projects / mesa.git / blob
? search:


b218595560b9743bfdb9bbc139d15ace70b9b515
[mesa.git] / src / panfrost / lib / 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 <midgard_pack.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 "midgard/disassemble.h"
38 #include "bifrost/disassemble.h"
39
40 #include "pan_encoder.h"
41
42 static void pandecode_swizzle(unsigned swizzle, enum mali_format format);
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 #define DUMP_CL(title, T, cl, indent) {\
61 fprintf(pandecode_dump_stream, "%s\n", title); \
62 struct MALI_ ## T temp; \
63 MALI_ ## T ## _unpack((const uint8_t *) cl, &temp); \
64 MALI_ ## T ## _print(pandecode_dump_stream, &temp, indent * 2); \
65 }
66
67 #define DUMP_ADDR(title, T, addr, indent) {\
68 struct pandecode_mapped_memory *mapped_mem = pandecode_find_mapped_gpu_mem_containing(addr); \
69 const uint8_t *cl = pandecode_fetch_gpu_mem(mapped_mem, addr, MALI_ ## T ## _LENGTH); \
70 DUMP_CL(title, T, cl, indent); \
71 }
72
73 FILE *pandecode_dump_stream;
74
75 /* Semantic logging type.
76 *
77 * Raw: for raw messages to be printed as is.
78 * Message: for helpful information to be commented out in replays.
79 * Property: for properties of a struct
80 *
81 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
82 */
83
84 enum pandecode_log_type {
85 PANDECODE_RAW,
86 PANDECODE_MESSAGE,
87 PANDECODE_PROPERTY
88 };
89
90 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
91 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
92 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
93
94 unsigned pandecode_indent = 0;
95
96 static void
97 pandecode_make_indent(void)
98 {
99 for (unsigned i = 0; i < pandecode_indent; ++i)
100 fprintf(pandecode_dump_stream, " ");
101 }
102
103 static void
104 pandecode_log_typed(enum pandecode_log_type type, const char *format, ...)
105 {
106 va_list ap;
107
108 pandecode_make_indent();
109
110 if (type == PANDECODE_MESSAGE)
111 fprintf(pandecode_dump_stream, "// ");
112 else if (type == PANDECODE_PROPERTY)
113 fprintf(pandecode_dump_stream, ".");
114
115 va_start(ap, format);
116 vfprintf(pandecode_dump_stream, format, ap);
117 va_end(ap);
118
119 if (type == PANDECODE_PROPERTY)
120 fprintf(pandecode_dump_stream, ",\n");
121 }
122
123 static void
124 pandecode_log_cont(const char *format, ...)
125 {
126 va_list ap;
127
128 va_start(ap, format);
129 vfprintf(pandecode_dump_stream, format, ap);
130 va_end(ap);
131 }
132
133 /* To check for memory safety issues, validates that the given pointer in GPU
134 * memory is valid, containing at least sz bytes. The goal is to eliminate
135 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
136 * overruns) by statically validating pointers.
137 */
138
139 static void
140 pandecode_validate_buffer(mali_ptr addr, size_t sz)
141 {
142 if (!addr) {
143 pandecode_msg("XXX: null pointer deref");
144 return;
145 }
146
147 /* Find a BO */
148
149 struct pandecode_mapped_memory *bo =
150 pandecode_find_mapped_gpu_mem_containing(addr);
151
152 if (!bo) {
153 pandecode_msg("XXX: invalid memory dereference\n");
154 return;
155 }
156
157 /* Bounds check */
158
159 unsigned offset = addr - bo->gpu_va;
160 unsigned total = offset + sz;
161
162 if (total > bo->length) {
163 pandecode_msg("XXX: buffer overrun. "
164 "Chunk of size %zu at offset %d in buffer of size %zu. "
165 "Overrun by %zu bytes. \n",
166 sz, offset, bo->length, total - bo->length);
167 return;
168 }
169 }
170
171 struct pandecode_flag_info {
172 u64 flag;
173 const char *name;
174 };
175
176 static void
177 pandecode_log_decoded_flags(const struct pandecode_flag_info *flag_info,
178 u64 flags)
179 {
180 bool decodable_flags_found = false;
181
182 for (int i = 0; flag_info[i].name; i++) {
183 if ((flags & flag_info[i].flag) != flag_info[i].flag)
184 continue;
185
186 if (!decodable_flags_found) {
187 decodable_flags_found = true;
188 } else {
189 pandecode_log_cont(" | ");
190 }
191
192 pandecode_log_cont("%s", flag_info[i].name);
193
194 flags &= ~flag_info[i].flag;
195 }
196
197 if (decodable_flags_found) {
198 if (flags)
199 pandecode_log_cont(" | 0x%" PRIx64, flags);
200 } else {
201 pandecode_log_cont("0x%" PRIx64, flags);
202 }
203 }
204
205 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
206 static const struct pandecode_flag_info gl_enable_flag_info[] = {
207 FLAG_INFO(OCCLUSION_QUERY),
208 FLAG_INFO(OCCLUSION_PRECISE),
209 FLAG_INFO(FRONT_CCW_TOP),
210 FLAG_INFO(CULL_FACE_FRONT),
211 FLAG_INFO(CULL_FACE_BACK),
212 {}
213 };
214 #undef FLAG_INFO
215
216 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
217 static const struct pandecode_flag_info clear_flag_info[] = {
218 FLAG_INFO(FAST),
219 FLAG_INFO(SLOW),
220 FLAG_INFO(SLOW_STENCIL),
221 {}
222 };
223 #undef FLAG_INFO
224
225 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
226 static const struct pandecode_flag_info mask_flag_info[] = {
227 FLAG_INFO(R),
228 FLAG_INFO(G),
229 FLAG_INFO(B),
230 FLAG_INFO(A),
231 {}
232 };
233 #undef FLAG_INFO
234
235 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
236 static const struct pandecode_flag_info u3_flag_info[] = {
237 FLAG_INFO(HAS_MSAA),
238 FLAG_INFO(PER_SAMPLE),
239 FLAG_INFO(CAN_DISCARD),
240 FLAG_INFO(HAS_BLEND_SHADER),
241 FLAG_INFO(DEPTH_WRITEMASK),
242 FLAG_INFO(DEPTH_CLIP_NEAR),
243 FLAG_INFO(DEPTH_CLIP_FAR),
244 {}
245 };
246
247 static const struct pandecode_flag_info u4_flag_info[] = {
248 FLAG_INFO(NO_MSAA),
249 FLAG_INFO(NO_DITHER),
250 FLAG_INFO(DEPTH_RANGE_A),
251 FLAG_INFO(DEPTH_RANGE_B),
252 FLAG_INFO(STENCIL_TEST),
253 FLAG_INFO(ALPHA_TO_COVERAGE),
254 {}
255 };
256 #undef FLAG_INFO
257
258 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
259 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
260 FLAG_INFO(SRGB),
261 {}
262 };
263 #undef FLAG_INFO
264
265 #define FLAG_INFO(flag) { MALI_AFBC_##flag, "MALI_AFBC_" #flag }
266 static const struct pandecode_flag_info afbc_fmt_flag_info[] = {
267 FLAG_INFO(YTR),
268 {}
269 };
270 #undef FLAG_INFO
271
272 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
273 static const struct pandecode_flag_info mfbd_extra_flag_hi_info[] = {
274 FLAG_INFO(PRESENT),
275 {}
276 };
277 #undef FLAG_INFO
278
279 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
280 static const struct pandecode_flag_info mfbd_extra_flag_lo_info[] = {
281 FLAG_INFO(ZS),
282 {}
283 };
284 #undef FLAG_INFO
285
286 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
287 static const struct pandecode_flag_info shader_midgard1_flag_lo_info [] = {
288 FLAG_INFO(WRITES_Z),
289 FLAG_INFO(EARLY_Z),
290 FLAG_INFO(READS_TILEBUFFER),
291 FLAG_INFO(WRITES_GLOBAL),
292 FLAG_INFO(READS_ZS),
293 {}
294 };
295
296 static const struct pandecode_flag_info shader_midgard1_flag_hi_info [] = {
297 FLAG_INFO(WRITES_S),
298 FLAG_INFO(SUPPRESS_INF_NAN),
299 {}
300 };
301 #undef FLAG_INFO
302
303 #define FLAG_INFO(flag) { MALI_BIFROST_##flag, "MALI_BIFROST_" #flag }
304 static const struct pandecode_flag_info shader_bifrost_info [] = {
305 FLAG_INFO(FULL_THREAD),
306 FLAG_INFO(EARLY_Z),
307 FLAG_INFO(FIRST_ATEST),
308 {}
309 };
310
311 #undef FLAG_INFO
312
313 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
314 static const struct pandecode_flag_info mfbd_flag_info [] = {
315 FLAG_INFO(DEPTH_WRITE),
316 FLAG_INFO(EXTRA),
317 {}
318 };
319 #undef FLAG_INFO
320
321 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
322 static const struct pandecode_flag_info sfbd_unk1_info [] = {
323 FLAG_INFO(MSAA_8),
324 FLAG_INFO(MSAA_A),
325 {}
326 };
327 #undef FLAG_INFO
328
329 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
330 static const struct pandecode_flag_info sfbd_unk2_info [] = {
331 FLAG_INFO(MSAA_B),
332 FLAG_INFO(SRGB),
333 {}
334 };
335 #undef FLAG_INFO
336
337 #define DEFINE_CASE(name) case MALI_## name: return "MALI_" #name
338 static char *pandecode_format(enum mali_format format)
339 {
340 static char unk_format_str[10];
341
342 switch (format) {
343 DEFINE_CASE(ETC2_RGB8);
344 DEFINE_CASE(ETC2_R11_UNORM);
345 DEFINE_CASE(ETC2_RGBA8);
346 DEFINE_CASE(ETC2_RG11_UNORM);
347 DEFINE_CASE(ETC2_R11_SNORM);
348 DEFINE_CASE(ETC2_RG11_SNORM);
349 DEFINE_CASE(ETC2_RGB8A1);
350 DEFINE_CASE(NXR);
351 DEFINE_CASE(BC1_UNORM);
352 DEFINE_CASE(BC2_UNORM);
353 DEFINE_CASE(BC3_UNORM);
354 DEFINE_CASE(BC4_UNORM);
355 DEFINE_CASE(BC4_SNORM);
356 DEFINE_CASE(BC5_UNORM);
357 DEFINE_CASE(BC5_SNORM);
358 DEFINE_CASE(BC6H_UF16);
359 DEFINE_CASE(BC6H_SF16);
360 DEFINE_CASE(BC7_UNORM);
361 DEFINE_CASE(ASTC_3D_LDR);
362 DEFINE_CASE(ASTC_3D_HDR);
363 DEFINE_CASE(ASTC_2D_LDR);
364 DEFINE_CASE(ASTC_2D_HDR);
365 DEFINE_CASE(RGB565);
366 DEFINE_CASE(RGB5_X1_UNORM);
367 DEFINE_CASE(RGB5_A1_UNORM);
368 DEFINE_CASE(RGB10_A2_UNORM);
369 DEFINE_CASE(RGB10_A2_SNORM);
370 DEFINE_CASE(RGB10_A2UI);
371 DEFINE_CASE(RGB10_A2I);
372 DEFINE_CASE(RGB332_UNORM);
373 DEFINE_CASE(RGB233_UNORM);
374 DEFINE_CASE(Z24X8_UNORM);
375 DEFINE_CASE(R32_FIXED);
376 DEFINE_CASE(RG32_FIXED);
377 DEFINE_CASE(RGB32_FIXED);
378 DEFINE_CASE(RGBA32_FIXED);
379 DEFINE_CASE(R11F_G11F_B10F);
380 DEFINE_CASE(R9F_G9F_B9F_E5F);
381 DEFINE_CASE(VARYING_POS);
382 DEFINE_CASE(VARYING_DISCARD);
383
384 DEFINE_CASE(R8_SNORM);
385 DEFINE_CASE(R16_SNORM);
386 DEFINE_CASE(R32_SNORM);
387 DEFINE_CASE(RG8_SNORM);
388 DEFINE_CASE(RG16_SNORM);
389 DEFINE_CASE(RG32_SNORM);
390 DEFINE_CASE(RGB8_SNORM);
391 DEFINE_CASE(RGB16_SNORM);
392 DEFINE_CASE(RGB32_SNORM);
393 DEFINE_CASE(RGBA8_SNORM);
394 DEFINE_CASE(RGBA16_SNORM);
395 DEFINE_CASE(RGBA32_SNORM);
396
397 DEFINE_CASE(R8UI);
398 DEFINE_CASE(R16UI);
399 DEFINE_CASE(R32UI);
400 DEFINE_CASE(RG8UI);
401 DEFINE_CASE(RG16UI);
402 DEFINE_CASE(RG32UI);
403 DEFINE_CASE(RGB8UI);
404 DEFINE_CASE(RGB16UI);
405 DEFINE_CASE(RGB32UI);
406 DEFINE_CASE(RGBA8UI);
407 DEFINE_CASE(RGBA16UI);
408 DEFINE_CASE(RGBA32UI);
409
410 DEFINE_CASE(R8_UNORM);
411 DEFINE_CASE(R16_UNORM);
412 DEFINE_CASE(R32_UNORM);
413 DEFINE_CASE(R32F);
414 DEFINE_CASE(RG8_UNORM);
415 DEFINE_CASE(RG16_UNORM);
416 DEFINE_CASE(RG32_UNORM);
417 DEFINE_CASE(RG32F);
418 DEFINE_CASE(RGB8_UNORM);
419 DEFINE_CASE(RGB16_UNORM);
420 DEFINE_CASE(RGB32_UNORM);
421 DEFINE_CASE(RGB32F);
422 DEFINE_CASE(RGBA4_UNORM);
423 DEFINE_CASE(RGBA8_UNORM);
424 DEFINE_CASE(RGBA16_UNORM);
425 DEFINE_CASE(RGBA32_UNORM);
426 DEFINE_CASE(RGBA32F);
427
428 DEFINE_CASE(R8I);
429 DEFINE_CASE(R16I);
430 DEFINE_CASE(R32I);
431 DEFINE_CASE(RG8I);
432 DEFINE_CASE(R16F);
433 DEFINE_CASE(RG16I);
434 DEFINE_CASE(RG32I);
435 DEFINE_CASE(RG16F);
436 DEFINE_CASE(RGB8I);
437 DEFINE_CASE(RGB16I);
438 DEFINE_CASE(RGB32I);
439 DEFINE_CASE(RGB16F);
440 DEFINE_CASE(RGBA8I);
441 DEFINE_CASE(RGBA16I);
