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