
1464 if (is_bifrost) {
1465 uniform_count = s->bifrost2.uniform_count;
1466 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
1467 } else {
1468 uniform_count = s->midgard1.uniform_count;
1469 uniform_buffer_count = s->midgard1.uniform_buffer_count;
1470 }
1471
1472 mali_ptr shader_ptr = pandecode_replay_shader_address("shader", s->shader);
1473
1474 pandecode_prop("texture_count = %" PRId16, s->texture_count);
1475 pandecode_prop("sampler_count = %" PRId16, s->sampler_count);
1476 pandecode_prop("attribute_count = %" PRId16, s->attribute_count);
1477 pandecode_prop("varying_count = %" PRId16, s->varying_count);
1478
1479 if (is_bifrost) {
1480 pandecode_log(".bifrost1 = {\n");
1481 pandecode_indent++;
1482
1483 pandecode_prop("uniform_buffer_count = %" PRId32, s->bifrost1.uniform_buffer_count);
1484 pandecode_prop("unk1 = 0x%" PRIx32, s->bifrost1.unk1);
1485
1486 pandecode_indent--;
1487 pandecode_log("},\n");
1488 } else {
1489 pandecode_log(".midgard1 = {\n");
1490 pandecode_indent++;
1491
1492 pandecode_prop("uniform_count = %" PRId16, s->midgard1.uniform_count);
1493 pandecode_prop("uniform_buffer_count = %" PRId16, s->midgard1.uniform_buffer_count);
1494 pandecode_prop("work_count = %" PRId16, s->midgard1.work_count);
1495
1496 pandecode_log(".flags = ");
1497 pandecode_log_decoded_flags(shader_midgard1_flag_info, s->midgard1.flags);
1498 pandecode_log_cont(",\n");
1499
1500 pandecode_prop("unknown2 = 0x%" PRIx32, s->midgard1.unknown2);
1501
1502 pandecode_indent--;
1503 pandecode_log("},\n");
1504 }
1505
1506 if (s->depth_units || s->depth_factor) {
1507 pandecode_prop("depth_factor = %f", s->depth_factor);
1508 pandecode_prop("depth_units = %f", s->depth_units);
1509 }
1510
1511 if (s->alpha_coverage) {
1512 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
1513 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
1514
1515 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
1516 invert_alpha_coverage ? "~" : "",
1517 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
1518 }
1519
1520 if (s->unknown2_3 || s->unknown2_4) {
1521 pandecode_log(".unknown2_3 = ");
1522
1523 int unknown2_3 = s->unknown2_3;
1524 int unknown2_4 = s->unknown2_4;
1525
1526 /* We're not quite sure what these flags mean without the depth test, if anything */
1527
1528 if (unknown2_3 & (MALI_DEPTH_TEST | MALI_DEPTH_FUNC_MASK)) {
1529 const char *func = pandecode_func_name(MALI_GET_DEPTH_FUNC(unknown2_3));
1530 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
1531
1532 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
1533 }
1534
1535 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
1536 pandecode_log_cont(",\n");
1537
1538 pandecode_log(".unknown2_4 = ");
1539 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
1540 pandecode_log_cont(",\n");
1541 }
1542
1543 if (s->stencil_mask_front || s->stencil_mask_back) {
1544 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
1545 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
1546 }
1547
1548 pandecode_replay_stencil("front", &s->stencil_front);
1549 pandecode_replay_stencil("back", &s->stencil_back);
1550
1551 if (is_bifrost) {
1552 pandecode_log(".bifrost2 = {\n");
1553 pandecode_indent++;
1554
1555 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
1556 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
1557 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
1558 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
1559
1560 pandecode_indent--;
1561 pandecode_log("},\n");
1562 } else if (s->midgard2.unknown2_7) {
1563 pandecode_log(".midgard2 = {\n");
1564 pandecode_indent++;
1565
