22bc165da8f43c6ccf5e15bb4d147d1aa3e9e78a
[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
492 static struct pandecode_fbd
493 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment, unsigned gpu_id)
494 {
495 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
496 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
497
498 struct pandecode_fbd info = {
499 .has_extra = false,
500 .rt_count = 1
501 };
502
503 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
504 pandecode_indent++;
505
506 pandecode_log(".shared_memory = {\n");
507 pandecode_indent++;
508 pandecode_shared_memory(&s->shared_memory, false);
509 pandecode_indent--;
510 pandecode_log("},\n");
511
512 pandecode_sfbd_format(s->format);
513
514 info.width = s->width + 1;
515 info.height = s->height + 1;
516
517 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
518 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
519
520 MEMORY_PROP(s, checksum);
521
522 if (s->checksum_stride)
523 pandecode_prop("checksum_stride = %d", s->checksum_stride);
524
525 MEMORY_PROP(s, framebuffer);
526 pandecode_prop("stride = %d", s->stride);
527
528 /* Earlier in the actual commandstream -- right before width -- but we
529 * delay to flow nicer */
530
531 pandecode_log(".clear_flags = ");
532 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
533 pandecode_log_cont(",\n");
534
535 if (s->depth_buffer) {
536 MEMORY_PROP(s, depth_buffer);
537 pandecode_prop("depth_stride = %d", s->depth_stride);
538 }
539
540 if (s->stencil_buffer) {
541 MEMORY_PROP(s, stencil_buffer);
542 pandecode_prop("stencil_stride = %d", s->stencil_stride);
543 }
544
545 if (s->depth_stride_zero ||
546 s->stencil_stride_zero ||
547 s->zero7 || s->zero8) {
548 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
549 pandecode_prop("depth_stride_zero = 0x%x",
550 s->depth_stride_zero);
551 pandecode_prop("stencil_stride_zero = 0x%x",
552 s->stencil_stride_zero);
553 pandecode_prop("zero7 = 0x%" PRIx32,
554 s->zero7);
555 pandecode_prop("zero8 = 0x%" PRIx32,
556 s->zero8);
557 }
558
559 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
560 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
561 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
562 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
563 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
564 }
565
566 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
567 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
568 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
569 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
570 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
571 }
572
573 if (s->clear_stencil) {
574 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
575 }
576
577 const struct midgard_tiler_descriptor t = s->tiler;
578
579 bool has_hierarchy = !(gpu_id == 0x0720 || gpu_id == 0x0820 || gpu_id == 0x0830);
580 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment, has_hierarchy);
581
582 pandecode_indent--;
583 pandecode_log("};\n");
584
585 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
586 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
587 pandecode_prop("zero5 = 0x%" PRIx32, s->zero5);
588
589 pandecode_log_cont(".zero3 = {");
590
591 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
592 pandecode_log_cont("%X, ", s->zero3[i]);
593
594 pandecode_log_cont("},\n");
595
596 pandecode_log_cont(".zero6 = {");
597
598 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
599 pandecode_log_cont("%X, ", s->zero6[i]);
600
601 pandecode_log_cont("},\n");
602
603 return info;
604 }
605
606 static void
607 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
608 {
609 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
610 const struct mali_shared_memory *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
611
612 pandecode_log("struct mali_shared_memory shared_%"PRIx64"_%d = {\n", gpu_va, job_no);
613 pandecode_indent++;
614 pandecode_shared_memory(s, true);
615 pandecode_indent--;
616 pandecode_log("},\n");
617 }
618
619 /* Extracts the number of components associated with a Mali format */
620
621 static unsigned
622 pandecode_format_component_count(enum mali_format fmt)
623 {
624 /* Mask out the format class */
625 unsigned top = fmt & 0b11100000;
626
627 switch (top) {
628 case MALI_FORMAT_SNORM:
629 case MALI_FORMAT_UINT:
630 case MALI_FORMAT_UNORM:
631 case MALI_FORMAT_SINT:
632 return ((fmt >> 3) & 3) + 1;
633 default:
634 /* TODO: Validate */
635 return 4;
636 }
637 }
638
639 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
640
641 static unsigned
642 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
643 {
644 unsigned mask = 0;
645 assert(MALI_CHANNEL_R == 0);
646
647 for (unsigned c = 0; c < 4; ++c) {
648 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
649
650 if (chan <= MALI_CHANNEL_A)
651 mask |= (1 << chan);
652 }
653
654 return mask;
655 }
656
657 /* Validates that a (format, swizzle) pair is valid, in the sense that the
658 * swizzle doesn't access any components that are undefined in the format.
659 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
660 * omitted */
661
662 static bool
663 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
664 {
665 unsigned nr_comp = pandecode_format_component_count(fmt);
666 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
667 unsigned valid_mask = (1 << nr_comp) - 1;
668 unsigned invalid_mask = ~valid_mask;
669
670 if (access_mask & invalid_mask) {
671 pandecode_msg("XXX: invalid components accessed\n");
672 return false;
673 }
674
675 /* Check for the default non-swizzling swizzle so we can suppress
676 * useless printing for the defaults */
677
678 unsigned default_swizzles[4] = {
679 MALI_CHANNEL_R | (MALI_CHANNEL_0 << 3) | (MALI_CHANNEL_0 << 6) | (MALI_CHANNEL_1 << 9),
680 MALI_CHANNEL_R | (MALI_CHANNEL_G << 3) | (MALI_CHANNEL_0 << 6) | (MALI_CHANNEL_1 << 9),
681 MALI_CHANNEL_R | (MALI_CHANNEL_G << 3) | (MALI_CHANNEL_B << 6) | (MALI_CHANNEL_1 << 9),
682 MALI_CHANNEL_R | (MALI_CHANNEL_G << 3) | (MALI_CHANNEL_B << 6) | (MALI_CHANNEL_A << 9)
683 };
684
685 return (swizzle == default_swizzles[nr_comp - 1]);
686 }
687
688 static void
689 pandecode_swizzle(unsigned swizzle, enum mali_format format)
690 {
691 /* First, do some validation */
692 bool trivial_swizzle = pandecode_validate_format_swizzle(
693 format, swizzle);
694
695 if (trivial_swizzle)
696 return;
697
698 /* Next, print the swizzle */
699 pandecode_log_cont(".");
700
701 static const char components[] = "rgba01";
702
703 for (unsigned c = 0; c < 4; ++c) {
704 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
705
706 if (chan > MALI_CHANNEL_1) {
707 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
708 continue;
709 }
710 pandecode_log_cont("%c", components[chan]);
711 }
712 }
713
714 static void
715 pandecode_rt_format(struct mali_rt_format format)
716 {
717 pandecode_log(".format = {\n");
718 pandecode_indent++;
719
720 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
721 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
722 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
723 pandecode_prop("unk4 = 0x%" PRIx32, format.unk4);
724
725 pandecode_prop("block = %s", mali_block_format_as_str(format.block));
726
727 /* TODO: Map formats so we can check swizzles and print nicely */
728 pandecode_log("swizzle");
729 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
730 pandecode_log_cont(",\n");
731
732 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
733 (format.nr_channels + 1));
734
735 pandecode_log(".flags = ");
736 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
737 pandecode_log_cont(",\n");
738
