projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
pan/decode: Handle VARYING_DISCARD
[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/macros.h"
34 #include "util/u_math.h"
35
36 #include "pan_pretty_print.h"
37 #include "midgard/disassemble.h"
38 #include "bifrost/disassemble.h"
39
40 #include "pan_encoder.h"
41
42 int pandecode_jc(mali_ptr jc_gpu_va, bool bifrost);
43
44 #define MEMORY_PROP(obj, p) {\
45 if (obj->p) { \
46 char *a = pointer_as_memory_reference(obj->p); \
47 pandecode_prop("%s = %s", #p, a); \
48 free(a); \
49 } \
50 }
51
52 #define MEMORY_PROP_DIR(obj, p) {\
53 if (obj.p) { \
54 char *a = pointer_as_memory_reference(obj.p); \
55 pandecode_prop("%s = %s", #p, a); \
56 free(a); \
57 } \
58 }
59
60 /* Semantic logging type.
61 *
62 * Raw: for raw messages to be printed as is.
63 * Message: for helpful information to be commented out in replays.
64 * Property: for properties of a struct
65 *
66 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
67 */
68
69 enum pandecode_log_type {
70 PANDECODE_RAW,
71 PANDECODE_MESSAGE,
72 PANDECODE_PROPERTY
73 };
74
75 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
76 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
77 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
78
79 unsigned pandecode_indent = 0;
80
81 static void
82 pandecode_make_indent(void)
83 {
84 for (unsigned i = 0; i < pandecode_indent; ++i)
85 printf(" ");
86 }
87
88 static void
89 pandecode_log_typed(enum pandecode_log_type type, const char *format, ...)
90 {
91 va_list ap;
92
93 pandecode_make_indent();
94
95 if (type == PANDECODE_MESSAGE)
96 printf("// ");
97 else if (type == PANDECODE_PROPERTY)
98 printf(".");
99
100 va_start(ap, format);
101 vprintf(format, ap);
102 va_end(ap);
103
104 if (type == PANDECODE_PROPERTY)
105 printf(",\n");
106 }
107
108 static void
109 pandecode_log_cont(const char *format, ...)
110 {
111 va_list ap;
112
113 va_start(ap, format);
114 vprintf(format, ap);
115 va_end(ap);
116 }
117
118 /* To check for memory safety issues, validates that the given pointer in GPU
119 * memory is valid, containing at least sz bytes. The goal is to eliminate
120 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
121 * overruns) by statically validating pointers.
122 */
123
124 static void
125 pandecode_validate_buffer(mali_ptr addr, size_t sz)
126 {
127 if (!addr) {
128 pandecode_msg("XXX: null pointer deref");
129 return;
130 }
131
132 /* Find a BO */
133
134 struct pandecode_mapped_memory *bo =
135 pandecode_find_mapped_gpu_mem_containing(addr);
136
137 if (!bo) {
138 pandecode_msg("XXX: invalid memory dereference\n");
139 return;
140 }
141
142 /* Bounds check */
143
144 unsigned offset = addr - bo->gpu_va;
145 unsigned total = offset + sz;
146
147 if (total > bo->length) {
148 pandecode_msg("XXX: buffer overrun."
149 "Chunk of size %d at offset %d in buffer of size %d. "
150 "Overrun by %d bytes.",
151 sz, offset, bo->length, total - bo->length);
152 return;
153 }
154 }
155
156 struct pandecode_flag_info {
157 u64 flag;
158 const char *name;
159 };
160
161 static void
162 pandecode_log_decoded_flags(const struct pandecode_flag_info *flag_info,
163 u64 flags)
164 {
165 bool decodable_flags_found = false;
166
167 for (int i = 0; flag_info[i].name; i++) {
168 if ((flags & flag_info[i].flag) != flag_info[i].flag)
169 continue;
170
171 if (!decodable_flags_found) {
172 decodable_flags_found = true;
173 } else {
174 pandecode_log_cont(" | ");
175 }
176
177 pandecode_log_cont("%s", flag_info[i].name);
178
179 flags &= ~flag_info[i].flag;
180 }
181
182 if (decodable_flags_found) {
183 if (flags)
184 pandecode_log_cont(" | 0x%" PRIx64, flags);
185 } else {
186 pandecode_log_cont("0x%" PRIx64, flags);
187 }
188 }
189
190 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
191 static const struct pandecode_flag_info gl_enable_flag_info[] = {
192 FLAG_INFO(OCCLUSION_QUERY),
193 FLAG_INFO(OCCLUSION_PRECISE),
194 FLAG_INFO(FRONT_CCW_TOP),
195 FLAG_INFO(CULL_FACE_FRONT),
196 FLAG_INFO(CULL_FACE_BACK),
197 {}
198 };
199 #undef FLAG_INFO
200
201 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
202 static const struct pandecode_flag_info clear_flag_info[] = {
203 FLAG_INFO(FAST),
204 FLAG_INFO(SLOW),
205 FLAG_INFO(SLOW_STENCIL),
206 {}
207 };
208 #undef FLAG_INFO
209
210 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
211 static const struct pandecode_flag_info mask_flag_info[] = {
212 FLAG_INFO(R),
213 FLAG_INFO(G),
214 FLAG_INFO(B),
215 FLAG_INFO(A),
216 {}
217 };
218 #undef FLAG_INFO
219
220 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
221 static const struct pandecode_flag_info u3_flag_info[] = {
222 FLAG_INFO(HAS_MSAA),
223 FLAG_INFO(CAN_DISCARD),
224 FLAG_INFO(HAS_BLEND_SHADER),
225 FLAG_INFO(DEPTH_TEST),
226 {}
227 };
228
229 static const struct pandecode_flag_info u4_flag_info[] = {
230 FLAG_INFO(NO_MSAA),
231 FLAG_INFO(NO_DITHER),
232 FLAG_INFO(DEPTH_RANGE_A),
233 FLAG_INFO(DEPTH_RANGE_B),
234 FLAG_INFO(STENCIL_TEST),
235 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER),
236 {}
237 };
238 #undef FLAG_INFO
239
240 #define FLAG_INFO(flag) { MALI_FRAMEBUFFER_##flag, "MALI_FRAMEBUFFER_" #flag }
241 static const struct pandecode_flag_info fb_fmt_flag_info[] = {
242 FLAG_INFO(MSAA_A),
243 FLAG_INFO(MSAA_B),
244 FLAG_INFO(MSAA_8),
245 {}
246 };
247 #undef FLAG_INFO
248
249 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
250 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
251 FLAG_INFO(MSAA),
252 FLAG_INFO(SRGB),
253 {}
254 };
255 #undef FLAG_INFO
256
257 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
258 static const struct pandecode_flag_info mfbd_extra_flag_info[] = {
259 FLAG_INFO(PRESENT),
260 FLAG_INFO(AFBC),
261 FLAG_INFO(ZS),
262 {}
263 };
264 #undef FLAG_INFO
265
266 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
267 static const struct pandecode_flag_info shader_midgard1_flag_info [] = {
268 FLAG_INFO(EARLY_Z),
269 FLAG_INFO(HELPER_INVOCATIONS),
270 FLAG_INFO(READS_TILEBUFFER),
271 FLAG_INFO(READS_ZS),
272 {}
273 };
274 #undef FLAG_INFO
275
276 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
277 static const struct pandecode_flag_info mfbd_flag_info [] = {
278 FLAG_INFO(DEPTH_WRITE),
279 FLAG_INFO(EXTRA),
280 {}
281 };
282 #undef FLAG_INFO
283
284 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
285 static const struct pandecode_flag_info sampler_flag_info [] = {
286 FLAG_INFO(MAG_NEAREST),
287 FLAG_INFO(MIN_NEAREST),
288 FLAG_INFO(MIP_LINEAR_1),
289 FLAG_INFO(MIP_LINEAR_2),
290 FLAG_INFO(NORM_COORDS),
291 {}
292 };
293 #undef FLAG_INFO
294
295 extern char *replace_fragment;
296 extern char *replace_vertex;
297
298 static char *
299 pandecode_job_type(enum mali_job_type type)
300 {
301 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
302
303 switch (type) {
304 DEFINE_CASE(NULL);
305 DEFINE_CASE(SET_VALUE);
306 DEFINE_CASE(CACHE_FLUSH);
307 DEFINE_CASE(COMPUTE);
308 DEFINE_CASE(VERTEX);
309 DEFINE_CASE(TILER);
310 DEFINE_CASE(FUSED);
311 DEFINE_CASE(FRAGMENT);
312
313 case JOB_NOT_STARTED:
314 return "NOT_STARTED";
315
316 default:
317 pandecode_log("Warning! Unknown job type %x\n", type);
318 return "!?!?!?";
319 }
320
321 #undef DEFINE_CASE
322 }
323
324 static char *
325 pandecode_draw_mode(enum mali_draw_mode mode)
326 {
327 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
328
329 switch (mode) {
330 DEFINE_CASE(DRAW_NONE);
331 DEFINE_CASE(POINTS);
332 DEFINE_CASE(LINES);
333 DEFINE_CASE(TRIANGLES);
334 DEFINE_CASE(TRIANGLE_STRIP);
335 DEFINE_CASE(TRIANGLE_FAN);
336 DEFINE_CASE(LINE_STRIP);
337 DEFINE_CASE(LINE_LOOP);
338 DEFINE_CASE(POLYGON);
339 DEFINE_CASE(QUADS);
340 DEFINE_CASE(QUAD_STRIP);
341
342 default:
343 pandecode_msg("XXX: invalid draw mode %X\n", mode);
344 return "";
345 }
346
347 #undef DEFINE_CASE
348 }
349
350 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
351 static char *
352 pandecode_func(enum mali_func mode)
353 {
354 switch (mode) {
355 DEFINE_CASE(NEVER);
356 DEFINE_CASE(LESS);
357 DEFINE_CASE(EQUAL);
358 DEFINE_CASE(LEQUAL);
359 DEFINE_CASE(GREATER);
360 DEFINE_CASE(NOTEQUAL);
361 DEFINE_CASE(GEQUAL);
362 DEFINE_CASE(ALWAYS);
363
364 default:
365 pandecode_msg("XXX: invalid func %X\n", mode);
366 return "";
367 }
368 }
369 #undef DEFINE_CASE
370
371 /* Why is this duplicated? Who knows... */
372 #define DEFINE_CASE(name) case MALI_ALT_FUNC_ ## name: return "MALI_ALT_FUNC_" #name
373 static char *
374 pandecode_alt_func(enum mali_alt_func mode)
375 {
376 switch (mode) {
377 DEFINE_CASE(NEVER);
378 DEFINE_CASE(LESS);
379 DEFINE_CASE(EQUAL);
380 DEFINE_CASE(LEQUAL);
381 DEFINE_CASE(GREATER);
382 DEFINE_CASE(NOTEQUAL);
383 DEFINE_CASE(GEQUAL);
384 DEFINE_CASE(ALWAYS);
385
386 default:
387 pandecode_msg("XXX: invalid alt func %X\n", mode);
388 return "";
389 }
390 }
391 #undef DEFINE_CASE
392
393 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
394 static char *
395 pandecode_stencil_op(enum mali_stencil_op op)
396 {
397 switch (op) {
398 DEFINE_CASE(KEEP);
399 DEFINE_CASE(REPLACE);
400 DEFINE_CASE(ZERO);
401 DEFINE_CASE(INVERT);
402 DEFINE_CASE(INCR_WRAP);
403 DEFINE_CASE(DECR_WRAP);
404 DEFINE_CASE(INCR);
405 DEFINE_CASE(DECR);
406
407 default:
408 pandecode_msg("XXX: invalid stencil op %X\n", op);
409 return "";
410 }
411 }
412
413 #undef DEFINE_CASE
414
415 #define DEFINE_CASE(name) case MALI_ATTR_ ## name: return "MALI_ATTR_" #name
416 static char *pandecode_attr_mode(enum mali_attr_mode mode)
417 {
418 switch(mode) {
419 DEFINE_CASE(UNUSED);
420 DEFINE_CASE(LINEAR);
421 DEFINE_CASE(POT_DIVIDE);
422 DEFINE_CASE(MODULO);
423 DEFINE_CASE(NPOT_DIVIDE);
424 DEFINE_CASE(IMAGE);
425 DEFINE_CASE(INTERNAL);