1566 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
1567 pandecode_indent--;
1568 pandecode_log("},\n");
1569 }
1570
1571 if (s->unknown2_8)
1572 pandecode_prop("unknown2_8 = 0x%" PRIx32, s->unknown2_8);
1573
1574 if (!is_bifrost) {
1575 /* TODO: Blend shaders routing/disasm */
1576
1577 union midgard_blend blend = s->blend;
1578 pandecode_midgard_blend(&blend, false);
1579 }
1580
1581 pandecode_indent--;
1582 pandecode_log("};\n");
1583
1584 /* MRT blend fields are used whenever MFBD is used, with
1585 * per-RT descriptors */
1586
1587 if (job_type == JOB_TYPE_TILER) {
1588 void* blend_base = (void *) (s + 1);
1589
1590 for (unsigned i = 0; i < rt_count; i++) {
1591 mali_ptr shader = 0;
1592
1593 if (is_bifrost)
1594 shader = pandecode_bifrost_blend(blend_base, job_no, i);
1595 else
1596 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
1597
1598 if (shader)
1599 pandecode_shader_disassemble(shader, job_no, job_type, false);
1600 }
1601 }
1602
1603 pandecode_shader_disassemble(shader_ptr, job_no, job_type, is_bifrost);
1604 } else
1605 pandecode_msg("<no shader>\n");
1606
1607 if (p->viewport) {
1608 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
1609 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
1610
1611 pandecode_log("struct mali_viewport viewport_%d%s = {\n", job_no, suffix);
1612 pandecode_indent++;
1613
1614 pandecode_prop("clip_minx = %f", f->clip_minx);
1615 pandecode_prop("clip_miny = %f", f->clip_miny);
1616 pandecode_prop("clip_minz = %f", f->clip_minz);
1617 pandecode_prop("clip_maxx = %f", f->clip_maxx);
1618 pandecode_prop("clip_maxy = %f", f->clip_maxy);
1619 pandecode_prop("clip_maxz = %f", f->clip_maxz);
1620
1621 /* Only the higher coordinates are MALI_POSITIVE scaled */
1622
1623 pandecode_prop("viewport0 = { %d, %d }",
1624 f->viewport0[0], f->viewport0[1]);
1625
1626 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
1627 f->viewport1[0] + 1, f->viewport1[1] + 1);
1628
1629 pandecode_indent--;
1630 pandecode_log("};\n");
1631 }
1632
1633 if (p->attribute_meta) {
1634 unsigned max_attr_index = pandecode_replay_attribute_meta(job_no, attribute_count, p, false, suffix);
1635
1636 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
1637 pandecode_replay_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index + 1, false);
1638 }
1639
1640 /* Varyings are encoded like attributes but not actually sent; we just
1641 * pass a zero buffer with the right stride/size set, (or whatever)
1642 * since the GPU will write to it itself */
1643
1644 if (p->varyings) {
1645 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
1646
1647 /* Number of descriptors depends on whether there are
1648 * non-internal varyings */
1649
1650 pandecode_replay_attributes(attr_mem, p->varyings, job_no, suffix, varying_count > 1 ? 4 : 1, true);
1651 }
1652
1653 if (p->varying_meta) {
1654 pandecode_replay_attribute_meta(job_no, varying_count, p, true, suffix);
1655 }
1656
1657 bool is_compute = job_type == JOB_TYPE_COMPUTE;
1658
1659 if (p->uniforms && !is_compute) {
1660 int rows = uniform_count, width = 4;
1661 size_t sz = rows * width * sizeof(float);
1662
1663 struct pandecode_mapped_memory *uniform_mem = pandecode_find_mapped_gpu_mem_containing(p->uniforms);
1664 pandecode_fetch_gpu_mem(uniform_mem, p->uniforms, sz);
1665 u32 *PANDECODE_PTR_VAR(uniforms, uniform_mem, p->uniforms);
1666
1667 pandecode_log("u32 uniforms_%d%s[] = {\n", job_no, suffix);
1668
1669 pandecode_indent++;
1670
1671 for (int row = 0; row < rows; row++) {
1672 for (int i = 0; i < width; i++) {
1673 u32 v = uniforms[i];
1674 float f;
1675 memcpy(&f, &v, sizeof(v));
