e51c7fb8289fd20a719819b38f370264aee2b174
[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 MAP_ADDR(T, addr, cl) \
68 const uint8_t *cl = 0; \
69 { \
70 struct pandecode_mapped_memory *mapped_mem = pandecode_find_mapped_gpu_mem_containing(addr); \
71 cl = pandecode_fetch_gpu_mem(mapped_mem, addr, MALI_ ## T ## _LENGTH); \
72 }
73
74 #define DUMP_ADDR(title, T, addr, indent) {\
75 MAP_ADDR(T, addr, cl) \
76 DUMP_CL(title, T, cl, indent); \
77 }
78
79 FILE *pandecode_dump_stream;
80
81 /* Semantic logging type.
82 *
83 * Raw: for raw messages to be printed as is.
84 * Message: for helpful information to be commented out in replays.
85 * Property: for properties of a struct
86 *
87 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
88 */
89
90 enum pandecode_log_type {
91 PANDECODE_RAW,
92 PANDECODE_MESSAGE,
93 PANDECODE_PROPERTY
94 };
95
96 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
97 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
98 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
99
100 unsigned pandecode_indent = 0;
101
102 static void
103 pandecode_make_indent(void)
104 {
105 for (unsigned i = 0; i < pandecode_indent; ++i)
106 fprintf(pandecode_dump_stream, " ");
107 }
108
109 static void
110 pandecode_log_typed(enum pandecode_log_type type, const char *format, ...)
111 {
112 va_list ap;
113
114 pandecode_make_indent();
115
116 if (type == PANDECODE_MESSAGE)
117 fprintf(pandecode_dump_stream, "// ");
118 else if (type == PANDECODE_PROPERTY)
119 fprintf(pandecode_dump_stream, ".");
120
121 va_start(ap, format);
122 vfprintf(pandecode_dump_stream, format, ap);
123 va_end(ap);
124
125 if (type == PANDECODE_PROPERTY)
126 fprintf(pandecode_dump_stream, ",\n");
127 }
128
129 static void
130 pandecode_log_cont(const char *format, ...)
131 {
132 va_list ap;
133
134 va_start(ap, format);
135 vfprintf(pandecode_dump_stream, format, ap);
136 va_end(ap);
137 }
138
139 /* To check for memory safety issues, validates that the given pointer in GPU
140 * memory is valid, containing at least sz bytes. The goal is to eliminate
141 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
142 * overruns) by statically validating pointers.
143 */
144
145 static void
146 pandecode_validate_buffer(mali_ptr addr, size_t sz)
147 {
148 if (!addr) {
149 pandecode_msg("XXX: null pointer deref");
150 return;
151 }
152
153 /* Find a BO */
154
155 struct pandecode_mapped_memory *bo =
156 pandecode_find_mapped_gpu_mem_containing(addr);
157
158 if (!bo) {
159 pandecode_msg("XXX: invalid memory dereference\n");
160 return;
161 }
162
163 /* Bounds check */
164
165 unsigned offset = addr - bo->gpu_va;
166 unsigned total = offset + sz;
167
168 if (total > bo->length) {
169 pandecode_msg("XXX: buffer overrun. "
170 "Chunk of size %zu at offset %d in buffer of size %zu. "
171 "Overrun by %zu bytes. \n",
172 sz, offset, bo->length, total - bo->length);
173 return;
174 }
175 }
176
177 struct pandecode_flag_info {
178 u64 flag;
179 const char *name;
180 };
181
182 static void
183 pandecode_log_decoded_flags(const struct pandecode_flag_info *flag_info,
184 u64 flags)
185 {
186 bool decodable_flags_found = false;
187
188 for (int i = 0; flag_info[i].name; i++) {
189 if ((flags & flag_info[i].flag) != flag_info[i].flag)
190 continue;
191
192 if (!decodable_flags_found) {
193 decodable_flags_found = true;
194 } else {
195 pandecode_log_cont(" | ");
196 }
197
198 pandecode_log_cont("%s", flag_info[i].name);
199
200 flags &= ~flag_info[i].flag;
201 }
202
203 if (decodable_flags_found) {
204 if (flags)
205 pandecode_log_cont(" | 0x%" PRIx64, flags);
206 } else {
207 pandecode_log_cont("0x%" PRIx64, flags);
208 }
209 }
210
211 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
212 static const struct pandecode_flag_info gl_enable_flag_info[] = {
213 FLAG_INFO(OCCLUSION_QUERY),
214 FLAG_INFO(OCCLUSION_PRECISE),
215 FLAG_INFO(FRONT_CCW_TOP),
216 FLAG_INFO(CULL_FACE_FRONT),
217 FLAG_INFO(CULL_FACE_BACK),
218 {}
219 };
220 #undef FLAG_INFO
221
222 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
223 static const struct pandecode_flag_info clear_flag_info[] = {
224 FLAG_INFO(FAST),
225 FLAG_INFO(SLOW),
226 FLAG_INFO(SLOW_STENCIL),
227 {}
228 };
229 #undef FLAG_INFO
230
231 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
232 static const struct pandecode_flag_info mask_flag_info[] = {
233 FLAG_INFO(R),
234 FLAG_INFO(G),
235 FLAG_INFO(B),
236 FLAG_INFO(A),
237 {}
238 };
239 #undef FLAG_INFO
240
241 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
242 static const struct pandecode_flag_info u3_flag_info[] = {
243 FLAG_INFO(HAS_MSAA),
244 FLAG_INFO(PER_SAMPLE),
245 FLAG_INFO(CAN_DISCARD),
246 FLAG_INFO(HAS_BLEND_SHADER),
247 FLAG_INFO(DEPTH_WRITEMASK),
248 FLAG_INFO(DEPTH_CLIP_NEAR),
249 FLAG_INFO(DEPTH_CLIP_FAR),
250 {}
251 };
252
253 static const struct pandecode_flag_info u4_flag_info[] = {
254 FLAG_INFO(NO_MSAA),
255 FLAG_INFO(NO_DITHER),
256 FLAG_INFO(DEPTH_RANGE_A),
257 FLAG_INFO(DEPTH_RANGE_B),
258 FLAG_INFO(STENCIL_TEST),
259 FLAG_INFO(ALPHA_TO_COVERAGE),
260 {}
261 };
262 #undef FLAG_INFO
263
264 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
265 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
266 FLAG_INFO(SRGB),
267 {}
268 };
269 #undef FLAG_INFO
270
271 #define FLAG_INFO(flag) { MALI_AFBC_##flag, "MALI_AFBC_" #flag }
272 static const struct pandecode_flag_info afbc_fmt_flag_info[] = {
273 FLAG_INFO(YTR),
274 {}
275 };
276 #undef FLAG_INFO
277
278 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
279 static const struct pandecode_flag_info mfbd_extra_flag_hi_info[] = {
280 FLAG_INFO(PRESENT),
281 {}
282 };
283 #undef FLAG_INFO
284
285 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
286 static const struct pandecode_flag_info mfbd_extra_flag_lo_info[] = {
287 FLAG_INFO(ZS),
288 {}
289 };
290 #undef FLAG_INFO
291
292 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
293 static const struct pandecode_flag_info shader_midgard1_flag_lo_info [] = {
294 FLAG_INFO(WRITES_Z),
295 FLAG_INFO(EARLY_Z),
296 FLAG_INFO(READS_TILEBUFFER),
297 FLAG_INFO(WRITES_GLOBAL),
298 FLAG_INFO(READS_ZS),
299 {}
300 };
301
302 static const struct pandecode_flag_info shader_midgard1_flag_hi_info [] = {
303 FLAG_INFO(WRITES_S),
304 FLAG_INFO(SUPPRESS_INF_NAN),
305 {}
306 };
307 #undef FLAG_INFO
308
309 #define FLAG_INFO(flag) { MALI_BIFROST_##flag, "MALI_BIFROST_" #flag }
310 static const struct pandecode_flag_info shader_bifrost_info [] = {
311 FLAG_INFO(FULL_THREAD),
312 FLAG_INFO(EARLY_Z),
313 FLAG_INFO(FIRST_ATEST),
314 {}
315 };
316
317 #undef FLAG_INFO
318
319 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
320 static const struct pandecode_flag_info mfbd_flag_info [] = {
321 FLAG_INFO(DEPTH_WRITE),
322 FLAG_INFO(EXTRA),
323 {}
324 };
325 #undef FLAG_INFO
326
327 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
328 static const struct pandecode_flag_info sfbd_unk1_info [] = {
329 FLAG_INFO(MSAA_8),
330 FLAG_INFO(MSAA_A),
331 {}
332 };
333 #undef FLAG_INFO
334
335 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
336 static const struct pandecode_flag_info sfbd_unk2_info [] = {
337 FLAG_INFO(MSAA_B),
338 FLAG_INFO(SRGB),
339 {}
340 };
341 #undef FLAG_INFO
342
343 /* Midgard's tiler descriptor is embedded within the
344 * larger FBD */
345
346 static void
347 pandecode_midgard_tiler_descriptor(
348 const struct midgard_tiler_descriptor *t,
349 unsigned width,
350 unsigned height,
351 bool is_fragment,
352 bool has_hierarchy)
353 {
354 pandecode_log(".tiler = {\n");
355 pandecode_indent++;
356
357 if (t->hierarchy_mask == MALI_TILER_DISABLED)
358 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
359 else
360 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
361
362 /* We know this name from the kernel, but we never see it nonzero */
363
364 if (t->flags)
365 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
366
367 MEMORY_PROP(t, polygon_list);
368
369 /* The body is offset from the base of the polygon list */
370 //assert(t->polygon_list_body > t->polygon_list);
371 unsigned body_offset = t->polygon_list_body - t->polygon_list;
372
373 /* It needs to fit inside the reported size */
374 //assert(t->polygon_list_size >= body_offset);
375
376 /* Now that we've sanity checked, we'll try to calculate the sizes
377 * ourselves for comparison */
378
379 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask, has_hierarchy);
380 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask, has_hierarchy);
381
382 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
383 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
384 ref_header, ref_size);
385 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
386 pandecode_msg("body offset %d\n", body_offset);
387 }
388
389 /* The tiler heap has a start and end specified -- it should be
390 * identical to what we have in the BO. The exception is if tiling is
391 * disabled. */
392
393 MEMORY_PROP(t, heap_start);
394 assert(t->heap_end >= t->heap_start);
395
396 unsigned heap_size = t->heap_end - t->heap_start;
397
398 /* Tiling is enabled with a special flag */
399 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
400 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
401
402 bool tiling_enabled = hierarchy_mask;
