projects / mesa.git / blob
? search:


7e7fed81f9976d73f35fc32d2937f172555f361e
[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_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 MEMORY_PROP_DIR(obj, p) {\
50 if (obj.p) { \
51 char *a = pointer_as_memory_reference(obj.p); \
52 pandecode_prop("%s = %s", #p, a); \
53 free(a); \
54 } \
55 }
56
57 #define DYN_MEMORY_PROP(obj, no, p) { \
58 if (obj->p) \
59 pandecode_prop("%s = %s_%d_p", #p, #p, no); \
60 }
61
62 /* Semantic logging type.
63 *
64 * Raw: for raw messages to be printed as is.
65 * Message: for helpful information to be commented out in replays.
66 * Property: for properties of a struct
67 *
68 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
69 */
70
71 enum pandecode_log_type {
72 PANDECODE_RAW,
73 PANDECODE_MESSAGE,
74 PANDECODE_PROPERTY
75 };
76
77 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
78 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
79 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
80
81 unsigned pandecode_indent = 0;
82
83 static void
84 pandecode_make_indent(void)
85 {
86 for (unsigned i = 0; i < pandecode_indent; ++i)
87 printf(" ");
88 }
89
90 static void
91 pandecode_log_typed(enum pandecode_log_type type, const char *format, ...)
92 {
93 va_list ap;
94
95 pandecode_make_indent();
96
97 if (type == PANDECODE_MESSAGE)
98 printf("// ");
99 else if (type == PANDECODE_PROPERTY)
100 printf(".");
101
102 va_start(ap, format);
103 vprintf(format, ap);
104 va_end(ap);
105
106 if (type == PANDECODE_PROPERTY)
107 printf(",\n");
108 }
109
110 static void
111 pandecode_log_cont(const char *format, ...)
112 {
113 va_list ap;
114
115 va_start(ap, format);
116 vprintf(format, ap);
117 va_end(ap);
118 }
119
120 struct pandecode_flag_info {
121 u64 flag;
122 const char *name;
123 };
124
125 static void
126 pandecode_log_decoded_flags(const struct pandecode_flag_info *flag_info,
127 u64 flags)
128 {
129 bool decodable_flags_found = false;
130
131 for (int i = 0; flag_info[i].name; i++) {
132 if ((flags & flag_info[i].flag) != flag_info[i].flag)
133 continue;
134
135 if (!decodable_flags_found) {
136 decodable_flags_found = true;
137 } else {
138 pandecode_log_cont(" | ");
139 }
140
141 pandecode_log_cont("%s", flag_info[i].name);
142
143 flags &= ~flag_info[i].flag;
144 }
145
146 if (decodable_flags_found) {
147 if (flags)
148 pandecode_log_cont(" | 0x%" PRIx64, flags);
149 } else {
150 pandecode_log_cont("0x%" PRIx64, flags);
151 }
152 }
153
154 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
155 static const struct pandecode_flag_info gl_enable_flag_info[] = {
156 FLAG_INFO(OCCLUSION_QUERY),
157 FLAG_INFO(OCCLUSION_PRECISE),
158 FLAG_INFO(FRONT_CCW_TOP),
159 FLAG_INFO(CULL_FACE_FRONT),
160 FLAG_INFO(CULL_FACE_BACK),
161 {}
162 };
163 #undef FLAG_INFO
164
165 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
166 static const struct pandecode_flag_info clear_flag_info[] = {
167 FLAG_INFO(FAST),
168 FLAG_INFO(SLOW),
169 FLAG_INFO(SLOW_STENCIL),
170 {}
171 };
172 #undef FLAG_INFO
173
174 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
175 static const struct pandecode_flag_info mask_flag_info[] = {
176 FLAG_INFO(R),
177 FLAG_INFO(G),
178 FLAG_INFO(B),
179 FLAG_INFO(A),
180 {}
181 };
182 #undef FLAG_INFO
183
184 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
185 static const struct pandecode_flag_info u3_flag_info[] = {
186 FLAG_INFO(HAS_MSAA),
187 FLAG_INFO(CAN_DISCARD),
188 FLAG_INFO(HAS_BLEND_SHADER),
189 FLAG_INFO(DEPTH_TEST),
190 {}
191 };
192
193 static const struct pandecode_flag_info u4_flag_info[] = {
194 FLAG_INFO(NO_MSAA),
195 FLAG_INFO(NO_DITHER),
196 FLAG_INFO(DEPTH_RANGE_A),
197 FLAG_INFO(DEPTH_RANGE_B),
198 FLAG_INFO(STENCIL_TEST),
199 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER),
200 {}
201 };
202 #undef FLAG_INFO
203
204 #define FLAG_INFO(flag) { MALI_FRAMEBUFFER_##flag, "MALI_FRAMEBUFFER_" #flag }
205 static const struct pandecode_flag_info fb_fmt_flag_info[] = {
206 FLAG_INFO(MSAA_A),
207 FLAG_INFO(MSAA_B),
208 FLAG_INFO(MSAA_8),
209 {}
210 };
211 #undef FLAG_INFO
212
213 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
214 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
215 FLAG_INFO(MSAA),
216 FLAG_INFO(SRGB),
217 {}
218 };
219 #undef FLAG_INFO
220
221 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
222 static const struct pandecode_flag_info mfbd_extra_flag_info[] = {
223 FLAG_INFO(PRESENT),
224 FLAG_INFO(AFBC),
225 FLAG_INFO(ZS),
226 {}
227 };
228 #undef FLAG_INFO
229
230 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
231 static const struct pandecode_flag_info shader_midgard1_flag_info [] = {
232 FLAG_INFO(EARLY_Z),
233 FLAG_INFO(HELPER_INVOCATIONS),
234 FLAG_INFO(READS_TILEBUFFER),
235 FLAG_INFO(READS_ZS),
236 {}
237 };
238 #undef FLAG_INFO
239
240 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
241 static const struct pandecode_flag_info mfbd_flag_info [] = {
242 FLAG_INFO(DEPTH_WRITE),
243 FLAG_INFO(EXTRA),
244 {}
245 };
246 #undef FLAG_INFO
247
248
249 extern char *replace_fragment;
250 extern char *replace_vertex;
251
252 static char *
253 pandecode_job_type(enum mali_job_type type)
254 {
255 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
256
257 switch (type) {
258 DEFINE_CASE(NULL);
259 DEFINE_CASE(SET_VALUE);
260 DEFINE_CASE(CACHE_FLUSH);
261 DEFINE_CASE(COMPUTE);
262 DEFINE_CASE(VERTEX);
263 DEFINE_CASE(TILER);
264 DEFINE_CASE(FUSED);
265 DEFINE_CASE(FRAGMENT);
266
267 case JOB_NOT_STARTED:
268 return "NOT_STARTED";
269
270 default:
271 pandecode_log("Warning! Unknown job type %x\n", type);
272 return "!?!?!?";
273 }
274
275 #undef DEFINE_CASE
276 }
277
278 static char *
279 pandecode_draw_mode(enum mali_draw_mode mode)
280 {
281 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
282
283 switch (mode) {
284 DEFINE_CASE(DRAW_NONE);
285 DEFINE_CASE(POINTS);
286 DEFINE_CASE(LINES);
287 DEFINE_CASE(TRIANGLES);
288 DEFINE_CASE(TRIANGLE_STRIP);
289 DEFINE_CASE(TRIANGLE_FAN);
290 DEFINE_CASE(LINE_STRIP);
291 DEFINE_CASE(LINE_LOOP);
292 DEFINE_CASE(POLYGON);
293 DEFINE_CASE(QUADS);
294 DEFINE_CASE(QUAD_STRIP);
295
296 default:
297 return "MALI_TRIANGLES /* XXX: Unknown GL mode, check dump */";
298 }
299
300 #undef DEFINE_CASE
301 }
302
303 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
304 static char *
305 pandecode_func(enum mali_func mode)
306 {
307 switch (mode) {
308 DEFINE_CASE(NEVER);
309 DEFINE_CASE(LESS);
310 DEFINE_CASE(EQUAL);
311 DEFINE_CASE(LEQUAL);
312 DEFINE_CASE(GREATER);
313 DEFINE_CASE(NOTEQUAL);
314 DEFINE_CASE(GEQUAL);
315 DEFINE_CASE(ALWAYS);
316
317 default:
318 return "MALI_FUNC_NEVER /* XXX: Unknown function, check dump */";
319 }
320 }
321 #undef DEFINE_CASE
322
323 /* Why is this duplicated? Who knows... */
324 #define DEFINE_CASE(name) case MALI_ALT_FUNC_ ## name: return "MALI_ALT_FUNC_" #name
325 static char *
326 pandecode_alt_func(enum mali_alt_func mode)
327 {
328 switch (mode) {
329 DEFINE_CASE(NEVER);
330 DEFINE_CASE(LESS);
331 DEFINE_CASE(EQUAL);
332 DEFINE_CASE(LEQUAL);
333 DEFINE_CASE(GREATER);
334 DEFINE_CASE(NOTEQUAL);
335 DEFINE_CASE(GEQUAL);
336 DEFINE_CASE(ALWAYS);
337
338 default:
339 return "MALI_FUNC_NEVER /* XXX: Unknown function, check dump */";
340 }
341 }
342 #undef DEFINE_CASE
343
344 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
345 static char *
346 pandecode_stencil_op(enum mali_stencil_op op)
347 {
348 switch (op) {
349 DEFINE_CASE(KEEP);
350 DEFINE_CASE(REPLACE);
351 DEFINE_CASE(ZERO);
352 DEFINE_CASE(INVERT);
353 DEFINE_CASE(INCR_WRAP);
354 DEFINE_CASE(DECR_WRAP);
355 DEFINE_CASE(INCR);
356 DEFINE_CASE(DECR);
357
358 default:
359 return "MALI_STENCIL_KEEP /* XXX: Unknown stencil op, check dump */";
360 }
361 }
362
363 #undef DEFINE_CASE
364
365 #define DEFINE_CASE(name) case MALI_ATTR_ ## name: return "MALI_ATTR_" #name
366 static char *pandecode_attr_mode(enum mali_attr_mode mode)
367 {
368 switch(mode) {
369 DEFINE_CASE(UNUSED);
370 DEFINE_CASE(LINEAR);
371 DEFINE_CASE(POT_DIVIDE);
372 DEFINE_CASE(MODULO);
373 DEFINE_CASE(NPOT_DIVIDE);
374 default:
375 return "MALI_ATTR_UNUSED /* XXX: Unknown stencil op, check dump */";
376 }
377 }
378
379 #undef DEFINE_CASE
380
381 #define DEFINE_CASE(name) case MALI_CHANNEL_## name: return "MALI_CHANNEL_" #name
382 static char *
383 pandecode_channel(enum mali_channel channel)
384 {
385 switch (channel) {
386 DEFINE_CASE(RED);
387 DEFINE_CASE(GREEN);
388 DEFINE_CASE(BLUE);
389 DEFINE_CASE(ALPHA);
390 DEFINE_CASE(ZERO);
391 DEFINE_CASE(ONE);
392 DEFINE_CASE(RESERVED_0);
393 DEFINE_CASE(RESERVED_1);
394
395 default:
396 return "MALI_CHANNEL_ZERO /* XXX: Unknown channel, check dump */";
397 }
398 }
399 #undef DEFINE_CASE
400
401 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
402 static char *
403 pandecode_wrap_mode(enum mali_wrap_mode op)
404 {
405 switch (op) {
406 DEFINE_CASE(REPEAT);
407 DEFINE_CASE(CLAMP_TO_EDGE);
408 DEFINE_CASE(CLAMP_TO_BORDER);
409 DEFINE_CASE(MIRRORED_REPEAT);
410
411 default:
412 return "MALI_WRAP_REPEAT /* XXX: Unknown wrap mode, check dump */";
413 }
414 }
415 #undef DEFINE_CASE
416
417 #define DEFINE_CASE(name) case MALI_TEX_## name: return "MALI_TEX_" #name