1676 pandecode_log_cont("%X /* %f */, ", v, f);
1677 }
1678
1679 pandecode_log_cont("\n");
1680
1681 uniforms += width;
1682 }
1683
1684 pandecode_indent--;
1685 pandecode_log("};\n");
1686 } else if (p->uniforms) {
1687 int rows = uniform_count * 2;
1688 size_t sz = rows * sizeof(mali_ptr);
1689
1690 struct pandecode_mapped_memory *uniform_mem = pandecode_find_mapped_gpu_mem_containing(p->uniforms);
1691 pandecode_fetch_gpu_mem(uniform_mem, p->uniforms, sz);
1692 mali_ptr *PANDECODE_PTR_VAR(uniforms, uniform_mem, p->uniforms);
1693
1694 pandecode_log("mali_ptr uniforms_%d%s[] = {\n", job_no, suffix);
1695
1696 pandecode_indent++;
1697
1698 for (int row = 0; row < rows; row++) {
1699 char *a = pointer_as_memory_reference(uniforms[row]);
1700 pandecode_log("%s,\n", a);
1701 free(a);
1702 }
1703
1704 pandecode_indent--;
1705 pandecode_log("};\n");
1706
1707 }
1708
1709 if (p->uniform_buffers) {
1710 pandecode_replay_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
1711 }
1712
1713 if (p->texture_trampoline) {
1714 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(p->texture_trampoline);
1715
1716 if (mmem) {
1717 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline);
1718
1719 pandecode_log("uint64_t texture_trampoline_%d[] = {\n", job_no);
1720 pandecode_indent++;
1721
1722 for (int tex = 0; tex < texture_count; ++tex) {
1723 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
1724 char *a = pointer_as_memory_reference(*u);
1725 pandecode_log("%s,\n", a);
1726 free(a);
1727 }
1728
1729 pandecode_indent--;
1730 pandecode_log("};\n");
1731
1732 /* Now, finally, descend down into the texture descriptor */
1733 for (int tex = 0; tex < texture_count; ++tex) {
1734 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
1735 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
1736
1737 if (tmem) {
1738 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, *u);
1739
1740 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", *u, job_no, tex);
1741 pandecode_indent++;
1742
1743 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", t->width + 1);
1744 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", t->height + 1);
1745 pandecode_prop("depth = MALI_POSITIVE(%" PRId16 ")", t->depth + 1);
1746 pandecode_prop("array_size = MALI_POSITIVE(%" PRId16 ")", t->array_size + 1);
1747 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
1748 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
1749 pandecode_prop("nr_mipmap_levels = %" PRId8, t->nr_mipmap_levels);
1750
1751 struct mali_texture_format f = t->format;
1752
1753 pandecode_log(".format = {\n");
1754 pandecode_indent++;
1755
1756 pandecode_replay_swizzle(f.swizzle);
1757 pandecode_prop("format = %s", pandecode_format_name(f.format));
1758 pandecode_prop("type = %s", pandecode_texture_type(f.type));
1759 pandecode_prop("srgb = %" PRId32, f.srgb);
1760 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
1761 pandecode_prop("usage2 = 0x%" PRIx32, f.usage2);
1762
1763 pandecode_indent--;
1764 pandecode_log("},\n");
1765
1766 pandecode_replay_swizzle(t->swizzle);
1767
1768 if (t->swizzle_zero) {
1769 /* Shouldn't happen */
1770 pandecode_msg("Swizzle zero tripped but replay will be fine anyway");
1771 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
1772 }
1773
1774 pandecode_prop("unknown3 = 0x%" PRIx32, t->unknown3);
1775
1776 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
1777 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
1778 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
1779
1780 pandecode_log(".payload = {\n");
1781 pandecode_indent++;
1782
1783 /* A bunch of bitmap pointers follow.