403
404 if (tiling_enabled) {
405 /* We should also have no other flags */
406 if (tiler_flags)
407 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
408 } else {
409 /* When tiling is disabled, we should have that flag and no others */
410
411 if (tiler_flags != MALI_TILER_DISABLED) {
412 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
413 tiler_flags);
414 }
415
416 /* We should also have an empty heap */
417 if (heap_size) {
418 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
419 heap_size);
420 }
421
422 /* Disabled tiling is used only for clear-only jobs, which are
423 * purely FRAGMENT, so we should never see this for
424 * non-FRAGMENT descriptors. */
425
426 if (!is_fragment)
427 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
428 }
429
430 /* We've never seen weights used in practice, but we know from the
431 * kernel these fields is there */
432
433 bool nonzero_weights = false;
434
435 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
436 nonzero_weights |= t->weights[w] != 0x0;
437 }
438
439 if (nonzero_weights) {
440 pandecode_log(".weights = { ");
441
442 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
443 pandecode_log_cont("%d, ", t->weights[w]);
444 }
445
446 pandecode_log("},");
447 }
448
449 pandecode_indent--;
450 pandecode_log("}\n");
451 }
452
453 /* TODO: The Bifrost tiler is not understood at all yet */
454
455 static void
456 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer *fb)
457 {
458 pandecode_log(".tiler = {\n");
459 pandecode_indent++;
460
461 MEMORY_PROP(fb, tiler_meta);
462
463 for (int i = 0; i < 16; i++) {
464 if (fb->zeros[i] != 0) {
465 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
466 i, fb->zeros[i]);
467 }
468 }
469
470 pandecode_log("},\n");
471
472 pandecode_indent--;
473 pandecode_log("}\n");
474
475 }
476
477 /* Information about the framebuffer passed back for
478 * additional analysis */
479
480 struct pandecode_fbd {
481 unsigned width;
482 unsigned height;
483 unsigned rt_count;
484 bool has_extra;
485 };
486
487 static void
488 pandecode_sfbd_format(struct mali_sfbd_format format)
489 {
490 pandecode_log(".format = {\n");
491 pandecode_indent++;
492
493 pandecode_log(".unk1 = ");
494 pandecode_log_decoded_flags(sfbd_unk1_info, format.unk1);
495 pandecode_log_cont(",\n");
496
497 /* TODO: Map formats so we can check swizzles and print nicely */
498 pandecode_log("swizzle");
499 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
500 pandecode_log_cont(",\n");
501
502 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
503 (format.nr_channels + 1));
504
505 pandecode_log(".unk2 = ");
506 pandecode_log_decoded_flags(sfbd_unk2_info, format.unk2);
507 pandecode_log_cont(",\n");
508
509 pandecode_prop("block = %s", mali_block_format_as_str(format.block));
510
511 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
512
513 pandecode_indent--;
514 pandecode_log("},\n");
515 }
516
517 static void
518 pandecode_shared_memory(const struct mali_shared_memory *desc, bool is_compute)
519 {
520 pandecode_prop("stack_shift = 0x%x", desc->stack_shift);
521
522 if (desc->unk0)
523 pandecode_prop("unk0 = 0x%x", desc->unk0);
524
525 if (desc->shared_workgroup_count != 0x1F) {
526 pandecode_prop("shared_workgroup_count = %d", desc->shared_workgroup_count);
527 if (!is_compute)
528 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
529 }
530
531 if (desc->shared_unk1 || desc->shared_shift) {
532 pandecode_prop("shared_unk1 = %X", desc->shared_unk1);
533 pandecode_prop("shared_shift = %X", desc->shared_shift);
534
535 if (!is_compute)
536 pandecode_msg("XXX: shared memory configured in noncompute shader");
537 }
538
539 if (desc->shared_zero) {
540 pandecode_msg("XXX: shared memory zero tripped\n");
541 pandecode_prop("shared_zero = 0x%" PRIx32, desc->shared_zero);
542 }
543
544 if (desc->shared_memory && !is_compute)
545 pandecode_msg("XXX: shared memory used in noncompute shader\n");
546
547 MEMORY_PROP(desc, scratchpad);
548 MEMORY_PROP(desc, shared_memory);
549 MEMORY_PROP(desc, unknown1);
550
551 if (desc->scratchpad) {
552 struct pandecode_mapped_memory *smem =
553 pandecode_find_mapped_gpu_mem_containing(desc->scratchpad);
554
555 pandecode_msg("scratchpad size %u\n", smem->length);
556 }
557
558 }
559
560 static struct pandecode_fbd
561 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment, unsigned gpu_id)
562 {
563 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
564 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
565
566 struct pandecode_fbd info = {
567 .has_extra = false,
568 .rt_count = 1
569 };
570
571 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
572 pandecode_indent++;
573
574 pandecode_log(".shared_memory = {\n");
575 pandecode_indent++;
576 pandecode_shared_memory(&s->shared_memory, false);
577 pandecode_indent--;
578 pandecode_log("},\n");
579
580 pandecode_sfbd_format(s->format);
581
582 info.width = s->width + 1;
583 info.height = s->height + 1;
584
585 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
586 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
587
588 MEMORY_PROP(s, checksum);
589
590 if (s->checksum_stride)
591 pandecode_prop("checksum_stride = %d", s->checksum_stride);
592
593 MEMORY_PROP(s, framebuffer);
594 pandecode_prop("stride = %d", s->stride);
595
596 /* Earlier in the actual commandstream -- right before width -- but we
597 * delay to flow nicer */
598
599 pandecode_log(".clear_flags = ");
600 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
601 pandecode_log_cont(",\n");
602
603 if (s->depth_buffer) {
604 MEMORY_PROP(s, depth_buffer);
605 pandecode_prop("depth_stride = %d", s->depth_stride);
606 }
607
608 if (s->stencil_buffer) {
609 MEMORY_PROP(s, stencil_buffer);
610 pandecode_prop("stencil_stride = %d", s->stencil_stride);
611 }
612
613 if (s->depth_stride_zero ||
614 s->stencil_stride_zero ||
615 s->zero7 || s->zero8) {
616 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
617 pandecode_prop("depth_stride_zero = 0x%x",
618 s->depth_stride_zero);
619 pandecode_prop("stencil_stride_zero = 0x%x",
620 s->stencil_stride_zero);
621 pandecode_prop("zero7 = 0x%" PRIx32,
622 s->zero7);
623 pandecode_prop("zero8 = 0x%" PRIx32,
624 s->zero8);
625 }
626
627 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
628 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
629 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
630 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
631 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
632 }
633
634 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
635 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
636 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
637 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
638 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
639 }
640
641 if (s->clear_stencil) {
642 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
643 }
644
645 const struct midgard_tiler_descriptor t = s->tiler;
646
647 bool has_hierarchy = !(gpu_id == 0x0720 || gpu_id == 0x0820 || gpu_id == 0x0830);
648 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment, has_hierarchy);
649
650 pandecode_indent--;
651 pandecode_log("};\n");
652
653 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
654 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
655 pandecode_prop("zero5 = 0x%" PRIx32, s->zero5);
656
657 pandecode_log_cont(".zero3 = {");
658
659 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
660 pandecode_log_cont("%X, ", s->zero3[i]);
661
662 pandecode_log_cont("},\n");
663
664 pandecode_log_cont(".zero6 = {");
665
666 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
667 pandecode_log_cont("%X, ", s->zero6[i]);
668
669 pandecode_log_cont("},\n");
670
671 return info;
672 }
673
674 static void
675 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
676 {
677 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
678 const struct mali_shared_memory *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
679
680 pandecode_log("struct mali_shared_memory shared_%"PRIx64"_%d = {\n", gpu_va, job_no);
681 pandecode_indent++;
682 pandecode_shared_memory(s, true);
683 pandecode_indent--;
684 pandecode_log("},\n");
685 }
686
687 /* Extracts the number of components associated with a Mali format */
688
689 static unsigned
690 pandecode_format_component_count(enum mali_format fmt)
691 {
692 /* Mask out the format class */
693 unsigned top = fmt & 0b11100000;
694
695 switch (top) {
696 case MALI_FORMAT_SNORM:
697 case MALI_FORMAT_UINT:
698 case MALI_FORMAT_UNORM:
699 case MALI_FORMAT_SINT:
700 return ((fmt >> 3) & 3) + 1;
701 default:
702 /* TODO: Validate */
703 return 4;
704 }
705 }
706
707 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
708
709 static unsigned
710 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
711 {
712 unsigned mask = 0;
713 assert(MALI_CHANNEL_R == 0);
714
715 for (unsigned c = 0; c < 4; ++c) {
716 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
717
718 if (chan <= MALI_CHANNEL_A)
719 mask |= (1 << chan);
720 }
721
722 return mask;
723 }
724
725 /* Validates that a (format, swizzle) pair is valid, in the sense that the
726 * swizzle doesn't access any components that are undefined in the format.