418 static char *
419 pandecode_texture_type(enum mali_texture_type type)
420 {
421 switch (type) {
422 DEFINE_CASE(1D);
423 DEFINE_CASE(2D);
424 DEFINE_CASE(3D);
425 DEFINE_CASE(CUBE);
426
427 default:
428 unreachable("Unknown case");
429 }
430 }
431 #undef DEFINE_CASE
432
433 #define DEFINE_CASE(name) case MALI_MFBD_BLOCK_## name: return "MALI_MFBD_BLOCK_" #name
434 static char *
435 pandecode_mfbd_block_format(enum mali_mfbd_block_format fmt)
436 {
437 switch (fmt) {
438 DEFINE_CASE(TILED);
439 DEFINE_CASE(UNKNOWN);
440 DEFINE_CASE(LINEAR);
441 DEFINE_CASE(AFBC);
442
443 default:
444 unreachable("Invalid case");
445 }
446 }
447 #undef DEFINE_CASE
448
449 /* Midgard's tiler descriptor is embedded within the
450 * larger FBD */
451
452 static void
453 pandecode_midgard_tiler_descriptor(const struct midgard_tiler_descriptor *t)
454 {
455 pandecode_log(".tiler = {\n");
456 pandecode_indent++;
457
458 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
459 pandecode_prop("flags = 0x%" PRIx16, t->flags);
460 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
461
462 MEMORY_PROP(t, polygon_list);
463 MEMORY_PROP(t, polygon_list_body);
464
465 MEMORY_PROP(t, heap_start);
466
467 {
468 /* Points to the end of a buffer */
469 char *a = pointer_as_memory_reference(t->heap_end - 1);
470 pandecode_prop("heap_end = %s + 1", a);
471 free(a);
472 }
473
474 bool nonzero_weights = false;
475
476 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
477 nonzero_weights |= t->weights[w] != 0x0;
478 }
479
480 if (nonzero_weights) {
481 pandecode_log(".weights = {");
482
483 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
484 pandecode_log("%d, ", t->weights[w]);
485 }
486
487 pandecode_log("},");
488 }
489
490 pandecode_indent--;
491 pandecode_log("}\n");
492 }
493
494 static void
495 pandecode_sfbd(uint64_t gpu_va, int job_no)
496 {
497 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
498 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
499
500 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
501 pandecode_indent++;
502
503 pandecode_prop("unknown1 = 0x%" PRIx32, s->unknown1);
504 pandecode_prop("unknown2 = 0x%" PRIx32, s->unknown2);
505
506 pandecode_log(".format = ");
507 pandecode_log_decoded_flags(fb_fmt_flag_info, s->format);
508 pandecode_log_cont(",\n");
509
510 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", s->width + 1);
511 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", s->height + 1);
512
513 MEMORY_PROP(s, framebuffer);
514 pandecode_prop("stride = %d", s->stride);
515
516 /* Earlier in the actual commandstream -- right before width -- but we
517 * delay to flow nicer */
518
519 pandecode_log(".clear_flags = ");
520 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
521 pandecode_log_cont(",\n");
522
523 if (s->depth_buffer | s->depth_buffer_enable) {
524 MEMORY_PROP(s, depth_buffer);
525 pandecode_prop("depth_buffer_enable = %s", DS_ENABLE(s->depth_buffer_enable));
526 }
527
528 if (s->stencil_buffer | s->stencil_buffer_enable) {
529 MEMORY_PROP(s, stencil_buffer);
530 pandecode_prop("stencil_buffer_enable = %s", DS_ENABLE(s->stencil_buffer_enable));
531 }
532
533 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
534 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
535 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
536 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
537 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
538 }
539
540 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
541 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
542 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
543 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
544 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
545 }
546
547 if (s->clear_stencil) {
548 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
549 }
550
551 MEMORY_PROP(s, unknown_address_0);
552 const struct midgard_tiler_descriptor t = s->tiler;
553 pandecode_midgard_tiler_descriptor(&t);
554
555 pandecode_indent--;
556 pandecode_log("};\n");
557
558 pandecode_prop("zero0 = 0x%" PRIx64, s->zero0);
559 pandecode_prop("zero1 = 0x%" PRIx64, s->zero1);
560 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
561 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
562
563 printf(".zero3 = {");
564
565 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
566 printf("%X, ", s->zero3[i]);
567
568 printf("},\n");
569
570 printf(".zero6 = {");
571
572 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
573 printf("%X, ", s->zero6[i]);
574
575 printf("},\n");
576 }
577
578 static void
579 pandecode_u32_slide(unsigned name, const u32 *slide, unsigned count)
580 {
581 pandecode_log(".unknown%d = {", name);
582
583 for (int i = 0; i < count; ++i)
584 printf("%X, ", slide[i]);
585
586 pandecode_log("},\n");
587 }
588
589 #define SHORT_SLIDE(num) \
590 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
591
592 static void
593 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
594 {
595 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
596 const struct mali_compute_fbd *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
597
598 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
599 pandecode_indent++;
600
601 SHORT_SLIDE(1);
602
603 pandecode_indent--;
604 printf("},\n");
605 }
606
607 static void
608 pandecode_swizzle(unsigned swizzle)
609 {
610 pandecode_prop("swizzle = %s | (%s << 3) | (%s << 6) | (%s << 9)",
611 pandecode_channel((swizzle >> 0) & 0x7),
612 pandecode_channel((swizzle >> 3) & 0x7),
613 pandecode_channel((swizzle >> 6) & 0x7),
614 pandecode_channel((swizzle >> 9) & 0x7));
615 }
616
617 static void
618 pandecode_rt_format(struct mali_rt_format format)
619 {
620 pandecode_log(".format = {\n");
621 pandecode_indent++;
622
623 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
624 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
625 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
626
627 pandecode_prop("block = %s",
628 pandecode_mfbd_block_format(format.block));
629
630 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
631 MALI_NEGATIVE(format.nr_channels));
632
633 pandecode_log(".flags = ");
634 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
635 pandecode_log_cont(",\n");
636
637 pandecode_swizzle(format.swizzle);
638
639 pandecode_prop("unk4 = 0x%" PRIx32, format.unk4);
640
641 pandecode_indent--;
642 pandecode_log("},\n");
643 }
644
645 static void
646 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct bifrost_framebuffer *fb)
647 {
648 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
649 pandecode_indent++;
650
651 for (int i = 0; i < MALI_NEGATIVE(fb->rt_count_1); i++) {
652 mali_ptr rt_va = gpu_va + i * sizeof(struct bifrost_render_target);
653 struct pandecode_mapped_memory *mem =
654 pandecode_find_mapped_gpu_mem_containing(rt_va);
655 const struct bifrost_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
656
657 pandecode_log("{\n");
658 pandecode_indent++;
659
660 pandecode_rt_format(rt->format);
661
662 if (rt->format.block == MALI_MFBD_BLOCK_AFBC) {
663 pandecode_log(".afbc = {\n");
664 pandecode_indent++;
665
666 char *a = pointer_as_memory_reference(rt->afbc.metadata);
667 pandecode_prop("metadata = %s", a);
668 free(a);
669
670 pandecode_prop("stride = %d", rt->afbc.stride);
671 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
672
673 pandecode_indent--;
674 pandecode_log("},\n");
675 } else {
676 pandecode_log(".chunknown = {\n");
677 pandecode_indent++;
678
679 pandecode_prop("unk = 0x%" PRIx64, rt->chunknown.unk);
680
681 char *a = pointer_as_memory_reference(rt->chunknown.pointer);
682 pandecode_prop("pointer = %s", a);
683 free(a);
684
685 pandecode_indent--;
686 pandecode_log("},\n");
687 }
688
689 MEMORY_PROP(rt, framebuffer);
690 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
691
692 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
693 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
694 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
695 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
696 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
697 }
698
699 if (rt->zero1 || rt->zero2 || rt->zero3) {
700 pandecode_msg("render target zeros tripped\n");
701 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
702 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
703 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
704 }
705
706 pandecode_indent--;
707 pandecode_log("},\n");
708 }
709
710 pandecode_indent--;
711 pandecode_log("};\n");
712 }
713
714 static unsigned
715 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool with_render_targets)
716 {
717 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
718 const struct bifrost_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
719
720 if (fb->sample_locations) {
721 /* The blob stores all possible sample locations in a single buffer
722 * allocated on startup, and just switches the pointer when switching
723 * MSAA state. For now, we just put the data into the cmdstream, but we
724 * should do something like what the blob does with a real driver.