739 pandecode_prop("msaa = %s", mali_msaa_as_str(format.msaa));
740
741 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
742 * which is a faster hardware-based alternative to the wallpaper method
743 * to preserve framebuffer contents across frames. In practice, MFBD
744 * preload is buggy on Midgard, and so this is a chicken bit. If this
745 * bit isn't set, most likely something broke unrelated to preload */
746
747 if (!format.no_preload) {
748 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
749 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
750 }
751
752 if (format.zero)
753 pandecode_prop("zero = 0x%" PRIx32, format.zero);
754
755 pandecode_indent--;
756 pandecode_log("},\n");
757 }
758
759 static void
760 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct mali_framebuffer *fb)
761 {
762 pandecode_log("struct mali_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
763 pandecode_indent++;
764
765 for (int i = 0; i < (fb->rt_count_1 + 1); i++) {
766 mali_ptr rt_va = gpu_va + i * sizeof(struct mali_render_target);
767 struct pandecode_mapped_memory *mem =
768 pandecode_find_mapped_gpu_mem_containing(rt_va);
769 const struct mali_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
770
771 pandecode_log("{\n");
772 pandecode_indent++;
773
774 pandecode_rt_format(rt->format);
775
776 if (rt->format.block == MALI_BLOCK_FORMAT_AFBC) {
777 pandecode_log(".afbc = {\n");
778 pandecode_indent++;
779
780 char *a = pointer_as_memory_reference(rt->afbc.metadata);
781 pandecode_prop("metadata = %s", a);
782 free(a);
783
784 pandecode_prop("stride = %d", rt->afbc.stride);
785
786 pandecode_log(".flags = ");
787 pandecode_log_decoded_flags(afbc_fmt_flag_info, rt->afbc.flags);
788 pandecode_log_cont(",\n");
789
790 pandecode_indent--;
791 pandecode_log("},\n");
792 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.flags) {
793 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
794 rt->afbc.metadata,
795 rt->afbc.stride,
796 rt->afbc.flags);
797 }
798
799 MEMORY_PROP(rt, framebuffer);
800 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
801
802 if (rt->layer_stride)
803 pandecode_prop("layer_stride = %d", rt->layer_stride);
804
805 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
806 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
807 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
808 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
809 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
810 }
811
812 if (rt->zero1 || rt->zero2) {
813 pandecode_msg("XXX: render target zeros tripped\n");
814 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
815 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
816 }
817
818 pandecode_indent--;
819 pandecode_log("},\n");
820 }
821
822 pandecode_indent--;
823 pandecode_log("};\n");
824 }
825
826 static struct pandecode_fbd
827 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment, bool is_compute, bool is_bifrost)
828 {
829 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
830 const struct mali_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
831
832 struct pandecode_fbd info;
833
834 if (is_bifrost && fb->msaa.sample_locations) {
835 /* The blob stores all possible sample locations in a single buffer
836 * allocated on startup, and just switches the pointer when switching
837 * MSAA state. For now, we just put the data into the cmdstream, but we
838 * should do something like what the blob does with a real driver.
839 *
840 * There seem to be 32 slots for sample locations, followed by another
841 * 16. The second 16 is just the center location followed by 15 zeros
842 * in all the cases I've identified (maybe shader vs. depth/color
843 * samples?).
844 */
845
846 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->msaa.sample_locations);
847
848 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->msaa.sample_locations);
849
850 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
851 pandecode_indent++;
852
853 for (int i = 0; i < 32 + 16; i++) {
854 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
855 }
856
857 pandecode_indent--;
858 pandecode_log("};\n");
859 }
860
861 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
862 pandecode_indent++;
863
864 if (is_bifrost) {
865 pandecode_log(".msaa = {\n");
866 pandecode_indent++;
867
868 if (fb->msaa.sample_locations)
869 pandecode_prop("sample_locations = sample_locations_%d", job_no);
870 else
871 pandecode_msg("XXX: sample_locations missing\n");
872
873 if (fb->msaa.zero1 || fb->msaa.zero2 || fb->msaa.zero4) {
874 pandecode_msg("XXX: multisampling zero tripped\n");
875 pandecode_prop("zero1 = %" PRIx64, fb->msaa.zero1);
876 pandecode_prop("zero2 = %" PRIx64, fb->msaa.zero2);
877 pandecode_prop("zero4 = %" PRIx64, fb->msaa.zero4);
878 }
879
880 pandecode_indent--;
881 pandecode_log("},\n");
882 } else {
883 pandecode_log(".shared_memory = {\n");
884 pandecode_indent++;
885 pandecode_shared_memory(&fb->shared_memory, is_compute);
886 pandecode_indent--;
887 pandecode_log("},\n");
888 }
889
890 info.width = fb->width1 + 1;
891 info.height = fb->height1 + 1;
892 info.rt_count = fb->rt_count_1 + 1;
893
894 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
895 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
896 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
897 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
898
899 pandecode_prop("unk1 = 0x%x", fb->unk1);
900 pandecode_prop("unk2 = 0x%x", fb->unk2);
901 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
902 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
903
904 pandecode_log(".mfbd_flags = ");
905 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
906 pandecode_log_cont(",\n");
907
908 if (fb->clear_stencil)
909 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
910
911 if (fb->clear_depth)
912 pandecode_prop("clear_depth = %f", fb->clear_depth);
913
914 if (!is_compute)
915 if (is_bifrost)
916 pandecode_bifrost_tiler_descriptor(fb);
917 else {
918 const struct midgard_tiler_descriptor t = fb->tiler;
919 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment, true);
920 }
921 else
922 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
923
924 if (fb->zero3 || fb->zero4) {
925 pandecode_msg("XXX: framebuffer zeros tripped\n");
926 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
927 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
928 }
929
930 pandecode_indent--;
931 pandecode_log("};\n");
932
933 gpu_va += sizeof(struct mali_framebuffer);
934
935 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
936
937 if (info.has_extra) {
938 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
939 const struct mali_framebuffer_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
940
941 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
942 pandecode_indent++;
943
944 MEMORY_PROP(fbx, checksum);
945
946 if (fbx->checksum_stride)
947 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
948
949 pandecode_log(".flags_hi = ");
950 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info, fbx->flags_hi);
951 pandecode_log_cont(",\n");
952
953 pandecode_log(".flags_lo = ");
954 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info, fbx->flags_lo);
955 pandecode_log_cont(",\n");
956
957 pandecode_prop("zs_block = %s", mali_block_format_as_str(fbx->zs_block));
958 pandecode_prop("zs_samples = MALI_POSITIVE(%u)", fbx->zs_samples + 1);
959
960 if (fbx->zs_block == MALI_BLOCK_FORMAT_AFBC) {
961 pandecode_log(".ds_afbc = {\n");
962 pandecode_indent++;
963
964 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
965 pandecode_prop("depth_stencil_afbc_stride = %d",