442 DEFINE_CASE(RGBA32I);
443 DEFINE_CASE(RGBA16F);
444
445 DEFINE_CASE(RGBA4);
446 DEFINE_CASE(RGBA8_2);
447 DEFINE_CASE(RGB10_A2_2);
448 default:
449 snprintf(unk_format_str, sizeof(unk_format_str), "MALI_0x%02x", format);
450 return unk_format_str;
451 }
452 }
453
454 #undef DEFINE_CASE
455
456 #define DEFINE_CASE(name) case MALI_MSAA_ ## name: return "MALI_MSAA_" #name
457 static char *
458 pandecode_msaa_mode(enum mali_msaa_mode mode)
459 {
460 switch (mode) {
461 DEFINE_CASE(SINGLE);
462 DEFINE_CASE(AVERAGE);
463 DEFINE_CASE(MULTIPLE);
464 DEFINE_CASE(LAYERED);
465 default:
466 unreachable("Impossible");
467 return "";
468 }
469 }
470 #undef DEFINE_CASE
471
472 static char *pandecode_attr_mode_short(enum mali_attr_mode mode)
473 {
474 switch(mode) {
475 /* TODO: Combine to just "instanced" once this can be done
476 * unambiguously in all known cases */
477 case MALI_ATTR_POT_DIVIDE:
478 return "instanced_pot";
479 case MALI_ATTR_MODULO:
480 return "instanced_mod";
481 case MALI_ATTR_NPOT_DIVIDE:
482 return "instanced_npot";
483 case MALI_ATTR_IMAGE:
484 return "image";
485 default:
486 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
487 return "";
488 }
489 }
490
491 static const char *
492 pandecode_special_record(uint64_t v, bool* attribute)
493 {
494 switch(v) {
495 case MALI_ATTR_VERTEXID:
496 *attribute = true;
497 return "gl_VertexID";
498 case MALI_ATTR_INSTANCEID:
499 *attribute = true;
500 return "gl_InstanceID";
501 case MALI_VARYING_FRAG_COORD:
502 return "gl_FragCoord";
503 case MALI_VARYING_FRONT_FACING:
504 return "gl_FrontFacing";
505 case MALI_VARYING_POINT_COORD:
506 return "gl_PointCoord";
507 default:
508 pandecode_msg("XXX: invalid special record %" PRIx64 "\n", v);
509 return "";
510 }
511 }
512
513 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
514 static char *
515 pandecode_block_format(enum mali_block_format fmt)
516 {
517 switch (fmt) {
518 DEFINE_CASE(TILED);
519 DEFINE_CASE(UNKNOWN);
520 DEFINE_CASE(LINEAR);
521 DEFINE_CASE(AFBC);
522
523 default:
524 unreachable("Invalid case");
525 }
526 }
527 #undef DEFINE_CASE
528
529 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
530 static char *
531 pandecode_exception_access(unsigned access)
532 {
533 switch (access) {
534 DEFINE_CASE(NONE);
535 DEFINE_CASE(EXECUTE);
536 DEFINE_CASE(READ);
537 DEFINE_CASE(WRITE);
538
539 default:
540 unreachable("Invalid case");
541 }
542 }
543 #undef DEFINE_CASE
544
545 /* Midgard's tiler descriptor is embedded within the
546 * larger FBD */
547
548 static void
549 pandecode_midgard_tiler_descriptor(
550 const struct midgard_tiler_descriptor *t,
551 unsigned width,
552 unsigned height,
553 bool is_fragment,
554 bool has_hierarchy)
555 {
556 pandecode_log(".tiler = {\n");
557 pandecode_indent++;
558
559 if (t->hierarchy_mask == MALI_TILER_DISABLED)
560 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
561 else
562 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
563
564 /* We know this name from the kernel, but we never see it nonzero */
565
566 if (t->flags)
567 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
568
569 MEMORY_PROP(t, polygon_list);
570
571 /* The body is offset from the base of the polygon list */
572 //assert(t->polygon_list_body > t->polygon_list);
573 unsigned body_offset = t->polygon_list_body - t->polygon_list;
574
575 /* It needs to fit inside the reported size */
576 //assert(t->polygon_list_size >= body_offset);
577
578 /* Now that we've sanity checked, we'll try to calculate the sizes
579 * ourselves for comparison */
580
581 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask, has_hierarchy);
582 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask, has_hierarchy);
583
584 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
585 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
586 ref_header, ref_size);
587 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
588 pandecode_msg("body offset %d\n", body_offset);
589 }
590
591 /* The tiler heap has a start and end specified -- it should be
592 * identical to what we have in the BO. The exception is if tiling is
593 * disabled. */
594
595 MEMORY_PROP(t, heap_start);
596 assert(t->heap_end >= t->heap_start);
597
598 struct pandecode_mapped_memory *heap =
599 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
600
601 unsigned heap_size = t->heap_end - t->heap_start;
602
603 /* Tiling is enabled with a special flag */
604 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
605 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
606
607 bool tiling_enabled = hierarchy_mask;
608
609 if (tiling_enabled) {
610 /* When tiling is enabled, the heap should be a tight fit */
611 unsigned heap_offset = t->heap_start - heap->gpu_va;
612 if ((heap_offset + heap_size) != heap->length) {
613 pandecode_msg("XXX: heap size %u (expected %zu)\n",
614 heap_size, heap->length - heap_offset);
615 }
616
617 /* We should also have no other flags */
618 if (tiler_flags)
619 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
620 } else {
621 /* When tiling is disabled, we should have that flag and no others */
622
623 if (tiler_flags != MALI_TILER_DISABLED) {
624 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
625 tiler_flags);
626 }
627
628 /* We should also have an empty heap */
629 if (heap_size) {
630 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
631 heap_size);
632 }
633
634 /* Disabled tiling is used only for clear-only jobs, which are
635 * purely FRAGMENT, so we should never see this for
636 * non-FRAGMENT descriptors. */
637
638 if (!is_fragment)
639 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
640 }
641
642 /* We've never seen weights used in practice, but we know from the
643 * kernel these fields is there */
644
645 bool nonzero_weights = false;
646
647 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
648 nonzero_weights |= t->weights[w] != 0x0;
649 }
650
651 if (nonzero_weights) {
652 pandecode_log(".weights = { ");
653
654 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
655 pandecode_log_cont("%d, ", t->weights[w]);
656 }
657
658 pandecode_log("},");
659 }
660
661 pandecode_indent--;
662 pandecode_log("}\n");
663 }
664
665 /* TODO: The Bifrost tiler is not understood at all yet */
666
667 static void
668 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer *fb)
669 {
670 pandecode_log(".tiler = {\n");
671 pandecode_indent++;
672
673 MEMORY_PROP(fb, tiler_meta);
674
675 for (int i = 0; i < 16; i++) {
676 if (fb->zeros[i] != 0) {
677 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
678 i, fb->zeros[i]);
679 }
680 }
681
682 pandecode_log("},\n");
683
684 pandecode_indent--;
685 pandecode_log("}\n");
686
687 }
688
689 /* Information about the framebuffer passed back for
690 * additional analysis */
691
692 struct pandecode_fbd {
693 unsigned width;
694 unsigned height;
695 unsigned rt_count;
696 bool has_extra;
697 };
698
699 static void
700 pandecode_sfbd_format(struct mali_sfbd_format format)
701 {
702 pandecode_log(".format = {\n");
703 pandecode_indent++;
704
705 pandecode_log(".unk1 = ");
706 pandecode_log_decoded_flags(sfbd_unk1_info, format.unk1);
707 pandecode_log_cont(",\n");
708
709 /* TODO: Map formats so we can check swizzles and print nicely */
710 pandecode_log("swizzle");
711 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
712 pandecode_log_cont(",\n");
713
714 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
715 (format.nr_channels + 1));
716
717 pandecode_log(".unk2 = ");
718 pandecode_log_decoded_flags(sfbd_unk2_info, format.unk2);
719 pandecode_log_cont(",\n");
720
721 pandecode_prop("block = %s", pandecode_block_format(format.block));
722
723 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
724
725 pandecode_indent--;
726 pandecode_log("},\n");
727 }
728
729 static void
730 pandecode_shared_memory(const struct mali_shared_memory *desc, bool is_compute)
731 {
732 pandecode_prop("stack_shift = 0x%x", desc->stack_shift);
733
734 if (desc->unk0)
735 pandecode_prop("unk0 = 0x%x", desc->unk0);
736
737 if (desc->shared_workgroup_count != 0x1F) {
738 pandecode_prop("shared_workgroup_count = %d", desc->shared_workgroup_count);
739 if (!is_compute)
740 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
741 }
742
743 if (desc->shared_unk1 || desc->shared_shift) {
744 pandecode_prop("shared_unk1 = %X", desc->shared_unk1);
745 pandecode_prop("shared_shift = %X", desc->shared_shift);
746
747 if (!is_compute)
748 pandecode_msg("XXX: shared memory configured in noncompute shader");
749 }
750
751 if (desc->shared_zero) {
752 pandecode_msg("XXX: shared memory zero tripped\n");
753 pandecode_prop("shared_zero = 0x%" PRIx32, desc->shared_zero);
754 }
755
756 if (desc->shared_memory && !is_compute)
757 pandecode_msg("XXX: shared memory used in noncompute shader\n");
758
759 MEMORY_PROP(desc, scratchpad);
760 MEMORY_PROP(desc, shared_memory);
761 MEMORY_PROP(desc, unknown1);
762
763 if (desc->scratchpad) {
764 struct pandecode_mapped_memory *smem =
765 pandecode_find_mapped_gpu_mem_containing(desc->scratchpad);
766
767 pandecode_msg("scratchpad size %u\n", smem->length);
768 }
769
770 }
771
772 static struct pandecode_fbd
773 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment, unsigned gpu_id)
774 {
775 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
776 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
777
778 struct pandecode_fbd info = {
779 .has_extra = false,
780 .rt_count = 1
781 };
782
783 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
784 pandecode_indent++;
785
786 pandecode_log(".shared_memory = {\n");
787 pandecode_indent++;
788 pandecode_shared_memory(&s->shared_memory, false);
789 pandecode_indent--;
790 pandecode_log("},\n");
791
792 pandecode_sfbd_format(s->format);
793
794 info.width = s->width + 1;
795 info.height = s->height + 1;
796
797 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
798 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
799
800 MEMORY_PROP(s, checksum);
801
802 if (s->checksum_stride)
803 pandecode_prop("checksum_stride = %d", s->checksum_stride);
804
805 MEMORY_PROP(s, framebuffer);
806 pandecode_prop("stride = %d", s->stride);
807
808 /* Earlier in the actual commandstream -- right before width -- but we
809 * delay to flow nicer */
810
811 pandecode_log(".clear_flags = ");
812 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
813 pandecode_log_cont(",\n");
814
815 if (s->depth_buffer) {
816 MEMORY_PROP(s, depth_buffer);
817 pandecode_prop("depth_stride = %d", s->depth_stride);
818 }
819
820 if (s->stencil_buffer) {
821 MEMORY_PROP(s, stencil_buffer);
822 pandecode_prop("stencil_stride = %d", s->stencil_stride);