725 *
726 * There seem to be 32 slots for sample locations, followed by another
727 * 16. The second 16 is just the center location followed by 15 zeros
728 * in all the cases I've identified (maybe shader vs. depth/color
729 * samples?).
730 */
731
732 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->sample_locations);
733
734 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->sample_locations);
735
736 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
737 pandecode_indent++;
738
739 for (int i = 0; i < 32 + 16; i++) {
740 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
741 }
742
743 pandecode_indent--;
744 pandecode_log("};\n");
745 }
746
747 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
748 pandecode_indent++;
749
750 pandecode_prop("unk0 = 0x%x", fb->unk0);
751
752 if (fb->sample_locations)
753 pandecode_prop("sample_locations = sample_locations_%d", job_no);
754
755 /* Assume that unknown1 was emitted in the last job for
756 * now */
757 MEMORY_PROP(fb, unknown1);
758
759 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
760 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
761 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
762 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
763
764 pandecode_prop("unk1 = 0x%x", fb->unk1);
765 pandecode_prop("unk2 = 0x%x", fb->unk2);
766 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
767 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
768
769 pandecode_log(".mfbd_flags = ");
770 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
771 pandecode_log_cont(",\n");
772
773 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
774 pandecode_prop("clear_depth = %f", fb->clear_depth);
775
776 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
777 MEMORY_PROP(fb, scratchpad);
778 const struct midgard_tiler_descriptor t = fb->tiler;
779 pandecode_midgard_tiler_descriptor(&t);
780
781 if (fb->zero3 || fb->zero4) {
782 pandecode_msg("framebuffer zeros tripped\n");
783 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
784 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
785 }
786
787 pandecode_indent--;
788 pandecode_log("};\n");
789
790 gpu_va += sizeof(struct bifrost_framebuffer);
791
792 if ((fb->mfbd_flags & MALI_MFBD_EXTRA) && with_render_targets) {
793 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
794 const struct bifrost_fb_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
795
796 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
797 pandecode_indent++;
798
799 MEMORY_PROP(fbx, checksum);
800
801 if (fbx->checksum_stride)
802 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
803
804 pandecode_log(".flags = ");
805 pandecode_log_decoded_flags(mfbd_extra_flag_info, fbx->flags);
806 pandecode_log_cont(",\n");
807
808 if (fbx->flags & MALI_EXTRA_AFBC_ZS) {
809 pandecode_log(".ds_afbc = {\n");
810 pandecode_indent++;
811
812 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
813 pandecode_prop("depth_stencil_afbc_stride = %d",
814 fbx->ds_afbc.depth_stencil_afbc_stride);
815 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
816
817 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
818 pandecode_msg("Depth/stencil AFBC zeros tripped\n");
819 pandecode_prop("zero1 = 0x%" PRIx32,
820 fbx->ds_afbc.zero1);
821 pandecode_prop("padding = 0x%" PRIx64,
822 fbx->ds_afbc.padding);
823 }
824
825 pandecode_indent--;
826 pandecode_log("},\n");
827 } else {
828 pandecode_log(".ds_linear = {\n");
829 pandecode_indent++;
830
831 if (fbx->ds_linear.depth) {
832 MEMORY_PROP_DIR(fbx->ds_linear, depth);
833 pandecode_prop("depth_stride = %d",
834 fbx->ds_linear.depth_stride);
835 }
836
837 if (fbx->ds_linear.stencil) {
838 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
839 pandecode_prop("stencil_stride = %d",
840 fbx->ds_linear.stencil_stride);
841 }
842
843 if (fbx->ds_linear.depth_stride_zero ||
844 fbx->ds_linear.stencil_stride_zero ||
845 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
846 pandecode_msg("Depth/stencil zeros tripped\n");
847 pandecode_prop("depth_stride_zero = 0x%x",
848 fbx->ds_linear.depth_stride_zero);
849 pandecode_prop("stencil_stride_zero = 0x%x",
850 fbx->ds_linear.stencil_stride_zero);
851 pandecode_prop("zero1 = 0x%" PRIx32,
852 fbx->ds_linear.zero1);
853 pandecode_prop("zero2 = 0x%" PRIx32,
854 fbx->ds_linear.zero2);
855 }
856
857 pandecode_indent--;
858 pandecode_log("},\n");
859 }
860
861 if (fbx->zero3 || fbx->zero4) {
862 pandecode_msg("fb_extra zeros tripped\n");
863 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
864 pandecode_prop("zero4 = 0x%" PRIx64, fbx->zero4);
865 }
866
867 pandecode_indent--;
868 pandecode_log("};\n");
869
870 gpu_va += sizeof(struct bifrost_fb_extra);
871 }
872
873 if (with_render_targets)
874 pandecode_render_target(gpu_va, job_no, fb);
875
876 /* Passback the render target count */
877 return MALI_NEGATIVE(fb->rt_count_1);
878 }
879
880 /* Just add a comment decoding the shift/odd fields forming the padded vertices
881 * count */
882
883 static void
884 pandecode_padded_vertices(unsigned shift, unsigned k)
885 {
886 unsigned odd = 2*k + 1;
887 unsigned pot = 1 << shift;
888 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
889 }
890
891 /* Given a magic divisor, recover what we were trying to divide by.
892 *
893 * Let m represent the magic divisor. By definition, m is an element on Z, whre
894 * 0 <= m < 2^N, for N bits in m.
895 *
896 * Let q represent the number we would like to divide by.
897 *
898 * By definition of a magic divisor for N-bit unsigned integers (a number you
899 * multiply by to magically get division), m is a number such that:
900 *
901 * (m * x) & (2^N - 1) = floor(x/q).
902 * for all x on Z where 0 <= x < 2^N
903 *
904 * Ignore the case where any of the above values equals zero; it is irrelevant
905 * for our purposes (instanced arrays).
906 *
907 * Choose x = q. Then:
908 *
909 * (m * x) & (2^N - 1) = floor(x/q).
910 * (m * q) & (2^N - 1) = floor(q/q).
911 *
912 * floor(q/q) = floor(1) = 1, therefore:
913 *
914 * (m * q) & (2^N - 1) = 1
915 *
916 * Recall the identity that the bitwise AND of one less than a power-of-two
917 * equals the modulo with that power of two, i.e. for all x:
918 *
919 * x & (2^N - 1) = x % N
920 *
921 * Therefore:
922 *
923 * mq % (2^N) = 1
924 *
925 * By definition, a modular multiplicative inverse of a number m is the number
926 * q such that with respect to a modulos M:
927 *
928 * mq % M = 1
929 *
930 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
931 *
932 */
933
934 static void
935 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
936 {
937 /* Compute the modular inverse of `magic` with respect to 2^(32 -
938 * shift) the most lame way possible... just repeatedly add.
