projects / mesa.git / blob
? search:


9f0867ba83af6119a950515d9f70e9ee984b2d39
[mesa.git] / src / gallium / drivers / freedreno / freedreno_screen.c
1 /*
2 * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 *
23 * Authors:
24 * Rob Clark <robclark@freedesktop.org>
25 */
26
27
28 #include "pipe/p_defines.h"
29 #include "pipe/p_screen.h"
30 #include "pipe/p_state.h"
31
32 #include "util/u_memory.h"
33 #include "util/u_inlines.h"
34 #include "util/format/u_format.h"
35 #include "util/format/u_format_s3tc.h"
36 #include "util/u_screen.h"
37 #include "util/u_string.h"
38 #include "util/u_debug.h"
39
40 #include "util/os_time.h"
41
42 #include "drm-uapi/drm_fourcc.h"
43 #include <errno.h>
44 #include <stdio.h>
45 #include <stdlib.h>
46 #include <sys/sysinfo.h>
47
48 #include "freedreno_screen.h"
49 #include "freedreno_resource.h"
50 #include "freedreno_fence.h"
51 #include "freedreno_query.h"
52 #include "freedreno_util.h"
53
54 #include "a2xx/fd2_screen.h"
55 #include "a3xx/fd3_screen.h"
56 #include "a4xx/fd4_screen.h"
57 #include "a5xx/fd5_screen.h"
58 #include "a6xx/fd6_screen.h"
59
60 /* for fd_get_driver/device_uuid() */
61 #include "common/freedreno_uuid.h"
62
63 #include "ir3/ir3_nir.h"
64 #include "ir3/ir3_compiler.h"
65 #include "a2xx/ir2.h"
66
67 static const struct debug_named_value debug_options[] = {
68 {"msgs", FD_DBG_MSGS, "Print debug messages"},
69 {"disasm", FD_DBG_DISASM, "Dump TGSI and adreno shader disassembly (a2xx only, see IR3_SHADER_DEBUG)"},
70 {"dclear", FD_DBG_DCLEAR, "Mark all state dirty after clear"},
71 {"ddraw", FD_DBG_DDRAW, "Mark all state dirty after draw"},
72 {"noscis", FD_DBG_NOSCIS, "Disable scissor optimization"},
73 {"direct", FD_DBG_DIRECT, "Force inline (SS_DIRECT) state loads"},
74 {"nobypass", FD_DBG_NOBYPASS, "Disable GMEM bypass"},
75 {"log", FD_DBG_LOG, "Enable GPU timestamp based logging (a6xx+)"},
76 {"nobin", FD_DBG_NOBIN, "Disable hw binning"},
77 {"nogmem", FD_DBG_NOGMEM, "Disable GMEM rendering (bypass only)"},
78 /* BIT(10) */
79 {"shaderdb", FD_DBG_SHADERDB, "Enable shaderdb output"},
80 {"flush", FD_DBG_FLUSH, "Force flush after every draw"},
81 {"deqp", FD_DBG_DEQP, "Enable dEQP hacks"},
82 {"inorder", FD_DBG_INORDER,"Disable reordering for draws/blits"},
83 {"bstat", FD_DBG_BSTAT, "Print batch stats at context destroy"},
84 {"nogrow", FD_DBG_NOGROW, "Disable \"growable\" cmdstream buffers, even if kernel supports it"},
85 {"lrz", FD_DBG_LRZ, "Enable experimental LRZ support (a5xx)"},
86 {"noindirect",FD_DBG_NOINDR, "Disable hw indirect draws (emulate on CPU)"},
87 {"noblit", FD_DBG_NOBLIT, "Disable blitter (fallback to generic blit path)"},
88 {"hiprio", FD_DBG_HIPRIO, "Force high-priority context"},
89 {"ttile", FD_DBG_TTILE, "Enable texture tiling (a2xx/a3xx/a5xx)"},
90 {"perfcntrs", FD_DBG_PERFC, "Expose performance counters"},
91 {"noubwc", FD_DBG_NOUBWC, "Disable UBWC for all internal buffers"},
92 {"nolrz", FD_DBG_NOLRZ, "Disable LRZ (a6xx)"},
93 {"notile", FD_DBG_NOTILE, "Disable tiling for all internal buffers"},
94 {"layout", FD_DBG_LAYOUT, "Dump resource layouts"},
95 {"nofp16", FD_DBG_NOFP16, "Disable mediump precision lowering"},
96 {"nohw", FD_DBG_NOHW, "Disable submitting commands to the HW"},
97 DEBUG_NAMED_VALUE_END
98 };
99
100 DEBUG_GET_ONCE_FLAGS_OPTION(fd_mesa_debug, "FD_MESA_DEBUG", debug_options, 0)
101
102 int fd_mesa_debug = 0;
103 bool fd_binning_enabled = true;
104
105 static const char *
106 fd_screen_get_name(struct pipe_screen *pscreen)
107 {
108 static char buffer[128];
109 snprintf(buffer, sizeof(buffer), "FD%03d",
110 fd_screen(pscreen)->device_id);
111 return buffer;
112 }
113
114 static const char *
115 fd_screen_get_vendor(struct pipe_screen *pscreen)
116 {
117 return "freedreno";
118 }
119
120 static const char *
121 fd_screen_get_device_vendor(struct pipe_screen *pscreen)
122 {
123 return "Qualcomm";
124 }
125
126
127 static uint64_t
128 fd_screen_get_timestamp(struct pipe_screen *pscreen)
129 {
130 struct fd_screen *screen = fd_screen(pscreen);
131
132 if (screen->has_timestamp) {
133 uint64_t n;
134 fd_pipe_get_param(screen->pipe, FD_TIMESTAMP, &n);
135 debug_assert(screen->max_freq > 0);
136 return n * 1000000000 / screen->max_freq;
137 } else {
138 int64_t cpu_time = os_time_get() * 1000;
139 return cpu_time + screen->cpu_gpu_time_delta;
140 }
141
142 }
143
144 static void
145 fd_screen_destroy(struct pipe_screen *pscreen)
146 {
147 struct fd_screen *screen = fd_screen(pscreen);
148
149 if (screen->pipe)
150 fd_pipe_del(screen->pipe);
151
152 if (screen->dev)
153 fd_device_del(screen->dev);
154
155 if (screen->ro)
156 FREE(screen->ro);
157
158 fd_bc_fini(&screen->batch_cache);
159 fd_gmem_screen_fini(pscreen);
160
161 slab_destroy_parent(&screen->transfer_pool);
162
163 simple_mtx_destroy(&screen->lock);
164
165 if (screen->compiler)
166 ir3_compiler_destroy(screen->compiler);
167
168 ralloc_free(screen->live_batches);
169
170 free(screen->perfcntr_queries);
171 free(screen);
172 }
173
174 /*
175 TODO either move caps to a2xx/a3xx specific code, or maybe have some
176 tables for things that differ if the delta is not too much..