823 }
824
825 if (s->depth_stride_zero ||
826 s->stencil_stride_zero ||
827 s->zero7 || s->zero8) {
828 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
829 pandecode_prop("depth_stride_zero = 0x%x",
830 s->depth_stride_zero);
831 pandecode_prop("stencil_stride_zero = 0x%x",
832 s->stencil_stride_zero);
833 pandecode_prop("zero7 = 0x%" PRIx32,
834 s->zero7);
835 pandecode_prop("zero8 = 0x%" PRIx32,
836 s->zero8);
837 }
838
839 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
840 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
841 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
842 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
843 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
844 }
845
846 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
847 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
848 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
849 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
850 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
851 }
852
853 if (s->clear_stencil) {
854 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
855 }
856
857 const struct midgard_tiler_descriptor t = s->tiler;
858
859 bool has_hierarchy = !(gpu_id == 0x0720 || gpu_id == 0x0820 || gpu_id == 0x0830);
860 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment, has_hierarchy);
861
862 pandecode_indent--;
863 pandecode_log("};\n");
864
865 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
866 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
867 pandecode_prop("zero5 = 0x%" PRIx32, s->zero5);
868
869 pandecode_log_cont(".zero3 = {");
870
871 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
872 pandecode_log_cont("%X, ", s->zero3[i]);
873
874 pandecode_log_cont("},\n");
875
876 pandecode_log_cont(".zero6 = {");
877
878 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
879 pandecode_log_cont("%X, ", s->zero6[i]);
880
881 pandecode_log_cont("},\n");
882
883 return info;
884 }
885
886 static void
887 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
888 {
889 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
890 const struct mali_shared_memory *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
891
892 pandecode_log("struct mali_shared_memory shared_%"PRIx64"_%d = {\n", gpu_va, job_no);
893 pandecode_indent++;
894 pandecode_shared_memory(s, true);
895 pandecode_indent--;
896 pandecode_log("},\n");
897 }
898
899 /* Extracts the number of components associated with a Mali format */
900
901 static unsigned
902 pandecode_format_component_count(enum mali_format fmt)
903 {
904 /* Mask out the format class */
905 unsigned top = fmt & 0b11100000;
906
907 switch (top) {
908 case MALI_FORMAT_SNORM:
909 case MALI_FORMAT_UINT:
910 case MALI_FORMAT_UNORM:
911 case MALI_FORMAT_SINT:
912 return ((fmt >> 3) & 3) + 1;
913 default:
914 /* TODO: Validate */
915 return 4;
916 }
917 }
918
919 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
920
921 static unsigned
922 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
923 {
924 unsigned mask = 0;
925 assert(MALI_CHANNEL_RED == 0);
926
927 for (unsigned c = 0; c < 4; ++c) {
928 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
929
930 if (chan <= MALI_CHANNEL_ALPHA)
931 mask |= (1 << chan);
932 }
933
934 return mask;
935 }
936
937 /* Validates that a (format, swizzle) pair is valid, in the sense that the
938 * swizzle doesn't access any components that are undefined in the format.
939 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
940 * omitted */
941
942 static bool
943 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
944 {
945 unsigned nr_comp = pandecode_format_component_count(fmt);
946 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
947 unsigned valid_mask = (1 << nr_comp) - 1;
948 unsigned invalid_mask = ~valid_mask;
949
950 if (access_mask & invalid_mask) {
951 pandecode_msg("XXX: invalid components accessed\n");
952 return false;
953 }
954
955 /* Check for the default non-swizzling swizzle so we can suppress
956 * useless printing for the defaults */
957
958 unsigned default_swizzles[4] = {
959 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
960 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
961 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
962 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
963 };
964
965 return (swizzle == default_swizzles[nr_comp - 1]);
966 }
967
968 /* Maps MALI_RGBA32F to rgba32f, etc */
969
970 static void
971 pandecode_format_short(enum mali_format fmt, bool srgb)
972 {
973 /* We want a type-like format, so cut off the initial MALI_ */
974 char *format = pandecode_format(fmt);
975 format += strlen("MALI_");
976
977 unsigned len = strlen(format);
978 char *lower_format = calloc(1, len + 1);
979
980 for (unsigned i = 0; i < len; ++i)
981 lower_format[i] = tolower(format[i]);
982
983 /* Sanity check sRGB flag is applied to RGB, per the name */
984 if (srgb && lower_format[0] != 'r')
985 pandecode_msg("XXX: sRGB applied to non-colour format\n");
986
987 /* Just prefix with an s, so you get formats like srgba8_unorm */
988 if (srgb)
989 pandecode_log_cont("s");
990
991 pandecode_log_cont("%s", lower_format);
992 free(lower_format);
993 }
994
995 static void
996 pandecode_swizzle(unsigned swizzle, enum mali_format format)
997 {
998 /* First, do some validation */
999 bool trivial_swizzle = pandecode_validate_format_swizzle(
1000 format, swizzle);
1001
1002 if (trivial_swizzle)
1003 return;
1004
1005 /* Next, print the swizzle */
1006 pandecode_log_cont(".");
1007
1008 static const char components[] = "rgba01";
1009
1010 for (unsigned c = 0; c < 4; ++c) {
1011 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
1012
1013 if (chan >= MALI_CHANNEL_RESERVED_0) {
1014 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
1015 continue;
1016 }
1017 pandecode_log_cont("%c", components[chan]);
1018 }
1019 }
1020
1021 static void
1022 pandecode_rt_format(struct mali_rt_format format)
1023 {
1024 pandecode_log(".format = {\n");
1025 pandecode_indent++;
1026
1027 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
1028 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
1029 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
1030 pandecode_prop("unk4 = 0x%" PRIx32, format.unk4);
1031
1032 pandecode_prop("block = %s", pandecode_block_format(format.block));
1033
1034 /* TODO: Map formats so we can check swizzles and print nicely */
1035 pandecode_log("swizzle");
1036 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
1037 pandecode_log_cont(",\n");
1038
1039 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
1040 (format.nr_channels + 1));
1041
1042 pandecode_log(".flags = ");
1043 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
1044 pandecode_log_cont(",\n");
1045
1046 pandecode_prop("msaa = %s", pandecode_msaa_mode(format.msaa));
1047
1048 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
1049 * which is a faster hardware-based alternative to the wallpaper method
1050 * to preserve framebuffer contents across frames. In practice, MFBD
1051 * preload is buggy on Midgard, and so this is a chicken bit. If this
1052 * bit isn't set, most likely something broke unrelated to preload */
1053
1054 if (!format.no_preload) {
1055 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
1056 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
1057 }
1058
1059 if (format.zero)
1060 pandecode_prop("zero = 0x%" PRIx32, format.zero);
1061
1062 pandecode_indent--;
1063 pandecode_log("},\n");
1064 }
1065
1066 static void
1067 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct mali_framebuffer *fb)
1068 {
1069 pandecode_log("struct mali_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
1070 pandecode_indent++;
1071
1072 for (int i = 0; i < (fb->rt_count_1 + 1); i++) {
1073 mali_ptr rt_va = gpu_va + i * sizeof(struct mali_render_target);
1074 struct pandecode_mapped_memory *mem =
1075 pandecode_find_mapped_gpu_mem_containing(rt_va);
1076 const struct mali_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
1077
1078 pandecode_log("{\n");
1079 pandecode_indent++;
1080
1081 pandecode_rt_format(rt->format);
1082
1083 if (rt->format.block == MALI_BLOCK_AFBC) {
1084 pandecode_log(".afbc = {\n");
1085 pandecode_indent++;
1086
1087 char *a = pointer_as_memory_reference(rt->afbc.metadata);
1088 pandecode_prop("metadata = %s", a);
1089 free(a);
1090
1091 pandecode_prop("stride = %d", rt->afbc.stride);
1092
1093 pandecode_log(".flags = ");
1094 pandecode_log_decoded_flags(afbc_fmt_flag_info, rt->afbc.flags);
1095 pandecode_log_cont(",\n");
1096
1097 pandecode_indent--;
1098 pandecode_log("},\n");
1099 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.flags) {
1100 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1101 rt->afbc.metadata,
1102 rt->afbc.stride,
1103 rt->afbc.flags);
1104 }
1105
1106 MEMORY_PROP(rt, framebuffer);
1107 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
1108
1109 if (rt->layer_stride)
1110 pandecode_prop("layer_stride = %d", rt->layer_stride);
1111
1112 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
1113 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
1114 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
1115 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
1116 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
1117 }
1118
1119 if (rt->zero1 || rt->zero2) {
1120 pandecode_msg("XXX: render target zeros tripped\n");
1121 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
1122 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
1123 }
1124
1125 pandecode_indent--;
1126 pandecode_log("},\n");
1127 }
1128
1129 pandecode_indent--;
1130 pandecode_log("};\n");
1131 }
1132
1133 static struct pandecode_fbd
1134 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment, bool is_compute, bool is_bifrost)
1135 {
1136 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1137 const struct mali_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
1138
1139 struct pandecode_fbd info;
1140
1141 if (is_bifrost && fb->msaa.sample_locations) {
1142 /* The blob stores all possible sample locations in a single buffer
1143 * allocated on startup, and just switches the pointer when switching
1144 * MSAA state. For now, we just put the data into the cmdstream, but we
1145 * should do something like what the blob does with a real driver.
1146 *
1147 * There seem to be 32 slots for sample locations, followed by another
1148 * 16. The second 16 is just the center location followed by 15 zeros
1149 * in all the cases I've identified (maybe shader vs. depth/color
1150 * samples?).