1784 * We work out the correct number,
1785 * based on the mipmap/cubemap
1786 * properties, but dump extra
1787 * possibilities to futureproof */
1788
1789 int bitmap_count = MALI_NEGATIVE(t->nr_mipmap_levels);
1790 bool manual_stride = f.usage2 & MALI_TEX_MANUAL_STRIDE;
1791
1792 /* Miptree for each face */
1793 if (f.type == MALI_TEX_CUBE)
1794 bitmap_count *= 6;
1795
1796 /* Array of textures */
1797 bitmap_count *= MALI_NEGATIVE(t->array_size);
1798
1799 /* Stride for each element */
1800 if (manual_stride)
1801 bitmap_count *= 2;
1802
1803 /* Sanity check the size */
1804 int max_count = sizeof(t->payload) / sizeof(t->payload[0]);
1805 assert (bitmap_count <= max_count);
1806
1807 /* Dump more to be safe, but not _that_ much more */
1808 int safe_count = MIN2(bitmap_count * 2, max_count);
1809
1810 for (int i = 0; i < safe_count; ++i) {
1811 char *prefix = (i >= bitmap_count) ? "// " : "";
1812
1813 /* How we dump depends if this is a stride or a pointer */
1814
1815 if ((f.usage2 & MALI_TEX_MANUAL_STRIDE) && (i & 1)) {
1816 /* signed 32-bit snuck in as a 64-bit pointer */
1817 uint64_t stride_set = t->payload[i];
1818 uint32_t clamped_stride = stride_set;
1819 int32_t stride = clamped_stride;
1820 assert(stride_set == clamped_stride);
1821 pandecode_log("%s(mali_ptr) %d /* stride */, \n", prefix, stride);
1822 } else {
1823 char *a = pointer_as_memory_reference(t->payload[i]);
1824 pandecode_log("%s%s, \n", prefix, a);
1825 free(a);
1826 }
1827 }
1828
1829 pandecode_indent--;
1830 pandecode_log("},\n");
1831
1832 pandecode_indent--;
1833 pandecode_log("};\n");
1834 }
1835 }
1836 }
1837 }
1838
1839 if (p->sampler_descriptor) {
1840 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->sampler_descriptor);
1841
1842 if (smem) {
1843 struct mali_sampler_descriptor *s;
1844
1845 mali_ptr d = p->sampler_descriptor;
1846
1847 for (int i = 0; i < sampler_count; ++i) {
1848 s = pandecode_fetch_gpu_mem(smem, d + sizeof(*s) * i, sizeof(*s));
1849
1850 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%d_%d = {\n", job_no, i);
1851 pandecode_indent++;
1852
1853 /* Only the lower two bits are understood right now; the rest we display as hex */
1854 pandecode_log(".filter_mode = MALI_TEX_MIN(%s) | MALI_TEX_MAG(%s) | 0x%" PRIx32",\n",
1855 MALI_FILTER_NAME(s->filter_mode & MALI_TEX_MIN_MASK),
1856 MALI_FILTER_NAME(s->filter_mode & MALI_TEX_MAG_MASK),
1857 s->filter_mode & ~3);
1858
1859 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
1860 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
1861
1862 pandecode_prop("wrap_s = %s", pandecode_wrap_mode_name(s->wrap_s));
1863 pandecode_prop("wrap_t = %s", pandecode_wrap_mode_name(s->wrap_t));
1864 pandecode_prop("wrap_r = %s", pandecode_wrap_mode_name(s->wrap_r));
1865
1866 pandecode_prop("compare_func = %s", pandecode_alt_func_name(s->compare_func));
1867
1868 if (s->zero || s->zero2) {
1869 pandecode_msg("Zero tripped\n");
1870 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
1871 }
1872
1873 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
1874
1875 pandecode_prop("border_color = { %f, %f, %f, %f }",
1876 s->border_color[0],
1877 s->border_color[1],
1878 s->border_color[2],
1879 s->border_color[3]);
1880
1881 pandecode_indent--;
1882 pandecode_log("};\n");
1883 }
1884 }
1885 }
1886 }
1887
1888 static void
1889 pandecode_replay_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
1890 {
1891 pandecode_log_cont("{\n");
1892 pandecode_indent++;
1893
1894 MEMORY_PROP(p, position_varying);
1895 DYN_MEMORY_PROP(p, job_no, uniform_buffers);
1896 DYN_MEMORY_PROP(p, job_no, texture_trampoline);
1897 DYN_MEMORY_PROP(p, job_no, sampler_descriptor);
1898 DYN_MEMORY_PROP(p, job_no, uniforms);
1899 DYN_MEMORY_PROP(p, job_no, attributes);
1900 DYN_MEMORY_PROP(p, job_no, attribute_meta);
1901 DYN_MEMORY_PROP(p, job_no, varyings);
1902 DYN_MEMORY_PROP(p, job_no, varying_meta);
1903 DYN_MEMORY_PROP(p, job_no, viewport);