966 fbx->ds_afbc.depth_stencil_afbc_stride);
967 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
968
969 pandecode_log(".flags = ");
970 pandecode_log_decoded_flags(afbc_fmt_flag_info, fbx->ds_afbc.flags);
971 pandecode_log_cont(",\n");
972
973 if (fbx->ds_afbc.padding) {
974 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
975 pandecode_prop("padding = 0x%" PRIx64, fbx->ds_afbc.padding);
976 }
977
978 pandecode_indent--;
979 pandecode_log("},\n");
980 } else {
981 pandecode_log(".ds_linear = {\n");
982 pandecode_indent++;
983
984 if (fbx->ds_linear.depth) {
985 MEMORY_PROP_DIR(fbx->ds_linear, depth);
986 pandecode_prop("depth_stride = %d",
987 fbx->ds_linear.depth_stride);
988 pandecode_prop("depth_layer_stride = %d",
989 fbx->ds_linear.depth_layer_stride);
990 } else if (fbx->ds_linear.depth_stride || fbx->ds_linear.depth_layer_stride) {
991 pandecode_msg("XXX: depth stride zero tripped %d %d\n", fbx->ds_linear.depth_stride, fbx->ds_linear.depth_layer_stride);
992 }
993
994 if (fbx->ds_linear.stencil) {
995 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
996 pandecode_prop("stencil_stride = %d",
997 fbx->ds_linear.stencil_stride);
998 pandecode_prop("stencil_layer_stride = %d",
999 fbx->ds_linear.stencil_layer_stride);
1000 } else if (fbx->ds_linear.stencil_stride || fbx->ds_linear.stencil_layer_stride) {
1001 pandecode_msg("XXX: stencil stride zero tripped %d %d\n", fbx->ds_linear.stencil_stride, fbx->ds_linear.stencil_layer_stride);
1002 }
1003
1004 if (fbx->ds_linear.depth_stride_zero ||
1005 fbx->ds_linear.stencil_stride_zero) {
1006 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1007 pandecode_prop("depth_stride_zero = 0x%x",
1008 fbx->ds_linear.depth_stride_zero);
1009 pandecode_prop("stencil_stride_zero = 0x%x",
1010 fbx->ds_linear.stencil_stride_zero);
1011 }
1012
1013 pandecode_indent--;
1014 pandecode_log("},\n");
1015 }
1016
1017 if (fbx->clear_color_1 | fbx->clear_color_2) {
1018 pandecode_prop("clear_color_1 = 0x%" PRIx32, fbx->clear_color_1);
1019 pandecode_prop("clear_color_2 = 0x%" PRIx32, fbx->clear_color_2);
1020 }
1021
1022 if (fbx->zero3) {
1023 pandecode_msg("XXX: fb_extra zeros tripped\n");
1024 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1025 }
1026
1027 pandecode_indent--;
1028 pandecode_log("};\n");
1029
1030 gpu_va += sizeof(struct mali_framebuffer_extra);
1031 }
1032
1033 if (is_fragment)
1034 pandecode_render_target(gpu_va, job_no, fb);
1035
1036 return info;
1037 }
1038
1039 static void
1040 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1041 mali_ptr addr, int job_no, char *suffix,
1042 int count, bool varying, enum mali_job_type job_type)
1043 {
1044 char *prefix = varying ? "Varying" : "Attribute";
1045 assert(addr);
1046
1047 if (!count) {
1048 pandecode_msg("warn: No %s records\n", prefix);
1049 return;
1050 }
1051
1052 MAP_ADDR(ATTRIBUTE_BUFFER, addr, cl);
1053
1054 for (int i = 0; i < count; ++i) {
1055 fprintf(pandecode_dump_stream, "%s\n", prefix);
1056
1057 struct MALI_ATTRIBUTE_BUFFER temp;
1058 MALI_ATTRIBUTE_BUFFER_unpack(cl + i * MALI_ATTRIBUTE_BUFFER_LENGTH, &temp);
1059 MALI_ATTRIBUTE_BUFFER_print(pandecode_dump_stream, &temp, 2);
1060
1061 if (temp.type == MALI_ATTRIBUTE_TYPE_1D_NPOT_DIVISOR) {
1062 struct MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT temp2;
1063 MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT_unpack(cl + (i + 1) * MALI_ATTRIBUTE_BUFFER_LENGTH, &temp2);
1064 MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT_print(pandecode_dump_stream, &temp2, 2);
1065 }
1066 }
1067 }
1068
1069 static mali_ptr
1070 pandecode_shader_address(const char *name, mali_ptr ptr)
1071 {
1072 /* TODO: Decode flags */
1073 mali_ptr shader_ptr = ptr & ~15;
1074
1075 char *a = pointer_as_memory_reference(shader_ptr);
1076 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1077 free(a);
1078
1079 return shader_ptr;
1080 }
1081
1082 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1083
1084 static unsigned
1085 decode_bifrost_constant(u16 constant)
1086 {
1087 float lo = (float) (constant & 0xFF);
1088 float hi = (float) (constant >> 8);
1089
1090 return (hi / 255.0) + (lo / 65535.0);
1091 }
1092
1093 static mali_ptr
1094 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1095 {
1096 struct bifrost_blend_rt *b =
1097 ((struct bifrost_blend_rt *) descs) + rt_no;
1098
1099 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1100 pandecode_indent++;
1101
1102 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1103 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1104 b->constant, decode_bifrost_constant(b->constant));
1105
1106 /* TODO figure out blend shader enable bit */
1107 DUMP_CL("Equation", BLEND_EQUATION, &b->equation, 2);
1108
1109 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1110 pandecode_prop("index = 0x%" PRIx16, b->index);
1111
1112 pandecode_log(".format = %s", mali_format_as_str(b->format));
1113 pandecode_swizzle(b->swizzle, b->format);
1114 pandecode_log_cont(",\n");
1115
1116 pandecode_prop("swizzle = 0x%" PRIx32, b->swizzle);
1117 pandecode_prop("format = 0x%" PRIx32, b->format);
1118
1119 if (b->zero1) {
1120 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1121 pandecode_prop("zero1 = 0x%" PRIx32, b->zero1);
1122 }
1123
1124 pandecode_log(".shader_type = ");
1125 switch(b->shader_type) {
1126 case BIFROST_BLEND_F16:
1127 pandecode_log_cont("BIFROST_BLEND_F16");
1128 break;
1129 case BIFROST_BLEND_F32:
1130 pandecode_log_cont("BIFROST_BLEND_F32");
1131 break;
1132 case BIFROST_BLEND_I32:
1133 pandecode_log_cont("BIFROST_BLEND_I32");
1134 break;
1135 case BIFROST_BLEND_U32:
1136 pandecode_log_cont("BIFROST_BLEND_U32");
1137 break;
1138 case BIFROST_BLEND_I16:
1139 pandecode_log_cont("BIFROST_BLEND_I16");
1140 break;
1141 case BIFROST_BLEND_U16:
1142 pandecode_log_cont("BIFROST_BLEND_U16");
1143 break;
1144 }
1145 pandecode_log_cont(",\n");
1146
1147 if (b->zero2) {
1148 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1149 pandecode_prop("zero2 = 0x%" PRIx32, b->zero2);
1150 }
1151
1152 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1153
1154 pandecode_indent--;
1155 pandecode_log("},\n");
1156
1157 return 0;
1158 }
1159
1160 static mali_ptr
1161 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1162 {
1163 /* constant/equation is in a union */
1164 if (!blend->shader)
1165 return 0;
1166
1167 pandecode_log(".blend = {\n");
1168 pandecode_indent++;
1169
1170 if (is_shader) {
1171 pandecode_shader_address("shader", blend->shader);
1172 } else {
1173 DUMP_CL("Equation", BLEND_EQUATION, &blend->equation, 2);
1174 pandecode_prop("constant = %f", blend->constant);
1175 }
1176
1177 pandecode_indent--;
1178 pandecode_log("},\n");
1179
1180 /* Return blend shader to disassemble if present */
1181 return is_shader ? (blend->shader & ~0xF) : 0;
1182 }
1183
1184 static mali_ptr
1185 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1186 {
1187 struct midgard_blend_rt *b =
1188 ((struct midgard_blend_rt *) descs) + rt_no;
1189
1190 /* Flags determine presence of blend shader */
1191 bool is_shader = b->flags.opaque[0] & 0x2;
1192
1193 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1194 pandecode_indent++;
1195
1196 DUMP_CL("Flags", BLEND_FLAGS, &b->flags, 2);
1197
1198 union midgard_blend blend = b->blend;
1199 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1200
1201 pandecode_indent--;
1202 pandecode_log("};\n");
1203
1204 return shader;
1205 }
1206
1207 /* Attributes and varyings have descriptor records, which contain information
1208 * about their format and ordering with the attribute/varying buffers. We'll
1209 * want to validate that the combinations specified are self-consistent.