1151 */
1152
1153 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->msaa.sample_locations);
1154
1155 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->msaa.sample_locations);
1156
1157 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
1158 pandecode_indent++;
1159
1160 for (int i = 0; i < 32 + 16; i++) {
1161 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
1162 }
1163
1164 pandecode_indent--;
1165 pandecode_log("};\n");
1166 }
1167
1168 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
1169 pandecode_indent++;
1170
1171 if (is_bifrost) {
1172 pandecode_log(".msaa = {\n");
1173 pandecode_indent++;
1174
1175 if (fb->msaa.sample_locations)
1176 pandecode_prop("sample_locations = sample_locations_%d", job_no);
1177 else
1178 pandecode_msg("XXX: sample_locations missing\n");
1179
1180 if (fb->msaa.zero1 || fb->msaa.zero2 || fb->msaa.zero4) {
1181 pandecode_msg("XXX: multisampling zero tripped\n");
1182 pandecode_prop("zero1 = %" PRIx64, fb->msaa.zero1);
1183 pandecode_prop("zero2 = %" PRIx64, fb->msaa.zero2);
1184 pandecode_prop("zero4 = %" PRIx64, fb->msaa.zero4);
1185 }
1186
1187 pandecode_indent--;
1188 pandecode_log("},\n");
1189 } else {
1190 pandecode_log(".shared_memory = {\n");
1191 pandecode_indent++;
1192 pandecode_shared_memory(&fb->shared_memory, is_compute);
1193 pandecode_indent--;
1194 pandecode_log("},\n");
1195 }
1196
1197 info.width = fb->width1 + 1;
1198 info.height = fb->height1 + 1;
1199 info.rt_count = fb->rt_count_1 + 1;
1200
1201 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
1202 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
1203 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
1204 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
1205
1206 pandecode_prop("unk1 = 0x%x", fb->unk1);
1207 pandecode_prop("unk2 = 0x%x", fb->unk2);
1208 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1209 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1210
1211 pandecode_log(".mfbd_flags = ");
1212 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1213 pandecode_log_cont(",\n");
1214
1215 if (fb->clear_stencil)
1216 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1217
1218 if (fb->clear_depth)
1219 pandecode_prop("clear_depth = %f", fb->clear_depth);
1220
1221 if (!is_compute)
1222 if (is_bifrost)
1223 pandecode_bifrost_tiler_descriptor(fb);
1224 else {
1225 const struct midgard_tiler_descriptor t = fb->tiler;
1226 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment, true);
1227 }
1228 else
1229 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
1230
1231 if (fb->zero3 || fb->zero4) {
1232 pandecode_msg("XXX: framebuffer zeros tripped\n");
1233 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1234 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1235 }
1236
1237 pandecode_indent--;
1238 pandecode_log("};\n");
1239
1240 gpu_va += sizeof(struct mali_framebuffer);
1241
1242 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
1243
1244 if (info.has_extra) {
1245 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1246 const struct mali_framebuffer_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1247
1248 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1249 pandecode_indent++;
1250
1251 MEMORY_PROP(fbx, checksum);
1252
1253 if (fbx->checksum_stride)
1254 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1255
1256 pandecode_log(".flags_hi = ");
1257 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info, fbx->flags_hi);
1258 pandecode_log_cont(",\n");
1259
1260 pandecode_log(".flags_lo = ");
1261 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info, fbx->flags_lo);
1262 pandecode_log_cont(",\n");
1263
1264 pandecode_prop("zs_block = %s", pandecode_block_format(fbx->zs_block));
1265 pandecode_prop("zs_samples = MALI_POSITIVE(%u)", fbx->zs_samples + 1);
1266
1267 if (fbx->zs_block == MALI_BLOCK_AFBC) {
1268 pandecode_log(".ds_afbc = {\n");
1269 pandecode_indent++;
1270
1271 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1272 pandecode_prop("depth_stencil_afbc_stride = %d",
1273 fbx->ds_afbc.depth_stencil_afbc_stride);
1274 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1275
1276 pandecode_log(".flags = ");
1277 pandecode_log_decoded_flags(afbc_fmt_flag_info, fbx->ds_afbc.flags);
1278 pandecode_log_cont(",\n");
1279
1280 if (fbx->ds_afbc.padding) {
1281 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1282 pandecode_prop("padding = 0x%" PRIx64, fbx->ds_afbc.padding);
1283 }
1284
1285 pandecode_indent--;
1286 pandecode_log("},\n");
1287 } else {
1288 pandecode_log(".ds_linear = {\n");
1289 pandecode_indent++;
1290
1291 if (fbx->ds_linear.depth) {
1292 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1293 pandecode_prop("depth_stride = %d",
1294 fbx->ds_linear.depth_stride);
1295 pandecode_prop("depth_layer_stride = %d",
1296 fbx->ds_linear.depth_layer_stride);
1297 } else if (fbx->ds_linear.depth_stride || fbx->ds_linear.depth_layer_stride) {
1298 pandecode_msg("XXX: depth stride zero tripped %d %d\n", fbx->ds_linear.depth_stride, fbx->ds_linear.depth_layer_stride);
1299 }
1300
1301 if (fbx->ds_linear.stencil) {
1302 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1303 pandecode_prop("stencil_stride = %d",
1304 fbx->ds_linear.stencil_stride);
1305 pandecode_prop("stencil_layer_stride = %d",
1306 fbx->ds_linear.stencil_layer_stride);
1307 } else if (fbx->ds_linear.stencil_stride || fbx->ds_linear.stencil_layer_stride) {
1308 pandecode_msg("XXX: stencil stride zero tripped %d %d\n", fbx->ds_linear.stencil_stride, fbx->ds_linear.stencil_layer_stride);
1309 }
1310
1311 if (fbx->ds_linear.depth_stride_zero ||
1312 fbx->ds_linear.stencil_stride_zero) {
1313 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1314 pandecode_prop("depth_stride_zero = 0x%x",
1315 fbx->ds_linear.depth_stride_zero);
1316 pandecode_prop("stencil_stride_zero = 0x%x",
1317 fbx->ds_linear.stencil_stride_zero);
1318 }
1319
1320 pandecode_indent--;
1321 pandecode_log("},\n");
1322 }
1323
1324 if (fbx->clear_color_1 | fbx->clear_color_2) {
1325 pandecode_prop("clear_color_1 = 0x%" PRIx32, fbx->clear_color_1);
1326 pandecode_prop("clear_color_2 = 0x%" PRIx32, fbx->clear_color_2);
1327 }
1328
1329 if (fbx->zero3) {
1330 pandecode_msg("XXX: fb_extra zeros tripped\n");
1331 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1332 }
1333
1334 pandecode_indent--;
1335 pandecode_log("};\n");
1336
1337 gpu_va += sizeof(struct mali_framebuffer_extra);
1338 }
1339
1340 if (is_fragment)
1341 pandecode_render_target(gpu_va, job_no, fb);
1342
1343 return info;
1344 }
1345
1346 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1347 * count */
1348
1349 static void
1350 pandecode_padded_vertices(unsigned shift, unsigned k)
1351 {
1352 unsigned odd = 2*k + 1;
1353 unsigned pot = 1 << shift;
1354 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1355 }
1356
1357 /* Given a magic divisor, recover what we were trying to divide by.
1358 *
1359 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1360 * 0 <= m < 2^N, for N bits in m.
1361 *
1362 * Let q represent the number we would like to divide by.
1363 *
1364 * By definition of a magic divisor for N-bit unsigned integers (a number you
1365 * multiply by to magically get division), m is a number such that:
1366 *
1367 * (m * x) & (2^N - 1) = floor(x/q).
1368 * for all x on Z where 0 <= x < 2^N
1369 *
1370 * Ignore the case where any of the above values equals zero; it is irrelevant
1371 * for our purposes (instanced arrays).
1372 *
1373 * Choose x = q. Then:
1374 *
1375 * (m * x) & (2^N - 1) = floor(x/q).
1376 * (m * q) & (2^N - 1) = floor(q/q).
1377 *
1378 * floor(q/q) = floor(1) = 1, therefore:
1379 *
1380 * (m * q) & (2^N - 1) = 1
1381 *
1382 * Recall the identity that the bitwise AND of one less than a power-of-two
1383 * equals the modulo with that power of two, i.e. for all x:
1384 *
1385 * x & (2^N - 1) = x % N
1386 *
1387 * Therefore:
1388 *
1389 * mq % (2^N) = 1
1390 *
1391 * By definition, a modular multiplicative inverse of a number m is the number
1392 * q such that with respect to a modulos M:
1393 *
1394 * mq % M = 1
1395 *
1396 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1397 *
1398 */
1399
1400 static void
1401 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1402 {
1403 #if 0
1404 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1405 * shift) the most lame way possible... just repeatedly add.
