projects / mesa.git / blob
? search:


c47b8f52b6cc5899f5e299ef9107f4f331cea1a2
[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_EXCEPTION_ACCESS_## name: return ""#name
514 static char *
515 pandecode_exception_access(unsigned access)
516 {
517 switch (access) {
518 DEFINE_CASE(NONE);
519 DEFINE_CASE(EXECUTE);
520 DEFINE_CASE(READ);
521 DEFINE_CASE(WRITE);
522
523 default:
524 unreachable("Invalid case");
525 }
526 }
527 #undef DEFINE_CASE
528
529 /* Midgard's tiler descriptor is embedded within the
530 * larger FBD */
531
532 static void
533 pandecode_midgard_tiler_descriptor(
534 const struct midgard_tiler_descriptor *t,
535 unsigned width,
536 unsigned height,
537 bool is_fragment,
538 bool has_hierarchy)
539 {
540 pandecode_log(".tiler = {\n");
541 pandecode_indent++;
542
543 if (t->hierarchy_mask == MALI_TILER_DISABLED)
544 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
545 else
546 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
547
548 /* We know this name from the kernel, but we never see it nonzero */
549
550 if (t->flags)
551 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
552
553 MEMORY_PROP(t, polygon_list);
554
555 /* The body is offset from the base of the polygon list */
556 //assert(t->polygon_list_body > t->polygon_list);
557 unsigned body_offset = t->polygon_list_body - t->polygon_list;
558
559 /* It needs to fit inside the reported size */
560 //assert(t->polygon_list_size >= body_offset);
561
562 /* Now that we've sanity checked, we'll try to calculate the sizes
563 * ourselves for comparison */
564
565 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask, has_hierarchy);
566 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask, has_hierarchy);
567
568 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
569 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
570 ref_header, ref_size);
571 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
572 pandecode_msg("body offset %d\n", body_offset);
573 }
574
575 /* The tiler heap has a start and end specified -- it should be
576 * identical to what we have in the BO. The exception is if tiling is
577 * disabled. */
578
579 MEMORY_PROP(t, heap_start);
580 assert(t->heap_end >= t->heap_start);
581
582 struct pandecode_mapped_memory *heap =
583 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
584
585 unsigned heap_size = t->heap_end - t->heap_start;
586
587 /* Tiling is enabled with a special flag */
588 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
589 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
590
591 bool tiling_enabled = hierarchy_mask;
592
593 if (tiling_enabled) {
594 /* When tiling is enabled, the heap should be a tight fit */
595 unsigned heap_offset = t->heap_start - heap->gpu_va;
596 if ((heap_offset + heap_size) != heap->length) {
597 pandecode_msg("XXX: heap size %u (expected %zu)\n",
598 heap_size, heap->length - heap_offset);
599 }
600
601 /* We should also have no other flags */
602 if (tiler_flags)
603 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
604 } else {
605 /* When tiling is disabled, we should have that flag and no others */
606
607 if (tiler_flags != MALI_TILER_DISABLED) {
608 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
609 tiler_flags);
610 }
611
612 /* We should also have an empty heap */
613 if (heap_size) {
614 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
615 heap_size);
616 }
617
618 /* Disabled tiling is used only for clear-only jobs, which are
619 * purely FRAGMENT, so we should never see this for
620 * non-FRAGMENT descriptors. */
621
622 if (!is_fragment)
623 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
624 }
625
626 /* We've never seen weights used in practice, but we know from the
627 * kernel these fields is there */
628
629 bool nonzero_weights = false;
630
631 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
632 nonzero_weights |= t->weights[w] != 0x0;
633 }
634
635 if (nonzero_weights) {
636 pandecode_log(".weights = { ");
637
638 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
639 pandecode_log_cont("%d, ", t->weights[w]);
640 }
641
642 pandecode_log("},");
643 }
644
645 pandecode_indent--;
646 pandecode_log("}\n");
647 }
648
649 /* TODO: The Bifrost tiler is not understood at all yet */
650
651 static void
652 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer *fb)
653 {
654 pandecode_log(".tiler = {\n");
655 pandecode_indent++;
656
657 MEMORY_PROP(fb, tiler_meta);
658
659 for (int i = 0; i < 16; i++) {
660 if (fb->zeros[i] != 0) {
661 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
662 i, fb->zeros[i]);
663 }
664 }
665
666 pandecode_log("},\n");
667
668 pandecode_indent--;
669 pandecode_log("}\n");
670
671 }
672
673 /* Information about the framebuffer passed back for
674 * additional analysis */
675
676 struct pandecode_fbd {
677 unsigned width;
678 unsigned height;
679 unsigned rt_count;
680 bool has_extra;
681 };
682
683 static void
684 pandecode_sfbd_format(struct mali_sfbd_format format)
685 {
686 pandecode_log(".format = {\n");
687 pandecode_indent++;
688
689 pandecode_log(".unk1 = ");
690 pandecode_log_decoded_flags(sfbd_unk1_info, format.unk1);
691 pandecode_log_cont(",\n");
692
693 /* TODO: Map formats so we can check swizzles and print nicely */
694 pandecode_log("swizzle");
695 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
696 pandecode_log_cont(",\n");
697
698 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
699 (format.nr_channels + 1));
700
701 pandecode_log(".unk2 = ");
702 pandecode_log_decoded_flags(sfbd_unk2_info, format.unk2);
703 pandecode_log_cont(",\n");
704
705 pandecode_prop("block = %s", mali_block_format_as_str(format.block));
706
707 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
708
709 pandecode_indent--;
710 pandecode_log("},\n");
711 }
712
713 static void
714 pandecode_shared_memory(const struct mali_shared_memory *desc, bool is_compute)
715 {
716 pandecode_prop("stack_shift = 0x%x", desc->stack_shift);
717
718 if (desc->unk0)
719 pandecode_prop("unk0 = 0x%x", desc->unk0);
720
721 if (desc->shared_workgroup_count != 0x1F) {
722 pandecode_prop("shared_workgroup_count = %d", desc->shared_workgroup_count);
723 if (!is_compute)
724 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
725 }
726
727 if (desc->shared_unk1 || desc->shared_shift) {
728 pandecode_prop("shared_unk1 = %X", desc->shared_unk1);
729 pandecode_prop("shared_shift = %X", desc->shared_shift);
730
731 if (!is_compute)
732 pandecode_msg("XXX: shared memory configured in noncompute shader");
733 }
734
735 if (desc->shared_zero) {
736 pandecode_msg("XXX: shared memory zero tripped\n");
737 pandecode_prop("shared_zero = 0x%" PRIx32, desc->shared_zero);
738 }
739
740 if (desc->shared_memory && !is_compute)
741 pandecode_msg("XXX: shared memory used in noncompute shader\n");
742
743 MEMORY_PROP(desc, scratchpad);
744 MEMORY_PROP(desc, shared_memory);
745 MEMORY_PROP(desc, unknown1);
746
747 if (desc->scratchpad) {
748 struct pandecode_mapped_memory *smem =
749 pandecode_find_mapped_gpu_mem_containing(desc->scratchpad);
750
751 pandecode_msg("scratchpad size %u\n", smem->length);
752 }
753
754 }
755
756 static struct pandecode_fbd
757 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment, unsigned gpu_id)
758 {
759 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
760 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
761
762 struct pandecode_fbd info = {
763 .has_extra = false,
764 .rt_count = 1
765 };
766
767 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
768 pandecode_indent++;
769
770 pandecode_log(".shared_memory = {\n");
771 pandecode_indent++;
772 pandecode_shared_memory(&s->shared_memory, false);
773 pandecode_indent--;
774 pandecode_log("},\n");
775
776 pandecode_sfbd_format(s->format);
777
778 info.width = s->width + 1;
779 info.height = s->height + 1;
780
781 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
782 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
783
784 MEMORY_PROP(s, checksum);
785
786 if (s->checksum_stride)
787 pandecode_prop("checksum_stride = %d", s->checksum_stride);
788
789 MEMORY_PROP(s, framebuffer);
790 pandecode_prop("stride = %d", s->stride);
791
792 /* Earlier in the actual commandstream -- right before width -- but we
793 * delay to flow nicer */
794
795 pandecode_log(".clear_flags = ");
796 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
797 pandecode_log_cont(",\n");
798
799 if (s->depth_buffer) {
800 MEMORY_PROP(s, depth_buffer);
801 pandecode_prop("depth_stride = %d", s->depth_stride);
802 }
803
804 if (s->stencil_buffer) {
805 MEMORY_PROP(s, stencil_buffer);
806 pandecode_prop("stencil_stride = %d", s->stencil_stride);
807 }
808
809 if (s->depth_stride_zero ||
810 s->stencil_stride_zero ||
811 s->zero7 || s->zero8) {
812 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
813 pandecode_prop("depth_stride_zero = 0x%x",
814 s->depth_stride_zero);
815 pandecode_prop("stencil_stride_zero = 0x%x",
816 s->stencil_stride_zero);
817 pandecode_prop("zero7 = 0x%" PRIx32,
818 s->zero7);
819 pandecode_prop("zero8 = 0x%" PRIx32,
820 s->zero8);
821 }
822
823 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
824 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
825 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
826 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
827 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
828 }
829
830 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
831 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
832 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
833 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
834 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
835 }
836
837 if (s->clear_stencil) {
838 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
839 }
840
841 const struct midgard_tiler_descriptor t = s->tiler;
842
843 bool has_hierarchy = !(gpu_id == 0x0720 || gpu_id == 0x0820 || gpu_id == 0x0830);
844 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment, has_hierarchy);
845
846 pandecode_indent--;
847 pandecode_log("};\n");
848
849 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
850 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
851 pandecode_prop("zero5 = 0x%" PRIx32, s->zero5);
852
853 pandecode_log_cont(".zero3 = {");
854
855 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
856 pandecode_log_cont("%X, ", s->zero3[i]);
857
858 pandecode_log_cont("},\n");
859
860 pandecode_log_cont(".zero6 = {");
861
862 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
863 pandecode_log_cont("%X, ", s->zero6[i]);
864
865 pandecode_log_cont("},\n");
866
867 return info;
868 }
869
870 static void
871 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
872 {
873 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
874 const struct mali_shared_memory *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
875
876 pandecode_log("struct mali_shared_memory shared_%"PRIx64"_%d = {\n", gpu_va, job_no);
877 pandecode_indent++;
878 pandecode_shared_memory(s, true);
879 pandecode_indent--;
880 pandecode_log("},\n");
881 }
882
883 /* Extracts the number of components associated with a Mali format */
884
885 static unsigned
886 pandecode_format_component_count(enum mali_format fmt)
887 {
888 /* Mask out the format class */
889 unsigned top = fmt & 0b11100000;
890
891 switch (top) {
892 case MALI_FORMAT_SNORM:
893 case MALI_FORMAT_UINT:
894 case MALI_FORMAT_UNORM:
895 case MALI_FORMAT_SINT:
896 return ((fmt >> 3) & 3) + 1;
897 default:
898 /* TODO: Validate */
899 return 4;
900 }
901 }
902
903 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
904
905 static unsigned
906 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
907 {
908 unsigned mask = 0;
909 assert(MALI_CHANNEL_RED == 0);
910
911 for (unsigned c = 0; c < 4; ++c) {
912 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
913
914 if (chan <= MALI_CHANNEL_ALPHA)
915 mask |= (1 << chan);
916 }
917
918 return mask;
919 }
920
921 /* Validates that a (format, swizzle) pair is valid, in the sense that the
922 * swizzle doesn't access any components that are undefined in the format.