939 * Asymptoptically slow but nobody cares in practice, unless you have
940 * massive numbers of vertices or high divisors. */
941
942 unsigned inverse = 0;
943
944 /* Magic implicitly has the highest bit set */
945 magic |= (1 << 31);
946
947 /* Depending on rounding direction */
948 if (extra)
949 magic++;
950
951 for (;;) {
952 uint32_t product = magic * inverse;
953
954 if (shift) {
955 product >>= shift;
956 }
957
958 if (product == 1)
959 break;
960
961 ++inverse;
962 }
963
964 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
965
966 /* Recall we're supposed to divide by (gl_level_divisor *
967 * padded_num_vertices) */
968
969 unsigned padded_num_vertices = inverse / orig_divisor;
970
971 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
972 }
973
974 static void
975 pandecode_attributes(const struct pandecode_mapped_memory *mem,
976 mali_ptr addr, int job_no, char *suffix,
977 int count, bool varying)
978 {
979 char *prefix = varying ? "varyings" : "attributes";
980
981 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
982
983 char base[128];
984 snprintf(base, sizeof(base), "%s_data_%d%s", prefix, job_no, suffix);
985
986 for (int i = 0; i < count; ++i) {
987 enum mali_attr_mode mode = attr[i].elements & 7;
988
989 if (mode == MALI_ATTR_UNUSED)
990 continue;
991
992 mali_ptr raw_elements = attr[i].elements & ~7;
993
994 /* TODO: Do we maybe want to dump the attribute values
995 * themselves given the specified format? Or is that too hard?
996 * */
997
998 char *a = pointer_as_memory_reference(raw_elements);
999 pandecode_log("mali_ptr %s_%d_p = %s;\n", base, i, a);
1000 free(a);
1001 }
1002
1003 pandecode_log("union mali_attr %s_%d[] = {\n", prefix, job_no);
1004 pandecode_indent++;
1005
1006 for (int i = 0; i < count; ++i) {
1007 pandecode_log("{\n");
1008 pandecode_indent++;
1009
1010 unsigned mode = attr[i].elements & 7;
1011 pandecode_prop("elements = (%s_%d_p) | %s", base, i, pandecode_attr_mode(mode));
1012 pandecode_prop("shift = %d", attr[i].shift);
1013 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1014 pandecode_prop("stride = 0x%" PRIx32, attr[i].stride);
1015 pandecode_prop("size = 0x%" PRIx32, attr[i].size);
1016
1017 /* Decode further where possible */
1018
1019 if (mode == MALI_ATTR_MODULO) {
1020 pandecode_padded_vertices(
1021 attr[i].shift,
1022 attr[i].extra_flags);
1023 }
1024
1025 pandecode_indent--;
1026 pandecode_log("}, \n");
1027
1028 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1029 i++;
1030 pandecode_log("{\n");
1031 pandecode_indent++;
1032 pandecode_prop("unk = 0x%x", attr[i].unk);
1033 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1034 if (attr[i].zero != 0)
1035 pandecode_prop("zero = 0x%x /* XXX zero tripped */", attr[i].zero);
1036 pandecode_prop("divisor = %d", attr[i].divisor);
1037 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1038 pandecode_indent--;
1039 pandecode_log("}, \n");
1040 }
1041
1042 }
1043
1044 pandecode_indent--;
1045 pandecode_log("};\n");
1046 }
1047
1048 static mali_ptr
1049 pandecode_shader_address(const char *name, mali_ptr ptr)
1050 {
1051 /* TODO: Decode flags */
1052 mali_ptr shader_ptr = ptr & ~15;
1053
1054 char *a = pointer_as_memory_reference(shader_ptr);
1055 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1056 free(a);
1057
1058 return shader_ptr;
1059 }
1060
1061 static bool
1062 all_zero(unsigned *buffer, unsigned count)
1063 {
1064 for (unsigned i = 0; i < count; ++i) {
1065 if (buffer[i])
1066 return false;
1067 }
1068
1069 return true;
1070 }
1071
1072 static void
1073 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1074 {
1075 if (all_zero((unsigned *) stencil, sizeof(stencil) / sizeof(unsigned)))
1076 return;
1077
1078 const char *func = pandecode_func(stencil->func);
1079 const char *sfail = pandecode_stencil_op(stencil->sfail);
1080 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1081 const char *dppass = pandecode_stencil_op(stencil->dppass);
1082
1083 if (stencil->zero)
1084 pandecode_msg("Stencil zero tripped: %X\n", stencil->zero);
1085
1086 pandecode_log(".stencil_%s = {\n", name);
1087 pandecode_indent++;
1088 pandecode_prop("ref = %d", stencil->ref);
1089 pandecode_prop("mask = 0x%02X", stencil->mask);
1090 pandecode_prop("func = %s", func);
1091 pandecode_prop("sfail = %s", sfail);
1092 pandecode_prop("dpfail = %s", dpfail);
1093 pandecode_prop("dppass = %s", dppass);
1094 pandecode_indent--;
1095 pandecode_log("},\n");
1096 }
1097
1098 static void
1099 pandecode_blend_equation(const struct mali_blend_equation *blend)
1100 {
1101 if (blend->zero1)
1102 pandecode_msg("Blend zero tripped: %X\n", blend->zero1);
1103
1104 pandecode_log(".equation = {\n");
1105 pandecode_indent++;
1106
1107 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1108 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1109
1110 pandecode_log(".color_mask = ");
1111 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1112 pandecode_log_cont(",\n");
1113
1114 pandecode_indent--;
1115 pandecode_log("},\n");
1116 }
1117
1118 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1119
1120 static unsigned
1121 decode_bifrost_constant(u16 constant)
1122 {
1123 float lo = (float) (constant & 0xFF);
1124 float hi = (float) (constant >> 8);
1125
1126 return (hi / 255.0) + (lo / 65535.0);
1127 }
1128
1129 static mali_ptr
1130 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1131 {
1132 struct bifrost_blend_rt *b =
1133 ((struct bifrost_blend_rt *) descs) + rt_no;
1134
1135 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1136 pandecode_indent++;
1137
1138 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1139 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1140 b->constant, decode_bifrost_constant(b->constant));
1141
1142 /* TODO figure out blend shader enable bit */
1143 pandecode_blend_equation(&b->equation);
1144 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1145 pandecode_prop("index = 0x%" PRIx16, b->index);
1146 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1147
1148 pandecode_indent--;
1149 pandecode_log("},\n");
1150
1151 return 0;
1152 }
1153
1154 static mali_ptr
1155 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1156 {
1157 if (all_zero((unsigned *) blend, sizeof(blend) / sizeof(unsigned)))
1158 return 0;
1159
1160 pandecode_log(".blend = {\n");
1161 pandecode_indent++;
1162
1163 if (is_shader) {
1164 pandecode_shader_address("shader", blend->shader);
1165 } else {
1166 pandecode_blend_equation(&blend->equation);
1167 pandecode_prop("constant = %f", blend->constant);
1168 }
1169
1170 pandecode_indent--;
1171 pandecode_log("},\n");
1172
1173 /* Return blend shader to disassemble if present */
1174 return is_shader ? (blend->shader & ~0xF) : 0;
1175 }
1176
1177 static mali_ptr
1178 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1179 {
1180 struct midgard_blend_rt *b =
1181 ((struct midgard_blend_rt *) descs) + rt_no;
1182
1183 /* Flags determine presence of blend shader */
1184 bool is_shader = (b->flags & 0xF) >= 0x2;
1185
1186 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1187 pandecode_indent++;
1188
1189 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1190
1191 union midgard_blend blend = b->blend;
1192 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1193
1194 pandecode_indent--;
1195 pandecode_log("};\n");
1196
1197 return shader;
1198 }
1199
1200 static int
1201 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1202 {
1203 char base[128];
1204 char *prefix = varying ? "varying" : "attribute";
1205 unsigned max_index = 0;
1206 snprintf(base, sizeof(base), "%s_meta", prefix);
1207
1208 pandecode_log("struct mali_attr_meta %s_%d%s[] = {\n", base, job_no, suffix);
1209 pandecode_indent++;
1210
1211 struct mali_attr_meta *attr_meta;
1212 mali_ptr p = varying ? (v->varying_meta & ~0xF) : v->attribute_meta;
1213
1214 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1215
1216 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1217 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1218 sizeof(*attr_mem));
1219
1220 pandecode_log("{\n");
1221 pandecode_indent++;
1222 pandecode_prop("index = %d", attr_meta->index);
1223
1224 if (attr_meta->index > max_index)
1225 max_index = attr_meta->index;
1226 pandecode_swizzle(attr_meta->swizzle);
1227 pandecode_prop("format = %s", pandecode_format(attr_meta->format));
1228
1229 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1230 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1231 pandecode_prop("src_offset = %d", attr_meta->src_offset);
1232 pandecode_indent--;
1233 pandecode_log("},\n");
1234
1235 }
1236
1237 pandecode_indent--;
1238 pandecode_log("};\n");
1239
1240 return max_index;
1241 }
1242
1243 static void
1244 pandecode_indices(uintptr_t pindices, uint32_t index_count, int job_no)
1245 {
1246 struct pandecode_mapped_memory *imem = pandecode_find_mapped_gpu_mem_containing(pindices);
1247
1248 if (imem) {
1249 /* Indices are literally just a u32 array :) */
1250
1251 uint32_t *PANDECODE_PTR_VAR(indices, imem, pindices);
1252
1253 pandecode_log("uint32_t indices_%d[] = {\n", job_no);
1254 pandecode_indent++;
1255
1256 for (unsigned i = 0; i < (index_count + 1); i += 3)
1257 pandecode_log("%d, %d, %d,\n",
1258 indices[i],
1259 indices[i + 1],
1260 indices[i + 2]);
1261
1262 pandecode_indent--;
1263 pandecode_log("};\n");
1264 }
1265 }
1266
1267 /* return bits [lo, hi) of word */
1268 static u32
1269 bits(u32 word, u32 lo, u32 hi)
1270 {
1271 if (hi - lo >= 32)
1272 return word; // avoid undefined behavior with the shift
1273
1274 return (word >> lo) & ((1 << (hi - lo)) - 1);
1275 }
1276
1277 static void
1278 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no)
1279 {
1280 pandecode_log_cont("{\n");
1281 pandecode_indent++;
1282
1283 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1284 pandecode_prop("size_y_shift = %d", p->size_y_shift);
1285 pandecode_prop("size_z_shift = %d", p->size_z_shift);
1286 pandecode_prop("workgroups_x_shift = %d", p->workgroups_x_shift);
1287 pandecode_prop("workgroups_y_shift = %d", p->workgroups_y_shift);
1288 pandecode_prop("workgroups_z_shift = %d", p->workgroups_z_shift);
1289 pandecode_prop("workgroups_x_shift_2 = 0x%" PRIx32, p->workgroups_x_shift_2);
1290
1291 /* Decode invocation_count. See the comment before the definition of
1292 * invocation_count for an explanation.