1406 * Asymptoptically slow but nobody cares in practice, unless you have
1407 * massive numbers of vertices or high divisors. */
1408
1409 unsigned inverse = 0;
1410
1411 /* Magic implicitly has the highest bit set */
1412 magic |= (1 << 31);
1413
1414 /* Depending on rounding direction */
1415 if (extra)
1416 magic++;
1417
1418 for (;;) {
1419 uint32_t product = magic * inverse;
1420
1421 if (shift) {
1422 product >>= shift;
1423 }
1424
1425 if (product == 1)
1426 break;
1427
1428 ++inverse;
1429 }
1430
1431 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1432
1433 /* Recall we're supposed to divide by (gl_level_divisor *
1434 * padded_num_vertices) */
1435
1436 unsigned padded_num_vertices = inverse / orig_divisor;
1437
1438 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1439 #endif
1440 }
1441
1442 static void
1443 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1444 mali_ptr addr, int job_no, char *suffix,
1445 int count, bool varying, enum mali_job_type job_type)
1446 {
1447 char *prefix = varying ? "varying" : "attribute";
1448 assert(addr);
1449
1450 if (!count) {
1451 pandecode_msg("warn: No %s records\n", prefix);
1452 return;
1453 }
1454
1455 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1456
1457 for (int i = 0; i < count; ++i) {
1458 /* First, check for special records */
1459 if (attr[i].elements < MALI_RECORD_SPECIAL) {
1460 if (attr[i].size)
1461 pandecode_msg("XXX: tripped size=%d\n", attr[i].size);
1462
1463 if (attr[i].stride) {
1464 /* gl_InstanceID passes a magic divisor in the
1465 * stride field to divide by the padded vertex
1466 * count. No other records should do so, so
1467 * stride should otherwise be zero. Note that
1468 * stride in the usual attribute sense doesn't
1469 * apply to special records. */
1470
1471 bool has_divisor = attr[i].elements == MALI_ATTR_INSTANCEID;
1472
1473 pandecode_log_cont("/* %smagic divisor = %X */ ",
1474 has_divisor ? "" : "XXX: ", attr[i].stride);
1475 }
1476
1477 if (attr[i].shift || attr[i].extra_flags) {
1478 /* Attributes use these fields for
1479 * instancing/padding/etc type issues, but
1480 * varyings don't */
1481
1482 pandecode_log_cont("/* %sshift=%d, extra=%d */ ",
1483 varying ? "XXX: " : "",
1484 attr[i].shift, attr[i].extra_flags);
1485 }
1486
1487 /* Print the special record name */
1488 bool attribute = false;
1489 pandecode_log("%s_%d = %s;\n", prefix, i, pandecode_special_record(attr[i].elements, &attribute));
1490
1491 /* Sanity check */
1492 if (attribute == varying)
1493 pandecode_msg("XXX: mismatched special record\n");
1494
1495 continue;
1496 }
1497
1498 enum mali_attr_mode mode = attr[i].elements & 7;
1499
1500 if (mode == MALI_ATTR_UNUSED)
1501 pandecode_msg("XXX: unused attribute record\n");
1502
1503 /* For non-linear records, we need to print the type of record */
1504 if (mode != MALI_ATTR_LINEAR)
1505 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode));
1506
1507 /* Print the name to link with attr_meta */
1508 pandecode_log_cont("%s_%d", prefix, i);
1509
1510 /* Print the stride and size */
1511 pandecode_log_cont("<%u>[%u]", attr[i].stride, attr[i].size);
1512
1513 /* TODO: Sanity check the quotient itself. It must be equal to
1514 * (or be greater than, if the driver added padding) the padded
1515 * vertex count. */
1516
1517 /* Finally, print the pointer */
1518 mali_ptr raw_elements = attr[i].elements & ~7;
1519 char *a = pointer_as_memory_reference(raw_elements);
1520 pandecode_log_cont(" = (%s);\n", a);
1521 free(a);
1522
1523 /* Check the pointer */
1524 pandecode_validate_buffer(raw_elements, attr[i].size);
1525
1526 /* shift/extra_flags exist only for instanced */
1527 if (attr[i].shift | attr[i].extra_flags) {
1528 /* These are set to random values by the blob for
1529 * varyings, most likely a symptom of uninitialized
1530 * memory where the hardware masked the bug. As such we
1531 * put this at a warning, not an error. */
1532
1533 if (mode == MALI_ATTR_LINEAR)
1534 pandecode_msg("warn: instancing fields set for linear\n");
1535
1536 pandecode_prop("shift = %d", attr[i].shift);
1537 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1538 }
1539
1540 /* Decode further where possible */
1541
1542 if (mode == MALI_ATTR_MODULO) {
1543 pandecode_padded_vertices(
1544 attr[i].shift,
1545 attr[i].extra_flags);
1546 }
1547
1548 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1549 i++;
1550 pandecode_log("{\n");
1551 pandecode_indent++;
1552 pandecode_prop("unk = 0x%x", attr[i].unk);
1553 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1554 if (attr[i].zero != 0)
1555 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1556 pandecode_prop("divisor = %d", attr[i].divisor);
1557 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1558 pandecode_indent--;
1559 pandecode_log("}, \n");
1560 }
1561
1562 }
1563
1564 pandecode_log("\n");
1565 }
1566
1567 static mali_ptr
1568 pandecode_shader_address(const char *name, mali_ptr ptr)
1569 {
1570 /* TODO: Decode flags */
1571 mali_ptr shader_ptr = ptr & ~15;
1572
1573 char *a = pointer_as_memory_reference(shader_ptr);
1574 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1575 free(a);
1576
1577 return shader_ptr;
1578 }
1579
1580 static void
1581 pandecode_blend_equation(const struct mali_blend_equation *blend)
1582 {
1583 if (blend->zero1)
1584 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1585
1586 pandecode_log(".equation = {\n");
1587 pandecode_indent++;
1588
1589 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1590 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1591
1592 pandecode_log(".color_mask = ");
1593 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1594 pandecode_log_cont(",\n");
1595
1596 pandecode_indent--;
1597 pandecode_log("},\n");
1598 }
1599
1600 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1601
1602 static unsigned
1603 decode_bifrost_constant(u16 constant)
1604 {
1605 float lo = (float) (constant & 0xFF);
1606 float hi = (float) (constant >> 8);
1607
1608 return (hi / 255.0) + (lo / 65535.0);
1609 }
1610
1611 static mali_ptr
1612 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1613 {
1614 struct bifrost_blend_rt *b =
1615 ((struct bifrost_blend_rt *) descs) + rt_no;
1616
1617 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1618 pandecode_indent++;
1619
1620 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1621 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1622 b->constant, decode_bifrost_constant(b->constant));
1623
1624 /* TODO figure out blend shader enable bit */
1625 pandecode_blend_equation(&b->equation);
1626
1627 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1628 pandecode_prop("index = 0x%" PRIx16, b->index);
1629
1630 pandecode_log(".format = ");
1631 pandecode_format_short(b->format, false);
1632 pandecode_swizzle(b->swizzle, b->format);
1633 pandecode_log_cont(",\n");
1634
1635 pandecode_prop("swizzle = 0x%" PRIx32, b->swizzle);
1636 pandecode_prop("format = 0x%" PRIx32, b->format);
1637
1638 if (b->zero1) {
1639 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1640 pandecode_prop("zero1 = 0x%" PRIx32, b->zero1);
1641 }
1642
1643 pandecode_log(".shader_type = ");
1644 switch(b->shader_type) {
1645 case BIFROST_BLEND_F16:
1646 pandecode_log_cont("BIFROST_BLEND_F16");
1647 break;
1648 case BIFROST_BLEND_F32:
1649 pandecode_log_cont("BIFROST_BLEND_F32");
1650 break;
1651 case BIFROST_BLEND_I32:
1652 pandecode_log_cont("BIFROST_BLEND_I32");
1653 break;
1654 case BIFROST_BLEND_U32:
1655 pandecode_log_cont("BIFROST_BLEND_U32");
1656 break;
1657 case BIFROST_BLEND_I16:
1658 pandecode_log_cont("BIFROST_BLEND_I16");
1659 break;
1660 case BIFROST_BLEND_U16:
1661 pandecode_log_cont("BIFROST_BLEND_U16");
1662 break;
1663 }
1664 pandecode_log_cont(",\n");
1665
1666 if (b->zero2) {
1667 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1668 pandecode_prop("zero2 = 0x%" PRIx32, b->zero2);
1669 }
1670
1671 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1672
1673 pandecode_indent--;
1674 pandecode_log("},\n");
1675
1676 return 0;
1677 }
1678
1679 static mali_ptr
1680 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1681 {
1682 /* constant/equation is in a union */
1683 if (!blend->shader)
1684 return 0;
1685
1686 pandecode_log(".blend = {\n");
1687 pandecode_indent++;
1688
1689 if (is_shader) {
1690 pandecode_shader_address("shader", blend->shader);
1691 } else {
1692 pandecode_blend_equation(&blend->equation);
1693 pandecode_prop("constant = %f", blend->constant);
1694 }
1695
1696 pandecode_indent--;
1697 pandecode_log("},\n");
1698
1699 /* Return blend shader to disassemble if present */
1700 return is_shader ? (blend->shader & ~0xF) : 0;
1701 }
1702
1703 static mali_ptr
1704 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1705 {
1706 struct midgard_blend_rt *b =
1707 ((struct midgard_blend_rt *) descs) + rt_no;
1708
1709 /* Flags determine presence of blend shader */
1710 bool is_shader = (b->flags & 0xF) >= 0x2;
1711
1712 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1713 pandecode_indent++;
1714
1715 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1716
1717 union midgard_blend blend = b->blend;
1718 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1719
1720 pandecode_indent--;
1721 pandecode_log("};\n");
1722
1723 return shader;
1724 }
1725
1726 /* Attributes and varyings have descriptor records, which contain information
1727 * about their format and ordering with the attribute/varying buffers. We'll
1728 * want to validate that the combinations specified are self-consistent.
1729 */
1730
1731 static int
1732 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1733 {
1734 char base[128];
1735 char *prefix = varying ? "varying" : "attribute";
1736 unsigned max_index = 0;
1737 snprintf(base, sizeof(base), "%s_meta", prefix);
1738
1739 struct mali_attr_meta *attr_meta;
1740 mali_ptr p = varying ? v->varying_meta : v->attribute_meta;
1741
1742 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1743
1744 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1745 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1746 sizeof(*attr_mem));
1747
1748 /* If the record is discard, it should be zero for everything else */
1749
1750 if (attr_meta->format == MALI_VARYING_DISCARD) {
1751 uint64_t zero =
1752 attr_meta->index |
1753 attr_meta->unknown1 |
1754 attr_meta->unknown3 |
1755 attr_meta->src_offset;
1756
1757 if (zero)
1758 pandecode_msg("XXX: expected empty record for varying discard\n");
1759
1760 /* We want to look for a literal 0000 swizzle -- this
1761 * is not encoded with all zeroes, however */
1762
1763 enum mali_channel z = MALI_CHANNEL_ZERO;
1764 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1765 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1766
1767 if (!good_swizzle)
1768 pandecode_msg("XXX: expected zero swizzle for discard\n");
1769
1770 if (!varying)
1771 pandecode_msg("XXX: cannot discard attribute\n");
1772
1773 /* If we're all good, omit the record */
1774 if (!zero && varying && good_swizzle) {
1775 pandecode_log("/* discarded varying */\n");
1776 continue;
1777 }
1778 }
1779
1780 if (attr_meta->index > max_index)
1781 max_index = attr_meta->index;
1782
1783 if (attr_meta->unknown1 != 0x2) {
1784 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1785 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1786 }
1787
1788 if (attr_meta->unknown3) {
1789 pandecode_msg("XXX: unexpected unknown3 set\n");
1790 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1791 }
1792
1793 pandecode_format_short(attr_meta->format, false);
1794 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1795
1796 if (attr_meta->src_offset)
1797 pandecode_log_cont("[%u]", attr_meta->src_offset);
1798
1799 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1800
1801 pandecode_log_cont(";\n");
1802 }
1803
1804 pandecode_log("\n");
1805
1806 return count ? (max_index + 1) : 0;
1807 }
1808
1809 /* return bits [lo, hi) of word */
1810 static u32
1811 bits(u32 word, u32 lo, u32 hi)
1812 {
1813 if (hi - lo >= 32)
1814 return word; // avoid undefined behavior with the shift
1815
1816 return (word >> lo) & ((1 << (hi - lo)) - 1);
1817 }
1818
1819 static void
1820 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool graphics)
1821 {
1822 pandecode_log(".prefix = {\n");
1823 pandecode_indent++;
1824
1825 /* Decode invocation_count. See the comment before the definition of
1826 * invocation_count for an explanation.