923 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
924 * omitted */
925
926 static bool
927 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
928 {
929 unsigned nr_comp = pandecode_format_component_count(fmt);
930 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
931 unsigned valid_mask = (1 << nr_comp) - 1;
932 unsigned invalid_mask = ~valid_mask;
933
934 if (access_mask & invalid_mask) {
935 pandecode_msg("XXX: invalid components accessed\n");
936 return false;
937 }
938
939 /* Check for the default non-swizzling swizzle so we can suppress
940 * useless printing for the defaults */
941
942 unsigned default_swizzles[4] = {
943 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
944 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
945 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
946 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
947 };
948
949 return (swizzle == default_swizzles[nr_comp - 1]);
950 }
951
952 /* Maps MALI_RGBA32F to rgba32f, etc */
953
954 static void
955 pandecode_format_short(enum mali_format fmt, bool srgb)
956 {
957 /* We want a type-like format, so cut off the initial MALI_ */
958 char *format = pandecode_format(fmt);
959 format += strlen("MALI_");
960
961 unsigned len = strlen(format);
962 char *lower_format = calloc(1, len + 1);
963
964 for (unsigned i = 0; i < len; ++i)
965 lower_format[i] = tolower(format[i]);
966
967 /* Sanity check sRGB flag is applied to RGB, per the name */
968 if (srgb && lower_format[0] != 'r')
969 pandecode_msg("XXX: sRGB applied to non-colour format\n");
970
971 /* Just prefix with an s, so you get formats like srgba8_unorm */
972 if (srgb)
973 pandecode_log_cont("s");
974
975 pandecode_log_cont("%s", lower_format);
976 free(lower_format);
977 }
978
979 static void
980 pandecode_swizzle(unsigned swizzle, enum mali_format format)
981 {
982 /* First, do some validation */
983 bool trivial_swizzle = pandecode_validate_format_swizzle(
984 format, swizzle);
985
986 if (trivial_swizzle)
987 return;
988
989 /* Next, print the swizzle */
990 pandecode_log_cont(".");
991
992 static const char components[] = "rgba01";
993
994 for (unsigned c = 0; c < 4; ++c) {
995 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
996
997 if (chan >= MALI_CHANNEL_RESERVED_0) {
998 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
999 continue;
1000 }
1001 pandecode_log_cont("%c", components[chan]);
1002 }
1003 }
1004
1005 static void
1006 pandecode_rt_format(struct mali_rt_format format)
1007 {
1008 pandecode_log(".format = {\n");
1009 pandecode_indent++;
1010
1011 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
1012 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
1013 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
1014 pandecode_prop("unk4 = 0x%" PRIx32, format.unk4);
1015
1016 pandecode_prop("block = %s", mali_block_format_as_str(format.block));
1017
1018 /* TODO: Map formats so we can check swizzles and print nicely */
1019 pandecode_log("swizzle");
1020 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
1021 pandecode_log_cont(",\n");
1022
1023 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
1024 (format.nr_channels + 1));
1025
1026 pandecode_log(".flags = ");
1027 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
1028 pandecode_log_cont(",\n");
1029
1030 pandecode_prop("msaa = %s", pandecode_msaa_mode(format.msaa));
1031
1032 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
1033 * which is a faster hardware-based alternative to the wallpaper method
1034 * to preserve framebuffer contents across frames. In practice, MFBD
1035 * preload is buggy on Midgard, and so this is a chicken bit. If this
1036 * bit isn't set, most likely something broke unrelated to preload */
1037
1038 if (!format.no_preload) {
1039 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
1040 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
1041 }
1042
1043 if (format.zero)
1044 pandecode_prop("zero = 0x%" PRIx32, format.zero);
1045
1046 pandecode_indent--;
1047 pandecode_log("},\n");
1048 }
1049
1050 static void
1051 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct mali_framebuffer *fb)
1052 {
1053 pandecode_log("struct mali_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
1054 pandecode_indent++;
1055
1056 for (int i = 0; i < (fb->rt_count_1 + 1); i++) {
1057 mali_ptr rt_va = gpu_va + i * sizeof(struct mali_render_target);
1058 struct pandecode_mapped_memory *mem =
1059 pandecode_find_mapped_gpu_mem_containing(rt_va);
1060 const struct mali_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
1061
1062 pandecode_log("{\n");
1063 pandecode_indent++;
1064
1065 pandecode_rt_format(rt->format);
1066
1067 if (rt->format.block == MALI_BLOCK_FORMAT_AFBC) {
1068 pandecode_log(".afbc = {\n");
1069 pandecode_indent++;
1070
1071 char *a = pointer_as_memory_reference(rt->afbc.metadata);
1072 pandecode_prop("metadata = %s", a);
1073 free(a);
1074
1075 pandecode_prop("stride = %d", rt->afbc.stride);
1076
1077 pandecode_log(".flags = ");
1078 pandecode_log_decoded_flags(afbc_fmt_flag_info, rt->afbc.flags);
1079 pandecode_log_cont(",\n");
1080
1081 pandecode_indent--;
1082 pandecode_log("},\n");
1083 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.flags) {
1084 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1085 rt->afbc.metadata,
1086 rt->afbc.stride,
1087 rt->afbc.flags);
1088 }
1089
1090 MEMORY_PROP(rt, framebuffer);
1091 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
1092
1093 if (rt->layer_stride)
1094 pandecode_prop("layer_stride = %d", rt->layer_stride);
1095
1096 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
1097 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
1098 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
1099 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
1100 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
1101 }
1102
1103 if (rt->zero1 || rt->zero2) {
1104 pandecode_msg("XXX: render target zeros tripped\n");
1105 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
1106 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
1107 }
1108
1109 pandecode_indent--;
1110 pandecode_log("},\n");
1111 }
1112
1113 pandecode_indent--;
1114 pandecode_log("};\n");
1115 }
1116
1117 static struct pandecode_fbd
1118 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment, bool is_compute, bool is_bifrost)
1119 {
1120 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1121 const struct mali_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
1122
1123 struct pandecode_fbd info;
1124
1125 if (is_bifrost && fb->msaa.sample_locations) {
1126 /* The blob stores all possible sample locations in a single buffer
1127 * allocated on startup, and just switches the pointer when switching
1128 * MSAA state. For now, we just put the data into the cmdstream, but we
1129 * should do something like what the blob does with a real driver.
1130 *
1131 * There seem to be 32 slots for sample locations, followed by another
1132 * 16. The second 16 is just the center location followed by 15 zeros
1133 * in all the cases I've identified (maybe shader vs. depth/color
1134 * samples?).
