• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright © 2017 Intel Corporation
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 shall be included
12  * in all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
15  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
19  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
20  * DEALINGS IN THE SOFTWARE.
21  */
22 
23 /**
24  * @file iris_screen.c
25  *
26  * Screen related driver hooks and capability lists.
27  *
28  * A program may use multiple rendering contexts (iris_context), but
29  * they all share a common screen (iris_screen).  Global driver state
30  * can be stored in the screen; it may be accessed by multiple threads.
31  */
32 
33 #include <stdio.h>
34 #include <errno.h>
35 #include <sys/ioctl.h>
36 #include "pipe/p_defines.h"
37 #include "pipe/p_state.h"
38 #include "pipe/p_context.h"
39 #include "pipe/p_screen.h"
40 #include "util/u_debug.h"
41 #include "util/os_file.h"
42 #include "util/u_cpu_detect.h"
43 #include "util/u_inlines.h"
44 #include "util/format/u_format.h"
45 #include "util/u_transfer_helper.h"
46 #include "util/u_upload_mgr.h"
47 #include "util/ralloc.h"
48 #include "util/xmlconfig.h"
49 #include "iris_context.h"
50 #include "iris_defines.h"
51 #include "iris_fence.h"
52 #include "iris_perf.h"
53 #include "iris_pipe.h"
54 #include "iris_resource.h"
55 #include "iris_screen.h"
56 #include "compiler/glsl_types.h"
57 #include "intel/common/intel_debug_identifier.h"
58 #include "intel/common/intel_gem.h"
59 #include "intel/common/intel_l3_config.h"
60 #include "intel/common/intel_uuid.h"
61 #include "iris_monitor.h"
62 
63 #define genX_call(devinfo, func, ...)             \
64    switch ((devinfo)->verx10) {                   \
65    case 300:                                      \
66       gfx30_##func(__VA_ARGS__);                  \
67       break;                                      \
68    case 200:                                      \
69       gfx20_##func(__VA_ARGS__);                  \
70       break;                                      \
71    case 125:                                      \
72       gfx125_##func(__VA_ARGS__);                 \
73       break;                                      \
74    case 120:                                      \
75       gfx12_##func(__VA_ARGS__);                  \
76       break;                                      \
77    case 110:                                      \
78       gfx11_##func(__VA_ARGS__);                  \
79       break;                                      \
80    case 90:                                       \
81       gfx9_##func(__VA_ARGS__);                   \
82       break;                                      \
83    case 80:                                       \
84       gfx8_##func(__VA_ARGS__);                   \
85       break;                                      \
86    default:                                       \
87       unreachable("Unknown hardware generation"); \
88    }
89 
90 static const char *
iris_get_vendor(struct pipe_screen * pscreen)91 iris_get_vendor(struct pipe_screen *pscreen)
92 {
93    return "Intel";
94 }
95 
96 static const char *
iris_get_device_vendor(struct pipe_screen * pscreen)97 iris_get_device_vendor(struct pipe_screen *pscreen)
98 {
99    return "Intel";
100 }
101 
102 static void
iris_get_device_uuid(struct pipe_screen * pscreen,char * uuid)103 iris_get_device_uuid(struct pipe_screen *pscreen, char *uuid)
104 {
105    struct iris_screen *screen = (struct iris_screen *)pscreen;
106 
107    intel_uuid_compute_device_id((uint8_t *)uuid, screen->devinfo, PIPE_UUID_SIZE);
108 }
109 
110 static void
iris_get_driver_uuid(struct pipe_screen * pscreen,char * uuid)111 iris_get_driver_uuid(struct pipe_screen *pscreen, char *uuid)
112 {
113    struct iris_screen *screen = (struct iris_screen *)pscreen;
114    const struct intel_device_info *devinfo = screen->devinfo;
115 
116    intel_uuid_compute_driver_id((uint8_t *)uuid, devinfo, PIPE_UUID_SIZE);
117 }
118 
119 static void
iris_warn_cl()120 iris_warn_cl()
121 {
122    static bool warned = false;
123    if (warned || INTEL_DEBUG(DEBUG_CL_QUIET))
124       return;
125 
126    warned = true;
127    fprintf(stderr, "WARNING: OpenCL support via iris driver is incomplete.\n"
128                    "For a complete and conformant OpenCL implementation, use\n"
129                    "https://github.com/intel/compute-runtime instead\n");
130 }
131 
132 static const char *
iris_get_name(struct pipe_screen * pscreen)133 iris_get_name(struct pipe_screen *pscreen)
134 {
135    struct iris_screen *screen = (struct iris_screen *)pscreen;
136    const struct intel_device_info *devinfo = screen->devinfo;
137    static char buf[128];
138 
139    snprintf(buf, sizeof(buf), "Mesa %s", devinfo->name);
140    return buf;
141 }
142 
143 static const char *
iris_get_cl_cts_version(struct pipe_screen * pscreen)144 iris_get_cl_cts_version(struct pipe_screen *pscreen)
145 {
146    struct iris_screen *screen = (struct iris_screen *)pscreen;
147    const struct intel_device_info *devinfo = screen->devinfo;
148 
149    /* https://www.khronos.org/conformance/adopters/conformant-products/opencl#submission_405 */
150    if (devinfo->verx10 == 120)
151       return "v2022-04-22-00";
152 
153    return NULL;
154 }
155 
156 static int
iris_get_video_memory(struct iris_screen * screen)157 iris_get_video_memory(struct iris_screen *screen)
158 {
159    uint64_t vram = iris_bufmgr_vram_size(screen->bufmgr);
160    uint64_t sram = iris_bufmgr_sram_size(screen->bufmgr);
161    if (vram) {
162       return vram / (1024 * 1024);
163    } else if (sram) {
164       return sram / (1024 * 1024);
165    } else {
166       /* This is the old code path, it get the GGTT size from the kernel
167        * (which should always be 4Gb on Gfx8+).