1210 */
1211
1212 static int
1213 pandecode_attribute_meta(int count, mali_ptr attribute, bool varying, char *suffix)
1214 {
1215 const char *prefix = varying ? "Varying" : "Attribute";
1216
1217 for (int i = 0; i < count; ++i, attribute += MALI_ATTRIBUTE_LENGTH)
1218 DUMP_ADDR(prefix, ATTRIBUTE, attribute, 1);
1219
1220 return count;
1221 }
1222
1223 /* return bits [lo, hi) of word */
1224 static u32
1225 bits(u32 word, u32 lo, u32 hi)
1226 {
1227 if (hi - lo >= 32)
1228 return word; // avoid undefined behavior with the shift
1229
1230 return (word >> lo) & ((1 << (hi - lo)) - 1);
1231 }
1232
1233 static void
1234 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool graphics)
1235 {
1236 pandecode_log(".prefix = {\n");
1237 pandecode_indent++;
1238
1239 /* Decode invocation_count. See the comment before the definition of
1240 * invocation_count for an explanation.
1241 */
1242 struct MALI_INVOCATION invocation;
1243 struct mali_invocation_packed invocation_packed = p->invocation;
1244 MALI_INVOCATION_unpack((const uint8_t *) &invocation_packed, &invocation);
1245
1246 unsigned size_x = bits(invocation.invocations, 0, invocation.size_y_shift) + 1;
1247 unsigned size_y = bits(invocation.invocations, invocation.size_y_shift, invocation.size_z_shift) + 1;
1248 unsigned size_z = bits(invocation.invocations, invocation.size_z_shift, invocation.workgroups_x_shift) + 1;
1249
1250 unsigned groups_x = bits(invocation.invocations, invocation.workgroups_x_shift, invocation.workgroups_y_shift) + 1;
1251 unsigned groups_y = bits(invocation.invocations, invocation.workgroups_y_shift, invocation.workgroups_z_shift) + 1;
1252 unsigned groups_z = bits(invocation.invocations, invocation.workgroups_z_shift, 32) + 1;
1253
1254 /* Even though we have this decoded, we want to ensure that the
1255 * representation is "unique" so we don't lose anything by printing only
1256 * the final result. More specifically, we need to check that we were
1257 * passed something in canonical form, since the definition per the
1258 * hardware is inherently not unique. How? Well, take the resulting
1259 * decode and pack it ourselves! If it is bit exact with what we
1260 * decoded, we're good to go. */
1261
1262 struct mali_invocation_packed ref;
1263 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, graphics);
1264
1265 if (memcmp(&ref, &invocation_packed, sizeof(ref))) {
1266 pandecode_msg("XXX: non-canonical workgroups packing\n");
1267 MALI_INVOCATION_print(pandecode_dump_stream, &invocation, 1 * 2);
1268 }
1269
1270 /* Regardless, print the decode */
1271 fprintf(pandecode_dump_stream,
1272 "Invocation (%d, %d, %d) x (%d, %d, %d)\n",
1273 size_x, size_y, size_z,
1274 groups_x, groups_y, groups_z);
1275
1276 fprintf(pandecode_dump_stream, "Primitive\n");
1277 struct MALI_PRIMITIVE primitive;
1278 struct mali_primitive_packed prim_packed = p->primitive;
1279 MALI_PRIMITIVE_unpack((const uint8_t *) &prim_packed, &primitive);
1280 MALI_PRIMITIVE_print(pandecode_dump_stream, &primitive, 1 * 2);
1281
1282 /* Validate an index buffer is present if we need one. TODO: verify
1283 * relationship between invocation_count and index_count */
1284
1285 if (primitive.indices) {
1286 /* Grab the size */
1287 unsigned size = (primitive.index_type == MALI_INDEX_TYPE_UINT32) ?
1288 sizeof(uint32_t) : primitive.index_type;
1289
1290 /* Ensure we got a size, and if so, validate the index buffer
1291 * is large enough to hold a full set of indices of the given
1292 * size */
1293
1294 if (!size)
1295 pandecode_msg("XXX: index size missing\n");
1296 else
1297 pandecode_validate_buffer(primitive.indices, primitive.index_count * size);
1298 } else if (primitive.index_type)
1299 pandecode_msg("XXX: unexpected index size\n");
1300
1301 pandecode_indent--;
1302 pandecode_log("},\n");
1303 }
1304
1305 static void
1306 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1307 {
1308 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1309 uint64_t *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1310
1311 for (int i = 0; i < ubufs_count; i++) {
1312 unsigned size = (ubufs[i] & ((1 << 10) - 1)) * 16;
1313 mali_ptr addr = (ubufs[i] >> 10) << 2;
1314
1315 pandecode_validate_buffer(addr, size);
1316
1317 char *ptr = pointer_as_memory_reference(addr);
1318 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
1319 free(ptr);
1320 }
1321
1322 pandecode_log("\n");
1323 }
1324
1325 static void
1326 pandecode_uniforms(mali_ptr uniforms, unsigned uniform_count)
1327 {
1328 pandecode_validate_buffer(uniforms, uniform_count * 16);
1329
1330 char *ptr = pointer_as_memory_reference(uniforms);
1331 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count, ptr);
1332 free(ptr);
1333 }
1334
1335 static const char *
1336 shader_type_for_job(unsigned type)
1337 {
1338 switch (type) {
1339 case MALI_JOB_TYPE_VERTEX: return "VERTEX";
1340 case MALI_JOB_TYPE_TILER: return "FRAGMENT";
1341 case MALI_JOB_TYPE_COMPUTE: return "COMPUTE";
1342 default:
1343 return "UNKNOWN";
1344 }
1345 }
1346
1347 static unsigned shader_id = 0;
1348
1349 static struct midgard_disasm_stats
1350 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1351 bool is_bifrost, unsigned gpu_id)
1352 {
1353 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1354 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1355
1356 /* Compute maximum possible size */
1357 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1358
1359 /* Print some boilerplate to clearly denote the assembly (which doesn't
1360 * obey indentation rules), and actually do the disassembly! */
1361
1362 pandecode_log_cont("\n\n");
1363
1364 struct midgard_disasm_stats stats;
1365
1366 if (is_bifrost) {
1367 disassemble_bifrost(pandecode_dump_stream, code, sz, true);
1368
1369 /* TODO: Extend stats to Bifrost */
1370 stats.texture_count = -128;
1371 stats.sampler_count = -128;
1372 stats.attribute_count = -128;
1373 stats.varying_count = -128;
1374 stats.uniform_count = -128;
1375 stats.uniform_buffer_count = -128;
1376 stats.work_count = -128;
1377
1378 stats.instruction_count = 0;
1379 stats.bundle_count = 0;
1380 stats.quadword_count = 0;
1381 stats.helper_invocations = false;
1382 } else {
1383 stats = disassemble_midgard(pandecode_dump_stream,
1384 code, sz, gpu_id,
1385 type == MALI_JOB_TYPE_TILER ?