426 default:
427 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
428 return "";
429 }
430 }
431
432 #undef DEFINE_CASE
433
434 #define DEFINE_CASE(name) case MALI_CHANNEL_## name: return "MALI_CHANNEL_" #name
435 static char *
436 pandecode_channel(enum mali_channel channel)
437 {
438 switch (channel) {
439 DEFINE_CASE(RED);
440 DEFINE_CASE(GREEN);
441 DEFINE_CASE(BLUE);
442 DEFINE_CASE(ALPHA);
443 DEFINE_CASE(ZERO);
444 DEFINE_CASE(ONE);
445 DEFINE_CASE(RESERVED_0);
446 DEFINE_CASE(RESERVED_1);
447
448 default:
449 pandecode_msg("XXX: invalid channel %X\n", channel);
450 return "";
451 }
452 }
453 #undef DEFINE_CASE
454
455 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
456 static char *
457 pandecode_wrap_mode(enum mali_wrap_mode op)
458 {
459 switch (op) {
460 DEFINE_CASE(REPEAT);
461 DEFINE_CASE(CLAMP_TO_EDGE);
462 DEFINE_CASE(CLAMP_TO_BORDER);
463 DEFINE_CASE(MIRRORED_REPEAT);
464
465 default:
466 pandecode_msg("XXX: invalid wrap mode %X\n", op);
467 return "";
468 }
469 }
470 #undef DEFINE_CASE
471
472 #define DEFINE_CASE(name) case MALI_TEX_## name: return "MALI_TEX_" #name
473 static char *
474 pandecode_texture_type(enum mali_texture_type type)
475 {
476 switch (type) {
477 DEFINE_CASE(1D);
478 DEFINE_CASE(2D);
479 DEFINE_CASE(3D);
480 DEFINE_CASE(CUBE);
481
482 default:
483 unreachable("Unknown case");
484 }
485 }
486 #undef DEFINE_CASE
487
488 #define DEFINE_CASE(name) case MALI_MFBD_BLOCK_## name: return "MALI_MFBD_BLOCK_" #name
489 static char *
490 pandecode_mfbd_block_format(enum mali_mfbd_block_format fmt)
491 {
492 switch (fmt) {
493 DEFINE_CASE(TILED);
494 DEFINE_CASE(UNKNOWN);
495 DEFINE_CASE(LINEAR);
496 DEFINE_CASE(AFBC);
497
498 default:
499 unreachable("Invalid case");
500 }
501 }
502 #undef DEFINE_CASE
503
504 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
505 static char *
506 pandecode_exception_access(enum mali_exception_access access)
507 {
508 switch (access) {
509 DEFINE_CASE(NONE);
510 DEFINE_CASE(EXECUTE);
511 DEFINE_CASE(READ);
512 DEFINE_CASE(WRITE);
513
514 default:
515 unreachable("Invalid case");
516 }
517 }
518 #undef DEFINE_CASE
519
520 /* Midgard's tiler descriptor is embedded within the
521 * larger FBD */
522
523 static void
524 pandecode_midgard_tiler_descriptor(
525 const struct midgard_tiler_descriptor *t,
526 unsigned width,
527 unsigned height,
528 bool is_fragment)
529 {
530 pandecode_log(".tiler = {\n");
531 pandecode_indent++;
532
533 if (t->hierarchy_mask == MALI_TILER_DISABLED)
534 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
535 else
536 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
537
538 /* We know this name from the kernel, but we never see it nonzero */
539
540 if (t->flags)
541 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
542
543 MEMORY_PROP(t, polygon_list);
544
545 /* The body is offset from the base of the polygon list */
546 assert(t->polygon_list_body > t->polygon_list);
547 unsigned body_offset = t->polygon_list_body - t->polygon_list;
548
549 /* It needs to fit inside the reported size */
550 assert(t->polygon_list_size >= body_offset);
551
552 /* Check that we fit */
553 struct pandecode_mapped_memory *plist =
554 pandecode_find_mapped_gpu_mem_containing(t->polygon_list);
555
556 assert(t->polygon_list_size <= plist->length);
557
558 /* Now that we've sanity checked, we'll try to calculate the sizes
559 * ourselves for comparison */
560
561 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask);
562 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask);
563
564 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
565 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
566 ref_header, ref_size);
567 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
568 pandecode_msg("body offset %d\n", body_offset);
569 }
570
571 /* The tiler heap has a start and end specified -- it should be
572 * identical to what we have in the BO. The exception is if tiling is
573 * disabled. */
574
575 MEMORY_PROP(t, heap_start);
576 assert(t->heap_end >= t->heap_start);
577
578 struct pandecode_mapped_memory *heap =
579 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
580
581 unsigned heap_size = t->heap_end - t->heap_start;
582
583 /* Tiling is enabled with a special flag */
584 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
585 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
586
587 bool tiling_enabled = hierarchy_mask;
588
589 if (tiling_enabled) {
590 /* When tiling is enabled, the heap should be a tight fit */
591 unsigned heap_offset = t->heap_start - heap->gpu_va;
592 if ((heap_offset + heap_size) != heap->length) {
593 pandecode_msg("XXX: heap size %d (expected %d)\n",
594 heap_size, heap->length - heap_offset);
595 }
596
597 /* We should also have no other flags */
598 if (tiler_flags)
599 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
600 } else {
601 /* When tiling is disabled, we should have that flag and no others */
602
603 if (tiler_flags != MALI_TILER_DISABLED) {
604 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
605 tiler_flags);
606 }
607
608 /* We should also have an empty heap */
609 if (heap_size) {
610 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
611 heap_size);
612 }
613
614 /* Disabled tiling is used only for clear-only jobs, which are
615 * purely FRAGMENT, so we should never see this for
616 * non-FRAGMENT descriptors. */
617
618 if (!is_fragment)
619 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
620 }
621
622 /* We've never seen weights used in practice, but we know from the
623 * kernel these fields is there */
624
625 bool nonzero_weights = false;
626
627 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
628 nonzero_weights |= t->weights[w] != 0x0;
629 }
630
631 if (nonzero_weights) {
632 pandecode_log(".weights = {");
633
634 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
635 pandecode_log("%d, ", t->weights[w]);
636 }
637
638 pandecode_log("},");
639 }
640
641 pandecode_indent--;
642 pandecode_log("}\n");
643 }
644
645 static void
646 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment)
647 {
648 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
649 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
650
651 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
652 pandecode_indent++;
653
654 pandecode_prop("unknown1 = 0x%" PRIx32, s->unknown1);
655 pandecode_prop("unknown2 = 0x%" PRIx32, s->unknown2);
656
657 pandecode_log(".format = ");
658 pandecode_log_decoded_flags(fb_fmt_flag_info, s->format);
659 pandecode_log_cont(",\n");
660
661 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", s->width + 1);
662 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", s->height + 1);
663
664 MEMORY_PROP(s, framebuffer);
665 pandecode_prop("stride = %d", s->stride);
666
667 /* Earlier in the actual commandstream -- right before width -- but we
668 * delay to flow nicer */
669
670 pandecode_log(".clear_flags = ");
671 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
672 pandecode_log_cont(",\n");
673
674 if (s->depth_buffer | s->depth_buffer_enable) {
675 MEMORY_PROP(s, depth_buffer);
676 pandecode_prop("depth_buffer_enable = %s", DS_ENABLE(s->depth_buffer_enable));
677 }
678
679 if (s->stencil_buffer | s->stencil_buffer_enable) {
680 MEMORY_PROP(s, stencil_buffer);
681 pandecode_prop("stencil_buffer_enable = %s", DS_ENABLE(s->stencil_buffer_enable));
682 }
683
684 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
685 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
686 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
687 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
688 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
689 }
690
691 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
692 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
693 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
694 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
695 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
696 }
697
698 if (s->clear_stencil) {
699 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
700 }
701
702 MEMORY_PROP(s, unknown_address_0);
703 const struct midgard_tiler_descriptor t = s->tiler;
704 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment);
705
706 pandecode_indent--;
707 pandecode_log("};\n");
708
709 pandecode_prop("zero0 = 0x%" PRIx64, s->zero0);
710 pandecode_prop("zero1 = 0x%" PRIx64, s->zero1);
711 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
712 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
713
714 printf(".zero3 = {");
715
716 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
717 printf("%X, ", s->zero3[i]);
718
719 printf("},\n");
720
721 printf(".zero6 = {");
722
723 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
724 printf("%X, ", s->zero6[i]);
725
726 printf("},\n");
727 }
728
729 static void
730 pandecode_u32_slide(unsigned name, const u32 *slide, unsigned count)
731 {
732 pandecode_log(".unknown%d = {", name);
733
734 for (int i = 0; i < count; ++i)
735 printf("%X, ", slide[i]);
736
737 pandecode_log("},\n");
738 }
739
740 #define SHORT_SLIDE(num) \
741 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
742
743 static void
744 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
745 {
746 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
747 const struct mali_compute_fbd *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
748
749 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
750 pandecode_indent++;
751
752 SHORT_SLIDE(1);
753
754 pandecode_indent--;
755 printf("},\n");
756 }
757
758 static void
759 pandecode_swizzle(unsigned swizzle)
760 {