727 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
728 * omitted */
729
730 static bool
731 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
732 {
733 unsigned nr_comp = pandecode_format_component_count(fmt);
734 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
735 unsigned valid_mask = (1 << nr_comp) - 1;
736 unsigned invalid_mask = ~valid_mask;
737
738 if (access_mask & invalid_mask) {
739 pandecode_msg("XXX: invalid components accessed\n");
740 return false;
741 }
742
743 /* Check for the default non-swizzling swizzle so we can suppress
744 * useless printing for the defaults */
745
746 unsigned default_swizzles[4] = {
747 MALI_CHANNEL_R | (MALI_CHANNEL_0 << 3) | (MALI_CHANNEL_0 << 6) | (MALI_CHANNEL_1 << 9),
748 MALI_CHANNEL_R | (MALI_CHANNEL_G << 3) | (MALI_CHANNEL_0 << 6) | (MALI_CHANNEL_1 << 9),
749 MALI_CHANNEL_R | (MALI_CHANNEL_G << 3) | (MALI_CHANNEL_B << 6) | (MALI_CHANNEL_1 << 9),
750 MALI_CHANNEL_R | (MALI_CHANNEL_G << 3) | (MALI_CHANNEL_B << 6) | (MALI_CHANNEL_A << 9)
751 };
752
753 return (swizzle == default_swizzles[nr_comp - 1]);
754 }
755
756 static void
757 pandecode_swizzle(unsigned swizzle, enum mali_format format)
758 {
759 /* First, do some validation */
760 bool trivial_swizzle = pandecode_validate_format_swizzle(
761 format, swizzle);
762
763 if (trivial_swizzle)
764 return;
765
766 /* Next, print the swizzle */
767 pandecode_log_cont(".");
768
769 static const char components[] = "rgba01";
770
771 for (unsigned c = 0; c < 4; ++c) {
772 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
773
774 if (chan > MALI_CHANNEL_1) {
775 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
776 continue;
777 }
778 pandecode_log_cont("%c", components[chan]);
779 }
780 }
781
782 static void
783 pandecode_rt_format(struct mali_rt_format format)
784 {
785 pandecode_log(".format = {\n");
786 pandecode_indent++;
787
788 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
789 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
790 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
791 pandecode_prop("unk4 = 0x%" PRIx32, format.unk4);
792
793 pandecode_prop("block = %s", mali_block_format_as_str(format.block));
794
795 /* TODO: Map formats so we can check swizzles and print nicely */
796 pandecode_log("swizzle");
797 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
798 pandecode_log_cont(",\n");
799
800 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
801 (format.nr_channels + 1));
802
803 pandecode_log(".flags = ");
804 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
805 pandecode_log_cont(",\n");
806
807 pandecode_prop("msaa = %s", mali_msaa_as_str(format.msaa));
808
809 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
810 * which is a faster hardware-based alternative to the wallpaper method
811 * to preserve framebuffer contents across frames. In practice, MFBD
812 * preload is buggy on Midgard, and so this is a chicken bit. If this
813 * bit isn't set, most likely something broke unrelated to preload */
814
815 if (!format.no_preload) {
816 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
817 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
818 }
819
820 if (format.zero)
821 pandecode_prop("zero = 0x%" PRIx32, format.zero);
822
823 pandecode_indent--;
824 pandecode_log("},\n");
825 }
826
827 static void
828 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct mali_framebuffer *fb)
829 {
830 pandecode_log("struct mali_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
831 pandecode_indent++;
832
833 for (int i = 0; i < (fb->rt_count_1 + 1); i++) {
834 mali_ptr rt_va = gpu_va + i * sizeof(struct mali_render_target);
835 struct pandecode_mapped_memory *mem =
836 pandecode_find_mapped_gpu_mem_containing(rt_va);
837 const struct mali_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
838
839 pandecode_log("{\n");
840 pandecode_indent++;
841
842 pandecode_rt_format(rt->format);
843
844 if (rt->format.block == MALI_BLOCK_FORMAT_AFBC) {
845 pandecode_log(".afbc = {\n");
846 pandecode_indent++;
847
848 char *a = pointer_as_memory_reference(rt->afbc.metadata);
849 pandecode_prop("metadata = %s", a);
850 free(a);
851
852 pandecode_prop("stride = %d", rt->afbc.stride);
853
854 pandecode_log(".flags = ");
855 pandecode_log_decoded_flags(afbc_fmt_flag_info, rt->afbc.flags);
856 pandecode_log_cont(",\n");
857
858 pandecode_indent--;
859 pandecode_log("},\n");
860 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.flags) {
861 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
862 rt->afbc.metadata,
863 rt->afbc.stride,
864 rt->afbc.flags);
865 }
866
867 MEMORY_PROP(rt, framebuffer);
868 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
869
870 if (rt->layer_stride)
871 pandecode_prop("layer_stride = %d", rt->layer_stride);
872
873 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
874 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
875 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
876 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
877 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
878 }
879
880 if (rt->zero1 || rt->zero2) {
881 pandecode_msg("XXX: render target zeros tripped\n");
882 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
883 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
884 }
885
886 pandecode_indent--;
887 pandecode_log("},\n");
888 }
889
890 pandecode_indent--;
891 pandecode_log("};\n");
892 }
893
894 static struct pandecode_fbd
895 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment, bool is_compute, bool is_bifrost)
896 {
897 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
898 const struct mali_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
899
900 struct pandecode_fbd info;
901
902 if (is_bifrost && fb->msaa.sample_locations) {
903 /* The blob stores all possible sample locations in a single buffer
904 * allocated on startup, and just switches the pointer when switching
905 * MSAA state. For now, we just put the data into the cmdstream, but we
906 * should do something like what the blob does with a real driver.
907 *
908 * There seem to be 32 slots for sample locations, followed by another
909 * 16. The second 16 is just the center location followed by 15 zeros
910 * in all the cases I've identified (maybe shader vs. depth/color
911 * samples?).
912 */
913
914 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->msaa.sample_locations);
915
916 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->msaa.sample_locations);
917
918 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
919 pandecode_indent++;
920
921 for (int i = 0; i < 32 + 16; i++) {
922 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
923 }
924
925 pandecode_indent--;
926 pandecode_log("};\n");
927 }
928
929 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
930 pandecode_indent++;
931
932 if (is_bifrost) {
933 pandecode_log(".msaa = {\n");
934 pandecode_indent++;
935
936 if (fb->msaa.sample_locations)
937 pandecode_prop("sample_locations = sample_locations_%d", job_no);
938 else
939 pandecode_msg("XXX: sample_locations missing\n");
940
941 if (fb->msaa.zero1 || fb->msaa.zero2 || fb->msaa.zero4) {
942 pandecode_msg("XXX: multisampling zero tripped\n");
943 pandecode_prop("zero1 = %" PRIx64, fb->msaa.zero1);
944 pandecode_prop("zero2 = %" PRIx64, fb->msaa.zero2);
945 pandecode_prop("zero4 = %" PRIx64, fb->msaa.zero4);
946 }
947
948 pandecode_indent--;
949 pandecode_log("},\n");
950 } else {
951 pandecode_log(".shared_memory = {\n");
952 pandecode_indent++;
953 pandecode_shared_memory(&fb->shared_memory, is_compute);
954 pandecode_indent--;
955 pandecode_log("},\n");
956 }
957
958 info.width = fb->width1 + 1;
959 info.height = fb->height1 + 1;
960 info.rt_count = fb->rt_count_1 + 1;
961
962 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
963 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
964 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
965 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
966
967 pandecode_prop("unk1 = 0x%x", fb->unk1);
968 pandecode_prop("unk2 = 0x%x", fb->unk2);
969 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
970 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
971
972 pandecode_log(".mfbd_flags = ");
973 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
974 pandecode_log_cont(",\n");
975
976 if (fb->clear_stencil)
977 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
978
979 if (fb->clear_depth)
980 pandecode_prop("clear_depth = %f", fb->clear_depth);
981
982 if (!is_compute)
983 if (is_bifrost)
984 pandecode_bifrost_tiler_descriptor(fb);
985 else {
986 const struct midgard_tiler_descriptor t = fb->tiler;
987 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment, true);
988 }
989 else
990 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
991
992 if (fb->zero3 || fb->zero4) {
993 pandecode_msg("XXX: framebuffer zeros tripped\n");
994 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
995 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
996 }
997
998 pandecode_indent--;
999 pandecode_log("};\n");
1000
1001 gpu_va += sizeof(struct mali_framebuffer);
1002
1003 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
1004
1005 if (info.has_extra) {