1293 */
1294 pandecode_msg("size: (%d, %d, %d)\n",
1295 bits(p->invocation_count, 0, p->size_y_shift) + 1,
1296 bits(p->invocation_count, p->size_y_shift, p->size_z_shift) + 1,
1297 bits(p->invocation_count, p->size_z_shift,
1298 p->workgroups_x_shift) + 1);
1299 pandecode_msg("workgroups: (%d, %d, %d)\n",
1300 bits(p->invocation_count, p->workgroups_x_shift,
1301 p->workgroups_y_shift) + 1,
1302 bits(p->invocation_count, p->workgroups_y_shift,
1303 p->workgroups_z_shift) + 1,
1304 bits(p->invocation_count, p->workgroups_z_shift,
1305 32) + 1);
1306
1307 /* TODO: Decode */
1308 if (p->unknown_draw)
1309 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1310
1311 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1312
1313 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
1314
1315 /* Index count only exists for tiler jobs anyway */
1316
1317 if (p->index_count)
1318 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1319
1320 if (p->negative_start)
1321 pandecode_prop("negative_start = %d", p->negative_start);
1322
1323 DYN_MEMORY_PROP(p, job_no, indices);
1324
1325 if (p->zero1) {
1326 pandecode_msg("Zero tripped\n");
1327 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1328 }
1329
1330 pandecode_indent--;
1331 pandecode_log("},\n");
1332 }
1333
1334 static void
1335 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1336 {
1337 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1338
1339 struct mali_uniform_buffer_meta *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1340
1341 for (int i = 0; i < ubufs_count; i++) {
1342 mali_ptr ptr = ubufs[i].ptr << 2;
1343 struct pandecode_mapped_memory *umem2 = pandecode_find_mapped_gpu_mem_containing(ptr);
1344 uint32_t *PANDECODE_PTR_VAR(ubuf, umem2, ptr);
1345 char name[50];
1346 snprintf(name, sizeof(name), "ubuf_%d", i);
1347 /* The blob uses ubuf 0 to upload internal stuff and
1348 * uniforms that won't fit/are accessed indirectly, so
1349 * it puts it in the batchbuffer.
1350 */
1351 pandecode_log("uint32_t %s_%d[] = {\n", name, job_no);
1352 pandecode_indent++;
1353
1354 for (int j = 0; j <= ubufs[i].size; j++) {
1355 for (int k = 0; k < 4; k++) {
1356 if (k == 0)
1357 pandecode_log("0x%"PRIx32", ", ubuf[4 * j + k]);
1358 else
1359 pandecode_log_cont("0x%"PRIx32", ", ubuf[4 * j + k]);
1360
1361 }
1362
1363 pandecode_log_cont("\n");
1364 }
1365
1366 pandecode_indent--;
1367 pandecode_log("};\n");
1368 }
1369
1370 pandecode_log("struct mali_uniform_buffer_meta uniform_buffers_%d[] = {\n",
1371 job_no);
1372 pandecode_indent++;
1373
1374 for (int i = 0; i < ubufs_count; i++) {
1375 pandecode_log("{\n");
1376 pandecode_indent++;
1377 pandecode_prop("size = MALI_POSITIVE(%d)", ubufs[i].size + 1);
1378 pandecode_prop("ptr = ubuf_%d_%d_p >> 2", i, job_no);
1379 pandecode_indent--;
1380 pandecode_log("},\n");
1381 }
1382
1383 pandecode_indent--;
1384 pandecode_log("};\n");
1385 }
1386
1387 static void
1388 pandecode_scratchpad(uintptr_t pscratchpad, int job_no, char *suffix)
1389 {
1390
1391 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(pscratchpad);
1392
1393 struct bifrost_scratchpad *PANDECODE_PTR_VAR(scratchpad, mem, pscratchpad);
1394
1395 if (scratchpad->zero)
1396 pandecode_msg("XXX scratchpad zero tripped");
1397
1398 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64"_%d%s = {\n", pscratchpad, job_no, suffix);
1399 pandecode_indent++;
1400
1401 pandecode_prop("flags = 0x%x", scratchpad->flags);
1402 MEMORY_PROP(scratchpad, gpu_scratchpad);
1403
1404 pandecode_indent--;
1405 pandecode_log("};\n");
1406 }
1407
1408 static void
1409 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1410 bool is_bifrost)
1411 {
1412 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1413 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1414
1415 /* Compute maximum possible size */
1416 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1417
1418 /* Print some boilerplate to clearly denote the assembly (which doesn't
1419 * obey indentation rules), and actually do the disassembly! */
1420
1421 printf("\n\n");
1422
1423 if (is_bifrost) {
1424 disassemble_bifrost(code, sz, false);
1425 } else {
1426 disassemble_midgard(code, sz);
1427 }
1428
1429 printf("\n\n");
1430 }
1431
1432 static void
1433 pandecode_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix *p,
1434 int job_no, enum mali_job_type job_type,
1435 char *suffix, bool is_bifrost)
1436 {
1437 mali_ptr shader_meta_ptr = (u64) (uintptr_t) (p->_shader_upper << 4);
1438 struct pandecode_mapped_memory *attr_mem;
1439
1440 unsigned rt_count = 1;
1441
1442 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1443 * are the only things actually needed from the FBD, vertex/tiler jobs
1444 * no longer reference the FBD -- instead, this field points to some
1445 * info about the scratchpad.
1446 */
1447 if (is_bifrost)
1448 pandecode_scratchpad(p->framebuffer & ~FBD_TYPE, job_no, suffix);
1449 else if (p->framebuffer & MALI_MFBD)
1450 rt_count = pandecode_mfbd_bfr((u64) ((uintptr_t) p->framebuffer) & FBD_MASK, job_no, false);
1451 else if (job_type == JOB_TYPE_COMPUTE)
1452 pandecode_compute_fbd((u64) (uintptr_t) p->framebuffer, job_no);
1453 else
1454 pandecode_sfbd((u64) (uintptr_t) p->framebuffer, job_no);
1455
1456 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
1457 int texture_count = 0, sampler_count = 0;
1458
1459 if (shader_meta_ptr) {
1460 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(shader_meta_ptr);
1461 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, shader_meta_ptr);
1462
1463 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", shader_meta_ptr, job_no, suffix);
1464 pandecode_indent++;
1465
1466 /* Save for dumps */
1467 attribute_count = s->attribute_count;
1468 varying_count = s->varying_count;
1469 texture_count = s->texture_count;
1470 sampler_count = s->sampler_count;
1471
1472 if (is_bifrost) {
1473 uniform_count = s->bifrost2.uniform_count;
1474 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
1475 } else {
1476 uniform_count = s->midgard1.uniform_count;
1477 uniform_buffer_count = s->midgard1.uniform_buffer_count;
1478 }
1479
1480 mali_ptr shader_ptr = pandecode_shader_address("shader", s->shader);
1481
1482 pandecode_prop("texture_count = %" PRId16, s->texture_count);
1483 pandecode_prop("sampler_count = %" PRId16, s->sampler_count);
1484 pandecode_prop("attribute_count = %" PRId16, s->attribute_count);
1485 pandecode_prop("varying_count = %" PRId16, s->varying_count);
1486
1487 if (is_bifrost) {
1488 pandecode_log(".bifrost1 = {\n");
1489 pandecode_indent++;
1490
1491 pandecode_prop("uniform_buffer_count = %" PRId32, s->bifrost1.uniform_buffer_count);
1492 pandecode_prop("unk1 = 0x%" PRIx32, s->bifrost1.unk1);
1493
1494 pandecode_indent--;
1495 pandecode_log("},\n");
1496 } else {
1497 pandecode_log(".midgard1 = {\n");
1498 pandecode_indent++;
1499
1500 pandecode_prop("uniform_count = %" PRId16, s->midgard1.uniform_count);
1501 pandecode_prop("uniform_buffer_count = %" PRId16, s->midgard1.uniform_buffer_count);
1502 pandecode_prop("work_count = %" PRId16, s->midgard1.work_count);
1503
1504 pandecode_log(".flags = ");
1505 pandecode_log_decoded_flags(shader_midgard1_flag_info, s->midgard1.flags);
1506 pandecode_log_cont(",\n");
1507
1508 pandecode_prop("unknown2 = 0x%" PRIx32, s->midgard1.unknown2);
1509
1510 pandecode_indent--;
1511 pandecode_log("},\n");
1512 }
1513
1514 if (s->depth_units || s->depth_factor) {
1515 pandecode_prop("depth_factor = %f", s->depth_factor);
1516 pandecode_prop("depth_units = %f", s->depth_units);
1517 }
1518
1519 if (s->alpha_coverage) {
1520 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
1521 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
1522
1523 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
1524 invert_alpha_coverage ? "~" : "",
1525 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
1526 }
1527
1528 if (s->unknown2_3 || s->unknown2_4) {
1529 pandecode_log(".unknown2_3 = ");
1530
1531 int unknown2_3 = s->unknown2_3;
1532 int unknown2_4 = s->unknown2_4;
1533
1534 /* We're not quite sure what these flags mean without the depth test, if anything */
1535
1536 if (unknown2_3 & (MALI_DEPTH_TEST | MALI_DEPTH_FUNC_MASK)) {
1537 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
