pan/decode: Downgrade shader property mismatch to warning
[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 <ctype.h>
33 #include "decode.h"
34 #include "util/macros.h"
35 #include "util/u_math.h"
36
37 #include "pan_pretty_print.h"
38 #include "midgard/disassemble.h"
39 #include "bifrost/disassemble.h"
40
41 #include "pan_encoder.h"
42
43 int pandecode_jc(mali_ptr jc_gpu_va, bool bifrost);
44
45 #define MEMORY_PROP(obj, p) {\
46 if (obj->p) { \
47 char *a = pointer_as_memory_reference(obj->p); \
48 pandecode_prop("%s = %s", #p, a); \
49 free(a); \
50 } \
51 }
52
53 #define MEMORY_PROP_DIR(obj, p) {\
54 if (obj.p) { \
55 char *a = pointer_as_memory_reference(obj.p); \
56 pandecode_prop("%s = %s", #p, a); \
57 free(a); \
58 } \
59 }
60
61 /* Semantic logging type.
62 *
63 * Raw: for raw messages to be printed as is.
64 * Message: for helpful information to be commented out in replays.
65 * Property: for properties of a struct
66 *
67 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
68 */
69
70 enum pandecode_log_type {
71 PANDECODE_RAW,
72 PANDECODE_MESSAGE,
73 PANDECODE_PROPERTY
74 };
75
76 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
77 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
78 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
79
80 unsigned pandecode_indent = 0;
81
82 static void
83 pandecode_make_indent(void)
84 {
85 for (unsigned i = 0; i < pandecode_indent; ++i)
86 printf(" ");
87 }
88
89 static void
90 pandecode_log_typed(enum pandecode_log_type type, const char *format, ...)
91 {
92 va_list ap;
93
94 pandecode_make_indent();
95
96 if (type == PANDECODE_MESSAGE)
97 printf("// ");
98 else if (type == PANDECODE_PROPERTY)
99 printf(".");
100
101 va_start(ap, format);
102 vprintf(format, ap);
103 va_end(ap);
104
105 if (type == PANDECODE_PROPERTY)
106 printf(",\n");
107 }
108
109 static void
110 pandecode_log_cont(const char *format, ...)
111 {
112 va_list ap;
113
114 va_start(ap, format);
115 vprintf(format, ap);
116 va_end(ap);
117 }
118
119 /* To check for memory safety issues, validates that the given pointer in GPU
120 * memory is valid, containing at least sz bytes. The goal is to eliminate
121 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
122 * overruns) by statically validating pointers.
123 */
124
125 static void
126 pandecode_validate_buffer(mali_ptr addr, size_t sz)
127 {
128 if (!addr) {
129 pandecode_msg("XXX: null pointer deref");
130 return;
131 }
132
133 /* Find a BO */
134
135 struct pandecode_mapped_memory *bo =
136 pandecode_find_mapped_gpu_mem_containing(addr);
137
138 if (!bo) {
139 pandecode_msg("XXX: invalid memory dereference\n");
140 return;
141 }
142
143 /* Bounds check */
144
145 unsigned offset = addr - bo->gpu_va;
146 unsigned total = offset + sz;
147
148 if (total > bo->length) {
149 pandecode_msg("XXX: buffer overrun. "
150 "Chunk of size %d at offset %d in buffer of size %d. "
151 "Overrun by %d bytes. \n",
152 sz, offset, bo->length, total - bo->length);
153 return;
154 }
155 }
156
157 struct pandecode_flag_info {
158 u64 flag;
159 const char *name;
160 };
161
162 static void
163 pandecode_log_decoded_flags(const struct pandecode_flag_info *flag_info,
164 u64 flags)
165 {
166 bool decodable_flags_found = false;
167
168 for (int i = 0; flag_info[i].name; i++) {
169 if ((flags & flag_info[i].flag) != flag_info[i].flag)
170 continue;
171
172 if (!decodable_flags_found) {
173 decodable_flags_found = true;
174 } else {
175 pandecode_log_cont(" | ");
176 }
177
178 pandecode_log_cont("%s", flag_info[i].name);
179
180 flags &= ~flag_info[i].flag;
181 }
182
183 if (decodable_flags_found) {
184 if (flags)
185 pandecode_log_cont(" | 0x%" PRIx64, flags);
186 } else {
187 pandecode_log_cont("0x%" PRIx64, flags);
188 }
189 }
190
191 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
192 static const struct pandecode_flag_info gl_enable_flag_info[] = {
193 FLAG_INFO(OCCLUSION_QUERY),
194 FLAG_INFO(OCCLUSION_PRECISE),
195 FLAG_INFO(FRONT_CCW_TOP),
196 FLAG_INFO(CULL_FACE_FRONT),
197 FLAG_INFO(CULL_FACE_BACK),
198 {}
199 };
200 #undef FLAG_INFO
201
202 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
203 static const struct pandecode_flag_info clear_flag_info[] = {
204 FLAG_INFO(FAST),
205 FLAG_INFO(SLOW),
206 FLAG_INFO(SLOW_STENCIL),
207 {}
208 };
209 #undef FLAG_INFO
210
211 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
212 static const struct pandecode_flag_info mask_flag_info[] = {
213 FLAG_INFO(R),
214 FLAG_INFO(G),
215 FLAG_INFO(B),
216 FLAG_INFO(A),
217 {}
218 };
219 #undef FLAG_INFO
220
221 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
222 static const struct pandecode_flag_info u3_flag_info[] = {
223 FLAG_INFO(HAS_MSAA),
224 FLAG_INFO(CAN_DISCARD),
225 FLAG_INFO(HAS_BLEND_SHADER),
226 FLAG_INFO(DEPTH_TEST),
227 {}
228 };
229
230 static const struct pandecode_flag_info u4_flag_info[] = {
231 FLAG_INFO(NO_MSAA),
232 FLAG_INFO(NO_DITHER),
233 FLAG_INFO(DEPTH_RANGE_A),
234 FLAG_INFO(DEPTH_RANGE_B),
235 FLAG_INFO(STENCIL_TEST),
236 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER),
237 {}
238 };
239 #undef FLAG_INFO
240
241 #define FLAG_INFO(flag) { MALI_FRAMEBUFFER_##flag, "MALI_FRAMEBUFFER_" #flag }
242 static const struct pandecode_flag_info fb_fmt_flag_info[] = {
243 FLAG_INFO(MSAA_A),
244 FLAG_INFO(MSAA_B),
245 FLAG_INFO(MSAA_8),
246 {}
247 };
248 #undef FLAG_INFO
249
250 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
251 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
252 FLAG_INFO(MSAA),
253 FLAG_INFO(SRGB),
254 {}
255 };
256 #undef FLAG_INFO
257
258 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
259 static const struct pandecode_flag_info mfbd_extra_flag_info[] = {
260 FLAG_INFO(PRESENT),
261 FLAG_INFO(AFBC),
262 FLAG_INFO(ZS),
263 {}
264 };
265 #undef FLAG_INFO
266
267 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
268 static const struct pandecode_flag_info shader_midgard1_flag_info [] = {
269 FLAG_INFO(EARLY_Z),
270 FLAG_INFO(HELPER_INVOCATIONS),
271 FLAG_INFO(READS_TILEBUFFER),
272 FLAG_INFO(READS_ZS),
273 {}
274 };
275 #undef FLAG_INFO
276
277 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
278 static const struct pandecode_flag_info mfbd_flag_info [] = {
279 FLAG_INFO(DEPTH_WRITE),
280 FLAG_INFO(EXTRA),
281 {}
282 };
283 #undef FLAG_INFO
284
285 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
286 static const struct pandecode_flag_info sampler_flag_info [] = {
287 FLAG_INFO(MAG_NEAREST),
288 FLAG_INFO(MIN_NEAREST),
289 FLAG_INFO(MIP_LINEAR_1),
290 FLAG_INFO(MIP_LINEAR_2),
291 FLAG_INFO(NORM_COORDS),
292 {}
293 };
294 #undef FLAG_INFO
295
296 extern char *replace_fragment;
297 extern char *replace_vertex;
298
299 static char *
300 pandecode_job_type(enum mali_job_type type)
301 {
302 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
303
304 switch (type) {
305 DEFINE_CASE(NULL);
306 DEFINE_CASE(SET_VALUE);
307 DEFINE_CASE(CACHE_FLUSH);
308 DEFINE_CASE(COMPUTE);
309 DEFINE_CASE(VERTEX);
310 DEFINE_CASE(TILER);
311 DEFINE_CASE(FUSED);
312 DEFINE_CASE(FRAGMENT);
313
314 case JOB_NOT_STARTED:
315 return "NOT_STARTED";
316
317 default:
318 pandecode_log("Warning! Unknown job type %x\n", type);
319 return "!?!?!?";
320 }
321
322 #undef DEFINE_CASE
323 }
324
325 static char *
326 pandecode_draw_mode(enum mali_draw_mode mode)
327 {
328 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
329
330 switch (mode) {
331 DEFINE_CASE(DRAW_NONE);
332 DEFINE_CASE(POINTS);
333 DEFINE_CASE(LINES);
334 DEFINE_CASE(TRIANGLES);
335 DEFINE_CASE(TRIANGLE_STRIP);
336 DEFINE_CASE(TRIANGLE_FAN);
337 DEFINE_CASE(LINE_STRIP);
338 DEFINE_CASE(LINE_LOOP);
339 DEFINE_CASE(POLYGON);
340 DEFINE_CASE(QUADS);
341 DEFINE_CASE(QUAD_STRIP);
342
343 default:
344 pandecode_msg("XXX: invalid draw mode %X\n", mode);
345 return "";
346 }
347
348 #undef DEFINE_CASE
349 }
350
351 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
352 static char *
353 pandecode_func(enum mali_func mode)
354 {
355 switch (mode) {
356 DEFINE_CASE(NEVER);
357 DEFINE_CASE(LESS);
358 DEFINE_CASE(EQUAL);
359 DEFINE_CASE(LEQUAL);
360 DEFINE_CASE(GREATER);
361 DEFINE_CASE(NOTEQUAL);
362 DEFINE_CASE(GEQUAL);
363 DEFINE_CASE(ALWAYS);
364
365 default:
366 pandecode_msg("XXX: invalid func %X\n", mode);
367 return "";
368 }
369 }
370 #undef DEFINE_CASE
371
372 /* Why is this duplicated? Who knows... */
373 #define DEFINE_CASE(name) case MALI_ALT_FUNC_ ## name: return "MALI_ALT_FUNC_" #name
374 static char *
375 pandecode_alt_func(enum mali_alt_func mode)
376 {
377 switch (mode) {
378 DEFINE_CASE(NEVER);
379 DEFINE_CASE(LESS);
380 DEFINE_CASE(EQUAL);
381 DEFINE_CASE(LEQUAL);
382 DEFINE_CASE(GREATER);
383 DEFINE_CASE(NOTEQUAL);
384 DEFINE_CASE(GEQUAL);
385 DEFINE_CASE(ALWAYS);
386
387 default:
388 pandecode_msg("XXX: invalid alt func %X\n", mode);
389 return "";
390 }
391 }
392 #undef DEFINE_CASE
393
394 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
395 static char *
396 pandecode_stencil_op(enum mali_stencil_op op)
397 {
398 switch (op) {
399 DEFINE_CASE(KEEP);
400 DEFINE_CASE(REPLACE);
401 DEFINE_CASE(ZERO);
402 DEFINE_CASE(INVERT);
403 DEFINE_CASE(INCR_WRAP);
404 DEFINE_CASE(DECR_WRAP);
405 DEFINE_CASE(INCR);
406 DEFINE_CASE(DECR);
407
408 default:
409 pandecode_msg("XXX: invalid stencil op %X\n", op);
410 return "";
411 }
412 }
413
414 #undef DEFINE_CASE
415
416 #define DEFINE_CASE(name) case MALI_ATTR_ ## name: return "MALI_ATTR_" #name
417 static char *pandecode_attr_mode_short(enum mali_attr_mode mode)
418 {
419 switch(mode) {
420 /* TODO: Combine to just "instanced" once this can be done
421 * unambiguously in all known cases */
422 case MALI_ATTR_POT_DIVIDE:
423 return "instanced_pot";
424 case MALI_ATTR_MODULO:
425 return "instanced_mod";
426 case MALI_ATTR_NPOT_DIVIDE:
427 return "instanced_npot";
428 case MALI_ATTR_IMAGE:
429 return "image";
430 case MALI_ATTR_INTERNAL:
431 return "internal";
432 default:
433 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
434 return "";
435 }
436 }
437