1386 MESA_SHADER_FRAGMENT : MESA_SHADER_VERTEX);
1387 }
1388
1389 unsigned nr_threads =
1390 (stats.work_count <= 4) ? 4 :
1391 (stats.work_count <= 8) ? 2 :
1392 1;
1393
1394 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
1395 "%u inst, %u bundles, %u quadwords, "
1396 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
1397 shader_id++,
1398 shader_type_for_job(type),
1399 stats.instruction_count, stats.bundle_count, stats.quadword_count,
1400 stats.work_count, nr_threads);
1401
1402 return stats;
1403 }
1404
1405 static void
1406 pandecode_texture_payload(mali_ptr payload,
1407 enum mali_texture_dimension dim,
1408 enum mali_texture_layout layout,
1409 bool manual_stride,
1410 uint8_t levels,
1411 uint16_t depth,
1412 uint16_t array_size,
1413 struct pandecode_mapped_memory *tmem)
1414 {
1415 pandecode_log(".payload = {\n");
1416 pandecode_indent++;
1417
1418 /* A bunch of bitmap pointers follow.
1419 * We work out the correct number,
1420 * based on the mipmap/cubemap
1421 * properties, but dump extra
1422 * possibilities to futureproof */
1423
1424 int bitmap_count = levels + 1;
1425
1426 /* Miptree for each face */
1427 if (dim == MALI_TEXTURE_DIMENSION_CUBE)
1428 bitmap_count *= 6;
1429
1430 /* Array of layers */
1431 bitmap_count *= depth;
1432
1433 /* Array of textures */
1434 bitmap_count *= array_size;
1435
1436 /* Stride for each element */
1437 if (manual_stride)
1438 bitmap_count *= 2;
1439
1440 mali_ptr *pointers_and_strides = pandecode_fetch_gpu_mem(tmem,
1441 payload, sizeof(mali_ptr) * bitmap_count);
1442 for (int i = 0; i < bitmap_count; ++i) {
1443 /* How we dump depends if this is a stride or a pointer */
1444
1445 if (manual_stride && (i & 1)) {
1446 /* signed 32-bit snuck in as a 64-bit pointer */
1447 uint64_t stride_set = pointers_and_strides[i];
1448 uint32_t clamped_stride = stride_set;
1449 int32_t stride = clamped_stride;
1450 assert(stride_set == clamped_stride);
1451 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
1452 } else {
1453 char *a = pointer_as_memory_reference(pointers_and_strides[i]);
1454 pandecode_log("%s, \n", a);
1455 free(a);
1456 }
1457 }
1458
1459 pandecode_indent--;
1460 pandecode_log("},\n");
1461 }
1462
1463 static void
1464 pandecode_texture(mali_ptr u,
1465 struct pandecode_mapped_memory *tmem,
1466 unsigned job_no, unsigned tex)
1467 {
1468 struct pandecode_mapped_memory *mapped_mem = pandecode_find_mapped_gpu_mem_containing(u);
1469 const uint8_t *cl = pandecode_fetch_gpu_mem(mapped_mem, u, MALI_MIDGARD_TEXTURE_LENGTH);
1470
1471 struct MALI_MIDGARD_TEXTURE temp;
1472 MALI_MIDGARD_TEXTURE_unpack(cl, &temp);
1473 MALI_MIDGARD_TEXTURE_print(pandecode_dump_stream, &temp, 2);
1474
1475 pandecode_texture_payload(u + MALI_MIDGARD_TEXTURE_LENGTH,
1476 temp.dimension, temp.texel_ordering, temp.manual_stride,
1477 temp.levels, temp.depth, temp.array_size, mapped_mem);
1478 }
1479
1480 static void
1481 pandecode_bifrost_texture(
1482 const void *cl,
1483 unsigned job_no,
1484 unsigned tex)
1485 {
1486 struct MALI_BIFROST_TEXTURE temp;
1487 MALI_BIFROST_TEXTURE_unpack(cl, &temp);
1488 MALI_BIFROST_TEXTURE_print(pandecode_dump_stream, &temp, 2);
1489
1490 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(temp.surfaces);
1491 pandecode_texture_payload(temp.surfaces, temp.dimension, temp.texel_ordering,
1492 true, temp.levels, 1, 1, tmem);
1493 }
1494
1495 /* 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 */
1496
1497 static void
1498 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
1499 {
1500 /* Nothing to do */
1501 if (claim == truth)
1502 return;
1503
1504 if (fuzzy && (truth < 0))
1505 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name, claim, truth);
1506
1507 if ((truth >= 0) && !fuzzy) {
1508 pandecode_msg("%s: expected %s = %d, claimed %u\n",
1509 (truth < claim) ? "warn" : "XXX",
1510 name, truth, claim);
1511 } else if ((claim > -truth) && !fuzzy) {
1512 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
1513 name, -truth, claim);
1514 } else if (fuzzy && (claim < truth))
1515 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
1516 name, truth, claim);
1517
1518 pandecode_log(".%s = %" PRId16, name, claim);
1519
1520 if (fuzzy)
1521 pandecode_log_cont(" /* %u used */", truth);
1522
1523 pandecode_log_cont(",\n");
1524 }
1525
1526 static void
1527 pandecode_blend_shader_disassemble(mali_ptr shader, int job_no, int job_type,
1528 bool is_bifrost, unsigned gpu_id)
1529 {
1530 struct midgard_disasm_stats stats =
1531 pandecode_shader_disassemble(shader, job_no, job_type, is_bifrost, gpu_id);
1532
1533 bool has_texture = (stats.texture_count > 0);
1534 bool has_sampler = (stats.sampler_count > 0);
1535 bool has_attribute = (stats.attribute_count > 0);
1536 bool has_varying = (stats.varying_count > 0);
1537 bool has_uniform = (stats.uniform_count > 0);
1538 bool has_ubo = (stats.uniform_buffer_count > 0);
1539
1540 if (has_texture || has_sampler)
1541 pandecode_msg("XXX: blend shader accessing textures\n");
1542
1543 if (has_attribute || has_varying)
1544 pandecode_msg("XXX: blend shader accessing interstage\n");
1545
1546 if (has_uniform || has_ubo)
1547 pandecode_msg("XXX: blend shader accessing uniforms\n");
1548 }
1549
1550 static void
1551 pandecode_textures(mali_ptr textures, unsigned texture_count, int job_no, bool is_bifrost)
1552 {
1553 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(textures);
1554
1555 if (!mmem)
1556 return;
1557
1558 pandecode_log("Textures (%"PRIx64"):\n", textures);
1559
1560 if (is_bifrost) {
1561 const void *cl = pandecode_fetch_gpu_mem(mmem,
1562 textures, MALI_BIFROST_TEXTURE_LENGTH *
1563 texture_count);
1564
1565 for (unsigned tex = 0; tex < texture_count; ++tex) {
1566 pandecode_bifrost_texture(cl +
1567 MALI_BIFROST_TEXTURE_LENGTH * tex,
1568 job_no, tex);
1569 }
1570 } else {
1571 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures);
1572
1573 for (int tex = 0; tex < texture_count; ++tex) {
1574 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
1575 char *a = pointer_as_memory_reference(*u);
1576 pandecode_log("%s,\n", a);
1577 free(a);
1578 }
1579
1580 /* Now, finally, descend down into the texture descriptor */
1581 for (unsigned tex = 0; tex < texture_count; ++tex) {
1582 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
1583 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
1584 if (tmem)
1585 pandecode_texture(*u, tmem, job_no, tex);
1586 }
1587 }
1588 }
1589
1590 static void
1591 pandecode_samplers(mali_ptr samplers, unsigned sampler_count, int job_no, bool is_bifrost)
1592 {
1593 for (int i = 0; i < sampler_count; ++i) {
1594 if (is_bifrost) {
1595 DUMP_ADDR("Sampler", BIFROST_SAMPLER, samplers + (MALI_BIFROST_SAMPLER_LENGTH * i), 1);
1596 } else {
1597 DUMP_ADDR("Sampler", MIDGARD_SAMPLER, samplers + (MALI_MIDGARD_SAMPLER_LENGTH * i), 1);
1598 }
1599 }
1600 }
1601
1602 static void
1603 pandecode_vertex_tiler_postfix_pre(
1604 const struct mali_vertex_tiler_postfix *p,
1605 int job_no, enum mali_job_type job_type,
1606 char *suffix, bool is_bifrost, unsigned gpu_id)
1607 {
1608 struct pandecode_mapped_memory *attr_mem;
1609
1610 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1611 * are the only things actually needed from the FBD, vertex/tiler jobs
1612 * no longer reference the FBD -- instead, this field points to some
1613 * info about the scratchpad.