1135 */
1136
1137 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->msaa.sample_locations);
1138
1139 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->msaa.sample_locations);
1140
1141 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
1142 pandecode_indent++;
1143
1144 for (int i = 0; i < 32 + 16; i++) {
1145 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
1146 }
1147
1148 pandecode_indent--;
1149 pandecode_log("};\n");
1150 }
1151
1152 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
1153 pandecode_indent++;
1154
1155 if (is_bifrost) {
1156 pandecode_log(".msaa = {\n");
1157 pandecode_indent++;
1158
1159 if (fb->msaa.sample_locations)
1160 pandecode_prop("sample_locations = sample_locations_%d", job_no);
1161 else
1162 pandecode_msg("XXX: sample_locations missing\n");
1163
1164 if (fb->msaa.zero1 || fb->msaa.zero2 || fb->msaa.zero4) {
1165 pandecode_msg("XXX: multisampling zero tripped\n");
1166 pandecode_prop("zero1 = %" PRIx64, fb->msaa.zero1);
1167 pandecode_prop("zero2 = %" PRIx64, fb->msaa.zero2);
1168 pandecode_prop("zero4 = %" PRIx64, fb->msaa.zero4);
1169 }
1170
1171 pandecode_indent--;
1172 pandecode_log("},\n");
1173 } else {
1174 pandecode_log(".shared_memory = {\n");
1175 pandecode_indent++;
1176 pandecode_shared_memory(&fb->shared_memory, is_compute);
1177 pandecode_indent--;
1178 pandecode_log("},\n");
1179 }
1180
1181 info.width = fb->width1 + 1;
1182 info.height = fb->height1 + 1;
1183 info.rt_count = fb->rt_count_1 + 1;
1184
1185 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
1186 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
1187 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
1188 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
1189
1190 pandecode_prop("unk1 = 0x%x", fb->unk1);
1191 pandecode_prop("unk2 = 0x%x", fb->unk2);
1192 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1193 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1194
1195 pandecode_log(".mfbd_flags = ");
1196 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1197 pandecode_log_cont(",\n");
1198
1199 if (fb->clear_stencil)
1200 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1201
1202 if (fb->clear_depth)
1203 pandecode_prop("clear_depth = %f", fb->clear_depth);
1204
1205 if (!is_compute)
1206 if (is_bifrost)
1207 pandecode_bifrost_tiler_descriptor(fb);
1208 else {
1209 const struct midgard_tiler_descriptor t = fb->tiler;
1210 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment, true);
1211 }
1212 else
1213 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
1214
1215 if (fb->zero3 || fb->zero4) {
1216 pandecode_msg("XXX: framebuffer zeros tripped\n");
1217 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1218 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1219 }
1220
1221 pandecode_indent--;
1222 pandecode_log("};\n");
1223
1224 gpu_va += sizeof(struct mali_framebuffer);
1225
1226 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
1227
1228 if (info.has_extra) {
1229 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1230 const struct mali_framebuffer_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1231
1232 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1233 pandecode_indent++;
1234
1235 MEMORY_PROP(fbx, checksum);
1236
1237 if (fbx->checksum_stride)
1238 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1239
1240 pandecode_log(".flags_hi = ");
1241 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info, fbx->flags_hi);
1242 pandecode_log_cont(",\n");
1243
1244 pandecode_log(".flags_lo = ");
1245 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info, fbx->flags_lo);
1246 pandecode_log_cont(",\n");
1247
1248 pandecode_prop("zs_block = %s", mali_block_format_as_str(fbx->zs_block));
1249 pandecode_prop("zs_samples = MALI_POSITIVE(%u)", fbx->zs_samples + 1);
1250
1251 if (fbx->zs_block == MALI_BLOCK_FORMAT_AFBC) {
1252 pandecode_log(".ds_afbc = {\n");
1253 pandecode_indent++;
1254
1255 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1256 pandecode_prop("depth_stencil_afbc_stride = %d",
1257 fbx->ds_afbc.depth_stencil_afbc_stride);
1258 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1259
1260 pandecode_log(".flags = ");
1261 pandecode_log_decoded_flags(afbc_fmt_flag_info, fbx->ds_afbc.flags);
1262 pandecode_log_cont(",\n");
1263
1264 if (fbx->ds_afbc.padding) {
1265 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1266 pandecode_prop("padding = 0x%" PRIx64, fbx->ds_afbc.padding);
1267 }
1268
1269 pandecode_indent--;
1270 pandecode_log("},\n");
1271 } else {
1272 pandecode_log(".ds_linear = {\n");
1273 pandecode_indent++;
1274
1275 if (fbx->ds_linear.depth) {
1276 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1277 pandecode_prop("depth_stride = %d",
1278 fbx->ds_linear.depth_stride);
1279 pandecode_prop("depth_layer_stride = %d",
1280 fbx->ds_linear.depth_layer_stride);
1281 } else if (fbx->ds_linear.depth_stride || fbx->ds_linear.depth_layer_stride) {
1282 pandecode_msg("XXX: depth stride zero tripped %d %d\n", fbx->ds_linear.depth_stride, fbx->ds_linear.depth_layer_stride);
1283 }
1284
1285 if (fbx->ds_linear.stencil) {
1286 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1287 pandecode_prop("stencil_stride = %d",
1288 fbx->ds_linear.stencil_stride);
1289 pandecode_prop("stencil_layer_stride = %d",
1290 fbx->ds_linear.stencil_layer_stride);
1291 } else if (fbx->ds_linear.stencil_stride || fbx->ds_linear.stencil_layer_stride) {
1292 pandecode_msg("XXX: stencil stride zero tripped %d %d\n", fbx->ds_linear.stencil_stride, fbx->ds_linear.stencil_layer_stride);
1293 }
1294
1295 if (fbx->ds_linear.depth_stride_zero ||
1296 fbx->ds_linear.stencil_stride_zero) {
1297 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1298 pandecode_prop("depth_stride_zero = 0x%x",
1299 fbx->ds_linear.depth_stride_zero);
1300 pandecode_prop("stencil_stride_zero = 0x%x",
1301 fbx->ds_linear.stencil_stride_zero);
1302 }
1303
1304 pandecode_indent--;
1305 pandecode_log("},\n");
1306 }
1307
1308 if (fbx->clear_color_1 | fbx->clear_color_2) {
1309 pandecode_prop("clear_color_1 = 0x%" PRIx32, fbx->clear_color_1);
1310 pandecode_prop("clear_color_2 = 0x%" PRIx32, fbx->clear_color_2);
1311 }
1312
1313 if (fbx->zero3) {
1314 pandecode_msg("XXX: fb_extra zeros tripped\n");
1315 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1316 }
1317
1318 pandecode_indent--;
1319 pandecode_log("};\n");
1320
1321 gpu_va += sizeof(struct mali_framebuffer_extra);
1322 }
1323
1324 if (is_fragment)
1325 pandecode_render_target(gpu_va, job_no, fb);
1326
1327 return info;
1328 }
1329
1330 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1331 * count */
1332
1333 static void
1334 pandecode_padded_vertices(unsigned shift, unsigned k)
1335 {
1336 unsigned odd = 2*k + 1;
1337 unsigned pot = 1 << shift;
1338 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1339 }
1340
1341 /* Given a magic divisor, recover what we were trying to divide by.
1342 *
1343 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1344 * 0 <= m < 2^N, for N bits in m.
1345 *
1346 * Let q represent the number we would like to divide by.
1347 *
1348 * By definition of a magic divisor for N-bit unsigned integers (a number you
1349 * multiply by to magically get division), m is a number such that:
1350 *
1351 * (m * x) & (2^N - 1) = floor(x/q).
1352 * for all x on Z where 0 <= x < 2^N
1353 *
1354 * Ignore the case where any of the above values equals zero; it is irrelevant
1355 * for our purposes (instanced arrays).
1356 *
1357 * Choose x = q. Then:
1358 *
1359 * (m * x) & (2^N - 1) = floor(x/q).
1360 * (m * q) & (2^N - 1) = floor(q/q).
1361 *
1362 * floor(q/q) = floor(1) = 1, therefore:
1363 *
1364 * (m * q) & (2^N - 1) = 1
1365 *
1366 * Recall the identity that the bitwise AND of one less than a power-of-two
1367 * equals the modulo with that power of two, i.e. for all x:
1368 *
1369 * x & (2^N - 1) = x % N
1370 *
1371 * Therefore:
1372 *
1373 * mq % (2^N) = 1
1374 *
1375 * By definition, a modular multiplicative inverse of a number m is the number
1376 * q such that with respect to a modulos M:
1377 *
1378 * mq % M = 1
1379 *
1380 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1381 *
1382 */
1383
1384 static void
1385 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1386 {
1387 #if 0
1388 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1389 * shift) the most lame way possible... just repeatedly add.