438 #undef DEFINE_CASE
439
440 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
441 static char *
442 pandecode_wrap_mode(enum mali_wrap_mode op)
443 {
444 switch (op) {
445 DEFINE_CASE(REPEAT);
446 DEFINE_CASE(CLAMP_TO_EDGE);
447 DEFINE_CASE(CLAMP_TO_BORDER);
448 DEFINE_CASE(MIRRORED_REPEAT);
449
450 default:
451 pandecode_msg("XXX: invalid wrap mode %X\n", op);
452 return "";
453 }
454 }
455 #undef DEFINE_CASE
456
457 #define DEFINE_CASE(name) case MALI_MFBD_BLOCK_## name: return "MALI_MFBD_BLOCK_" #name
458 static char *
459 pandecode_mfbd_block_format(enum mali_mfbd_block_format fmt)
460 {
461 switch (fmt) {
462 DEFINE_CASE(TILED);
463 DEFINE_CASE(UNKNOWN);
464 DEFINE_CASE(LINEAR);
465 DEFINE_CASE(AFBC);
466
467 default:
468 unreachable("Invalid case");
469 }
470 }
471 #undef DEFINE_CASE
472
473 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
474 static char *
475 pandecode_exception_access(enum mali_exception_access access)
476 {
477 switch (access) {
478 DEFINE_CASE(NONE);
479 DEFINE_CASE(EXECUTE);
480 DEFINE_CASE(READ);
481 DEFINE_CASE(WRITE);
482
483 default:
484 unreachable("Invalid case");
485 }
486 }
487 #undef DEFINE_CASE
488
489 /* Midgard's tiler descriptor is embedded within the
490 * larger FBD */
491
492 static void
493 pandecode_midgard_tiler_descriptor(
494 const struct midgard_tiler_descriptor *t,
495 unsigned width,
496 unsigned height,
497 bool is_fragment)
498 {
499 pandecode_log(".tiler = {\n");
500 pandecode_indent++;
501
502 if (t->hierarchy_mask == MALI_TILER_DISABLED)
503 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
504 else
505 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
506
507 /* We know this name from the kernel, but we never see it nonzero */
508
509 if (t->flags)
510 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
511
512 MEMORY_PROP(t, polygon_list);
513
514 /* The body is offset from the base of the polygon list */
515 assert(t->polygon_list_body > t->polygon_list);
516 unsigned body_offset = t->polygon_list_body - t->polygon_list;
517
518 /* It needs to fit inside the reported size */
519 assert(t->polygon_list_size >= body_offset);
520
521 /* Check that we fit */
522 struct pandecode_mapped_memory *plist =
523 pandecode_find_mapped_gpu_mem_containing(t->polygon_list);
524
525 assert(t->polygon_list_size <= plist->length);
526
527 /* Now that we've sanity checked, we'll try to calculate the sizes
528 * ourselves for comparison */
529
530 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask);
531 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask);
532
533 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
534 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
535 ref_header, ref_size);
536 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
537 pandecode_msg("body offset %d\n", body_offset);
538 }
539
540 /* The tiler heap has a start and end specified -- it should be
541 * identical to what we have in the BO. The exception is if tiling is
542 * disabled. */
543
544 MEMORY_PROP(t, heap_start);
545 assert(t->heap_end >= t->heap_start);
546
547 struct pandecode_mapped_memory *heap =
548 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
549
550 unsigned heap_size = t->heap_end - t->heap_start;
551
552 /* Tiling is enabled with a special flag */
553 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
554 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
555
556 bool tiling_enabled = hierarchy_mask;
557
558 if (tiling_enabled) {
559 /* When tiling is enabled, the heap should be a tight fit */
560 unsigned heap_offset = t->heap_start - heap->gpu_va;
561 if ((heap_offset + heap_size) != heap->length) {
562 pandecode_msg("XXX: heap size %d (expected %d)\n",
563 heap_size, heap->length - heap_offset);
564 }
565
566 /* We should also have no other flags */
567 if (tiler_flags)
568 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
569 } else {
570 /* When tiling is disabled, we should have that flag and no others */
571
572 if (tiler_flags != MALI_TILER_DISABLED) {
573 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
574 tiler_flags);
575 }
576
577 /* We should also have an empty heap */
578 if (heap_size) {
579 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
580 heap_size);
581 }
582
583 /* Disabled tiling is used only for clear-only jobs, which are
584 * purely FRAGMENT, so we should never see this for
585 * non-FRAGMENT descriptors. */
586
587 if (!is_fragment)
588 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
589 }
590
591 /* We've never seen weights used in practice, but we know from the
592 * kernel these fields is there */
593
594 bool nonzero_weights = false;
595
596 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
597 nonzero_weights |= t->weights[w] != 0x0;
598 }
599
600 if (nonzero_weights) {
601 pandecode_log(".weights = {");
602
603 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
604 pandecode_log("%d, ", t->weights[w]);
605 }
606
607 pandecode_log("},");
608 }
609
610 pandecode_indent--;
611 pandecode_log("}\n");
612 }
613
614 /* Information about the framebuffer passed back for
615 * additional analysis */
616
617 struct pandecode_fbd {
618 unsigned width;
619 unsigned height;
620 unsigned rt_count;
621 bool has_extra;
622 };
623
624 static struct pandecode_fbd
625 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment)
626 {
627 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
628 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
629
630 struct pandecode_fbd info;
631
632 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
633 pandecode_indent++;
634
635 pandecode_prop("unknown1 = 0x%" PRIx32, s->unknown1);
636 pandecode_prop("unknown2 = 0x%" PRIx32, s->unknown2);
637
638 pandecode_log(".format = ");
639 pandecode_log_decoded_flags(fb_fmt_flag_info, s->format);
640 pandecode_log_cont(",\n");
641
642 info.width = s->width + 1;
643 info.height = s->height + 1;
644 info.rt_count = 1;
645
646 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
647 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
648
649 MEMORY_PROP(s, framebuffer);
650 pandecode_prop("stride = %d", s->stride);
651
652 /* Earlier in the actual commandstream -- right before width -- but we
653 * delay to flow nicer */
654
655 pandecode_log(".clear_flags = ");
656 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
657 pandecode_log_cont(",\n");
658
659 if (s->depth_buffer | s->depth_buffer_enable) {
660 MEMORY_PROP(s, depth_buffer);
661 pandecode_prop("depth_buffer_enable = %s", DS_ENABLE(s->depth_buffer_enable));
662 }
663
664 if (s->stencil_buffer | s->stencil_buffer_enable) {
665 MEMORY_PROP(s, stencil_buffer);
666 pandecode_prop("stencil_buffer_enable = %s", DS_ENABLE(s->stencil_buffer_enable));
667 }
668
669 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
670 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
671 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
672 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
673 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
674 }
675
676 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
677 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
678 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
679 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
680 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
681 }
682
683 if (s->clear_stencil) {
684 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
685 }
686
687 MEMORY_PROP(s, unknown_address_0);
688 const struct midgard_tiler_descriptor t = s->tiler;
689 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment);
690
691 pandecode_indent--;
692 pandecode_log("};\n");
693
694 pandecode_prop("zero0 = 0x%" PRIx64, s->zero0);
695 pandecode_prop("zero1 = 0x%" PRIx64, s->zero1);
696 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
697 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
698
699 printf(".zero3 = {");
700
701 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
702 printf("%X, ", s->zero3[i]);
703
704 printf("},\n");
705
706 printf(".zero6 = {");
707
708 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
709 printf("%X, ", s->zero6[i]);
710
711 printf("},\n");
712
713 return info;
714 }
715
716 static void
717 pandecode_u32_slide(unsigned name, const u32 *slide, unsigned count)
718 {
719 pandecode_log(".unknown%d = {", name);
720
721 for (int i = 0; i < count; ++i)
722 printf("%X, ", slide[i]);
723
724 pandecode_log("},\n");
725 }
726
727 #define SHORT_SLIDE(num) \
728 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
729
730 static void
731 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
732 {
733 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
734 const struct mali_compute_fbd *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
735
736 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
737 pandecode_indent++;
738
739 SHORT_SLIDE(1);
740
741 pandecode_indent--;
742 printf("},\n");
743 }
744
745 /* Extracts the number of components associated with a Mali format */
746
747 static unsigned
748 pandecode_format_component_count(enum mali_format fmt)
749 {
750 /* Mask out the format class */
751 unsigned top = fmt & 0b11100000;
752
753 switch (top) {
754 case MALI_FORMAT_SNORM:
755 case MALI_FORMAT_UINT:
756 case MALI_FORMAT_UNORM:
757 case MALI_FORMAT_SINT:
758 return ((fmt >> 3) & 3) + 1;
759 default:
760 /* TODO: Validate */
761 return 4;
762 }
763 }
764
765 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
766
767 static unsigned
768 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
769 {
770 unsigned mask = 0;
771 assert(MALI_CHANNEL_RED == 0);
772
773 for (unsigned c = 0; c < 4; ++c) {
774 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
775
776 if (chan <= MALI_CHANNEL_ALPHA)
777 mask |= (1 << chan);
778 }
779
780 return mask;
781 }
782
783 /* Validates that a (format, swizzle) pair is valid, in the sense that the
784 * swizzle doesn't access any components that are undefined in the format.