1904 DYN_MEMORY_PROP(p, job_no, occlusion_counter);
1905
1906 if (is_bifrost)
1907 pandecode_prop("framebuffer = scratchpad_%d_p", job_no);
1908 else
1909 pandecode_prop("framebuffer = framebuffer_%d_p | %s", job_no, p->framebuffer & MALI_MFBD ? "MALI_MFBD" : "0");
1910
1911 pandecode_prop("_shader_upper = (shader_meta_%d_p) >> 4", job_no);
1912 pandecode_prop("flags = %d", p->flags);
1913
1914 pandecode_indent--;
1915 pandecode_log("},\n");
1916 }
1917
1918 static void
1919 pandecode_replay_vertex_only_bfr(struct bifrost_vertex_only *v)
1920 {
1921 pandecode_log_cont("{\n");
1922 pandecode_indent++;
1923
1924 pandecode_prop("unk2 = 0x%x", v->unk2);
1925
1926 if (v->zero0 || v->zero1) {
1927 pandecode_msg("vertex only zero tripped");
1928 pandecode_prop("zero0 = 0x%" PRIx32, v->zero0);
1929 pandecode_prop("zero1 = 0x%" PRIx64, v->zero1);
1930 }
1931
1932 pandecode_indent--;
1933 pandecode_log("}\n");
1934 }
1935
1936 static void
1937 pandecode_replay_tiler_heap_meta(mali_ptr gpu_va, int job_no)
1938 {
1939
1940 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1941 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
1942
1943 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no);
1944 pandecode_indent++;
1945
1946 if (h->zero) {
1947 pandecode_msg("tiler heap zero tripped\n");
1948 pandecode_prop("zero = 0x%x", h->zero);
1949 }
1950
1951 for (int i = 0; i < 12; i++) {
1952 if (h->zeros[i] != 0) {
1953 pandecode_msg("tiler heap zero %d tripped, value %x\n",
1954 i, h->zeros[i]);
1955 }
1956 }
1957
1958 pandecode_prop("heap_size = 0x%x", h->heap_size);
1959 MEMORY_PROP(h, tiler_heap_start);
1960 MEMORY_PROP(h, tiler_heap_free);
1961
1962 /* this might point to the beginning of another buffer, when it's
1963 * really the end of the tiler heap buffer, so we have to be careful
1964 * here.
1965 */
1966 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
1967 pandecode_prop("tiler_heap_end = %s + 1", a);
1968 free(a);
1969
1970 pandecode_indent--;
1971 pandecode_log("};\n");
1972 }
1973
1974 static void
1975 pandecode_replay_tiler_meta(mali_ptr gpu_va, int job_no)
1976 {
1977 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1978 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
1979
1980 pandecode_replay_tiler_heap_meta(t->tiler_heap_meta, job_no);
1981
1982 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no);
1983 pandecode_indent++;
1984
1985 if (t->zero0 || t->zero1) {
1986 pandecode_msg("tiler meta zero tripped");
1987 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
1988 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
1989 }
1990
1991 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
1992 pandecode_prop("flags = 0x%" PRIx16, t->flags);
1993
1994 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
1995 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
1996 DYN_MEMORY_PROP(t, job_no, tiler_heap_meta);
1997
1998 for (int i = 0; i < 12; i++) {
1999 if (t->zeros[i] != 0) {
2000 pandecode_msg("tiler heap zero %d tripped, value %" PRIx64 "\n",
2001 i, t->zeros[i]);
2002 }
2003 }
2004
2005 pandecode_indent--;
2006 pandecode_log("};\n");
2007 }
2008
2009 static void
2010 pandecode_replay_gl_enables(uint32_t gl_enables, int job_type)
2011 {
2012 pandecode_log(".gl_enables = ");
2013
2014 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2015
2016 pandecode_log_cont(",\n");
2017 }
2018
2019 static void
2020 pandecode_replay_primitive_size(union midgard_primitive_size u, bool constant)
2021 {
2022 if (u.pointer == 0x0)
2023 return;
2024
2025 pandecode_log(".primitive_size = {\n");
2026 pandecode_indent++;
2027
2028 if (constant) {
2029 pandecode_prop("constant = %f", u.constant);
2030 } else {
2031 MEMORY_PROP((&u), pointer);
2032 }
2033
2034 pandecode_indent--;
2035 pandecode_log("},\n");
2036 }
2037
2038 static void
2039 pandecode_replay_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2040 {
2041 pandecode_log_cont("{\n");