168        *
169        * We should probably never end up here. This is just a fallback to get
170        * some kind of value in case os_get_available_system_memory fails.
171        */
172       const struct intel_device_info *devinfo = screen->devinfo;
173       /* Once a batch uses more than 75% of the maximum mappable size, we
174        * assume that there's some fragmentation, and we start doing extra
175        * flushing, etc.  That's the big cliff apps will care about.
176        */
177       const unsigned gpu_mappable_megabytes =
178          (devinfo->aperture_bytes * 3 / 4) / (1024 * 1024);
179 
180       const long system_memory_pages = sysconf(_SC_PHYS_PAGES);
181       const long system_page_size = sysconf(_SC_PAGE_SIZE);
182 
183       if (system_memory_pages <= 0 || system_page_size <= 0)
184          return -1;
185 
186       const uint64_t system_memory_bytes =
187          (uint64_t) system_memory_pages * (uint64_t) system_page_size;
188 
189       const unsigned system_memory_megabytes =
190          (unsigned) (system_memory_bytes / (1024 * 1024));
191 
192       return MIN2(system_memory_megabytes, gpu_mappable_megabytes);
193    }
194 }
195 
196 static int
iris_get_shader_param(struct pipe_screen * pscreen,enum pipe_shader_type p_stage,enum pipe_shader_cap param)197 iris_get_shader_param(struct pipe_screen *pscreen,
198                       enum pipe_shader_type p_stage,
199                       enum pipe_shader_cap param)
200 {
201    gl_shader_stage stage = stage_from_pipe(p_stage);
202 
203    if (p_stage == PIPE_SHADER_MESH ||
204        p_stage == PIPE_SHADER_TASK)
205       return 0;
206 
207    /* this is probably not totally correct.. but it's a start: */
208    switch (param) {
209    case PIPE_SHADER_CAP_MAX_INSTRUCTIONS:
210       return stage == MESA_SHADER_FRAGMENT ? 1024 : 16384;
211    case PIPE_SHADER_CAP_MAX_ALU_INSTRUCTIONS:
212    case PIPE_SHADER_CAP_MAX_TEX_INSTRUCTIONS:
213    case PIPE_SHADER_CAP_MAX_TEX_INDIRECTIONS:
214       return stage == MESA_SHADER_FRAGMENT ? 1024 : 0;
215 
216    case PIPE_SHADER_CAP_MAX_CONTROL_FLOW_DEPTH:
217       return UINT_MAX;
218 
219    case PIPE_SHADER_CAP_MAX_INPUTS:
220       return stage == MESA_SHADER_VERTEX ? 16 : 32;
221    case PIPE_SHADER_CAP_MAX_OUTPUTS:
222       return 32;
223    case PIPE_SHADER_CAP_MAX_CONST_BUFFER0_SIZE:
224       return 16 * 1024 * sizeof(float);
225    case PIPE_SHADER_CAP_MAX_CONST_BUFFERS:
226       return 16;
227    case PIPE_SHADER_CAP_MAX_TEMPS:
228       return 256; /* GL_MAX_PROGRAM_TEMPORARIES_ARB */
229    case PIPE_SHADER_CAP_CONT_SUPPORTED:
230       return 0;
231    case PIPE_SHADER_CAP_INDIRECT_TEMP_ADDR:
232    case PIPE_SHADER_CAP_INDIRECT_CONST_ADDR:
233       /* Lie about these to avoid st/mesa's GLSL IR lowering of indirects,
234        * which we don't want.  Our compiler backend will check brw_compiler's
235        * options and call nir_lower_indirect_derefs appropriately anyway.