785 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
786 * omitted */
787
788 static bool
789 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
790 {
791 unsigned nr_comp = pandecode_format_component_count(fmt);
792 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
793 unsigned valid_mask = (1 << nr_comp) - 1;
794 unsigned invalid_mask = ~valid_mask;
795
796 if (access_mask & invalid_mask) {
797 pandecode_msg("XXX: invalid components accessed\n");
798 return false;
799 }
800
801 /* Check for the default non-swizzling swizzle so we can suppress
802 * useless printing for the defaults */
803
804 unsigned default_swizzles[4] = {
805 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
806 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
807 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
808 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
809 };
810
811 return (swizzle == default_swizzles[nr_comp - 1]);
812 }
813
814 /* Maps MALI_RGBA32F to rgba32f, etc */
815
816 static void
817 pandecode_format_short(enum mali_format fmt, bool srgb)
818 {
819 /* We want a type-like format, so cut off the initial MALI_ */
820 char *format = pandecode_format(fmt);
821 format += strlen("MALI_");
822
823 unsigned len = strlen(format);
824 char *lower_format = calloc(1, len + 1);
825
826 for (unsigned i = 0; i < len; ++i)
827 lower_format[i] = tolower(format[i]);
828
829 /* Sanity check sRGB flag is applied to RGB, per the name */
830 if (srgb && lower_format[0] != 'r')
831 pandecode_msg("XXX: sRGB applied to non-colour format\n");
832
833 /* Just prefix with an s, so you get formats like srgba8_unorm */
834 if (srgb)
835 pandecode_log_cont("s");
836
837 pandecode_log_cont("%s", lower_format);
838 free(lower_format);
839 }
840
841 static void
842 pandecode_swizzle(unsigned swizzle, enum mali_format format)
843 {
844 /* First, do some validation */
845 bool trivial_swizzle = pandecode_validate_format_swizzle(
846 format, swizzle);
847
848 if (trivial_swizzle)
849 return;
850
851 /* Next, print the swizzle */
852 pandecode_log_cont(".");
853
854 static const char components[] = "rgba01";
855
856 for (unsigned c = 0; c < 4; ++c) {
857 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
858
859 if (chan >= MALI_CHANNEL_RESERVED_0) {
860 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
861 continue;
862 }
863 pandecode_log_cont("%c", components[chan]);
864 }
865 }
866
867 static void
868 pandecode_rt_format(struct mali_rt_format format)
869 {
870 pandecode_log(".format = {\n");
871 pandecode_indent++;
872
873 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
874 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
875 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
876
877 pandecode_prop("block = %s",
878 pandecode_mfbd_block_format(format.block));
879
880 /* TODO: Map formats so we can check swizzles and print nicely */
881 pandecode_log("swizzle");
882 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
883 pandecode_log_cont(",\n");
884
885 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
886 MALI_NEGATIVE(format.nr_channels));
887
888 pandecode_log(".flags = ");
889 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
890 pandecode_log_cont(",\n");
891
892 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
893 * which is a faster hardware-based alternative to the wallpaper method
894 * to preserve framebuffer contents across frames. In practice, MFBD
895 * preload is buggy on Midgard, and so this is a chicken bit. If this
896 * bit isn't set, most likely something broke unrelated to preload */
897
898 if (!format.no_preload) {
899 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
900 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
901 }
902
903 if (format.zero)
904 pandecode_prop("zero = 0x%" PRIx32, format.zero);
905
906 pandecode_indent--;
907 pandecode_log("},\n");
908 }
909
910 static void
911 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct bifrost_framebuffer *fb)
912 {
913 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
914 pandecode_indent++;
915
916 for (int i = 0; i < MALI_NEGATIVE(fb->rt_count_1); i++) {
917 mali_ptr rt_va = gpu_va + i * sizeof(struct bifrost_render_target);
918 struct pandecode_mapped_memory *mem =
919 pandecode_find_mapped_gpu_mem_containing(rt_va);
920 const struct bifrost_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
921
922 pandecode_log("{\n");
923 pandecode_indent++;
924
925 pandecode_rt_format(rt->format);
926
927 if (rt->format.block == MALI_MFBD_BLOCK_AFBC) {
928 pandecode_log(".afbc = {\n");
929 pandecode_indent++;
930
931 char *a = pointer_as_memory_reference(rt->afbc.metadata);
932 pandecode_prop("metadata = %s", a);
933 free(a);
934
935 pandecode_prop("stride = %d", rt->afbc.stride);
936 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
937
938 pandecode_indent--;
939 pandecode_log("},\n");
940 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.unk) {
941 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
942 rt->afbc.metadata,
943 rt->afbc.stride,
944 rt->afbc.unk);
945 }
946
947 MEMORY_PROP(rt, framebuffer);
948 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
949
950 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
951 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
952 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
953 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
954 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
955 }
956
957 if (rt->zero1 || rt->zero2 || rt->zero3) {
958 pandecode_msg("XXX: render target zeros tripped\n");
959 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
960 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
961 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
962 }
963
964 pandecode_indent--;
965 pandecode_log("},\n");
966 }
967
968 pandecode_indent--;
969 pandecode_log("};\n");
970 }
971
972 static struct pandecode_fbd
973 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment)
974 {
975 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
976 const struct bifrost_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
977
978 struct pandecode_fbd info;
979
980 if (fb->sample_locations) {
981 /* The blob stores all possible sample locations in a single buffer
982 * allocated on startup, and just switches the pointer when switching
983 * MSAA state. For now, we just put the data into the cmdstream, but we
984 * should do something like what the blob does with a real driver.
985 *
986 * There seem to be 32 slots for sample locations, followed by another
987 * 16. The second 16 is just the center location followed by 15 zeros
988 * in all the cases I've identified (maybe shader vs. depth/color
989 * samples?).
990 */
991
992 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->sample_locations);
993
994 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->sample_locations);
995
996 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
997 pandecode_indent++;
998
999 for (int i = 0; i < 32 + 16; i++) {
1000 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
1001 }
1002
1003 pandecode_indent--;
1004 pandecode_log("};\n");
1005 }
1006
1007 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
1008 pandecode_indent++;
1009
1010 pandecode_prop("unk0 = 0x%x", fb->unk0);
1011
1012 if (fb->sample_locations)
1013 pandecode_prop("sample_locations = sample_locations_%d", job_no);
1014
1015 /* Assume that unknown1 was emitted in the last job for
1016 * now */
1017 MEMORY_PROP(fb, unknown1);
1018
1019 info.width = fb->width1 + 1;
1020 info.height = fb->height1 + 1;
1021 info.rt_count = fb->rt_count_1 + 1;
1022
1023 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
1024 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
1025 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
1026 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
1027
1028 pandecode_prop("unk1 = 0x%x", fb->unk1);
1029 pandecode_prop("unk2 = 0x%x", fb->unk2);
1030 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1031 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1032
1033 pandecode_log(".mfbd_flags = ");
1034 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1035 pandecode_log_cont(",\n");
1036
1037 if (fb->clear_stencil)
1038 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1039
1040 if (fb->clear_depth)
1041 pandecode_prop("clear_depth = %f", fb->clear_depth);
1042
1043 /* TODO: What is this? Let's not blow up.. */
1044 if (fb->unknown2 != 0x1F)
1045 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1046
1047 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1048 MEMORY_PROP(fb, scratchpad);
1049 const struct midgard_tiler_descriptor t = fb->tiler;
1050 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment);
1051
1052 if (fb->zero3 || fb->zero4) {
1053 pandecode_msg("XXX: framebuffer zeros tripped\n");
1054 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1055 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1056 }
1057
1058 pandecode_indent--;
1059 pandecode_log("};\n");
1060
1061 gpu_va += sizeof(struct bifrost_framebuffer);
1062
1063 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
1064
1065 if (info.has_extra) {
1066 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1067 const struct bifrost_fb_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1068
1069 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1070 pandecode_indent++;
1071
1072 MEMORY_PROP(fbx, checksum);
1073
1074 if (fbx->checksum_stride)
1075 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1076
1077 pandecode_log(".flags = ");
1078 pandecode_log_decoded_flags(mfbd_extra_flag_info, fbx->flags);
1079 pandecode_log_cont(",\n");
1080
1081 if (fbx->flags & MALI_EXTRA_AFBC_ZS) {
1082 pandecode_log(".ds_afbc = {\n");
1083 pandecode_indent++;
1084
1085 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1086 pandecode_prop("depth_stencil_afbc_stride = %d",
1087 fbx->ds_afbc.depth_stencil_afbc_stride);
1088 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1089
1090 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
1091 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1092 pandecode_prop("zero1 = 0x%" PRIx32,
1093 fbx->ds_afbc.zero1);
1094 pandecode_prop("padding = 0x%" PRIx64,
1095 fbx->ds_afbc.padding);
1096 }
1097
1098 pandecode_indent--;
1099 pandecode_log("},\n");
1100 } else {
1101 pandecode_log(".ds_linear = {\n");
1102 pandecode_indent++;
1103
1104 if (fbx->ds_linear.depth) {
1105 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1106 pandecode_prop("depth_stride = %d",
1107 fbx->ds_linear.depth_stride);
1108 }
1109
1110 if (fbx->ds_linear.stencil) {
1111 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1112 pandecode_prop("stencil_stride = %d",
1113 fbx->ds_linear.stencil_stride);
1114 }
1115
1116 if (fbx->ds_linear.depth_stride_zero ||
1117 fbx->ds_linear.stencil_stride_zero ||
1118 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
1119 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1120 pandecode_prop("depth_stride_zero = 0x%x",
1121 fbx->ds_linear.depth_stride_zero);
1122 pandecode_prop("stencil_stride_zero = 0x%x",
1123 fbx->ds_linear.stencil_stride_zero);
1124 pandecode_prop("zero1 = 0x%" PRIx32,
1125 fbx->ds_linear.zero1);
1126 pandecode_prop("zero2 = 0x%" PRIx32,
1127 fbx->ds_linear.zero2);
1128 }
1129
1130 pandecode_indent--;
1131 pandecode_log("},\n");
1132 }
1133
1134 if (fbx->zero3 || fbx->zero4) {
1135 pandecode_msg("XXX: fb_extra zeros tripped\n");
1136 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1137 pandecode_prop("zero4 = 0x%" PRIx64, fbx->zero4);
1138 }
1139
1140 pandecode_indent--;
1141 pandecode_log("};\n");
1142
1143 gpu_va += sizeof(struct bifrost_fb_extra);
1144 }
1145
1146 if (is_fragment)
1147 pandecode_render_target(gpu_va, job_no, fb);
1148
1149 return info;
1150 }
1151
1152 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1153 * count */
1154
1155 static void
1156 pandecode_padded_vertices(unsigned shift, unsigned k)
1157 {
1158 unsigned odd = 2*k + 1;
1159 unsigned pot = 1 << shift;
1160 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1161 }
1162
1163 /* Given a magic divisor, recover what we were trying to divide by.