1006 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1007 const struct mali_framebuffer_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1008
1009 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1010 pandecode_indent++;
1011
1012 MEMORY_PROP(fbx, checksum);
1013
1014 if (fbx->checksum_stride)
1015 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1016
1017 pandecode_log(".flags_hi = ");
1018 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info, fbx->flags_hi);
1019 pandecode_log_cont(",\n");
1020
1021 pandecode_log(".flags_lo = ");
1022 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info, fbx->flags_lo);
1023 pandecode_log_cont(",\n");
1024
1025 pandecode_prop("zs_block = %s", mali_block_format_as_str(fbx->zs_block));
1026 pandecode_prop("zs_samples = MALI_POSITIVE(%u)", fbx->zs_samples + 1);
1027
1028 if (fbx->zs_block == MALI_BLOCK_FORMAT_AFBC) {
1029 pandecode_log(".ds_afbc = {\n");
1030 pandecode_indent++;
1031
1032 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1033 pandecode_prop("depth_stencil_afbc_stride = %d",
1034 fbx->ds_afbc.depth_stencil_afbc_stride);
1035 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1036
1037 pandecode_log(".flags = ");
1038 pandecode_log_decoded_flags(afbc_fmt_flag_info, fbx->ds_afbc.flags);
1039 pandecode_log_cont(",\n");
1040
1041 if (fbx->ds_afbc.padding) {
1042 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1043 pandecode_prop("padding = 0x%" PRIx64, fbx->ds_afbc.padding);
1044 }
1045
1046 pandecode_indent--;
1047 pandecode_log("},\n");
1048 } else {
1049 pandecode_log(".ds_linear = {\n");
1050 pandecode_indent++;
1051
1052 if (fbx->ds_linear.depth) {
1053 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1054 pandecode_prop("depth_stride = %d",
1055 fbx->ds_linear.depth_stride);
1056 pandecode_prop("depth_layer_stride = %d",
1057 fbx->ds_linear.depth_layer_stride);
1058 } else if (fbx->ds_linear.depth_stride || fbx->ds_linear.depth_layer_stride) {
1059 pandecode_msg("XXX: depth stride zero tripped %d %d\n", fbx->ds_linear.depth_stride, fbx->ds_linear.depth_layer_stride);
1060 }
1061
1062 if (fbx->ds_linear.stencil) {
1063 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1064 pandecode_prop("stencil_stride = %d",
1065 fbx->ds_linear.stencil_stride);
1066 pandecode_prop("stencil_layer_stride = %d",
1067 fbx->ds_linear.stencil_layer_stride);
1068 } else if (fbx->ds_linear.stencil_stride || fbx->ds_linear.stencil_layer_stride) {
1069 pandecode_msg("XXX: stencil stride zero tripped %d %d\n", fbx->ds_linear.stencil_stride, fbx->ds_linear.stencil_layer_stride);
1070 }
1071
1072 if (fbx->ds_linear.depth_stride_zero ||
1073 fbx->ds_linear.stencil_stride_zero) {
1074 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1075 pandecode_prop("depth_stride_zero = 0x%x",
1076 fbx->ds_linear.depth_stride_zero);
1077 pandecode_prop("stencil_stride_zero = 0x%x",
1078 fbx->ds_linear.stencil_stride_zero);
1079 }
1080
1081 pandecode_indent--;
1082 pandecode_log("},\n");
1083 }
1084
1085 if (fbx->clear_color_1 | fbx->clear_color_2) {
1086 pandecode_prop("clear_color_1 = 0x%" PRIx32, fbx->clear_color_1);
1087 pandecode_prop("clear_color_2 = 0x%" PRIx32, fbx->clear_color_2);
1088 }
1089
1090 if (fbx->zero3) {
1091 pandecode_msg("XXX: fb_extra zeros tripped\n");
1092 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1093 }
1094
1095 pandecode_indent--;
1096 pandecode_log("};\n");
1097
1098 gpu_va += sizeof(struct mali_framebuffer_extra);
1099 }
1100
1101 if (is_fragment)
1102 pandecode_render_target(gpu_va, job_no, fb);
1103
1104 return info;
1105 }
1106
1107 static void
1108 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1109 mali_ptr addr, int job_no, char *suffix,
1110 int count, bool varying, enum mali_job_type job_type)
1111 {
1112 char *prefix = varying ? "Varying" : "Attribute";
1113 assert(addr);
1114
1115 if (!count) {
1116 pandecode_msg("warn: No %s records\n", prefix);
1117 return;
1118 }
1119
1120 MAP_ADDR(ATTRIBUTE_BUFFER, addr, cl);
1121
1122 for (int i = 0; i < count; ++i) {
1123 fprintf(pandecode_dump_stream, "%s\n", prefix);
1124
1125 struct MALI_ATTRIBUTE_BUFFER temp;
1126 MALI_ATTRIBUTE_BUFFER_unpack(cl + i * MALI_ATTRIBUTE_BUFFER_LENGTH, &temp);
1127 MALI_ATTRIBUTE_BUFFER_print(pandecode_dump_stream, &temp, 2);
1128
1129 if (temp.type == MALI_ATTRIBUTE_TYPE_1D_NPOT_DIVISOR) {
1130 struct MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT temp2;
1131 MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT_unpack(cl + (i + 1) * MALI_ATTRIBUTE_BUFFER_LENGTH, &temp2);
1132 MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT_print(pandecode_dump_stream, &temp2, 2);
1133 }
1134 }
1135 }
1136
1137 static mali_ptr
1138 pandecode_shader_address(const char *name, mali_ptr ptr)
1139 {
1140 /* TODO: Decode flags */
1141 mali_ptr shader_ptr = ptr & ~15;
1142
1143 char *a = pointer_as_memory_reference(shader_ptr);
1144 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1145 free(a);
1146
1147 return shader_ptr;
1148 }
1149
1150 static void
1151 pandecode_blend_equation(const struct mali_blend_equation *blend)
1152 {
1153 if (blend->zero1)
1154 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1155
1156 pandecode_log(".equation = {\n");
1157 pandecode_indent++;
1158
1159 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1160 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1161
1162 pandecode_log(".color_mask = ");
1163 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1164 pandecode_log_cont(",\n");
1165
1166 pandecode_indent--;
1167 pandecode_log("},\n");
1168 }
1169
1170 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1171
1172 static unsigned
1173 decode_bifrost_constant(u16 constant)
1174 {
1175 float lo = (float) (constant & 0xFF);
1176 float hi = (float) (constant >> 8);
1177
1178 return (hi / 255.0) + (lo / 65535.0);
1179 }
1180
1181 static mali_ptr
1182 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1183 {
1184 struct bifrost_blend_rt *b =
1185 ((struct bifrost_blend_rt *) descs) + rt_no;
1186
1187 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1188 pandecode_indent++;
1189
1190 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1191 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1192 b->constant, decode_bifrost_constant(b->constant));
1193
1194 /* TODO figure out blend shader enable bit */
1195 pandecode_blend_equation(&b->equation);
1196
1197 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1198 pandecode_prop("index = 0x%" PRIx16, b->index);
1199
1200 pandecode_log(".format = %s", mali_format_as_str(b->format));
1201 pandecode_swizzle(b->swizzle, b->format);
1202 pandecode_log_cont(",\n");
1203
1204 pandecode_prop("swizzle = 0x%" PRIx32, b->swizzle);
1205 pandecode_prop("format = 0x%" PRIx32, b->format);
1206
1207 if (b->zero1) {
1208 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1209 pandecode_prop("zero1 = 0x%" PRIx32, b->zero1);
1210 }
1211
1212 pandecode_log(".shader_type = ");
1213 switch(b->shader_type) {
1214 case BIFROST_BLEND_F16:
1215 pandecode_log_cont("BIFROST_BLEND_F16");
1216 break;
1217 case BIFROST_BLEND_F32:
1218 pandecode_log_cont("BIFROST_BLEND_F32");
1219 break;
1220 case BIFROST_BLEND_I32:
1221 pandecode_log_cont("BIFROST_BLEND_I32");
1222 break;
1223 case BIFROST_BLEND_U32:
1224 pandecode_log_cont("BIFROST_BLEND_U32");
1225 break;
1226 case BIFROST_BLEND_I16:
1227 pandecode_log_cont("BIFROST_BLEND_I16");
1228 break;
1229 case BIFROST_BLEND_U16:
1230 pandecode_log_cont("BIFROST_BLEND_U16");
1231 break;
1232 }
1233 pandecode_log_cont(",\n");
1234
1235 if (b->zero2) {
1236 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1237 pandecode_prop("zero2 = 0x%" PRIx32, b->zero2);
1238 }
1239
1240 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1241
1242 pandecode_indent--;
1243 pandecode_log("},\n");
1244
1245 return 0;
1246 }
1247
1248 static mali_ptr
1249 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1250 {
1251 /* constant/equation is in a union */
1252 if (!blend->shader)
1253 return 0;
1254
1255 pandecode_log(".blend = {\n");
1256 pandecode_indent++;
1257
1258 if (is_shader) {
1259 pandecode_shader_address("shader", blend->shader);
1260 } else {
1261 pandecode_blend_equation(&blend->equation);
1262 pandecode_prop("constant = %f", blend->constant);
1263 }
1264
1265 pandecode_indent--;
1266 pandecode_log("},\n");
1267
1268 /* Return blend shader to disassemble if present */
1269 return is_shader ? (blend->shader & ~0xF) : 0;
1270 }
1271
1272 static mali_ptr
1273 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1274 {
1275 struct midgard_blend_rt *b =
1276 ((struct midgard_blend_rt *) descs) + rt_no;
1277
1278 /* Flags determine presence of blend shader */
1279 bool is_shader = (b->flags & 0xF) >= 0x2;
1280
1281 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1282 pandecode_indent++;
1283
1284 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1285
1286 union midgard_blend blend = b->blend;
1287 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1288
1289 pandecode_indent--;
1290 pandecode_log("};\n");
1291
1292 return shader;
1293 }
1294
1295 /* Attributes and varyings have descriptor records, which contain information
1296 * about their format and ordering with the attribute/varying buffers. We'll
1297 * want to validate that the combinations specified are self-consistent.