1827 */
1828
1829 unsigned size_y_shift = bits(p->invocation_shifts, 0, 5);
1830 unsigned size_z_shift = bits(p->invocation_shifts, 5, 10);
1831 unsigned workgroups_x_shift = bits(p->invocation_shifts, 10, 16);
1832 unsigned workgroups_y_shift = bits(p->invocation_shifts, 16, 22);
1833 unsigned workgroups_z_shift = bits(p->invocation_shifts, 22, 28);
1834 unsigned workgroups_x_shift_2 = bits(p->invocation_shifts, 28, 32);
1835
1836 unsigned size_x = bits(p->invocation_count, 0, size_y_shift) + 1;
1837 unsigned size_y = bits(p->invocation_count, size_y_shift, size_z_shift) + 1;
1838 unsigned size_z = bits(p->invocation_count, size_z_shift, workgroups_x_shift) + 1;
1839
1840 unsigned groups_x = bits(p->invocation_count, workgroups_x_shift, workgroups_y_shift) + 1;
1841 unsigned groups_y = bits(p->invocation_count, workgroups_y_shift, workgroups_z_shift) + 1;
1842 unsigned groups_z = bits(p->invocation_count, workgroups_z_shift, 32) + 1;
1843
1844 /* Even though we have this decoded, we want to ensure that the
1845 * representation is "unique" so we don't lose anything by printing only
1846 * the final result. More specifically, we need to check that we were
1847 * passed something in canonical form, since the definition per the
1848 * hardware is inherently not unique. How? Well, take the resulting
1849 * decode and pack it ourselves! If it is bit exact with what we
1850 * decoded, we're good to go. */
1851
1852 struct mali_vertex_tiler_prefix ref;
1853 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, graphics);
1854
1855 bool canonical =
1856 (p->invocation_count == ref.invocation_count) &&
1857 (p->invocation_shifts == ref.invocation_shifts);
1858
1859 if (!canonical) {
1860 pandecode_msg("XXX: non-canonical workgroups packing\n");
1861 pandecode_msg("expected: %X, %X",
1862 ref.invocation_count,
1863 ref.invocation_shifts);
1864
1865 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1866 pandecode_prop("size_y_shift = %d", size_y_shift);
1867 pandecode_prop("size_z_shift = %d", size_z_shift);
1868 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift);
1869 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift);
1870 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift);
1871 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2);
1872 }
1873
1874 /* Regardless, print the decode */
1875 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1876 size_x, size_y, size_z,
1877 groups_x, groups_y, groups_z);
1878
1879 /* TODO: Decode */
1880 if (p->unknown_draw)
1881 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1882
1883 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1884
1885 if (p->draw_mode != MALI_DRAW_MODE_NONE)
1886 pandecode_prop("draw_mode = %s", mali_draw_mode_as_str(p->draw_mode));
1887
1888 /* Index count only exists for tiler jobs anyway */
1889
1890 if (p->index_count)
1891 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1892
1893
1894 unsigned index_raw_size = (p->unknown_draw & MALI_DRAW_INDEXED_SIZE);
1895 index_raw_size >>= MALI_DRAW_INDEXED_SHIFT;
1896
1897 /* Validate an index buffer is present if we need one. TODO: verify
1898 * relationship between invocation_count and index_count */
1899
1900 if (p->indices) {
1901 unsigned count = p->index_count;
1902
1903 /* Grab the size */
1904 unsigned size = (index_raw_size == 0x3) ? 4 : index_raw_size;
1905
1906 /* Ensure we got a size, and if so, validate the index buffer
1907 * is large enough to hold a full set of indices of the given
1908 * size */
1909
1910 if (!index_raw_size)
1911 pandecode_msg("XXX: index size missing\n");
1912 else
1913 pandecode_validate_buffer(p->indices, count * size);
1914 } else if (index_raw_size)
1915 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size);
1916
1917 if (p->offset_bias_correction)
1918 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1919
1920 /* TODO: Figure out what this is. It's not zero */
1921 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1922
1923 pandecode_indent--;
1924 pandecode_log("},\n");
1925 }
1926
1927 static void
1928 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1929 {
1930 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1931 uint64_t *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1932
1933 for (int i = 0; i < ubufs_count; i++) {
1934 unsigned size = (ubufs[i] & ((1 << 10) - 1)) * 16;
1935 mali_ptr addr = (ubufs[i] >> 10) << 2;
1936
1937 pandecode_validate_buffer(addr, size);
1938
1939 char *ptr = pointer_as_memory_reference(addr);
1940 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
1941 free(ptr);
1942 }
1943
1944 pandecode_log("\n");
1945 }
1946
1947 static void
1948 pandecode_uniforms(mali_ptr uniforms, unsigned uniform_count)
1949 {
1950 pandecode_validate_buffer(uniforms, uniform_count * 16);
1951
1952 char *ptr = pointer_as_memory_reference(uniforms);
1953 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count, ptr);
1954 free(ptr);
1955 }
1956
1957 static const char *
1958 shader_type_for_job(unsigned type)
1959 {
1960 switch (type) {
1961 case MALI_JOB_TYPE_VERTEX: return "VERTEX";
1962 case MALI_JOB_TYPE_TILER: return "FRAGMENT";
1963 case MALI_JOB_TYPE_COMPUTE: return "COMPUTE";
1964 default:
1965 return "UNKNOWN";
1966 }
1967 }
1968
1969 static unsigned shader_id = 0;
1970
1971 static struct midgard_disasm_stats
1972 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1973 bool is_bifrost, unsigned gpu_id)
1974 {
1975 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1976 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1977
1978 /* Compute maximum possible size */
1979 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1980
1981 /* Print some boilerplate to clearly denote the assembly (which doesn't
1982 * obey indentation rules), and actually do the disassembly! */
1983
1984 pandecode_log_cont("\n\n");
1985
1986 struct midgard_disasm_stats stats;
1987
1988 if (is_bifrost) {
1989 disassemble_bifrost(pandecode_dump_stream, code, sz, true);
1990
1991 /* TODO: Extend stats to Bifrost */
1992 stats.texture_count = -128;
1993 stats.sampler_count = -128;
1994 stats.attribute_count = -128;
1995 stats.varying_count = -128;
1996 stats.uniform_count = -128;
1997 stats.uniform_buffer_count = -128;
1998 stats.work_count = -128;
1999
2000 stats.instruction_count = 0;
2001 stats.bundle_count = 0;
2002 stats.quadword_count = 0;
2003 stats.helper_invocations = false;
2004 } else {
2005 stats = disassemble_midgard(pandecode_dump_stream,
2006 code, sz, gpu_id,
2007 type == MALI_JOB_TYPE_TILER ?
2008 MESA_SHADER_FRAGMENT : MESA_SHADER_VERTEX);
2009 }
2010
2011 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
2012 * driver-internal purposes with the blob and interfere */
2013
2014 bool should_shaderdb = type != MALI_JOB_TYPE_COMPUTE;
2015
2016 if (should_shaderdb) {
2017 unsigned nr_threads =
2018 (stats.work_count <= 4) ? 4 :
2019 (stats.work_count <= 8) ? 2 :
2020 1;
2021
2022 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
2023 "%u inst, %u bundles, %u quadwords, "
2024 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
2025 shader_id++,
2026 shader_type_for_job(type),
2027 stats.instruction_count, stats.bundle_count, stats.quadword_count,
2028 stats.work_count, nr_threads);
2029 }
2030
2031
2032 return stats;
2033 }
2034
2035 static void
2036 pandecode_texture_payload(mali_ptr payload,
2037 enum mali_texture_dimension dim,
2038 enum mali_texture_layout layout,
2039 bool manual_stride,
2040 uint8_t levels,
2041 uint16_t depth,
2042 uint16_t array_size,
2043 struct pandecode_mapped_memory *tmem)
2044 {
2045 pandecode_log(".payload = {\n");
2046 pandecode_indent++;
2047
2048 /* A bunch of bitmap pointers follow.
2049 * We work out the correct number,
2050 * based on the mipmap/cubemap
2051 * properties, but dump extra
2052 * possibilities to futureproof */
2053
2054 int bitmap_count = levels + 1;
2055
2056 /* Miptree for each face */
2057 if (dim == MALI_TEXTURE_DIMENSION_CUBE)
2058 bitmap_count *= 6;
2059
2060 /* Array of layers */
2061 bitmap_count *= depth;
2062
2063 /* Array of textures */
2064 bitmap_count *= array_size;
2065
2066 /* Stride for each element */
2067 if (manual_stride)
2068 bitmap_count *= 2;
2069
2070 mali_ptr *pointers_and_strides = pandecode_fetch_gpu_mem(tmem,
2071 payload, sizeof(mali_ptr) * bitmap_count);
2072 for (int i = 0; i < bitmap_count; ++i) {
2073 /* How we dump depends if this is a stride or a pointer */
2074
2075 if (manual_stride && (i & 1)) {
2076 /* signed 32-bit snuck in as a 64-bit pointer */
2077 uint64_t stride_set = pointers_and_strides[i];
2078 uint32_t clamped_stride = stride_set;
2079 int32_t stride = clamped_stride;
2080 assert(stride_set == clamped_stride);
2081 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
2082 } else {
2083 char *a = pointer_as_memory_reference(pointers_and_strides[i]);
2084 pandecode_log("%s, \n", a);
2085 free(a);
2086 }
2087 }
2088
2089 pandecode_indent--;
2090 pandecode_log("},\n");
2091 }
2092
2093 static void
2094 pandecode_texture(mali_ptr u,
2095 struct pandecode_mapped_memory *tmem,
2096 unsigned job_no, unsigned tex)
2097 {
2098 struct pandecode_mapped_memory *mapped_mem = pandecode_find_mapped_gpu_mem_containing(u);
2099 const uint8_t *cl = pandecode_fetch_gpu_mem(mapped_mem, u, MALI_MIDGARD_TEXTURE_LENGTH);
2100
2101 struct MALI_MIDGARD_TEXTURE temp;
2102 MALI_MIDGARD_TEXTURE_unpack(cl, &temp);
2103 MALI_MIDGARD_TEXTURE_print(pandecode_dump_stream, &temp, 2);
2104
2105 pandecode_texture_payload(u + MALI_MIDGARD_TEXTURE_LENGTH,
2106 temp.dimension, temp.texel_ordering, temp.manual_stride,
2107 temp.levels, temp.depth, temp.array_size, mapped_mem);
2108 }
2109
2110 static void
2111 pandecode_bifrost_texture(
2112 const void *cl,
2113 unsigned job_no,
2114 unsigned tex)
2115 {
2116 struct MALI_BIFROST_TEXTURE temp;
2117 MALI_BIFROST_TEXTURE_unpack(cl, &temp);
2118 MALI_BIFROST_TEXTURE_print(pandecode_dump_stream, &temp, 2);
2119
2120 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(temp.surfaces);
2121 pandecode_texture_payload(temp.surfaces, temp.dimension, temp.texel_ordering,
2122 true, temp.levels, 1, 1, tmem);
2123 }
2124
2125 /* 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 */
2126
2127 static void
2128 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
2129 {
2130 /* Nothing to do */
2131 if (claim == truth)
2132 return;
2133
2134 if (fuzzy && (truth < 0))
2135 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name, claim, truth);
2136
2137 if ((truth >= 0) && !fuzzy) {
2138 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2139 (truth < claim) ? "warn" : "XXX",
2140 name, truth, claim);
2141 } else if ((claim > -truth) && !fuzzy) {
2142 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2143 name, -truth, claim);
2144 } else if (fuzzy && (claim < truth))
2145 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2146 name, truth, claim);
2147
2148 pandecode_log(".%s = %" PRId16, name, claim);
2149
2150 if (fuzzy)
2151 pandecode_log_cont(" /* %u used */", truth);
2152
2153 pandecode_log_cont(",\n");
2154 }
2155
2156 static void
2157 pandecode_blend_shader_disassemble(mali_ptr shader, int job_no, int job_type,
2158 bool is_bifrost, unsigned gpu_id)
2159 {
2160 struct midgard_disasm_stats stats =
2161 pandecode_shader_disassemble(shader, job_no, job_type, is_bifrost, gpu_id);
2162
2163 bool has_texture = (stats.texture_count > 0);
2164 bool has_sampler = (stats.sampler_count > 0);
2165 bool has_attribute = (stats.attribute_count > 0);
2166 bool has_varying = (stats.varying_count > 0);
2167 bool has_uniform = (stats.uniform_count > 0);
2168 bool has_ubo = (stats.uniform_buffer_count > 0);
2169
2170 if (has_texture || has_sampler)
2171 pandecode_msg("XXX: blend shader accessing textures\n");
2172
2173 if (has_attribute || has_varying)
2174 pandecode_msg("XXX: blend shader accessing interstage\n");
2175
2176 if (has_uniform || has_ubo)
2177 pandecode_msg("XXX: blend shader accessing uniforms\n");
2178 }
2179
2180 static void
2181 pandecode_textures(mali_ptr textures, unsigned texture_count, int job_no, bool is_bifrost)
2182 {
2183 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(textures);
2184
2185 if (!mmem)
2186 return;
2187
2188 pandecode_log("Textures (%"PRIx64"):\n", textures);
2189
2190 if (is_bifrost) {
2191 const void *cl = pandecode_fetch_gpu_mem(mmem,
2192 textures, MALI_BIFROST_TEXTURE_LENGTH *
2193 texture_count);
2194
2195 for (unsigned tex = 0; tex < texture_count; ++tex) {
2196 pandecode_bifrost_texture(cl +
2197 MALI_BIFROST_TEXTURE_LENGTH * tex,
2198 job_no, tex);
2199 }
2200 } else {
2201 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures);
2202
2203 for (int tex = 0; tex < texture_count; ++tex) {
2204 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
2205 char *a = pointer_as_memory_reference(*u);
2206 pandecode_log("%s,\n", a);
2207 free(a);
2208 }
2209
2210 /* Now, finally, descend down into the texture descriptor */
2211 for (unsigned tex = 0; tex < texture_count; ++tex) {
2212 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
2213 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2214 if (tmem)
2215 pandecode_texture(*u, tmem, job_no, tex);
2216 }
2217 }
2218 }
2219
2220 static void
2221 pandecode_samplers(mali_ptr samplers, unsigned sampler_count, int job_no, bool is_bifrost)
2222 {
2223 for (int i = 0; i < sampler_count; ++i) {
2224 if (is_bifrost) {
2225 DUMP_ADDR("Sampler", BIFROST_SAMPLER, samplers + (MALI_BIFROST_SAMPLER_LENGTH * i), 1);
2226 } else {
2227 DUMP_ADDR("Sampler", MIDGARD_SAMPLER, samplers + (MALI_MIDGARD_SAMPLER_LENGTH * i), 1);
2228 }
2229 }
2230 }
2231
2232 static void
2233 pandecode_vertex_tiler_postfix_pre(
2234 const struct mali_vertex_tiler_postfix *p,
2235 int job_no, enum mali_job_type job_type,
2236 char *suffix, bool is_bifrost, unsigned gpu_id)
2237 {
2238 struct pandecode_mapped_memory *attr_mem;
2239
2240 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2241 * are the only things actually needed from the FBD, vertex/tiler jobs
2242 * no longer reference the FBD -- instead, this field points to some
2243 * info about the scratchpad.