177 */
178 static int
179 fd_screen_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
180 {
181 struct fd_screen *screen = fd_screen(pscreen);
182
183 /* this is probably not totally correct.. but it's a start: */
184 switch (param) {
185 /* Supported features (boolean caps). */
186 case PIPE_CAP_NPOT_TEXTURES:
187 case PIPE_CAP_MIXED_FRAMEBUFFER_SIZES:
188 case PIPE_CAP_ANISOTROPIC_FILTER:
189 case PIPE_CAP_POINT_SPRITE:
190 case PIPE_CAP_BLEND_EQUATION_SEPARATE:
191 case PIPE_CAP_TEXTURE_SWIZZLE:
192 case PIPE_CAP_MIXED_COLORBUFFER_FORMATS:
193 case PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT:
194 case PIPE_CAP_SEAMLESS_CUBE_MAP:
195 case PIPE_CAP_VERTEX_COLOR_UNCLAMPED:
196 case PIPE_CAP_QUADS_FOLLOW_PROVOKING_VERTEX_CONVENTION:
197 case PIPE_CAP_BUFFER_MAP_PERSISTENT_COHERENT:
198 case PIPE_CAP_STRING_MARKER:
199 case PIPE_CAP_MIXED_COLOR_DEPTH_BITS:
200 case PIPE_CAP_TEXTURE_BARRIER:
201 case PIPE_CAP_INVALIDATE_BUFFER:
202 case PIPE_CAP_RGB_OVERRIDE_DST_ALPHA_BLEND:
203 case PIPE_CAP_GLSL_TESS_LEVELS_AS_INPUTS:
204 case PIPE_CAP_NIR_COMPACT_ARRAYS:
205 return 1;
206
207 case PIPE_CAP_VERTEX_BUFFER_OFFSET_4BYTE_ALIGNED_ONLY:
208 case PIPE_CAP_VERTEX_BUFFER_STRIDE_4BYTE_ALIGNED_ONLY:
209 case PIPE_CAP_VERTEX_ELEMENT_SRC_OFFSET_4BYTE_ALIGNED_ONLY:
210 return !is_a2xx(screen);
211
212 case PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER:
213 return is_a2xx(screen);
214 case PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER:
215 return !is_a2xx(screen);
216
217 case PIPE_CAP_PACKED_UNIFORMS:
218 return !is_a2xx(screen);
219
220 case PIPE_CAP_ROBUST_BUFFER_ACCESS_BEHAVIOR:
221 case PIPE_CAP_DEVICE_RESET_STATUS_QUERY:
222 return screen->has_robustness;
223
224 case PIPE_CAP_VERTEXID_NOBASE:
225 return is_a3xx(screen) || is_a4xx(screen);
226
227 case PIPE_CAP_COMPUTE:
228 return has_compute(screen);
229
230 case PIPE_CAP_PREFER_BLIT_BASED_TEXTURE_TRANSFER:
231 case PIPE_CAP_PCI_GROUP:
232 case PIPE_CAP_PCI_BUS:
233 case PIPE_CAP_PCI_DEVICE:
234 case PIPE_CAP_PCI_FUNCTION:
235 case PIPE_CAP_DEPTH_CLIP_DISABLE_SEPARATE:
236 return 0;
237
238 case PIPE_CAP_FRAGMENT_SHADER_TEXTURE_LOD:
239 case PIPE_CAP_FRAGMENT_SHADER_DERIVATIVES:
240 case PIPE_CAP_VERTEX_SHADER_SATURATE:
241 case PIPE_CAP_PRIMITIVE_RESTART:
242 case PIPE_CAP_PRIMITIVE_RESTART_FIXED_INDEX:
243 case PIPE_CAP_TGSI_INSTANCEID:
244 case PIPE_CAP_VERTEX_ELEMENT_INSTANCE_DIVISOR:
245 case PIPE_CAP_INDEP_BLEND_ENABLE:
246 case PIPE_CAP_INDEP_BLEND_FUNC:
247 case PIPE_CAP_TEXTURE_BUFFER_OBJECTS:
248 case PIPE_CAP_TEXTURE_HALF_FLOAT_LINEAR:
249 case PIPE_CAP_CONDITIONAL_RENDER:
250 case PIPE_CAP_CONDITIONAL_RENDER_INVERTED:
251 case PIPE_CAP_SEAMLESS_CUBE_MAP_PER_TEXTURE:
252 case PIPE_CAP_CLIP_HALFZ:
253 return is_a3xx(screen) || is_a4xx(screen) || is_a5xx(screen) || is_a6xx(screen);
254
255 case PIPE_CAP_FAKE_SW_MSAA:
256 return !fd_screen_get_param(pscreen, PIPE_CAP_TEXTURE_MULTISAMPLE);
257
258 case PIPE_CAP_TEXTURE_MULTISAMPLE:
259 return is_a5xx(screen) || is_a6xx(screen);
260
261 case PIPE_CAP_SURFACE_SAMPLE_COUNT:
262 return is_a6xx(screen);
263
264 case PIPE_CAP_DEPTH_CLIP_DISABLE:
265 return is_a3xx(screen) || is_a4xx(screen);
266
267 case PIPE_CAP_POLYGON_OFFSET_CLAMP:
268 return is_a4xx(screen) || is_a5xx(screen) || is_a6xx(screen);
269
270 case PIPE_CAP_TEXTURE_BUFFER_OFFSET_ALIGNMENT:
271 if (is_a3xx(screen)) return 16;
272 if (is_a4xx(screen)) return 32;
273 if (is_a5xx(screen)) return 32;
274 if (is_a6xx(screen)) return 64;
275 return 0;
276 case PIPE_CAP_MAX_TEXTURE_BUFFER_SIZE:
277 /* We could possibly emulate more by pretending 2d/rect textures and
278 * splitting high bits of index into 2nd dimension..