1164 *
1165 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1166 * 0 <= m < 2^N, for N bits in m.
1167 *
1168 * Let q represent the number we would like to divide by.
1169 *
1170 * By definition of a magic divisor for N-bit unsigned integers (a number you
1171 * multiply by to magically get division), m is a number such that:
1172 *
1173 * (m * x) & (2^N - 1) = floor(x/q).
1174 * for all x on Z where 0 <= x < 2^N
1175 *
1176 * Ignore the case where any of the above values equals zero; it is irrelevant
1177 * for our purposes (instanced arrays).
1178 *
1179 * Choose x = q. Then:
1180 *
1181 * (m * x) & (2^N - 1) = floor(x/q).
1182 * (m * q) & (2^N - 1) = floor(q/q).
1183 *
1184 * floor(q/q) = floor(1) = 1, therefore:
1185 *
1186 * (m * q) & (2^N - 1) = 1
1187 *
1188 * Recall the identity that the bitwise AND of one less than a power-of-two
1189 * equals the modulo with that power of two, i.e. for all x:
1190 *
1191 * x & (2^N - 1) = x % N
1192 *
1193 * Therefore:
1194 *
1195 * mq % (2^N) = 1
1196 *
1197 * By definition, a modular multiplicative inverse of a number m is the number
1198 * q such that with respect to a modulos M:
1199 *
1200 * mq % M = 1
1201 *
1202 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1203 *
1204 */
1205
1206 static void
1207 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1208 {
1209 #if 0
1210 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1211 * shift) the most lame way possible... just repeatedly add.
1212 * Asymptoptically slow but nobody cares in practice, unless you have
1213 * massive numbers of vertices or high divisors. */
1214
1215 unsigned inverse = 0;
1216
1217 /* Magic implicitly has the highest bit set */
1218 magic |= (1 << 31);
1219
1220 /* Depending on rounding direction */
1221 if (extra)
1222 magic++;
1223
1224 for (;;) {
1225 uint32_t product = magic * inverse;
1226
1227 if (shift) {
1228 product >>= shift;
1229 }
1230
1231 if (product == 1)
1232 break;
1233
1234 ++inverse;
1235 }
1236
1237 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1238
1239 /* Recall we're supposed to divide by (gl_level_divisor *
1240 * padded_num_vertices) */
1241
1242 unsigned padded_num_vertices = inverse / orig_divisor;
1243
1244 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1245 #endif
1246 }
1247
1248 static void
1249 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1250 mali_ptr addr, int job_no, char *suffix,
1251 int count, bool varying)
1252 {
1253 char *prefix = varying ? "varying" : "attribute";
1254
1255 if (!addr) {
1256 pandecode_msg("no %s\n", prefix);
1257 return;
1258 }
1259
1260 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1261
1262 for (int i = 0; i < count; ++i) {
1263 enum mali_attr_mode mode = attr[i].elements & 7;
1264
1265 if (mode == MALI_ATTR_UNUSED)
1266 pandecode_msg("XXX: unused attribute record\n");
1267
1268 /* For non-linear records, we need to print the type of record */
1269 if (mode != MALI_ATTR_LINEAR)
1270 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode));
1271
1272 /* Print the name to link with attr_meta */
1273 pandecode_log_cont("%s_%d", prefix, i);
1274
1275 /* Print the stride and size */
1276 pandecode_log_cont("<%u>[%u]", attr[i].stride, attr[i].size);
1277
1278 /* Check: the size must be divisible by the stride */
1279 if (attr[i].size % attr[i].stride)
1280 pandecode_msg("XXX: size not divisible by stride\n");
1281
1282 /* TODO: Sanity check the quotient itself -- it should equal
1283 * vertex count (or something computed from it for instanced)
1284 * which means we can check and elide */
1285
1286 /* Finally, print the pointer */
1287 mali_ptr raw_elements = attr[i].elements & ~7;
1288 char *a = pointer_as_memory_reference(raw_elements);
1289 pandecode_log_cont(" = (%s);\n", a);
1290 free(a);
1291
1292 /* Check the pointer */
1293 pandecode_validate_buffer(raw_elements, attr[i].size);
1294
1295 /* shift/extra_flags exist only for instanced */
1296 if (attr[i].shift | attr[i].extra_flags) {
1297 /* These are set to random values by the blob for
1298 * varyings, most likely a symptom of uninitialized
1299 * memory where the hardware masked the bug. As such we
1300 * put this at a warning, not an error. */
1301
1302 if (mode == MALI_ATTR_LINEAR)
1303 pandecode_msg("warn: instancing fields set for linear\n");
1304
1305 pandecode_prop("shift = %d", attr[i].shift);
1306 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1307 }
1308
1309 /* Decode further where possible */
1310
1311 if (mode == MALI_ATTR_MODULO) {
1312 pandecode_padded_vertices(
1313 attr[i].shift,
1314 attr[i].extra_flags);
1315 }
1316
1317 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1318 i++;
1319 pandecode_log("{\n");
1320 pandecode_indent++;
1321 pandecode_prop("unk = 0x%x", attr[i].unk);
1322 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1323 if (attr[i].zero != 0)
1324 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1325 pandecode_prop("divisor = %d", attr[i].divisor);
1326 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1327 pandecode_indent--;
1328 pandecode_log("}, \n");
1329 }
1330
1331 }
1332
1333 pandecode_log("\n");
1334 }
1335
1336 static mali_ptr
1337 pandecode_shader_address(const char *name, mali_ptr ptr)
1338 {
1339 /* TODO: Decode flags */
1340 mali_ptr shader_ptr = ptr & ~15;
1341
1342 char *a = pointer_as_memory_reference(shader_ptr);
1343 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1344 free(a);
1345
1346 return shader_ptr;
1347 }
1348
1349 static void
1350 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1351 {
1352 unsigned any_nonzero =
1353 stencil->ref | stencil->mask | stencil->func |
1354 stencil->sfail | stencil->dpfail | stencil->dppass;
1355
1356 if (any_nonzero == 0)
1357 return;
1358
1359 const char *func = pandecode_func(stencil->func);
1360 const char *sfail = pandecode_stencil_op(stencil->sfail);
1361 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1362 const char *dppass = pandecode_stencil_op(stencil->dppass);
1363
1364 if (stencil->zero)
1365 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero);
1366
1367 pandecode_log(".stencil_%s = {\n", name);
1368 pandecode_indent++;
1369 pandecode_prop("ref = %d", stencil->ref);
1370 pandecode_prop("mask = 0x%02X", stencil->mask);
1371 pandecode_prop("func = %s", func);
1372 pandecode_prop("sfail = %s", sfail);
1373 pandecode_prop("dpfail = %s", dpfail);
1374 pandecode_prop("dppass = %s", dppass);
1375 pandecode_indent--;
1376 pandecode_log("},\n");
1377 }
1378
1379 static void
1380 pandecode_blend_equation(const struct mali_blend_equation *blend)
1381 {
1382 if (blend->zero1)
1383 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1384
1385 pandecode_log(".equation = {\n");
1386 pandecode_indent++;
1387
1388 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1389 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1390
1391 pandecode_log(".color_mask = ");
1392 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1393 pandecode_log_cont(",\n");
1394
1395 pandecode_indent--;
1396 pandecode_log("},\n");
1397 }
1398
1399 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1400
1401 static unsigned
1402 decode_bifrost_constant(u16 constant)
1403 {
1404 float lo = (float) (constant & 0xFF);
1405 float hi = (float) (constant >> 8);
1406
1407 return (hi / 255.0) + (lo / 65535.0);
1408 }
1409
1410 static mali_ptr
1411 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1412 {
1413 struct bifrost_blend_rt *b =
1414 ((struct bifrost_blend_rt *) descs) + rt_no;
1415
1416 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1417 pandecode_indent++;
1418
1419 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1420 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1421 b->constant, decode_bifrost_constant(b->constant));
1422
1423 /* TODO figure out blend shader enable bit */
1424 pandecode_blend_equation(&b->equation);
1425 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1426 pandecode_prop("index = 0x%" PRIx16, b->index);
1427 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1428
1429 pandecode_indent--;
1430 pandecode_log("},\n");
1431
1432 return 0;
1433 }
1434
1435 static mali_ptr
1436 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1437 {
1438 /* constant/equation is in a union */
1439 if (!blend->shader)
1440 return 0;
1441
1442 pandecode_log(".blend = {\n");
1443 pandecode_indent++;
1444
1445 if (is_shader) {
1446 pandecode_shader_address("shader", blend->shader);
1447 } else {
1448 pandecode_blend_equation(&blend->equation);
1449 pandecode_prop("constant = %f", blend->constant);
1450 }
1451
1452 pandecode_indent--;
1453 pandecode_log("},\n");
1454
1455 /* Return blend shader to disassemble if present */
1456 return is_shader ? (blend->shader & ~0xF) : 0;
1457 }
1458
1459 static mali_ptr
1460 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1461 {
1462 struct midgard_blend_rt *b =
1463 ((struct midgard_blend_rt *) descs) + rt_no;
1464
1465 /* Flags determine presence of blend shader */
1466 bool is_shader = (b->flags & 0xF) >= 0x2;
1467
1468 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1469 pandecode_indent++;
1470
1471 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1472
1473 union midgard_blend blend = b->blend;
1474 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1475
1476 pandecode_indent--;
1477 pandecode_log("};\n");
1478
1479 return shader;
1480 }
1481
1482 /* Attributes and varyings have descriptor records, which contain information
1483 * about their format and ordering with the attribute/varying buffers. We'll
1484 * want to validate that the combinations specified are self-consistent.