236        */
237       return true;
238    case PIPE_SHADER_CAP_SUBROUTINES:
239       return 0;
240    case PIPE_SHADER_CAP_INTEGERS:
241       return 1;
242    case PIPE_SHADER_CAP_INT64_ATOMICS:
243    case PIPE_SHADER_CAP_FP16:
244    case PIPE_SHADER_CAP_FP16_DERIVATIVES:
245    case PIPE_SHADER_CAP_FP16_CONST_BUFFERS:
246    case PIPE_SHADER_CAP_INT16:
247    case PIPE_SHADER_CAP_GLSL_16BIT_CONSTS:
248       return 0;
249    case PIPE_SHADER_CAP_MAX_TEXTURE_SAMPLERS:
250       return IRIS_MAX_SAMPLERS;
251    case PIPE_SHADER_CAP_MAX_SAMPLER_VIEWS:
252       return IRIS_MAX_TEXTURES;
253    case PIPE_SHADER_CAP_MAX_SHADER_IMAGES:
254       return IRIS_MAX_IMAGES;
255    case PIPE_SHADER_CAP_MAX_SHADER_BUFFERS:
256       return IRIS_MAX_ABOS + IRIS_MAX_SSBOS;
257    case PIPE_SHADER_CAP_MAX_HW_ATOMIC_COUNTERS:
258    case PIPE_SHADER_CAP_MAX_HW_ATOMIC_COUNTER_BUFFERS:
259       return 0;
260    case PIPE_SHADER_CAP_SUPPORTED_IRS:
261       return 1 << PIPE_SHADER_IR_NIR;
262    case PIPE_SHADER_CAP_TGSI_ANY_INOUT_DECL_RANGE:
263    case PIPE_SHADER_CAP_TGSI_SQRT_SUPPORTED:
264       return 0;
265    default:
266       unreachable("unknown shader param");
267    }
268 }
269 
270 static int
iris_get_compute_param(struct pipe_screen * pscreen,enum pipe_shader_ir ir_type,enum pipe_compute_cap param,void * ret)271 iris_get_compute_param(struct pipe_screen *pscreen,
272                        enum pipe_shader_ir ir_type,
273                        enum pipe_compute_cap param,
274                        void *ret)
275 {
276    struct iris_screen *screen = (struct iris_screen *)pscreen;
277    const struct intel_device_info *devinfo = screen->devinfo;
278 
279    const uint32_t max_invocations =
280       MIN2(1024, 32 * devinfo->max_cs_workgroup_threads);
281 
282 #define RET(x) do {                  \
283    if (ret)                          \
284       memcpy(ret, x, sizeof(x));     \
285    return sizeof(x);                 \
286 } while (0)
287 
288    switch (param) {
289    case PIPE_COMPUTE_CAP_ADDRESS_BITS:
290       /* This gets queried on OpenCL device init and is never queried by the
291        * OpenGL state tracker.
292        */
293       iris_warn_cl();
294       RET((uint32_t []){ 64 });
295 
296    case PIPE_COMPUTE_CAP_IR_TARGET:
297       if (ret)
298          strcpy(ret, "gen");
299       return 4;
300 
301    case PIPE_COMPUTE_CAP_GRID_DIMENSION:
302       RET((uint64_t []) { 3 });
303 
304    case PIPE_COMPUTE_CAP_MAX_GRID_SIZE:
305       RET(((uint64_t []) { UINT32_MAX, UINT32_MAX, UINT32_MAX }));
306 
307    case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE:
308       /* MaxComputeWorkGroupSize[0..2] */
309       RET(((uint64_t []) {max_invocations, max_invocations, max_invocations}));
310 
311    case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK:
312       /* MaxComputeWorkGroupInvocations */
313    case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK:
314       /* MaxComputeVariableGroupInvocations */
315       RET((uint64_t []) { max_invocations });
316 
317    case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE:
318       /* MaxComputeSharedMemorySize */
319       RET((uint64_t []) { 64 * 1024 });
320 
321    case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED:
322       RET((uint32_t []) { 1 });
323 
324    case PIPE_COMPUTE_CAP_SUBGROUP_SIZES:
325       RET((uint32_t []) { 32 | 16 | 8 });
326 
327    case PIPE_COMPUTE_CAP_MAX_SUBGROUPS:
328       RET((uint32_t []) { devinfo->max_cs_workgroup_threads });
329 
330    case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE:
331    case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE:
332       RET((uint64_t []) { 1 << 30 }); /* TODO */
333 
334    case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY:
335       RET((uint32_t []) { 400 }); /* TODO */
336 
337    case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS: {
338       RET((uint32_t []) { intel_device_info_subslice_total(devinfo) });
339    }
340 
341    case PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE:
342       /* MaxComputeSharedMemorySize */
343       RET((uint64_t []) { 64 * 1024 });
344 
