projects / mesa.git / blob
? search:


90b2579a436357ebcbaf3a95002d69eb4841aee3
[mesa.git] / src / gallium / drivers / panfrost / pandecode / decode.c
1 /*
2 * Copyright (C) 2017-2019 Alyssa Rosenzweig
3 * Copyright (C) 2017-2019 Connor Abbott
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24
25 #include <panfrost-job.h>
26 #include <stdio.h>
27 #include <stdlib.h>
28 #include <memory.h>
29 #include <stdbool.h>
30 #include <stdarg.h>
31 #include "decode.h"
32 #include "util/u_math.h"
33
34 #include "../pan_pretty_print.h"
35 #include "../midgard/disassemble.h"
36 #include "../bifrost/disassemble.h"
37 int pandecode_replay_jc(mali_ptr jc_gpu_va, bool bifrost);
38
39 #define MEMORY_PROP(obj, p) {\
40 char *a = pointer_as_memory_reference(obj->p); \
41 pandecode_prop("%s = %s", #p, a); \
42 free(a); \
43 }
44
45 #define MEMORY_COMMENT(obj, p) {\
46 char *a = pointer_as_memory_reference(obj->p); \
47 pandecode_msg("%s = %s\n", #p, a); \
48 free(a); \
49 }
50
51 #define DYN_MEMORY_PROP(obj, no, p) { \
52 if (obj->p) \
53 pandecode_prop("%s = %s_%d_p", #p, #p, no); \
54 }
55
56 /* Semantic logging type.
57 *
58 * Raw: for raw messages to be printed as is.
59 * Message: for helpful information to be commented out in replays.
60 * Property: for properties of a struct
61 *
62 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
63 */
64
65 enum pandecode_log_type {
66 PANDECODE_RAW,
67 PANDECODE_MESSAGE,
68 PANDECODE_PROPERTY
69 };
70
71 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
72 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
73 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
74
75 unsigned pandecode_indent = 0;
76
77 static void
78 pandecode_make_indent(void)
79 {
80 for (unsigned i = 0; i < pandecode_indent; ++i)
81 printf(" ");
82 }
83
84 static void
85 pandecode_log_typed(enum pandecode_log_type type, const char *format, ...)
86 {
87 va_list ap;
88
89 pandecode_make_indent();
90
91 if (type == PANDECODE_MESSAGE)
92 printf("// ");
93 else if (type == PANDECODE_PROPERTY)
94 printf(".");
95
96 va_start(ap, format);
97 vprintf(format, ap);
98 va_end(ap);
99
100 if (type == PANDECODE_PROPERTY)
101 printf(",\n");
102 }
103
104 static void
105 pandecode_log_cont(const char *format, ...)
106 {
107 va_list ap;
108
109 va_start(ap, format);
110 vprintf(format, ap);
111 va_end(ap);
112 }
113
114 struct pandecode_flag_info {
115 u64 flag;
116 const char *name;
117 };
118
119 static void
120 pandecode_log_decoded_flags(const struct pandecode_flag_info *flag_info,
121 u64 flags)
122 {
123 bool decodable_flags_found = false;
124
125 for (int i = 0; flag_info[i].name; i++) {
126 if ((flags & flag_info[i].flag) != flag_info[i].flag)
127 continue;
128
129 if (!decodable_flags_found) {
130 decodable_flags_found = true;
131 } else {
132 pandecode_log_cont(" | ");
133 }
134
135 pandecode_log_cont("%s", flag_info[i].name);
136
137 flags &= ~flag_info[i].flag;
138 }
139
140 if (decodable_flags_found) {
141 if (flags)
142 pandecode_log_cont(" | 0x%" PRIx64, flags);
143 } else {
144 pandecode_log_cont("0x%" PRIx64, flags);
145 }
146 }
147
148 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
149 static const struct pandecode_flag_info gl_enable_flag_info[] = {
150 FLAG_INFO(OCCLUSION_QUERY),
151 FLAG_INFO(OCCLUSION_PRECISE),
152 FLAG_INFO(FRONT_CCW_TOP),
153 FLAG_INFO(CULL_FACE_FRONT),
154 FLAG_INFO(CULL_FACE_BACK),
155 {}
156 };
157 #undef FLAG_INFO
158
159 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
160 static const struct pandecode_flag_info clear_flag_info[] = {
161 FLAG_INFO(FAST),
162 FLAG_INFO(SLOW),
163 FLAG_INFO(SLOW_STENCIL),
164 {}
165 };
166 #undef FLAG_INFO
167
168 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
169 static const struct pandecode_flag_info mask_flag_info[] = {
170 FLAG_INFO(R),
171 FLAG_INFO(G),
172 FLAG_INFO(B),
173 FLAG_INFO(A),
174 {}
175 };
176 #undef FLAG_INFO
177
178 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
179 static const struct pandecode_flag_info u3_flag_info[] = {
180 FLAG_INFO(HAS_MSAA),
181 FLAG_INFO(CAN_DISCARD),
182 FLAG_INFO(HAS_BLEND_SHADER),
183 FLAG_INFO(DEPTH_TEST),
184 {}
185 };
186
187 static const struct pandecode_flag_info u4_flag_info[] = {
188 FLAG_INFO(NO_MSAA),
189 FLAG_INFO(NO_DITHER),
190 FLAG_INFO(DEPTH_RANGE_A),
191 FLAG_INFO(DEPTH_RANGE_B),
192 FLAG_INFO(STENCIL_TEST),
193 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER),
194 {}
195 };
196 #undef FLAG_INFO
197
198 #define FLAG_INFO(flag) { MALI_FRAMEBUFFER_##flag, "MALI_FRAMEBUFFER_" #flag }
199 static const struct pandecode_flag_info fb_fmt_flag_info[] = {
200 FLAG_INFO(MSAA_A),
201 FLAG_INFO(MSAA_B),
202 FLAG_INFO(MSAA_8),
203 {}
204 };
205 #undef FLAG_INFO
206
207 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
208 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
209 FLAG_INFO(AFBC),
210 FLAG_INFO(MSAA),
211 {}
212 };
213 #undef FLAG_INFO
214
215 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
216 static const struct pandecode_flag_info mfbd_extra_flag_info[] = {
217 FLAG_INFO(PRESENT),
218 FLAG_INFO(AFBC),
219 FLAG_INFO(ZS),
220 {}
221 };
222 #undef FLAG_INFO
223
224 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
225 static const struct pandecode_flag_info shader_unknown1_flag_info [] = {
226 FLAG_INFO(EARLY_Z),
227 FLAG_INFO(HELPER_INVOCATIONS),
228 FLAG_INFO(READS_TILEBUFFER),
229 FLAG_INFO(READS_ZS),
230 {}
231 };
232 #undef FLAG_INFO
233
234 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
235 static const struct pandecode_flag_info mfbd_flag_info [] = {
236 FLAG_INFO(DEPTH_WRITE),
237 FLAG_INFO(EXTRA),
238 {}
239 };
240 #undef FLAG_INFO
241
242
243 extern char *replace_fragment;
244 extern char *replace_vertex;
245
246 static char *
247 pandecode_job_type_name(enum mali_job_type type)
248 {
249 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
250
251 switch (type) {
252 DEFINE_CASE(NULL);
253 DEFINE_CASE(SET_VALUE);
254 DEFINE_CASE(CACHE_FLUSH);
255 DEFINE_CASE(COMPUTE);
256 DEFINE_CASE(VERTEX);
257 DEFINE_CASE(TILER);
258 DEFINE_CASE(FUSED);
259 DEFINE_CASE(FRAGMENT);
260
261 case JOB_NOT_STARTED:
262 return "NOT_STARTED";
263
264 default:
265 pandecode_log("Warning! Unknown job type %x\n", type);
266 return "!?!?!?";
267 }
268
269 #undef DEFINE_CASE
270 }
271
272 static char *
273 pandecode_draw_mode_name(enum mali_draw_mode mode)
274 {
275 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
276
277 switch (mode) {
278 DEFINE_CASE(DRAW_NONE);
279 DEFINE_CASE(POINTS);
280 DEFINE_CASE(LINES);
281 DEFINE_CASE(TRIANGLES);
282 DEFINE_CASE(TRIANGLE_STRIP);
283 DEFINE_CASE(TRIANGLE_FAN);
284 DEFINE_CASE(LINE_STRIP);
285 DEFINE_CASE(LINE_LOOP);
286 DEFINE_CASE(POLYGON);
287 DEFINE_CASE(QUADS);
288 DEFINE_CASE(QUAD_STRIP);
289
290 default:
291 return "MALI_TRIANGLES /* XXX: Unknown GL mode, check dump */";
292 }
293
294 #undef DEFINE_CASE
295 }
296
297 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
298 static char *
299 pandecode_func_name(enum mali_func mode)
300 {
301 switch (mode) {
302 DEFINE_CASE(NEVER);
303 DEFINE_CASE(LESS);
304 DEFINE_CASE(EQUAL);
305 DEFINE_CASE(LEQUAL);
306 DEFINE_CASE(GREATER);
307 DEFINE_CASE(NOTEQUAL);
308 DEFINE_CASE(GEQUAL);
309 DEFINE_CASE(ALWAYS);
310
311 default:
312 return "MALI_FUNC_NEVER /* XXX: Unknown function, check dump */";
313 }
314 }
315 #undef DEFINE_CASE
316
317 /* Why is this duplicated? Who knows... */
318 #define DEFINE_CASE(name) case MALI_ALT_FUNC_ ## name: return "MALI_ALT_FUNC_" #name
319 static char *
320 pandecode_alt_func_name(enum mali_alt_func mode)
321 {
322 switch (mode) {
323 DEFINE_CASE(NEVER);
324 DEFINE_CASE(LESS);
325 DEFINE_CASE(EQUAL);
326 DEFINE_CASE(LEQUAL);
327 DEFINE_CASE(GREATER);
328 DEFINE_CASE(NOTEQUAL);
329 DEFINE_CASE(GEQUAL);
330 DEFINE_CASE(ALWAYS);
331
332 default:
333 return "MALI_FUNC_NEVER /* XXX: Unknown function, check dump */";
334 }
335 }
336 #undef DEFINE_CASE
337
338 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
339 static char *
340 pandecode_stencil_op_name(enum mali_stencil_op op)
341 {
342 switch (op) {
343 DEFINE_CASE(KEEP);
344 DEFINE_CASE(REPLACE);
345 DEFINE_CASE(ZERO);
346 DEFINE_CASE(INVERT);
347 DEFINE_CASE(INCR_WRAP);
348 DEFINE_CASE(DECR_WRAP);
349 DEFINE_CASE(INCR);
350 DEFINE_CASE(DECR);
351
352 default:
353 return "MALI_STENCIL_KEEP /* XXX: Unknown stencil op, check dump */";
354 }
355 }
356
357 #undef DEFINE_CASE
358
