1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
4 * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
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:
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
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
26 * Rob Clark <robclark@freedesktop.org>
30 #include "pipe/p_defines.h"
31 #include "pipe/p_screen.h"
32 #include "pipe/p_state.h"
34 #include "util/u_memory.h"
35 #include "util/u_inlines.h"
36 #include "util/u_format.h"
37 #include "util/u_format_s3tc.h"
38 #include "util/u_string.h"
39 #include "util/u_debug.h"
41 #include "util/os_time.h"
46 #include <sys/sysinfo.h>
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"
54 #include "a2xx/fd2_screen.h"
55 #include "a3xx/fd3_screen.h"
56 #include "a4xx/fd4_screen.h"
57 #include "a5xx/fd5_screen.h"
59 #include "ir3/ir3_nir.h"
61 /* XXX this should go away */
62 #include "state_tracker/drm_driver.h"
64 static const struct debug_named_value debug_options
[] = {
65 {"msgs", FD_DBG_MSGS
, "Print debug messages"},
66 {"disasm", FD_DBG_DISASM
, "Dump TGSI and adreno shader disassembly"},
67 {"dclear", FD_DBG_DCLEAR
, "Mark all state dirty after clear"},
68 {"ddraw", FD_DBG_DDRAW
, "Mark all state dirty after draw"},
69 {"noscis", FD_DBG_NOSCIS
, "Disable scissor optimization"},
70 {"direct", FD_DBG_DIRECT
, "Force inline (SS_DIRECT) state loads"},
71 {"nobypass", FD_DBG_NOBYPASS
, "Disable GMEM bypass"},
72 {"fraghalf", FD_DBG_FRAGHALF
, "Use half-precision in fragment shader"},
73 {"nobin", FD_DBG_NOBIN
, "Disable hw binning"},
74 {"optmsgs", FD_DBG_OPTMSGS
,"Enable optimizer debug messages"},
75 {"glsl120", FD_DBG_GLSL120
,"Temporary flag to force GLSL 1.20 (rather than 1.30) on a3xx+"},
76 {"shaderdb", FD_DBG_SHADERDB
, "Enable shaderdb output"},
77 {"flush", FD_DBG_FLUSH
, "Force flush after every draw"},
78 {"deqp", FD_DBG_DEQP
, "Enable dEQP hacks"},
79 {"inorder", FD_DBG_INORDER
,"Disable reordering for draws/blits"},
80 {"bstat", FD_DBG_BSTAT
, "Print batch stats at context destroy"},
81 {"nogrow", FD_DBG_NOGROW
, "Disable \"growable\" cmdstream buffers, even if kernel supports it"},
82 {"lrz", FD_DBG_LRZ
, "Enable experimental LRZ support (a5xx+)"},
83 {"noindirect",FD_DBG_NOINDR
, "Disable hw indirect draws (emulate on CPU)"},
84 {"noblit", FD_DBG_NOBLIT
, "Disable blitter (fallback to generic blit path)"},
85 {"hiprio", FD_DBG_HIPRIO
, "Force high-priority context"},
86 {"ttile", FD_DBG_TTILE
, "Enable texture tiling (a5xx)"},
90 DEBUG_GET_ONCE_FLAGS_OPTION(fd_mesa_debug
, "FD_MESA_DEBUG", debug_options
, 0)
92 int fd_mesa_debug
= 0;
93 bool fd_binning_enabled
= true;
94 static bool glsl120
= false;
96 static const struct debug_named_value shader_debug_options
[] = {
97 {"vs", FD_DBG_SHADER_VS
, "Print shader disasm for vertex shaders"},
98 {"fs", FD_DBG_SHADER_FS
, "Print shader disasm for fragment shaders"},
99 {"cs", FD_DBG_SHADER_CS
, "Print shader disasm for compute shaders"},
100 DEBUG_NAMED_VALUE_END
103 DEBUG_GET_ONCE_FLAGS_OPTION(fd_shader_debug
, "FD_SHADER_DEBUG", shader_debug_options
, 0)
105 enum fd_shader_debug fd_shader_debug
= 0;
108 fd_screen_get_name(struct pipe_screen
*pscreen
)
110 static char buffer
[128];
111 util_snprintf(buffer
, sizeof(buffer
), "FD%03d",
112 fd_screen(pscreen
)->device_id
);
117 fd_screen_get_vendor(struct pipe_screen
*pscreen
)
123 fd_screen_get_device_vendor(struct pipe_screen
*pscreen
)
130 fd_screen_get_timestamp(struct pipe_screen
*pscreen
)
132 struct fd_screen
*screen
= fd_screen(pscreen
);
134 if (screen
->has_timestamp
) {
136 fd_pipe_get_param(screen
->pipe
, FD_TIMESTAMP
, &n
);
137 debug_assert(screen
->max_freq
> 0);
138 return n
* 1000000000 / screen
->max_freq
;
140 int64_t cpu_time
= os_time_get() * 1000;
141 return cpu_time
+ screen
->cpu_gpu_time_delta
;
147 fd_screen_destroy(struct pipe_screen
*pscreen
)
149 struct fd_screen
*screen
= fd_screen(pscreen
);
152 fd_pipe_del(screen
->pipe
);
155 fd_device_del(screen
->dev
);
157 fd_bc_fini(&screen
->batch_cache
);
159 slab_destroy_parent(&screen
->transfer_pool
);
161 mtx_destroy(&screen
->lock
);
163 ralloc_free(screen
->compiler
);
165 free(screen
->perfcntr_queries
);
170 TODO either move caps to a2xx/a3xx specific code, or maybe have some
171 tables for things that differ if the delta is not too much..