1485 */
1486
1487 static int
1488 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1489 {
1490 char base[128];
1491 char *prefix = varying ? "varying" : "attribute";
1492 unsigned max_index = 0;
1493 snprintf(base, sizeof(base), "%s_meta", prefix);
1494
1495 struct mali_attr_meta *attr_meta;
1496 mali_ptr p = varying ? (v->varying_meta & ~0xF) : v->attribute_meta;
1497
1498 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1499
1500 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1501 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1502 sizeof(*attr_mem));
1503
1504 /* If the record is discard, it should be zero for everything else */
1505
1506 if (attr_meta->format == MALI_VARYING_DISCARD) {
1507 uint64_t zero =
1508 attr_meta->index |
1509 attr_meta->unknown1 |
1510 attr_meta->unknown3 |
1511 attr_meta->src_offset;
1512
1513 if (zero)
1514 pandecode_msg("XXX: expected empty record for varying discard\n");
1515
1516 /* We want to look for a literal 0000 swizzle -- this
1517 * is not encoded with all zeroes, however */
1518
1519 enum mali_channel z = MALI_CHANNEL_ZERO;
1520 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1521 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1522
1523 if (!good_swizzle)
1524 pandecode_msg("XXX: expected zero swizzle for discard\n");
1525
1526 if (!varying)
1527 pandecode_msg("XXX: cannot discard attribute\n");
1528
1529 /* If we're all good, omit the record */
1530 if (!zero && varying && good_swizzle) {
1531 pandecode_log("/* discarded varying */\n");
1532 continue;
1533 }
1534 }
1535
1536 if (attr_meta->index > max_index)
1537 max_index = attr_meta->index;
1538
1539 if (attr_meta->unknown1 != 0x2) {
1540 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1541 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1542 }
1543
1544 if (attr_meta->unknown3) {
1545 pandecode_msg("XXX: unexpected unknown3 set\n");
1546 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1547 }
1548
1549 pandecode_format_short(attr_meta->format, false);
1550 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1551
1552 if (attr_meta->src_offset)
1553 pandecode_log_cont("[%u]", attr_meta->src_offset);
1554
1555 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1556
1557 pandecode_log_cont(";\n");
1558 }
1559
1560 pandecode_log("\n");
1561
1562 return count ? (max_index + 1) : 0;
1563 }
1564
1565 /* return bits [lo, hi) of word */
1566 static u32
1567 bits(u32 word, u32 lo, u32 hi)
1568 {
1569 if (hi - lo >= 32)
1570 return word; // avoid undefined behavior with the shift
1571
1572 return (word >> lo) & ((1 << (hi - lo)) - 1);
1573 }
1574
1575 static void
1576 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool noninstanced)
1577 {
1578 pandecode_log_cont("{\n");
1579 pandecode_indent++;
1580
1581 /* Decode invocation_count. See the comment before the definition of
1582 * invocation_count for an explanation.
1583 */
1584
1585 unsigned size_x = bits(p->invocation_count, 0, p->size_y_shift) + 1;
1586 unsigned size_y = bits(p->invocation_count, p->size_y_shift, p->size_z_shift) + 1;
1587 unsigned size_z = bits(p->invocation_count, p->size_z_shift, p->workgroups_x_shift) + 1;
1588
1589 unsigned groups_x = bits(p->invocation_count, p->workgroups_x_shift, p->workgroups_y_shift) + 1;
1590 unsigned groups_y = bits(p->invocation_count, p->workgroups_y_shift, p->workgroups_z_shift) + 1;
1591 unsigned groups_z = bits(p->invocation_count, p->workgroups_z_shift, 32) + 1;
1592
1593 /* Even though we have this decoded, we want to ensure that the
1594 * representation is "unique" so we don't lose anything by printing only
1595 * the final result. More specifically, we need to check that we were
1596 * passed something in canonical form, since the definition per the
1597 * hardware is inherently not unique. How? Well, take the resulting
1598 * decode and pack it ourselves! If it is bit exact with what we
1599 * decoded, we're good to go. */
1600
1601 struct mali_vertex_tiler_prefix ref;
1602 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, noninstanced);
1603
1604 bool canonical =
1605 (p->invocation_count == ref.invocation_count) &&
1606 (p->size_y_shift == ref.size_y_shift) &&
1607 (p->size_z_shift == ref.size_z_shift) &&
1608 (p->workgroups_x_shift == ref.workgroups_x_shift) &&
1609 (p->workgroups_y_shift == ref.workgroups_y_shift) &&
1610 (p->workgroups_z_shift == ref.workgroups_z_shift) &&
1611 (p->workgroups_x_shift_2 == ref.workgroups_x_shift_2);
1612
1613 if (!canonical) {
1614 pandecode_msg("XXX: non-canonical workgroups packing\n");
1615 pandecode_msg("expected: %X, %d, %d, %d, %d, %d\n",
1616 ref.invocation_count,
1617 ref.size_y_shift,
1618 ref.size_z_shift,
1619 ref.workgroups_x_shift,
1620 ref.workgroups_y_shift,
1621 ref.workgroups_z_shift,
1622 ref.workgroups_x_shift_2);
1623
1624 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1625 pandecode_prop("size_y_shift = %d", p->size_y_shift);
1626 pandecode_prop("size_z_shift = %d", p->size_z_shift);
1627 pandecode_prop("workgroups_x_shift = %d", p->workgroups_x_shift);
1628 pandecode_prop("workgroups_y_shift = %d", p->workgroups_y_shift);
1629 pandecode_prop("workgroups_z_shift = %d", p->workgroups_z_shift);
1630 pandecode_prop("workgroups_x_shift_2 = %d", p->workgroups_x_shift_2);
1631 }
1632
1633 /* Regardless, print the decode */
1634 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1635 size_x, size_y, size_z,
1636 groups_x, groups_y, groups_z);
1637
1638 /* TODO: Decode */
1639 if (p->unknown_draw)
1640 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1641
1642 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1643
1644 if (p->draw_mode != MALI_DRAW_NONE)
1645 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
1646
1647 /* Index count only exists for tiler jobs anyway */
1648
1649 if (p->index_count)
1650 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1651
1652
1653 unsigned index_raw_size = (p->unknown_draw & MALI_DRAW_INDEXED_SIZE);
1654 index_raw_size >>= MALI_DRAW_INDEXED_SHIFT;
1655
1656 /* Validate an index buffer is present if we need one. TODO: verify
1657 * relationship between invocation_count and index_count */
1658
1659 if (p->indices) {
1660 unsigned count = p->index_count;
1661
1662 /* Grab the size */
1663 unsigned size = (index_raw_size == 0x3) ? 4 : index_raw_size;
1664
1665 /* Ensure we got a size, and if so, validate the index buffer
1666 * is large enough to hold a full set of indices of the given
1667 * size */
1668
1669 if (!index_raw_size)
1670 pandecode_msg("XXX: index size missing\n");
1671 else
1672 pandecode_validate_buffer(p->indices, count * size);
1673 } else if (index_raw_size)
1674 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size);
1675
1676 if (p->offset_bias_correction)
1677 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1678
1679 /* TODO: Figure out what this is. It's not zero */
1680 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1681
1682 pandecode_indent--;
1683 pandecode_log("},\n");
1684 }
1685
1686 static void
1687 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1688 {
1689 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1690 struct mali_uniform_buffer_meta *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1691
1692 for (int i = 0; i < ubufs_count; i++) {
1693 unsigned size = (ubufs[i].size + 1) * 16;
1694 mali_ptr addr = ubufs[i].ptr << 2;
1695
1696 pandecode_validate_buffer(addr, size);
1697
1698 char *ptr = pointer_as_memory_reference(ubufs[i].ptr << 2);
1699 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
1700 free(ptr);
1701 }
1702
1703 pandecode_log("\n");
1704 }
1705
1706 static void
1707 pandecode_scratchpad(uintptr_t pscratchpad, int job_no, char *suffix)
1708 {
1709
1710 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(pscratchpad);
1711
1712 struct bifrost_scratchpad *PANDECODE_PTR_VAR(scratchpad, mem, pscratchpad);
1713
1714 if (scratchpad->zero) {
1715 pandecode_msg("XXX: scratchpad zero tripped");
1716 pandecode_prop("zero = 0x%x\n", scratchpad->zero);
1717 }
1718
1719 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64"_%d%s = {\n", pscratchpad, job_no, suffix);
1720 pandecode_indent++;
1721
1722 pandecode_prop("flags = 0x%x", scratchpad->flags);
1723 MEMORY_PROP(scratchpad, gpu_scratchpad);
1724
1725 pandecode_indent--;
1726 pandecode_log("};\n");
1727 }
1728
1729 static unsigned shader_id = 0;
1730
1731 static struct midgard_disasm_stats
1732 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1733 bool is_bifrost)
1734 {
1735 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1736 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1737
1738 /* Compute maximum possible size */
1739 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1740
1741 /* Print some boilerplate to clearly denote the assembly (which doesn't
1742 * obey indentation rules), and actually do the disassembly! */
1743
1744 printf("\n\n");
1745
1746 struct midgard_disasm_stats stats;
1747
1748 if (is_bifrost) {
1749 disassemble_bifrost(code, sz, false);
1750
1751 /* TODO: Extend stats to Bifrost */
1752 stats.texture_count = -128;
1753 stats.sampler_count = -128;
1754 stats.attribute_count = -128;
1755 stats.varying_count = -128;
1756 stats.uniform_count = -128;
1757 stats.uniform_buffer_count = -128;
1758 stats.work_count = -128;
1759
1760 stats.instruction_count = 0;
1761 stats.bundle_count = 0;
1762 stats.quadword_count = 0;
1763 } else {
1764 stats = disassemble_midgard(code, sz);
1765 }
1766
1767 /* Print shader-db stats */
1768
1769 unsigned nr_threads =
1770 (stats.work_count <= 4) ? 4 :
1771 (stats.work_count <= 8) ? 2 :
1772 1;
1773
1774 printf("shader%d - %s shader: "
1775 "%u inst, %u bundles, %u quadwords, "
1776 "%u registers, %u threads, 0 loops\n\n\n",
1777 shader_id++,
1778 (type == JOB_TYPE_TILER) ? "FRAGMENT" : "VERTEX",
1779 stats.instruction_count, stats.bundle_count, stats.quadword_count,
1780 stats.work_count, nr_threads);
1781
1782
1783 return stats;
1784 }
1785
1786 static void