1538 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
1539
1540 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
1541 }
1542
1543 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
1544 pandecode_log_cont(",\n");
1545
1546 pandecode_log(".unknown2_4 = ");
1547 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
1548 pandecode_log_cont(",\n");
1549 }
1550
1551 if (s->stencil_mask_front || s->stencil_mask_back) {
1552 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
1553 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
1554 }
1555
1556 pandecode_stencil("front", &s->stencil_front);
1557 pandecode_stencil("back", &s->stencil_back);
1558
1559 if (is_bifrost) {
1560 pandecode_log(".bifrost2 = {\n");
1561 pandecode_indent++;
1562
1563 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
1564 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
1565 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
1566 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
1567
1568 pandecode_indent--;
1569 pandecode_log("},\n");
1570 } else if (s->midgard2.unknown2_7) {
1571 pandecode_log(".midgard2 = {\n");
1572 pandecode_indent++;
1573
1574 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
1575 pandecode_indent--;
1576 pandecode_log("},\n");
1577 }
1578
1579 if (s->unknown2_8)
1580 pandecode_prop("unknown2_8 = 0x%" PRIx32, s->unknown2_8);
1581
1582 if (!is_bifrost) {
1583 /* TODO: Blend shaders routing/disasm */
1584
1585 union midgard_blend blend = s->blend;
1586 pandecode_midgard_blend(&blend, false);
1587 }
1588
1589 pandecode_indent--;
1590 pandecode_log("};\n");
1591
1592 /* MRT blend fields are used whenever MFBD is used, with
1593 * per-RT descriptors */
1594
1595 if (job_type == JOB_TYPE_TILER) {
1596 void* blend_base = (void *) (s + 1);
1597
1598 for (unsigned i = 0; i < rt_count; i++) {
1599 mali_ptr shader = 0;
1600
1601 if (is_bifrost)
1602 shader = pandecode_bifrost_blend(blend_base, job_no, i);
1603 else
1604 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
1605
1606 if (shader)
1607 pandecode_shader_disassemble(shader, job_no, job_type, false);
1608 }
1609 }
1610
1611 pandecode_shader_disassemble(shader_ptr, job_no, job_type, is_bifrost);
1612 } else
1613 pandecode_msg("<no shader>\n");
1614
1615 if (p->viewport) {
1616 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
1617 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
1618
1619 pandecode_log("struct mali_viewport viewport_%d%s = {\n", job_no, suffix);
1620 pandecode_indent++;
1621
1622 pandecode_prop("clip_minx = %f", f->clip_minx);
1623 pandecode_prop("clip_miny = %f", f->clip_miny);
1624 pandecode_prop("clip_minz = %f", f->clip_minz);
1625 pandecode_prop("clip_maxx = %f", f->clip_maxx);
1626 pandecode_prop("clip_maxy = %f", f->clip_maxy);
1627 pandecode_prop("clip_maxz = %f", f->clip_maxz);
1628
1629 /* Only the higher coordinates are MALI_POSITIVE scaled */
1630
1631 pandecode_prop("viewport0 = { %d, %d }",
1632 f->viewport0[0], f->viewport0[1]);
1633
1634 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
1635 f->viewport1[0] + 1, f->viewport1[1] + 1);
1636
1637 pandecode_indent--;
1638 pandecode_log("};\n");
1639 }
1640
1641 if (p->attribute_meta) {
1642 unsigned max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
1643
1644 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
1645 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index + 1, false);
1646 }
1647
1648 /* Varyings are encoded like attributes but not actually sent; we just
1649 * pass a zero buffer with the right stride/size set, (or whatever)
1650 * since the GPU will write to it itself */
1651
1652 if (p->varyings) {
1653 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
1654
1655 /* Number of descriptors depends on whether there are
1656 * non-internal varyings */
1657
1658 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count > 1 ? 4 : 1, true);
1659 }
1660
1661 if (p->varying_meta) {
1662 pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
1663 }
1664
1665 bool is_compute = job_type == JOB_TYPE_COMPUTE;
1666
1667 if (p->uniforms && !is_compute) {
1668 int rows = uniform_count, width = 4;
1669 size_t sz = rows * width * sizeof(float);
1670
1671 struct pandecode_mapped_memory *uniform_mem = pandecode_find_mapped_gpu_mem_containing(p->uniforms);
1672 pandecode_fetch_gpu_mem(uniform_mem, p->uniforms, sz);
1673 u32 *PANDECODE_PTR_VAR(uniforms, uniform_mem, p->uniforms);
1674
1675 pandecode_log("u32 uniforms_%d%s[] = {\n", job_no, suffix);
1676
1677 pandecode_indent++;
1678
1679 for (int row = 0; row < rows; row++) {
1680 for (int i = 0; i < width; i++) {
1681 u32 v = uniforms[i];
1682 float f;
1683 memcpy(&f, &v, sizeof(v));
1684 pandecode_log_cont("%X /* %f */, ", v, f);
1685 }
1686
1687 pandecode_log_cont("\n");
1688
1689 uniforms += width;
1690 }
1691
1692 pandecode_indent--;
1693 pandecode_log("};\n");
1694 } else if (p->uniforms) {
1695 int rows = uniform_count * 2;
1696 size_t sz = rows * sizeof(mali_ptr);
1697
1698 struct pandecode_mapped_memory *uniform_mem = pandecode_find_mapped_gpu_mem_containing(p->uniforms);
1699 pandecode_fetch_gpu_mem(uniform_mem, p->uniforms, sz);
1700 mali_ptr *PANDECODE_PTR_VAR(uniforms, uniform_mem, p->uniforms);
1701
1702 pandecode_log("mali_ptr uniforms_%d%s[] = {\n", job_no, suffix);
1703
1704 pandecode_indent++;
1705
1706 for (int row = 0; row < rows; row++) {
1707 char *a = pointer_as_memory_reference(uniforms[row]);
1708 pandecode_log("%s,\n", a);
1709 free(a);
1710 }
1711
1712 pandecode_indent--;
1713 pandecode_log("};\n");
1714
1715 }
1716
1717 if (p->uniform_buffers) {
1718 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
1719 }
1720
1721 if (p->texture_trampoline) {
1722 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(p->texture_trampoline);
1723
1724 if (mmem) {
1725 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline);
1726
1727 pandecode_log("uint64_t texture_trampoline_%d[] = {\n", job_no);
1728 pandecode_indent++;
1729
1730 for (int tex = 0; tex < texture_count; ++tex) {
1731 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
1732 char *a = pointer_as_memory_reference(*u);
1733 pandecode_log("%s,\n", a);
1734 free(a);
1735 }
1736
1737 pandecode_indent--;
1738 pandecode_log("};\n");
1739
1740 /* Now, finally, descend down into the texture descriptor */
1741 for (int tex = 0; tex < texture_count; ++tex) {
1742 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
1743 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
1744
1745 if (tmem) {
1746 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, *u);
1747
1748 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", *u, job_no, tex);
1749 pandecode_indent++;
1750
1751 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", t->width + 1);
1752 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", t->height + 1);
1753 pandecode_prop("depth = MALI_POSITIVE(%" PRId16 ")", t->depth + 1);
1754 pandecode_prop("array_size = MALI_POSITIVE(%" PRId16 ")", t->array_size + 1);
1755 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
1756 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
1757 pandecode_prop("nr_mipmap_levels = %" PRId8, t->nr_mipmap_levels);
1758
1759 struct mali_texture_format f = t->format;
1760
1761 pandecode_log(".format = {\n");
1762 pandecode_indent++;
1763
1764 pandecode_swizzle(f.swizzle);
1765 pandecode_prop("format = %s", pandecode_format(f.format));
1766 pandecode_prop("type = %s", pandecode_texture_type(f.type));
1767 pandecode_prop("srgb = %" PRId32, f.srgb);
1768 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
1769 pandecode_prop("usage2 = 0x%" PRIx32, f.usage2);
1770
1771 pandecode_indent--;
1772 pandecode_log("},\n");
1773
1774 pandecode_swizzle(t->swizzle);
1775
1776 if (t->swizzle_zero) {
1777 /* Shouldn't happen */
1778 pandecode_msg("Swizzle zero tripped but replay will be fine anyway");
1779 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
1780 }
1781
1782 pandecode_prop("unknown3 = 0x%" PRIx32, t->unknown3);
1783
1784 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
1785 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
1786 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
1787
1788 pandecode_log(".payload = {\n");
1789 pandecode_indent++;
1790
1791 /* A bunch of bitmap pointers follow.