345    case PIPE_COMPUTE_CAP_MAX_INPUT_SIZE:
346       /* We could probably allow more; this is the OpenCL minimum */
347       RET((uint64_t []) { 1024 });
348 
349    default:
350       unreachable("unknown compute param");
351    }
352 }
353 
354 static void
iris_init_screen_caps(struct iris_screen * screen)355 iris_init_screen_caps(struct iris_screen *screen)
356 {
357    struct pipe_caps *caps = (struct pipe_caps *)&screen->base.caps;
358 
359    u_init_pipe_screen_caps(&screen->base, 1);
360 
361    const struct intel_device_info *devinfo = screen->devinfo;
362 
363    caps->npot_textures = true;
364    caps->anisotropic_filter = true;
365    caps->occlusion_query = true;
366    caps->query_time_elapsed = true;
367    caps->texture_swizzle = true;
368    caps->texture_mirror_clamp_to_edge = true;
369    caps->blend_equation_separate = true;
370    caps->fragment_shader_texture_lod = true;
371    caps->fragment_shader_derivatives = true;
372    caps->primitive_restart = true;
373    caps->primitive_restart_fixed_index = true;
374    caps->indep_blend_enable = true;
375    caps->indep_blend_func = true;
376    caps->fs_coord_origin_upper_left = true;
377    caps->fs_coord_pixel_center_integer = true;
378    caps->depth_clip_disable = true;
379    caps->vs_instanceid = true;
380    caps->vertex_element_instance_divisor = true;
381    caps->seamless_cube_map = true;
382    caps->seamless_cube_map_per_texture = true;
383    caps->conditional_render = true;
384    caps->texture_barrier = true;
385    caps->stream_output_pause_resume = true;
386    caps->vertex_color_unclamped = true;
387    caps->compute = true;
388    caps->start_instance = true;
389    caps->query_timestamp = true;
390    caps->texture_multisample = true;
391    caps->cube_map_array = true;
392    caps->texture_buffer_objects = true;
393    caps->query_pipeline_statistics_single = true;
394    caps->texture_query_lod = true;
395    caps->sample_shading = true;
396    caps->force_persample_interp = true;
397    caps->draw_indirect = true;
398    caps->multi_draw_indirect = true;
399    caps->multi_draw_indirect_params = true;
400    caps->mixed_framebuffer_sizes = true;
401    caps->vs_layer_viewport = true;
402    caps->tes_layer_viewport = true;
403    caps->fs_fine_derivative = true;
404    caps->shader_pack_half_float = true;
405    caps->conditional_render_inverted = true;
406    caps->clip_halfz = true;
407    caps->tgsi_texcoord = true;
408    caps->stream_output_interleave_buffers = true;
409    caps->doubles = true;
410    caps->int64 = true;
411    caps->sampler_view_target = true;
412    caps->robust_buffer_access_behavior = true;
413    caps->device_reset_status_query = true;
414    caps->copy_between_compressed_and_plain_formats = true;
415    caps->framebuffer_no_attachment = true;
416    caps->cull_distance = true;
417    caps->packed_uniforms = true;
418    caps->signed_vertex_buffer_offset = true;
419    caps->texture_float_linear = true;
420    caps->texture_half_float_linear = true;
421    caps->polygon_offset_clamp = true;
422    caps->query_so_overflow = true;
423    caps->query_buffer_object = true;
424    caps->tgsi_tex_txf_lz = true;
425    caps->texture_query_samples = true;
426    caps->shader_clock = true;
427    caps->shader_ballot = true;
428    caps->multisample_z_resolve = true;
429    caps->clear_scissored = true;
430    caps->shader_group_vote = true;
431    caps->vs_window_space_position = true;
432    caps->texture_gather_sm5 = true;
433    caps->shader_array_components = true;
434    caps->glsl_tess_levels_as_inputs = true;
435    caps->load_constbuf = true;
436    caps->draw_parameters = true;
437    caps->fs_position_is_sysval = true;
438    caps->fs_face_is_integer_sysval = true;
439    caps->compute_shader_derivatives = true;
440    caps->invalidate_buffer = true;
441    caps->surface_reinterpret_blocks = true;
442    caps->texture_shadow_lod = true;
443    caps->shader_samples_identical = true;
444    caps->gl_spirv = true;
445    caps->gl_spirv_variable_pointers = true;
446    caps->demote_to_helper_invocation = true;
447    caps->native_fence_fd = true;
448    caps->memobj = true;