1298 */
1299
1300 static int
1301 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1302 {
1303 const char *prefix = varying ? "Varying" : "Attribute";
1304 mali_ptr p = varying ? v->varying_meta : v->attribute_meta;
1305
1306 for (int i = 0; i < count; ++i, p += MALI_ATTRIBUTE_LENGTH)
1307 DUMP_ADDR(prefix, ATTRIBUTE, p, 1);
1308
1309 return count;
1310 }
1311
1312 /* return bits [lo, hi) of word */
1313 static u32
1314 bits(u32 word, u32 lo, u32 hi)
1315 {
1316 if (hi - lo >= 32)
1317 return word; // avoid undefined behavior with the shift
1318
1319 return (word >> lo) & ((1 << (hi - lo)) - 1);
1320 }
1321
1322 static void
1323 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool graphics)
1324 {
1325 pandecode_log(".prefix = {\n");
1326 pandecode_indent++;
1327
1328 /* Decode invocation_count. See the comment before the definition of
1329 * invocation_count for an explanation.
1330 */
1331
1332 unsigned size_y_shift = bits(p->invocation_shifts, 0, 5);
1333 unsigned size_z_shift = bits(p->invocation_shifts, 5, 10);
1334 unsigned workgroups_x_shift = bits(p->invocation_shifts, 10, 16);
1335 unsigned workgroups_y_shift = bits(p->invocation_shifts, 16, 22);
1336 unsigned workgroups_z_shift = bits(p->invocation_shifts, 22, 28);
1337 unsigned workgroups_x_shift_2 = bits(p->invocation_shifts, 28, 32);
1338
1339 unsigned size_x = bits(p->invocation_count, 0, size_y_shift) + 1;
1340 unsigned size_y = bits(p->invocation_count, size_y_shift, size_z_shift) + 1;
1341 unsigned size_z = bits(p->invocation_count, size_z_shift, workgroups_x_shift) + 1;
1342
1343 unsigned groups_x = bits(p->invocation_count, workgroups_x_shift, workgroups_y_shift) + 1;
1344 unsigned groups_y = bits(p->invocation_count, workgroups_y_shift, workgroups_z_shift) + 1;
1345 unsigned groups_z = bits(p->invocation_count, workgroups_z_shift, 32) + 1;
1346
1347 /* Even though we have this decoded, we want to ensure that the
1348 * representation is "unique" so we don't lose anything by printing only
1349 * the final result. More specifically, we need to check that we were
1350 * passed something in canonical form, since the definition per the
1351 * hardware is inherently not unique. How? Well, take the resulting
1352 * decode and pack it ourselves! If it is bit exact with what we
1353 * decoded, we're good to go. */
1354
1355 struct mali_vertex_tiler_prefix ref;
1356 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, graphics);
1357
1358 bool canonical =
1359 (p->invocation_count == ref.invocation_count) &&
1360 (p->invocation_shifts == ref.invocation_shifts);
1361
1362 if (!canonical) {
1363 pandecode_msg("XXX: non-canonical workgroups packing\n");
1364 pandecode_msg("expected: %X, %X",
1365 ref.invocation_count,
1366 ref.invocation_shifts);
1367
1368 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1369 pandecode_prop("size_y_shift = %d", size_y_shift);
1370 pandecode_prop("size_z_shift = %d", size_z_shift);
1371 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift);
1372 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift);
1373 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift);
1374 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2);
1375 }
1376
1377 /* Regardless, print the decode */
1378 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1379 size_x, size_y, size_z,
1380 groups_x, groups_y, groups_z);
1381
1382 /* TODO: Decode */
1383 if (p->unknown_draw)
1384 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1385
1386 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1387
1388 if (p->draw_mode != MALI_DRAW_MODE_NONE)
1389 pandecode_prop("draw_mode = %s", mali_draw_mode_as_str(p->draw_mode));
1390
1391 /* Index count only exists for tiler jobs anyway */
1392
1393 if (p->index_count)
1394 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1395
1396
1397 unsigned index_raw_size = (p->unknown_draw & MALI_DRAW_INDEXED_SIZE);
1398 index_raw_size >>= MALI_DRAW_INDEXED_SHIFT;
1399
1400 /* Validate an index buffer is present if we need one. TODO: verify
1401 * relationship between invocation_count and index_count */
1402
1403 if (p->indices) {
1404 unsigned count = p->index_count;
1405
1406 /* Grab the size */
1407 unsigned size = (index_raw_size == 0x3) ? 4 : index_raw_size;
1408
1409 /* Ensure we got a size, and if so, validate the index buffer
1410 * is large enough to hold a full set of indices of the given
1411 * size */
1412
1413 if (!index_raw_size)
1414 pandecode_msg("XXX: index size missing\n");
1415 else
1416 pandecode_validate_buffer(p->indices, count * size);
1417 } else if (index_raw_size)
1418 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size);
1419
1420 if (p->offset_bias_correction)
1421 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1422
1423 /* TODO: Figure out what this is. It's not zero */
1424 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1425
1426 pandecode_indent--;
1427 pandecode_log("},\n");
1428 }
1429
1430 static void
1431 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1432 {
1433 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1434 uint64_t *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1435
1436 for (int i = 0; i < ubufs_count; i++) {
1437 unsigned size = (ubufs[i] & ((1 << 10) - 1)) * 16;
1438 mali_ptr addr = (ubufs[i] >> 10) << 2;
1439
1440 pandecode_validate_buffer(addr, size);
1441
1442 char *ptr = pointer_as_memory_reference(addr);
1443 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
1444 free(ptr);
1445 }
1446
1447 pandecode_log("\n");
1448 }
1449
1450 static void
1451 pandecode_uniforms(mali_ptr uniforms, unsigned uniform_count)
1452 {
1453 pandecode_validate_buffer(uniforms, uniform_count * 16);
1454
1455 char *ptr = pointer_as_memory_reference(uniforms);
1456 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count, ptr);
1457 free(ptr);
1458 }
1459
1460 static const char *
1461 shader_type_for_job(unsigned type)
1462 {
1463 switch (type) {
1464 case MALI_JOB_TYPE_VERTEX: return "VERTEX";
1465 case MALI_JOB_TYPE_TILER: return "FRAGMENT";
1466 case MALI_JOB_TYPE_COMPUTE: return "COMPUTE";
1467 default:
1468 return "UNKNOWN";
1469 }
1470 }
1471
1472 static unsigned shader_id = 0;
1473
1474 static struct midgard_disasm_stats
1475 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1476 bool is_bifrost, unsigned gpu_id)
1477 {
1478 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1479 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1480
1481 /* Compute maximum possible size */
1482 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1483
1484 /* Print some boilerplate to clearly denote the assembly (which doesn't
1485 * obey indentation rules), and actually do the disassembly! */
1486
1487 pandecode_log_cont("\n\n");
1488
1489 struct midgard_disasm_stats stats;
1490
1491 if (is_bifrost) {
1492 disassemble_bifrost(pandecode_dump_stream, code, sz, true);
1493
1494 /* TODO: Extend stats to Bifrost */
1495 stats.texture_count = -128;
1496 stats.sampler_count = -128;
1497 stats.attribute_count = -128;
1498 stats.varying_count = -128;
1499 stats.uniform_count = -128;
1500 stats.uniform_buffer_count = -128;
1501 stats.work_count = -128;
1502
1503 stats.instruction_count = 0;
1504 stats.bundle_count = 0;
1505 stats.quadword_count = 0;
1506 stats.helper_invocations = false;
1507 } else {
1508 stats = disassemble_midgard(pandecode_dump_stream,
1509 code, sz, gpu_id,
1510 type == MALI_JOB_TYPE_TILER ?
1511 MESA_SHADER_FRAGMENT : MESA_SHADER_VERTEX);
1512 }
1513
1514 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
1515 * driver-internal purposes with the blob and interfere */
1516
1517 bool should_shaderdb = type != MALI_JOB_TYPE_COMPUTE;
1518
1519 if (should_shaderdb) {
1520 unsigned nr_threads =
1521 (stats.work_count <= 4) ? 4 :
1522 (stats.work_count <= 8) ? 2 :
1523 1;
1524
1525 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
1526 "%u inst, %u bundles, %u quadwords, "
1527 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
1528 shader_id++,
1529 shader_type_for_job(type),
1530 stats.instruction_count, stats.bundle_count, stats.quadword_count,
1531 stats.work_count, nr_threads);
1532 }
1533
1534
1535 return stats;
1536 }
1537
1538 static void
1539 pandecode_texture_payload(mali_ptr payload,
1540 enum mali_texture_dimension dim,
1541 enum mali_texture_layout layout,
1542 bool manual_stride,
1543 uint8_t levels,
1544 uint16_t depth,
1545 uint16_t array_size,
1546 struct pandecode_mapped_memory *tmem)
1547 {
1548 pandecode_log(".payload = {\n");
1549 pandecode_indent++;
1550
1551 /* A bunch of bitmap pointers follow.