174 fd_screen_get_param(struct pipe_screen
*pscreen
, enum pipe_cap param
)
176 struct fd_screen
*screen
= fd_screen(pscreen
);
178 /* this is probably not totally correct.. but it's a start: */
180 /* Supported features (boolean caps). */
181 case PIPE_CAP_NPOT_TEXTURES
:
182 case PIPE_CAP_MIXED_FRAMEBUFFER_SIZES
:
183 case PIPE_CAP_ANISOTROPIC_FILTER
:
184 case PIPE_CAP_POINT_SPRITE
:
185 case PIPE_CAP_BLEND_EQUATION_SEPARATE
:
186 case PIPE_CAP_TEXTURE_SWIZZLE
:
187 case PIPE_CAP_MIXED_COLORBUFFER_FORMATS
:
188 case PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
:
189 case PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
:
190 case PIPE_CAP_SEAMLESS_CUBE_MAP
:
191 case PIPE_CAP_VERTEX_COLOR_UNCLAMPED
:
192 case PIPE_CAP_QUADS_FOLLOW_PROVOKING_VERTEX_CONVENTION
:
193 case PIPE_CAP_VERTEX_BUFFER_OFFSET_4BYTE_ALIGNED_ONLY
:
194 case PIPE_CAP_VERTEX_BUFFER_STRIDE_4BYTE_ALIGNED_ONLY
:
195 case PIPE_CAP_VERTEX_ELEMENT_SRC_OFFSET_4BYTE_ALIGNED_ONLY
:
196 case PIPE_CAP_BUFFER_MAP_PERSISTENT_COHERENT
:
197 case PIPE_CAP_STRING_MARKER
:
198 case PIPE_CAP_MIXED_COLOR_DEPTH_BITS
:
199 case PIPE_CAP_TEXTURE_BARRIER
:
200 case PIPE_CAP_INVALIDATE_BUFFER
:
203 case PIPE_CAP_VERTEXID_NOBASE
:
204 return is_a3xx(screen
) || is_a4xx(screen
);
206 case PIPE_CAP_COMPUTE
:
207 return has_compute(screen
);
209 case PIPE_CAP_SHADER_STENCIL_EXPORT
:
210 case PIPE_CAP_TGSI_TEXCOORD
:
211 case PIPE_CAP_PREFER_BLIT_BASED_TEXTURE_TRANSFER
:
212 case PIPE_CAP_TEXTURE_MIRROR_CLAMP
:
213 case PIPE_CAP_QUERY_MEMORY_INFO
:
214 case PIPE_CAP_PCI_GROUP
:
215 case PIPE_CAP_PCI_BUS
:
216 case PIPE_CAP_PCI_DEVICE
:
217 case PIPE_CAP_PCI_FUNCTION
:
221 case PIPE_CAP_PRIMITIVE_RESTART
:
222 case PIPE_CAP_TGSI_INSTANCEID
:
223 case PIPE_CAP_VERTEX_ELEMENT_INSTANCE_DIVISOR
:
224 case PIPE_CAP_INDEP_BLEND_ENABLE
:
225 case PIPE_CAP_INDEP_BLEND_FUNC
:
226 case PIPE_CAP_TEXTURE_BUFFER_OBJECTS
:
227 case PIPE_CAP_TEXTURE_HALF_FLOAT_LINEAR
:
228 case PIPE_CAP_CONDITIONAL_RENDER
:
229 case PIPE_CAP_CONDITIONAL_RENDER_INVERTED
:
230 case PIPE_CAP_SEAMLESS_CUBE_MAP_PER_TEXTURE
:
231 case PIPE_CAP_CLIP_HALFZ
:
232 return is_a3xx(screen
) || is_a4xx(screen
) || is_a5xx(screen
);
234 case PIPE_CAP_FAKE_SW_MSAA
:
235 return !fd_screen_get_param(pscreen
, PIPE_CAP_TEXTURE_MULTISAMPLE
);
237 case PIPE_CAP_TEXTURE_MULTISAMPLE
:
238 return is_a5xx(screen
);
240 case PIPE_CAP_DEPTH_CLIP_DISABLE
:
241 return is_a3xx(screen
) || is_a4xx(screen
);
243 case PIPE_CAP_POLYGON_OFFSET_CLAMP
:
244 return is_a5xx(screen
);
246 case PIPE_CAP_BUFFER_SAMPLER_VIEW_RGBA_ONLY
:
248 case PIPE_CAP_TEXTURE_BUFFER_OFFSET_ALIGNMENT
:
249 if (is_a3xx(screen
)) return 16;
250 if (is_a4xx(screen
)) return 32;
251 if (is_a5xx(screen
)) return 32;
253 case PIPE_CAP_MAX_TEXTURE_BUFFER_SIZE
:
254 /* We could possibly emulate more by pretending 2d/rect textures and
255 * splitting high bits of index into 2nd dimension..