449    caps->mixed_color_depth_bits = true;
450    caps->fence_signal = true;
451    caps->image_store_formatted = true;
452    caps->legacy_math_rules = true;
453    caps->alpha_to_coverage_dither_control = true;
454    caps->map_unsynchronized_thread_safe = true;
455    caps->has_const_bw = true;
456    caps->cl_gl_sharing = true;
457    caps->uma = iris_bufmgr_vram_size(screen->bufmgr) == 0;
458    caps->query_memory_info = iris_bufmgr_vram_size(screen->bufmgr) != 0;
459    caps->prefer_back_buffer_reuse = false;
460    caps->fbfetch = IRIS_MAX_DRAW_BUFFERS;
461    caps->fbfetch_coherent = devinfo->ver >= 9 && devinfo->ver < 20;
462    caps->conservative_raster_inner_coverage =
463    caps->post_depth_coverage =
464    caps->shader_stencil_export =
465    caps->depth_clip_disable_separate =
466    caps->fragment_shader_interlock =
467    caps->atomic_float_minmax = devinfo->ver >= 9;
468    caps->depth_bounds_test = devinfo->ver >= 12;
469    caps->max_dual_source_render_targets = 1;
470    caps->max_render_targets = IRIS_MAX_DRAW_BUFFERS;
471    caps->max_texture_2d_size = 16384;
472    caps->max_texture_cube_levels = IRIS_MAX_MIPLEVELS; /* 16384x16384 */
473    caps->max_texture_3d_levels = 12; /* 2048x2048 */
474    caps->max_stream_output_buffers = 4;
475    caps->max_texture_array_layers = 2048;
476    caps->max_stream_output_separate_components =
477       IRIS_MAX_SOL_BINDINGS / IRIS_MAX_SOL_BUFFERS;
478    caps->max_stream_output_interleaved_components = IRIS_MAX_SOL_BINDINGS;
479    caps->glsl_feature_level =
480    caps->glsl_feature_level_compatibility = 460;
481    /* 3DSTATE_CONSTANT_XS requires the start of UBOs to be 32B aligned */
482    caps->constant_buffer_offset_alignment = 32;
483    caps->min_map_buffer_alignment = IRIS_MAP_BUFFER_ALIGNMENT;
484    caps->shader_buffer_offset_alignment = 4;
485    caps->max_shader_buffer_size = 1 << 27;
486    caps->texture_buffer_offset_alignment = 16; // XXX: u_screen says 256 is the minimum value...
487    caps->linear_image_pitch_alignment = 1;
488    caps->linear_image_base_address_alignment = 1;
489    caps->texture_transfer_modes = PIPE_TEXTURE_TRANSFER_BLIT;
490    caps->max_texel_buffer_elements = IRIS_MAX_TEXTURE_BUFFER_SIZE;
491    caps->max_viewports = 16;
492    caps->max_geometry_output_vertices = 256;
493    caps->max_geometry_total_output_components = 1024;
494    caps->max_gs_invocations = 32;
495    caps->max_texture_gather_components = 4;
496    caps->min_texture_gather_offset = -32;
497    caps->max_texture_gather_offset = 31;
498    caps->max_vertex_streams = 4;
499    caps->vendor_id = 0x8086;
500    caps->device_id = screen->devinfo->pci_device_id;
501    caps->video_memory = iris_get_video_memory(screen);
502    caps->max_shader_patch_varyings =
503    caps->max_varyings = 32;
504    /* We want immediate arrays to go get uploaded as nir->constant_data by
505     * nir_opt_large_constants() instead.
506     */
507    caps->prefer_imm_arrays_as_constbuf = false;
508    /* AMD_pinned_memory assumes the flexibility of using client memory
509     * for any buffer (incl. vertex buffers) which rules out the prospect
510     * of using snooped buffers, as using snooped buffers without
511     * cogniscience is likely to be detrimental to performance and require
512     * extensive checking in the driver for correctness, e.g. to prevent
513     * illegal snoop <-> snoop transfers.
514     */
515    caps->resource_from_user_memory = devinfo->has_llc;
516    caps->throttle = !screen->driconf.disable_throttling;
517 
518    caps->context_priority_mask =
519       PIPE_CONTEXT_PRIORITY_LOW |
520       PIPE_CONTEXT_PRIORITY_MEDIUM |
521       PIPE_CONTEXT_PRIORITY_HIGH;
522 
523    caps->frontend_noop = true;
524 
525    // XXX: don't hardcode 00:00:02.0 PCI here
526    caps->pci_group = 0;
527    caps->pci_bus = 0;
528    caps->pci_device = 2;
529    caps->pci_function = 0;
530 
531    caps->opencl_integer_functions =
532    caps->integer_multiply_32x16 = true;
533 
534    /* Internal details of VF cache make this optimization harmful on GFX
535     * version 8 and 9, because generated VERTEX_BUFFER_STATEs are cached
536     * separately.