279 */
280 if (is_a3xx(screen)) return 8192;
281 if (is_a4xx(screen)) return 16384;
282 if (is_a5xx(screen)) return 16384;
283 if (is_a6xx(screen)) return 1 << 27;
284 return 0;
285
286 case PIPE_CAP_TEXTURE_FLOAT_LINEAR:
287 case PIPE_CAP_CUBE_MAP_ARRAY:
288 case PIPE_CAP_SAMPLER_VIEW_TARGET:
289 case PIPE_CAP_TEXTURE_QUERY_LOD:
290 return is_a4xx(screen) || is_a5xx(screen) || is_a6xx(screen);
291
292 case PIPE_CAP_START_INSTANCE:
293 /* Note that a5xx can do this, it just can't (at least with
294 * current firmware) do draw_indirect with base_instance.
295 * Since draw_indirect is needed sooner (gles31 and gl40 vs
296 * gl42), hide base_instance on a5xx. :-/
297 */
298 return is_a4xx(screen);
299
300 case PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT:
301 return 64;
302
303 case PIPE_CAP_GLSL_FEATURE_LEVEL:
304 case PIPE_CAP_GLSL_FEATURE_LEVEL_COMPATIBILITY:
305 return is_ir3(screen) ? 140 : 120;
306
307 case PIPE_CAP_ESSL_FEATURE_LEVEL:
308 /* we can probably enable 320 for a5xx too, but need to test: */
309 if (is_a6xx(screen)) return 320;
310 if (is_a5xx(screen)) return 310;
311 if (is_ir3(screen)) return 300;
312 return 120;
313
314 case PIPE_CAP_SHADER_BUFFER_OFFSET_ALIGNMENT:
315 if (is_a6xx(screen)) return 64;
316 if (is_a5xx(screen)) return 4;
317 return 0;
318
319 case PIPE_CAP_MAX_TEXTURE_GATHER_COMPONENTS:
320 if (is_a4xx(screen) || is_a5xx(screen) || is_a6xx(screen))
321 return 4;
322 return 0;
323
324 /* TODO if we need this, do it in nir/ir3 backend to avoid breaking precompile: */
325 case PIPE_CAP_FORCE_PERSAMPLE_INTERP:
326 return 0;
327
328 case PIPE_CAP_FBFETCH:
329 if (fd_device_version(screen->dev) >= FD_VERSION_GMEM_BASE &&
330 is_a6xx(screen))
331 return 1;
332 return 0;
333 case PIPE_CAP_SAMPLE_SHADING:
334 if (is_a6xx(screen)) return 1;
335 return 0;
336
337 case PIPE_CAP_CONTEXT_PRIORITY_MASK:
338 return screen->priority_mask;
339
340 case PIPE_CAP_DRAW_INDIRECT:
341 if (is_a4xx(screen) || is_a5xx(screen) || is_a6xx(screen))
342 return 1;
343 return 0;
344
345 case PIPE_CAP_FRAMEBUFFER_NO_ATTACHMENT:
346 if (is_a4xx(screen) || is_a5xx(screen) || is_a6xx(screen))
347 return 1;
348 return 0;
349
350 case PIPE_CAP_LOAD_CONSTBUF:
351 /* name is confusing, but this turns on std430 packing */
352 if (is_ir3(screen))
353 return 1;
354 return 0;
355
356 case PIPE_CAP_NIR_IMAGES_AS_DEREF:
357 return 0;
358
359 case PIPE_CAP_MAX_VIEWPORTS:
360 return 1;
361
362 case PIPE_CAP_MAX_VARYINGS:
363 return 16;
364
365 case PIPE_CAP_MAX_SHADER_PATCH_VARYINGS:
366 /* We don't really have a limit on this, it all goes into the main
367 * memory buffer. Needs to be at least 120 / 4 (minimum requirement
368 * for GL_MAX_TESS_PATCH_COMPONENTS).