257 if (is_a3xx(screen
)) return 8192;
258 if (is_a4xx(screen
)) return 16384;
259 if (is_a5xx(screen
)) return 16384;
262 case PIPE_CAP_TEXTURE_FLOAT_LINEAR
:
263 case PIPE_CAP_CUBE_MAP_ARRAY
:
264 case PIPE_CAP_SAMPLER_VIEW_TARGET
:
265 case PIPE_CAP_TEXTURE_QUERY_LOD
:
266 return is_a4xx(screen
) || is_a5xx(screen
);
268 case PIPE_CAP_START_INSTANCE
:
269 /* Note that a5xx can do this, it just can't (at least with
270 * current firmware) do draw_indirect with base_instance.
271 * Since draw_indirect is needed sooner (gles31 and gl40 vs
272 * gl42), hide base_instance on a5xx. :-/
274 return is_a4xx(screen
);
276 case PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT
:
279 case PIPE_CAP_GLSL_FEATURE_LEVEL
:
280 case PIPE_CAP_GLSL_FEATURE_LEVEL_COMPATIBILITY
:
283 return is_ir3(screen
) ? 140 : 120;
285 case PIPE_CAP_SHADER_BUFFER_OFFSET_ALIGNMENT
:
290 case PIPE_CAP_MAX_TEXTURE_GATHER_COMPONENTS
:
291 if (is_a4xx(screen
) || is_a5xx(screen
))
295 /* Unsupported features. */
296 case PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
:
297 case PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
:
298 case PIPE_CAP_TGSI_CAN_COMPACT_CONSTANTS
:
299 case PIPE_CAP_USER_VERTEX_BUFFERS
:
300 case PIPE_CAP_QUERY_PIPELINE_STATISTICS
:
301 case PIPE_CAP_TEXTURE_BORDER_COLOR_QUIRK
:
302 case PIPE_CAP_TGSI_VS_LAYER_VIEWPORT
:
303 case PIPE_CAP_TEXTURE_GATHER_SM5
:
304 case PIPE_CAP_SAMPLE_SHADING
:
305 case PIPE_CAP_TEXTURE_GATHER_OFFSETS
:
306 case PIPE_CAP_TGSI_VS_WINDOW_SPACE_POSITION
:
307 case PIPE_CAP_MULTI_DRAW_INDIRECT
:
308 case PIPE_CAP_MULTI_DRAW_INDIRECT_PARAMS
:
309 case PIPE_CAP_TGSI_FS_FINE_DERIVATIVE
:
310 case PIPE_CAP_MULTISAMPLE_Z_RESOLVE
:
311 case PIPE_CAP_RESOURCE_FROM_USER_MEMORY
:
312 case PIPE_CAP_DEVICE_RESET_STATUS_QUERY
:
313 case PIPE_CAP_MAX_SHADER_PATCH_VARYINGS
:
314 case PIPE_CAP_DEPTH_BOUNDS_TEST
:
315 case PIPE_CAP_TGSI_TXQS
:
316 /* TODO if we need this, do it in nir/ir3 backend to avoid breaking precompile: */
317 case PIPE_CAP_FORCE_PERSAMPLE_INTERP
:
318 case PIPE_CAP_COPY_BETWEEN_COMPRESSED_AND_PLAIN_FORMATS
:
319 case PIPE_CAP_CLEAR_TEXTURE
:
320 case PIPE_CAP_DRAW_PARAMETERS
:
321 case PIPE_CAP_TGSI_PACK_HALF_FLOAT
:
322 case PIPE_CAP_TGSI_FS_POSITION_IS_SYSVAL
:
323 case PIPE_CAP_TGSI_FS_FACE_IS_INTEGER_SYSVAL
:
324 case PIPE_CAP_GENERATE_MIPMAP
:
325 case PIPE_CAP_SURFACE_REINTERPRET_BLOCKS
:
326 case PIPE_CAP_ROBUST_BUFFER_ACCESS_BEHAVIOR
:
327 case PIPE_CAP_CULL_DISTANCE
:
328 case PIPE_CAP_PRIMITIVE_RESTART_FOR_PATCHES
:
329 case PIPE_CAP_TGSI_VOTE
:
330 case PIPE_CAP_MAX_WINDOW_RECTANGLES
:
331 case PIPE_CAP_POLYGON_OFFSET_UNITS_UNSCALED
:
332 case PIPE_CAP_VIEWPORT_SUBPIXEL_BITS
:
333 case PIPE_CAP_TGSI_ARRAY_COMPONENTS
:
334 case PIPE_CAP_TGSI_CAN_READ_OUTPUTS
:
335 case PIPE_CAP_TGSI_FS_FBFETCH
:
336 case PIPE_CAP_TGSI_MUL_ZERO_WINS
:
337 case PIPE_CAP_DOUBLES
:
339 case PIPE_CAP_INT64_DIVMOD
:
340 case PIPE_CAP_TGSI_TEX_TXF_LZ
:
341 case PIPE_CAP_TGSI_CLOCK
:
342 case PIPE_CAP_POLYGON_MODE_FILL_RECTANGLE
:
343 case PIPE_CAP_SPARSE_BUFFER_PAGE_SIZE
:
344 case PIPE_CAP_TGSI_BALLOT
:
345 case PIPE_CAP_TGSI_TES_LAYER_VIEWPORT