537     */
538    caps->allow_dynamic_vao_fastpath = devinfo->ver >= 11;
539 
540    caps->timer_resolution = DIV_ROUND_UP(1000000000ull, devinfo->timestamp_frequency);
541 
542    caps->device_protected_context =
543       screen->kernel_features & KERNEL_HAS_PROTECTED_CONTEXT;
544 
545    caps->astc_void_extents_need_denorm_flush =
546       devinfo->ver == 9 && !intel_device_info_is_9lp(devinfo);
547 
548    caps->min_line_width =
549    caps->min_line_width_aa =
550    caps->min_point_size =
551    caps->min_point_size_aa = 1;
552 
553    caps->point_size_granularity =
554    caps->line_width_granularity = 0.1;
555 
556    caps->max_line_width =
557    caps->max_line_width_aa = 7.375f;
558 
559    caps->max_point_size =
560    caps->max_point_size_aa = 255.0f;
561 
562    caps->max_texture_anisotropy = 16.0f;
563    caps->max_texture_lod_bias = 15.0f;
564 }
565 
566 static uint64_t
iris_get_timestamp(struct pipe_screen * pscreen)567 iris_get_timestamp(struct pipe_screen *pscreen)
568 {
569    struct iris_screen *screen = (struct iris_screen *) pscreen;
570    uint64_t result;
571 
572    if (!intel_gem_read_render_timestamp(iris_bufmgr_get_fd(screen->bufmgr),
573                                         screen->devinfo->kmd_type, &result))
574       return 0;
575 
576    result = intel_device_info_timebase_scale(screen->devinfo, result);
577 
578    return result;
579 }
580 
581 void
iris_screen_destroy(struct iris_screen * screen)582 iris_screen_destroy(struct iris_screen *screen)
583 {
584    intel_perf_free(screen->perf_cfg);
585    iris_destroy_screen_measure(screen);
586    util_queue_destroy(&screen->shader_compiler_queue);
587    glsl_type_singleton_decref();
588    iris_bo_unreference(screen->workaround_bo);
589    iris_bo_unreference(screen->breakpoint_bo);
590    u_transfer_helper_destroy(screen->base.transfer_helper);
591    iris_bufmgr_unref(screen->bufmgr);
592    disk_cache_destroy(screen->disk_cache);
593    close(screen->winsys_fd);
594    ralloc_free(screen);
595 }
596 
597 static void
iris_screen_unref(struct pipe_screen * pscreen)598 iris_screen_unref(struct pipe_screen *pscreen)
599 {
600    iris_pscreen_unref(pscreen);
601 }
602 
603 static void
iris_query_memory_info(struct pipe_screen * pscreen,struct pipe_memory_info * info)604 iris_query_memory_info(struct pipe_screen *pscreen,
605                        struct pipe_memory_info *info)
606 {
607    struct iris_screen *screen = (struct iris_screen *)pscreen;
608    struct intel_device_info di;
609    memcpy(&di, screen->devinfo, sizeof(di));
610 
611    if (!intel_device_info_update_memory_info(&di, screen->fd))
612       return;
613 
614    info->total_device_memory =
615       (di.mem.vram.mappable.size + di.mem.vram.unmappable.size) / 1024;
616    info->avail_device_memory =
617       (di.mem.vram.mappable.free + di.mem.vram.unmappable.free) / 1024;
618    info->total_staging_memory = di.mem.sram.mappable.size / 1024;
619    info->avail_staging_memory = di.mem.sram.mappable.free / 1024;
620 
621    /* Neither kernel gives us any way to calculate this information */
622    info->device_memory_evicted = 0;
623    info->nr_device_memory_evictions = 0;
624 }
625 
626 static struct disk_cache *
iris_get_disk_shader_cache(struct pipe_screen * pscreen)627 iris_get_disk_shader_cache(struct pipe_screen *pscreen)
628 {
629    struct iris_screen *screen = (struct iris_screen *) pscreen;
630    return screen->disk_cache;
631 }
632 
633 static const struct intel_l3_config *
iris_get_default_l3_config(const struct intel_device_info * devinfo,bool compute)634 iris_get_default_l3_config(const struct intel_device_info *devinfo,
635                            bool compute)
636 {
637    bool wants_dc_cache = true;
638    bool has_slm = compute;
639    const struct intel_l3_weights w =
640       intel_get_default_l3_weights(devinfo, wants_dc_cache, has_slm);
641    return intel_get_l3_config(devinfo, w);
642 }
643 
644 static void
iris_detect_kernel_features(struct iris_screen * screen)645 iris_detect_kernel_features(struct iris_screen *screen)
646 {