369 */
370 return 128;
371
372 case PIPE_CAP_MAX_TEXTURE_UPLOAD_MEMORY_BUDGET:
373 return 64 * 1024 * 1024;
374
375 case PIPE_CAP_SHAREABLE_SHADERS:
376 case PIPE_CAP_GLSL_OPTIMIZE_CONSERVATIVELY:
377 /* manage the variants for these ourself, to avoid breaking precompile: */
378 case PIPE_CAP_FRAGMENT_COLOR_CLAMPED:
379 case PIPE_CAP_VERTEX_COLOR_CLAMPED:
380 if (is_ir3(screen))
381 return 1;
382 return 0;
383
384 /* Geometry shaders.. */
385 case PIPE_CAP_MAX_GEOMETRY_OUTPUT_VERTICES:
386 return 512;
387 case PIPE_CAP_MAX_GEOMETRY_TOTAL_OUTPUT_COMPONENTS:
388 return 2048;
389 case PIPE_CAP_MAX_GS_INVOCATIONS:
390 return 32;
391
392 /* Stream output. */
393 case PIPE_CAP_MAX_STREAM_OUTPUT_BUFFERS:
394 if (is_ir3(screen))
395 return PIPE_MAX_SO_BUFFERS;
396 return 0;
397 case PIPE_CAP_STREAM_OUTPUT_PAUSE_RESUME:
398 case PIPE_CAP_STREAM_OUTPUT_INTERLEAVE_BUFFERS:
399 case PIPE_CAP_TGSI_FS_POSITION_IS_SYSVAL:
400 case PIPE_CAP_TGSI_TEXCOORD:
401 if (is_ir3(screen))
402 return 1;
403 return 0;
404 case PIPE_CAP_TGSI_FS_FACE_IS_INTEGER_SYSVAL:
405 return 1;
406 case PIPE_CAP_TGSI_FS_POINT_IS_SYSVAL:
407 return is_a2xx(screen);
408 case PIPE_CAP_MAX_STREAM_OUTPUT_SEPARATE_COMPONENTS:
409 case PIPE_CAP_MAX_STREAM_OUTPUT_INTERLEAVED_COMPONENTS:
410 if (is_ir3(screen))
411 return 16 * 4; /* should only be shader out limit? */
412 return 0;
413
414 /* Texturing. */
415 case PIPE_CAP_MAX_TEXTURE_2D_SIZE:
416 if (is_a6xx(screen) || is_a5xx(screen) || is_a4xx(screen))
417 return 16384;
418 else
419 return 8192;
420 case PIPE_CAP_MAX_TEXTURE_CUBE_LEVELS:
421 if (is_a6xx(screen) || is_a5xx(screen) || is_a4xx(screen))
422 return 15;
423 else
424 return 14;
425 case PIPE_CAP_MAX_TEXTURE_3D_LEVELS:
426 return 11;
427
428 case PIPE_CAP_MAX_TEXTURE_ARRAY_LAYERS:
429 return (is_a3xx(screen) || is_a4xx(screen) || is_a5xx(screen) || is_a6xx(screen)) ? 256 : 0;
430
431 /* Render targets. */
432 case PIPE_CAP_MAX_RENDER_TARGETS:
433 return screen->max_rts;
434 case PIPE_CAP_MAX_DUAL_SOURCE_RENDER_TARGETS:
435 return is_a3xx(screen) ? 1 : 0;
436
437 /* Queries. */
438 case PIPE_CAP_OCCLUSION_QUERY:
439 return is_a3xx(screen) || is_a4xx(screen) || is_a5xx(screen) || is_a6xx(screen);
440 case PIPE_CAP_QUERY_TIMESTAMP:
441 case PIPE_CAP_QUERY_TIME_ELAPSED:
442 /* only a4xx, requires new enough kernel so we know max_freq: */
443 return (screen->max_freq > 0) && (is_a4xx(screen) || is_a5xx(screen) || is_a6xx(screen));
444
445 case PIPE_CAP_VENDOR_ID:
446 return 0x5143;
447 case PIPE_CAP_DEVICE_ID:
448 return 0xFFFFFFFF;
449 case PIPE_CAP_ACCELERATED:
450 return 1;
451 case PIPE_CAP_VIDEO_MEMORY:
452 DBG("FINISHME: The value returned is incorrect\n");
453 return 10;
454 case PIPE_CAP_UMA:
455 return 1;
456 case PIPE_CAP_MEMOBJ:
457 return fd_device_version(screen->dev) >= FD_VERSION_MEMORY_FD;
458 case PIPE_CAP_NATIVE_FENCE_FD:
459 return fd_device_version(screen->dev) >= FD_VERSION_FENCE_FD;
460 default:
461 return u_pipe_screen_get_param_defaults(pscreen, param);
462 }
463 }
464
465 static float
466 fd_screen_get_paramf(struct pipe_screen *pscreen, enum pipe_capf param)
467 {
468 switch (param) {
469 case PIPE_CAPF_MAX_LINE_WIDTH:
470 case PIPE_CAPF_MAX_LINE_WIDTH_AA:
471 /* NOTE: actual value is 127.0f, but this is working around a deqp
472 * bug.. dEQP-GLES3.functional.rasterization.primitives.lines_wide
473 * uses too small of a render target size, and gets confused when
474 * the lines start going offscreen.