761 pandecode_prop("swizzle = %s | (%s << 3) | (%s << 6) | (%s << 9)",
762 pandecode_channel((swizzle >> 0) & 0x7),
763 pandecode_channel((swizzle >> 3) & 0x7),
764 pandecode_channel((swizzle >> 6) & 0x7),
765 pandecode_channel((swizzle >> 9) & 0x7));
766 }
767
768 static void
769 pandecode_rt_format(struct mali_rt_format format)
770 {
771 pandecode_log(".format = {\n");
772 pandecode_indent++;
773
774 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
775 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
776 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
777
778 pandecode_prop("block = %s",
779 pandecode_mfbd_block_format(format.block));
780
781 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
782 MALI_NEGATIVE(format.nr_channels));
783
784 pandecode_log(".flags = ");
785 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
786 pandecode_log_cont(",\n");
787
788 pandecode_swizzle(format.swizzle);
789
790 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
791 * which is a faster hardware-based alternative to the wallpaper method
792 * to preserve framebuffer contents across frames. In practice, MFBD
793 * preload is buggy on Midgard, and so this is a chicken bit. If this
794 * bit isn't set, most likely something broke unrelated to preload */
795
796 if (!format.no_preload) {
797 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
798 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
799 }
800
801 if (format.zero)
802 pandecode_prop("zero = 0x%" PRIx32, format.zero);
803
804 pandecode_indent--;
805 pandecode_log("},\n");
806 }
807
808 static void
809 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct bifrost_framebuffer *fb)
810 {
811 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
812 pandecode_indent++;
813
814 for (int i = 0; i < MALI_NEGATIVE(fb->rt_count_1); i++) {
815 mali_ptr rt_va = gpu_va + i * sizeof(struct bifrost_render_target);
816 struct pandecode_mapped_memory *mem =
817 pandecode_find_mapped_gpu_mem_containing(rt_va);
818 const struct bifrost_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
819
820 pandecode_log("{\n");
821 pandecode_indent++;
822
823 pandecode_rt_format(rt->format);
824
825 if (rt->format.block == MALI_MFBD_BLOCK_AFBC) {
826 pandecode_log(".afbc = {\n");
827 pandecode_indent++;
828
829 char *a = pointer_as_memory_reference(rt->afbc.metadata);
830 pandecode_prop("metadata = %s", a);
831 free(a);
832
833 pandecode_prop("stride = %d", rt->afbc.stride);
834 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
835
836 pandecode_indent--;
837 pandecode_log("},\n");
838 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.unk) {
839 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
840 rt->afbc.metadata,
841 rt->afbc.stride,
842 rt->afbc.unk);
843 }
844
845 MEMORY_PROP(rt, framebuffer);
846 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
847
848 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
849 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
850 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
851 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
852 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
853 }
854
855 if (rt->zero1 || rt->zero2 || rt->zero3) {
856 pandecode_msg("XXX: render target zeros tripped\n");
857 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
858 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
859 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
860 }
861
862 pandecode_indent--;
863 pandecode_log("},\n");
864 }
865
866 pandecode_indent--;
867 pandecode_log("};\n");
868 }
869
870 static unsigned
871 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment)
872 {
873 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
874 const struct bifrost_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
875
876 if (fb->sample_locations) {
877 /* The blob stores all possible sample locations in a single buffer
878 * allocated on startup, and just switches the pointer when switching
879 * MSAA state. For now, we just put the data into the cmdstream, but we
880 * should do something like what the blob does with a real driver.
881 *
882 * There seem to be 32 slots for sample locations, followed by another
883 * 16. The second 16 is just the center location followed by 15 zeros
884 * in all the cases I've identified (maybe shader vs. depth/color
885 * samples?).
886 */
887
888 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->sample_locations);
889
890 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->sample_locations);
891
892 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
893 pandecode_indent++;
894
895 for (int i = 0; i < 32 + 16; i++) {
896 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
897 }
898
899 pandecode_indent--;
900 pandecode_log("};\n");
901 }
902
903 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
904 pandecode_indent++;
905
906 pandecode_prop("unk0 = 0x%x", fb->unk0);
907
908 if (fb->sample_locations)
909 pandecode_prop("sample_locations = sample_locations_%d", job_no);
910
911 /* Assume that unknown1 was emitted in the last job for
912 * now */
913 MEMORY_PROP(fb, unknown1);
914
915 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
916 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
917 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
918 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
919
920 pandecode_prop("unk1 = 0x%x", fb->unk1);
921 pandecode_prop("unk2 = 0x%x", fb->unk2);
922 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
923 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
924
925 pandecode_log(".mfbd_flags = ");
926 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
927 pandecode_log_cont(",\n");
928
929 if (fb->clear_stencil)
930 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
931
932 if (fb->clear_depth)
933 pandecode_prop("clear_depth = %f", fb->clear_depth);
934
935 /* TODO: What is this? Let's not blow up.. */
936 if (fb->unknown2 != 0x1F)
937 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
938
939 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
940 MEMORY_PROP(fb, scratchpad);
941 const struct midgard_tiler_descriptor t = fb->tiler;
942 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment);
943
944 if (fb->zero3 || fb->zero4) {
945 pandecode_msg("XXX: framebuffer zeros tripped\n");
946 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
947 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
948 }
949
950 pandecode_indent--;
951 pandecode_log("};\n");
952
953 gpu_va += sizeof(struct bifrost_framebuffer);
954
955 if ((fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment) {
956 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
957 const struct bifrost_fb_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
958
959 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
960 pandecode_indent++;
961
962 MEMORY_PROP(fbx, checksum);
963
964 if (fbx->checksum_stride)
965 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
966
967 pandecode_log(".flags = ");
968 pandecode_log_decoded_flags(mfbd_extra_flag_info, fbx->flags);
969 pandecode_log_cont(",\n");
970
971 if (fbx->flags & MALI_EXTRA_AFBC_ZS) {
972 pandecode_log(".ds_afbc = {\n");
973 pandecode_indent++;
974
975 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
976 pandecode_prop("depth_stencil_afbc_stride = %d",
977 fbx->ds_afbc.depth_stencil_afbc_stride);
978 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
979
980 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
981 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
982 pandecode_prop("zero1 = 0x%" PRIx32,
983 fbx->ds_afbc.zero1);
984 pandecode_prop("padding = 0x%" PRIx64,
985 fbx->ds_afbc.padding);
986 }
987
988 pandecode_indent--;
989 pandecode_log("},\n");
990 } else {
991 pandecode_log(".ds_linear = {\n");
992 pandecode_indent++;
993
994 if (fbx->ds_linear.depth) {
995 MEMORY_PROP_DIR(fbx->ds_linear, depth);
996 pandecode_prop("depth_stride = %d",
997 fbx->ds_linear.depth_stride);
998 }
999
1000 if (fbx->ds_linear.stencil) {
1001 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1002 pandecode_prop("stencil_stride = %d",
1003 fbx->ds_linear.stencil_stride);
1004 }
1005
1006 if (fbx->ds_linear.depth_stride_zero ||
1007 fbx->ds_linear.stencil_stride_zero ||
1008 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
1009 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1010 pandecode_prop("depth_stride_zero = 0x%x",
1011 fbx->ds_linear.depth_stride_zero);
1012 pandecode_prop("stencil_stride_zero = 0x%x",
1013 fbx->ds_linear.stencil_stride_zero);
1014 pandecode_prop("zero1 = 0x%" PRIx32,
1015 fbx->ds_linear.zero1);
1016 pandecode_prop("zero2 = 0x%" PRIx32,
1017 fbx->ds_linear.zero2);
1018 }
1019
1020 pandecode_indent--;
1021 pandecode_log("},\n");
1022 }
1023
1024 if (fbx->zero3 || fbx->zero4) {
1025 pandecode_msg("XXX: fb_extra zeros tripped\n");
1026 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1027 pandecode_prop("zero4 = 0x%" PRIx64, fbx->zero4);
1028 }
1029
1030 pandecode_indent--;
1031 pandecode_log("};\n");
1032
1033 gpu_va += sizeof(struct bifrost_fb_extra);
1034 }
1035
1036 if (is_fragment)
1037 pandecode_render_target(gpu_va, job_no, fb);
1038
1039 /* Passback the render target count */
1040 return MALI_NEGATIVE(fb->rt_count_1);
1041 }
1042
1043 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1044 * count */
1045
1046 static void
1047 pandecode_padded_vertices(unsigned shift, unsigned k)
1048 {
1049 unsigned odd = 2*k + 1;
1050 unsigned pot = 1 << shift;
1051 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1052 }
1053
1054 /* Given a magic divisor, recover what we were trying to divide by.