647    const struct intel_device_info *devinfo = screen->devinfo;
648    /* Kernel 5.2+ */
649    if (intel_gem_supports_syncobj_wait(screen->fd))
650       screen->kernel_features |= KERNEL_HAS_WAIT_FOR_SUBMIT;
651    if (intel_gem_supports_protected_context(screen->fd, devinfo->kmd_type))
652       screen->kernel_features |= KERNEL_HAS_PROTECTED_CONTEXT;
653 }
654 
655 static bool
iris_init_identifier_bo(struct iris_screen * screen)656 iris_init_identifier_bo(struct iris_screen *screen)
657 {
658    void *bo_map;
659 
660    bo_map = iris_bo_map(NULL, screen->workaround_bo, MAP_READ | MAP_WRITE);
661    if (!bo_map)
662       return false;
663 
664    assert(iris_bo_is_real(screen->workaround_bo));
665 
666    screen->workaround_address = (struct iris_address) {
667       .bo = screen->workaround_bo,
668       .offset = ALIGN(
669          intel_debug_write_identifiers(bo_map, 4096, "Iris"), 32),
670    };
671 
672    iris_bo_unmap(screen->workaround_bo);
673 
674    return true;
675 }
676 
677 static int
iris_screen_get_fd(struct pipe_screen * pscreen)678 iris_screen_get_fd(struct pipe_screen *pscreen)
679 {
680    struct iris_screen *screen = (struct iris_screen *) pscreen;
681 
682    return screen->winsys_fd;
683 }
684 
685 static void
iris_set_damage_region(struct pipe_screen * pscreen,struct pipe_resource * pres,unsigned int nrects,const struct pipe_box * rects)686 iris_set_damage_region(struct pipe_screen *pscreen, struct pipe_resource *pres,
687                        unsigned int nrects, const struct pipe_box *rects)
688 {
689    struct iris_resource *res = (struct iris_resource *)pres;
690 
691    res->use_damage = nrects > 0;
692    if (!res->use_damage)
693       return;
694 
695    res->damage.x = INT32_MAX;
696    res->damage.y = INT32_MAX;
697    res->damage.width = 0;
698    res->damage.height = 0;
699 
700    for (unsigned i = 0; i < nrects; i++) {
701       res->damage.x = MIN2(res->damage.x, rects[i].x);
702       res->damage.y = MIN2(res->damage.y, rects[i].y);
703       res->damage.width = MAX2(res->damage.width, rects[i].width + rects[i].x);
704       res->damage.height = MAX2(res->damage.height, rects[i].height + rects[i].y);
705 
706       if (unlikely(res->damage.x == 0 &&
707                    res->damage.y == 0 &&
708                    res->damage.width == res->base.b.width0 &&
709                    res->damage.height == res->base.b.height0))
710          break;
711    }
712 
713    res->damage.x = MAX2(res->damage.x, 0);
714    res->damage.y = MAX2(res->damage.y, 0);
715    res->damage.width = MIN2(res->damage.width, res->base.b.width0);
716    res->damage.height = MIN2(res->damage.height, res->base.b.height0);
717 }
718 
719 struct pipe_screen *
iris_screen_create(int fd,const struct pipe_screen_config * config)720 iris_screen_create(int fd, const struct pipe_screen_config *config)
721 {
722    struct iris_screen *screen = rzalloc(NULL, struct iris_screen);
723    if (!screen)
724       return NULL;
725 
726    driParseConfigFiles(config->options, config->options_info, 0, "iris",
727                        NULL, NULL, NULL, 0, NULL, 0);
728 
729    bool bo_reuse = false;
730    int bo_reuse_mode = driQueryOptioni(config->options, "bo_reuse");
731    switch (bo_reuse_mode) {
732    case DRI_CONF_BO_REUSE_DISABLED:
733       break;
734    case DRI_CONF_BO_REUSE_ALL:
735       bo_reuse = true;
736       break;
737    }
738 
739    process_intel_debug_variable();
740 
741    screen->bufmgr = iris_bufmgr_get_for_fd(fd, bo_reuse);
742    if (!screen->bufmgr)
743       return NULL;
744 
745    screen->devinfo = iris_bufmgr_get_device_info(screen->bufmgr);
746    p_atomic_set(&screen->refcount, 1);
747 
748    /* Here are the i915 features we need for Iris (in chronological order) :
749     *    - I915_PARAM_HAS_EXEC_NO_RELOC     (3.10)
750     *    - I915_PARAM_HAS_EXEC_HANDLE_LUT   (3.10)
751     *    - I915_PARAM_HAS_EXEC_BATCH_FIRST  (4.13)
752     *    - I915_PARAM_HAS_EXEC_FENCE_ARRAY  (4.14)
753     *    - I915_PARAM_HAS_CONTEXT_ISOLATION (4.16)
754     *
755     * Checking the last feature availability will include all previous ones.