475 *
476 * See: https://code.google.com/p/android/issues/detail?id=206513
477 */
478 if (fd_mesa_debug & FD_DBG_DEQP)
479 return 48.0f;
480 return 127.0f;
481 case PIPE_CAPF_MAX_POINT_WIDTH:
482 case PIPE_CAPF_MAX_POINT_WIDTH_AA:
483 return 4092.0f;
484 case PIPE_CAPF_MAX_TEXTURE_ANISOTROPY:
485 return 16.0f;
486 case PIPE_CAPF_MAX_TEXTURE_LOD_BIAS:
487 return 15.0f;
488 case PIPE_CAPF_MIN_CONSERVATIVE_RASTER_DILATE:
489 case PIPE_CAPF_MAX_CONSERVATIVE_RASTER_DILATE:
490 case PIPE_CAPF_CONSERVATIVE_RASTER_DILATE_GRANULARITY:
491 return 0.0f;
492 }
493 debug_printf("unknown paramf %d\n", param);
494 return 0;
495 }
496
497 static int
498 fd_screen_get_shader_param(struct pipe_screen *pscreen,
499 enum pipe_shader_type shader,
500 enum pipe_shader_cap param)
501 {
502 struct fd_screen *screen = fd_screen(pscreen);
503
504 switch(shader)
505 {
506 case PIPE_SHADER_FRAGMENT:
507 case PIPE_SHADER_VERTEX:
508 break;
509 case PIPE_SHADER_TESS_CTRL:
510 case PIPE_SHADER_TESS_EVAL:
511 case PIPE_SHADER_GEOMETRY:
512 if (is_a6xx(screen))
513 break;
514 return 0;
515 case PIPE_SHADER_COMPUTE:
516 if (has_compute(screen))
517 break;
518 return 0;
519 default:
520 DBG("unknown shader type %d", shader);
521 return 0;
522 }
523
524 /* this is probably not totally correct.. but it's a start: */
525 switch (param) {
526 case PIPE_SHADER_CAP_MAX_INSTRUCTIONS:
527 case PIPE_SHADER_CAP_MAX_ALU_INSTRUCTIONS:
528 case PIPE_SHADER_CAP_MAX_TEX_INSTRUCTIONS:
529 case PIPE_SHADER_CAP_MAX_TEX_INDIRECTIONS:
530 return 16384;
531 case PIPE_SHADER_CAP_MAX_CONTROL_FLOW_DEPTH:
532 return 8; /* XXX */
533 case PIPE_SHADER_CAP_MAX_INPUTS:
534 case PIPE_SHADER_CAP_MAX_OUTPUTS:
535 return 16;
536 case PIPE_SHADER_CAP_MAX_TEMPS:
537 return 64; /* Max native temporaries. */
538 case PIPE_SHADER_CAP_MAX_CONST_BUFFER_SIZE:
539 /* NOTE: seems to be limit for a3xx is actually 512 but
540 * split between VS and FS. Use lower limit of 256 to
541 * avoid getting into impossible situations:
542 */
543 return ((is_a3xx(screen) || is_a4xx(screen) || is_a5xx(screen) || is_a6xx(screen)) ? 4096 : 64) * sizeof(float[4]);
544 case PIPE_SHADER_CAP_MAX_CONST_BUFFERS:
545 return is_ir3(screen) ? 16 : 1;
546 case PIPE_SHADER_CAP_TGSI_CONT_SUPPORTED:
547 return 1;
548 case PIPE_SHADER_CAP_INDIRECT_INPUT_ADDR:
549 case PIPE_SHADER_CAP_INDIRECT_OUTPUT_ADDR:
550 /* Technically this should be the same as for TEMP/CONST, since
551 * everything is just normal registers. This is just temporary
552 * hack until load_input/store_output handle arrays in a similar
553 * way as load_var/store_var..
554 *
555 * For tessellation stages, inputs are loaded using ldlw or ldg, both
556 * of which support indirection.
557 */
558 return shader == PIPE_SHADER_TESS_CTRL || shader == PIPE_SHADER_TESS_EVAL;
559 case PIPE_SHADER_CAP_INDIRECT_TEMP_ADDR:
560 case PIPE_SHADER_CAP_INDIRECT_CONST_ADDR:
561 /* a2xx compiler doesn't handle indirect: */
562 return is_ir3(screen) ? 1 : 0;
563 case PIPE_SHADER_CAP_SUBROUTINES:
564 case PIPE_SHADER_CAP_TGSI_DROUND_SUPPORTED:
565 case PIPE_SHADER_CAP_TGSI_DFRACEXP_DLDEXP_SUPPORTED:
566 case PIPE_SHADER_CAP_TGSI_LDEXP_SUPPORTED:
567 case PIPE_SHADER_CAP_TGSI_FMA_SUPPORTED:
568 case PIPE_SHADER_CAP_TGSI_ANY_INOUT_DECL_RANGE:
569 case PIPE_SHADER_CAP_MAX_HW_ATOMIC_COUNTERS:
570 case PIPE_SHADER_CAP_LOWER_IF_THRESHOLD:
571 case PIPE_SHADER_CAP_TGSI_SKIP_MERGE_REGISTERS:
572 case PIPE_SHADER_CAP_MAX_HW_ATOMIC_COUNTER_BUFFERS:
573 return 0;
574 case PIPE_SHADER_CAP_TGSI_SQRT_SUPPORTED:
575 return 1;
576 case PIPE_SHADER_CAP_INTEGERS:
577 return is_ir3(screen) ? 1 : 0;
578 case PIPE_SHADER_CAP_INT64_ATOMICS:
579 case PIPE_SHADER_CAP_FP16_DERIVATIVES:
580 case PIPE_SHADER_CAP_INT16:
581 case PIPE_SHADER_CAP_GLSL_16BIT_CONSTS:
582 return 0;
583 case PIPE_SHADER_CAP_FP16:
584 return ((is_a5xx(screen) || is_a6xx(screen)) &&
585 (shader == PIPE_SHADER_COMPUTE ||
586 shader == PIPE_SHADER_FRAGMENT) &&
587 !(fd_mesa_debug & FD_DBG_NOFP16));
588 case PIPE_SHADER_CAP_MAX_TEXTURE_SAMPLERS:
589 case PIPE_SHADER_CAP_MAX_SAMPLER_VIEWS:
590 return 16;
591 case PIPE_SHADER_CAP_PREFERRED_IR:
592 return PIPE_SHADER_IR_NIR;
593 case PIPE_SHADER_CAP_SUPPORTED_IRS:
594 return (1 << PIPE_SHADER_IR_NIR) | (1 << PIPE_SHADER_IR_TGSI);
595 case PIPE_SHADER_CAP_MAX_UNROLL_ITERATIONS_HINT:
596 return 32;
597 case PIPE_SHADER_CAP_MAX_SHADER_BUFFERS:
598 case PIPE_SHADER_CAP_MAX_SHADER_IMAGES:
599 if (is_a5xx(screen) || is_a6xx(screen)) {
600 /* a5xx (and a4xx for that matter) has one state-block
601 * for compute-shader SSBO's and another that is shared
602 * by VS/HS/DS/GS/FS.. so to simplify things for now
603 * just advertise SSBOs for FS and CS. We could possibly
604 * do what blob does, and partition the space for
605 * VS/HS/DS/GS/FS. The blob advertises:
606 *
607 * GL_MAX_VERTEX_SHADER_STORAGE_BLOCKS: 4
608 * GL_MAX_GEOMETRY_SHADER_STORAGE_BLOCKS: 4
609 * GL_MAX_TESS_CONTROL_SHADER_STORAGE_BLOCKS: 4
610 * GL_MAX_TESS_EVALUATION_SHADER_STORAGE_BLOCKS: 4
611 * GL_MAX_FRAGMENT_SHADER_STORAGE_BLOCKS: 4
612 * GL_MAX_COMPUTE_SHADER_STORAGE_BLOCKS: 24
613 * GL_MAX_COMBINED_SHADER_STORAGE_BLOCKS: 24
614 *
615 * I think that way we could avoid having to patch shaders
616 * for actual SSBO indexes by using a static partitioning.