1055 *
1056 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1057 * 0 <= m < 2^N, for N bits in m.
1058 *
1059 * Let q represent the number we would like to divide by.
1060 *
1061 * By definition of a magic divisor for N-bit unsigned integers (a number you
1062 * multiply by to magically get division), m is a number such that:
1063 *
1064 * (m * x) & (2^N - 1) = floor(x/q).
1065 * for all x on Z where 0 <= x < 2^N
1066 *
1067 * Ignore the case where any of the above values equals zero; it is irrelevant
1068 * for our purposes (instanced arrays).
1069 *
1070 * Choose x = q. Then:
1071 *
1072 * (m * x) & (2^N - 1) = floor(x/q).
1073 * (m * q) & (2^N - 1) = floor(q/q).
1074 *
1075 * floor(q/q) = floor(1) = 1, therefore:
1076 *
1077 * (m * q) & (2^N - 1) = 1
1078 *
1079 * Recall the identity that the bitwise AND of one less than a power-of-two
1080 * equals the modulo with that power of two, i.e. for all x:
1081 *
1082 * x & (2^N - 1) = x % N
1083 *
1084 * Therefore:
1085 *
1086 * mq % (2^N) = 1
1087 *
1088 * By definition, a modular multiplicative inverse of a number m is the number
1089 * q such that with respect to a modulos M:
1090 *
1091 * mq % M = 1
1092 *
1093 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1094 *
1095 */
1096
1097 static void
1098 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1099 {
1100 #if 0
1101 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1102 * shift) the most lame way possible... just repeatedly add.
1103 * Asymptoptically slow but nobody cares in practice, unless you have
1104 * massive numbers of vertices or high divisors. */
1105
1106 unsigned inverse = 0;
1107
1108 /* Magic implicitly has the highest bit set */
1109 magic |= (1 << 31);
1110
1111 /* Depending on rounding direction */
1112 if (extra)
1113 magic++;
1114
1115 for (;;) {
1116 uint32_t product = magic * inverse;
1117
1118 if (shift) {
1119 product >>= shift;
1120 }
1121
1122 if (product == 1)
1123 break;
1124
1125 ++inverse;
1126 }
1127
1128 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1129
1130 /* Recall we're supposed to divide by (gl_level_divisor *
1131 * padded_num_vertices) */
1132
1133 unsigned padded_num_vertices = inverse / orig_divisor;
1134
1135 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1136 #endif
1137 }
1138
1139 static void
1140 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1141 mali_ptr addr, int job_no, char *suffix,
1142 int count, bool varying)
1143 {
1144 char *prefix = varying ? "varyings" : "attributes";
1145
1146 if (!addr) {
1147 pandecode_msg("no %s\n", prefix);
1148 return;
1149 }
1150
1151 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1152
1153 char base[128];
1154 snprintf(base, sizeof(base), "%s_data_%d%s", prefix, job_no, suffix);
1155
1156 for (int i = 0; i < count; ++i) {
1157 enum mali_attr_mode mode = attr[i].elements & 7;
1158
1159 if (mode == MALI_ATTR_UNUSED)
1160 continue;
1161
1162 mali_ptr raw_elements = attr[i].elements & ~7;
1163
1164 /* TODO: Do we maybe want to dump the attribute values
1165 * themselves given the specified format? Or is that too hard?
1166 * */
1167
1168 char *a = pointer_as_memory_reference(raw_elements);
1169 pandecode_log("mali_ptr %s_%d_p = %s;\n", base, i, a);
1170 free(a);
1171 }
1172
1173 pandecode_log("union mali_attr %s_%d[] = {\n", prefix, job_no);
1174 pandecode_indent++;
1175
1176 for (int i = 0; i < count; ++i) {
1177 pandecode_log("{\n");
1178 pandecode_indent++;
1179
1180 unsigned mode = attr[i].elements & 7;
1181 pandecode_prop("elements = (%s_%d_p) | %s", base, i, pandecode_attr_mode(mode));
1182 pandecode_prop("shift = %d", attr[i].shift);
1183 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1184 pandecode_prop("stride = 0x%" PRIx32, attr[i].stride);
1185 pandecode_prop("size = 0x%" PRIx32, attr[i].size);
1186
1187 /* Decode further where possible */
1188
1189 if (mode == MALI_ATTR_MODULO) {
1190 pandecode_padded_vertices(
1191 attr[i].shift,
1192 attr[i].extra_flags);
1193 }
1194
1195 pandecode_indent--;
1196 pandecode_log("}, \n");
1197
1198 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1199 i++;
1200 pandecode_log("{\n");
1201 pandecode_indent++;
1202 pandecode_prop("unk = 0x%x", attr[i].unk);
1203 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1204 if (attr[i].zero != 0)
1205 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1206 pandecode_prop("divisor = %d", attr[i].divisor);
1207 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1208 pandecode_indent--;
1209 pandecode_log("}, \n");
1210 }
1211
1212 }
1213
1214 pandecode_indent--;
1215 pandecode_log("};\n");
1216 }
1217
1218 static mali_ptr
1219 pandecode_shader_address(const char *name, mali_ptr ptr)
1220 {
1221 /* TODO: Decode flags */
1222 mali_ptr shader_ptr = ptr & ~15;
1223
1224 char *a = pointer_as_memory_reference(shader_ptr);
1225 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1226 free(a);
1227
1228 return shader_ptr;
1229 }
1230
1231 static bool
1232 all_zero(unsigned *buffer, unsigned count)
1233 {
1234 for (unsigned i = 0; i < count; ++i) {
1235 if (buffer[i])
1236 return false;
1237 }
1238
1239 return true;
1240 }
1241
1242 static void
1243 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1244 {
1245 if (all_zero((unsigned *) stencil, sizeof(stencil) / sizeof(unsigned)))
1246 return;
1247
1248 const char *func = pandecode_func(stencil->func);
1249 const char *sfail = pandecode_stencil_op(stencil->sfail);
1250 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1251 const char *dppass = pandecode_stencil_op(stencil->dppass);
1252
1253 if (stencil->zero)
1254 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero);
1255
1256 pandecode_log(".stencil_%s = {\n", name);
1257 pandecode_indent++;
1258 pandecode_prop("ref = %d", stencil->ref);
1259 pandecode_prop("mask = 0x%02X", stencil->mask);
1260 pandecode_prop("func = %s", func);
1261 pandecode_prop("sfail = %s", sfail);
1262 pandecode_prop("dpfail = %s", dpfail);
1263 pandecode_prop("dppass = %s", dppass);
1264 pandecode_indent--;
1265 pandecode_log("},\n");
1266 }
1267
1268 static void
1269 pandecode_blend_equation(const struct mali_blend_equation *blend)
1270 {
1271 if (blend->zero1)
1272 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1273
1274 pandecode_log(".equation = {\n");
1275 pandecode_indent++;
1276
1277 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1278 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1279
1280 pandecode_log(".color_mask = ");
1281 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1282 pandecode_log_cont(",\n");
1283
1284 pandecode_indent--;
1285 pandecode_log("},\n");
1286 }
1287
1288 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1289
1290 static unsigned
1291 decode_bifrost_constant(u16 constant)
1292 {
1293 float lo = (float) (constant & 0xFF);
1294 float hi = (float) (constant >> 8);
1295
1296 return (hi / 255.0) + (lo / 65535.0);
1297 }
1298
1299 static mali_ptr
1300 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1301 {
1302 struct bifrost_blend_rt *b =
1303 ((struct bifrost_blend_rt *) descs) + rt_no;
1304
1305 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1306 pandecode_indent++;
1307
1308 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1309 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1310 b->constant, decode_bifrost_constant(b->constant));
1311
1312 /* TODO figure out blend shader enable bit */
1313 pandecode_blend_equation(&b->equation);
1314 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1315 pandecode_prop("index = 0x%" PRIx16, b->index);
1316 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1317
1318 pandecode_indent--;
1319 pandecode_log("},\n");
1320
1321 return 0;
1322 }
1323
1324 static mali_ptr
1325 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1326 {
1327 if (all_zero((unsigned *) blend, sizeof(blend) / sizeof(unsigned)))
1328 return 0;
1329
1330 pandecode_log(".blend = {\n");
1331 pandecode_indent++;
1332
1333 if (is_shader) {
1334 pandecode_shader_address("shader", blend->shader);
1335 } else {
1336 pandecode_blend_equation(&blend->equation);
1337 pandecode_prop("constant = %f", blend->constant);
1338 }
1339
1340 pandecode_indent--;
1341 pandecode_log("},\n");
1342
1343 /* Return blend shader to disassemble if present */
1344 return is_shader ? (blend->shader & ~0xF) : 0;
1345 }
1346
1347 static mali_ptr
1348 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1349 {
1350 struct midgard_blend_rt *b =
1351 ((struct midgard_blend_rt *) descs) + rt_no;
1352
1353 /* Flags determine presence of blend shader */
1354 bool is_shader = (b->flags & 0xF) >= 0x2;
1355
1356 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1357 pandecode_indent++;
1358
1359 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1360
1361 union midgard_blend blend = b->blend;
1362 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1363
1364 pandecode_indent--;
1365 pandecode_log("};\n");
1366
1367 return shader;
1368 }
1369
1370 static int
1371 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1372 {
1373 char base[128];
1374 char *prefix = varying ? "varying" : "attribute";
1375 unsigned max_index = 0;
1376 snprintf(base, sizeof(base), "%s_meta", prefix);
1377
1378 pandecode_log("struct mali_attr_meta %s_%d%s[] = {\n", base, job_no, suffix);
1379 pandecode_indent++;
1380
1381 struct mali_attr_meta *attr_meta;
1382 mali_ptr p = varying ? (v->varying_meta & ~0xF) : v->attribute_meta;
1383
1384 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1385