359 #define DEFINE_CASE(name) case MALI_ATTR_ ## name: return "MALI_ATTR_" #name
360 static char *pandecode_attr_mode_name(enum mali_attr_mode mode)
361 {
362 switch(mode) {
363 DEFINE_CASE(UNUSED);
364 DEFINE_CASE(LINEAR);
365 DEFINE_CASE(POT_DIVIDE);
366 DEFINE_CASE(MODULO);
367 DEFINE_CASE(NPOT_DIVIDE);
368 default: return "MALI_ATTR_UNUSED /* XXX: Unknown stencil op, check dump */";
369 }
370 }
371
372 #undef DEFINE_CASE
373
374 #define DEFINE_CASE(name) case MALI_CHANNEL_## name: return "MALI_CHANNEL_" #name
375 static char *
376 pandecode_channel_name(enum mali_channel channel)
377 {
378 switch (channel) {
379 DEFINE_CASE(RED);
380 DEFINE_CASE(GREEN);
381 DEFINE_CASE(BLUE);
382 DEFINE_CASE(ALPHA);
383 DEFINE_CASE(ZERO);
384 DEFINE_CASE(ONE);
385 DEFINE_CASE(RESERVED_0);
386 DEFINE_CASE(RESERVED_1);
387
388 default:
389 return "MALI_CHANNEL_ZERO /* XXX: Unknown channel, check dump */";
390 }
391 }
392 #undef DEFINE_CASE
393
394 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
395 static char *
396 pandecode_wrap_mode_name(enum mali_wrap_mode op)
397 {
398 switch (op) {
399 DEFINE_CASE(REPEAT);
400 DEFINE_CASE(CLAMP_TO_EDGE);
401 DEFINE_CASE(CLAMP_TO_BORDER);
402 DEFINE_CASE(MIRRORED_REPEAT);
403
404 default:
405 return "MALI_WRAP_REPEAT /* XXX: Unknown wrap mode, check dump */";
406 }
407 }
408 #undef DEFINE_CASE
409
410 static inline char *
411 pandecode_decode_fbd_type(enum mali_fbd_type type)
412 {
413 if (type == MALI_SFBD) return "SFBD";
414 else if (type == MALI_MFBD) return "MFBD";
415 else return "WATFBD /* XXX */";
416 }
417
418 static void
419 pandecode_replay_sfbd(uint64_t gpu_va, int job_no)
420 {
421 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
422 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
423
424 pandecode_log("struct mali_single_framebuffer framebuffer_%d = {\n", job_no);
425 pandecode_indent++;
426
427 pandecode_prop("unknown1 = 0x%" PRIx32, s->unknown1);
428 pandecode_prop("unknown2 = 0x%" PRIx32, s->unknown2);
429
430 pandecode_log(".format = ");
431 pandecode_log_decoded_flags(fb_fmt_flag_info, s->format);
432 pandecode_log_cont(",\n");
433
434 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", s->width + 1);
435 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", s->height + 1);
436
437 MEMORY_PROP(s, framebuffer);
438 pandecode_prop("stride = %d", s->stride);
439
440 /* Earlier in the actual commandstream -- right before width -- but we
441 * delay to flow nicer */
442
443 pandecode_log(".clear_flags = ");
444 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
445 pandecode_log_cont(",\n");
446
447 if (s->depth_buffer | s->depth_buffer_enable) {
448 MEMORY_PROP(s, depth_buffer);
449 pandecode_prop("depth_buffer_enable = %s", DS_ENABLE(s->depth_buffer_enable));
450 }
451
452 if (s->stencil_buffer | s->stencil_buffer_enable) {
453 MEMORY_PROP(s, stencil_buffer);
454 pandecode_prop("stencil_buffer_enable = %s", DS_ENABLE(s->stencil_buffer_enable));
455 }
456
457 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
458 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
459 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
460 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
461 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
462 }
463
464 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
465 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
466 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
467 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
468 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
469 }
470
471 if (s->clear_stencil) {
472 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
473 }
474
475 MEMORY_PROP(s, unknown_address_0);
476 MEMORY_PROP(s, tiler_polygon_list);
477 MEMORY_PROP(s, tiler_polygon_list_body);
478
479 pandecode_prop("tiler_resolution_check = 0x%" PRIx32, s->tiler_resolution_check);
480 pandecode_prop("tiler_hierarchy_mask = 0x%" PRIx16, s->tiler_hierarchy_mask);
481 pandecode_prop("tiler_flags = 0x%" PRIx16, s->tiler_flags);
482
483 MEMORY_PROP(s, tiler_heap_free);
484 MEMORY_PROP(s, tiler_heap_end);
485
486 pandecode_indent--;
487 pandecode_log("};\n");
488
489 pandecode_prop("zero0 = 0x%" PRIx64, s->zero0);
490 pandecode_prop("zero1 = 0x%" PRIx64, s->zero1);
491 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
492 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
493
494 printf(".zero3 = {");
495
496 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
497 printf("%X, ", s->zero3[i]);
498
499 printf("},\n");
500
501 printf(".zero6 = {");
502
503 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
504 printf("%X, ", s->zero6[i]);
505
506 printf("},\n");
507 }
508
509 static void
510 pandecode_replay_swizzle(unsigned swizzle)
511 {
512 pandecode_prop("swizzle = %s | (%s << 3) | (%s << 6) | (%s << 9)",
513 pandecode_channel_name((swizzle >> 0) & 0x7),
514 pandecode_channel_name((swizzle >> 3) & 0x7),
515 pandecode_channel_name((swizzle >> 6) & 0x7),
516 pandecode_channel_name((swizzle >> 9) & 0x7));
517 }
518
519 static void
520 pandecode_rt_format(struct mali_rt_format format)
521 {
522 pandecode_log(".format = {\n");
523 pandecode_indent++;
524
525 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
526 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
527
528 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
529 MALI_NEGATIVE(format.nr_channels));
530
531 pandecode_log(".flags = ");
532 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
533 pandecode_log_cont(",\n");
534
535 pandecode_replay_swizzle(format.swizzle);
536
537 pandecode_prop("unk4 = 0x%" PRIx32, format.unk4);
538
539 pandecode_indent--;
540 pandecode_log("},\n");
541 }
542
543 static void
544 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct bifrost_framebuffer *fb)
545 {
546 pandecode_log("struct bifrost_render_target rts_list_%d[] = {\n", job_no);
547 pandecode_indent++;
548
549 for (int i = 0; i < MALI_NEGATIVE(fb->rt_count_1); i++) {
550 mali_ptr rt_va = gpu_va + i * sizeof(struct bifrost_render_target);
551 struct pandecode_mapped_memory *mem =
552 pandecode_find_mapped_gpu_mem_containing(rt_va);
553 const struct bifrost_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
554
555 pandecode_log("{\n");
556 pandecode_indent++;
557
558 pandecode_rt_format(rt->format);
559
560 /* TODO: How the actual heck does AFBC enabling work here? */
561 if (0) {
562 pandecode_log(".afbc = {\n");
563 pandecode_indent++;
564
565 char *a = pointer_as_memory_reference(rt->afbc.metadata);
566 pandecode_prop("metadata = %s", a);
567 free(a);
568
569 pandecode_prop("stride = %d", rt->afbc.stride);
570 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
571
572 pandecode_indent--;
573 pandecode_log("},\n");
574 } else {
575 pandecode_log(".chunknown = {\n");
576 pandecode_indent++;
577
578 pandecode_prop("unk = 0x%" PRIx64, rt->chunknown.unk);
579
580 char *a = pointer_as_memory_reference(rt->chunknown.pointer);
581 pandecode_prop("pointer = %s", a);
582 free(a);
583
584 pandecode_indent--;
585 pandecode_log("},\n");
586 }
587
588 MEMORY_PROP(rt, framebuffer);
589 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
590
591 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
592 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
593 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
594 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
595 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
596 }
597
598 if (rt->zero1 || rt->zero2 || rt->zero3) {
599 pandecode_msg("render target zeros tripped\n");
600 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
601 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
602 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
603 }
604
605 pandecode_indent--;
606 pandecode_log("},\n");
607 }
608
609 pandecode_indent--;
610 pandecode_log("};\n");
611 }
612
613 static void
614 pandecode_replay_mfbd_bfr(uint64_t gpu_va, int job_no, bool with_render_targets)
615 {
616 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
617 const struct bifrost_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
618
619 if (fb->sample_locations) {
620 /* The blob stores all possible sample locations in a single buffer
621 * allocated on startup, and just switches the pointer when switching
622 * MSAA state. For now, we just put the data into the cmdstream, but we
623 * should do something like what the blob does with a real driver.