617 *
618 * Note same state block is used for images and buffers,
619 * but images also need texture state for read access
620 * (isam/isam.3d)
621 */
622 switch(shader)
623 {
624 case PIPE_SHADER_FRAGMENT:
625 case PIPE_SHADER_COMPUTE:
626 return 24;
627 default:
628 return 0;
629 }
630 }
631 return 0;
632 }
633 debug_printf("unknown shader param %d\n", param);
634 return 0;
635 }
636
637 /* TODO depending on how much the limits differ for a3xx/a4xx, maybe move this
638 * into per-generation backend?
639 */
640 static int
641 fd_get_compute_param(struct pipe_screen *pscreen, enum pipe_shader_ir ir_type,
642 enum pipe_compute_cap param, void *ret)
643 {
644 struct fd_screen *screen = fd_screen(pscreen);
645 const char * const ir = "ir3";
646
647 if (!has_compute(screen))
648 return 0;
649
650 #define RET(x) do { \
651 if (ret) \
652 memcpy(ret, x, sizeof(x)); \
653 return sizeof(x); \
654 } while (0)
655
656 switch (param) {
657 case PIPE_COMPUTE_CAP_ADDRESS_BITS:
658 // don't expose 64b pointer support yet, until ir3 supports 64b
659 // math, otherwise spir64 target is used and we get 64b pointer
660 // calculations that we can't do yet
661 // if (is_a5xx(screen))
662 // RET((uint32_t []){ 64 });
663 RET((uint32_t []){ 32 });
664
665 case PIPE_COMPUTE_CAP_IR_TARGET:
666 if (ret)
667 sprintf(ret, "%s", ir);
668 return strlen(ir) * sizeof(char);
669
670 case PIPE_COMPUTE_CAP_GRID_DIMENSION:
671 RET((uint64_t []) { 3 });
672
673 case PIPE_COMPUTE_CAP_MAX_GRID_SIZE:
674 RET(((uint64_t []) { 65535, 65535, 65535 }));
675
676 case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE:
677 RET(((uint64_t []) { 1024, 1024, 64 }));
678
679 case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK:
680 RET((uint64_t []) { 1024 });
681
682 case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE:
683 RET((uint64_t []) { screen->ram_size });
684
685 case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE:
686 RET((uint64_t []) { 32768 });
687
688 case PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE:
689 case PIPE_COMPUTE_CAP_MAX_INPUT_SIZE:
690 RET((uint64_t []) { 4096 });
691
692 case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE:
693 RET((uint64_t []) { screen->ram_size });
694
695 case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY:
696 RET((uint32_t []) { screen->max_freq / 1000000 });
697
698 case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS:
699 RET((uint32_t []) { 9999 }); // TODO
700
701 case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED:
702 RET((uint32_t []) { 1 });
703
704 case PIPE_COMPUTE_CAP_SUBGROUP_SIZE:
705 RET((uint32_t []) { 32 }); // TODO
706
707 case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK:
708 RET((uint64_t []) { 1024 }); // TODO
709 }
710
711 return 0;
712 }
713
714 static const void *
715 fd_get_compiler_options(struct pipe_screen *pscreen,
716 enum pipe_shader_ir ir, unsigned shader)
717 {
718 struct fd_screen *screen = fd_screen(pscreen);
719
720 if (is_ir3(screen))
721 return ir3_get_compiler_options(screen->compiler);
722
723 return ir2_get_compiler_options();
724 }
725
726 static struct disk_cache *
727 fd_get_disk_shader_cache(struct pipe_screen *pscreen)
728 {
729 struct fd_screen *screen = fd_screen(pscreen);
730
731 if (is_ir3(screen)) {
732 struct ir3_compiler *compiler = screen->compiler;
733 return compiler->disk_cache;
734 }
735
736 return NULL;
737 }
738
739 bool
740 fd_screen_bo_get_handle(struct pipe_screen *pscreen,
741 struct fd_bo *bo,
742 struct renderonly_scanout *scanout,
743 unsigned stride,
744 struct winsys_handle *whandle)
745 {
746 whandle->stride = stride;
747
748 if (whandle->type == WINSYS_HANDLE_TYPE_SHARED) {
749 return fd_bo_get_name(bo, &whandle->handle) == 0;
750 } else if (whandle->type == WINSYS_HANDLE_TYPE_KMS) {
751 if (renderonly_get_handle(scanout, whandle))
752 return true;
753 whandle->handle = fd_bo_handle(bo);
754 return true;
755 } else if (whandle->type == WINSYS_HANDLE_TYPE_FD) {
756 whandle->handle = fd_bo_dmabuf(bo);
757 return true;
758 } else {
759 return false;
760 }
761 }
762
763 static void
764 fd_screen_query_dmabuf_modifiers(struct pipe_screen *pscreen,