1386 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1387 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1388 sizeof(*attr_mem));
1389
1390 /* If the record is discard, it should be zero for everything else */
1391
1392 if (attr_meta->format == MALI_VARYING_DISCARD) {
1393 uint64_t zero =
1394 attr_meta->index |
1395 attr_meta->unknown1 |
1396 attr_meta->unknown3 |
1397 attr_meta->src_offset;
1398
1399 if (zero)
1400 pandecode_msg("XXX: expected empty record for varying discard\n");
1401
1402 /* We want to look for a literal 0000 swizzle -- this
1403 * is not encoded with all zeroes, however */
1404
1405 enum mali_channel z = MALI_CHANNEL_ZERO;
1406 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1407 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1408
1409 if (!good_swizzle)
1410 pandecode_msg("XXX: expected zero swizzle for discard\n");
1411
1412 if (!varying)
1413 pandecode_msg("XXX: cannot discard attribute\n");
1414
1415 /* If we're all good, omit the record */
1416 if (!zero && varying && good_swizzle) {
1417 pandecode_log("/* discarded varying */\n");
1418 continue;
1419 }
1420 }
1421
1422 pandecode_log("{\n");
1423 pandecode_indent++;
1424 pandecode_prop("index = %d", attr_meta->index);
1425
1426 if (attr_meta->index > max_index)
1427 max_index = attr_meta->index;
1428 pandecode_swizzle(attr_meta->swizzle);
1429 pandecode_prop("format = %s", pandecode_format(attr_meta->format));
1430
1431 if (attr_meta->unknown1 != 0x2) {
1432 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1433 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1434 }
1435
1436 if (attr_meta->unknown3) {
1437 pandecode_msg("XXX: unexpected unknown3 set\n");
1438 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1439 }
1440
1441 pandecode_prop("src_offset = %d", attr_meta->src_offset);
1442 pandecode_indent--;
1443 pandecode_log("},\n");
1444
1445 }
1446
1447 pandecode_indent--;
1448 pandecode_log("};\n");
1449
1450 return count ? (max_index + 1) : 0;
1451 }
1452
1453 static void
1454 pandecode_indices(uintptr_t pindices, uint32_t index_count, int job_no)
1455 {
1456 struct pandecode_mapped_memory *imem = pandecode_find_mapped_gpu_mem_containing(pindices);
1457
1458 if (imem) {
1459 /* Indices are literally just a u32 array :) */
1460
1461 uint32_t *PANDECODE_PTR_VAR(indices, imem, pindices);
1462
1463 pandecode_log("uint32_t indices_%d[] = {\n", job_no);
1464 pandecode_indent++;
1465
1466 for (unsigned i = 0; i < (index_count + 1); i += 3)
1467 pandecode_log("%d, %d, %d,\n",
1468 indices[i],
1469 indices[i + 1],
1470 indices[i + 2]);
1471
1472 pandecode_indent--;
1473 pandecode_log("};\n");
1474 }
1475 }
1476
1477 /* return bits [lo, hi) of word */
1478 static u32
1479 bits(u32 word, u32 lo, u32 hi)
1480 {
1481 if (hi - lo >= 32)
1482 return word; // avoid undefined behavior with the shift
1483
1484 return (word >> lo) & ((1 << (hi - lo)) - 1);
1485 }
1486
1487 static void
1488 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool noninstanced)
1489 {
1490 pandecode_log_cont("{\n");
1491 pandecode_indent++;
1492
1493 /* Decode invocation_count. See the comment before the definition of
1494 * invocation_count for an explanation.
1495 */
1496
1497 unsigned size_x = bits(p->invocation_count, 0, p->size_y_shift) + 1;
1498 unsigned size_y = bits(p->invocation_count, p->size_y_shift, p->size_z_shift) + 1;
1499 unsigned size_z = bits(p->invocation_count, p->size_z_shift, p->workgroups_x_shift) + 1;
1500
1501 unsigned groups_x = bits(p->invocation_count, p->workgroups_x_shift, p->workgroups_y_shift) + 1;
1502 unsigned groups_y = bits(p->invocation_count, p->workgroups_y_shift, p->workgroups_z_shift) + 1;
1503 unsigned groups_z = bits(p->invocation_count, p->workgroups_z_shift, 32) + 1;
1504
1505 /* Even though we have this decoded, we want to ensure that the
1506 * representation is "unique" so we don't lose anything by printing only
1507 * the final result. More specifically, we need to check that we were
1508 * passed something in canonical form, since the definition per the
1509 * hardware is inherently not unique. How? Well, take the resulting
1510 * decode and pack it ourselves! If it is bit exact with what we
1511 * decoded, we're good to go. */
1512
1513 struct mali_vertex_tiler_prefix ref;
1514 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, noninstanced);
1515
1516 bool canonical =
1517 (p->invocation_count == ref.invocation_count) &&
1518 (p->size_y_shift == ref.size_y_shift) &&
1519 (p->size_z_shift == ref.size_z_shift) &&
1520 (p->workgroups_x_shift == ref.workgroups_x_shift) &&
1521 (p->workgroups_y_shift == ref.workgroups_y_shift) &&
1522 (p->workgroups_z_shift == ref.workgroups_z_shift) &&
1523 (p->workgroups_x_shift_2 == ref.workgroups_x_shift_2);
1524
1525 if (!canonical) {
1526 pandecode_msg("XXX: non-canonical workgroups packing\n");
1527 pandecode_msg("expected: %X, %d, %d, %d, %d, %d\n",
1528 ref.invocation_count,
1529 ref.size_y_shift,
1530 ref.size_z_shift,
1531 ref.workgroups_x_shift,
1532 ref.workgroups_y_shift,
1533 ref.workgroups_z_shift,
1534 ref.workgroups_x_shift_2);
1535
1536 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1537 pandecode_prop("size_y_shift = %d", p->size_y_shift);
1538 pandecode_prop("size_z_shift = %d", p->size_z_shift);
1539 pandecode_prop("workgroups_x_shift = %d", p->workgroups_x_shift);
1540 pandecode_prop("workgroups_y_shift = %d", p->workgroups_y_shift);
1541 pandecode_prop("workgroups_z_shift = %d", p->workgroups_z_shift);
1542 pandecode_prop("workgroups_x_shift_2 = %d", p->workgroups_x_shift_2);
1543 }
1544
1545 /* Regardless, print the decode */
1546 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1547 size_x, size_y, size_z,
1548 groups_x, groups_y, groups_z);
1549
1550 /* TODO: Decode */
1551 if (p->unknown_draw)
1552 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1553
1554 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1555
1556 if (p->draw_mode != MALI_DRAW_NONE)
1557 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
1558
1559 /* Index count only exists for tiler jobs anyway */
1560
1561 if (p->index_count)
1562 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1563
1564 if (p->offset_bias_correction)
1565 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1566
1567 /* TODO: Figure out what this is. It's not zero */
1568 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1569
1570 pandecode_indent--;
1571 pandecode_log("},\n");
1572 }
1573
1574 static void
1575 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1576 {
1577 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1578 struct mali_uniform_buffer_meta *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1579
1580 pandecode_log("struct mali_uniform_buffer_meta uniform_buffers_%"PRIx64"_%d[] = {\n",
1581 pubufs, job_no);
1582 pandecode_indent++;
1583
1584 for (int i = 0; i < ubufs_count; i++) {
1585 pandecode_log("{\n");
1586 pandecode_indent++;
1587
1588 unsigned size = (ubufs[i].size + 1) * 16;
1589 mali_ptr addr = ubufs[i].ptr << 2;
1590
1591 pandecode_validate_buffer(addr, size);
1592
1593 char *ptr = pointer_as_memory_reference(ubufs[i].ptr << 2);
1594 pandecode_prop("size = %u", size);
1595 pandecode_prop("ptr = (%s) >> 2", ptr);
1596 pandecode_indent--;
1597 pandecode_log("},\n");
1598 free(ptr);
1599 }
1600
1601 pandecode_indent--;
1602 pandecode_log("};\n");
1603 }
1604
1605 static void
1606 pandecode_scratchpad(uintptr_t pscratchpad, int job_no, char *suffix)
1607 {
1608
1609 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(pscratchpad);
1610
1611 struct bifrost_scratchpad *PANDECODE_PTR_VAR(scratchpad, mem, pscratchpad);
1612
1613 if (scratchpad->zero) {
1614 pandecode_msg("XXX: scratchpad zero tripped");
1615 pandecode_prop("zero = 0x%x\n", scratchpad->zero);
1616 }
1617
1618 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64"_%d%s = {\n", pscratchpad, job_no, suffix);
1619 pandecode_indent++;
1620
1621 pandecode_prop("flags = 0x%x", scratchpad->flags);
1622 MEMORY_PROP(scratchpad, gpu_scratchpad);
1623
1624 pandecode_indent--;
1625 pandecode_log("};\n");
1626 }
1627
1628 static unsigned shader_id = 0;
1629
1630 static void
1631 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1632 bool is_bifrost, unsigned nr_regs)
1633 {
1634 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1635 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1636
1637 /* Compute maximum possible size */
1638 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1639
1640 /* Print some boilerplate to clearly denote the assembly (which doesn't
1641 * obey indentation rules), and actually do the disassembly! */
1642
1643 printf("\n\n");
1644
1645 char prefix[512];
1646
1647 snprintf(prefix, sizeof(prefix) - 1, "shader%d - %s shader: ",
1648 shader_id++,
1649 (type == JOB_TYPE_TILER) ? "FRAGMENT" : "VERTEX");
1650
1651 if (is_bifrost) {
1652 disassemble_bifrost(code, sz, false);
1653 } else {
1654 disassemble_midgard(code, sz, true, nr_regs, prefix);
1655 }
1656
1657 printf("\n\n");
1658 }
1659
1660 static void
1661 pandecode_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix *p,
1662 int job_no, enum mali_job_type job_type,
1663 char *suffix, bool is_bifrost)
1664 {
1665 mali_ptr shader_meta_ptr = (u64) (uintptr_t) (p->_shader_upper << 4);
1666 struct pandecode_mapped_memory *attr_mem;
1667
1668 unsigned rt_count = 1;
1669
1670 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1671 * are the only things actually needed from the FBD, vertex/tiler jobs
1672 * no longer reference the FBD -- instead, this field points to some
1673 * info about the scratchpad.