1390 * Asymptoptically slow but nobody cares in practice, unless you have
1391 * massive numbers of vertices or high divisors. */
1392
1393 unsigned inverse = 0;
1394
1395 /* Magic implicitly has the highest bit set */
1396 magic |= (1 << 31);
1397
1398 /* Depending on rounding direction */
1399 if (extra)
1400 magic++;
1401
1402 for (;;) {
1403 uint32_t product = magic * inverse;
1404
1405 if (shift) {
1406 product >>= shift;
1407 }
1408
1409 if (product == 1)
1410 break;
1411
1412 ++inverse;
1413 }
1414
1415 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1416
1417 /* Recall we're supposed to divide by (gl_level_divisor *
1418 * padded_num_vertices) */
1419
1420 unsigned padded_num_vertices = inverse / orig_divisor;
1421
1422 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1423 #endif
1424 }
1425
1426 static void
1427 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1428 mali_ptr addr, int job_no, char *suffix,
1429 int count, bool varying, enum mali_job_type job_type)
1430 {
1431 char *prefix = varying ? "varying" : "attribute";
1432 assert(addr);
1433
1434 if (!count) {
1435 pandecode_msg("warn: No %s records\n", prefix);
1436 return;
1437 }
1438
1439 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1440
1441 for (int i = 0; i < count; ++i) {
1442 /* First, check for special records */
1443 if (attr[i].elements < MALI_RECORD_SPECIAL) {
1444 if (attr[i].size)
1445 pandecode_msg("XXX: tripped size=%d\n", attr[i].size);
1446
1447 if (attr[i].stride) {
1448 /* gl_InstanceID passes a magic divisor in the
1449 * stride field to divide by the padded vertex
1450 * count. No other records should do so, so
1451 * stride should otherwise be zero. Note that
1452 * stride in the usual attribute sense doesn't
1453 * apply to special records. */
1454
1455 bool has_divisor = attr[i].elements == MALI_ATTR_INSTANCEID;
1456
1457 pandecode_log_cont("/* %smagic divisor = %X */ ",
1458 has_divisor ? "" : "XXX: ", attr[i].stride);
1459 }
1460
1461 if (attr[i].shift || attr[i].extra_flags) {
1462 /* Attributes use these fields for
1463 * instancing/padding/etc type issues, but
1464 * varyings don't */
1465
1466 pandecode_log_cont("/* %sshift=%d, extra=%d */ ",
1467 varying ? "XXX: " : "",
1468 attr[i].shift, attr[i].extra_flags);
1469 }
1470
1471 /* Print the special record name */
1472 bool attribute = false;
1473 pandecode_log("%s_%d = %s;\n", prefix, i, pandecode_special_record(attr[i].elements, &attribute));
1474
1475 /* Sanity check */
1476 if (attribute == varying)
1477 pandecode_msg("XXX: mismatched special record\n");
1478
1479 continue;
1480 }
1481
1482 enum mali_attr_mode mode = attr[i].elements & 7;
1483
1484 if (mode == MALI_ATTR_UNUSED)
1485 pandecode_msg("XXX: unused attribute record\n");
1486
1487 /* For non-linear records, we need to print the type of record */
1488 if (mode != MALI_ATTR_LINEAR)
1489 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode));
1490
1491 /* Print the name to link with attr_meta */
1492 pandecode_log_cont("%s_%d", prefix, i);
1493
1494 /* Print the stride and size */
1495 pandecode_log_cont("<%u>[%u]", attr[i].stride, attr[i].size);
1496
1497 /* TODO: Sanity check the quotient itself. It must be equal to
1498 * (or be greater than, if the driver added padding) the padded
1499 * vertex count. */
1500
1501 /* Finally, print the pointer */
1502 mali_ptr raw_elements = attr[i].elements & ~7;
1503 char *a = pointer_as_memory_reference(raw_elements);
1504 pandecode_log_cont(" = (%s);\n", a);
1505 free(a);
1506
1507 /* Check the pointer */
1508 pandecode_validate_buffer(raw_elements, attr[i].size);
1509
1510 /* shift/extra_flags exist only for instanced */
1511 if (attr[i].shift | attr[i].extra_flags) {
1512 /* These are set to random values by the blob for
1513 * varyings, most likely a symptom of uninitialized
1514 * memory where the hardware masked the bug. As such we
1515 * put this at a warning, not an error. */
1516
1517 if (mode == MALI_ATTR_LINEAR)
1518 pandecode_msg("warn: instancing fields set for linear\n");
1519
1520 pandecode_prop("shift = %d", attr[i].shift);
1521 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1522 }
1523
1524 /* Decode further where possible */
1525
1526 if (mode == MALI_ATTR_MODULO) {
1527 pandecode_padded_vertices(
1528 attr[i].shift,
1529 attr[i].extra_flags);
1530 }
1531
1532 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1533 i++;
1534 pandecode_log("{\n");
1535 pandecode_indent++;
1536 pandecode_prop("unk = 0x%x", attr[i].unk);
1537 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1538 if (attr[i].zero != 0)
1539 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1540 pandecode_prop("divisor = %d", attr[i].divisor);
1541 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1542 pandecode_indent--;
1543 pandecode_log("}, \n");
1544 }
1545
1546 }
1547
1548 pandecode_log("\n");
1549 }
1550
1551 static mali_ptr
1552 pandecode_shader_address(const char *name, mali_ptr ptr)
1553 {
1554 /* TODO: Decode flags */
1555 mali_ptr shader_ptr = ptr & ~15;
1556
1557 char *a = pointer_as_memory_reference(shader_ptr);
1558 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1559 free(a);
1560
1561 return shader_ptr;
1562 }
1563
1564 static void
1565 pandecode_blend_equation(const struct mali_blend_equation *blend)
1566 {
1567 if (blend->zero1)
1568 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1569
1570 pandecode_log(".equation = {\n");
1571 pandecode_indent++;
1572
1573 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1574 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1575
1576 pandecode_log(".color_mask = ");
1577 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1578 pandecode_log_cont(",\n");
1579
1580 pandecode_indent--;
1581 pandecode_log("},\n");
1582 }
1583
1584 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1585
1586 static unsigned
1587 decode_bifrost_constant(u16 constant)
1588 {
1589 float lo = (float) (constant & 0xFF);
1590 float hi = (float) (constant >> 8);
1591
1592 return (hi / 255.0) + (lo / 65535.0);
1593 }
1594
1595 static mali_ptr
1596 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1597 {
1598 struct bifrost_blend_rt *b =
1599 ((struct bifrost_blend_rt *) descs) + rt_no;
1600
1601 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1602 pandecode_indent++;
1603
1604 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1605 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1606 b->constant, decode_bifrost_constant(b->constant));
1607
1608 /* TODO figure out blend shader enable bit */
1609 pandecode_blend_equation(&b->equation);
1610
1611 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1612 pandecode_prop("index = 0x%" PRIx16, b->index);
1613
1614 pandecode_log(".format = ");
1615 pandecode_format_short(b->format, false);
1616 pandecode_swizzle(b->swizzle, b->format);
1617 pandecode_log_cont(",\n");
1618
1619 pandecode_prop("swizzle = 0x%" PRIx32, b->swizzle);
1620 pandecode_prop("format = 0x%" PRIx32, b->format);
1621
1622 if (b->zero1) {
1623 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1624 pandecode_prop("zero1 = 0x%" PRIx32, b->zero1);
1625 }
1626
1627 pandecode_log(".shader_type = ");
1628 switch(b->shader_type) {
1629 case BIFROST_BLEND_F16:
1630 pandecode_log_cont("BIFROST_BLEND_F16");
1631 break;
1632 case BIFROST_BLEND_F32:
1633 pandecode_log_cont("BIFROST_BLEND_F32");
1634 break;
1635 case BIFROST_BLEND_I32:
1636 pandecode_log_cont("BIFROST_BLEND_I32");
1637 break;
1638 case BIFROST_BLEND_U32:
1639 pandecode_log_cont("BIFROST_BLEND_U32");
1640 break;
1641 case BIFROST_BLEND_I16:
1642 pandecode_log_cont("BIFROST_BLEND_I16");
1643 break;
1644 case BIFROST_BLEND_U16:
1645 pandecode_log_cont("BIFROST_BLEND_U16");
1646 break;
1647 }
1648 pandecode_log_cont(",\n");
1649
1650 if (b->zero2) {
1651 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1652 pandecode_prop("zero2 = 0x%" PRIx32, b->zero2);
1653 }
1654
1655 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1656
1657 pandecode_indent--;
1658 pandecode_log("},\n");
1659
1660 return 0;
1661 }
1662
1663 static mali_ptr
1664 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1665 {
1666 /* constant/equation is in a union */
1667 if (!blend->shader)
1668 return 0;
1669
1670 pandecode_log(".blend = {\n");
1671 pandecode_indent++;
1672
1673 if (is_shader) {
1674 pandecode_shader_address("shader", blend->shader);
1675 } else {
1676 pandecode_blend_equation(&blend->equation);
1677 pandecode_prop("constant = %f", blend->constant);
1678 }
1679
1680 pandecode_indent--;
1681 pandecode_log("},\n");
1682
1683 /* Return blend shader to disassemble if present */
1684 return is_shader ? (blend->shader & ~0xF) : 0;
1685 }
1686
1687 static mali_ptr
1688 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1689 {
1690 struct midgard_blend_rt *b =
1691 ((struct midgard_blend_rt *) descs) + rt_no;
1692
1693 /* Flags determine presence of blend shader */
1694 bool is_shader = (b->flags & 0xF) >= 0x2;
1695
1696 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1697 pandecode_indent++;
1698
1699 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1700
1701 union midgard_blend blend = b->blend;
1702 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1703
1704 pandecode_indent--;
1705 pandecode_log("};\n");
1706
1707 return shader;
1708 }
1709
1710 /* Attributes and varyings have descriptor records, which contain information
1711 * about their format and ordering with the attribute/varying buffers. We'll
1712 * want to validate that the combinations specified are self-consistent.
1713 */
1714
1715 static int
1716 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1717 {
1718 char base[128];
1719 char *prefix = varying ? "varying" : "attribute";
1720 unsigned max_index = 0;
1721 snprintf(base, sizeof(base), "%s_meta", prefix);
1722
1723 struct mali_attr_meta *attr_meta;
1724 mali_ptr p = varying ? v->varying_meta : v->attribute_meta;
1725
1726 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1727
1728 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1729 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1730 sizeof(*attr_mem));
1731
1732 /* If the record is discard, it should be zero for everything else */
1733
1734 if (attr_meta->format == MALI_VARYING_DISCARD) {
1735 uint64_t zero =
1736 attr_meta->index |
1737 attr_meta->unknown1 |
1738 attr_meta->unknown3 |
1739 attr_meta->src_offset;
1740
1741 if (zero)
1742 pandecode_msg("XXX: expected empty record for varying discard\n");
1743
1744 /* We want to look for a literal 0000 swizzle -- this
1745 * is not encoded with all zeroes, however */
1746
1747 enum mali_channel z = MALI_CHANNEL_ZERO;
1748 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1749 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1750
1751 if (!good_swizzle)
1752 pandecode_msg("XXX: expected zero swizzle for discard\n");
1753
1754 if (!varying)
1755 pandecode_msg("XXX: cannot discard attribute\n");
1756
1757 /* If we're all good, omit the record */
1758 if (!zero && varying && good_swizzle) {
1759 pandecode_log("/* discarded varying */\n");
1760 continue;
1761 }
1762 }
1763
1764 if (attr_meta->index > max_index)
1765 max_index = attr_meta->index;
1766
1767 if (attr_meta->unknown1 != 0x2) {
1768 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1769 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1770 }
1771
1772 if (attr_meta->unknown3) {
1773 pandecode_msg("XXX: unexpected unknown3 set\n");
1774 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1775 }
1776
1777 pandecode_format_short(attr_meta->format, false);
1778 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1779
1780 if (attr_meta->src_offset)
1781 pandecode_log_cont("[%u]", attr_meta->src_offset);
1782
1783 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1784
1785 pandecode_log_cont(";\n");
1786 }
1787
1788 pandecode_log("\n");
1789
1790 return count ? (max_index + 1) : 0;
1791 }
1792
1793 /* return bits [lo, hi) of word */
1794 static u32
1795 bits(u32 word, u32 lo, u32 hi)
1796 {
1797 if (hi - lo >= 32)
1798 return word; // avoid undefined behavior with the shift
1799
1800 return (word >> lo) & ((1 << (hi - lo)) - 1);
1801 }
1802
1803 static void
1804 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool graphics)
1805 {
1806 pandecode_log(".prefix = {\n");
1807 pandecode_indent++;
1808
1809 /* Decode invocation_count. See the comment before the definition of
1810 * invocation_count for an explanation.