624 *
625 * There seem to be 32 slots for sample locations, followed by another
626 * 16. The second 16 is just the center location followed by 15 zeros
627 * in all the cases I've identified (maybe shader vs. depth/color
628 * samples?).
629 */
630
631 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->sample_locations);
632
633 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->sample_locations);
634
635 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
636 pandecode_indent++;
637
638 for (int i = 0; i < 32 + 16; i++) {
639 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
640 }
641
642 pandecode_indent--;
643 pandecode_log("};\n");
644 }
645
646 pandecode_log("struct bifrost_framebuffer framebuffer_%d = {\n", job_no);
647 pandecode_indent++;
648
649 pandecode_prop("unk0 = 0x%x", fb->unk0);
650
651 if (fb->sample_locations)
652 pandecode_prop("sample_locations = sample_locations_%d", job_no);
653
654 /* Assume that unknown1 was emitted in the last job for
655 * now */
656 MEMORY_PROP(fb, unknown1);
657
658 pandecode_prop("tiler_polygon_list_size = 0x%x", fb->tiler_polygon_list_size);
659 pandecode_prop("tiler_hierarchy_mask = 0x%" PRIx16, fb->tiler_hierarchy_mask);
660 pandecode_prop("tiler_flags = 0x%" PRIx16, fb->tiler_flags);
661
662 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
663 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
664 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
665 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
666
667 pandecode_prop("unk1 = 0x%x", fb->unk1);
668 pandecode_prop("unk2 = 0x%x", fb->unk2);
669 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
670 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
671
672 pandecode_log(".mfbd_flags = ");
673 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
674 pandecode_log_cont(",\n");
675
676 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
677 pandecode_prop("clear_depth = %f", fb->clear_depth);
678
679 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
680 MEMORY_PROP(fb, scratchpad);
681 MEMORY_PROP(fb, tiler_polygon_list);
682 MEMORY_PROP(fb, tiler_polygon_list_body);
683 MEMORY_PROP(fb, tiler_heap_start);
684 MEMORY_PROP(fb, tiler_heap_end);
685
686 if (fb->zero3 || fb->zero4) {
687 pandecode_msg("framebuffer zeros tripped\n");
688 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
689 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
690 }
691
692 bool nonzero_weights = false;
693
694 for (unsigned w = 0; w < ARRAY_SIZE(fb->tiler_weights); ++w) {
695 nonzero_weights |= fb->tiler_weights[w] != 0x0;
696 }
697
698 if (nonzero_weights) {
699 pandecode_log(".tiler_weights = {");
700
701 for (unsigned w = 0; w < ARRAY_SIZE(fb->tiler_weights); ++w) {
702 pandecode_log("%d, ", fb->tiler_weights[w]);
703 }
704
705 pandecode_log("},");
706 }
707
708 pandecode_indent--;
709 pandecode_log("};\n");
710
711 gpu_va += sizeof(struct bifrost_framebuffer);
712
713 if ((fb->mfbd_flags & MALI_MFBD_EXTRA) && with_render_targets) {
714 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
715 const struct bifrost_fb_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
716
717 pandecode_log("struct bifrost_fb_extra fb_extra_%d = {\n", job_no);
718 pandecode_indent++;
719
720 MEMORY_PROP(fbx, checksum);
721
722 if (fbx->checksum_stride)
723 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
724
725 pandecode_log(".flags = ");
726 pandecode_log_decoded_flags(mfbd_extra_flag_info, fbx->flags);
727 pandecode_log_cont(",\n");
728
729 if (fbx->flags & MALI_EXTRA_AFBC_ZS) {
730 pandecode_log(".ds_afbc = {\n");
731 pandecode_indent++;
732
733 MEMORY_PROP((&fbx->ds_afbc), depth_stencil_afbc_metadata);
734 pandecode_prop("depth_stencil_afbc_stride = %d",
735 fbx->ds_afbc.depth_stencil_afbc_stride);
736 MEMORY_PROP((&fbx->ds_afbc), depth_stencil);
737
738 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
739 pandecode_msg("Depth/stencil AFBC zeros tripped\n");
740 pandecode_prop("zero1 = 0x%" PRIx32,
741 fbx->ds_afbc.zero1);
742 pandecode_prop("padding = 0x%" PRIx64,
743 fbx->ds_afbc.padding);
744 }
745
746 pandecode_indent--;
747 pandecode_log("},\n");
748 } else {
749 pandecode_log(".ds_linear = {\n");
750 pandecode_indent++;
751
752 if (fbx->ds_linear.depth) {
753 MEMORY_PROP((&fbx->ds_linear), depth);
754 pandecode_prop("depth_stride = %d",
755 fbx->ds_linear.depth_stride);
756 }
757
758 if (fbx->ds_linear.stencil) {
759 MEMORY_PROP((&fbx->ds_linear), stencil);
760 pandecode_prop("stencil_stride = %d",
761 fbx->ds_linear.stencil_stride);
762 }
763
764 if (fbx->ds_linear.depth_stride_zero ||
765 fbx->ds_linear.stencil_stride_zero ||
766 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
767 pandecode_msg("Depth/stencil zeros tripped\n");
768 pandecode_prop("depth_stride_zero = 0x%x",
769 fbx->ds_linear.depth_stride_zero);
770 pandecode_prop("stencil_stride_zero = 0x%x",
771 fbx->ds_linear.stencil_stride_zero);
772 pandecode_prop("zero1 = 0x%" PRIx32,
773 fbx->ds_linear.zero1);
774 pandecode_prop("zero2 = 0x%" PRIx32,
775 fbx->ds_linear.zero2);
776 }
777
778 pandecode_indent--;
779 pandecode_log("},\n");
780 }
781
782 if (fbx->zero3 || fbx->zero4) {
783 pandecode_msg("fb_extra zeros tripped\n");
784 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
785 pandecode_prop("zero4 = 0x%" PRIx64, fbx->zero4);
786 }
787
788 pandecode_indent--;
789 pandecode_log("};\n");
790
791 gpu_va += sizeof(struct bifrost_fb_extra);
792 }
793
794 if (with_render_targets)
795 pandecode_render_target(gpu_va, job_no, fb);
796 }
797
798 static void
799 pandecode_replay_attributes(const struct pandecode_mapped_memory *mem,
800 mali_ptr addr, int job_no, char *suffix,
801 int count, bool varying)
802 {
803 char *prefix = varying ? "varyings" : "attributes";
804
805 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
806
807 char base[128];
808 snprintf(base, sizeof(base), "%s_data_%d%s", prefix, job_no, suffix);
809
810 for (int i = 0; i < count; ++i) {
811 enum mali_attr_mode mode = attr[i].elements & 7;
812
813 if (mode == MALI_ATTR_UNUSED)
814 continue;
815
816 mali_ptr raw_elements = attr[i].elements & ~7;
817
818 /* TODO: Do we maybe want to dump the attribute values
819 * themselves given the specified format? Or is that too hard?