1674 */
1675 if (is_bifrost)
1676 pandecode_scratchpad(p->framebuffer & ~FBD_TYPE, job_no, suffix);
1677 else if (p->framebuffer & MALI_MFBD)
1678 rt_count = pandecode_mfbd_bfr((u64) ((uintptr_t) p->framebuffer) & FBD_MASK, job_no, false);
1679 else if (job_type == JOB_TYPE_COMPUTE)
1680 pandecode_compute_fbd((u64) (uintptr_t) p->framebuffer, job_no);
1681 else
1682 pandecode_sfbd((u64) (uintptr_t) p->framebuffer, job_no, false);
1683
1684 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
1685 int texture_count = 0, sampler_count = 0;
1686
1687 if (shader_meta_ptr) {
1688 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(shader_meta_ptr);
1689 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, shader_meta_ptr);
1690
1691 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", shader_meta_ptr, job_no, suffix);
1692 pandecode_indent++;
1693
1694 /* Save for dumps */
1695 attribute_count = s->attribute_count;
1696 varying_count = s->varying_count;
1697 texture_count = s->texture_count;
1698 sampler_count = s->sampler_count;
1699
1700 if (is_bifrost) {
1701 uniform_count = s->bifrost2.uniform_count;
1702 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
1703 } else {
1704 uniform_count = s->midgard1.uniform_buffer_count;
1705 uniform_buffer_count = s->midgard1.uniform_buffer_count;
1706 }
1707
1708 mali_ptr shader_ptr = pandecode_shader_address("shader", s->shader);
1709
1710 pandecode_prop("texture_count = %" PRId16, s->texture_count);
1711 pandecode_prop("sampler_count = %" PRId16, s->sampler_count);
1712 pandecode_prop("attribute_count = %" PRId16, s->attribute_count);
1713 pandecode_prop("varying_count = %" PRId16, s->varying_count);
1714
1715 unsigned nr_registers = 0;
1716
1717 if (is_bifrost) {
1718 pandecode_log(".bifrost1 = {\n");
1719 pandecode_indent++;
1720
1721 pandecode_prop("uniform_buffer_count = %" PRId32, s->bifrost1.uniform_buffer_count);
1722 pandecode_prop("unk1 = 0x%" PRIx32, s->bifrost1.unk1);
1723
1724 pandecode_indent--;
1725 pandecode_log("},\n");
1726 } else {
1727 pandecode_log(".midgard1 = {\n");
1728 pandecode_indent++;
1729
1730 pandecode_prop("uniform_count = %" PRId16, s->midgard1.uniform_count);
1731 pandecode_prop("uniform_buffer_count = %" PRId16, s->midgard1.uniform_buffer_count);
1732 pandecode_prop("work_count = %" PRId16, s->midgard1.work_count);
1733 nr_registers = s->midgard1.work_count;
1734
1735 pandecode_log(".flags = ");
1736 pandecode_log_decoded_flags(shader_midgard1_flag_info, s->midgard1.flags);
1737 pandecode_log_cont(",\n");
1738
1739 pandecode_prop("unknown2 = 0x%" PRIx32, s->midgard1.unknown2);
1740
1741 pandecode_indent--;
1742 pandecode_log("},\n");
1743 }
1744
1745 if (s->depth_units || s->depth_factor) {
1746 pandecode_prop("depth_factor = %f", s->depth_factor);
1747 pandecode_prop("depth_units = %f", s->depth_units);
1748 }
1749
1750 if (s->alpha_coverage) {
1751 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
1752 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
1753
1754 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
1755 invert_alpha_coverage ? "~" : "",
1756 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
1757 }
1758
1759 if (s->unknown2_3 || s->unknown2_4) {
1760 pandecode_log(".unknown2_3 = ");
1761
1762 int unknown2_3 = s->unknown2_3;
1763 int unknown2_4 = s->unknown2_4;
1764
1765 /* We're not quite sure what these flags mean without the depth test, if anything */
1766
1767 if (unknown2_3 & (MALI_DEPTH_TEST | MALI_DEPTH_FUNC_MASK)) {
1768 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
1769 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
1770
1771 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
1772 }
1773
1774 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
1775 pandecode_log_cont(",\n");
1776
1777 pandecode_log(".unknown2_4 = ");
1778 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
1779 pandecode_log_cont(",\n");
1780 }
1781
1782 if (s->stencil_mask_front || s->stencil_mask_back) {
1783 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
1784 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
1785 }
1786
1787 pandecode_stencil("front", &s->stencil_front);
1788 pandecode_stencil("back", &s->stencil_back);
1789
1790 if (is_bifrost) {
1791 pandecode_log(".bifrost2 = {\n");
1792 pandecode_indent++;
1793
1794 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
1795 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
1796 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
1797 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
1798
1799 pandecode_indent--;
1800 pandecode_log("},\n");
1801 } else if (s->midgard2.unknown2_7) {
1802 pandecode_log(".midgard2 = {\n");
1803 pandecode_indent++;
1804
1805 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
1806 pandecode_indent--;
1807 pandecode_log("},\n");
1808 }
1809
1810 if (s->unknown2_8)
1811 pandecode_prop("unknown2_8 = 0x%" PRIx32, s->unknown2_8);
1812
1813 if (!is_bifrost) {
1814 /* TODO: Blend shaders routing/disasm */
1815
1816 union midgard_blend blend = s->blend;
1817 pandecode_midgard_blend(&blend, false);
1818 }
1819
1820 pandecode_indent--;
1821 pandecode_log("};\n");
1822
1823 /* MRT blend fields are used whenever MFBD is used, with
1824 * per-RT descriptors */
1825
1826 if (job_type == JOB_TYPE_TILER) {
1827 void* blend_base = (void *) (s + 1);
1828
1829 for (unsigned i = 0; i < rt_count; i++) {
1830 mali_ptr shader = 0;
1831
1832 if (is_bifrost)
1833 shader = pandecode_bifrost_blend(blend_base, job_no, i);
1834 else
1835 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
1836
1837 if (shader & ~0xF)
1838 pandecode_shader_disassemble(shader, job_no, job_type, false, 0);
1839 }
1840 }
1841
1842 if (shader_ptr & ~0xF)
1843 pandecode_shader_disassemble(shader_ptr, job_no, job_type, is_bifrost, nr_registers);
1844 } else
1845 pandecode_msg("<no shader>\n");
1846
1847 if (p->viewport) {
1848 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
1849 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
1850
1851 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
1852 pandecode_indent++;
1853
1854 pandecode_prop("clip_minx = %f", f->clip_minx);
1855 pandecode_prop("clip_miny = %f", f->clip_miny);
1856 pandecode_prop("clip_minz = %f", f->clip_minz);
1857 pandecode_prop("clip_maxx = %f", f->clip_maxx);
1858 pandecode_prop("clip_maxy = %f", f->clip_maxy);
1859 pandecode_prop("clip_maxz = %f", f->clip_maxz);
1860
1861 /* Only the higher coordinates are MALI_POSITIVE scaled */
1862
1863 pandecode_prop("viewport0 = { %d, %d }",
1864 f->viewport0[0], f->viewport0[1]);
1865
1866 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
1867 f->viewport1[0] + 1, f->viewport1[1] + 1);
1868
1869 pandecode_indent--;
1870 pandecode_log("};\n");
1871 }
1872
1873 if (p->attribute_meta) {
1874 unsigned max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
1875
1876 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
1877 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false);
1878 }
1879
1880 /* Varyings are encoded like attributes but not actually sent; we just
1881 * pass a zero buffer with the right stride/size set, (or whatever)
1882 * since the GPU will write to it itself */
1883
1884 if (p->varying_meta) {
1885 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
1886 }
1887
1888 if (p->varyings) {
1889 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
1890
1891 /* Number of descriptors depends on whether there are
1892 * non-internal varyings */
1893
1894 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true);
1895 }
1896
1897 if (p->uniform_buffers) {
1898 if (uniform_buffer_count)
1899 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
1900 else
1901 pandecode_msg("XXX: UBOs specified but not referenced\n");
1902 } else if (uniform_buffer_count)
1903 pandecode_msg("XXX: UBOs referenced but not specified\n");
1904
1905 /* We don't want to actually dump uniforms, but we do need to validate
1906 * that the counts we were given are sane */
1907
1908 if (p->uniforms) {
1909 if (uniform_count)
1910 pandecode_validate_buffer(p->uniforms, uniform_count * 16);
1911 else
1912 pandecode_msg("XXX: Uniforms specified but not referenced");
1913 } else if (uniform_count)
1914 pandecode_msg("XXX: UBOs referenced but not specified\n");
1915
1916 if (p->texture_trampoline) {
1917 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(p->texture_trampoline);
1918
1919 if (mmem) {
1920 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline);
1921
1922 pandecode_log("uint64_t texture_trampoline_%"PRIx64"_%d[] = {\n", p->texture_trampoline, job_no);
1923 pandecode_indent++;
1924
1925 for (int tex = 0; tex < texture_count; ++tex) {
1926 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
1927 char *a = pointer_as_memory_reference(*u);
1928 pandecode_log("%s,\n", a);
1929 free(a);
1930 }
1931
1932 pandecode_indent--;
1933 pandecode_log("};\n");
1934
1935 /* Now, finally, descend down into the texture descriptor */
1936 for (int tex = 0; tex < texture_count; ++tex) {
1937 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
1938 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
1939
1940 if (tmem) {
1941 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, *u);
1942
1943 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", *u, job_no, tex);
1944 pandecode_indent++;
1945
1946 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", t->width + 1);
1947 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", t->height + 1);
1948 pandecode_prop("depth = MALI_POSITIVE(%" PRId16 ")", t->depth + 1);
1949 pandecode_prop("array_size = MALI_POSITIVE(%" PRId16 ")", t->array_size + 1);
1950 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
1951 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
1952 pandecode_prop("nr_mipmap_levels = %" PRId8, t->nr_mipmap_levels);
1953
1954 struct mali_texture_format f = t->format;
1955
1956 pandecode_log(".format = {\n");
1957 pandecode_indent++;
1958
1959 pandecode_swizzle(f.swizzle);
1960 pandecode_prop("format = %s", pandecode_format(f.format));
1961 pandecode_prop("type = %s", pandecode_texture_type(f.type));
1962 pandecode_prop("srgb = %" PRId32, f.srgb);
1963 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
1964 pandecode_prop("usage2 = 0x%" PRIx32, f.usage2);
1965
1966 pandecode_indent--;
1967 pandecode_log("},\n");
1968
1969 pandecode_swizzle(t->swizzle);
1970
1971 if (t->swizzle_zero) {
1972 /* Shouldn't happen */
1973 pandecode_msg("XXX: swizzle zero tripped\n");
1974 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
1975 }
1976
1977 pandecode_prop("unknown3 = 0x%" PRIx32, t->unknown3);
1978
1979 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
1980 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
1981 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
1982
1983 pandecode_log(".payload = {\n");
1984 pandecode_indent++;
1985
1986 /* A bunch of bitmap pointers follow.