1811 */
1812
1813 unsigned size_y_shift = bits(p->invocation_shifts, 0, 5);
1814 unsigned size_z_shift = bits(p->invocation_shifts, 5, 10);
1815 unsigned workgroups_x_shift = bits(p->invocation_shifts, 10, 16);
1816 unsigned workgroups_y_shift = bits(p->invocation_shifts, 16, 22);
1817 unsigned workgroups_z_shift = bits(p->invocation_shifts, 22, 28);
1818 unsigned workgroups_x_shift_2 = bits(p->invocation_shifts, 28, 32);
1819
1820 unsigned size_x = bits(p->invocation_count, 0, size_y_shift) + 1;
1821 unsigned size_y = bits(p->invocation_count, size_y_shift, size_z_shift) + 1;
1822 unsigned size_z = bits(p->invocation_count, size_z_shift, workgroups_x_shift) + 1;
1823
1824 unsigned groups_x = bits(p->invocation_count, workgroups_x_shift, workgroups_y_shift) + 1;
1825 unsigned groups_y = bits(p->invocation_count, workgroups_y_shift, workgroups_z_shift) + 1;
1826 unsigned groups_z = bits(p->invocation_count, workgroups_z_shift, 32) + 1;
1827
1828 /* Even though we have this decoded, we want to ensure that the
1829 * representation is "unique" so we don't lose anything by printing only
1830 * the final result. More specifically, we need to check that we were
1831 * passed something in canonical form, since the definition per the
1832 * hardware is inherently not unique. How? Well, take the resulting
1833 * decode and pack it ourselves! If it is bit exact with what we
1834 * decoded, we're good to go. */
1835
1836 struct mali_vertex_tiler_prefix ref;
1837 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, graphics);
1838
1839 bool canonical =
1840 (p->invocation_count == ref.invocation_count) &&
1841 (p->invocation_shifts == ref.invocation_shifts);
1842
1843 if (!canonical) {
1844 pandecode_msg("XXX: non-canonical workgroups packing\n");
1845 pandecode_msg("expected: %X, %X",
1846 ref.invocation_count,
1847 ref.invocation_shifts);
1848
1849 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1850 pandecode_prop("size_y_shift = %d", size_y_shift);
1851 pandecode_prop("size_z_shift = %d", size_z_shift);
1852 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift);
1853 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift);
1854 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift);
1855 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2);
1856 }
1857
1858 /* Regardless, print the decode */
1859 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1860 size_x, size_y, size_z,
1861 groups_x, groups_y, groups_z);
1862
1863 /* TODO: Decode */
1864 if (p->unknown_draw)
1865 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1866
1867 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1868
1869 if (p->draw_mode != MALI_DRAW_MODE_NONE)
1870 pandecode_prop("draw_mode = %s", mali_draw_mode_as_str(p->draw_mode));
1871
1872 /* Index count only exists for tiler jobs anyway */
1873
1874 if (p->index_count)
1875 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1876
1877
1878 unsigned index_raw_size = (p->unknown_draw & MALI_DRAW_INDEXED_SIZE);
1879 index_raw_size >>= MALI_DRAW_INDEXED_SHIFT;
1880
1881 /* Validate an index buffer is present if we need one. TODO: verify
1882 * relationship between invocation_count and index_count */
1883
1884 if (p->indices) {
1885 unsigned count = p->index_count;
1886
1887 /* Grab the size */
1888 unsigned size = (index_raw_size == 0x3) ? 4 : index_raw_size;
1889
1890 /* Ensure we got a size, and if so, validate the index buffer
1891 * is large enough to hold a full set of indices of the given
1892 * size */
1893
1894 if (!index_raw_size)
1895 pandecode_msg("XXX: index size missing\n");
1896 else
1897 pandecode_validate_buffer(p->indices, count * size);
1898 } else if (index_raw_size)
1899 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size);
1900
1901 if (p->offset_bias_correction)
1902 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1903
1904 /* TODO: Figure out what this is. It's not zero */
1905 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1906
1907 pandecode_indent--;
1908 pandecode_log("},\n");
1909 }
1910
1911 static void
1912 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1913 {
1914 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1915 uint64_t *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1916
1917 for (int i = 0; i < ubufs_count; i++) {
1918 unsigned size = (ubufs[i] & ((1 << 10) - 1)) * 16;
1919 mali_ptr addr = (ubufs[i] >> 10) << 2;
1920
1921 pandecode_validate_buffer(addr, size);
1922
1923 char *ptr = pointer_as_memory_reference(addr);
1924 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
1925 free(ptr);
1926 }
1927
1928 pandecode_log("\n");
1929 }
1930
1931 static void
1932 pandecode_uniforms(mali_ptr uniforms, unsigned uniform_count)
1933 {
1934 pandecode_validate_buffer(uniforms, uniform_count * 16);
1935
1936 char *ptr = pointer_as_memory_reference(uniforms);
1937 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count, ptr);
1938 free(ptr);
1939 }
1940
1941 static const char *
1942 shader_type_for_job(unsigned type)
1943 {
1944 switch (type) {
1945 case MALI_JOB_TYPE_VERTEX: return "VERTEX";
1946 case MALI_JOB_TYPE_TILER: return "FRAGMENT";
1947 case MALI_JOB_TYPE_COMPUTE: return "COMPUTE";
1948 default:
1949 return "UNKNOWN";
1950 }
1951 }
1952
1953 static unsigned shader_id = 0;
1954
1955 static struct midgard_disasm_stats
1956 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1957 bool is_bifrost, unsigned gpu_id)
1958 {
1959 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1960 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1961
1962 /* Compute maximum possible size */
1963 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1964
1965 /* Print some boilerplate to clearly denote the assembly (which doesn't
1966 * obey indentation rules), and actually do the disassembly! */
1967
1968 pandecode_log_cont("\n\n");
1969
1970 struct midgard_disasm_stats stats;
1971
1972 if (is_bifrost) {
1973 disassemble_bifrost(pandecode_dump_stream, code, sz, true);
1974
1975 /* TODO: Extend stats to Bifrost */
1976 stats.texture_count = -128;
1977 stats.sampler_count = -128;
1978 stats.attribute_count = -128;
1979 stats.varying_count = -128;
1980 stats.uniform_count = -128;
1981 stats.uniform_buffer_count = -128;
1982 stats.work_count = -128;
1983
1984 stats.instruction_count = 0;
1985 stats.bundle_count = 0;
1986 stats.quadword_count = 0;
1987 stats.helper_invocations = false;
1988 } else {
1989 stats = disassemble_midgard(pandecode_dump_stream,
1990 code, sz, gpu_id,
1991 type == MALI_JOB_TYPE_TILER ?
1992 MESA_SHADER_FRAGMENT : MESA_SHADER_VERTEX);
1993 }
1994
1995 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
1996 * driver-internal purposes with the blob and interfere */
1997
1998 bool should_shaderdb = type != MALI_JOB_TYPE_COMPUTE;
1999
2000 if (should_shaderdb) {
2001 unsigned nr_threads =
2002 (stats.work_count <= 4) ? 4 :
2003 (stats.work_count <= 8) ? 2 :
2004 1;
2005
2006 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
2007 "%u inst, %u bundles, %u quadwords, "
2008 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
2009 shader_id++,
2010 shader_type_for_job(type),
2011 stats.instruction_count, stats.bundle_count, stats.quadword_count,
2012 stats.work_count, nr_threads);
2013 }
2014
2015
2016 return stats;
2017 }
2018
2019 static void
2020 pandecode_texture_payload(mali_ptr payload,
2021 enum mali_texture_dimension dim,
2022 enum mali_texture_layout layout,
2023 bool manual_stride,
2024 uint8_t levels,
2025 uint16_t depth,
2026 uint16_t array_size,
2027 struct pandecode_mapped_memory *tmem)
2028 {
2029 pandecode_log(".payload = {\n");
2030 pandecode_indent++;
2031
2032 /* A bunch of bitmap pointers follow.