2042 pandecode_indent++;
2043
2044 /* TODO: gl_PointSize on Bifrost */
2045 pandecode_replay_primitive_size(t->primitive_size, true);
2046
2047 DYN_MEMORY_PROP(t, job_no, tiler_meta);
2048 pandecode_replay_gl_enables(t->gl_enables, JOB_TYPE_TILER);
2049
2050 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2051 || t->zero6 || t->zero7 || t->zero8) {
2052 pandecode_msg("tiler only zero tripped");
2053 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2054 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2055 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2056 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2057 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2058 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2059 pandecode_prop("zero7 = 0x%" PRIx32, t->zero7);
2060 pandecode_prop("zero8 = 0x%" PRIx64, t->zero8);
2061 }
2062
2063 pandecode_indent--;
2064 pandecode_log("},\n");
2065 }
2066
2067 static int
2068 pandecode_replay_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2069 const struct pandecode_mapped_memory *mem,
2070 mali_ptr payload, int job_no)
2071 {
2072 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2073
2074 pandecode_replay_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true);
2075
2076 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no);
2077 pandecode_indent++;
2078
2079 pandecode_log(".prefix = ");
2080 pandecode_replay_vertex_tiler_prefix(&v->prefix, job_no);
2081
2082 pandecode_log(".vertex = ");
2083 pandecode_replay_vertex_only_bfr(&v->vertex);
2084
2085 pandecode_log(".postfix = ");
2086 pandecode_replay_vertex_tiler_postfix(&v->postfix, job_no, true);
2087
2088 pandecode_indent--;
2089 pandecode_log("};\n");
2090
2091 return sizeof(*v);
2092 }
2093
2094 static int
2095 pandecode_replay_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2096 const struct pandecode_mapped_memory *mem,
2097 mali_ptr payload, int job_no)
2098 {
2099 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2100
2101 pandecode_replay_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true);
2102
2103 pandecode_replay_indices(t->prefix.indices, t->prefix.index_count, job_no);
2104 pandecode_replay_tiler_meta(t->tiler.tiler_meta, job_no);
2105
2106 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no);
2107 pandecode_indent++;
2108
2109 pandecode_log(".prefix = ");
2110 pandecode_replay_vertex_tiler_prefix(&t->prefix, job_no);
2111
2112 pandecode_log(".tiler = ");
2113 pandecode_replay_tiler_only_bfr(&t->tiler, job_no);
2114
2115 pandecode_log(".postfix = ");
2116 pandecode_replay_vertex_tiler_postfix(&t->postfix, job_no, true);
2117
2118 pandecode_indent--;
2119 pandecode_log("};\n");
2120
2121 return sizeof(*t);
2122 }
2123
2124 static int
2125 pandecode_replay_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2126 const struct pandecode_mapped_memory *mem,
2127 mali_ptr payload, int job_no)
2128 {
2129 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2130
2131 pandecode_replay_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false);
2132
2133 pandecode_replay_indices(v->prefix.indices, v->prefix.index_count, job_no);
2134
2135 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2136 pandecode_indent++;
2137
2138 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2139 pandecode_replay_primitive_size(v->primitive_size, !has_primitive_pointer);
2140
2141 pandecode_log(".prefix = ");
2142 pandecode_replay_vertex_tiler_prefix(&v->prefix, job_no);
2143
2144 pandecode_replay_gl_enables(v->gl_enables, h->job_type);
2145
2146 if (v->instance_shift || v->instance_odd) {
2147 pandecode_prop("instance_shift = 0x%d /* %d */",
2148 v->instance_shift, 1 << v->instance_shift);
2149 pandecode_prop("instance_odd = 0x%X /* %d */",
2150 v->instance_odd, (2 * v->instance_odd) + 1);
2151
2152 pandecode_padded_vertices(v->instance_shift, v->instance_odd);
2153 }
2154
2155 if (v->draw_start)