1987 * We work out the correct number,
1988 * based on the mipmap/cubemap
1989 * properties, but dump extra
1990 * possibilities to futureproof */
1991
1992 int bitmap_count = MALI_NEGATIVE(t->nr_mipmap_levels);
1993 bool manual_stride = f.usage2 & MALI_TEX_MANUAL_STRIDE;
1994
1995 /* Miptree for each face */
1996 if (f.type == MALI_TEX_CUBE)
1997 bitmap_count *= 6;
1998
1999 /* Array of textures */
2000 bitmap_count *= MALI_NEGATIVE(t->array_size);
2001
2002 /* Stride for each element */
2003 if (manual_stride)
2004 bitmap_count *= 2;
2005
2006 /* Sanity check the size */
2007 int max_count = sizeof(t->payload) / sizeof(t->payload[0]);
2008 assert (bitmap_count <= max_count);
2009
2010 for (int i = 0; i < bitmap_count; ++i) {
2011 /* How we dump depends if this is a stride or a pointer */
2012
2013 if ((f.usage2 & MALI_TEX_MANUAL_STRIDE) && (i & 1)) {
2014 /* signed 32-bit snuck in as a 64-bit pointer */
2015 uint64_t stride_set = t->payload[i];
2016 uint32_t clamped_stride = stride_set;
2017 int32_t stride = clamped_stride;
2018 assert(stride_set == clamped_stride);
2019 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
2020 } else {
2021 char *a = pointer_as_memory_reference(t->payload[i]);
2022 pandecode_log("%s, \n", a);
2023 free(a);
2024 }
2025 }
2026
2027 pandecode_indent--;
2028 pandecode_log("},\n");
2029
2030 pandecode_indent--;
2031 pandecode_log("};\n");
2032 }
2033 }
2034 }
2035 }
2036
2037 if (p->sampler_descriptor) {
2038 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->sampler_descriptor);
2039
2040 if (smem) {
2041 struct mali_sampler_descriptor *s;
2042
2043 mali_ptr d = p->sampler_descriptor;
2044
2045 for (int i = 0; i < sampler_count; ++i) {
2046 s = pandecode_fetch_gpu_mem(smem, d + sizeof(*s) * i, sizeof(*s));
2047
2048 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", d + sizeof(*s) * i, job_no, i);
2049 pandecode_indent++;
2050
2051 pandecode_log(".filter_mode = ");
2052 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2053 pandecode_log_cont(",\n");
2054
2055 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2056 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2057
2058 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2059 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2060 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2061
2062 pandecode_prop("compare_func = %s", pandecode_alt_func(s->compare_func));
2063
2064 if (s->zero || s->zero2) {
2065 pandecode_msg("XXX: sampler zero tripped\n");
2066 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2067 }
2068
2069 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2070
2071 pandecode_prop("border_color = { %f, %f, %f, %f }",
2072 s->border_color[0],
2073 s->border_color[1],
2074 s->border_color[2],
2075 s->border_color[3]);
2076
2077 pandecode_indent--;
2078 pandecode_log("};\n");
2079 }
2080 }
2081 }
2082 }
2083
2084 static void
2085 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2086 {
2087 if (!(p->position_varying || p->occlusion_counter || p->flags))
2088 return;
2089
2090 pandecode_log(".postfix = {\n");
2091 pandecode_indent++;
2092
2093 MEMORY_PROP(p, position_varying);
2094 MEMORY_PROP(p, occlusion_counter);
2095
2096 if (p->flags)
2097 pandecode_prop("flags = %d", p->flags);
2098
2099 pandecode_indent--;
2100 pandecode_log("},\n");
2101 }
2102
2103 static void
2104 pandecode_vertex_only_bfr(struct bifrost_vertex_only *v)
2105 {
2106 pandecode_log_cont("{\n");
2107 pandecode_indent++;
2108
2109 pandecode_prop("unk2 = 0x%x", v->unk2);
2110
2111 if (v->zero0 || v->zero1) {
2112 pandecode_msg("XXX: vertex only zero tripped");
2113 pandecode_prop("zero0 = 0x%" PRIx32, v->zero0);
2114 pandecode_prop("zero1 = 0x%" PRIx64, v->zero1);
2115 }
2116
2117 pandecode_indent--;
2118 pandecode_log("}\n");
2119 }
2120
2121 static void
2122 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2123 {
2124
2125 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2126 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2127
2128 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no);
2129 pandecode_indent++;
2130
2131 if (h->zero) {
2132 pandecode_msg("XXX: tiler heap zero tripped\n");
2133 pandecode_prop("zero = 0x%x", h->zero);
2134 }
2135
2136 for (int i = 0; i < 12; i++) {
2137 if (h->zeros[i] != 0) {
2138 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2139 i, h->zeros[i]);
2140 }
2141 }
2142
2143 pandecode_prop("heap_size = 0x%x", h->heap_size);
2144 MEMORY_PROP(h, tiler_heap_start);
2145 MEMORY_PROP(h, tiler_heap_free);
2146
2147 /* this might point to the beginning of another buffer, when it's
2148 * really the end of the tiler heap buffer, so we have to be careful
2149 * here. but for zero length, we need the same pointer.