820 * */
821
822 char *a = pointer_as_memory_reference(raw_elements);
823 pandecode_log("mali_ptr %s_%d_p = %s;\n", base, i, a);
824 free(a);
825 }
826
827 pandecode_log("union mali_attr %s_%d[] = {\n", prefix, job_no);
828 pandecode_indent++;
829
830 for (int i = 0; i < count; ++i) {
831 pandecode_log("{\n");
832 pandecode_indent++;
833
834 pandecode_prop("elements = (%s_%d_p) | %s", base, i, pandecode_attr_mode_name(attr[i].elements & 7));
835 pandecode_prop("shift = %d", attr[i].shift);
836 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
837 pandecode_prop("stride = 0x%" PRIx32, attr[i].stride);
838 pandecode_prop("size = 0x%" PRIx32, attr[i].size);
839 pandecode_indent--;
840 pandecode_log("}, \n");
841
842 if ((attr[i].elements & 7) == MALI_ATTR_NPOT_DIVIDE) {
843 i++;
844 pandecode_log("{\n");
845 pandecode_indent++;
846 pandecode_prop("unk = 0x%x", attr[i].unk);
847 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
848 if (attr[i].zero != 0)
849 pandecode_prop("zero = 0x%x /* XXX zero tripped */", attr[i].zero);
850 pandecode_prop("divisor = %d", attr[i].divisor);
851 pandecode_indent--;
852 pandecode_log("}, \n");
853 }
854
855 }
856
857 pandecode_indent--;
858 pandecode_log("};\n");
859 }
860
861 static mali_ptr
862 pandecode_replay_shader_address(const char *name, mali_ptr ptr)
863 {
864 /* TODO: Decode flags */
865 mali_ptr shader_ptr = ptr & ~15;
866
867 char *a = pointer_as_memory_reference(shader_ptr);
868 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
869 free(a);
870
871 return shader_ptr;
872 }
873
874 static void
875 pandecode_replay_stencil(const char *name, const struct mali_stencil_test *stencil)
876 {
877 const char *func = pandecode_func_name(stencil->func);
878 const char *sfail = pandecode_stencil_op_name(stencil->sfail);
879 const char *dpfail = pandecode_stencil_op_name(stencil->dpfail);
880 const char *dppass = pandecode_stencil_op_name(stencil->dppass);
881
882 if (stencil->zero)
883 pandecode_msg("Stencil zero tripped: %X\n", stencil->zero);
884
885 pandecode_log(".stencil_%s = {\n", name);
886 pandecode_indent++;
887 pandecode_prop("ref = %d", stencil->ref);
888 pandecode_prop("mask = 0x%02X", stencil->mask);
889 pandecode_prop("func = %s", func);
890 pandecode_prop("sfail = %s", sfail);
891 pandecode_prop("dpfail = %s", dpfail);
892 pandecode_prop("dppass = %s", dppass);
893 pandecode_indent--;
894 pandecode_log("},\n");
895 }
896
897 static void
898 pandecode_replay_blend_equation(const struct mali_blend_equation *blend)
899 {
900 if (blend->zero1)
901 pandecode_msg("Blend zero tripped: %X\n", blend->zero1);
902
903 pandecode_log(".equation = {\n");
904 pandecode_indent++;
905
906 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
907 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
908
909 pandecode_log(".color_mask = ");
910 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
911 pandecode_log_cont(",\n");
912
913 pandecode_indent--;
914 pandecode_log("},\n");
915 }
916
917 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
918
919 static unsigned
920 decode_bifrost_constant(u16 constant)
921 {
922 float lo = (float) (constant & 0xFF);
923 float hi = (float) (constant >> 8);
924
925 return (hi / 255.0) + (lo / 65535.0);
926 }
927
928 static mali_ptr
929 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
930 {
931 struct bifrost_blend_rt *b =
932 ((struct bifrost_blend_rt *) descs) + rt_no;
933
934 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
935 pandecode_indent++;
936
937 pandecode_prop("flags = 0x%" PRIx16, b->flags);
938 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
939 b->constant, decode_bifrost_constant(b->constant));
940
941 /* TODO figure out blend shader enable bit */
942 pandecode_replay_blend_equation(&b->equation);
943 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
944 pandecode_prop("index = 0x%" PRIx16, b->index);
945 pandecode_prop("shader = 0x%" PRIx32, b->shader);
946
947 pandecode_indent--;
948 pandecode_log("},\n");
949
950 return 0;
951 }
952
953 static mali_ptr
954 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
955 {
956 pandecode_log(".blend = {\n");
957 pandecode_indent++;
958
959 if (is_shader) {
960 pandecode_replay_shader_address("shader", blend->shader);
961 } else {
962 pandecode_replay_blend_equation(&blend->equation);
963 pandecode_prop("constant = %f", blend->constant);
964 }
965
966 pandecode_indent--;
967 pandecode_log("},\n");
968
969 /* Return blend shader to disassemble if present */
970 return is_shader ? (blend->shader & ~0xF) : 0;
971 }
972
973 static mali_ptr
974 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
975 {
976 struct midgard_blend_rt *b =
977 ((struct midgard_blend_rt *) descs) + rt_no;
978
979 /* Flags determine presence of blend shader */
980 bool is_shader = (b->flags & 0xF) >= 0x2;
981
982 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
983 pandecode_indent++;
984
985 pandecode_prop("flags = 0x%" PRIx64, b->flags);
986
987 mali_ptr shader = pandecode_midgard_blend(&b->blend, is_shader);
988
989 pandecode_indent--;
990 pandecode_log("};\n");
991
992 return shader;
993 }
994
995 static int
996 pandecode_replay_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
997 {
998 char base[128];
999 char *prefix = varying ? "varying" : "attribute";
1000 unsigned max_index = 0;
1001 snprintf(base, sizeof(base), "%s_meta", prefix);
1002
1003 pandecode_log("struct mali_attr_meta %s_%d%s[] = {\n", base, job_no, suffix);
1004 pandecode_indent++;
1005
1006 struct mali_attr_meta *attr_meta;
1007 mali_ptr p = varying ? (v->varying_meta & ~0xF) : v->attribute_meta;
1008
1009 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1010
1011 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1012 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1013 sizeof(*attr_mem));
1014
1015 pandecode_log("{\n");
1016 pandecode_indent++;
1017 pandecode_prop("index = %d", attr_meta->index);
1018
1019 if (attr_meta->index > max_index)
1020 max_index = attr_meta->index;
1021 pandecode_replay_swizzle(attr_meta->swizzle);
1022 pandecode_prop("format = %s", pandecode_format_name(attr_meta->format));
1023
1024 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1025 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1026 pandecode_prop("src_offset = 0x%" PRIx64, (u64) attr_meta->src_offset);
1027 pandecode_indent--;
1028 pandecode_log("},\n");
1029
1030 }
1031
1032 pandecode_indent--;
1033 pandecode_log("};\n");
1034
1035 return max_index;
1036 }
1037
1038 static void
1039 pandecode_replay_indices(uintptr_t pindices, uint32_t index_count, int job_no)
1040 {
1041 struct pandecode_mapped_memory *imem = pandecode_find_mapped_gpu_mem_containing(pindices);
1042
1043 if (imem) {
1044 /* Indices are literally just a u32 array :) */
1045
1046 uint32_t *PANDECODE_PTR_VAR(indices, imem, pindices);
1047
1048 pandecode_log("uint32_t indices_%d[] = {\n", job_no);
1049 pandecode_indent++;
1050
1051 for (unsigned i = 0; i < (index_count + 1); i += 3)
1052 pandecode_log("%d, %d, %d,\n",
1053 indices[i],
1054 indices[i + 1],
1055 indices[i + 2]);
1056
1057 pandecode_indent--;
1058 pandecode_log("};\n");
1059 }
1060 }
1061
1062 /* return bits [lo, hi) of word */
1063 static u32
1064 bits(u32 word, u32 lo, u32 hi)
1065 {
1066 if (hi - lo >= 32)
1067 return word; // avoid undefined behavior with the shift
1068
1069 return (word >> lo) & ((1 << (hi - lo)) - 1);
1070 }
1071
1072 static void
1073 pandecode_replay_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no)
1074 {
1075 pandecode_log_cont("{\n");
1076 pandecode_indent++;
1077
1078 pandecode_prop("invocation_count = %" PRIx32, p->invocation_count);
1079 pandecode_prop("size_y_shift = %d", p->size_y_shift);
1080 pandecode_prop("size_z_shift = %d", p->size_z_shift);
1081 pandecode_prop("workgroups_x_shift = %d", p->workgroups_x_shift);
1082 pandecode_prop("workgroups_y_shift = %d", p->workgroups_y_shift);
1083 pandecode_prop("workgroups_z_shift = %d", p->workgroups_z_shift);
1084 pandecode_prop("workgroups_x_shift_2 = 0x%" PRIx32, p->workgroups_x_shift_2);
1085
1086 /* Decode invocation_count. See the comment before the definition of
1087 * invocation_count for an explanation.
1088 */
1089 pandecode_msg("size: (%d, %d, %d)\n",
1090 bits(p->invocation_count, 0, p->size_y_shift) + 1,
1091 bits(p->invocation_count, p->size_y_shift, p->size_z_shift) + 1,
1092 bits(p->invocation_count, p->size_z_shift,
1093 p->workgroups_x_shift) + 1);
1094 pandecode_msg("workgroups: (%d, %d, %d)\n",
1095 bits(p->invocation_count, p->workgroups_x_shift,
1096 p->workgroups_y_shift) + 1,
1097 bits(p->invocation_count, p->workgroups_y_shift,
1098 p->workgroups_z_shift) + 1,
1099 bits(p->invocation_count, p->workgroups_z_shift,
1100 32) + 1);
1101
1102 /* TODO: Decode */
1103 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1104 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1105
1106 pandecode_prop("draw_mode = %s", pandecode_draw_mode_name(p->draw_mode));
1107
1108 /* Index count only exists for tiler jobs anyway */
1109
1110 if (p->index_count)
1111 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1112
1113 pandecode_prop("negative_start = %d", p->negative_start);
1114
1115 DYN_MEMORY_PROP(p, job_no, indices);
1116
1117 if (p->zero1) {
1118 pandecode_msg("Zero tripped\n");
1119 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1120 }
1121
1122 pandecode_indent--;
1123 pandecode_log("},\n");
1124 }
1125
1126 static void
1127 pandecode_replay_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1128 {
1129 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1130
1131 struct mali_uniform_buffer_meta *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1132
1133 for (int i = 0; i < ubufs_count; i++) {
1134 mali_ptr ptr = ubufs[i].ptr << 2;
1135 struct pandecode_mapped_memory *umem2 = pandecode_find_mapped_gpu_mem_containing(ptr);
1136 uint32_t *PANDECODE_PTR_VAR(ubuf, umem2, ptr);
1137 char name[50];
1138 snprintf(name, sizeof(name), "ubuf_%d", i);
1139 /* The blob uses ubuf 0 to upload internal stuff and
1140 * uniforms that won't fit/are accessed indirectly, so
1141 * it puts it in the batchbuffer.