:
346 case PIPE_CAP_CAN_BIND_CONST_BUFFER_AS_VERTEX
:
347 case PIPE_CAP_ALLOW_MAPPED_BUFFERS_DURING_EXECUTION
:
348 case PIPE_CAP_POST_DEPTH_COVERAGE
:
349 case PIPE_CAP_BINDLESS_TEXTURE
:
350 case PIPE_CAP_NIR_SAMPLERS_AS_DEREF
:
351 case PIPE_CAP_QUERY_SO_OVERFLOW
:
352 case PIPE_CAP_MEMOBJ
:
353 case PIPE_CAP_TGSI_ANY_REG_AS_ADDRESS
:
354 case PIPE_CAP_TILE_RASTER_ORDER
:
355 case PIPE_CAP_MAX_COMBINED_SHADER_OUTPUT_RESOURCES
:
356 case PIPE_CAP_SIGNED_VERTEX_BUFFER_OFFSET
:
357 case PIPE_CAP_FENCE_SIGNAL
:
358 case PIPE_CAP_CONSTBUF0_FLAGS
:
359 case PIPE_CAP_PACKED_UNIFORMS
:
360 case PIPE_CAP_CONSERVATIVE_RASTER_POST_SNAP_TRIANGLES
:
361 case PIPE_CAP_CONSERVATIVE_RASTER_POST_SNAP_POINTS_LINES
:
362 case PIPE_CAP_CONSERVATIVE_RASTER_PRE_SNAP_TRIANGLES
:
363 case PIPE_CAP_CONSERVATIVE_RASTER_PRE_SNAP_POINTS_LINES
:
364 case PIPE_CAP_CONSERVATIVE_RASTER_POST_DEPTH_COVERAGE
:
365 case PIPE_CAP_MAX_CONSERVATIVE_RASTER_SUBPIXEL_PRECISION_BIAS
:
366 case PIPE_CAP_PROGRAMMABLE_SAMPLE_LOCATIONS
:
369 case PIPE_CAP_CONTEXT_PRIORITY_MASK
:
370 return screen
->priority_mask
;
372 case PIPE_CAP_DRAW_INDIRECT
:
373 if (is_a4xx(screen
) || is_a5xx(screen
))
377 case PIPE_CAP_FRAMEBUFFER_NO_ATTACHMENT
:
378 if (is_a4xx(screen
) || is_a5xx(screen
))
382 case PIPE_CAP_LOAD_CONSTBUF
:
383 /* name is confusing, but this turns on std430 packing */
388 case PIPE_CAP_MAX_VIEWPORTS
:
391 case PIPE_CAP_SHAREABLE_SHADERS
:
392 case PIPE_CAP_GLSL_OPTIMIZE_CONSERVATIVELY
:
393 /* manage the variants for these ourself, to avoid breaking precompile: */
394 case PIPE_CAP_FRAGMENT_COLOR_CLAMPED
:
395 case PIPE_CAP_VERTEX_COLOR_CLAMPED
:
401 case PIPE_CAP_MAX_STREAM_OUTPUT_BUFFERS
:
403 return PIPE_MAX_SO_BUFFERS
;
405 case PIPE_CAP_STREAM_OUTPUT_PAUSE_RESUME
:
406 case PIPE_CAP_STREAM_OUTPUT_INTERLEAVE_BUFFERS
:
410 case PIPE_CAP_MAX_STREAM_OUTPUT_SEPARATE_COMPONENTS
:
411 case PIPE_CAP_MAX_STREAM_OUTPUT_INTERLEAVED_COMPONENTS
:
413 return 16 * 4; /* should only be shader out limit? */
416 /* Geometry shader output, unsupported. */
417 case PIPE_CAP_MAX_GEOMETRY_OUTPUT_VERTICES
:
418 case PIPE_CAP_MAX_GEOMETRY_TOTAL_OUTPUT_COMPONENTS
:
419 case PIPE_CAP_MAX_VERTEX_STREAMS
:
422 case PIPE_CAP_MAX_VERTEX_ATTRIB_STRIDE
:
426 case PIPE_CAP_MAX_TEXTURE_2D_LEVELS
:
427 case PIPE_CAP_MAX_TEXTURE_CUBE_LEVELS
:
428 return MAX_MIP_LEVELS
;
429 case PIPE_CAP_MAX_TEXTURE_3D_LEVELS
:
432 case PIPE_CAP_MAX_TEXTURE_ARRAY_LAYERS
:
433 return (is_a3xx(screen
) || is_a4xx(screen
) || is_a5xx(screen
)) ? 256 : 0;
435 /* Render targets. */
436 case PIPE_CAP_MAX_RENDER_TARGETS
:
437 return screen
->max_rts
;
438 case PIPE_CAP_MAX_DUAL_SOURCE_RENDER_TARGETS
:
439 return is_a3xx(screen
) ? 1 : 0;
442 case PIPE_CAP_QUERY_BUFFER_OBJECT
:
444 case PIPE_CAP_OCCLUSION_QUERY
:
445 return is_a3xx(screen
) || is_a4xx(screen
) || is_a5xx(screen
);
446 case PIPE_CAP_QUERY_TIMESTAMP
:
447 case PIPE_CAP_QUERY_TIME_ELAPSED
:
448 /* only a4xx, requires new enough kernel so we know max_freq: */
449 return (screen
->max_freq
> 0) && (is_a4xx(screen
) || is_a5xx(screen
));
451 case PIPE_CAP_MIN_TEXTURE_GATHER_OFFSET