1552 * We work out the correct number,
1553 * based on the mipmap/cubemap
1554 * properties, but dump extra
1555 * possibilities to futureproof */
1556
1557 int bitmap_count = levels + 1;
1558
1559 /* Miptree for each face */
1560 if (dim == MALI_TEXTURE_DIMENSION_CUBE)
1561 bitmap_count *= 6;
1562
1563 /* Array of layers */
1564 bitmap_count *= depth;
1565
1566 /* Array of textures */
1567 bitmap_count *= array_size;
1568
1569 /* Stride for each element */
1570 if (manual_stride)
1571 bitmap_count *= 2;
1572
1573 mali_ptr *pointers_and_strides = pandecode_fetch_gpu_mem(tmem,
1574 payload, sizeof(mali_ptr) * bitmap_count);
1575 for (int i = 0; i < bitmap_count; ++i) {
1576 /* How we dump depends if this is a stride or a pointer */
1577
1578 if (manual_stride && (i & 1)) {
1579 /* signed 32-bit snuck in as a 64-bit pointer */
1580 uint64_t stride_set = pointers_and_strides[i];
1581 uint32_t clamped_stride = stride_set;
1582 int32_t stride = clamped_stride;
1583 assert(stride_set == clamped_stride);
1584 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
1585 } else {
1586 char *a = pointer_as_memory_reference(pointers_and_strides[i]);
1587 pandecode_log("%s, \n", a);
1588 free(a);
1589 }
1590 }
1591
1592 pandecode_indent--;
1593 pandecode_log("},\n");
1594 }
1595
1596 static void
1597 pandecode_texture(mali_ptr u,
1598 struct pandecode_mapped_memory *tmem,
1599 unsigned job_no, unsigned tex)
1600 {
1601 struct pandecode_mapped_memory *mapped_mem = pandecode_find_mapped_gpu_mem_containing(u);
1602 const uint8_t *cl = pandecode_fetch_gpu_mem(mapped_mem, u, MALI_MIDGARD_TEXTURE_LENGTH);
1603
1604 struct MALI_MIDGARD_TEXTURE temp;
1605 MALI_MIDGARD_TEXTURE_unpack(cl, &temp);
1606 MALI_MIDGARD_TEXTURE_print(pandecode_dump_stream, &temp, 2);
1607
1608 pandecode_texture_payload(u + MALI_MIDGARD_TEXTURE_LENGTH,
1609 temp.dimension, temp.texel_ordering, temp.manual_stride,
1610 temp.levels, temp.depth, temp.array_size, mapped_mem);
1611 }
1612
1613 static void
1614 pandecode_bifrost_texture(
1615 const void *cl,
1616 unsigned job_no,
1617 unsigned tex)
1618 {
1619 struct MALI_BIFROST_TEXTURE temp;
1620 MALI_BIFROST_TEXTURE_unpack(cl, &temp);
1621 MALI_BIFROST_TEXTURE_print(pandecode_dump_stream, &temp, 2);
1622
1623 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(temp.surfaces);
1624 pandecode_texture_payload(temp.surfaces, temp.dimension, temp.texel_ordering,
1625 true, temp.levels, 1, 1, tmem);
1626 }
1627
1628 /* 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 */
1629
1630 static void
1631 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
1632 {
1633 /* Nothing to do */
1634 if (claim == truth)
1635 return;
1636
1637 if (fuzzy && (truth < 0))
1638 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name, claim, truth);
1639
1640 if ((truth >= 0) && !fuzzy) {
1641 pandecode_msg("%s: expected %s = %d, claimed %u\n",
1642 (truth < claim) ? "warn" : "XXX",
1643 name, truth, claim);
1644 } else if ((claim > -truth) && !fuzzy) {
1645 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
1646 name, -truth, claim);
1647 } else if (fuzzy && (claim < truth))
1648 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
1649 name, truth, claim);
1650
1651 pandecode_log(".%s = %" PRId16, name, claim);
1652
1653 if (fuzzy)
1654 pandecode_log_cont(" /* %u used */", truth);
1655
1656 pandecode_log_cont(",\n");
1657 }
1658
1659 static void
1660 pandecode_blend_shader_disassemble(mali_ptr shader, int job_no, int job_type,
1661 bool is_bifrost, unsigned gpu_id)
1662 {
1663 struct midgard_disasm_stats stats =
1664 pandecode_shader_disassemble(shader, job_no, job_type, is_bifrost, gpu_id);
1665
1666 bool has_texture = (stats.texture_count > 0);
1667 bool has_sampler = (stats.sampler_count > 0);
1668 bool has_attribute = (stats.attribute_count > 0);
1669 bool has_varying = (stats.varying_count > 0);
1670 bool has_uniform = (stats.uniform_count > 0);
1671 bool has_ubo = (stats.uniform_buffer_count > 0);
1672
1673 if (has_texture || has_sampler)
1674 pandecode_msg("XXX: blend shader accessing textures\n");
1675
1676 if (has_attribute || has_varying)
1677 pandecode_msg("XXX: blend shader accessing interstage\n");
1678
1679 if (has_uniform || has_ubo)
1680 pandecode_msg("XXX: blend shader accessing uniforms\n");
1681 }
1682
1683 static void
1684 pandecode_textures(mali_ptr textures, unsigned texture_count, int job_no, bool is_bifrost)
1685 {
1686 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(textures);
1687
1688 if (!mmem)
1689 return;
1690
1691 pandecode_log("Textures (%"PRIx64"):\n", textures);
1692
1693 if (is_bifrost) {
1694 const void *cl = pandecode_fetch_gpu_mem(mmem,
1695 textures, MALI_BIFROST_TEXTURE_LENGTH *
1696 texture_count);
1697
1698 for (unsigned tex = 0; tex < texture_count; ++tex) {
1699 pandecode_bifrost_texture(cl +
1700 MALI_BIFROST_TEXTURE_LENGTH * tex,
1701 job_no, tex);
1702 }
1703 } else {
1704 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures);
1705
1706 for (int tex = 0; tex < texture_count; ++tex) {
1707 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
1708 char *a = pointer_as_memory_reference(*u);
1709 pandecode_log("%s,\n", a);
1710 free(a);
1711 }
1712
1713 /* Now, finally, descend down into the texture descriptor */
1714 for (unsigned tex = 0; tex < texture_count; ++tex) {
1715 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
1716 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
1717 if (tmem)
1718 pandecode_texture(*u, tmem, job_no, tex);
1719 }
1720 }
1721 }
1722
1723 static void
1724 pandecode_samplers(mali_ptr samplers, unsigned sampler_count, int job_no, bool is_bifrost)
1725 {
1726 for (int i = 0; i < sampler_count; ++i) {
1727 if (is_bifrost) {
1728 DUMP_ADDR("Sampler", BIFROST_SAMPLER, samplers + (MALI_BIFROST_SAMPLER_LENGTH * i), 1);
1729 } else {
1730 DUMP_ADDR("Sampler", MIDGARD_SAMPLER, samplers + (MALI_MIDGARD_SAMPLER_LENGTH * i), 1);
1731 }
1732 }
1733 }
1734
1735 static void
1736 pandecode_vertex_tiler_postfix_pre(
1737 const struct mali_vertex_tiler_postfix *p,
1738 int job_no, enum mali_job_type job_type,
1739 char *suffix, bool is_bifrost, unsigned gpu_id)
1740 {
1741 struct pandecode_mapped_memory *attr_mem;
1742
1743 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1744 * are the only things actually needed from the FBD, vertex/tiler jobs
1745 * no longer reference the FBD -- instead, this field points to some
1746 * info about the scratchpad.
1747 */
1748
1749 struct pandecode_fbd fbd_info = {
1750 /* Default for Bifrost */
1751 .rt_count = 1
1752 };
1753
1754 if (is_bifrost) {
1755 pandecode_log_cont("\t/* %X %/\n", p->shared_memory & 1);
1756 pandecode_compute_fbd(p->shared_memory & ~1, job_no);
1757 } else if (p->shared_memory & MALI_MFBD)
1758 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->shared_memory) & FBD_MASK, job_no, false, job_type == MALI_JOB_TYPE_COMPUTE, false);
1759 else if (job_type == MALI_JOB_TYPE_COMPUTE)
1760 pandecode_compute_fbd((u64) (uintptr_t) p->shared_memory, job_no);
1761 else
1762 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->shared_memory, job_no, false, gpu_id);
1763
1764 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
1765 int texture_count = 0, sampler_count = 0;
1766
1767 if (p->shader) {
1768 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->shader);
1769 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, p->shader);
1770
1771 /* Disassemble ahead-of-time to get stats. Initialize with
1772 * stats for the missing-shader case so we get validation
1773 * there, too */
1774
1775 struct midgard_disasm_stats info = {
1776 .texture_count = 0,
1777 .sampler_count = 0,
1778 .attribute_count = 0,
1779 .varying_count = 0,
1780 .work_count = 1,
1781
1782 .uniform_count = -128,
1783 .uniform_buffer_count = 0
1784 };
1785
1786 if (s->shader & ~0xF)
1787 info = pandecode_shader_disassemble(s->shader & ~0xF, job_no, job_type, is_bifrost, gpu_id);
1788
1789 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", p->shader, job_no, suffix);
1790 pandecode_indent++;
1791
1792 /* Save for dumps */
1793 attribute_count = s->attribute_count;
1794 varying_count = s->varying_count;
1795 texture_count = s->texture_count;
1796 sampler_count = s->sampler_count;
1797
1798 if (is_bifrost) {
1799 uniform_count = s->bifrost2.uniform_count;
1800 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
1801 } else {
1802 uniform_count = s->midgard1.uniform_count;
1803 uniform_buffer_count = s->midgard1.uniform_buffer_count;
1804 }
1805
1806 pandecode_shader_address("shader", s->shader);
1807
1808 pandecode_shader_prop("texture_count", s->texture_count, info.texture_count, false);