1614 */
1615
1616 struct pandecode_fbd fbd_info = {
1617 /* Default for Bifrost */
1618 .rt_count = 1
1619 };
1620
1621 if (is_bifrost) {
1622 pandecode_log_cont("\t/* %X %/\n", p->shared_memory & 1);
1623 pandecode_compute_fbd(p->shared_memory & ~1, job_no);
1624 } else if (p->shared_memory & MALI_MFBD)
1625 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->shared_memory) & FBD_MASK, job_no, false, job_type == MALI_JOB_TYPE_COMPUTE, false);
1626 else if (job_type == MALI_JOB_TYPE_COMPUTE)
1627 pandecode_compute_fbd((u64) (uintptr_t) p->shared_memory, job_no);
1628 else
1629 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->shared_memory, job_no, false, gpu_id);
1630
1631 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
1632 int texture_count = 0, sampler_count = 0;
1633
1634 if (p->shader) {
1635 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->shader);
1636 uint32_t *cl = pandecode_fetch_gpu_mem(smem, p->shader, MALI_STATE_LENGTH);
1637
1638 /* Disassemble ahead-of-time to get stats. Initialize with
1639 * stats for the missing-shader case so we get validation
1640 * there, too */
1641
1642 struct midgard_disasm_stats info = {
1643 .texture_count = 0,
1644 .sampler_count = 0,
1645 .attribute_count = 0,
1646 .varying_count = 0,
1647 .work_count = 1,
1648
1649 .uniform_count = -128,
1650 .uniform_buffer_count = 0
1651 };
1652
1653 struct MALI_STATE state;
1654 struct MALI_MIDGARD_PROPERTIES midg_props;
1655 struct MALI_BIFROST_PROPERTIES bi_props;
1656
1657 MALI_STATE_unpack((const uint8_t *) cl, &state);
1658
1659 if (state.shader.shader & ~0xF)
1660 info = pandecode_shader_disassemble(state.shader.shader & ~0xF, job_no, job_type, is_bifrost, gpu_id);
1661
1662 fprintf(pandecode_dump_stream, "State %"PRIx64"\n", p->shader);
1663 MALI_STATE_print(pandecode_dump_stream, &state, 1 * 2);
1664
1665 /* Save for dumps */
1666 attribute_count = state.shader.attribute_count;
1667 varying_count = state.shader.varying_count;
1668 texture_count = state.shader.texture_count;
1669 sampler_count = state.shader.sampler_count;
1670
1671 fprintf(pandecode_dump_stream, " Properties\n");
1672 if (is_bifrost) {
1673 MALI_BIFROST_PROPERTIES_unpack((const uint8_t *) &state.properties, &bi_props);
1674 MALI_BIFROST_PROPERTIES_print(pandecode_dump_stream, &bi_props, 2 * 2);
1675
1676 uniform_count = state.preload.uniform_count;
1677 uniform_buffer_count = bi_props.uniform_buffer_count;
1678 } else {
1679 MALI_MIDGARD_PROPERTIES_unpack((const uint8_t *) &state.properties, &midg_props);
1680 MALI_MIDGARD_PROPERTIES_print(pandecode_dump_stream, &midg_props, 2 * 2);
1681
1682 uniform_count = midg_props.uniform_count;
1683 uniform_buffer_count = midg_props.uniform_buffer_count;
1684 }
1685
1686 pandecode_shader_prop("texture_count", texture_count, info.texture_count, false);
1687 pandecode_shader_prop("sampler_count", sampler_count, info.sampler_count, false);
1688 pandecode_shader_prop("attribute_count", attribute_count, info.attribute_count, false);
1689 pandecode_shader_prop("varying_count", varying_count, info.varying_count, false);
1690
1691 if (is_bifrost) {
1692 uint32_t opaque = state.preload.uniform_count << 15
1693 | state.preload.untyped;
1694
1695 switch (job_type) {
1696 case MALI_JOB_TYPE_VERTEX:
1697 DUMP_CL("Preload", PRELOAD_VERTEX, &opaque, 2);
1698 break;
1699 case MALI_JOB_TYPE_TILER:
1700 DUMP_CL("Preload", PRELOAD_FRAGMENT, &opaque, 2);
1701 break;
1702 case MALI_JOB_TYPE_COMPUTE:
1703 DUMP_CL("Preload", PRELOAD_COMPUTE, &opaque, 2);
1704 break;
1705 default:
1706 DUMP_CL("Preload", PRELOAD, &opaque, 2);
1707 break;
1708 }
1709 }
1710
1711 if (!is_bifrost) {
1712 /* TODO: Blend shaders routing/disasm */
1713 union midgard_blend blend;
1714 memcpy(&blend, &state.sfbd_blend, sizeof(blend));
1715 mali_ptr shader = pandecode_midgard_blend(&blend, state.multisample_misc.sfbd_blend_shader);
1716 if (shader & ~0xF)
1717 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
1718 }
1719
1720 /* MRT blend fields are used whenever MFBD is used, with
1721 * per-RT descriptors */
1722
1723 if (job_type == MALI_JOB_TYPE_TILER && (is_bifrost || p->shared_memory & MALI_MFBD)) {
1724 void* blend_base = ((void *) cl) + MALI_STATE_LENGTH;
1725
1726 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
1727 mali_ptr shader = 0;
1728
1729 if (is_bifrost)
1730 shader = pandecode_bifrost_blend(blend_base, job_no, i);
1731 else
1732 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
1733
1734 if (shader & ~0xF)
1735 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
1736
1737 }
1738 }
1739 } else
1740 pandecode_msg("XXX: missing shader descriptor\n");
1741
1742 if (p->viewport)
1743 DUMP_ADDR("Viewport", VIEWPORT, p->viewport, 1);
1744
1745 unsigned max_attr_index = 0;
1746
1747 if (p->attribute_meta)
1748 max_attr_index = pandecode_attribute_meta(attribute_count, p->attribute_meta, false, suffix);
1749
1750 if (p->attributes) {
1751 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
1752 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false, job_type);
1753 }
1754
1755 /* Varyings are encoded like attributes but not actually sent; we just
1756 * pass a zero buffer with the right stride/size set, (or whatever)
1757 * since the GPU will write to it itself */
1758
1759 if (p->varying_meta) {
1760 varying_count = pandecode_attribute_meta(varying_count, p->varying_meta, true, suffix);
1761 }
1762
1763 if (p->varyings) {
1764 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
1765
1766 /* Number of descriptors depends on whether there are
1767 * non-internal varyings */
1768
1769 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true, job_type);
1770 }
1771
1772 if (p->uniform_buffers) {
1773 if (uniform_buffer_count)
1774 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
1775 else
1776 pandecode_msg("warn: UBOs specified but not referenced\n");
1777 } else if (uniform_buffer_count)
1778 pandecode_msg("XXX: UBOs referenced but not specified\n");
1779
1780 /* We don't want to actually dump uniforms, but we do need to validate
1781 * that the counts we were given are sane */
1782
1783 if (p->uniforms) {
1784 if (uniform_count)
1785 pandecode_uniforms(p->uniforms, uniform_count);
1786 else
1787 pandecode_msg("warn: Uniforms specified but not referenced\n");
1788 } else if (uniform_count)
1789 pandecode_msg("XXX: Uniforms referenced but not specified\n");
1790
1791 if (p->textures)
1792 pandecode_textures(p->textures, texture_count, job_no, is_bifrost);
1793
1794 if (p->sampler_descriptor)
1795 pandecode_samplers(p->sampler_descriptor, sampler_count, job_no, is_bifrost);
1796 }
1797
1798 static void
1799 pandecode_gl_enables(uint32_t gl_enables, int job_type)
1800 {
1801 pandecode_log(".gl_enables = ");
1802
1803 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
1804
1805 pandecode_log_cont(",\n");
1806 }
1807
1808 static void
1809 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
1810 {
1811 if (p->shader & 0xF)
1812 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p->shader & 0xF));
1813
1814 pandecode_log(".postfix = {\n");