2244 */
2245
2246 struct pandecode_fbd fbd_info = {
2247 /* Default for Bifrost */
2248 .rt_count = 1
2249 };
2250
2251 if (is_bifrost) {
2252 pandecode_log_cont("\t/* %X %/\n", p->shared_memory & 1);
2253 pandecode_compute_fbd(p->shared_memory & ~1, job_no);
2254 } else if (p->shared_memory & MALI_MFBD)
2255 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->shared_memory) & FBD_MASK, job_no, false, job_type == MALI_JOB_TYPE_COMPUTE, false);
2256 else if (job_type == MALI_JOB_TYPE_COMPUTE)
2257 pandecode_compute_fbd((u64) (uintptr_t) p->shared_memory, job_no);
2258 else
2259 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->shared_memory, job_no, false, gpu_id);
2260
2261 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
2262 int texture_count = 0, sampler_count = 0;
2263
2264 if (p->shader) {
2265 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->shader);
2266 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, p->shader);
2267
2268 /* Disassemble ahead-of-time to get stats. Initialize with
2269 * stats for the missing-shader case so we get validation
2270 * there, too */
2271
2272 struct midgard_disasm_stats info = {
2273 .texture_count = 0,
2274 .sampler_count = 0,
2275 .attribute_count = 0,
2276 .varying_count = 0,
2277 .work_count = 1,
2278
2279 .uniform_count = -128,
2280 .uniform_buffer_count = 0
2281 };
2282
2283 if (s->shader & ~0xF)
2284 info = pandecode_shader_disassemble(s->shader & ~0xF, job_no, job_type, is_bifrost, gpu_id);
2285
2286 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", p->shader, job_no, suffix);
2287 pandecode_indent++;
2288
2289 /* Save for dumps */
2290 attribute_count = s->attribute_count;
2291 varying_count = s->varying_count;
2292 texture_count = s->texture_count;
2293 sampler_count = s->sampler_count;
2294
2295 if (is_bifrost) {
2296 uniform_count = s->bifrost2.uniform_count;
2297 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
2298 } else {
2299 uniform_count = s->midgard1.uniform_count;
2300 uniform_buffer_count = s->midgard1.uniform_buffer_count;
2301 }
2302
2303 pandecode_shader_address("shader", s->shader);
2304
2305 pandecode_shader_prop("texture_count", s->texture_count, info.texture_count, false);
2306 pandecode_shader_prop("sampler_count", s->sampler_count, info.sampler_count, false);
2307 pandecode_shader_prop("attribute_count", s->attribute_count, info.attribute_count, false);
2308 pandecode_shader_prop("varying_count", s->varying_count, info.varying_count, false);
2309 pandecode_shader_prop("uniform_buffer_count",
2310 uniform_buffer_count,
2311 info.uniform_buffer_count, true);
2312
2313 if (!is_bifrost) {
2314 pandecode_shader_prop("uniform_count",
2315 uniform_count,
2316 info.uniform_count, false);
2317
2318 pandecode_shader_prop("work_count",
2319 s->midgard1.work_count, info.work_count, false);
2320 }
2321
2322 if (is_bifrost) {
2323 pandecode_log("bifrost1.unk1 = ");
2324 pandecode_log_decoded_flags(shader_bifrost_info, s->bifrost1.unk1);
2325 pandecode_log_cont(",\n");
2326 } else {
2327 bool helpers = s->midgard1.flags_lo & MALI_HELPER_INVOCATIONS;
2328
2329 if (helpers != info.helper_invocations) {
2330 pandecode_msg("XXX: expected helpers %u but got %u\n",
2331 info.helper_invocations, helpers);
2332 }
2333
2334 pandecode_log(".midgard1.flags_lo = ");
2335 pandecode_log_decoded_flags(shader_midgard1_flag_lo_info,
2336 s->midgard1.flags_lo & ~MALI_HELPER_INVOCATIONS);
2337 pandecode_log_cont(",\n");
2338
2339 pandecode_log(".midgard1.flags_hi = ");
2340 pandecode_log_decoded_flags(shader_midgard1_flag_hi_info, s->midgard1.flags_hi);
2341 pandecode_log_cont(",\n");
2342 }
2343
2344 if (s->depth_units || s->depth_factor) {
2345 pandecode_prop("depth_factor = %f", s->depth_factor);
2346 pandecode_prop("depth_units = %f", s->depth_units);
2347 }
2348
2349 if (s->coverage_mask)
2350 pandecode_prop("coverage_mask = 0x%X", s->coverage_mask);
2351
2352 if (s->unknown2_2)
2353 pandecode_prop(".unknown2_2 = %X", s->unknown2_2);
2354
2355 if (s->unknown2_3 || s->unknown2_4) {
2356 pandecode_log(".unknown2_3 = ");
2357
2358 int unknown2_3 = s->unknown2_3;
2359 int unknown2_4 = s->unknown2_4;
2360
2361 /* We're not quite sure what these flags mean without the depth test, if anything */
2362
2363 if (unknown2_3 & (MALI_DEPTH_WRITEMASK | MALI_DEPTH_FUNC_MASK)) {
2364 const char *func = mali_func_as_str(MALI_GET_DEPTH_FUNC(unknown2_3));
2365 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
2366
2367 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
2368 }
2369
2370 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
2371 pandecode_log_cont(",\n");
2372
2373 pandecode_log(".unknown2_4 = ");
2374 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
2375 pandecode_log_cont(",\n");
2376 }
2377
2378 if (s->stencil_mask_front || s->stencil_mask_back) {
2379 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
2380 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
2381 }
2382
2383 DUMP_CL("Stencil front", STENCIL, &s->stencil_front, 1);
2384 DUMP_CL("Stencil back", STENCIL, &s->stencil_back, 1);
2385
2386 if (is_bifrost) {
2387 pandecode_log(".bifrost2 = {\n");
2388 pandecode_indent++;
2389
2390 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
2391 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
2392 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
2393 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
2394
2395 pandecode_indent--;
2396 pandecode_log("},\n");
2397 } else if (s->midgard2.unknown2_7) {
2398 pandecode_log(".midgard2 = {\n");
2399 pandecode_indent++;
2400
2401 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
2402 pandecode_indent--;
2403 pandecode_log("},\n");
2404 }
2405
2406 if (s->padding) {
2407 pandecode_msg("XXX: shader padding tripped\n");
2408 pandecode_prop("padding = 0x%" PRIx32, s->padding);
2409 }
2410
2411 if (!is_bifrost) {
2412 /* TODO: Blend shaders routing/disasm */
2413 union midgard_blend blend = s->blend;
2414 mali_ptr shader = pandecode_midgard_blend(&blend, s->unknown2_3 & MALI_HAS_BLEND_SHADER);
2415 if (shader & ~0xF)
2416 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2417 } else {
2418 pandecode_msg("mdg_blend = %" PRIx64 "\n", s->blend.shader);
2419 }
2420
2421 pandecode_indent--;
2422 pandecode_log("};\n");
2423
2424 /* MRT blend fields are used whenever MFBD is used, with
2425 * per-RT descriptors */
2426
2427 if (job_type == MALI_JOB_TYPE_TILER && (is_bifrost || p->shared_memory & MALI_MFBD)) {
2428 void* blend_base = (void *) (s + 1);
2429
2430 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
2431 mali_ptr shader = 0;
2432
2433 if (is_bifrost)
2434 shader = pandecode_bifrost_blend(blend_base, job_no, i);
2435 else
2436 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
2437
2438 if (shader & ~0xF)
2439 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2440
2441 }
2442 }
2443 } else
2444 pandecode_msg("XXX: missing shader descriptor\n");
2445
2446 if (p->viewport)
2447 DUMP_ADDR("Viewport", VIEWPORT, p->viewport, 1);
2448
2449 unsigned max_attr_index = 0;
2450
2451 if (p->attribute_meta)
2452 max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
2453
2454 if (p->attributes) {
2455 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
2456 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false, job_type);
2457 }
2458
2459 /* Varyings are encoded like attributes but not actually sent; we just
2460 * pass a zero buffer with the right stride/size set, (or whatever)
2461 * since the GPU will write to it itself */
2462
2463 if (p->varying_meta) {
2464 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
2465 }
2466
2467 if (p->varyings) {
2468 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
2469
2470 /* Number of descriptors depends on whether there are
2471 * non-internal varyings */
2472
2473 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true, job_type);
2474 }
2475
2476 if (p->uniform_buffers) {
2477 if (uniform_buffer_count)
2478 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
2479 else
2480 pandecode_msg("warn: UBOs specified but not referenced\n");
2481 } else if (uniform_buffer_count)
2482 pandecode_msg("XXX: UBOs referenced but not specified\n");
2483
2484 /* We don't want to actually dump uniforms, but we do need to validate
2485 * that the counts we were given are sane */
2486
2487 if (p->uniforms) {
2488 if (uniform_count)
2489 pandecode_uniforms(p->uniforms, uniform_count);
2490 else
2491 pandecode_msg("warn: Uniforms specified but not referenced\n");
2492 } else if (uniform_count)
2493 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2494
2495 if (p->textures)
2496 pandecode_textures(p->textures, texture_count, job_no, is_bifrost);
2497
2498 if (p->sampler_descriptor)
2499 pandecode_samplers(p->sampler_descriptor, sampler_count, job_no, is_bifrost);
2500 }
2501
2502 static void
2503 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2504 {
2505 pandecode_log(".gl_enables = ");
2506
2507 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2508
2509 pandecode_log_cont(",\n");
2510 }
2511
2512 static void
2513 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2514 {
2515 if (p->shader & 0xF)
2516 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p->shader & 0xF));
2517
2518 pandecode_log(".postfix = {\n");
2519 pandecode_indent++;
2520
2521 pandecode_gl_enables(p->gl_enables, MALI_JOB_TYPE_TILER);
2522 pandecode_prop("instance_shift = 0x%x", p->instance_shift);
2523 pandecode_prop("instance_odd = 0x%x", p->instance_odd);
2524
2525 if (p->zero4) {
2526 pandecode_msg("XXX: vertex only zero tripped");
2527 pandecode_prop("zero4 = 0x%" PRIx32, p->zero4);
2528 }
2529
2530 pandecode_prop("offset_start = 0x%x", p->offset_start);
2531
2532 if (p->zero5) {
2533 pandecode_msg("XXX: vertex only zero tripped");
2534 pandecode_prop("zero5 = 0x%" PRIx64, p->zero5);
2535 }
2536
2537 MEMORY_PROP(p, position_varying);
2538 MEMORY_PROP(p, occlusion_counter);
2539
2540 pandecode_indent--;
2541 pandecode_log("},\n");
2542 }
2543
2544 static void
2545 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2546 {
2547 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2548 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2549
2550 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
2551 pandecode_indent++;
2552
2553 if (h->zero) {
2554 pandecode_msg("XXX: tiler heap zero tripped\n");
2555 pandecode_prop("zero = 0x%x", h->zero);
2556 }
2557
2558 pandecode_prop("heap_size = 0x%x", h->heap_size);
2559 MEMORY_PROP(h, tiler_heap_start);
2560 MEMORY_PROP(h, tiler_heap_free);
2561
2562 /* this might point to the beginning of another buffer, when it's
2563 * really the end of the tiler heap buffer, so we have to be careful
2564 * here. but for zero length, we need the same pointer.