2156 pandecode_prop("draw_start = %d", v->draw_start);
2157
2158 #ifndef __LP64__
2159
2160 if (v->zero3) {
2161 pandecode_msg("Zero tripped\n");
2162 pandecode_prop("zero3 = 0x%" PRIx32, v->zero3);
2163 }
2164
2165 #endif
2166
2167 if (v->zero5) {
2168 pandecode_msg("Zero tripped\n");
2169 pandecode_prop("zero5 = 0x%" PRIx64, v->zero5);
2170 }
2171
2172 pandecode_log(".postfix = ");
2173 pandecode_replay_vertex_tiler_postfix(&v->postfix, job_no, false);
2174
2175 pandecode_indent--;
2176 pandecode_log("};\n");
2177
2178 return sizeof(*v);
2179 }
2180
2181 static int
2182 pandecode_replay_fragment_job(const struct pandecode_mapped_memory *mem,
2183 mali_ptr payload, int job_no,
2184 bool is_bifrost)
2185 {
2186 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2187
2188 bool fbd_dumped = false;
2189
2190 if (!is_bifrost && (s->framebuffer & FBD_TYPE) == MALI_SFBD) {
2191 /* Only SFBDs are understood, not MFBDs. We're speculating,
2192 * based on the versioning, kernel code, etc, that the
2193 * difference is between Single FrameBuffer Descriptor and
2194 * Multiple FrmaeBuffer Descriptor; the change apparently lines
2195 * up with multi-framebuffer support being added (T7xx onwards,
2196 * including Gxx). In any event, there's some field shuffling
2197 * that we haven't looked into yet. */
2198
2199 pandecode_replay_sfbd(s->framebuffer & FBD_MASK, job_no);
2200 fbd_dumped = true;
2201 } else if ((s->framebuffer & FBD_TYPE) == MALI_MFBD) {
2202 /* We don't know if Bifrost supports SFBD's at all, since the
2203 * driver never uses them. And the format is different from
2204 * Midgard anyways, due to the tiler heap and scratchpad being
2205 * moved out into separate structures, so it's not clear what a
2206 * Bifrost SFBD would even look like without getting an actual
2207 * trace, which appears impossible.
2208 */
2209
2210 pandecode_replay_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true);
2211 fbd_dumped = true;
2212 }
2213
2214 uintptr_t p = (uintptr_t) s->framebuffer & FBD_MASK;
2215 pandecode_log("struct mali_payload_fragment payload_%"PRIx64"_%d = {\n", payload, job_no);
2216 pandecode_indent++;
2217
2218 /* See the comments by the macro definitions for mathematical context
2219 * on why this is so weird */
2220
2221 if (MALI_TILE_COORD_FLAGS(s->max_tile_coord) || MALI_TILE_COORD_FLAGS(s->min_tile_coord))
2222 pandecode_msg("Tile coordinate flag missed, replay wrong\n");
2223
2224 pandecode_prop("min_tile_coord = MALI_COORDINATE_TO_TILE_MIN(%d, %d)",
2225 MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT,
2226 MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT);
2227
2228 pandecode_prop("max_tile_coord = MALI_COORDINATE_TO_TILE_MAX(%d, %d)",
2229 (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT,
2230 (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT);
2231
2232 /* If the FBD was just decoded, we can refer to it by pointer. If not,
2233 * we have to fallback on offsets. */
2234
2235 const char *fbd_type = s->framebuffer & MALI_MFBD ? "MALI_MFBD" : "MALI_SFBD";
2236
2237 if (fbd_dumped)
2238 pandecode_prop("framebuffer = framebuffer_%d_p | %s", job_no, fbd_type);
2239 else
2240 pandecode_prop("framebuffer = %s | %s", pointer_as_memory_reference(p), fbd_type);
2241
2242 pandecode_indent--;
2243 pandecode_log("};\n");
2244
2245 return sizeof(*s);
2246 }
2247
2248 static int job_descriptor_number = 0;
2249
2250 int
2251 pandecode_replay_jc(mali_ptr jc_gpu_va, bool bifrost)
2252 {
2253 struct mali_job_descriptor_header *h;
2254
2255 int start_number = 0;
2256
2257 bool first = true;
2258 bool last_size;
2259
2260 do {
2261 struct pandecode_mapped_memory *mem =
2262 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2263
2264 void *payload;
2265
2266 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2267
2268 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2269 * high 32-bits of the 64-bit pointer are reused to store
2270 * something else.