1792 * We work out the correct number,
1793 * based on the mipmap/cubemap
1794 * properties, but dump extra
1795 * possibilities to futureproof */
1796
1797 int bitmap_count = MALI_NEGATIVE(t->nr_mipmap_levels);
1798 bool manual_stride = f.usage2 & MALI_TEX_MANUAL_STRIDE;
1799
1800 /* Miptree for each face */
1801 if (f.type == MALI_TEX_CUBE)
1802 bitmap_count *= 6;
1803
1804 /* Array of textures */
1805 bitmap_count *= MALI_NEGATIVE(t->array_size);
1806
1807 /* Stride for each element */
1808 if (manual_stride)
1809 bitmap_count *= 2;
1810
1811 /* Sanity check the size */
1812 int max_count = sizeof(t->payload) / sizeof(t->payload[0]);
1813 assert (bitmap_count <= max_count);
1814
1815 /* Dump more to be safe, but not _that_ much more */
1816 int safe_count = MIN2(bitmap_count * 2, max_count);
1817
1818 for (int i = 0; i < safe_count; ++i) {
1819 char *prefix = (i >= bitmap_count) ? "// " : "";
1820
1821 /* How we dump depends if this is a stride or a pointer */
1822
1823 if ((f.usage2 & MALI_TEX_MANUAL_STRIDE) && (i & 1)) {
1824 /* signed 32-bit snuck in as a 64-bit pointer */
1825 uint64_t stride_set = t->payload[i];
1826 uint32_t clamped_stride = stride_set;
1827 int32_t stride = clamped_stride;
1828 assert(stride_set == clamped_stride);
1829 pandecode_log("%s(mali_ptr) %d /* stride */, \n", prefix, stride);
1830 } else {
1831 char *a = pointer_as_memory_reference(t->payload[i]);
1832 pandecode_log("%s%s, \n", prefix, a);
1833 free(a);
1834 }
1835 }
1836
1837 pandecode_indent--;
1838 pandecode_log("},\n");
1839
1840 pandecode_indent--;
1841 pandecode_log("};\n");
1842 }
1843 }
1844 }
1845 }
1846
1847 if (p->sampler_descriptor) {
1848 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->sampler_descriptor);
1849
1850 if (smem) {
1851 struct mali_sampler_descriptor *s;
1852
1853 mali_ptr d = p->sampler_descriptor;
1854
1855 for (int i = 0; i < sampler_count; ++i) {
1856 s = pandecode_fetch_gpu_mem(smem, d + sizeof(*s) * i, sizeof(*s));
1857
1858 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%d_%d = {\n", job_no, i);
1859 pandecode_indent++;
1860
1861 /* Only the lower two bits are understood right now; the rest we display as hex */
1862 pandecode_log(".filter_mode = MALI_TEX_MIN(%s) | MALI_TEX_MAG(%s) | 0x%" PRIx32",\n",
1863 MALI_FILTER_NAME(s->filter_mode & MALI_TEX_MIN_MASK),
1864 MALI_FILTER_NAME(s->filter_mode & MALI_TEX_MAG_MASK),
1865 s->filter_mode & ~3);
1866
1867 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
1868 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
1869
1870 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
1871 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
1872 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
1873
1874 pandecode_prop("compare_func = %s", pandecode_alt_func(s->compare_func));
1875
1876 if (s->zero || s->zero2) {
1877 pandecode_msg("Zero tripped\n");
1878 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
1879 }
1880
1881 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
1882
1883 pandecode_prop("border_color = { %f, %f, %f, %f }",
1884 s->border_color[0],
1885 s->border_color[1],
1886 s->border_color[2],
1887 s->border_color[3]);
1888
1889 pandecode_indent--;
1890 pandecode_log("};\n");
1891 }
1892 }
1893 }
1894 }
1895
1896 static void
1897 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
1898 {
1899 pandecode_log_cont("{\n");
1900 pandecode_indent++;
1901
1902 MEMORY_PROP(p, position_varying);
1903 DYN_MEMORY_PROP(p, job_no, uniform_buffers);
1904 DYN_MEMORY_PROP(p, job_no, texture_trampoline);
1905 DYN_MEMORY_PROP(p, job_no, sampler_descriptor);
1906 DYN_MEMORY_PROP(p, job_no, uniforms);
1907 DYN_MEMORY_PROP(p, job_no, attributes);
1908 DYN_MEMORY_PROP(p, job_no, attribute_meta);
1909 DYN_MEMORY_PROP(p, job_no, varyings);
1910 DYN_MEMORY_PROP(p, job_no, varying_meta);
1911 DYN_MEMORY_PROP(p, job_no, viewport);
1912 DYN_MEMORY_PROP(p, job_no, occlusion_counter);
1913
1914 if (is_bifrost)
1915 pandecode_prop("framebuffer = scratchpad_%d_p", job_no);
1916 else
1917 pandecode_prop("framebuffer = framebuffer_%d_p | %s", job_no, p->framebuffer & MALI_MFBD ? "MALI_MFBD" : "0");
1918
1919 pandecode_prop("_shader_upper = (shader_meta_%d_p) >> 4", job_no);
1920 pandecode_prop("flags = %d", p->flags);
1921
1922 pandecode_indent--;
1923 pandecode_log("},\n");
1924 }
1925
1926 static void
1927 pandecode_vertex_only_bfr(struct bifrost_vertex_only *v)
1928 {
1929 pandecode_log_cont("{\n");
1930 pandecode_indent++;
1931
1932 pandecode_prop("unk2 = 0x%x", v->unk2);
1933
1934 if (v->zero0 || v->zero1) {
1935 pandecode_msg("vertex only zero tripped");
1936 pandecode_prop("zero0 = 0x%" PRIx32, v->zero0);
1937 pandecode_prop("zero1 = 0x%" PRIx64, v->zero1);
1938 }
1939
1940 pandecode_indent--;
1941 pandecode_log("}\n");
1942 }
1943
1944 static void
1945 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
1946 {
1947
1948 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1949 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
1950
1951 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no);
1952 pandecode_indent++;
1953
1954 if (h->zero) {
1955 pandecode_msg("tiler heap zero tripped\n");
1956 pandecode_prop("zero = 0x%x", h->zero);
1957 }
1958
1959 for (int i = 0; i < 12; i++) {
1960 if (h->zeros[i] != 0) {
1961 pandecode_msg("tiler heap zero %d tripped, value %x\n",
1962 i, h->zeros[i]);
1963 }
1964 }
1965
1966 pandecode_prop("heap_size = 0x%x", h->heap_size);
1967 MEMORY_PROP(h, tiler_heap_start);
1968 MEMORY_PROP(h, tiler_heap_free);
1969
1970 /* this might point to the beginning of another buffer, when it's
1971 * really the end of the tiler heap buffer, so we have to be careful
1972 * here.