1809 pandecode_shader_prop("sampler_count", s->sampler_count, info.sampler_count, false);
1810 pandecode_shader_prop("attribute_count", s->attribute_count, info.attribute_count, false);
1811 pandecode_shader_prop("varying_count", s->varying_count, info.varying_count, false);
1812 pandecode_shader_prop("uniform_buffer_count",
1813 uniform_buffer_count,
1814 info.uniform_buffer_count, true);
1815
1816 if (!is_bifrost) {
1817 pandecode_shader_prop("uniform_count",
1818 uniform_count,
1819 info.uniform_count, false);
1820
1821 pandecode_shader_prop("work_count",
1822 s->midgard1.work_count, info.work_count, false);
1823 }
1824
1825 if (is_bifrost) {
1826 pandecode_log("bifrost1.unk1 = ");
1827 pandecode_log_decoded_flags(shader_bifrost_info, s->bifrost1.unk1);
1828 pandecode_log_cont(",\n");
1829 } else {
1830 bool helpers = s->midgard1.flags_lo & MALI_HELPER_INVOCATIONS;
1831
1832 if (helpers != info.helper_invocations) {
1833 pandecode_msg("XXX: expected helpers %u but got %u\n",
1834 info.helper_invocations, helpers);
1835 }
1836
1837 pandecode_log(".midgard1.flags_lo = ");
1838 pandecode_log_decoded_flags(shader_midgard1_flag_lo_info,
1839 s->midgard1.flags_lo & ~MALI_HELPER_INVOCATIONS);
1840 pandecode_log_cont(",\n");
1841
1842 pandecode_log(".midgard1.flags_hi = ");
1843 pandecode_log_decoded_flags(shader_midgard1_flag_hi_info, s->midgard1.flags_hi);
1844 pandecode_log_cont(",\n");
1845 }
1846
1847 if (s->depth_units || s->depth_factor) {
1848 pandecode_prop("depth_factor = %f", s->depth_factor);
1849 pandecode_prop("depth_units = %f", s->depth_units);
1850 }
1851
1852 if (s->coverage_mask)
1853 pandecode_prop("coverage_mask = 0x%X", s->coverage_mask);
1854
1855 if (s->unknown2_2)
1856 pandecode_prop(".unknown2_2 = %X", s->unknown2_2);
1857
1858 if (s->unknown2_3 || s->unknown2_4) {
1859 pandecode_log(".unknown2_3 = ");
1860
1861 int unknown2_3 = s->unknown2_3;
1862 int unknown2_4 = s->unknown2_4;
1863
1864 /* We're not quite sure what these flags mean without the depth test, if anything */
1865
1866 if (unknown2_3 & (MALI_DEPTH_WRITEMASK | MALI_DEPTH_FUNC_MASK)) {
1867 const char *func = mali_func_as_str(MALI_GET_DEPTH_FUNC(unknown2_3));
1868 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
1869
1870 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
1871 }
1872
1873 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
1874 pandecode_log_cont(",\n");
1875
1876 pandecode_log(".unknown2_4 = ");
1877 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
1878 pandecode_log_cont(",\n");
1879 }
1880
1881 if (s->stencil_mask_front || s->stencil_mask_back) {
1882 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
1883 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
1884 }
1885
1886 DUMP_CL("Stencil front", STENCIL, &s->stencil_front, 1);
1887 DUMP_CL("Stencil back", STENCIL, &s->stencil_back, 1);
1888
1889 if (is_bifrost) {
1890 pandecode_log(".bifrost2 = {\n");
1891 pandecode_indent++;
1892
1893 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
1894 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
1895 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
1896 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
1897
1898 pandecode_indent--;
1899 pandecode_log("},\n");
1900 } else if (s->midgard2.unknown2_7) {
1901 pandecode_log(".midgard2 = {\n");
1902 pandecode_indent++;
1903
1904 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
1905 pandecode_indent--;
1906 pandecode_log("},\n");
1907 }
1908
1909 if (s->padding) {
1910 pandecode_msg("XXX: shader padding tripped\n");
1911 pandecode_prop("padding = 0x%" PRIx32, s->padding);
1912 }
1913
1914 if (!is_bifrost) {
1915 /* TODO: Blend shaders routing/disasm */
1916 union midgard_blend blend = s->blend;
1917 mali_ptr shader = pandecode_midgard_blend(&blend, s->unknown2_3 & MALI_HAS_BLEND_SHADER);
1918 if (shader & ~0xF)
1919 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
1920 } else {
1921 pandecode_msg("mdg_blend = %" PRIx64 "\n", s->blend.shader);
1922 }
1923
1924 pandecode_indent--;
1925 pandecode_log("};\n");
1926
1927 /* MRT blend fields are used whenever MFBD is used, with
1928 * per-RT descriptors */
1929
1930 if (job_type == MALI_JOB_TYPE_TILER && (is_bifrost || p->shared_memory & MALI_MFBD)) {
1931 void* blend_base = (void *) (s + 1);
1932
1933 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
1934 mali_ptr shader = 0;
1935
1936 if (is_bifrost)
1937 shader = pandecode_bifrost_blend(blend_base, job_no, i);
1938 else
1939 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
1940
1941 if (shader & ~0xF)
1942 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
1943
1944 }
1945 }
1946 } else
1947 pandecode_msg("XXX: missing shader descriptor\n");
1948
1949 if (p->viewport)
1950 DUMP_ADDR("Viewport", VIEWPORT, p->viewport, 1);
1951
1952 unsigned max_attr_index = 0;
1953
1954 if (p->attribute_meta)
1955 max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
1956
1957 if (p->attributes) {
1958 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
1959 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false, job_type);
1960 }
1961
1962 /* Varyings are encoded like attributes but not actually sent; we just
1963 * pass a zero buffer with the right stride/size set, (or whatever)
1964 * since the GPU will write to it itself */
1965
1966 if (p->varying_meta) {
1967 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
1968 }
1969
1970 if (p->varyings) {
1971 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
1972
1973 /* Number of descriptors depends on whether there are
1974 * non-internal varyings */
1975
1976 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true, job_type);
1977 }
1978
1979 if (p->uniform_buffers) {
1980 if (uniform_buffer_count)
1981 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
1982 else
1983 pandecode_msg("warn: UBOs specified but not referenced\n");
1984 } else if (uniform_buffer_count)
1985 pandecode_msg("XXX: UBOs referenced but not specified\n");
1986
1987 /* We don't want to actually dump uniforms, but we do need to validate
1988 * that the counts we were given are sane */
1989
1990 if (p->uniforms) {
1991 if (uniform_count)
1992 pandecode_uniforms(p->uniforms, uniform_count);
1993 else
1994 pandecode_msg("warn: Uniforms specified but not referenced\n");
1995 } else if (uniform_count)
1996 pandecode_msg("XXX: Uniforms referenced but not specified\n");
1997
1998 if (p->textures)
1999 pandecode_textures(p->textures, texture_count, job_no, is_bifrost);
2000
2001 if (p->sampler_descriptor)
2002 pandecode_samplers(p->sampler_descriptor, sampler_count, job_no, is_bifrost);
2003 }
2004
2005 static void
2006 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2007 {
2008 pandecode_log(".gl_enables = ");
2009
2010 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2011
2012 pandecode_log_cont(",\n");
2013 }
2014
2015 static void
2016 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2017 {
2018 if (p->shader & 0xF)
2019 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p->shader & 0xF));
2020
2021 pandecode_log(".postfix = {\n");
2022 pandecode_indent++;
2023
2024 pandecode_gl_enables(p->gl_enables, MALI_JOB_TYPE_TILER);
2025 pandecode_prop("instance_shift = 0x%x", p->instance_shift);
2026 pandecode_prop("instance_odd = 0x%x", p->instance_odd);
2027
2028 if (p->zero4) {
2029 pandecode_msg("XXX: vertex only zero tripped");
2030 pandecode_prop("zero4 = 0x%" PRIx32, p->zero4);
2031 }
2032
2033 pandecode_prop("offset_start = 0x%x", p->offset_start);
2034
2035 if (p->zero5) {
2036 pandecode_msg("XXX: vertex only zero tripped");
2037 pandecode_prop("zero5 = 0x%" PRIx64, p->zero5);
2038 }
2039
2040 MEMORY_PROP(p, position_varying);
2041 MEMORY_PROP(p, occlusion_counter);
2042
2043 pandecode_indent--;
2044 pandecode_log("},\n");
2045 }
2046
2047 static void
2048 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2049 {
2050 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2051 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2052
2053 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
2054 pandecode_indent++;
2055
2056 if (h->zero) {
2057 pandecode_msg("XXX: tiler heap zero tripped\n");
2058 pandecode_prop("zero = 0x%x", h->zero);
2059 }
2060
2061 pandecode_prop("heap_size = 0x%x", h->heap_size);
2062 MEMORY_PROP(h, tiler_heap_start);
2063 MEMORY_PROP(h, tiler_heap_free);
2064
2065 /* this might point to the beginning of another buffer, when it's
2066 * really the end of the tiler heap buffer, so we have to be careful
2067 * here. but for zero length, we need the same pointer.