765 enum pipe_format format,
766 int max, uint64_t *modifiers,
767 unsigned int *external_only,
768 int *count)
769 {
770 struct fd_screen *screen = fd_screen(pscreen);
771 int i, num = 0;
772
773 max = MIN2(max, screen->num_supported_modifiers);
774
775 if (!max) {
776 max = screen->num_supported_modifiers;
777 external_only = NULL;
778 modifiers = NULL;
779 }
780
781 for (i = 0; i < max; i++) {
782 if (modifiers)
783 modifiers[num] = screen->supported_modifiers[i];
784
785 if (external_only)
786 external_only[num] = 0;
787
788 num++;
789 }
790
791 *count = num;
792 }
793
794 struct fd_bo *
795 fd_screen_bo_from_handle(struct pipe_screen *pscreen,
796 struct winsys_handle *whandle)
797 {
798 struct fd_screen *screen = fd_screen(pscreen);
799 struct fd_bo *bo;
800
801 if (whandle->type == WINSYS_HANDLE_TYPE_SHARED) {
802 bo = fd_bo_from_name(screen->dev, whandle->handle);
803 } else if (whandle->type == WINSYS_HANDLE_TYPE_KMS) {
804 bo = fd_bo_from_handle(screen->dev, whandle->handle, 0);
805 } else if (whandle->type == WINSYS_HANDLE_TYPE_FD) {
806 bo = fd_bo_from_dmabuf(screen->dev, whandle->handle);
807 } else {
808 DBG("Attempt to import unsupported handle type %d", whandle->type);
809 return NULL;
810 }
811
812 if (!bo) {
813 DBG("ref name 0x%08x failed", whandle->handle);
814 return NULL;
815 }
816
817 return bo;
818 }
819
820 static void _fd_fence_ref(struct pipe_screen *pscreen,
821 struct pipe_fence_handle **ptr,
822 struct pipe_fence_handle *pfence)
823 {
824 fd_fence_ref(ptr, pfence);
825 }
826
827 static void
828 fd_screen_get_device_uuid(struct pipe_screen *pscreen, char *uuid)
829 {
830 struct fd_screen *screen = fd_screen(pscreen);
831
832 fd_get_device_uuid(uuid, screen->gpu_id);
833 }
834
835 static void
836 fd_screen_get_driver_uuid(struct pipe_screen *pscreen, char *uuid)
837 {
838 fd_get_driver_uuid(uuid);
839 }
840
841 struct pipe_screen *
842 fd_screen_create(struct fd_device *dev, struct renderonly *ro)
843 {
844 struct fd_screen *screen = CALLOC_STRUCT(fd_screen);
845 struct pipe_screen *pscreen;
846 uint64_t val;
847
848 fd_mesa_debug = debug_get_option_fd_mesa_debug();
849
850 if (fd_mesa_debug & FD_DBG_NOBIN)
851 fd_binning_enabled = false;
852
853 if (!screen)
854 return NULL;
855
856 pscreen = &screen->base;
857
858 screen->dev = dev;
859 screen->refcnt = 1;
860
861 if (ro) {
862 screen->ro = renderonly_dup(ro);
863 if (!screen->ro) {
864 DBG("could not create renderonly object");
865 goto fail;
866 }
867 }
868
869 // maybe this should be in context?
870 screen->pipe = fd_pipe_new(screen->dev, FD_PIPE_3D);
871 if (!screen->pipe) {
872 DBG("could not create 3d pipe");
873 goto fail;
874 }
875
876 if (fd_pipe_get_param(screen->pipe, FD_GMEM_SIZE, &val)) {
877 DBG("could not get GMEM size");
878 goto fail;
879 }
880 screen->gmemsize_bytes = val;
881
882 if (fd_device_version(dev) >= FD_VERSION_GMEM_BASE) {
883 fd_pipe_get_param(screen->pipe, FD_GMEM_BASE, &screen->gmem_base);
884 }
885
886 if (fd_pipe_get_param(screen->pipe, FD_DEVICE_ID, &val)) {
887 DBG("could not get device-id");
888 goto fail;
889 }
890 screen->device_id = val;
891
892 if (fd_pipe_get_param(screen->pipe, FD_MAX_FREQ, &val)) {
893 DBG("could not get gpu freq");
894 /* this limits what performance related queries are
895 * supported but is not fatal
896 */
897 screen->max_freq = 0;
898 } else {
899 screen->max_freq = val;
900 if (fd_pipe_get_param(screen->pipe, FD_TIMESTAMP, &val) == 0)
901 screen->has_timestamp = true;
902 }
903
904 if (fd_pipe_get_param(screen->pipe, FD_GPU_ID, &val)) {
905 DBG("could not get gpu-id");
906 goto fail;
907 }
908 screen->gpu_id = val;
909
910 if (fd_pipe_get_param(screen->pipe, FD_CHIP_ID, &val)) {
911 DBG("could not get chip-id");
912 /* older kernels may not have this property: */
913 unsigned core = screen->gpu_id / 100;
914 unsigned major = (screen->gpu_id % 100) / 10;
915 unsigned minor = screen->gpu_id % 10;
916 unsigned patch = 0; /* assume the worst */
917 val = (patch & 0xff) | ((minor & 0xff) << 8) |
918 ((major & 0xff) << 16) | ((core & 0xff) << 24);
919 }
920 screen->chip_id = val;
921
922 if (fd_pipe_get_param(screen->pipe, FD_NR_RINGS, &val)) {
923 DBG("could not get # of rings");