2033 * We work out the correct number,
2034 * based on the mipmap/cubemap
2035 * properties, but dump extra
2036 * possibilities to futureproof */
2037
2038 int bitmap_count = levels + 1;
2039
2040 /* Miptree for each face */
2041 if (dim == MALI_TEXTURE_DIMENSION_CUBE)
2042 bitmap_count *= 6;
2043
2044 /* Array of layers */
2045 bitmap_count *= depth;
2046
2047 /* Array of textures */
2048 bitmap_count *= array_size;
2049
2050 /* Stride for each element */
2051 if (manual_stride)
2052 bitmap_count *= 2;
2053
2054 mali_ptr *pointers_and_strides = pandecode_fetch_gpu_mem(tmem,
2055 payload, sizeof(mali_ptr) * bitmap_count);
2056 for (int i = 0; i < bitmap_count; ++i) {
2057 /* How we dump depends if this is a stride or a pointer */
2058
2059 if (manual_stride && (i & 1)) {
2060 /* signed 32-bit snuck in as a 64-bit pointer */
2061 uint64_t stride_set = pointers_and_strides[i];
2062 uint32_t clamped_stride = stride_set;
2063 int32_t stride = clamped_stride;
2064 assert(stride_set == clamped_stride);
2065 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
2066 } else {
2067 char *a = pointer_as_memory_reference(pointers_and_strides[i]);
2068 pandecode_log("%s, \n", a);
2069 free(a);
2070 }
2071 }
2072
2073 pandecode_indent--;
2074 pandecode_log("},\n");
2075 }
2076
2077 static void
2078 pandecode_texture(mali_ptr u,
2079 struct pandecode_mapped_memory *tmem,
2080 unsigned job_no, unsigned tex)
2081 {
2082 struct pandecode_mapped_memory *mapped_mem = pandecode_find_mapped_gpu_mem_containing(u);
2083 const uint8_t *cl = pandecode_fetch_gpu_mem(mapped_mem, u, MALI_MIDGARD_TEXTURE_LENGTH);
2084
2085 struct MALI_MIDGARD_TEXTURE temp;
2086 MALI_MIDGARD_TEXTURE_unpack(cl, &temp);
2087 MALI_MIDGARD_TEXTURE_print(pandecode_dump_stream, &temp, 2);
2088
2089 pandecode_texture_payload(u + MALI_MIDGARD_TEXTURE_LENGTH,
2090 temp.dimension, temp.texel_ordering, temp.manual_stride,
2091 temp.levels, temp.depth, temp.array_size, mapped_mem);
2092 }
2093
2094 static void
2095 pandecode_bifrost_texture(
2096 const void *cl,
2097 unsigned job_no,
2098 unsigned tex)
2099 {
2100 struct MALI_BIFROST_TEXTURE temp;
2101 MALI_BIFROST_TEXTURE_unpack(cl, &temp);
2102 MALI_BIFROST_TEXTURE_print(pandecode_dump_stream, &temp, 2);
2103
2104 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(temp.surfaces);
2105 pandecode_texture_payload(temp.surfaces, temp.dimension, temp.texel_ordering,
2106 true, temp.levels, 1, 1, tmem);
2107 }
2108
2109 /* 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 */
2110
2111 static void
2112 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
2113 {
2114 /* Nothing to do */
2115 if (claim == truth)
2116 return;
2117
2118 if (fuzzy && (truth < 0))
2119 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name, claim, truth);
2120
2121 if ((truth >= 0) && !fuzzy) {
2122 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2123 (truth < claim) ? "warn" : "XXX",
2124 name, truth, claim);
2125 } else if ((claim > -truth) && !fuzzy) {
2126 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2127 name, -truth, claim);
2128 } else if (fuzzy && (claim < truth))
2129 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2130 name, truth, claim);
2131
2132 pandecode_log(".%s = %" PRId16, name, claim);
2133
2134 if (fuzzy)
2135 pandecode_log_cont(" /* %u used */", truth);
2136
2137 pandecode_log_cont(",\n");
2138 }
2139
2140 static void
2141 pandecode_blend_shader_disassemble(mali_ptr shader, int job_no, int job_type,
2142 bool is_bifrost, unsigned gpu_id)
2143 {
2144 struct midgard_disasm_stats stats =
2145 pandecode_shader_disassemble(shader, job_no, job_type, is_bifrost, gpu_id);
2146
2147 bool has_texture = (stats.texture_count > 0);
2148 bool has_sampler = (stats.sampler_count > 0);
2149 bool has_attribute = (stats.attribute_count > 0);
2150 bool has_varying = (stats.varying_count > 0);
2151 bool has_uniform = (stats.uniform_count > 0);
2152 bool has_ubo = (stats.uniform_buffer_count > 0);
2153
2154 if (has_texture || has_sampler)
2155 pandecode_msg("XXX: blend shader accessing textures\n");
2156
2157 if (has_attribute || has_varying)
2158 pandecode_msg("XXX: blend shader accessing interstage\n");
2159
2160 if (has_uniform || has_ubo)
2161 pandecode_msg("XXX: blend shader accessing uniforms\n");
2162 }
2163
2164 static void
2165 pandecode_textures(mali_ptr textures, unsigned texture_count, int job_no, bool is_bifrost)
2166 {
2167 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(textures);
2168
2169 if (!mmem)
2170 return;
2171
2172 pandecode_log("Textures (%"PRIx64"):\n", textures);
2173
2174 if (is_bifrost) {
2175 const void *cl = pandecode_fetch_gpu_mem(mmem,
2176 textures, MALI_BIFROST_TEXTURE_LENGTH *
2177 texture_count);
2178
2179 for (unsigned tex = 0; tex < texture_count; ++tex) {
2180 pandecode_bifrost_texture(cl +
2181 MALI_BIFROST_TEXTURE_LENGTH * tex,
2182 job_no, tex);
2183 }
2184 } else {
2185 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures);
2186
2187 for (int tex = 0; tex < texture_count; ++tex) {
2188 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
2189 char *a = pointer_as_memory_reference(*u);
2190 pandecode_log("%s,\n", a);
2191 free(a);
2192 }
2193
2194 /* Now, finally, descend down into the texture descriptor */
2195 for (unsigned tex = 0; tex < texture_count; ++tex) {
2196 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
2197 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2198 if (tmem)
2199 pandecode_texture(*u, tmem, job_no, tex);
2200 }
2201 }
2202 }
2203
2204 static void
2205 pandecode_samplers(mali_ptr samplers, unsigned sampler_count, int job_no, bool is_bifrost)
2206 {
2207 for (int i = 0; i < sampler_count; ++i) {
2208 if (is_bifrost) {
2209 DUMP_ADDR("Sampler", BIFROST_SAMPLER, samplers + (MALI_BIFROST_SAMPLER_LENGTH * i), 1);
2210 } else {
2211 DUMP_ADDR("Sampler", MIDGARD_SAMPLER, samplers + (MALI_MIDGARD_SAMPLER_LENGTH * i), 1);
2212 }
2213 }
2214 }
2215
2216 static void
2217 pandecode_vertex_tiler_postfix_pre(
2218 const struct mali_vertex_tiler_postfix *p,
2219 int job_no, enum mali_job_type job_type,
2220 char *suffix, bool is_bifrost, unsigned gpu_id)
2221 {
2222 struct pandecode_mapped_memory *attr_mem;
2223
2224 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2225 * are the only things actually needed from the FBD, vertex/tiler jobs
2226 * no longer reference the FBD -- instead, this field points to some
2227 * info about the scratchpad.