1815 pandecode_indent++;
1816
1817 pandecode_gl_enables(p->gl_enables, MALI_JOB_TYPE_TILER);
1818 pandecode_prop("instance_shift = 0x%x", p->instance_shift);
1819 pandecode_prop("instance_odd = 0x%x", p->instance_odd);
1820
1821 if (p->zero4) {
1822 pandecode_msg("XXX: vertex only zero tripped");
1823 pandecode_prop("zero4 = 0x%" PRIx32, p->zero4);
1824 }
1825
1826 pandecode_prop("offset_start = 0x%x", p->offset_start);
1827
1828 if (p->zero5) {
1829 pandecode_msg("XXX: vertex only zero tripped");
1830 pandecode_prop("zero5 = 0x%" PRIx64, p->zero5);
1831 }
1832
1833 MEMORY_PROP(p, position_varying);
1834 MEMORY_PROP(p, occlusion_counter);
1835
1836 pandecode_indent--;
1837 pandecode_log("},\n");
1838 }
1839
1840 static void
1841 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
1842 {
1843 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1844 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
1845
1846 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
1847 pandecode_indent++;
1848
1849 if (h->zero) {
1850 pandecode_msg("XXX: tiler heap zero tripped\n");
1851 pandecode_prop("zero = 0x%x", h->zero);
1852 }
1853
1854 pandecode_prop("heap_size = 0x%x", h->heap_size);
1855 MEMORY_PROP(h, tiler_heap_start);
1856 MEMORY_PROP(h, tiler_heap_free);
1857
1858 /* this might point to the beginning of another buffer, when it's
1859 * really the end of the tiler heap buffer, so we have to be careful
1860 * here. but for zero length, we need the same pointer.
1861 */
1862
1863 if (h->tiler_heap_end == h->tiler_heap_start) {
1864 MEMORY_PROP(h, tiler_heap_start);
1865 } else {
1866 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
1867 pandecode_prop("tiler_heap_end = %s + 1", a);
1868 free(a);
1869 }
1870
1871 for (int i = 0; i < 10; i++) {
1872 if (h->zeros[i] != 0) {
1873 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
1874 i, h->zeros[i]);
1875 }
1876 }
1877
1878 if (h->unk1 != 0x1) {
1879 pandecode_msg("XXX: tiler heap unk1 tripped\n");
1880 pandecode_prop("unk1 = 0x%x", h->unk1);
1881 }
1882
1883 if (h->unk7e007e != 0x7e007e) {
1884 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
1885 pandecode_prop("unk7e007e = 0x%x", h->unk7e007e);
1886 }
1887
1888 pandecode_indent--;
1889 pandecode_log("};\n");
1890 }
1891
1892 static void
1893 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
1894 {
1895 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1896 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
1897
1898 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
1899
1900 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
1901 pandecode_indent++;
1902
1903 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32, t->tiler_heap_next_start);
1904 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32, t->used_hierarchy_mask);
1905
1906 if (t->hierarchy_mask != 0xa &&
1907 t->hierarchy_mask != 0x14 &&
1908 t->hierarchy_mask != 0x28 &&
1909 t->hierarchy_mask != 0x50 &&
1910 t->hierarchy_mask != 0xa0)
1911 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
1912
1913 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
1914
1915 pandecode_prop("flags = 0x%" PRIx16, t->flags);
1916
1917 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
1918 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
1919
1920 if (t->zero0) {
1921 pandecode_msg("XXX: tiler meta zero tripped\n");
1922 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
1923 }
1924
1925 for (int i = 0; i < 12; i++) {
1926 if (t->zeros[i] != 0) {
1927 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
1928 i, t->zeros[i]);
1929 }
1930 }
1931
1932 pandecode_indent--;
1933 pandecode_log("};\n");
1934 }
1935
1936 static void
1937 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
1938 {
1939 if (u.pointer == 0x0)
1940 return;
1941
1942 pandecode_log(".primitive_size = {\n");
1943 pandecode_indent++;
1944
1945 if (constant) {
1946 pandecode_prop("constant = %f", u.constant);
1947 } else {
1948 MEMORY_PROP((&u), pointer);
1949 }
1950
1951 pandecode_indent--;
1952 pandecode_log("},\n");
1953 }
1954
1955 static int
1956 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
1957 const struct pandecode_mapped_memory *mem,
1958 mali_ptr payload, int job_no, unsigned gpu_id)
1959 {
1960 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
1961
1962 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true, gpu_id);
1963
1964 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64"_%d = {\n", payload, job_no);
1965 pandecode_indent++;
1966
1967 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
1968 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
1969
1970 pandecode_indent--;
1971 pandecode_log("};\n");
1972
1973 return sizeof(*v);
1974 }
1975
1976 static int
1977 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
1978 const struct pandecode_mapped_memory *mem,
1979 mali_ptr payload, int job_no, unsigned gpu_id)
1980 {
1981 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
1982
1983 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true, gpu_id);
1984 pandecode_tiler_meta(t->tiler_meta, job_no);
1985
1986 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64"_%d = {\n", payload, job_no);
1987 pandecode_indent++;
1988
1989 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
1990
1991 /* TODO: gl_PointSize on Bifrost */
1992 pandecode_primitive_size(t->primitive_size, true);
1993
1994 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
1995 || t->zero6) {
1996 pandecode_msg("XXX: tiler only zero tripped\n");
1997 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
1998 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
1999 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2000 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2001 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2002 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2003 }
2004
2005 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2006
2007 pandecode_indent--;
2008 pandecode_log("};\n");
2009
2010 return sizeof(*t);
2011 }
2012
2013 static int
2014 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2015 const struct pandecode_mapped_memory *mem,
2016 mali_ptr payload, int job_no, unsigned gpu_id)
2017 {
2018 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2019 bool is_graphics = (h->job_type == MALI_JOB_TYPE_VERTEX) || (h->job_type == MALI_JOB_TYPE_TILER);
2020
2021 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false, gpu_id);
2022
2023 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2024 pandecode_indent++;
2025
2026 pandecode_vertex_tiler_prefix(&v->prefix, job_no, is_graphics);
2027 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2028
2029 struct MALI_PRIMITIVE primitive;
2030 struct mali_primitive_packed prim_packed = v->prefix.primitive;
2031 MALI_PRIMITIVE_unpack((const uint8_t *) &prim_packed, &primitive);
2032
2033 pandecode_primitive_size(v->primitive_size, primitive.point_size_array == 0);
2034
2035 pandecode_indent--;
2036 pandecode_log("};\n");
2037
2038 return sizeof(*v);
2039 }
2040
2041 static int