924 screen->priority_mask = 0;
925 } else {
926 /* # of rings equates to number of unique priority values: */
927 screen->priority_mask = (1 << val) - 1;
928 }
929
930 if (fd_device_version(dev) >= FD_VERSION_ROBUSTNESS)
931 screen->has_robustness = true;
932
933 struct sysinfo si;
934 sysinfo(&si);
935 screen->ram_size = si.totalram;
936
937 DBG("Pipe Info:");
938 DBG(" GPU-id: %d", screen->gpu_id);
939 DBG(" Chip-id: 0x%08x", screen->chip_id);
940 DBG(" GMEM size: 0x%08x", screen->gmemsize_bytes);
941
942 /* explicitly checking for GPU revisions that are known to work. This
943 * may be overly conservative for a3xx, where spoofing the gpu_id with
944 * the blob driver seems to generate identical cmdstream dumps. But
945 * on a2xx, there seem to be small differences between the GPU revs
946 * so it is probably better to actually test first on real hardware
947 * before enabling:
948 *
949 * If you have a different adreno version, feel free to add it to one
950 * of the cases below and see what happens. And if it works, please
951 * send a patch ;-)
952 */
953 switch (screen->gpu_id) {
954 case 200:
955 case 201:
956 case 205:
957 case 220:
958 fd2_screen_init(pscreen);
959 break;
960 case 305:
961 case 307:
962 case 320:
963 case 330:
964 fd3_screen_init(pscreen);
965 break;
966 case 405:
967 case 420:
968 case 430:
969 fd4_screen_init(pscreen);
970 break;
971 case 510:
972 case 530:
973 case 540:
974 fd5_screen_init(pscreen);
975 break;
976 case 618:
977 case 630:
978 case 640:
979 case 650:
980 fd6_screen_init(pscreen);
981 break;
982 default:
983 debug_printf("unsupported GPU: a%03d\n", screen->gpu_id);
984 goto fail;
985 }
986
987 if (screen->gpu_id >= 600) {
988 screen->gmem_alignw = 16;
989 screen->gmem_alignh = 4;
990 screen->tile_alignw = is_a650(screen) ? 96 : 32;
991 screen->tile_alignh = 32;
992 screen->num_vsc_pipes = 32;
993 } else if (screen->gpu_id >= 500) {
994 screen->gmem_alignw = screen->tile_alignw = 64;
995 screen->gmem_alignh = screen->tile_alignh = 32;
996 screen->num_vsc_pipes = 16;
997 } else {
998 screen->gmem_alignw = screen->tile_alignw = 32;
999 screen->gmem_alignh = screen->tile_alignh = 32;
1000 screen->num_vsc_pipes = 8;
1001 }
1002
1003 if (fd_mesa_debug & FD_DBG_PERFC) {
1004 screen->perfcntr_groups = fd_perfcntrs(screen->gpu_id,
1005 &screen->num_perfcntr_groups);
1006 }
1007
1008 /* NOTE: don't enable if we have too old of a kernel to support
1009 * growable cmdstream buffers, since memory requirement for cmdstream
1010 * buffers would be too much otherwise.
1011 */
1012 if (fd_device_version(dev) >= FD_VERSION_UNLIMITED_CMDS)
1013 screen->reorder = !(fd_mesa_debug & FD_DBG_INORDER);
1014
1015 if (BATCH_DEBUG)
1016 screen->live_batches = _mesa_pointer_set_create(NULL);
1017
1018 fd_bc_init(&screen->batch_cache);
1019
1020 list_inithead(&screen->context_list);
1021
1022 (void) simple_mtx_init(&screen->lock, mtx_plain);
1023
1024 pscreen->destroy = fd_screen_destroy;
1025 pscreen->get_param = fd_screen_get_param;
1026 pscreen->get_paramf = fd_screen_get_paramf;
1027 pscreen->get_shader_param = fd_screen_get_shader_param;
1028 pscreen->get_compute_param = fd_get_compute_param;
1029 pscreen->get_compiler_options = fd_get_compiler_options;
1030 pscreen->get_disk_shader_cache = fd_get_disk_shader_cache;
1031
1032 fd_resource_screen_init(pscreen);
1033 fd_query_screen_init(pscreen);
1034 fd_gmem_screen_init(pscreen);
1035
1036 pscreen->get_name = fd_screen_get_name;
1037 pscreen->get_vendor = fd_screen_get_vendor;
1038 pscreen->get_device_vendor = fd_screen_get_device_vendor;
1039
1040 pscreen->get_timestamp = fd_screen_get_timestamp;
1041
1042 pscreen->fence_reference = _fd_fence_ref;
1043 pscreen->fence_finish = fd_fence_finish;
1044 pscreen->fence_get_fd = fd_fence_get_fd;
1045
1046 pscreen->query_dmabuf_modifiers = fd_screen_query_dmabuf_modifiers;
1047
1048 pscreen->get_device_uuid = fd_screen_get_device_uuid;
1049 pscreen->get_driver_uuid = fd_screen_get_driver_uuid;
1050
1051 slab_create_parent(&screen->transfer_pool, sizeof(struct fd_transfer), 16);
1052
1053 return pscreen;
1054
1055 fail:
1056 fd_screen_destroy(pscreen);
1057 return NULL;
1058 }