2150 */
2151
2152 if (h->tiler_heap_end == h->tiler_heap_start) {
2153 MEMORY_PROP(h, tiler_heap_start);
2154 } else {
2155 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2156 pandecode_prop("tiler_heap_end = %s + 1", a);
2157 free(a);
2158 }
2159
2160 pandecode_indent--;
2161 pandecode_log("};\n");
2162 }
2163
2164 static void
2165 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2166 {
2167 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2168 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2169
2170 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2171
2172 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no);
2173 pandecode_indent++;
2174
2175 if (t->zero0 || t->zero1) {
2176 pandecode_msg("XXX: tiler meta zero tripped\n");
2177 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2178 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2179 }
2180
2181 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2182 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2183
2184 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2185 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2186
2187 for (int i = 0; i < 12; i++) {
2188 if (t->zeros[i] != 0) {
2189 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2190 i, t->zeros[i]);
2191 }
2192 }
2193
2194 pandecode_indent--;
2195 pandecode_log("};\n");
2196 }
2197
2198 static void
2199 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2200 {
2201 pandecode_log(".gl_enables = ");
2202
2203 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2204
2205 pandecode_log_cont(",\n");
2206 }
2207
2208 static void
2209 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2210 {
2211 if (u.pointer == 0x0)
2212 return;
2213
2214 pandecode_log(".primitive_size = {\n");
2215 pandecode_indent++;
2216
2217 if (constant) {
2218 pandecode_prop("constant = %f", u.constant);
2219 } else {
2220 MEMORY_PROP((&u), pointer);
2221 }
2222
2223 pandecode_indent--;
2224 pandecode_log("},\n");
2225 }
2226
2227 static void
2228 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2229 {
2230 pandecode_log_cont("{\n");
2231 pandecode_indent++;
2232
2233 /* TODO: gl_PointSize on Bifrost */
2234 pandecode_primitive_size(t->primitive_size, true);
2235
2236 pandecode_gl_enables(t->gl_enables, JOB_TYPE_TILER);
2237
2238 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2239 || t->zero6 || t->zero7 || t->zero8) {
2240 pandecode_msg("XXX: tiler only zero tripped\n");
2241 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2242 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2243 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2244 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2245 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2246 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2247 pandecode_prop("zero7 = 0x%" PRIx32, t->zero7);
2248 pandecode_prop("zero8 = 0x%" PRIx64, t->zero8);
2249 }
2250
2251 pandecode_indent--;
2252 pandecode_log("},\n");
2253 }
2254
2255 static int
2256 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2257 const struct pandecode_mapped_memory *mem,
2258 mali_ptr payload, int job_no)
2259 {
2260 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2261
2262 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true);
2263
2264 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no);
2265 pandecode_indent++;
2266
2267 pandecode_log(".prefix = ");
2268 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2269
2270 pandecode_log(".vertex = ");
2271 pandecode_vertex_only_bfr(&v->vertex);
2272
2273 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2274
2275 pandecode_indent--;
2276 pandecode_log("};\n");
2277
2278 return sizeof(*v);
2279 }
2280
2281 static int
2282 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2283 const struct pandecode_mapped_memory *mem,
2284 mali_ptr payload, int job_no)
2285 {
2286 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2287
2288 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true);
2289
2290 pandecode_indices(t->prefix.indices, t->prefix.index_count, job_no);
2291 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2292
2293 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no);
2294 pandecode_indent++;
2295
2296 pandecode_log(".prefix = ");
2297 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
2298
2299 pandecode_log(".tiler = ");
2300 pandecode_tiler_only_bfr(&t->tiler, job_no);
2301
2302 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2303
2304 pandecode_indent--;
2305 pandecode_log("};\n");
2306
2307 return sizeof(*t);
2308 }
2309
2310 static int
2311 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2312 const struct pandecode_mapped_memory *mem,
2313 mali_ptr payload, int job_no)
2314 {
2315 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2316
2317 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false);
2318
2319 pandecode_indices(v->prefix.indices, v->prefix.index_count, job_no);
2320
2321 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2322 pandecode_indent++;
2323
2324 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2325 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
2326
2327 bool instanced = v->instance_shift || v->instance_odd;
2328 bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER);
2329
2330 pandecode_log(".prefix = ");
2331 pandecode_vertex_tiler_prefix(&v->prefix, job_no, !instanced && is_graphics);
2332
2333 pandecode_gl_enables(v->gl_enables, h->job_type);
2334
2335 if (v->instance_shift || v->instance_odd) {
2336 pandecode_prop("instance_shift = 0x%d /* %d */",
2337 v->instance_shift, 1 << v->instance_shift);
2338 pandecode_prop("instance_odd = 0x%X /* %d */",
2339 v->instance_odd, (2 * v->instance_odd) + 1);
2340
2341 pandecode_padded_vertices(v->instance_shift, v->instance_odd);
2342 }
2343
2344 if (v->offset_start)
2345 pandecode_prop("offset_start = %d", v->offset_start);
2346
2347 if (v->zero5) {
2348 pandecode_msg("XXX: midgard payload zero tripped\n");
2349 pandecode_prop("zero5 = 0x%" PRIx64, v->zero5);
2350 }
2351
2352 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2353
2354 pandecode_indent--;
2355 pandecode_log("};\n");
2356
2357 return sizeof(*v);
2358 }
2359
2360 static int
2361 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2362 mali_ptr payload, int job_no,
2363 bool is_bifrost)
2364 {
2365 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2366
2367 bool fbd_dumped = false;
2368
2369 if (!is_bifrost && (s->framebuffer & FBD_TYPE) == MALI_SFBD) {
2370 /* Only SFBDs are understood, not MFBDs. We're speculating,
2371 * based on the versioning, kernel code, etc, that the
2372 * difference is between Single FrameBuffer Descriptor and
2373 * Multiple FrmaeBuffer Descriptor; the change apparently lines
2374 * up with multi-framebuffer support being added (T7xx onwards,
2375 * including Gxx). In any event, there's some field shuffling
2376 * that we haven't looked into yet. */
2377
2378 pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true);
2379 fbd_dumped = true;
2380 } else if ((s->framebuffer & FBD_TYPE) == MALI_MFBD) {
2381 /* We don't know if Bifrost supports SFBD's at all, since the
2382 * driver never uses them. And the format is different from
2383 * Midgard anyways, due to the tiler heap and scratchpad being
2384 * moved out into separate structures, so it's not clear what a
2385 * Bifrost SFBD would even look like without getting an actual
2386 * trace, which appears impossible.
2387 */
2388
2389 pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true);
2390 fbd_dumped = true;
2391 }
2392
2393 uintptr_t p = (uintptr_t) s->framebuffer & FBD_MASK;
2394 pandecode_log("struct mali_payload_fragment payload_%"PRIx64"_%d = {\n", payload, job_no);
2395 pandecode_indent++;
2396
2397 /* See the comments by the macro definitions for mathematical context
2398 * on why this is so weird */
2399
2400 if (MALI_TILE_COORD_FLAGS(s->max_tile_coord) || MALI_TILE_COORD_FLAGS(s->min_tile_coord))
2401 pandecode_msg("Tile coordinate flag missed, replay wrong\n");
2402
2403 pandecode_prop("min_tile_coord = MALI_COORDINATE_TO_TILE_MIN(%d, %d)",
2404 MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT,
2405 MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT);
2406
2407 pandecode_prop("max_tile_coord = MALI_COORDINATE_TO_TILE_MAX(%d, %d)",
2408 (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT,
2409 (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT);
2410
2411 /* If the FBD was just decoded, we can refer to it by pointer. If not,
2412 * we have to fallback on offsets. */
2413
2414 const char *fbd_type = s->framebuffer & MALI_MFBD ? "MALI_MFBD" : "MALI_SFBD";
2415
2416 /* TODO: Decode */
2417 unsigned extra_flags = (s->framebuffer & ~FBD_MASK) & ~MALI_MFBD;
2418
2419 if (fbd_dumped)
2420 pandecode_prop("framebuffer = framebuffer_%d_p | %s | 0x%X", job_no,
2421 fbd_type, extra_flags);
2422 else
2423 pandecode_prop("framebuffer = %s | %s | 0x%X", pointer_as_memory_reference(p),
2424 fbd_type, extra_flags);
2425
2426 pandecode_indent--;
2427 pandecode_log("};\n");
2428
2429 return sizeof(*s);
2430 }
2431
2432 static int job_descriptor_number = 0;
2433
2434 int
2435 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost)
2436 {
2437 struct mali_job_descriptor_header *h;
2438
2439 int start_number = 0;
2440
2441 bool first = true;
2442 bool last_size;
2443
2444 do {
2445 struct pandecode_mapped_memory *mem =
2446 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2447
2448 void *payload;
2449
2450 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2451
2452 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2453 * high 32-bits of the 64-bit pointer are reused to store
2454 * something else.
2455 */
2456 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2457 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
2458 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2459
2460 payload = pandecode_fetch_gpu_mem(mem, payload_ptr,
2461 MALI_PAYLOAD_SIZE);
2462
2463 int job_no = job_descriptor_number++;
2464
2465 if (first)
2466 start_number = job_no;
2467
2468 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2469 pandecode_indent++;
2470
2471 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
2472
2473 /* Save for next job fixing */
2474 last_size = h->job_descriptor_size;
2475
2476 if (h->job_descriptor_size)
2477 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2478
2479 if (h->exception_status && h->exception_status != 0x1)
2480 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2481 h->exception_status,
2482 (h->exception_status >> 16) & 0xFFFF,
2483 pandecode_exception_access((h->exception_status >> 8) & 0x3),
2484 h->exception_status & 0xFF);
2485
2486 if (h->first_incomplete_task)
2487 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2488
2489 if (h->fault_pointer)
2490 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2491
2492 if (h->job_barrier)
2493 pandecode_prop("job_barrier = %d", h->job_barrier);
2494
2495 pandecode_prop("job_index = %d", h->job_index);
2496
2497 if (h->unknown_flags)
2498 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2499
2500 if (h->job_dependency_index_1)
2501 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2502
2503 if (h->job_dependency_index_2)
2504 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2505
2506 pandecode_indent--;
2507 pandecode_log("};\n");
2508
2509 /* Do not touch the field yet -- decode the payload first, and
2510 * don't touch that either. This is essential for the uploads
2511 * to occur in sequence and therefore be dynamically allocated
2512 * correctly. Do note the size, however, for that related
2513 * reason. */
2514
2515 switch (h->job_type) {
2516 case JOB_TYPE_SET_VALUE: {
2517 struct mali_payload_set_value *s = payload;
2518 pandecode_log("struct mali_payload_set_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2519 pandecode_indent++;
2520 MEMORY_PROP(s, out);
2521 pandecode_prop("unknown = 0x%" PRIX64, s->unknown);
2522 pandecode_indent--;
2523 pandecode_log("};\n");
2524
2525 break;
2526 }
2527
2528 case JOB_TYPE_TILER:
2529 case JOB_TYPE_VERTEX:
2530 case JOB_TYPE_COMPUTE:
2531 if (bifrost) {
2532 if (h->job_type == JOB_TYPE_TILER)
2533 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no);
2534 else
2535 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no);
2536 } else
2537 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no);
2538
2539 break;
2540
2541 case JOB_TYPE_FRAGMENT:
2542 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost);
2543 break;
2544
2545 default:
2546 break;
2547 }
2548
2549 /* Handle linkage */
2550
2551 if (!first) {
2552 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no - 1);
2553
2554 if (last_size)
2555 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no);
2556 else
2557 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no);
2558 }
2559
2560 first = false;
2561
2562 } while ((jc_gpu_va = h->job_descriptor_size ? h->next_job_64 : h->next_job_32));
2563
2564 return start_number;
2565 }