1142 */
1143 pandecode_log("uint32_t %s_%d[] = {\n", name, job_no);
1144 pandecode_indent++;
1145
1146 for (int j = 0; j <= ubufs[i].size; j++) {
1147 for (int k = 0; k < 4; k++) {
1148 if (k == 0)
1149 pandecode_log("0x%"PRIx32", ", ubuf[4 * j + k]);
1150 else
1151 pandecode_log_cont("0x%"PRIx32", ", ubuf[4 * j + k]);
1152
1153 }
1154
1155 pandecode_log_cont("\n");
1156 }
1157
1158 pandecode_indent--;
1159 pandecode_log("};\n");
1160 }
1161
1162 pandecode_log("struct mali_uniform_buffer_meta uniform_buffers_%d[] = {\n",
1163 job_no);
1164 pandecode_indent++;
1165
1166 for (int i = 0; i < ubufs_count; i++) {
1167 pandecode_log("{\n");
1168 pandecode_indent++;
1169 pandecode_prop("size = MALI_POSITIVE(%d)", ubufs[i].size + 1);
1170 pandecode_prop("ptr = ubuf_%d_%d_p >> 2", i, job_no);
1171 pandecode_indent--;
1172 pandecode_log("},\n");
1173 }
1174
1175 pandecode_indent--;
1176 pandecode_log("};\n");
1177 }
1178
1179 static void
1180 pandecode_replay_scratchpad(uintptr_t pscratchpad, int job_no, char *suffix)
1181 {
1182
1183 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(pscratchpad);
1184
1185 struct bifrost_scratchpad *PANDECODE_PTR_VAR(scratchpad, mem, pscratchpad);
1186
1187 if (scratchpad->zero)
1188 pandecode_msg("XXX scratchpad zero tripped");
1189
1190 pandecode_log("struct bifrost_scratchpad scratchpad_%d%s = {\n", job_no, suffix);
1191 pandecode_indent++;
1192
1193 pandecode_prop("flags = 0x%x", scratchpad->flags);
1194 MEMORY_PROP(scratchpad, gpu_scratchpad);
1195
1196 pandecode_indent--;
1197 pandecode_log("};\n");
1198 }
1199
1200 static void
1201 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1202 bool is_bifrost)
1203 {
1204 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1205 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1206
1207 /* Compute maximum possible size */
1208 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1209
1210 /* Print some boilerplate to clearly denote the assembly (which doesn't
1211 * obey indentation rules), and actually do the disassembly! */
1212
1213 printf("\n\n");
1214
1215 if (is_bifrost) {
1216 disassemble_bifrost(code, sz, false);
1217 } else {
1218 disassemble_midgard(code, sz);
1219 }
1220
1221 printf("\n\n");
1222 }
1223
1224 static void
1225 pandecode_replay_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix *p,
1226 int job_no, enum mali_job_type job_type,
1227 char *suffix, bool is_bifrost)
1228 {
1229 mali_ptr shader_meta_ptr = (u64) (uintptr_t) (p->_shader_upper << 4);
1230 struct pandecode_mapped_memory *attr_mem;
1231
1232 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1233 * are the only things actually needed from the FBD, vertex/tiler jobs
1234 * no longer reference the FBD -- instead, this field points to some
1235 * info about the scratchpad.
1236 */
1237 if (is_bifrost)
1238 pandecode_replay_scratchpad(p->framebuffer & ~FBD_TYPE, job_no, suffix);
1239 else if (p->framebuffer & MALI_MFBD)
1240 pandecode_replay_mfbd_bfr((u64) ((uintptr_t) p->framebuffer) & FBD_MASK, job_no, false);
1241 else
1242 pandecode_replay_sfbd((u64) (uintptr_t) p->framebuffer, job_no);
1243
1244 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
1245 int texture_count = 0, sampler_count = 0;
1246
1247 if (shader_meta_ptr) {
1248 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(shader_meta_ptr);
1249 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, shader_meta_ptr);
1250
1251 pandecode_log("struct mali_shader_meta shader_meta_%d%s = {\n", job_no, suffix);
1252 pandecode_indent++;
1253
1254 /* Save for dumps */
1255 attribute_count = s->attribute_count;
1256 varying_count = s->varying_count;
1257 texture_count = s->texture_count;
1258 sampler_count = s->sampler_count;
1259
1260 if (is_bifrost) {
1261 uniform_count = s->bifrost2.uniform_count;
1262 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
1263 } else {
1264 uniform_count = s->midgard1.uniform_count;
1265 /* TODO figure this out */
1266 uniform_buffer_count = 1;
1267 }
1268
1269 mali_ptr shader_ptr = pandecode_replay_shader_address("shader", s->shader);
1270
1271 pandecode_prop("texture_count = %" PRId16, s->texture_count);
1272 pandecode_prop("sampler_count = %" PRId16, s->sampler_count);
1273 pandecode_prop("attribute_count = %" PRId16, s->attribute_count);
1274 pandecode_prop("varying_count = %" PRId16, s->varying_count);
1275
1276 if (is_bifrost) {
1277 pandecode_log(".bifrost1 = {\n");
1278 pandecode_indent++;
1279
1280 pandecode_prop("uniform_buffer_count = %" PRId32, s->bifrost1.uniform_buffer_count);
1281 pandecode_prop("unk1 = 0x%" PRIx32, s->bifrost1.unk1);
1282
1283 pandecode_indent--;
1284 pandecode_log("},\n");
1285 } else {
1286 pandecode_log(".midgard1 = {\n");
1287 pandecode_indent++;
1288
1289 pandecode_prop("uniform_count = %" PRId16, s->midgard1.uniform_count);
1290 pandecode_prop("work_count = %" PRId16, s->midgard1.work_count);
1291
1292 pandecode_log(".unknown1 = ");
1293 pandecode_log_decoded_flags(shader_unknown1_flag_info, s->midgard1.unknown1);
1294 pandecode_log_cont(",\n");
1295
1296 pandecode_prop("unknown2 = 0x%" PRIx32, s->midgard1.unknown2);
1297
1298 pandecode_indent--;
1299 pandecode_log("},\n");
1300 }
1301
1302 if (s->depth_units || s->depth_factor) {
1303 if (is_bifrost)
1304 pandecode_prop("depth_units = %f", s->depth_units);
1305 else
1306 pandecode_prop("depth_units = MALI_NEGATIVE(%f)", s->depth_units - 1.0f);
1307
1308 pandecode_prop("depth_factor = %f", s->depth_factor);
1309 }
1310
1311 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
1312 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
1313
1314 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
1315 invert_alpha_coverage ? "~" : "",
1316 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
1317
1318 pandecode_log(".unknown2_3 = ");
1319
1320 int unknown2_3 = s->unknown2_3;
1321 int unknown2_4 = s->unknown2_4;
1322
1323 /* We're not quite sure what these flags mean without the depth test, if anything */
1324
1325 if (unknown2_3 & (MALI_DEPTH_TEST | MALI_DEPTH_FUNC_MASK)) {
1326 const char *func = pandecode_func_name(MALI_GET_DEPTH_FUNC(unknown2_3));
1327 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
1328
1329 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
1330 }
1331
1332 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
1333 pandecode_log_cont(",\n");
1334
1335 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
1336 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
1337
1338 pandecode_log(".unknown2_4 = ");
1339 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
1340 pandecode_log_cont(",\n");
1341
1342 pandecode_replay_stencil("front", &s->stencil_front);
1343 pandecode_replay_stencil("back", &s->stencil_back);
1344
1345 if (is_bifrost) {
1346 pandecode_log(".bifrost2 = {\n");
1347 pandecode_indent++;
1348
1349 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
1350 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
1351 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
1352 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
1353
1354 pandecode_indent--;
1355 pandecode_log("},\n");
1356 } else {
1357 pandecode_log(".midgard2 = {\n");
1358 pandecode_indent++;
1359
1360 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
1361 pandecode_indent--;
1362 pandecode_log("},\n");
1363 }
1364
1365 pandecode_prop("unknown2_8 = 0x%" PRIx32, s->unknown2_8);
1366
1367 if (!is_bifrost) {
1368 /* TODO: Blend shaders routing/disasm */
1369
1370 pandecode_midgard_blend(&s->blend, false);
1371 }
1372
1373 pandecode_indent--;
1374 pandecode_log("};\n");
1375
1376 /* MRT blend fields are used whenever MFBD is used, with
1377 * per-RT descriptors */
1378
1379 if (job_type == JOB_TYPE_TILER) {
1380 void* blend_base = (void *) (s + 1);
1381
1382 for (unsigned i = 0; i < 4; i++) {
1383 mali_ptr shader = 0;
1384
1385 if (is_bifrost)
1386 shader = pandecode_bifrost_blend(blend_base, job_no, i);
1387 else
1388 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
1389
1390 if (shader)
1391 pandecode_shader_disassemble(shader, job_no, job_type, false);
1392 }
1393 }
1394
1395 pandecode_shader_disassemble(shader_ptr, job_no, job_type, is_bifrost);