:
452 case PIPE_CAP_MIN_TEXEL_OFFSET
:
455 case PIPE_CAP_MAX_TEXTURE_GATHER_OFFSET
:
456 case PIPE_CAP_MAX_TEXEL_OFFSET
:
459 case PIPE_CAP_ENDIANNESS
:
460 return PIPE_ENDIAN_LITTLE
;
462 case PIPE_CAP_MIN_MAP_BUFFER_ALIGNMENT
:
465 case PIPE_CAP_VENDOR_ID
:
467 case PIPE_CAP_DEVICE_ID
:
469 case PIPE_CAP_ACCELERATED
:
471 case PIPE_CAP_VIDEO_MEMORY
:
472 DBG("FINISHME: The value returned is incorrect\n");
476 case PIPE_CAP_NATIVE_FENCE_FD
:
477 return fd_device_version(screen
->dev
) >= FD_VERSION_FENCE_FD
;
479 debug_printf("unknown param %d\n", param
);
484 fd_screen_get_paramf(struct pipe_screen
*pscreen
, enum pipe_capf param
)
487 case PIPE_CAPF_MAX_LINE_WIDTH
:
488 case PIPE_CAPF_MAX_LINE_WIDTH_AA
:
489 /* NOTE: actual value is 127.0f, but this is working around a deqp
490 * bug.. dEQP-GLES3.functional.rasterization.primitives.lines_wide
491 * uses too small of a render target size, and gets confused when
492 * the lines start going offscreen.
494 * See: https://code.google.com/p/android/issues/detail?id=206513
496 if (fd_mesa_debug
& FD_DBG_DEQP
)
499 case PIPE_CAPF_MAX_POINT_WIDTH
:
500 case PIPE_CAPF_MAX_POINT_WIDTH_AA
:
502 case PIPE_CAPF_MAX_TEXTURE_ANISOTROPY
:
504 case PIPE_CAPF_MAX_TEXTURE_LOD_BIAS
:
506 case PIPE_CAPF_MIN_CONSERVATIVE_RASTER_DILATE
:
507 case PIPE_CAPF_MAX_CONSERVATIVE_RASTER_DILATE
:
508 case PIPE_CAPF_CONSERVATIVE_RASTER_DILATE_GRANULARITY
:
511 debug_printf("unknown paramf %d\n", param
);
516 fd_screen_get_shader_param(struct pipe_screen
*pscreen
,
517 enum pipe_shader_type shader
,
518 enum pipe_shader_cap param
)
520 struct fd_screen
*screen
= fd_screen(pscreen
);
524 case PIPE_SHADER_FRAGMENT
:
525 case PIPE_SHADER_VERTEX
:
527 case PIPE_SHADER_COMPUTE
:
528 if (has_compute(screen
))
531 case PIPE_SHADER_GEOMETRY
:
532 /* maye we could emulate.. */
535 DBG("unknown shader type %d", shader
);
539 /* this is probably not totally correct.. but it's a start: */
541 case PIPE_SHADER_CAP_MAX_INSTRUCTIONS
:
542 case PIPE_SHADER_CAP_MAX_ALU_INSTRUCTIONS
:
543 case PIPE_SHADER_CAP_MAX_TEX_INSTRUCTIONS
:
544 case PIPE_SHADER_CAP_MAX_TEX_INDIRECTIONS
:
546 case PIPE_SHADER_CAP_MAX_CONTROL_FLOW_DEPTH
:
548 case PIPE_SHADER_CAP_MAX_INPUTS
:
549 case PIPE_SHADER_CAP_MAX_OUTPUTS
:
551 case PIPE_SHADER_CAP_MAX_TEMPS
:
552 return 64; /* Max native temporaries. */
553 case PIPE_SHADER_CAP_MAX_CONST_BUFFER_SIZE
:
554 /* NOTE: seems to be limit for a3xx is actually 512 but
555 * split between VS and FS. Use lower limit of 256 to
556 * avoid getting into impossible situations:
558 return ((is_a3xx(screen
) || is_a4xx(screen
) || is_a5xx(screen
)) ? 4096 : 64) * sizeof(float[4]);
559 case PIPE_SHADER_CAP_MAX_CONST_BUFFERS
:
560 return is_ir3(screen
) ? 16 : 1;
561 case PIPE_SHADER_CAP_TGSI_CONT_SUPPORTED
:
563 case PIPE_SHADER_CAP_INDIRECT_INPUT_ADDR
:
564 case PIPE_SHADER_CAP_INDIRECT_OUTPUT_ADDR
:
565 /* Technically this should be the same as for TEMP/CONST, since
566 * everything is just normal registers. This is just temporary
567 * hack until load_input/store_output handle arrays in a similar
568 * way as load_var/store_var..