2042 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2043 mali_ptr payload, int job_no,
2044 bool is_bifrost, unsigned gpu_id)
2045 {
2046 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2047
2048 bool is_mfbd = s->framebuffer & MALI_MFBD;
2049
2050 if (!is_mfbd && is_bifrost)
2051 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2052
2053 struct pandecode_fbd info;
2054
2055 if (is_mfbd)
2056 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true, false, is_bifrost);
2057 else
2058 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true, gpu_id);
2059
2060 /* Compute the tag for the tagged pointer. This contains the type of
2061 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2062 * additional structures follow the MFBD header (an extra payload or
2063 * not, as well as a count of render targets) */
2064
2065 unsigned expected_tag = is_mfbd ? MALI_MFBD : 0;
2066
2067 if (is_mfbd) {
2068 if (info.has_extra)
2069 expected_tag |= MALI_MFBD_TAG_EXTRA;
2070
2071 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
2072 }
2073
2074 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
2075 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2076 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
2077 pandecode_prop("max_tile_coord = 0x%X\n", s->max_tile_coord);
2078 }
2079
2080 /* Extract tile coordinates */
2081
2082 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
2083 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
2084
2085 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2086 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2087
2088 /* For the max, we also want the floored (rather than ceiled) version for checking */
2089
2090 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
2091 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
2092
2093 /* Validate the coordinates are well-ordered */
2094
2095 if (min_x == max_x)
2096 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2097 else if (min_x > max_x)
2098 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2099
2100 if (min_y == max_y)
2101 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2102 else if (min_y > max_y)
2103 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2104
2105 /* Validate the coordinates fit inside the framebuffer. We use floor,
2106 * rather than ceil, for the max coordinates, since the tile
2107 * coordinates for something like an 800x600 framebuffer will actually
2108 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2109
2110 if ((min_x > info.width) || (max_x_f > info.width))
2111 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2112
2113 if ((min_y > info.height) || (max_y_f > info.height))
2114 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2115
2116 /* After validation, we print */
2117
2118 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
2119
2120 /* The FBD is a tagged pointer */
2121
2122 unsigned tag = (s->framebuffer & ~FBD_MASK);
2123
2124 if (tag != expected_tag)
2125 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
2126
2127 return sizeof(*s);
2128 }
2129
2130 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
2131 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
2132 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
2133 * might wish to combine this with the bifrost parameter) because some details
2134 * are model-specific even within a particular architecture. Minimal traces
2135 * *only* examine the job descriptors, skipping printing entirely if there is
2136 * no faults, and only descends into the payload if there are faults. This is
2137 * useful for looking for faults without the overhead of invasive traces. */
2138
2139 void
2140 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost, unsigned gpu_id, bool minimal)
2141 {
2142 pandecode_dump_file_open();
2143
2144 struct mali_job_descriptor_header *h;
2145 unsigned job_descriptor_number = 0;
2146
2147 do {
2148 struct pandecode_mapped_memory *mem =
2149 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2150
2151 void *payload;
2152
2153 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2154
2155 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h);
2156 payload = pandecode_fetch_gpu_mem(mem, payload_ptr, 64);
2157
2158 int job_no = job_descriptor_number++;
2159
2160 /* If the job is good to go, skip it in minimal mode */
2161 if (minimal && (h->exception_status == 0x0 || h->exception_status == 0x1))
2162 continue;
2163
2164 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2165 pandecode_indent++;
2166
2167 pandecode_prop("job_type = %s", mali_job_type_as_str(h->job_type));
2168
2169 if (h->job_descriptor_size)
2170 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2171
2172 if (h->exception_status && h->exception_status != 0x1)
2173 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2174 h->exception_status,
2175 (h->exception_status >> 16) & 0xFFFF,
2176 mali_exception_access_as_str((h->exception_status >> 8) & 0x3),
2177 h->exception_status & 0xFF);
2178
2179 if (h->first_incomplete_task)
2180 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2181
2182 if (h->fault_pointer)
2183 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2184
2185 if (h->job_barrier)
2186 pandecode_prop("job_barrier = %d", h->job_barrier);
2187
2188 pandecode_prop("job_index = %d", h->job_index);
2189
2190 if (h->unknown_flags)
2191 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2192
2193 if (h->job_dependency_index_1)
2194 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2195
2196 if (h->job_dependency_index_2)
2197 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2198
2199 pandecode_indent--;
2200 pandecode_log("};\n");
2201
2202 switch (h->job_type) {
2203 case MALI_JOB_TYPE_WRITE_VALUE: {
2204 struct mali_payload_write_value *s = payload;
2205 pandecode_log("struct mali_payload_write_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2206 pandecode_indent++;
2207 MEMORY_PROP(s, address);
2208
2209 if (s->value_descriptor != MALI_WRITE_VALUE_ZERO) {
2210 pandecode_msg("XXX: unknown value descriptor\n");
2211 pandecode_prop("value_descriptor = 0x%" PRIX32, s->value_descriptor);
2212 }
2213
2214 if (s->reserved) {
2215 pandecode_msg("XXX: set value tripped\n");
2216 pandecode_prop("reserved = 0x%" PRIX32, s->reserved);
2217 }
2218
2219 pandecode_prop("immediate = 0x%" PRIX64, s->immediate);
2220 pandecode_indent--;
2221 pandecode_log("};\n");
2222
2223 break;
2224 }
2225
2226 case MALI_JOB_TYPE_TILER:
2227 case MALI_JOB_TYPE_VERTEX:
2228 case MALI_JOB_TYPE_COMPUTE:
2229 if (bifrost) {
2230 if (h->job_type == MALI_JOB_TYPE_TILER)
2231 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2232 else
2233 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2234 } else
2235 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no, gpu_id);
2236
2237 break;
2238
2239 case MALI_JOB_TYPE_FRAGMENT:
2240 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost, gpu_id);
2241 break;
2242
2243 default:
2244 break;
2245 }
2246 } while ((jc_gpu_va = h->next_job));
2247
2248 pandecode_map_read_write();
2249 }