1396 } else
1397 pandecode_msg("<no shader>\n");
1398
1399 if (p->viewport) {
1400 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
1401 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
1402
1403 pandecode_log("struct mali_viewport viewport_%d%s = {\n", job_no, suffix);
1404 pandecode_indent++;
1405
1406 pandecode_prop("clip_minx = %f", f->clip_minx);
1407 pandecode_prop("clip_miny = %f", f->clip_miny);
1408 pandecode_prop("clip_minz = %f", f->clip_minz);
1409 pandecode_prop("clip_maxx = %f", f->clip_maxx);
1410 pandecode_prop("clip_maxy = %f", f->clip_maxy);
1411 pandecode_prop("clip_maxz = %f", f->clip_maxz);
1412
1413 /* Only the higher coordinates are MALI_POSITIVE scaled */
1414
1415 pandecode_prop("viewport0 = { %d, %d }",
1416 f->viewport0[0], f->viewport0[1]);
1417
1418 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
1419 f->viewport1[0] + 1, f->viewport1[1] + 1);
1420
1421 pandecode_indent--;
1422 pandecode_log("};\n");
1423 }
1424
1425 if (p->attribute_meta) {
1426 unsigned max_attr_index = pandecode_replay_attribute_meta(job_no, attribute_count, p, false, suffix);
1427
1428 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
1429 pandecode_replay_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index + 1, false);
1430 }
1431
1432 /* Varyings are encoded like attributes but not actually sent; we just
1433 * pass a zero buffer with the right stride/size set, (or whatever)
1434 * since the GPU will write to it itself */
1435
1436 if (p->varyings) {
1437 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
1438
1439 /* Number of descriptors depends on whether there are
1440 * non-internal varyings */
1441
1442 pandecode_replay_attributes(attr_mem, p->varyings, job_no, suffix, varying_count > 1 ? 4 : 1, true);
1443 }
1444
1445 if (p->varying_meta) {
1446 pandecode_replay_attribute_meta(job_no, varying_count, p, true, suffix);
1447 }
1448
1449 if (p->uniforms) {
1450 int rows = uniform_count, width = 4;
1451 size_t sz = rows * width * sizeof(float);
1452
1453 struct pandecode_mapped_memory *uniform_mem = pandecode_find_mapped_gpu_mem_containing(p->uniforms);
1454 pandecode_fetch_gpu_mem(uniform_mem, p->uniforms, sz);
1455 float *PANDECODE_PTR_VAR(uniforms, uniform_mem, p->uniforms);
1456
1457 pandecode_log("float uniforms_%d%s[] = {\n", job_no, suffix);
1458
1459 pandecode_indent++;
1460
1461 for (int row = 0; row < rows; row++) {
1462 for (int i = 0; i < width; i++)
1463 pandecode_log_cont("%ff, ", uniforms[i]);
1464
1465 pandecode_log_cont("\n");
1466
1467 uniforms += width;
1468 }
1469
1470 pandecode_indent--;
1471 pandecode_log("};\n");
1472 }
1473
1474 if (p->uniform_buffers) {
1475 pandecode_replay_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
1476 }
1477
1478 if (p->texture_trampoline) {
1479 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(p->texture_trampoline);
1480
1481 if (mmem) {
1482 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline);
1483
1484 pandecode_log("uint64_t texture_trampoline_%d[] = {\n", job_no);
1485 pandecode_indent++;
1486
1487 for (int tex = 0; tex < texture_count; ++tex) {
1488 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
1489 char *a = pointer_as_memory_reference(*u);
1490 pandecode_log("%s,\n", a);
1491 free(a);
1492 }
1493
1494 pandecode_indent--;
1495 pandecode_log("};\n");
1496
1497 /* Now, finally, descend down into the texture descriptor */
1498 for (int tex = 0; tex < texture_count; ++tex) {
1499 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
1500 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
1501
1502 if (tmem) {
1503 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, *u);
1504
1505 pandecode_log("struct mali_texture_descriptor texture_descriptor_%d_%d = {\n", job_no, tex);
1506 pandecode_indent++;
1507
1508 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", t->width + 1);
1509 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", t->height + 1);
1510 pandecode_prop("depth = MALI_POSITIVE(%" PRId16 ")", t->depth + 1);
1511
1512 pandecode_prop("unknown1 = %" PRId16, t->unknown1);
1513 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
1514 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
1515 pandecode_prop("nr_mipmap_levels = %" PRId8, t->nr_mipmap_levels);
1516
1517 struct mali_texture_format f = t->format;
1518
1519 pandecode_log(".format = {\n");
1520 pandecode_indent++;
1521
1522 pandecode_replay_swizzle(f.swizzle);
1523 pandecode_prop("format = %s", pandecode_format_name(f.format));
1524
1525 pandecode_prop("usage1 = 0x%" PRIx32, f.usage1);
1526 pandecode_prop("is_not_cubemap = %" PRId32, f.is_not_cubemap);
1527 pandecode_prop("usage2 = 0x%" PRIx32, f.usage2);
1528
1529 pandecode_indent--;
1530 pandecode_log("},\n");
1531
1532 pandecode_replay_swizzle(t->swizzle);
1533
1534 if (t->swizzle_zero) {
1535 /* Shouldn't happen */
1536 pandecode_msg("Swizzle zero tripped but replay will be fine anyway");
1537 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
1538 }
1539
1540 pandecode_prop("unknown3 = 0x%" PRIx32, t->unknown3);
1541
1542 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
1543 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
1544 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
1545
1546 pandecode_log(".payload = {\n");
1547 pandecode_indent++;
1548
1549 /* A bunch of bitmap pointers follow.
1550 * We work out the correct number,
1551 * based on the mipmap/cubemap
1552 * properties, but dump extra
1553 * possibilities to futureproof */
1554
1555 int bitmap_count = MALI_NEGATIVE(t->nr_mipmap_levels);
1556 bool manual_stride = f.usage2 & MALI_TEX_MANUAL_STRIDE;
1557
1558 /* Miptree for each face */
1559 if (!f.is_not_cubemap)
1560 bitmap_count *= 6;
1561
1562 /* Stride for each element */
1563 if (manual_stride)
1564 bitmap_count *= 2;
1565
1566 /* Sanity check the size */
1567 int max_count = sizeof(t->payload) / sizeof(t->payload[0]);
1568 assert (bitmap_count <= max_count);
1569
1570 /* Dump more to be safe, but not _that_ much more */
1571 int safe_count = MIN2(bitmap_count * 2, max_count);
1572
1573 for (int i = 0; i < safe_count; ++i) {
1574 char *prefix = (i >= bitmap_count) ? "// " : "";
1575
1576 /* How we dump depends if this is a stride or a pointer */
1577
1578 if ((f.usage2 & MALI_TEX_MANUAL_STRIDE) && (i & 1)) {
1579 /* signed 32-bit snuck in as a 64-bit pointer */
1580 uint64_t stride_set = t->payload[i];
1581 uint32_t clamped_stride = stride_set;
1582 int32_t stride = clamped_stride;
1583 assert(stride_set == clamped_stride);
1584 pandecode_log("%s(mali_ptr) %d /* stride */, \n", prefix, stride);
1585 } else {
1586 char *a = pointer_as_memory_reference(t->payload[i]);
1587 pandecode_log("%s%s, \n", prefix, a);
1588 free(a);
1589 }
1590 }
1591
1592 pandecode_indent--;
1593 pandecode_log("},\n");
1594
1595 pandecode_indent--;
1596 pandecode_log("};\n");
1597 }
1598 }
1599 }
1600 }
1601
1602 if (p->sampler_descriptor) {
1603 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->sampler_descriptor);
1604
1605 if (smem) {
1606 struct mali_sampler_descriptor *s;
1607
1608 mali_ptr d = p->sampler_descriptor;
1609
1610 for (int i = 0; i < sampler_count; ++i) {
1611 s = pandecode_fetch_gpu_mem(smem, d + sizeof(*s) * i, sizeof(*s));
1612
1613 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%d_%d = {\n", job_no, i);
1614 pandecode_indent++;
1615
1616 /* Only the lower two bits are understood right now; the rest we display as hex */
1617 pandecode_log(".filter_mode = MALI_TEX_MIN(%s) | MALI_TEX_MAG(%s) | 0x%" PRIx32",\n",
1618 MALI_FILTER_NAME(s->filter_mode & MALI_TEX_MIN_MASK),
1619 MALI_FILTER_NAME(s->filter_mode & MALI_TEX_MAG_MASK),
1620 s->filter_mode & ~3);
1621
1622 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
1623 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
1624
1625 pandecode_prop("wrap_s = %s", pandecode_wrap_mode_name(s->wrap_s));
1626 pandecode_prop("wrap_t = %s", pandecode_wrap_mode_name(s->wrap_t));
1627 pandecode_prop("wrap_r = %s", pandecode_wrap_mode_name(s->wrap_r));
1628
1629 pandecode_prop("compare_func = %s", pandecode_alt_func_name(s->compare_func));
1630
1631 if (s->zero || s->zero2) {