571 case PIPE_SHADER_CAP_INDIRECT_TEMP_ADDR
:
572 case PIPE_SHADER_CAP_INDIRECT_CONST_ADDR
:
573 /* a2xx compiler doesn't handle indirect: */
574 return is_ir3(screen
) ? 1 : 0;
575 case PIPE_SHADER_CAP_SUBROUTINES
:
576 case PIPE_SHADER_CAP_TGSI_DROUND_SUPPORTED
:
577 case PIPE_SHADER_CAP_TGSI_DFRACEXP_DLDEXP_SUPPORTED
:
578 case PIPE_SHADER_CAP_TGSI_LDEXP_SUPPORTED
:
579 case PIPE_SHADER_CAP_TGSI_FMA_SUPPORTED
:
580 case PIPE_SHADER_CAP_TGSI_ANY_INOUT_DECL_RANGE
:
581 case PIPE_SHADER_CAP_MAX_HW_ATOMIC_COUNTERS
:
583 case PIPE_SHADER_CAP_TGSI_SQRT_SUPPORTED
:
585 case PIPE_SHADER_CAP_INTEGERS
:
588 return is_ir3(screen
) ? 1 : 0;
589 case PIPE_SHADER_CAP_INT64_ATOMICS
:
591 case PIPE_SHADER_CAP_FP16
:
593 case PIPE_SHADER_CAP_MAX_TEXTURE_SAMPLERS
:
594 case PIPE_SHADER_CAP_MAX_SAMPLER_VIEWS
:
596 case PIPE_SHADER_CAP_PREFERRED_IR
:
598 return PIPE_SHADER_IR_NIR
;
599 return PIPE_SHADER_IR_TGSI
;
600 case PIPE_SHADER_CAP_SUPPORTED_IRS
:
601 if (is_ir3(screen
)) {
602 return (1 << PIPE_SHADER_IR_NIR
) | (1 << PIPE_SHADER_IR_TGSI
);
604 return (1 << PIPE_SHADER_IR_TGSI
);
607 case PIPE_SHADER_CAP_MAX_UNROLL_ITERATIONS_HINT
:
609 case PIPE_SHADER_CAP_LOWER_IF_THRESHOLD
:
610 case PIPE_SHADER_CAP_TGSI_SKIP_MERGE_REGISTERS
:
611 case PIPE_SHADER_CAP_MAX_HW_ATOMIC_COUNTER_BUFFERS
:
612 case PIPE_SHADER_CAP_SCALAR_ISA
:
614 case PIPE_SHADER_CAP_MAX_SHADER_BUFFERS
:
615 case PIPE_SHADER_CAP_MAX_SHADER_IMAGES
:
616 if (is_a5xx(screen
)) {
617 /* a5xx (and a4xx for that matter) has one state-block
618 * for compute-shader SSBO's and another that is shared
619 * by VS/HS/DS/GS/FS.. so to simplify things for now
620 * just advertise SSBOs for FS and CS. We could possibly
621 * do what blob does, and partition the space for
622 * VS/HS/DS/GS/FS. The blob advertises:
624 * GL_MAX_VERTEX_SHADER_STORAGE_BLOCKS: 4
625 * GL_MAX_GEOMETRY_SHADER_STORAGE_BLOCKS: 4
626 * GL_MAX_TESS_CONTROL_SHADER_STORAGE_BLOCKS: 4
627 * GL_MAX_TESS_EVALUATION_SHADER_STORAGE_BLOCKS: 4
628 * GL_MAX_FRAGMENT_SHADER_STORAGE_BLOCKS: 4
629 * GL_MAX_COMPUTE_SHADER_STORAGE_BLOCKS: 24
630 * GL_MAX_COMBINED_SHADER_STORAGE_BLOCKS: 24
632 * I think that way we could avoid having to patch shaders
633 * for actual SSBO indexes by using a static partitioning.
635 * Note same state block is used for images and buffers,
636 * but images also need texture state for read access
641 case PIPE_SHADER_FRAGMENT
:
642 case PIPE_SHADER_COMPUTE
:
650 debug_printf("unknown shader param %d\n", param
);
654 /* TODO depending on how much the limits differ for a3xx/a4xx, maybe move this
655 * into per-generation backend?