1787 pandecode_texture(mali_ptr u,
1788 struct pandecode_mapped_memory *tmem,
1789 unsigned job_no, unsigned tex)
1790 {
1791 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, u);
1792
1793 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", u, job_no, tex);
1794 pandecode_indent++;
1795
1796 struct mali_texture_format f = t->format;
1797
1798 /* See the definiton of enum mali_texture_type */
1799
1800 bool is_cube = f.type == MALI_TEX_CUBE;
1801 unsigned dimension = is_cube ? 2 : f.type;
1802
1803 pandecode_make_indent();
1804
1805 /* TODO: Are there others? */
1806 bool is_zs = f.format == MALI_Z32_UNORM;
1807
1808 /* Recall Z/S switched the meaning of linear/tiled .. */
1809 if (is_zs && f.layout == MALI_TEXTURE_LINEAR)
1810 pandecode_msg("XXX: depth/stencil cannot be tiled\n");
1811
1812 /* Print the layout. Default is linear; a modifier can denote AFBC or
1813 * u-interleaved/tiled modes */
1814
1815 if (f.layout == MALI_TEXTURE_AFBC)
1816 pandecode_log_cont("afbc");
1817 else if (f.layout == MALI_TEXTURE_TILED)
1818 pandecode_log_cont(is_zs ? "linear" : "tiled");
1819 else if (f.layout == MALI_TEXTURE_LINEAR)
1820 pandecode_log_cont("linear");
1821 else
1822 pandecode_msg("XXX: invalid texture layout 0x%X\n", f.layout);
1823
1824 pandecode_swizzle(t->swizzle, f.format);
1825 pandecode_log_cont(" ");
1826
1827 /* Distinguish cube/2D with modifier */
1828
1829 if (is_cube)
1830 pandecode_log_cont("cube ");
1831
1832 pandecode_format_short(f.format, f.srgb);
1833 pandecode_swizzle(f.swizzle, f.format);
1834
1835 /* All four width/height/depth/array_size dimensions are present
1836 * regardless of the type of texture, but it is an error to have
1837 * non-zero dimensions for unused dimensions. Verify this. array_size
1838 * can always be set, as can width. */
1839
1840 if (t->height && dimension < 2)
1841 pandecode_msg("XXX: nonzero height for <2D texture\n");
1842
1843 if (t->depth && dimension < 3)
1844 pandecode_msg("XXX: nonzero depth for <2D texture\n");
1845
1846 /* Print only the dimensions that are actually there */
1847
1848 pandecode_log_cont(": %d", t->width + 1);
1849
1850 if (dimension >= 2)
1851 pandecode_log_cont("x%u", t->height + 1);
1852
1853 if (dimension >= 3)
1854 pandecode_log_cont("x%u", t->depth + 1);
1855
1856 if (t->array_size)
1857 pandecode_log_cont("[%u]", t->array_size + 1);
1858
1859 if (t->levels)
1860 pandecode_log_cont(" mip %u", t->levels);
1861
1862 pandecode_log_cont("\n");
1863
1864 if (f.unknown1 | f.zero) {
1865 pandecode_msg("XXX: texture format zero tripped\n");
1866 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
1867 pandecode_prop("zero = %" PRId32, f.zero);
1868 }
1869
1870 if (!f.unknown2) {
1871 pandecode_msg("XXX: expected unknown texture bit set\n");
1872 pandecode_prop("unknown2 = %" PRId32, f.unknown1);
1873 }
1874
1875 if (t->swizzle_zero) {
1876 pandecode_msg("XXX: swizzle zero tripped\n");
1877 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
1878 }
1879
1880 if (t->unknown3 | t->unknown3A | t->unknown5 | t->unknown6 | t->unknown7) {
1881 pandecode_msg("XXX: texture zero tripped\n");
1882 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
1883 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
1884 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
1885 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
1886 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
1887 }
1888
1889 pandecode_log(".payload = {\n");
1890 pandecode_indent++;
1891
1892 /* A bunch of bitmap pointers follow.
1893 * We work out the correct number,
1894 * based on the mipmap/cubemap
1895 * properties, but dump extra
1896 * possibilities to futureproof */
1897
1898 int bitmap_count = MALI_NEGATIVE(t->levels);
1899
1900 /* Miptree for each face */
1901 if (f.type == MALI_TEX_CUBE)
1902 bitmap_count *= 6;
1903
1904 /* Array of textures */
1905 bitmap_count *= MALI_NEGATIVE(t->array_size);
1906
1907 /* Stride for each element */
1908 if (f.manual_stride)
1909 bitmap_count *= 2;
1910
1911 /* Sanity check the size */
1912 int max_count = sizeof(t->payload) / sizeof(t->payload[0]);
1913 assert (bitmap_count <= max_count);
1914
1915 for (int i = 0; i < bitmap_count; ++i) {
1916 /* How we dump depends if this is a stride or a pointer */
1917
1918 if (f.manual_stride && (i & 1)) {
1919 /* signed 32-bit snuck in as a 64-bit pointer */
1920 uint64_t stride_set = t->payload[i];
1921 uint32_t clamped_stride = stride_set;
1922 int32_t stride = clamped_stride;
1923 assert(stride_set == clamped_stride);
1924 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
1925 } else {
1926 char *a = pointer_as_memory_reference(t->payload[i]);
1927 pandecode_log("%s, \n", a);
1928 free(a);
1929 }
1930 }
1931
1932 pandecode_indent--;
1933 pandecode_log("},\n");
1934
1935 pandecode_indent--;
1936 pandecode_log("};\n");
1937 }
1938
1939 /* For shader properties like texture_count, we have a claimed property in the shader_meta, and the actual Truth from static analysis (this may just be an upper limit). We validate accordingly */
1940
1941 static void
1942 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
1943 {
1944 /* Nothing to do */
1945 if (claim == truth)
1946 return;
1947
1948 if (fuzzy)
1949 assert(truth >= 0);
1950
1951 if ((truth >= 0) && !fuzzy) {
1952 pandecode_msg("%s: expected %s = %d, claimed %u\n",
1953 (truth < claim) ? "warn" : "XXX",
1954 name, truth, claim);
1955 } else if ((claim > -truth) && !fuzzy) {
1956 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
1957 name, -truth, claim);
1958 } else if (fuzzy && (claim < truth))
1959 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
1960 name, truth, claim);
1961
1962 pandecode_log(".%s = %" PRId16, name, claim);
1963
1964 if (fuzzy)
1965 pandecode_log_cont(" /* %u used */", truth);
1966
1967 pandecode_log_cont(",\n");
1968 }
1969
1970 static void
1971 pandecode_vertex_tiler_postfix_pre(
1972 const struct mali_vertex_tiler_postfix *p,
1973 int job_no, enum mali_job_type job_type,
1974 char *suffix, bool is_bifrost)
1975 {
1976 mali_ptr shader_meta_ptr = (u64) (uintptr_t) (p->_shader_upper << 4);
1977 struct pandecode_mapped_memory *attr_mem;
1978
1979 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1980 * are the only things actually needed from the FBD, vertex/tiler jobs
1981 * no longer reference the FBD -- instead, this field points to some
1982 * info about the scratchpad.
1983 */
1984
1985 struct pandecode_fbd fbd_info = {
1986 /* Default for Bifrost */
1987 .rt_count = 1
1988 };
1989
1990 if (is_bifrost)
1991 pandecode_scratchpad(p->framebuffer & ~FBD_TYPE, job_no, suffix);
1992 else if (p->framebuffer & MALI_MFBD)
1993 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->framebuffer) & FBD_MASK, job_no, false);
1994 else if (job_type == JOB_TYPE_COMPUTE)
1995 pandecode_compute_fbd((u64) (uintptr_t) p->framebuffer, job_no);
1996 else
1997 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->framebuffer, job_no, false);
1998
1999 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
2000 int texture_count = 0, sampler_count = 0;
2001
2002 if (shader_meta_ptr) {
2003 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(shader_meta_ptr);
2004 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, shader_meta_ptr);
2005
2006 /* Disassemble ahead-of-time to get stats. Initialize with
2007 * stats for the missing-shader case so we get validation
2008 * there, too */
2009
2010 struct midgard_disasm_stats info = {
2011 .texture_count = 0,
2012 .sampler_count = 0,
2013 .attribute_count = 0,
2014 .varying_count = 0,
2015 .work_count = 1,
2016
2017 .uniform_count = -128,
2018 .uniform_buffer_count = 0
2019 };
2020
2021 if (s->shader & ~0xF)
2022 info = pandecode_shader_disassemble(s->shader & ~0xF, job_no, job_type, is_bifrost);
2023
2024 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", shader_meta_ptr, job_no, suffix);
2025 pandecode_indent++;
2026
2027 /* Save for dumps */
2028 attribute_count = s->attribute_count;
2029 varying_count = s->varying_count;
2030 texture_count = s->texture_count;
2031 sampler_count = s->sampler_count;
2032
2033 if (is_bifrost) {
2034 uniform_count = s->bifrost2.uniform_count;
2035 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
2036 } else {
2037 uniform_count = s->midgard1.uniform_count;
2038 uniform_buffer_count = s->midgard1.uniform_buffer_count;
2039 }
2040
2041 pandecode_shader_address("shader", s->shader);
2042
2043 pandecode_shader_prop("texture_count", s->texture_count, info.texture_count, false);
2044 pandecode_shader_prop("sampler_count", s->sampler_count, info.sampler_count, false);
2045 pandecode_shader_prop("attribute_count", s->attribute_count, info.attribute_count, false);
2046 pandecode_shader_prop("varying_count", s->varying_count, info.varying_count, false);
2047 pandecode_shader_prop("uniform_buffer_count",
2048 uniform_buffer_count,
2049 info.uniform_buffer_count, true);
2050
2051 if (!is_bifrost) {
2052 pandecode_shader_prop("uniform_count",
2053 uniform_count,
2054 info.uniform_count, false);
2055
2056 pandecode_shader_prop("work_count",
2057 s->midgard1.work_count, info.work_count, false);
2058 }
2059
2060 if (is_bifrost) {
2061 pandecode_prop("bifrost1.unk1 = 0x%" PRIx32, s->bifrost1.unk1);
2062 } else {
2063 pandecode_log(".midgard1.flags = ");
2064 pandecode_log_decoded_flags(shader_midgard1_flag_info, s->midgard1.flags);
2065 pandecode_log_cont(",\n");
2066
2067 pandecode_prop("midgard1.unknown2 = 0x%" PRIx32, s->midgard1.unknown2);
2068 }
2069
2070 if (s->depth_units || s->depth_factor) {
2071 pandecode_prop("depth_factor = %f", s->depth_factor);
2072 pandecode_prop("depth_units = %f", s->depth_units);
2073 }
2074
2075 if (s->alpha_coverage) {
2076 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
2077 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
2078
2079 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
2080 invert_alpha_coverage ? "~" : "",
2081 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