1632 pandecode_msg("Zero tripped\n");
1633 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
1634 }
1635
1636 pandecode_prop("unknown2 = %d", s->unknown2);
1637
1638 pandecode_prop("border_color = { %f, %f, %f, %f }",
1639 s->border_color[0],
1640 s->border_color[1],
1641 s->border_color[2],
1642 s->border_color[3]);
1643
1644 pandecode_indent--;
1645 pandecode_log("};\n");
1646 }
1647 }
1648 }
1649 }
1650
1651 static void
1652 pandecode_replay_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
1653 {
1654 pandecode_log_cont("{\n");
1655 pandecode_indent++;
1656
1657 MEMORY_PROP(p, position_varying);
1658 MEMORY_COMMENT(p, position_varying);
1659 DYN_MEMORY_PROP(p, job_no, uniform_buffers);
1660 MEMORY_COMMENT(p, uniform_buffers);
1661 DYN_MEMORY_PROP(p, job_no, texture_trampoline);
1662 MEMORY_COMMENT(p, texture_trampoline);
1663 DYN_MEMORY_PROP(p, job_no, sampler_descriptor);
1664 MEMORY_COMMENT(p, sampler_descriptor);
1665 DYN_MEMORY_PROP(p, job_no, uniforms);
1666 MEMORY_COMMENT(p, uniforms);
1667 DYN_MEMORY_PROP(p, job_no, attributes);
1668 MEMORY_COMMENT(p, attributes);
1669 DYN_MEMORY_PROP(p, job_no, attribute_meta);
1670 MEMORY_COMMENT(p, attribute_meta);
1671 DYN_MEMORY_PROP(p, job_no, varyings);
1672 MEMORY_COMMENT(p, varyings);
1673 DYN_MEMORY_PROP(p, job_no, varying_meta);
1674 MEMORY_COMMENT(p, varying_meta);
1675 DYN_MEMORY_PROP(p, job_no, viewport);
1676 MEMORY_COMMENT(p, viewport);
1677 DYN_MEMORY_PROP(p, job_no, occlusion_counter);
1678 MEMORY_COMMENT(p, occlusion_counter);
1679 MEMORY_COMMENT(p, framebuffer & ~1);
1680 pandecode_msg("%" PRIx64 "\n", p->viewport);
1681 pandecode_msg("%" PRIx64 "\n", p->framebuffer);
1682
1683 if (is_bifrost)
1684 pandecode_prop("framebuffer = scratchpad_%d_p", job_no);
1685 else
1686 pandecode_prop("framebuffer = framebuffer_%d_p | %s", job_no, p->framebuffer & MALI_MFBD ? "MALI_MFBD" : "0");
1687
1688 pandecode_prop("_shader_upper = (shader_meta_%d_p) >> 4", job_no);
1689 pandecode_prop("flags = %d", p->flags);
1690
1691 pandecode_indent--;
1692 pandecode_log("},\n");
1693 }
1694
1695 static void
1696 pandecode_replay_vertex_only_bfr(struct bifrost_vertex_only *v)
1697 {
1698 pandecode_log_cont("{\n");
1699 pandecode_indent++;
1700
1701 pandecode_prop("unk2 = 0x%x", v->unk2);
1702
1703 if (v->zero0 || v->zero1) {
1704 pandecode_msg("vertex only zero tripped");
1705 pandecode_prop("zero0 = 0x%" PRIx32, v->zero0);
1706 pandecode_prop("zero1 = 0x%" PRIx64, v->zero1);
1707 }
1708
1709 pandecode_indent--;
1710 pandecode_log("}\n");
1711 }
1712
1713 static void
1714 pandecode_replay_tiler_heap_meta(mali_ptr gpu_va, int job_no)
1715 {
1716
1717 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1718 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
1719
1720 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no);
1721 pandecode_indent++;
1722
1723 if (h->zero) {
1724 pandecode_msg("tiler heap zero tripped\n");
1725 pandecode_prop("zero = 0x%x", h->zero);
1726 }
1727
1728 for (int i = 0; i < 12; i++) {
1729 if (h->zeros[i] != 0) {
1730 pandecode_msg("tiler heap zero %d tripped, value %x\n",
1731 i, h->zeros[i]);
1732 }
1733 }
1734
1735 pandecode_prop("heap_size = 0x%x", h->heap_size);
1736 MEMORY_PROP(h, tiler_heap_start);
1737 MEMORY_PROP(h, tiler_heap_free);
1738
1739 /* this might point to the beginning of another buffer, when it's
1740 * really the end of the tiler heap buffer, so we have to be careful
1741 * here.
1742 */
1743 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
1744 pandecode_prop("tiler_heap_end = %s + 1", a);
1745 free(a);
1746
1747 pandecode_indent--;
1748 pandecode_log("};\n");
1749 }
1750
1751 static void
1752 pandecode_replay_tiler_meta(mali_ptr gpu_va, int job_no)
1753 {
1754 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1755 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
1756
1757 pandecode_replay_tiler_heap_meta(t->tiler_heap_meta, job_no);
1758
1759 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no);
1760 pandecode_indent++;
1761
1762 if (t->zero0 || t->zero1) {
1763 pandecode_msg("tiler meta zero tripped");
1764 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
1765 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
1766 }
1767
1768 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
1769 pandecode_prop("flags = 0x%" PRIx16, t->flags);
1770
1771 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
1772 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
1773 DYN_MEMORY_PROP(t, job_no, tiler_heap_meta);
1774
1775 for (int i = 0; i < 12; i++) {
1776 if (t->zeros[i] != 0) {
1777 pandecode_msg("tiler heap zero %d tripped, value %" PRIx64 "\n",
1778 i, t->zeros[i]);
1779 }
1780 }
1781
1782 pandecode_indent--;
1783 pandecode_log("};\n");
1784 }
1785
1786 static void
1787 pandecode_replay_gl_enables(uint32_t gl_enables, int job_type)
1788 {
1789 pandecode_log(".gl_enables = ");
1790
1791 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
1792
1793 pandecode_log_cont(",\n");
1794 }
1795
1796 static void
1797 pandecode_replay_primitive_size(union midgard_primitive_size u, bool constant)
1798 {
1799 pandecode_log(".primitive_size = {\n");
1800 pandecode_indent++;
1801
1802 if (constant) {
1803 pandecode_prop("constant = %f", u.constant);
1804 } else {
1805 MEMORY_PROP((&u), pointer);
1806 }
1807
1808 pandecode_indent--;
1809 pandecode_log("},\n");
1810 }
1811
1812 static void
1813 pandecode_replay_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
1814 {
1815 pandecode_log_cont("{\n");
1816 pandecode_indent++;
1817
1818 /* TODO: gl_PointSize on Bifrost */
1819 pandecode_replay_primitive_size(t->primitive_size, true);
1820
1821 DYN_MEMORY_PROP(t, job_no, tiler_meta);
1822 pandecode_replay_gl_enables(t->gl_enables, JOB_TYPE_TILER);
1823
1824 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
1825 || t->zero6 || t->zero7 || t->zero8) {
1826 pandecode_msg("tiler only zero tripped");
1827 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
1828 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
1829 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
1830 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
1831 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
1832 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
1833 pandecode_prop("zero7 = 0x%" PRIx32, t->zero7);
1834 pandecode_prop("zero8 = 0x%" PRIx64, t->zero8);
1835 }
1836
1837 pandecode_indent--;
1838 pandecode_log("},\n");
1839 }
1840
1841 static int
1842 pandecode_replay_vertex_job_bfr(const struct mali_job_descriptor_header *h,
1843 const struct pandecode_mapped_memory *mem,
1844 mali_ptr payload, int job_no)
1845 {
1846 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
1847
1848 pandecode_replay_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true);
1849
1850 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no);
1851 pandecode_indent++;
1852
1853 pandecode_log(".prefix = ");
1854 pandecode_replay_vertex_tiler_prefix(&v->prefix, job_no);
1855
1856 pandecode_log(".vertex = ");
1857 pandecode_replay_vertex_only_bfr(&v->vertex);
1858
1859 pandecode_log(".postfix = ");
1860 pandecode_replay_vertex_tiler_postfix(&v->postfix, job_no, true);
1861
1862 pandecode_indent--;
1863 pandecode_log("};\n");
1864
1865 return sizeof(*v);
1866 }
1867
1868 static int
1869 pandecode_replay_tiler_job_bfr(const struct mali_job_descriptor_header *h,
1870 const struct pandecode_mapped_memory *mem,
1871 mali_ptr payload, int job_no)
1872 {
1873 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
1874
1875 pandecode_replay_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true);
1876
1877 pandecode_replay_indices(t->prefix.indices, t->prefix.index_count, job_no);
1878 pandecode_replay_tiler_meta(t->tiler.tiler_meta, job_no);
1879
1880 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no);
1881 pandecode_indent++;
1882
1883 pandecode_log(".prefix = ");
1884 pandecode_replay_vertex_tiler_prefix(&t->prefix, job_no);
1885
1886 pandecode_log(".tiler = ");
1887 pandecode_replay_tiler_only_bfr(&t->tiler, job_no);
1888
1889 pandecode_log(".postfix = ");
1890 pandecode_replay_vertex_tiler_postfix(&t->postfix, job_no, true);