2228 */
2229
2230 struct pandecode_fbd fbd_info = {
2231 /* Default for Bifrost */
2232 .rt_count = 1
2233 };
2234
2235 if (is_bifrost) {
2236 pandecode_log_cont("\t/* %X %/\n", p->shared_memory & 1);
2237 pandecode_compute_fbd(p->shared_memory & ~1, job_no);
2238 } else if (p->shared_memory & MALI_MFBD)
2239 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->shared_memory) & FBD_MASK, job_no, false, job_type == MALI_JOB_TYPE_COMPUTE, false);
2240 else if (job_type == MALI_JOB_TYPE_COMPUTE)
2241 pandecode_compute_fbd((u64) (uintptr_t) p->shared_memory, job_no);
2242 else
2243 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->shared_memory, job_no, false, gpu_id);
2244
2245 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
2246 int texture_count = 0, sampler_count = 0;
2247
2248 if (p->shader) {
2249 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->shader);
2250 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, p->shader);
2251
2252 /* Disassemble ahead-of-time to get stats. Initialize with
2253 * stats for the missing-shader case so we get validation
2254 * there, too */
2255
2256 struct midgard_disasm_stats info = {
2257 .texture_count = 0,
2258 .sampler_count = 0,
2259 .attribute_count = 0,
2260 .varying_count = 0,
2261 .work_count = 1,
2262
2263 .uniform_count = -128,
2264 .uniform_buffer_count = 0
2265 };
2266
2267 if (s->shader & ~0xF)
2268 info = pandecode_shader_disassemble(s->shader & ~0xF, job_no, job_type, is_bifrost, gpu_id);
2269
2270 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", p->shader, job_no, suffix);
2271 pandecode_indent++;
2272
2273 /* Save for dumps */
2274 attribute_count = s->attribute_count;
2275 varying_count = s->varying_count;
2276 texture_count = s->texture_count;
2277 sampler_count = s->sampler_count;
2278
2279 if (is_bifrost) {
2280 uniform_count = s->bifrost2.uniform_count;
2281 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
2282 } else {
2283 uniform_count = s->midgard1.uniform_count;
2284 uniform_buffer_count = s->midgard1.uniform_buffer_count;
2285 }
2286
2287 pandecode_shader_address("shader", s->shader);
2288
2289 pandecode_shader_prop("texture_count", s->texture_count, info.texture_count, false);
2290 pandecode_shader_prop("sampler_count", s->sampler_count, info.sampler_count, false);
2291 pandecode_shader_prop("attribute_count", s->attribute_count, info.attribute_count, false);
2292 pandecode_shader_prop("varying_count", s->varying_count, info.varying_count, false);
2293 pandecode_shader_prop("uniform_buffer_count",
2294 uniform_buffer_count,
2295 info.uniform_buffer_count, true);
2296
2297 if (!is_bifrost) {
2298 pandecode_shader_prop("uniform_count",
2299 uniform_count,
2300 info.uniform_count, false);
2301
2302 pandecode_shader_prop("work_count",
2303 s->midgard1.work_count, info.work_count, false);
2304 }
2305
2306 if (is_bifrost) {
2307 pandecode_log("bifrost1.unk1 = ");
2308 pandecode_log_decoded_flags(shader_bifrost_info, s->bifrost1.unk1);
2309 pandecode_log_cont(",\n");
2310 } else {
2311 bool helpers = s->midgard1.flags_lo & MALI_HELPER_INVOCATIONS;
2312
2313 if (helpers != info.helper_invocations) {
2314 pandecode_msg("XXX: expected helpers %u but got %u\n",
2315 info.helper_invocations, helpers);
2316 }
2317
2318 pandecode_log(".midgard1.flags_lo = ");
2319 pandecode_log_decoded_flags(shader_midgard1_flag_lo_info,
2320 s->midgard1.flags_lo & ~MALI_HELPER_INVOCATIONS);
2321 pandecode_log_cont(",\n");
2322
2323 pandecode_log(".midgard1.flags_hi = ");
2324 pandecode_log_decoded_flags(shader_midgard1_flag_hi_info, s->midgard1.flags_hi);
2325 pandecode_log_cont(",\n");
2326 }
2327
2328 if (s->depth_units || s->depth_factor) {
2329 pandecode_prop("depth_factor = %f", s->depth_factor);
2330 pandecode_prop("depth_units = %f", s->depth_units);
2331 }
2332
2333 if (s->coverage_mask)
2334 pandecode_prop("coverage_mask = 0x%X", s->coverage_mask);
2335
2336 if (s->unknown2_2)
2337 pandecode_prop(".unknown2_2 = %X", s->unknown2_2);
2338
2339 if (s->unknown2_3 || s->unknown2_4) {
2340 pandecode_log(".unknown2_3 = ");
2341
2342 int unknown2_3 = s->unknown2_3;
2343 int unknown2_4 = s->unknown2_4;
2344
2345 /* We're not quite sure what these flags mean without the depth test, if anything */
2346
2347 if (unknown2_3 & (MALI_DEPTH_WRITEMASK | MALI_DEPTH_FUNC_MASK)) {
2348 const char *func = mali_func_as_str(MALI_GET_DEPTH_FUNC(unknown2_3));
2349 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
2350
2351 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
2352 }
2353
2354 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
2355 pandecode_log_cont(",\n");
2356
2357 pandecode_log(".unknown2_4 = ");
2358 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
2359 pandecode_log_cont(",\n");
2360 }
2361
2362 if (s->stencil_mask_front || s->stencil_mask_back) {
2363 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
2364 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
2365 }
2366
2367 DUMP_CL("Stencil front", STENCIL, &s->stencil_front, 1);
2368 DUMP_CL("Stencil back", STENCIL, &s->stencil_back, 1);
2369
2370 if (is_bifrost) {
2371 pandecode_log(".bifrost2 = {\n");
2372 pandecode_indent++;
2373
2374 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
2375 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
2376 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
2377 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
2378
2379 pandecode_indent--;
2380 pandecode_log("},\n");
2381 } else if (s->midgard2.unknown2_7) {
2382 pandecode_log(".midgard2 = {\n");
2383 pandecode_indent++;
2384
2385 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
2386 pandecode_indent--;
2387 pandecode_log("},\n");
2388 }
2389
2390 if (s->padding) {
2391 pandecode_msg("XXX: shader padding tripped\n");
2392 pandecode_prop("padding = 0x%" PRIx32, s->padding);
2393 }
2394
2395 if (!is_bifrost) {
2396 /* TODO: Blend shaders routing/disasm */
2397 union midgard_blend blend = s->blend;
2398 mali_ptr shader = pandecode_midgard_blend(&blend, s->unknown2_3 & MALI_HAS_BLEND_SHADER);
2399 if (shader & ~0xF)
2400 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2401 } else {
2402 pandecode_msg("mdg_blend = %" PRIx64 "\n", s->blend.shader);
2403 }
2404
2405 pandecode_indent--;
2406 pandecode_log("};\n");
2407
2408 /* MRT blend fields are used whenever MFBD is used, with
2409 * per-RT descriptors */
2410
2411 if (job_type == MALI_JOB_TYPE_TILER && (is_bifrost || p->shared_memory & MALI_MFBD)) {
2412 void* blend_base = (void *) (s + 1);
2413
2414 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
2415 mali_ptr shader = 0;
2416
2417 if (is_bifrost)
2418 shader = pandecode_bifrost_blend(blend_base, job_no, i);
2419 else
2420 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
2421
2422 if (shader & ~0xF)
2423 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2424
2425 }
2426 }
2427 } else
2428 pandecode_msg("XXX: missing shader descriptor\n");
2429
2430 if (p->viewport)
2431 DUMP_ADDR("Viewport", VIEWPORT, p->viewport, 1);
2432
2433 unsigned max_attr_index = 0;
2434
2435 if (p->attribute_meta)
2436 max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
2437
2438 if (p->attributes) {
2439 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
2440 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false, job_type);
2441 }
2442
2443 /* Varyings are encoded like attributes but not actually sent; we just
2444 * pass a zero buffer with the right stride/size set, (or whatever)
2445 * since the GPU will write to it itself */
2446
2447 if (p->varying_meta) {
2448 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
2449 }
2450
2451 if (p->varyings) {
2452 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
2453
2454 /* Number of descriptors depends on whether there are
2455 * non-internal varyings */
2456
2457 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true, job_type);
2458 }
2459
2460 if (p->uniform_buffers) {
2461 if (uniform_buffer_count)
2462 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
2463 else
2464 pandecode_msg("warn: UBOs specified but not referenced\n");
2465 } else if (uniform_buffer_count)
2466 pandecode_msg("XXX: UBOs referenced but not specified\n");
2467
2468 /* We don't want to actually dump uniforms, but we do need to validate
2469 * that the counts we were given are sane */
2470
2471 if (p->uniforms) {
2472 if (uniform_count)
2473 pandecode_uniforms(p->uniforms, uniform_count);
2474 else
2475 pandecode_msg("warn: Uniforms specified but not referenced\n");
2476 } else if (uniform_count)
2477 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2478
2479 if (p->textures)
2480 pandecode_textures(p->textures, texture_count, job_no, is_bifrost);
2481
2482 if (p->sampler_descriptor)
2483 pandecode_samplers(p->sampler_descriptor, sampler_count, job_no, is_bifrost);
2484 }
2485
2486 static void
2487 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2488 {
2489 pandecode_log(".gl_enables = ");
2490
2491 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2492
2493 pandecode_log_cont(",\n");
2494 }
2495
2496 static void
2497 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2498 {
2499 if (p->shader & 0xF)
2500 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p->shader & 0xF));
2501
2502 pandecode_log(".postfix = {\n");
2503 pandecode_indent++;
2504
2505 pandecode_gl_enables(p->gl_enables, MALI_JOB_TYPE_TILER);
2506 pandecode_prop("instance_shift = 0x%x", p->instance_shift);
2507 pandecode_prop("instance_odd = 0x%x", p->instance_odd);
2508
2509 if (p->zero4) {
2510 pandecode_msg("XXX: vertex only zero tripped");
2511 pandecode_prop("zero4 = 0x%" PRIx32, p->zero4);
2512 }
2513
2514 pandecode_prop("offset_start = 0x%x", p->offset_start);
2515
2516 if (p->zero5) {
2517 pandecode_msg("XXX: vertex only zero tripped");
2518 pandecode_prop("zero5 = 0x%" PRIx64, p->zero5);
2519 }
2520
2521 MEMORY_PROP(p, position_varying);
2522 MEMORY_PROP(p, occlusion_counter);
2523
2524 pandecode_indent--;
2525 pandecode_log("},\n");
2526 }
2527
2528 static void
2529 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2530 {
2531 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2532 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2533
2534 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
2535 pandecode_indent++;
2536
2537 if (h->zero) {
2538 pandecode_msg("XXX: tiler heap zero tripped\n");
2539 pandecode_prop("zero = 0x%x", h->zero);
2540 }
2541
2542 pandecode_prop("heap_size = 0x%x", h->heap_size);
2543 MEMORY_PROP(h, tiler_heap_start);
2544 MEMORY_PROP(h, tiler_heap_free);
2545
2546 /* this might point to the beginning of another buffer, when it's
2547 * really the end of the tiler heap buffer, so we have to be careful
2548 * here. but for zero length, we need the same pointer.