2082 }
2083
2084 if (s->unknown2_3 || s->unknown2_4) {
2085 pandecode_log(".unknown2_3 = ");
2086
2087 int unknown2_3 = s->unknown2_3;
2088 int unknown2_4 = s->unknown2_4;
2089
2090 /* We're not quite sure what these flags mean without the depth test, if anything */
2091
2092 if (unknown2_3 & (MALI_DEPTH_TEST | MALI_DEPTH_FUNC_MASK)) {
2093 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
2094 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
2095
2096 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
2097 }
2098
2099 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
2100 pandecode_log_cont(",\n");
2101
2102 pandecode_log(".unknown2_4 = ");
2103 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
2104 pandecode_log_cont(",\n");
2105 }
2106
2107 if (s->stencil_mask_front || s->stencil_mask_back) {
2108 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
2109 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
2110 }
2111
2112 pandecode_stencil("front", &s->stencil_front);
2113 pandecode_stencil("back", &s->stencil_back);
2114
2115 if (is_bifrost) {
2116 pandecode_log(".bifrost2 = {\n");
2117 pandecode_indent++;
2118
2119 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
2120 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
2121 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
2122 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
2123
2124 pandecode_indent--;
2125 pandecode_log("},\n");
2126 } else if (s->midgard2.unknown2_7) {
2127 pandecode_log(".midgard2 = {\n");
2128 pandecode_indent++;
2129
2130 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
2131 pandecode_indent--;
2132 pandecode_log("},\n");
2133 }
2134
2135 if (s->unknown2_8)
2136 pandecode_prop("unknown2_8 = 0x%" PRIx32, s->unknown2_8);
2137
2138 if (!is_bifrost) {
2139 /* TODO: Blend shaders routing/disasm */
2140
2141 union midgard_blend blend = s->blend;
2142 pandecode_midgard_blend(&blend, false);
2143 }
2144
2145 pandecode_indent--;
2146 pandecode_log("};\n");
2147
2148 /* MRT blend fields are used whenever MFBD is used, with
2149 * per-RT descriptors */
2150
2151 if (job_type == JOB_TYPE_TILER) {
2152 void* blend_base = (void *) (s + 1);
2153
2154 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
2155 mali_ptr shader = 0;
2156
2157 if (is_bifrost)
2158 shader = pandecode_bifrost_blend(blend_base, job_no, i);
2159 else
2160 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
2161
2162 if (shader & ~0xF) {
2163 struct midgard_disasm_stats stats =
2164 pandecode_shader_disassemble(shader, job_no, job_type, false);
2165
2166 bool has_texture = (stats.texture_count > 0);
2167 bool has_sampler = (stats.sampler_count > 0);
2168 bool has_attribute = (stats.attribute_count > 0);
2169 bool has_varying = (stats.varying_count > 0);
2170 bool has_uniform = (stats.uniform_count > 0);
2171 bool has_ubo = (stats.uniform_buffer_count > 0);
2172
2173 if (has_texture || has_sampler)
2174 pandecode_msg("XXX: blend shader accessing textures\n");
2175
2176 if (has_attribute || has_varying)
2177 pandecode_msg("XXX: blend shader accessing interstage\n");
2178
2179 if (has_uniform || has_ubo)
2180 pandecode_msg("XXX: blend shader accessing uniforms\n");
2181 }
2182
2183 }
2184 }
2185 } else
2186 pandecode_msg("XXX: missing shader descriptor\n");
2187
2188 if (p->viewport) {
2189 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
2190 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
2191
2192 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
2193 pandecode_indent++;
2194
2195 pandecode_prop("clip_minx = %f", f->clip_minx);
2196 pandecode_prop("clip_miny = %f", f->clip_miny);
2197 pandecode_prop("clip_minz = %f", f->clip_minz);
2198 pandecode_prop("clip_maxx = %f", f->clip_maxx);
2199 pandecode_prop("clip_maxy = %f", f->clip_maxy);
2200 pandecode_prop("clip_maxz = %f", f->clip_maxz);
2201
2202 /* Only the higher coordinates are MALI_POSITIVE scaled */
2203
2204 pandecode_prop("viewport0 = { %d, %d }",
2205 f->viewport0[0], f->viewport0[1]);
2206
2207 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2208 f->viewport1[0] + 1, f->viewport1[1] + 1);
2209
2210 pandecode_indent--;
2211 pandecode_log("};\n");
2212 }
2213
2214 if (p->attribute_meta) {
2215 unsigned max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
2216
2217 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
2218 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false);
2219 }
2220
2221 /* Varyings are encoded like attributes but not actually sent; we just
2222 * pass a zero buffer with the right stride/size set, (or whatever)
2223 * since the GPU will write to it itself */
2224
2225 if (p->varying_meta) {
2226 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
2227 }
2228
2229 if (p->varyings) {
2230 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
2231
2232 /* Number of descriptors depends on whether there are
2233 * non-internal varyings */
2234
2235 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true);
2236 }
2237
2238 if (p->uniform_buffers) {
2239 if (uniform_buffer_count)
2240 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
2241 else
2242 pandecode_msg("warn: UBOs specified but not referenced\n");
2243 } else if (uniform_buffer_count)
2244 pandecode_msg("XXX: UBOs referenced but not specified\n");
2245
2246 /* We don't want to actually dump uniforms, but we do need to validate
2247 * that the counts we were given are sane */
2248
2249 if (p->uniforms) {
2250 if (uniform_count)
2251 pandecode_validate_buffer(p->uniforms, uniform_count * 16);
2252 else
2253 pandecode_msg("warn: Uniforms specified but not referenced\n");
2254 } else if (uniform_count)
2255 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2256
2257 if (p->texture_trampoline) {
2258 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(p->texture_trampoline);
2259
2260 if (mmem) {
2261 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline);
2262
2263 pandecode_log("uint64_t texture_trampoline_%"PRIx64"_%d[] = {\n", p->texture_trampoline, job_no);
2264 pandecode_indent++;
2265
2266 for (int tex = 0; tex < texture_count; ++tex) {
2267 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2268 char *a = pointer_as_memory_reference(*u);
2269 pandecode_log("%s,\n", a);
2270 free(a);
2271 }
2272
2273 pandecode_indent--;
2274 pandecode_log("};\n");
2275
2276 /* Now, finally, descend down into the texture descriptor */
2277 for (unsigned tex = 0; tex < texture_count; ++tex) {
2278 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2279 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2280 if (tmem)
2281 pandecode_texture(*u, tmem, job_no, tex);
2282 }
2283 }
2284 }
2285
2286 if (p->sampler_descriptor) {
2287 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->sampler_descriptor);
2288
2289 if (smem) {
2290 struct mali_sampler_descriptor *s;
2291
2292 mali_ptr d = p->sampler_descriptor;
2293
2294 for (int i = 0; i < sampler_count; ++i) {
2295 s = pandecode_fetch_gpu_mem(smem, d + sizeof(*s) * i, sizeof(*s));
2296
2297 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", d + sizeof(*s) * i, job_no, i);
2298 pandecode_indent++;
2299
2300 pandecode_log(".filter_mode = ");
2301 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2302 pandecode_log_cont(",\n");
2303
2304 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2305 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2306
2307 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2308 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2309 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2310
2311 pandecode_prop("compare_func = %s", pandecode_alt_func(s->compare_func));
2312
2313 if (s->zero || s->zero2) {
2314 pandecode_msg("XXX: sampler zero tripped\n");
2315 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2316 }
2317
2318 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2319
2320 pandecode_prop("border_color = { %f, %f, %f, %f }",
2321 s->border_color[0],
2322 s->border_color[1],
2323 s->border_color[2],
2324 s->border_color[3]);
2325
2326 pandecode_indent--;
2327 pandecode_log("};\n");
2328 }
2329 }
2330 }
2331 }
2332
2333 static void
2334 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2335 {
2336 if (!(p->position_varying || p->occlusion_counter || p->flags))
2337 return;
2338
2339 pandecode_log(".postfix = {\n");
2340 pandecode_indent++;
2341
2342 MEMORY_PROP(p, position_varying);
2343 MEMORY_PROP(p, occlusion_counter);
2344
2345 if (p->flags)
2346 pandecode_prop("flags = %d", p->flags);
2347
2348 pandecode_indent--;
2349 pandecode_log("},\n");
2350 }
2351
2352 static void
2353 pandecode_vertex_only_bfr(struct bifrost_vertex_only *v)
2354 {
2355 pandecode_log_cont("{\n");
2356 pandecode_indent++;
2357
2358 pandecode_prop("unk2 = 0x%x", v->unk2);
2359
2360 if (v->zero0 || v->zero1) {
2361 pandecode_msg("XXX: vertex only zero tripped");
2362 pandecode_prop("zero0 = 0x%" PRIx32, v->zero0);
2363 pandecode_prop("zero1 = 0x%" PRIx64, v->zero1);
2364 }
2365
2366 pandecode_indent--;
2367 pandecode_log("}\n");
2368 }
2369
2370 static void
2371 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2372 {
2373
2374 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2375 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2376
2377 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no);
2378 pandecode_indent++;
2379
2380 if (h->zero) {
2381 pandecode_msg("XXX: tiler heap zero tripped\n");
2382 pandecode_prop("zero = 0x%x", h->zero);
2383 }
2384
2385 for (int i = 0; i < 12; i++) {
2386 if (h->zeros[i] != 0) {
2387 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2388 i, h->zeros[i]);
2389 }
2390 }
2391
2392 pandecode_prop("heap_size = 0x%x", h->heap_size);
2393 MEMORY_PROP(h, tiler_heap_start);
2394 MEMORY_PROP(h, tiler_heap_free);
2395
2396 /* this might point to the beginning of another buffer, when it's
2397 * really the end of the tiler heap buffer, so we have to be careful
2398 * here. but for zero length, we need the same pointer.