2565 */
2566
2567 if (h->tiler_heap_end == h->tiler_heap_start) {
2568 MEMORY_PROP(h, tiler_heap_start);
2569 } else {
2570 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2571 pandecode_prop("tiler_heap_end = %s + 1", a);
2572 free(a);
2573 }
2574
2575 for (int i = 0; i < 10; i++) {
2576 if (h->zeros[i] != 0) {
2577 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2578 i, h->zeros[i]);
2579 }
2580 }
2581
2582 if (h->unk1 != 0x1) {
2583 pandecode_msg("XXX: tiler heap unk1 tripped\n");
2584 pandecode_prop("unk1 = 0x%x", h->unk1);
2585 }
2586
2587 if (h->unk7e007e != 0x7e007e) {
2588 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
2589 pandecode_prop("unk7e007e = 0x%x", h->unk7e007e);
2590 }
2591
2592 pandecode_indent--;
2593 pandecode_log("};\n");
2594 }
2595
2596 static void
2597 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2598 {
2599 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2600 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2601
2602 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2603
2604 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
2605 pandecode_indent++;
2606
2607 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32, t->tiler_heap_next_start);
2608 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32, t->used_hierarchy_mask);
2609
2610 if (t->hierarchy_mask != 0xa &&
2611 t->hierarchy_mask != 0x14 &&
2612 t->hierarchy_mask != 0x28 &&
2613 t->hierarchy_mask != 0x50 &&
2614 t->hierarchy_mask != 0xa0)
2615 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
2616
2617 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2618
2619 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2620
2621 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2622 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2623
2624 if (t->zero0) {
2625 pandecode_msg("XXX: tiler meta zero tripped\n");
2626 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2627 }
2628
2629 for (int i = 0; i < 12; i++) {
2630 if (t->zeros[i] != 0) {
2631 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2632 i, t->zeros[i]);
2633 }
2634 }
2635
2636 pandecode_indent--;
2637 pandecode_log("};\n");
2638 }
2639
2640 static void
2641 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2642 {
2643 if (u.pointer == 0x0)
2644 return;
2645
2646 pandecode_log(".primitive_size = {\n");
2647 pandecode_indent++;
2648
2649 if (constant) {
2650 pandecode_prop("constant = %f", u.constant);
2651 } else {
2652 MEMORY_PROP((&u), pointer);
2653 }
2654
2655 pandecode_indent--;
2656 pandecode_log("},\n");
2657 }
2658
2659 static void
2660 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2661 {
2662 pandecode_log_cont("{\n");
2663 pandecode_indent++;
2664
2665 /* TODO: gl_PointSize on Bifrost */
2666 pandecode_primitive_size(t->primitive_size, true);
2667
2668 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2669 || t->zero6) {
2670 pandecode_msg("XXX: tiler only zero tripped\n");
2671 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2672 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2673 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2674 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2675 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2676 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2677 }
2678
2679 pandecode_indent--;
2680 pandecode_log("},\n");
2681 }
2682
2683 static int
2684 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2685 const struct pandecode_mapped_memory *mem,
2686 mali_ptr payload, int job_no, unsigned gpu_id)
2687 {
2688 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2689
2690 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true, gpu_id);
2691
2692 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64"_%d = {\n", payload, job_no);
2693 pandecode_indent++;
2694
2695 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2696 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2697
2698 pandecode_indent--;
2699 pandecode_log("};\n");
2700
2701 return sizeof(*v);
2702 }
2703
2704 static int
2705 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2706 const struct pandecode_mapped_memory *mem,
2707 mali_ptr payload, int job_no, unsigned gpu_id)
2708 {
2709 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2710
2711 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true, gpu_id);
2712 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2713
2714 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64"_%d = {\n", payload, job_no);
2715 pandecode_indent++;
2716
2717 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
2718
2719 pandecode_log(".tiler = ");
2720 pandecode_tiler_only_bfr(&t->tiler, job_no);
2721
2722 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2723
2724 pandecode_indent--;
2725 pandecode_log("};\n");
2726
2727 return sizeof(*t);
2728 }
2729
2730 static int
2731 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2732 const struct pandecode_mapped_memory *mem,
2733 mali_ptr payload, int job_no, unsigned gpu_id)
2734 {
2735 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2736 bool is_graphics = (h->job_type == MALI_JOB_TYPE_VERTEX) || (h->job_type == MALI_JOB_TYPE_TILER);
2737
2738 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false, gpu_id);
2739
2740 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2741 pandecode_indent++;
2742
2743 pandecode_vertex_tiler_prefix(&v->prefix, job_no, is_graphics);
2744 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2745
2746 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2747 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
2748
2749 pandecode_indent--;
2750 pandecode_log("};\n");
2751
2752 return sizeof(*v);
2753 }
2754
2755 static int
2756 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2757 mali_ptr payload, int job_no,
2758 bool is_bifrost, unsigned gpu_id)
2759 {
2760 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2761
2762 bool is_mfbd = s->framebuffer & MALI_MFBD;
2763
2764 if (!is_mfbd && is_bifrost)
2765 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2766
2767 struct pandecode_fbd info;
2768
2769 if (is_mfbd)
2770 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true, false, is_bifrost);
2771 else
2772 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true, gpu_id);
2773
2774 /* Compute the tag for the tagged pointer. This contains the type of
2775 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2776 * additional structures follow the MFBD header (an extra payload or
2777 * not, as well as a count of render targets) */
2778
2779 unsigned expected_tag = is_mfbd ? MALI_MFBD : 0;
2780
2781 if (is_mfbd) {
2782 if (info.has_extra)
2783 expected_tag |= MALI_MFBD_TAG_EXTRA;
2784
2785 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
2786 }
2787
2788 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
2789 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2790 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
2791 pandecode_prop("max_tile_coord = 0x%X\n", s->max_tile_coord);
2792 }
2793
2794 /* Extract tile coordinates */
2795
2796 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
2797 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
2798
2799 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2800 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2801
2802 /* For the max, we also want the floored (rather than ceiled) version for checking */
2803
2804 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
2805 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
2806
2807 /* Validate the coordinates are well-ordered */
2808
2809 if (min_x == max_x)
2810 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2811 else if (min_x > max_x)
2812 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2813
2814 if (min_y == max_y)
2815 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2816 else if (min_y > max_y)
2817 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2818
2819 /* Validate the coordinates fit inside the framebuffer. We use floor,
2820 * rather than ceil, for the max coordinates, since the tile
2821 * coordinates for something like an 800x600 framebuffer will actually
2822 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2823
2824 if ((min_x > info.width) || (max_x_f > info.width))
2825 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2826
2827 if ((min_y > info.height) || (max_y_f > info.height))
2828 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2829
2830 /* After validation, we print */
2831
2832 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
2833
2834 /* The FBD is a tagged pointer */
2835
2836 unsigned tag = (s->framebuffer & ~FBD_MASK);
2837
2838 if (tag != expected_tag)
2839 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
2840
2841 return sizeof(*s);
2842 }
2843
2844 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
2845 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
2846 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
2847 * might wish to combine this with the bifrost parameter) because some details
2848 * are model-specific even within a particular architecture. Minimal traces
2849 * *only* examine the job descriptors, skipping printing entirely if there is
2850 * no faults, and only descends into the payload if there are faults. This is
2851 * useful for looking for faults without the overhead of invasive traces. */
2852
2853 void
2854 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost, unsigned gpu_id, bool minimal)
2855 {
2856 pandecode_dump_file_open();
2857
2858 struct mali_job_descriptor_header *h;
2859 unsigned job_descriptor_number = 0;
2860
2861 do {
2862 struct pandecode_mapped_memory *mem =
2863 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2864
2865 void *payload;
2866
2867 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2868
2869 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2870 * high 32-bits of the 64-bit pointer are reused to store
2871 * something else.
2872 */
2873 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2874 h->job_type != MALI_JOB_TYPE_FRAGMENT ? 4 : 0;
2875 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2876
2877 payload = pandecode_fetch_gpu_mem(mem, payload_ptr, 256);
2878
2879 int job_no = job_descriptor_number++;
2880
2881 /* If the job is good to go, skip it in minimal mode */
2882 if (minimal && (h->exception_status == 0x0 || h->exception_status == 0x1))
2883 continue;
2884
2885 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2886 pandecode_indent++;
2887
2888 pandecode_prop("job_type = %s", mali_job_type_as_str(h->job_type));
2889
2890 if (h->job_descriptor_size)
2891 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2892
2893 if (h->exception_status && h->exception_status != 0x1)
2894 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2895 h->exception_status,
2896 (h->exception_status >> 16) & 0xFFFF,
2897 pandecode_exception_access((h->exception_status >> 8) & 0x3),
2898 h->exception_status & 0xFF);
2899
2900 if (h->first_incomplete_task)
2901 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2902
2903 if (h->fault_pointer)
2904 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2905
2906 if (h->job_barrier)
2907 pandecode_prop("job_barrier = %d", h->job_barrier);
2908
2909 pandecode_prop("job_index = %d", h->job_index);
2910
2911 if (h->unknown_flags)
2912 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2913
2914 if (h->job_dependency_index_1)
2915 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2916
2917 if (h->job_dependency_index_2)
2918 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2919
2920 pandecode_indent--;
2921 pandecode_log("};\n");
2922
2923 switch (h->job_type) {
2924 case MALI_JOB_TYPE_WRITE_VALUE: {
2925 struct mali_payload_write_value *s = payload;
2926 pandecode_log("struct mali_payload_write_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2927 pandecode_indent++;
2928 MEMORY_PROP(s, address);
2929
2930 if (s->value_descriptor != MALI_WRITE_VALUE_ZERO) {
2931 pandecode_msg("XXX: unknown value descriptor\n");
2932 pandecode_prop("value_descriptor = 0x%" PRIX32, s->value_descriptor);
2933 }
2934
2935 if (s->reserved) {
2936 pandecode_msg("XXX: set value tripped\n");
2937 pandecode_prop("reserved = 0x%" PRIX32, s->reserved);
2938 }
2939
2940 pandecode_prop("immediate = 0x%" PRIX64, s->immediate);
2941 pandecode_indent--;
2942 pandecode_log("};\n");
2943
2944 break;
2945 }
2946
2947 case MALI_JOB_TYPE_TILER:
2948 case MALI_JOB_TYPE_VERTEX:
2949 case MALI_JOB_TYPE_COMPUTE:
2950 if (bifrost) {
2951 if (h->job_type == MALI_JOB_TYPE_TILER)
2952 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2953 else
2954 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2955 } else
2956 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no, gpu_id);
2957
2958 break;
2959
2960 case MALI_JOB_TYPE_FRAGMENT:
2961 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost, gpu_id);
2962 break;
2963
2964 default:
2965 break;
2966 }
2967 } while ((jc_gpu_va = h->next_job));
2968
2969 pandecode_map_read_write();
2970 }