658 fd_get_compute_param(struct pipe_screen
*pscreen
, enum pipe_shader_ir ir_type
,
659 enum pipe_compute_cap param
, void *ret
)
661 struct fd_screen
*screen
= fd_screen(pscreen
);
662 const char * const ir
= "ir3";
664 if (!has_compute(screen
))
667 #define RET(x) do { \
669 memcpy(ret, x, sizeof(x)); \
674 case PIPE_COMPUTE_CAP_ADDRESS_BITS
:
675 // don't expose 64b pointer support yet, until ir3 supports 64b
676 // math, otherwise spir64 target is used and we get 64b pointer
677 // calculations that we can't do yet
678 // if (is_a5xx(screen))
679 // RET((uint32_t []){ 64 });
680 RET((uint32_t []){ 32 });
682 case PIPE_COMPUTE_CAP_IR_TARGET
:
685 return strlen(ir
) * sizeof(char);
687 case PIPE_COMPUTE_CAP_GRID_DIMENSION
:
688 RET((uint64_t []) { 3 });
690 case PIPE_COMPUTE_CAP_MAX_GRID_SIZE
:
691 RET(((uint64_t []) { 65535, 65535, 65535 }));
693 case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE
:
694 RET(((uint64_t []) { 1024, 1024, 64 }));
696 case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK
:
697 RET((uint64_t []) { 1024 });
699 case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE
:
700 RET((uint64_t []) { screen
->ram_size
});
702 case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE
:
703 RET((uint64_t []) { 32768 });
705 case PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE
:
706 case PIPE_COMPUTE_CAP_MAX_INPUT_SIZE
:
707 RET((uint64_t []) { 4096 });
709 case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE
:
710 RET((uint64_t []) { screen
->ram_size
});
712 case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY
:
713 RET((uint32_t []) { screen
->max_freq
/ 1000000 });
715 case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS
:
716 RET((uint32_t []) { 9999 }); // TODO
718 case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED
:
719 RET((uint32_t []) { 1 });
721 case PIPE_COMPUTE_CAP_SUBGROUP_SIZE
:
722 RET((uint32_t []) { 32 }); // TODO
724 case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK
:
725 RET((uint64_t []) { 1024 }); // TODO
732 fd_get_compiler_options(struct pipe_screen
*pscreen
,
733 enum pipe_shader_ir ir
, unsigned shader
)
735 struct fd_screen
*screen
= fd_screen(pscreen
);
738 return ir3_get_compiler_options(screen
->compiler
);
744 fd_screen_bo_get_handle(struct pipe_screen
*pscreen
,
747 struct winsys_handle
*whandle
)
749 whandle
->stride
= stride
;
751 if (whandle
->type
== WINSYS_HANDLE_TYPE_SHARED
) {
752 return fd_bo_get_name(bo
, &whandle
->handle
) == 0;
753 } else if (whandle
->type
== WINSYS_HANDLE_TYPE_KMS
) {
754 whandle
->handle
= fd_bo_handle(bo
);
756 } else if (whandle
->type
== WINSYS_HANDLE_TYPE_FD
) {
757 whandle
->handle
= fd_bo_dmabuf(bo
);
765 fd_screen_bo_from_handle(struct pipe_screen
*pscreen
,
766 struct winsys_handle
*whandle
)
768 struct fd_screen
*screen
= fd_screen(pscreen
);
771 if (whandle
->type
== WINSYS_HANDLE_TYPE_SHARED
) {
772 bo
= fd_bo_from_name(screen
->dev
, whandle
->handle
);
773 } else if (whandle
->type
== WINSYS_HANDLE_TYPE_KMS
) {
774 bo
= fd_bo_from_handle(screen
->dev
, whandle
->handle
, 0);
775 } else if (whandle
->type
== WINSYS_HANDLE_TYPE_FD
) {
776 bo
= fd_bo_from_dmabuf(screen
->dev
, whandle
->handle
);
778 DBG("Attempt to import unsupported handle type %d", whandle
->type
);
783 DBG("ref name 0x%08x failed", whandle
->handle
);
791 fd_screen_create(struct fd_device
*dev
)
793 struct fd_screen
*screen
= CALLOC_STRUCT(fd_screen
);
794 struct pipe_screen
*pscreen
;
797 fd_mesa_debug
= debug_get_option_fd_mesa_debug();
798 fd_shader_debug
= debug_get_option_fd_shader_debug();
800 if (fd_mesa_debug
& FD_DBG_NOBIN
)
801 fd_binning_enabled
= false;
803 glsl120
= !!(fd_mesa_debug
& FD_DBG_GLSL120
);
808 pscreen
= &screen
->base
;
813 // maybe this should be in context?