1973 */
1974 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
1975 pandecode_prop("tiler_heap_end = %s + 1", a);
1976 free(a);
1977
1978 pandecode_indent--;
1979 pandecode_log("};\n");
1980 }
1981
1982 static void
1983 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
1984 {
1985 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1986 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
1987
1988 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
1989
1990 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no);
1991 pandecode_indent++;
1992
1993 if (t->zero0 || t->zero1) {
1994 pandecode_msg("tiler meta zero tripped");
1995 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
1996 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
1997 }
1998
1999 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2000 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2001
2002 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2003 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2004 DYN_MEMORY_PROP(t, job_no, tiler_heap_meta);
2005
2006 for (int i = 0; i < 12; i++) {
2007 if (t->zeros[i] != 0) {
2008 pandecode_msg("tiler heap zero %d tripped, value %" PRIx64 "\n",
2009 i, t->zeros[i]);
2010 }
2011 }
2012
2013 pandecode_indent--;
2014 pandecode_log("};\n");
2015 }
2016
2017 static void
2018 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2019 {
2020 pandecode_log(".gl_enables = ");
2021
2022 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2023
2024 pandecode_log_cont(",\n");
2025 }
2026
2027 static void
2028 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2029 {
2030 if (u.pointer == 0x0)
2031 return;
2032
2033 pandecode_log(".primitive_size = {\n");
2034 pandecode_indent++;
2035
2036 if (constant) {
2037 pandecode_prop("constant = %f", u.constant);
2038 } else {
2039 MEMORY_PROP((&u), pointer);
2040 }
2041
2042 pandecode_indent--;
2043 pandecode_log("},\n");
2044 }
2045
2046 static void
2047 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2048 {
2049 pandecode_log_cont("{\n");
2050 pandecode_indent++;
2051
2052 /* TODO: gl_PointSize on Bifrost */
2053 pandecode_primitive_size(t->primitive_size, true);
2054
2055 DYN_MEMORY_PROP(t, job_no, tiler_meta);
2056 pandecode_gl_enables(t->gl_enables, JOB_TYPE_TILER);
2057
2058 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2059 || t->zero6 || t->zero7 || t->zero8) {
2060 pandecode_msg("tiler only zero tripped");
2061 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2062 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2063 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2064 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2065 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2066 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2067 pandecode_prop("zero7 = 0x%" PRIx32, t->zero7);
2068 pandecode_prop("zero8 = 0x%" PRIx64, t->zero8);
2069 }
2070
2071 pandecode_indent--;
2072 pandecode_log("},\n");
2073 }
2074
2075 static int
2076 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2077 const struct pandecode_mapped_memory *mem,
2078 mali_ptr payload, int job_no)
2079 {
2080 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2081
2082 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true);
2083
2084 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no);
2085 pandecode_indent++;
2086
2087 pandecode_log(".prefix = ");
2088 pandecode_vertex_tiler_prefix(&v->prefix, job_no);
2089
2090 pandecode_log(".vertex = ");
2091 pandecode_vertex_only_bfr(&v->vertex);
2092
2093 pandecode_log(".postfix = ");
2094 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2095
2096 pandecode_indent--;
2097 pandecode_log("};\n");
2098
2099 return sizeof(*v);
2100 }
2101
2102 static int
2103 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2104 const struct pandecode_mapped_memory *mem,
2105 mali_ptr payload, int job_no)
2106 {
2107 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2108
2109 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true);
2110
2111 pandecode_indices(t->prefix.indices, t->prefix.index_count, job_no);
2112 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2113
2114 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no);
2115 pandecode_indent++;
2116
2117 pandecode_log(".prefix = ");
2118 pandecode_vertex_tiler_prefix(&t->prefix, job_no);
2119
2120 pandecode_log(".tiler = ");
2121 pandecode_tiler_only_bfr(&t->tiler, job_no);
2122
2123 pandecode_log(".postfix = ");
2124 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2125
2126 pandecode_indent--;
2127 pandecode_log("};\n");
2128
2129 return sizeof(*t);
2130 }
2131
2132 static int
2133 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2134 const struct pandecode_mapped_memory *mem,
2135 mali_ptr payload, int job_no)
2136 {
2137 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2138
2139 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false);
2140
2141 pandecode_indices(v->prefix.indices, v->prefix.index_count, job_no);
2142
2143 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2144 pandecode_indent++;
2145
2146 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2147 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
2148
2149 pandecode_log(".prefix = ");
2150 pandecode_vertex_tiler_prefix(&v->prefix, job_no);
2151
2152 pandecode_gl_enables(v->gl_enables, h->job_type);
2153
2154 if (v->instance_shift || v->instance_odd) {
2155 pandecode_prop("instance_shift = 0x%d /* %d */",
2156 v->instance_shift, 1 << v->instance_shift);
2157 pandecode_prop("instance_odd = 0x%X /* %d */",
2158 v->instance_odd, (2 * v->instance_odd) + 1);
2159
2160 pandecode_padded_vertices(v->instance_shift, v->instance_odd);
2161 }
2162
2163 if (v->draw_start)
2164 pandecode_prop("draw_start = %d", v->draw_start);
2165
2166 if (v->zero5) {
2167 pandecode_msg("Zero tripped\n");
2168 pandecode_prop("zero5 = 0x%" PRIx64, v->zero5);
2169 }
2170
2171 pandecode_log(".postfix = ");
2172 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2173
2174 pandecode_indent--;
2175 pandecode_log("};\n");
2176
2177 return sizeof(*v);
2178 }
2179
2180 static int
2181 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2182 mali_ptr payload, int job_no,
2183 bool is_bifrost)
2184 {
2185 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2186
2187 bool fbd_dumped = false;
2188
2189 if (!is_bifrost && (s->framebuffer & FBD_TYPE) == MALI_SFBD) {
2190 /* Only SFBDs are understood, not MFBDs. We're speculating,
2191 * based on the versioning, kernel code, etc, that the
2192 * difference is between Single FrameBuffer Descriptor and
2193 * Multiple FrmaeBuffer Descriptor; the change apparently lines
2194 * up with multi-framebuffer support being added (T7xx onwards,
2195 * including Gxx). In any event, there's some field shuffling
2196 * that we haven't looked into yet. */
2197
2198 pandecode_sfbd(s->framebuffer & FBD_MASK, job_no);
2199 fbd_dumped = true;
2200 } else if ((s->framebuffer & FBD_TYPE) == MALI_MFBD) {
2201 /* We don't know if Bifrost supports SFBD's at all, since the
2202 * driver never uses them. And the format is different from
2203 * Midgard anyways, due to the tiler heap and scratchpad being
2204 * moved out into separate structures, so it's not clear what a
2205 * Bifrost SFBD would even look like without getting an actual
2206 * trace, which appears impossible.
2207 */
2208
2209 pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true);
2210 fbd_dumped = true;
2211 }
2212
2213 uintptr_t p = (uintptr_t) s->framebuffer & FBD_MASK;
2214 pandecode_log("struct mali_payload_fragment payload_%"PRIx64"_%d = {\n", payload, job_no);
2215 pandecode_indent++;
2216
2217 /* See the comments by the macro definitions for mathematical context
2218 * on why this is so weird */
2219
2220 if (MALI_TILE_COORD_FLAGS(s->max_tile_coord) || MALI_TILE_COORD_FLAGS(s->min_tile_coord))
2221 pandecode_msg("Tile coordinate flag missed, replay wrong\n");
2222
2223 pandecode_prop("min_tile_coord = MALI_COORDINATE_TO_TILE_MIN(%d, %d)",
2224 MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT,
2225 MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT);
2226
2227 pandecode_prop("max_tile_coord = MALI_COORDINATE_TO_TILE_MAX(%d, %d)",
2228 (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT,
2229 (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT);
2230
2231 /* If the FBD was just decoded, we can refer to it by pointer. If not,
2232 * we have to fallback on offsets. */
2233
2234 const char *fbd_type = s->framebuffer & MALI_MFBD ? "MALI_MFBD" : "MALI_SFBD";
2235
2236 if (fbd_dumped)
2237 pandecode_prop("framebuffer = framebuffer_%d_p | %s", job_no, fbd_type);
2238 else
2239 pandecode_prop("framebuffer = %s | %s", pointer_as_memory_reference(p), fbd_type);
2240
2241 pandecode_indent--;
2242 pandecode_log("};\n");
2243
2244 return sizeof(*s);
2245 }
2246
2247 static int job_descriptor_number = 0;
2248
2249 int
2250 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost)
2251 {
2252 struct mali_job_descriptor_header *h;
2253
2254 int start_number = 0;
2255
2256 bool first = true;
2257 bool last_size;
2258
2259 do {
2260 struct pandecode_mapped_memory *mem =
2261 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2262
2263 void *payload;
2264
2265 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2266
2267 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2268 * high 32-bits of the 64-bit pointer are reused to store
2269 * something else.
2270 */
2271 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2272 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
2273 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2274
2275 payload = pandecode_fetch_gpu_mem(mem, payload_ptr,
2276 MALI_PAYLOAD_SIZE);
2277
2278 int job_no = job_descriptor_number++;
2279
2280 if (first)
2281 start_number = job_no;
2282
2283 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2284 pandecode_indent++;
2285
2286 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
2287
2288 /* Save for next job fixing */
2289 last_size = h->job_descriptor_size;
2290
2291 if (h->job_descriptor_size)
2292 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2293
2294 if (h->exception_status != 0x1)
2295 pandecode_prop("exception_status = %x (source ID: 0x%x access: 0x%x exception: 0x%x)",
2296 h->exception_status,
2297 (h->exception_status >> 16) & 0xFFFF,
2298 (h->exception_status >> 8) & 0x3,
2299 h->exception_status & 0xFF);
2300
2301 if (h->first_incomplete_task)
2302 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2303
2304 if (h->fault_pointer)
2305 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2306
2307 if (h->job_barrier)
2308 pandecode_prop("job_barrier = %d", h->job_barrier);
2309
2310 pandecode_prop("job_index = %d", h->job_index);
2311
2312 if (h->unknown_flags)
2313 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2314
2315 if (h->job_dependency_index_1)
2316 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2317
2318 if (h->job_dependency_index_2)
2319 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2320
2321 pandecode_indent--;
2322 pandecode_log("};\n");
2323
2324 /* Do not touch the field yet -- decode the payload first, and
2325 * don't touch that either. This is essential for the uploads
2326 * to occur in sequence and therefore be dynamically allocated
2327 * correctly. Do note the size, however, for that related
2328 * reason. */
2329
2330 switch (h->job_type) {
2331 case JOB_TYPE_SET_VALUE: {
2332 struct mali_payload_set_value *s = payload;
2333 pandecode_log("struct mali_payload_set_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2334 pandecode_indent++;
2335 MEMORY_PROP(s, out);
2336 pandecode_prop("unknown = 0x%" PRIX64, s->unknown);
2337 pandecode_indent--;
2338 pandecode_log("};\n");
2339
2340 break;
2341 }
2342
2343 case JOB_TYPE_TILER:
2344 case JOB_TYPE_VERTEX:
2345 case JOB_TYPE_COMPUTE:
2346 if (bifrost) {
2347 if (h->job_type == JOB_TYPE_TILER)
2348 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no);
2349 else
2350 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no);
2351 } else
2352 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no);
2353
2354 break;
2355
2356 case JOB_TYPE_FRAGMENT:
2357 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost);
2358 break;
2359
2360 default:
2361 break;
2362 }
2363
2364 /* Handle linkage */
2365
2366 if (!first) {
2367 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no - 1);
2368
2369 if (last_size)
2370 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no);
2371 else
2372 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no);
2373 }
2374
2375 first = false;
2376
2377 } while ((jc_gpu_va = h->job_descriptor_size ? h->next_job_64 : h->next_job_32));
2378
2379 return start_number;
2380 }