2068 */
2069
2070 if (h->tiler_heap_end == h->tiler_heap_start) {
2071 MEMORY_PROP(h, tiler_heap_start);
2072 } else {
2073 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2074 pandecode_prop("tiler_heap_end = %s + 1", a);
2075 free(a);
2076 }
2077
2078 for (int i = 0; i < 10; i++) {
2079 if (h->zeros[i] != 0) {
2080 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2081 i, h->zeros[i]);
2082 }
2083 }
2084
2085 if (h->unk1 != 0x1) {
2086 pandecode_msg("XXX: tiler heap unk1 tripped\n");
2087 pandecode_prop("unk1 = 0x%x", h->unk1);
2088 }
2089
2090 if (h->unk7e007e != 0x7e007e) {
2091 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
2092 pandecode_prop("unk7e007e = 0x%x", h->unk7e007e);
2093 }
2094
2095 pandecode_indent--;
2096 pandecode_log("};\n");
2097 }
2098
2099 static void
2100 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2101 {
2102 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2103 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2104
2105 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2106
2107 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
2108 pandecode_indent++;
2109
2110 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32, t->tiler_heap_next_start);
2111 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32, t->used_hierarchy_mask);
2112
2113 if (t->hierarchy_mask != 0xa &&
2114 t->hierarchy_mask != 0x14 &&
2115 t->hierarchy_mask != 0x28 &&
2116 t->hierarchy_mask != 0x50 &&
2117 t->hierarchy_mask != 0xa0)
2118 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
2119
2120 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2121
2122 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2123
2124 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2125 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2126
2127 if (t->zero0) {
2128 pandecode_msg("XXX: tiler meta zero tripped\n");
2129 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2130 }
2131
2132 for (int i = 0; i < 12; i++) {
2133 if (t->zeros[i] != 0) {
2134 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2135 i, t->zeros[i]);
2136 }
2137 }
2138
2139 pandecode_indent--;
2140 pandecode_log("};\n");
2141 }
2142
2143 static void
2144 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2145 {
2146 if (u.pointer == 0x0)
2147 return;
2148
2149 pandecode_log(".primitive_size = {\n");
2150 pandecode_indent++;
2151
2152 if (constant) {
2153 pandecode_prop("constant = %f", u.constant);
2154 } else {
2155 MEMORY_PROP((&u), pointer);
2156 }
2157
2158 pandecode_indent--;
2159 pandecode_log("},\n");
2160 }
2161
2162 static void
2163 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2164 {
2165 pandecode_log_cont("{\n");
2166 pandecode_indent++;
2167
2168 /* TODO: gl_PointSize on Bifrost */
2169 pandecode_primitive_size(t->primitive_size, true);
2170
2171 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2172 || t->zero6) {
2173 pandecode_msg("XXX: tiler only zero tripped\n");
2174 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2175 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2176 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2177 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2178 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2179 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2180 }
2181
2182 pandecode_indent--;
2183 pandecode_log("},\n");
2184 }
2185
2186 static int
2187 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2188 const struct pandecode_mapped_memory *mem,
2189 mali_ptr payload, int job_no, unsigned gpu_id)
2190 {
2191 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2192
2193 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true, gpu_id);
2194
2195 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64"_%d = {\n", payload, job_no);
2196 pandecode_indent++;
2197
2198 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2199 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2200
2201 pandecode_indent--;
2202 pandecode_log("};\n");
2203
2204 return sizeof(*v);
2205 }
2206
2207 static int
2208 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2209 const struct pandecode_mapped_memory *mem,
2210 mali_ptr payload, int job_no, unsigned gpu_id)
2211 {
2212 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2213
2214 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true, gpu_id);
2215 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2216
2217 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64"_%d = {\n", payload, job_no);
2218 pandecode_indent++;
2219
2220 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
2221
2222 pandecode_log(".tiler = ");
2223 pandecode_tiler_only_bfr(&t->tiler, job_no);
2224
2225 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2226
2227 pandecode_indent--;
2228 pandecode_log("};\n");
2229
2230 return sizeof(*t);
2231 }
2232
2233 static int
2234 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2235 const struct pandecode_mapped_memory *mem,
2236 mali_ptr payload, int job_no, unsigned gpu_id)
2237 {
2238 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2239 bool is_graphics = (h->job_type == MALI_JOB_TYPE_VERTEX) || (h->job_type == MALI_JOB_TYPE_TILER);
2240
2241 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false, gpu_id);
2242
2243 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2244 pandecode_indent++;
2245
2246 pandecode_vertex_tiler_prefix(&v->prefix, job_no, is_graphics);
2247 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2248
2249 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2250 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
2251
2252 pandecode_indent--;
2253 pandecode_log("};\n");
2254
2255 return sizeof(*v);
2256 }
2257
2258 static int
2259 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2260 mali_ptr payload, int job_no,
2261 bool is_bifrost, unsigned gpu_id)
2262 {
2263 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2264
2265 bool is_mfbd = s->framebuffer & MALI_MFBD;
2266
2267 if (!is_mfbd && is_bifrost)
2268 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2269
2270 struct pandecode_fbd info;
2271
2272 if (is_mfbd)
2273 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true, false, is_bifrost);
2274 else
2275 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true, gpu_id);
2276
2277 /* Compute the tag for the tagged pointer. This contains the type of
2278 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2279 * additional structures follow the MFBD header (an extra payload or
2280 * not, as well as a count of render targets) */
2281
2282 unsigned expected_tag = is_mfbd ? MALI_MFBD : 0;
2283
2284 if (is_mfbd) {
2285 if (info.has_extra)
2286 expected_tag |= MALI_MFBD_TAG_EXTRA;
2287
2288 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
2289 }
2290
2291 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
2292 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2293 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
2294 pandecode_prop("max_tile_coord = 0x%X\n", s->max_tile_coord);
2295 }
2296
2297 /* Extract tile coordinates */
2298
2299 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
2300 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
2301
2302 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2303 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2304
2305 /* For the max, we also want the floored (rather than ceiled) version for checking */
2306
2307 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
2308 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
2309
2310 /* Validate the coordinates are well-ordered */
2311
2312 if (min_x == max_x)
2313 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2314 else if (min_x > max_x)
2315 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2316
2317 if (min_y == max_y)
2318 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2319 else if (min_y > max_y)
2320 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2321
2322 /* Validate the coordinates fit inside the framebuffer. We use floor,
2323 * rather than ceil, for the max coordinates, since the tile
2324 * coordinates for something like an 800x600 framebuffer will actually
2325 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2326
2327 if ((min_x > info.width) || (max_x_f > info.width))
2328 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2329
2330 if ((min_y > info.height) || (max_y_f > info.height))
2331 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2332
2333 /* After validation, we print */
2334
2335 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
2336
2337 /* The FBD is a tagged pointer */
2338
2339 unsigned tag = (s->framebuffer & ~FBD_MASK);
2340
2341 if (tag != expected_tag)
2342 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
2343
2344 return sizeof(*s);
2345 }
2346
2347 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
2348 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
2349 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
2350 * might wish to combine this with the bifrost parameter) because some details
2351 * are model-specific even within a particular architecture. Minimal traces
2352 * *only* examine the job descriptors, skipping printing entirely if there is
2353 * no faults, and only descends into the payload if there are faults. This is
2354 * useful for looking for faults without the overhead of invasive traces. */
2355
2356 void
2357 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost, unsigned gpu_id, bool minimal)
2358 {
2359 pandecode_dump_file_open();
2360
2361 struct mali_job_descriptor_header *h;
2362 unsigned job_descriptor_number = 0;
2363
2364 do {
2365 struct pandecode_mapped_memory *mem =
2366 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2367
2368 void *payload;
2369
2370 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2371
2372 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2373 * high 32-bits of the 64-bit pointer are reused to store
2374 * something else.
2375 */
2376 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2377 h->job_type != MALI_JOB_TYPE_FRAGMENT ? 4 : 0;
2378 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2379
2380 payload = pandecode_fetch_gpu_mem(mem, payload_ptr, 256);
2381
2382 int job_no = job_descriptor_number++;
2383
2384 /* If the job is good to go, skip it in minimal mode */
2385 if (minimal && (h->exception_status == 0x0 || h->exception_status == 0x1))
2386 continue;
2387
2388 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2389 pandecode_indent++;
2390
2391 pandecode_prop("job_type = %s", mali_job_type_as_str(h->job_type));
2392
2393 if (h->job_descriptor_size)
2394 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2395
2396 if (h->exception_status && h->exception_status != 0x1)
2397 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2398 h->exception_status,
2399 (h->exception_status >> 16) & 0xFFFF,
2400 mali_exception_access_as_str((h->exception_status >> 8) & 0x3),
2401 h->exception_status & 0xFF);
2402
2403 if (h->first_incomplete_task)
2404 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2405
2406 if (h->fault_pointer)
2407 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2408
2409 if (h->job_barrier)
2410 pandecode_prop("job_barrier = %d", h->job_barrier);
2411
2412 pandecode_prop("job_index = %d", h->job_index);
2413
2414 if (h->unknown_flags)
2415 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2416
2417 if (h->job_dependency_index_1)
2418 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2419
2420 if (h->job_dependency_index_2)
2421 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2422
2423 pandecode_indent--;
2424 pandecode_log("};\n");
2425
2426 switch (h->job_type) {
2427 case MALI_JOB_TYPE_WRITE_VALUE: {
2428 struct mali_payload_write_value *s = payload;
2429 pandecode_log("struct mali_payload_write_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2430 pandecode_indent++;
2431 MEMORY_PROP(s, address);
2432
2433 if (s->value_descriptor != MALI_WRITE_VALUE_ZERO) {
2434 pandecode_msg("XXX: unknown value descriptor\n");
2435 pandecode_prop("value_descriptor = 0x%" PRIX32, s->value_descriptor);
2436 }
2437
2438 if (s->reserved) {
2439 pandecode_msg("XXX: set value tripped\n");
2440 pandecode_prop("reserved = 0x%" PRIX32, s->reserved);
2441 }
2442
2443 pandecode_prop("immediate = 0x%" PRIX64, s->immediate);
2444 pandecode_indent--;
2445 pandecode_log("};\n");
2446
2447 break;
2448 }
2449
2450 case MALI_JOB_TYPE_TILER:
2451 case MALI_JOB_TYPE_VERTEX:
2452 case MALI_JOB_TYPE_COMPUTE:
2453 if (bifrost) {
2454 if (h->job_type == MALI_JOB_TYPE_TILER)
2455 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2456 else
2457 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2458 } else
2459 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no, gpu_id);
2460
2461 break;
2462
2463 case MALI_JOB_TYPE_FRAGMENT:
2464 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost, gpu_id);
2465 break;
2466
2467 default:
2468 break;
2469 }
2470 } while ((jc_gpu_va = h->next_job));
2471
2472 pandecode_map_read_write();
2473 }