1891
1892 pandecode_indent--;
1893 pandecode_log("};\n");
1894
1895 return sizeof(*t);
1896 }
1897
1898 static int
1899 pandecode_replay_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
1900 const struct pandecode_mapped_memory *mem,
1901 mali_ptr payload, int job_no)
1902 {
1903 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
1904
1905 char *a = pointer_as_memory_reference(payload);
1906 pandecode_msg("vt payload: %s\n", a);
1907 free(a);
1908
1909 pandecode_replay_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false);
1910
1911 pandecode_replay_indices(v->prefix.indices, v->prefix.index_count, job_no);
1912
1913 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
1914 pandecode_indent++;
1915
1916 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
1917 pandecode_replay_primitive_size(v->primitive_size, !has_primitive_pointer);
1918
1919 pandecode_log(".prefix = ");
1920 pandecode_replay_vertex_tiler_prefix(&v->prefix, job_no);
1921
1922 pandecode_replay_gl_enables(v->gl_enables, h->job_type);
1923 pandecode_prop("draw_start = %d", v->draw_start);
1924
1925 #ifndef __LP64__
1926
1927 if (v->zero3) {
1928 pandecode_msg("Zero tripped\n");
1929 pandecode_prop("zero3 = 0x%" PRIx32, v->zero3);
1930 }
1931
1932 #endif
1933
1934 if (v->zero5) {
1935 pandecode_msg("Zero tripped\n");
1936 pandecode_prop("zero5 = 0x%" PRIx64, v->zero5);
1937 }
1938
1939 pandecode_log(".postfix = ");
1940 pandecode_replay_vertex_tiler_postfix(&v->postfix, job_no, false);
1941
1942 pandecode_indent--;
1943 pandecode_log("};\n");
1944
1945 return sizeof(*v);
1946 }
1947
1948 static int
1949 pandecode_replay_fragment_job(const struct pandecode_mapped_memory *mem,
1950 mali_ptr payload, int job_no,
1951 bool is_bifrost)
1952 {
1953 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
1954
1955 bool fbd_dumped = false;
1956
1957 if (!is_bifrost && (s->framebuffer & FBD_TYPE) == MALI_SFBD) {
1958 /* Only SFBDs are understood, not MFBDs. We're speculating,
1959 * based on the versioning, kernel code, etc, that the
1960 * difference is between Single FrameBuffer Descriptor and
1961 * Multiple FrmaeBuffer Descriptor; the change apparently lines
1962 * up with multi-framebuffer support being added (T7xx onwards,
1963 * including Gxx). In any event, there's some field shuffling
1964 * that we haven't looked into yet. */
1965
1966 pandecode_replay_sfbd(s->framebuffer & FBD_MASK, job_no);
1967 fbd_dumped = true;
1968 } else if ((s->framebuffer & FBD_TYPE) == MALI_MFBD) {
1969 /* We don't know if Bifrost supports SFBD's at all, since the
1970 * driver never uses them. And the format is different from
1971 * Midgard anyways, due to the tiler heap and scratchpad being
1972 * moved out into separate structures, so it's not clear what a
1973 * Bifrost SFBD would even look like without getting an actual
1974 * trace, which appears impossible.
1975 */
1976
1977 pandecode_replay_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true);
1978 fbd_dumped = true;
1979 }
1980
1981 uintptr_t p = (uintptr_t) s->framebuffer & FBD_MASK;
1982
1983 pandecode_log("struct mali_payload_fragment payload_%d = {\n", job_no);
1984 pandecode_indent++;
1985
1986 /* See the comments by the macro definitions for mathematical context
1987 * on why this is so weird */
1988
1989 if (MALI_TILE_COORD_FLAGS(s->max_tile_coord) || MALI_TILE_COORD_FLAGS(s->min_tile_coord))
1990 pandecode_msg("Tile coordinate flag missed, replay wrong\n");
1991
1992 pandecode_prop("min_tile_coord = MALI_COORDINATE_TO_TILE_MIN(%d, %d)",
1993 MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT,
1994 MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT);
1995
1996 pandecode_prop("max_tile_coord = MALI_COORDINATE_TO_TILE_MAX(%d, %d)",
1997 (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT,
1998 (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT);
1999
2000 /* If the FBD was just decoded, we can refer to it by pointer. If not,
2001 * we have to fallback on offsets. */
2002
2003 const char *fbd_type = s->framebuffer & MALI_MFBD ? "MALI_MFBD" : "MALI_SFBD";
2004
2005 if (fbd_dumped)
2006 pandecode_prop("framebuffer = framebuffer_%d_p | %s", job_no, fbd_type);
2007 else
2008 pandecode_prop("framebuffer = %s | %s", pointer_as_memory_reference(p), fbd_type);
2009
2010 pandecode_indent--;
2011 pandecode_log("};\n");
2012
2013 return sizeof(*s);
2014 }
2015
2016 static int job_descriptor_number = 0;
2017
2018 int
2019 pandecode_replay_jc(mali_ptr jc_gpu_va, bool bifrost)
2020 {
2021 struct mali_job_descriptor_header *h;
2022
2023 int start_number = 0;
2024
2025 bool first = true;
2026 bool last_size;
2027
2028 do {
2029 struct pandecode_mapped_memory *mem =
2030 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2031
2032 void *payload;
2033
2034 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2035
2036 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2037 * high 32-bits of the 64-bit pointer are reused to store
2038 * something else.
2039 */
2040 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2041 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
2042 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2043
2044 payload = pandecode_fetch_gpu_mem(mem, payload_ptr,
2045 MALI_PAYLOAD_SIZE);
2046
2047 int job_no = job_descriptor_number++;
2048
2049 if (first)
2050 start_number = job_no;
2051
2052 pandecode_log("struct mali_job_descriptor_header job_%d = {\n", job_no);
2053 pandecode_indent++;
2054
2055 pandecode_prop("job_type = %s", pandecode_job_type_name(h->job_type));
2056
2057 /* Save for next job fixing */
2058 last_size = h->job_descriptor_size;
2059
2060 if (h->job_descriptor_size)
2061 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2062
2063 if (h->exception_status)
2064 pandecode_prop("exception_status = %d", h->exception_status);
2065
2066 if (h->first_incomplete_task)
2067 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2068
2069 if (h->fault_pointer)
2070 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2071
2072 if (h->job_barrier)
2073 pandecode_prop("job_barrier = %d", h->job_barrier);
2074
2075 pandecode_prop("job_index = %d", h->job_index);
2076
2077 if (h->unknown_flags)
2078 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2079
2080 if (h->job_dependency_index_1)
2081 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2082
2083 if (h->job_dependency_index_2)
2084 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2085
2086 pandecode_indent--;
2087 pandecode_log("};\n");
2088
2089 /* Do not touch the field yet -- decode the payload first, and
2090 * don't touch that either. This is essential for the uploads
2091 * to occur in sequence and therefore be dynamically allocated
2092 * correctly. Do note the size, however, for that related
2093 * reason. */
2094
2095 switch (h->job_type) {
2096 case JOB_TYPE_SET_VALUE: {
2097 struct mali_payload_set_value *s = payload;
2098
2099 pandecode_log("struct mali_payload_set_value payload_%d = {\n", job_no);
2100 pandecode_indent++;
2101 MEMORY_PROP(s, out);
2102 pandecode_prop("unknown = 0x%" PRIX64, s->unknown);
2103 pandecode_indent--;
2104 pandecode_log("};\n");
2105
2106 break;
2107 }
2108
2109 case JOB_TYPE_TILER:
2110 case JOB_TYPE_VERTEX:
2111 case JOB_TYPE_COMPUTE:
2112 if (bifrost) {
2113 if (h->job_type == JOB_TYPE_TILER)
2114 pandecode_replay_tiler_job_bfr(h, mem, payload_ptr, job_no);
2115 else
2116 pandecode_replay_vertex_job_bfr(h, mem, payload_ptr, job_no);
2117 } else
2118 pandecode_replay_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no);
2119
2120 break;
2121
2122 case JOB_TYPE_FRAGMENT:
2123 pandecode_replay_fragment_job(mem, payload_ptr, job_no, bifrost);
2124 break;
2125
2126 default:
2127 break;
2128 }
2129
2130 /* Handle linkage */
2131
2132 if (!first) {
2133 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no - 1);
2134
2135 if (last_size)
2136 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no);
2137 else
2138 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no);
2139 }
2140
2141 first = false;
2142
2143 } while ((jc_gpu_va = h->job_descriptor_size ? h->next_job_64 : h->next_job_32));
2144
2145 return start_number;
2146 }