2549 */
2550
2551 if (h->tiler_heap_end == h->tiler_heap_start) {
2552 MEMORY_PROP(h, tiler_heap_start);
2553 } else {
2554 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2555 pandecode_prop("tiler_heap_end = %s + 1", a);
2556 free(a);
2557 }
2558
2559 for (int i = 0; i < 10; i++) {
2560 if (h->zeros[i] != 0) {
2561 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2562 i, h->zeros[i]);
2563 }
2564 }
2565
2566 if (h->unk1 != 0x1) {
2567 pandecode_msg("XXX: tiler heap unk1 tripped\n");
2568 pandecode_prop("unk1 = 0x%x", h->unk1);
2569 }
2570
2571 if (h->unk7e007e != 0x7e007e) {
2572 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
2573 pandecode_prop("unk7e007e = 0x%x", h->unk7e007e);
2574 }
2575
2576 pandecode_indent--;
2577 pandecode_log("};\n");
2578 }
2579
2580 static void
2581 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2582 {
2583 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2584 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2585
2586 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2587
2588 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
2589 pandecode_indent++;
2590
2591 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32, t->tiler_heap_next_start);
2592 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32, t->used_hierarchy_mask);
2593
2594 if (t->hierarchy_mask != 0xa &&
2595 t->hierarchy_mask != 0x14 &&
2596 t->hierarchy_mask != 0x28 &&
2597 t->hierarchy_mask != 0x50 &&
2598 t->hierarchy_mask != 0xa0)
2599 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
2600
2601 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2602
2603 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2604
2605 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2606 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2607
2608 if (t->zero0) {
2609 pandecode_msg("XXX: tiler meta zero tripped\n");
2610 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2611 }
2612
2613 for (int i = 0; i < 12; i++) {
2614 if (t->zeros[i] != 0) {
2615 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2616 i, t->zeros[i]);
2617 }
2618 }
2619
2620 pandecode_indent--;
2621 pandecode_log("};\n");
2622 }
2623
2624 static void
2625 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2626 {
2627 if (u.pointer == 0x0)
2628 return;
2629
2630 pandecode_log(".primitive_size = {\n");
2631 pandecode_indent++;
2632
2633 if (constant) {
2634 pandecode_prop("constant = %f", u.constant);
2635 } else {
2636 MEMORY_PROP((&u), pointer);
2637 }
2638
2639 pandecode_indent--;
2640 pandecode_log("},\n");
2641 }
2642
2643 static void
2644 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2645 {
2646 pandecode_log_cont("{\n");
2647 pandecode_indent++;
2648
2649 /* TODO: gl_PointSize on Bifrost */
2650 pandecode_primitive_size(t->primitive_size, true);
2651
2652 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2653 || t->zero6) {
2654 pandecode_msg("XXX: tiler only zero tripped\n");
2655 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2656 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2657 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2658 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2659 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2660 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2661 }
2662
2663 pandecode_indent--;
2664 pandecode_log("},\n");
2665 }
2666
2667 static int
2668 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2669 const struct pandecode_mapped_memory *mem,
2670 mali_ptr payload, int job_no, unsigned gpu_id)
2671 {
2672 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2673
2674 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true, gpu_id);
2675
2676 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64"_%d = {\n", payload, job_no);
2677 pandecode_indent++;
2678
2679 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2680 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2681
2682 pandecode_indent--;
2683 pandecode_log("};\n");
2684
2685 return sizeof(*v);
2686 }
2687
2688 static int
2689 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2690 const struct pandecode_mapped_memory *mem,
2691 mali_ptr payload, int job_no, unsigned gpu_id)
2692 {
2693 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2694
2695 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true, gpu_id);
2696 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2697
2698 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64"_%d = {\n", payload, job_no);
2699 pandecode_indent++;
2700
2701 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
2702
2703 pandecode_log(".tiler = ");
2704 pandecode_tiler_only_bfr(&t->tiler, job_no);
2705
2706 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2707
2708 pandecode_indent--;
2709 pandecode_log("};\n");
2710
2711 return sizeof(*t);
2712 }
2713
2714 static int
2715 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2716 const struct pandecode_mapped_memory *mem,
2717 mali_ptr payload, int job_no, unsigned gpu_id)
2718 {
2719 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2720 bool is_graphics = (h->job_type == MALI_JOB_TYPE_VERTEX) || (h->job_type == MALI_JOB_TYPE_TILER);
2721
2722 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false, gpu_id);
2723
2724 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2725 pandecode_indent++;
2726
2727 pandecode_vertex_tiler_prefix(&v->prefix, job_no, is_graphics);
2728 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2729
2730 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2731 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
2732
2733 pandecode_indent--;
2734 pandecode_log("};\n");
2735
2736 return sizeof(*v);
2737 }
2738
2739 static int
2740 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2741 mali_ptr payload, int job_no,
2742 bool is_bifrost, unsigned gpu_id)
2743 {
2744 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2745
2746 bool is_mfbd = s->framebuffer & MALI_MFBD;
2747
2748 if (!is_mfbd && is_bifrost)
2749 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2750
2751 struct pandecode_fbd info;
2752
2753 if (is_mfbd)
2754 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true, false, is_bifrost);
2755 else
2756 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true, gpu_id);
2757
2758 /* Compute the tag for the tagged pointer. This contains the type of
2759 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2760 * additional structures follow the MFBD header (an extra payload or
2761 * not, as well as a count of render targets) */
2762
2763 unsigned expected_tag = is_mfbd ? MALI_MFBD : 0;
2764
2765 if (is_mfbd) {
2766 if (info.has_extra)
2767 expected_tag |= MALI_MFBD_TAG_EXTRA;
2768
2769 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
2770 }
2771
2772 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
2773 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2774 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
2775 pandecode_prop("max_tile_coord = 0x%X\n", s->max_tile_coord);
2776 }
2777
2778 /* Extract tile coordinates */
2779
2780 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
2781 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
2782
2783 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2784 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2785
2786 /* For the max, we also want the floored (rather than ceiled) version for checking */
2787
2788 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
2789 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
2790
2791 /* Validate the coordinates are well-ordered */
2792
2793 if (min_x == max_x)
2794 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2795 else if (min_x > max_x)
2796 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2797
2798 if (min_y == max_y)
2799 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2800 else if (min_y > max_y)
2801 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2802
2803 /* Validate the coordinates fit inside the framebuffer. We use floor,
2804 * rather than ceil, for the max coordinates, since the tile
2805 * coordinates for something like an 800x600 framebuffer will actually
2806 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2807
2808 if ((min_x > info.width) || (max_x_f > info.width))
2809 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2810
2811 if ((min_y > info.height) || (max_y_f > info.height))
2812 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2813
2814 /* After validation, we print */
2815
2816 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
2817
2818 /* The FBD is a tagged pointer */
2819
2820 unsigned tag = (s->framebuffer & ~FBD_MASK);
2821
2822 if (tag != expected_tag)
2823 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
2824
2825 return sizeof(*s);
2826 }
2827
2828 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
2829 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
2830 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
2831 * might wish to combine this with the bifrost parameter) because some details
2832 * are model-specific even within a particular architecture. Minimal traces
2833 * *only* examine the job descriptors, skipping printing entirely if there is
2834 * no faults, and only descends into the payload if there are faults. This is
2835 * useful for looking for faults without the overhead of invasive traces. */
2836
2837 void
2838 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost, unsigned gpu_id, bool minimal)
2839 {
2840 pandecode_dump_file_open();
2841
2842 struct mali_job_descriptor_header *h;
2843 unsigned job_descriptor_number = 0;
2844
2845 do {
2846 struct pandecode_mapped_memory *mem =
2847 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2848
2849 void *payload;
2850
2851 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2852
2853 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2854 * high 32-bits of the 64-bit pointer are reused to store
2855 * something else.
2856 */
2857 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2858 h->job_type != MALI_JOB_TYPE_FRAGMENT ? 4 : 0;
2859 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2860
2861 payload = pandecode_fetch_gpu_mem(mem, payload_ptr, 256);
2862
2863 int job_no = job_descriptor_number++;
2864
2865 /* If the job is good to go, skip it in minimal mode */
2866 if (minimal && (h->exception_status == 0x0 || h->exception_status == 0x1))
2867 continue;
2868
2869 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2870 pandecode_indent++;
2871
2872 pandecode_prop("job_type = %s", mali_job_type_as_str(h->job_type));
2873
2874 if (h->job_descriptor_size)
2875 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2876
2877 if (h->exception_status && h->exception_status != 0x1)
2878 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2879 h->exception_status,
2880 (h->exception_status >> 16) & 0xFFFF,
2881 pandecode_exception_access((h->exception_status >> 8) & 0x3),
2882 h->exception_status & 0xFF);
2883
2884 if (h->first_incomplete_task)
2885 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2886
2887 if (h->fault_pointer)
2888 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2889
2890 if (h->job_barrier)
2891 pandecode_prop("job_barrier = %d", h->job_barrier);
2892
2893 pandecode_prop("job_index = %d", h->job_index);
2894
2895 if (h->unknown_flags)
2896 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2897
2898 if (h->job_dependency_index_1)
2899 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2900
2901 if (h->job_dependency_index_2)
2902 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2903
2904 pandecode_indent--;
2905 pandecode_log("};\n");
2906
2907 switch (h->job_type) {
2908 case MALI_JOB_TYPE_WRITE_VALUE: {
2909 struct mali_payload_write_value *s = payload;
2910 pandecode_log("struct mali_payload_write_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2911 pandecode_indent++;
2912 MEMORY_PROP(s, address);
2913
2914 if (s->value_descriptor != MALI_WRITE_VALUE_ZERO) {
2915 pandecode_msg("XXX: unknown value descriptor\n");
2916 pandecode_prop("value_descriptor = 0x%" PRIX32, s->value_descriptor);
2917 }
2918
2919 if (s->reserved) {
2920 pandecode_msg("XXX: set value tripped\n");
2921 pandecode_prop("reserved = 0x%" PRIX32, s->reserved);
2922 }
2923
2924 pandecode_prop("immediate = 0x%" PRIX64, s->immediate);
2925 pandecode_indent--;
2926 pandecode_log("};\n");
2927
2928 break;
2929 }
2930
2931 case MALI_JOB_TYPE_TILER:
2932 case MALI_JOB_TYPE_VERTEX:
2933 case MALI_JOB_TYPE_COMPUTE:
2934 if (bifrost) {
2935 if (h->job_type == MALI_JOB_TYPE_TILER)
2936 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2937 else
2938 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2939 } else
2940 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no, gpu_id);
2941
2942 break;
2943
2944 case MALI_JOB_TYPE_FRAGMENT:
2945 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost, gpu_id);
2946 break;
2947
2948 default:
2949 break;
2950 }
2951 } while ((jc_gpu_va = h->next_job));
2952
2953 pandecode_map_read_write();
2954 }