2399 */
2400
2401 if (h->tiler_heap_end == h->tiler_heap_start) {
2402 MEMORY_PROP(h, tiler_heap_start);
2403 } else {
2404 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2405 pandecode_prop("tiler_heap_end = %s + 1", a);
2406 free(a);
2407 }
2408
2409 pandecode_indent--;
2410 pandecode_log("};\n");
2411 }
2412
2413 static void
2414 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2415 {
2416 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2417 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2418
2419 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2420
2421 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no);
2422 pandecode_indent++;
2423
2424 if (t->zero0 || t->zero1) {
2425 pandecode_msg("XXX: tiler meta zero tripped\n");
2426 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2427 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2428 }
2429
2430 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2431 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2432
2433 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2434 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2435
2436 for (int i = 0; i < 12; i++) {
2437 if (t->zeros[i] != 0) {
2438 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2439 i, t->zeros[i]);
2440 }
2441 }
2442
2443 pandecode_indent--;
2444 pandecode_log("};\n");
2445 }
2446
2447 static void
2448 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2449 {
2450 pandecode_log(".gl_enables = ");
2451
2452 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2453
2454 pandecode_log_cont(",\n");
2455 }
2456
2457 static void
2458 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2459 {
2460 if (u.pointer == 0x0)
2461 return;
2462
2463 pandecode_log(".primitive_size = {\n");
2464 pandecode_indent++;
2465
2466 if (constant) {
2467 pandecode_prop("constant = %f", u.constant);
2468 } else {
2469 MEMORY_PROP((&u), pointer);
2470 }
2471
2472 pandecode_indent--;
2473 pandecode_log("},\n");
2474 }
2475
2476 static void
2477 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2478 {
2479 pandecode_log_cont("{\n");
2480 pandecode_indent++;
2481
2482 /* TODO: gl_PointSize on Bifrost */
2483 pandecode_primitive_size(t->primitive_size, true);
2484
2485 pandecode_gl_enables(t->gl_enables, JOB_TYPE_TILER);
2486
2487 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2488 || t->zero6 || t->zero7 || t->zero8) {
2489 pandecode_msg("XXX: tiler only zero tripped\n");
2490 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2491 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2492 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2493 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2494 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2495 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2496 pandecode_prop("zero7 = 0x%" PRIx32, t->zero7);
2497 pandecode_prop("zero8 = 0x%" PRIx64, t->zero8);
2498 }
2499
2500 pandecode_indent--;
2501 pandecode_log("},\n");
2502 }
2503
2504 static int
2505 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2506 const struct pandecode_mapped_memory *mem,
2507 mali_ptr payload, int job_no)
2508 {
2509 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2510
2511 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true);
2512
2513 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no);
2514 pandecode_indent++;
2515
2516 pandecode_log(".prefix = ");
2517 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2518
2519 pandecode_log(".vertex = ");
2520 pandecode_vertex_only_bfr(&v->vertex);
2521
2522 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2523
2524 pandecode_indent--;
2525 pandecode_log("};\n");
2526
2527 return sizeof(*v);
2528 }
2529
2530 static int
2531 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2532 const struct pandecode_mapped_memory *mem,
2533 mali_ptr payload, int job_no)
2534 {
2535 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2536
2537 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true);
2538 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2539
2540 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no);
2541 pandecode_indent++;
2542
2543 pandecode_log(".prefix = ");
2544 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
2545
2546 pandecode_log(".tiler = ");
2547 pandecode_tiler_only_bfr(&t->tiler, job_no);
2548
2549 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2550
2551 pandecode_indent--;
2552 pandecode_log("};\n");
2553
2554 return sizeof(*t);
2555 }
2556
2557 static int
2558 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2559 const struct pandecode_mapped_memory *mem,
2560 mali_ptr payload, int job_no)
2561 {
2562 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2563
2564 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false);
2565
2566 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2567 pandecode_indent++;
2568
2569 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2570 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
2571
2572 bool instanced = v->instance_shift || v->instance_odd;
2573 bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER);
2574
2575 pandecode_log(".prefix = ");
2576 pandecode_vertex_tiler_prefix(&v->prefix, job_no, !instanced && is_graphics);
2577
2578 pandecode_gl_enables(v->gl_enables, h->job_type);
2579
2580 if (v->instance_shift || v->instance_odd) {
2581 pandecode_prop("instance_shift = 0x%d /* %d */",
2582 v->instance_shift, 1 << v->instance_shift);
2583 pandecode_prop("instance_odd = 0x%X /* %d */",
2584 v->instance_odd, (2 * v->instance_odd) + 1);
2585
2586 pandecode_padded_vertices(v->instance_shift, v->instance_odd);
2587 }
2588
2589 if (v->offset_start)
2590 pandecode_prop("offset_start = %d", v->offset_start);
2591
2592 if (v->zero5) {
2593 pandecode_msg("XXX: midgard payload zero tripped\n");
2594 pandecode_prop("zero5 = 0x%" PRIx64, v->zero5);
2595 }
2596
2597 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2598
2599 pandecode_indent--;
2600 pandecode_log("};\n");
2601
2602 return sizeof(*v);
2603 }
2604
2605 static int
2606 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2607 mali_ptr payload, int job_no,
2608 bool is_bifrost)
2609 {
2610 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2611
2612 bool is_mfbd = (s->framebuffer & FBD_TYPE) == MALI_MFBD;
2613
2614 /* Bifrost theoretically may retain support for SFBD on compute jobs,
2615 * but for graphics workloads with a FRAGMENT payload, use MFBD */
2616
2617 if (!is_mfbd && is_bifrost)
2618 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2619
2620 struct pandecode_fbd info;
2621
2622 if (is_mfbd)
2623 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true);
2624 else
2625 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true);
2626
2627 /* Compute the tag for the tagged pointer. This contains the type of
2628 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2629 * additional structures follow the MFBD header (an extra payload or
2630 * not, as well as a count of render targets) */
2631
2632 unsigned expected_tag = is_mfbd ? MALI_MFBD : MALI_SFBD;
2633
2634 if (is_mfbd) {
2635 if (info.has_extra)
2636 expected_tag |= MALI_MFBD_TAG_EXTRA;
2637
2638 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
2639 }
2640
2641 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
2642 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2643 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
2644 pandecode_prop("max_tile_coord = 0x%X\n", s->min_tile_coord);
2645 }
2646
2647 /* Extract tile coordinates */
2648
2649 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
2650 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
2651
2652 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2653 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2654
2655 /* For the max, we also want the floored (rather than ceiled) version for checking */
2656
2657 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
2658 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
2659
2660 /* Validate the coordinates are well-ordered */
2661
2662 if (min_x == max_x)
2663 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2664 else if (min_x > max_x)
2665 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2666
2667 if (min_y == max_y)
2668 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2669 else if (min_y > max_y)
2670 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2671
2672 /* Validate the coordinates fit inside the framebuffer. We use floor,
2673 * rather than ceil, for the max coordinates, since the tile
2674 * coordinates for something like an 800x600 framebuffer will actually
2675 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2676
2677 if ((min_x > info.width) || (max_x_f > info.width))
2678 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2679
2680 if ((min_y > info.height) || (max_y_f > info.height))
2681 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2682
2683 /* After validation, we print */
2684
2685 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
2686
2687 /* The FBD is a tagged pointer */
2688
2689 unsigned tag = (s->framebuffer & ~FBD_MASK);
2690
2691 if (tag != expected_tag)
2692 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
2693
2694 return sizeof(*s);
2695 }
2696
2697 static int job_descriptor_number = 0;
2698
2699 int
2700 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost)
2701 {
2702 struct mali_job_descriptor_header *h;
2703
2704 int start_number = 0;
2705
2706 bool first = true;
2707 bool last_size;
2708
2709 do {
2710 struct pandecode_mapped_memory *mem =
2711 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2712
2713 void *payload;
2714
2715 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2716
2717 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2718 * high 32-bits of the 64-bit pointer are reused to store
2719 * something else.
2720 */
2721 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2722 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
2723 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2724
2725 payload = pandecode_fetch_gpu_mem(mem, payload_ptr,
2726 MALI_PAYLOAD_SIZE);
2727
2728 int job_no = job_descriptor_number++;
2729
2730 if (first)
2731 start_number = job_no;
2732
2733 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2734 pandecode_indent++;
2735
2736 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
2737
2738 /* Save for next job fixing */
2739 last_size = h->job_descriptor_size;
2740
2741 if (h->job_descriptor_size)
2742 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2743
2744 if (h->exception_status && h->exception_status != 0x1)
2745 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2746 h->exception_status,
2747 (h->exception_status >> 16) & 0xFFFF,
2748 pandecode_exception_access((h->exception_status >> 8) & 0x3),
2749 h->exception_status & 0xFF);
2750
2751 if (h->first_incomplete_task)
2752 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2753
2754 if (h->fault_pointer)
2755 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2756
2757 if (h->job_barrier)
2758 pandecode_prop("job_barrier = %d", h->job_barrier);
2759
2760 pandecode_prop("job_index = %d", h->job_index);
2761
2762 if (h->unknown_flags)
2763 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2764
2765 if (h->job_dependency_index_1)
2766 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2767
2768 if (h->job_dependency_index_2)
2769 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2770
2771 pandecode_indent--;
2772 pandecode_log("};\n");
2773
2774 /* Do not touch the field yet -- decode the payload first, and
2775 * don't touch that either. This is essential for the uploads
2776 * to occur in sequence and therefore be dynamically allocated
2777 * correctly. Do note the size, however, for that related
2778 * reason. */
2779
2780 switch (h->job_type) {
2781 case JOB_TYPE_SET_VALUE: {
2782 struct mali_payload_set_value *s = payload;
2783 pandecode_log("struct mali_payload_set_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2784 pandecode_indent++;
2785 MEMORY_PROP(s, out);
2786 pandecode_prop("unknown = 0x%" PRIX64, s->unknown);
2787 pandecode_indent--;
2788 pandecode_log("};\n");
2789
2790 break;
2791 }
2792
2793 case JOB_TYPE_TILER:
2794 case JOB_TYPE_VERTEX:
2795 case JOB_TYPE_COMPUTE:
2796 if (bifrost) {
2797 if (h->job_type == JOB_TYPE_TILER)
2798 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no);
2799 else
2800 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no);
2801 } else
2802 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no);
2803
2804 break;
2805
2806 case JOB_TYPE_FRAGMENT:
2807 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost);
2808 break;
2809
2810 default:
2811 break;
2812 }
2813
2814 /* Handle linkage */
2815
2816 if (!first) {
2817 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no - 1);
2818
2819 if (last_size)
2820 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no);
2821 else
2822 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no);
2823 }
2824
2825 first = false;
2826
2827 } while ((jc_gpu_va = h->job_descriptor_size ? h->next_job_64 : h->next_job_32));
2828
2829 return start_number;
2830 }