756     */
757    if (!screen->devinfo->has_context_isolation) {
758       debug_error("Kernel is too old (4.16+ required) or unusable for Iris.\n"
759                   "Check your dmesg logs for loading failures.\n");
760       return NULL;
761    }
762 
763    screen->fd = iris_bufmgr_get_fd(screen->bufmgr);
764    screen->winsys_fd = os_dupfd_cloexec(fd);
765 
766    screen->id = iris_bufmgr_create_screen_id(screen->bufmgr);
767 
768    screen->workaround_bo =
769       iris_bo_alloc(screen->bufmgr, "workaround", 4096, 4096,
770                     IRIS_MEMZONE_OTHER, BO_ALLOC_NO_SUBALLOC | BO_ALLOC_CAPTURE);
771    if (!screen->workaround_bo)
772       return NULL;
773 
774    screen->breakpoint_bo = iris_bo_alloc(screen->bufmgr, "breakpoint", 4, 4,
775                                          IRIS_MEMZONE_OTHER, BO_ALLOC_ZEROED);
776    if (!screen->breakpoint_bo)
777       return NULL;
778 
779    if (!iris_init_identifier_bo(screen))
780       return NULL;
781 
782    screen->driconf.dual_color_blend_by_location =
783       driQueryOptionb(config->options, "dual_color_blend_by_location");
784    screen->driconf.disable_throttling =
785       driQueryOptionb(config->options, "disable_throttling");
786    screen->driconf.always_flush_cache = INTEL_DEBUG(DEBUG_STALL) ||
787       driQueryOptionb(config->options, "always_flush_cache");
788    screen->driconf.sync_compile =
789       driQueryOptionb(config->options, "sync_compile");
790    screen->driconf.limit_trig_input_range =
791       driQueryOptionb(config->options, "limit_trig_input_range");
792    screen->driconf.lower_depth_range_rate =
793       driQueryOptionf(config->options, "lower_depth_range_rate");
794    screen->driconf.intel_enable_wa_14018912822 =
795       driQueryOptionb(config->options, "intel_enable_wa_14018912822");
796    screen->driconf.enable_tbimr =
797       driQueryOptionb(config->options, "intel_tbimr");
798    screen->driconf.generated_indirect_threshold =
799       driQueryOptioni(config->options, "generated_indirect_threshold");
800 
801    screen->precompile = debug_get_bool_option("shader_precompile", true);
802 
803    isl_device_init(&screen->isl_dev, screen->devinfo);
804    screen->isl_dev.dummy_aux_address = iris_bufmgr_get_dummy_aux_address(screen->bufmgr);
805 
806    screen->isl_dev.sampler_route_to_lsc =
807       driQueryOptionb(config->options, "intel_sampler_route_to_lsc");
808 
809    iris_compiler_init(screen);
810 
811    screen->l3_config_3d = iris_get_default_l3_config(screen->devinfo, false);
812    screen->l3_config_cs = iris_get_default_l3_config(screen->devinfo, true);
813 
814    iris_disk_cache_init(screen);
815 
816    slab_create_parent(&screen->transfer_pool,
817                       sizeof(struct iris_transfer), 64);
818 
819    iris_detect_kernel_features(screen);
820 
821    struct pipe_screen *pscreen = &screen->base;
822 
823    iris_init_screen_fence_functions(pscreen);
824    iris_init_screen_resource_functions(pscreen);
825    iris_init_screen_measure(screen);
826 
827    pscreen->destroy = iris_screen_unref;
828    pscreen->get_name = iris_get_name;
829    pscreen->get_vendor = iris_get_vendor;
830    pscreen->get_device_vendor = iris_get_device_vendor;
831    pscreen->get_cl_cts_version = iris_get_cl_cts_version;
832    pscreen->get_screen_fd = iris_screen_get_fd;
833    pscreen->get_shader_param = iris_get_shader_param;
834    pscreen->get_compute_param = iris_get_compute_param;
835    pscreen->get_compiler_options = iris_get_compiler_options;
836    pscreen->get_device_uuid = iris_get_device_uuid;
837    pscreen->get_driver_uuid = iris_get_driver_uuid;
838    pscreen->get_disk_shader_cache = iris_get_disk_shader_cache;
839    pscreen->is_format_supported = iris_is_format_supported;
840    pscreen->context_create = iris_create_context;
841    pscreen->get_timestamp = iris_get_timestamp;
842    pscreen->query_memory_info = iris_query_memory_info;
843    pscreen->get_driver_query_group_info = iris_get_monitor_group_info;
844    pscreen->get_driver_query_info = iris_get_monitor_info;
845    pscreen->set_damage_region = iris_set_damage_region;
846    iris_init_screen_program_functions(pscreen);
847 
848    iris_init_screen_caps(screen);
849 
850    genX_call(screen->devinfo, init_screen_state, screen);
851    genX_call(screen->devinfo, init_screen_gen_state, screen);
852 
853    glsl_type_singleton_init_or_ref();
854 
855    intel_driver_ds_init();
856 
857    /* FINISHME: Big core vs little core (for CPUs that have both kinds of
858     * cores) and, possibly, thread vs core should be considered here too.
859     */
860    unsigned compiler_threads = 1;
861    const struct util_cpu_caps_t *caps = util_get_cpu_caps();
862    unsigned hw_threads = caps->nr_cpus;
863 
864    if (hw_threads >= 12) {
865       compiler_threads = hw_threads * 3 / 4;
866    } else if (hw_threads >= 6) {
867       compiler_threads = hw_threads - 2;
868    } else if (hw_threads >= 2) {
869       compiler_threads = hw_threads - 1;
870    }
871 
872    if (!util_queue_init(&screen->shader_compiler_queue,
873                         "sh", 64, compiler_threads,
874                         UTIL_QUEUE_INIT_RESIZE_IF_FULL |
875                         UTIL_QUEUE_INIT_SET_FULL_THREAD_AFFINITY,
876                         NULL)) {
877       iris_screen_destroy(screen);
878       return NULL;
879    }
880 
881    return pscreen;
882 }
883