814 screen
->pipe
= fd_pipe_new(screen
->dev
, FD_PIPE_3D
);
816 DBG("could not create 3d pipe");
820 if (fd_pipe_get_param(screen
->pipe
, FD_GMEM_SIZE
, &val
)) {
821 DBG("could not get GMEM size");
824 screen
->gmemsize_bytes
= val
;
826 if (fd_pipe_get_param(screen
->pipe
, FD_DEVICE_ID
, &val
)) {
827 DBG("could not get device-id");
830 screen
->device_id
= val
;
832 if (fd_pipe_get_param(screen
->pipe
, FD_MAX_FREQ
, &val
)) {
833 DBG("could not get gpu freq");
834 /* this limits what performance related queries are
835 * supported but is not fatal
837 screen
->max_freq
= 0;
839 screen
->max_freq
= val
;
840 if (fd_pipe_get_param(screen
->pipe
, FD_TIMESTAMP
, &val
) == 0)
841 screen
->has_timestamp
= true;
844 if (fd_pipe_get_param(screen
->pipe
, FD_GPU_ID
, &val
)) {
845 DBG("could not get gpu-id");
848 screen
->gpu_id
= val
;
850 if (fd_pipe_get_param(screen
->pipe
, FD_CHIP_ID
, &val
)) {
851 DBG("could not get chip-id");
852 /* older kernels may not have this property: */
853 unsigned core
= screen
->gpu_id
/ 100;
854 unsigned major
= (screen
->gpu_id
% 100) / 10;
855 unsigned minor
= screen
->gpu_id
% 10;
856 unsigned patch
= 0; /* assume the worst */
857 val
= (patch
& 0xff) | ((minor
& 0xff) << 8) |
858 ((major
& 0xff) << 16) | ((core
& 0xff) << 24);
860 screen
->chip_id
= val
;
862 if (fd_pipe_get_param(screen
->pipe
, FD_NR_RINGS
, &val
)) {
863 DBG("could not get # of rings");
864 screen
->priority_mask
= 0;
866 /* # of rings equates to number of unique priority values: */
867 screen
->priority_mask
= (1 << val
) - 1;
872 screen
->ram_size
= si
.totalram
;
875 DBG(" GPU-id: %d", screen
->gpu_id
);
876 DBG(" Chip-id: 0x%08x", screen
->chip_id
);
877 DBG(" GMEM size: 0x%08x", screen
->gmemsize_bytes
);
879 /* explicitly checking for GPU revisions that are known to work. This
880 * may be overly conservative for a3xx, where spoofing the gpu_id with
881 * the blob driver seems to generate identical cmdstream dumps. But
882 * on a2xx, there seem to be small differences between the GPU revs
883 * so it is probably better to actually test first on real hardware
886 * If you have a different adreno version, feel free to add it to one
887 * of the cases below and see what happens. And if it works, please
890 switch (screen
->gpu_id
) {
893 fd2_screen_init(pscreen
);
899 fd3_screen_init(pscreen
);
903 fd4_screen_init(pscreen
);
906 fd5_screen_init(pscreen
);
909 debug_printf("unsupported GPU: a%03d\n", screen
->gpu_id
);
913 if (screen
->gpu_id
>= 500) {
914 screen
->gmem_alignw
= 64;
915 screen
->gmem_alignh
= 32;
916 screen
->num_vsc_pipes
= 16;
918 screen
->gmem_alignw
= 32;
919 screen
->gmem_alignh
= 32;
920 screen
->num_vsc_pipes
= 8;
923 /* NOTE: don't enable reordering on a2xx, since completely untested.
924 * Also, don't enable if we have too old of a kernel to support
925 * growable cmdstream buffers, since memory requirement for cmdstream
926 * buffers would be too much otherwise.
928 if ((screen
->gpu_id
>= 300) && (fd_device_version(dev
) >= FD_VERSION_UNLIMITED_CMDS
))
929 screen
->reorder
= !(fd_mesa_debug
& FD_DBG_INORDER
);
931 fd_bc_init(&screen
->batch_cache
);
933 (void) mtx_init(&screen
->lock
, mtx_plain
);
935 pscreen
->destroy
= fd_screen_destroy
;
936 pscreen
->get_param
= fd_screen_get_param
;
937 pscreen
->get_paramf
= fd_screen_get_paramf
;
938 pscreen
->get_shader_param
= fd_screen_get_shader_param
;
939 pscreen
->get_compute_param
= fd_get_compute_param
;
940 pscreen
->get_compiler_options
= fd_get_compiler_options
;
942 fd_resource_screen_init(pscreen
);
943 fd_query_screen_init(pscreen
);
945 pscreen
->get_name
= fd_screen_get_name
;
946 pscreen
->get_vendor
= fd_screen_get_vendor
;
947 pscreen
->get_device_vendor
= fd_screen_get_device_vendor
;
949 pscreen
->get_timestamp
= fd_screen_get_timestamp
;
951 pscreen
->fence_reference
= fd_fence_ref
;
952 pscreen
->fence_finish
= fd_fence_finish
;
953 pscreen
->fence_get_fd
= fd_fence_get_fd
;
955 slab_create_parent(&screen
->transfer_pool
, sizeof(struct fd_transfer
), 16);
960 fd_screen_destroy(pscreen
);