2271 */
2272 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2273 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
2274 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2275
2276 payload = pandecode_fetch_gpu_mem(mem, payload_ptr,
2277 MALI_PAYLOAD_SIZE);
2278
2279 int job_no = job_descriptor_number++;
2280
2281 if (first)
2282 start_number = job_no;
2283
2284 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2285 pandecode_indent++;
2286
2287 pandecode_prop("job_type = %s", pandecode_job_type_name(h->job_type));
2288
2289 /* Save for next job fixing */
2290 last_size = h->job_descriptor_size;
2291
2292 if (h->job_descriptor_size)
2293 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2294
2295 if (h->exception_status != 0x1)
2296 pandecode_prop("exception_status = %x (source ID: 0x%x access: 0x%x exception: 0x%x)",
2297 h->exception_status,
2298 (h->exception_status >> 16) & 0xFFFF,
2299 (h->exception_status >> 8) & 0x3,
2300 h->exception_status & 0xFF);
2301
2302 if (h->first_incomplete_task)
2303 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2304
2305 if (h->fault_pointer)
2306 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2307
2308 if (h->job_barrier)
2309 pandecode_prop("job_barrier = %d", h->job_barrier);
2310
2311 pandecode_prop("job_index = %d", h->job_index);
2312
2313 if (h->unknown_flags)
2314 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2315
2316 if (h->job_dependency_index_1)
2317 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2318
2319 if (h->job_dependency_index_2)
2320 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2321
2322 pandecode_indent--;
2323 pandecode_log("};\n");
2324
2325 /* Do not touch the field yet -- decode the payload first, and
2326 * don't touch that either. This is essential for the uploads
2327 * to occur in sequence and therefore be dynamically allocated
2328 * correctly. Do note the size, however, for that related
2329 * reason. */
2330
2331 switch (h->job_type) {
2332 case JOB_TYPE_SET_VALUE: {
2333 struct mali_payload_set_value *s = payload;
2334 pandecode_log("struct mali_payload_set_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2335 pandecode_indent++;
2336 MEMORY_PROP(s, out);
2337 pandecode_prop("unknown = 0x%" PRIX64, s->unknown);
2338 pandecode_indent--;
2339 pandecode_log("};\n");
2340
2341 break;
2342 }
2343
2344 case JOB_TYPE_TILER:
2345 case JOB_TYPE_VERTEX:
2346 case JOB_TYPE_COMPUTE:
2347 if (bifrost) {
2348 if (h->job_type == JOB_TYPE_TILER)
2349 pandecode_replay_tiler_job_bfr(h, mem, payload_ptr, job_no);
2350 else
2351 pandecode_replay_vertex_job_bfr(h, mem, payload_ptr, job_no);
2352 } else
2353 pandecode_replay_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no);
2354
2355 break;
2356
2357 case JOB_TYPE_FRAGMENT:
2358 pandecode_replay_fragment_job(mem, payload_ptr, job_no, bifrost);
2359 break;
2360
2361 default:
2362 break;
2363 }
2364
2365 /* Handle linkage */
2366
2367 if (!first) {
2368 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no - 1);
2369
2370 if (last_size)
2371 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no);
2372 else
2373 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no);
2374 }
2375
2376 first = false;
2377
2378 } while ((jc_gpu_va = h->job_descriptor_size ? h->next_job_64 : h->next_job_32));
2379
2380 return start_number;
2381 }