radv/gfx10: determine if a pipeline is eligible for NGG passthrough
[mesa.git] / src / amd / vulkan / radv_shader.h
1 /*
2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
4 *
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
7 *
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
17 * Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 * IN THE SOFTWARE.
26 */
27
28 #ifndef RADV_SHADER_H
29 #define RADV_SHADER_H
30
31 #include "ac_binary.h"
32 #include "amd_family.h"
33 #include "radv_constants.h"
34
35 #include "nir/nir.h"
36 #include "vulkan/vulkan.h"
37
38 struct radv_device;
39
40 struct radv_shader_module {
41 struct nir_shader *nir;
42 unsigned char sha1[20];
43 uint32_t size;
44 char data[0];
45 };
46
47 enum {
48 RADV_ALPHA_ADJUST_NONE = 0,
49 RADV_ALPHA_ADJUST_SNORM = 1,
50 RADV_ALPHA_ADJUST_SINT = 2,
51 RADV_ALPHA_ADJUST_SSCALED = 3,
52 };
53
54 struct radv_vs_out_key {
55 uint32_t as_es:1;
56 uint32_t as_ls:1;
57 uint32_t as_ngg:1;
58 uint32_t as_ngg_passthrough:1;
59 uint32_t export_prim_id:1;
60 uint32_t export_layer_id:1;
61 uint32_t export_clip_dists:1;
62 };
63
64 struct radv_vs_variant_key {
65 struct radv_vs_out_key out;
66
67 uint32_t instance_rate_inputs;
68 uint32_t instance_rate_divisors[MAX_VERTEX_ATTRIBS];
69 uint8_t vertex_attribute_formats[MAX_VERTEX_ATTRIBS];
70 uint32_t vertex_attribute_bindings[MAX_VERTEX_ATTRIBS];
71 uint32_t vertex_attribute_offsets[MAX_VERTEX_ATTRIBS];
72 uint32_t vertex_attribute_strides[MAX_VERTEX_ATTRIBS];
73
74 /* For 2_10_10_10 formats the alpha is handled as unsigned by pre-vega HW.
75 * so we may need to fix it up. */
76 uint64_t alpha_adjust;
77
78 /* For some formats the channels have to be shuffled. */
79 uint32_t post_shuffle;
80
81 /* Output primitive type. */
82 uint8_t outprim;
83 };
84
85 struct radv_tes_variant_key {
86 struct radv_vs_out_key out;
87
88 uint8_t num_patches;
89 uint8_t tcs_num_outputs;
90 };
91
92 struct radv_tcs_variant_key {
93 struct radv_vs_variant_key vs_key;
94 unsigned primitive_mode;
95 unsigned input_vertices;
96 unsigned num_inputs;
97 uint32_t tes_reads_tess_factors:1;
98 };
99
100 struct radv_fs_variant_key {
101 uint32_t col_format;
102 uint8_t log2_ps_iter_samples;
103 uint8_t num_samples;
104 uint32_t is_int8;
105 uint32_t is_int10;
106 };
107
108 struct radv_cs_variant_key {
109 uint8_t subgroup_size;
110 };
111
112 struct radv_shader_variant_key {
113 union {
114 struct radv_vs_variant_key vs;
115 struct radv_fs_variant_key fs;
116 struct radv_tes_variant_key tes;
117 struct radv_tcs_variant_key tcs;
118 struct radv_cs_variant_key cs;
119
120 /* A common prefix of the vs and tes keys. */
121 struct radv_vs_out_key vs_common_out;
122 };
123 bool has_multiview_view_index;
124 };
125
126 struct radv_nir_compiler_options {
127 struct radv_pipeline_layout *layout;
128 struct radv_shader_variant_key key;
129 bool explicit_scratch_args;
130 bool clamp_shadow_reference;
131 bool robust_buffer_access;
132 bool dump_shader;
133 bool dump_preoptir;
134 bool record_ir;
135 bool check_ir;
136 bool has_ls_vgpr_init_bug;
137 bool use_ngg_streamout;
138 enum radeon_family family;
139 enum chip_class chip_class;
140 uint32_t tess_offchip_block_dw_size;
141 uint32_t address32_hi;
142 };
143
144 enum radv_ud_index {
145 AC_UD_SCRATCH_RING_OFFSETS = 0,
146 AC_UD_PUSH_CONSTANTS = 1,
147 AC_UD_INLINE_PUSH_CONSTANTS = 2,
148 AC_UD_INDIRECT_DESCRIPTOR_SETS = 3,
149 AC_UD_VIEW_INDEX = 4,
150 AC_UD_STREAMOUT_BUFFERS = 5,
151 AC_UD_SHADER_START = 6,
152 AC_UD_VS_VERTEX_BUFFERS = AC_UD_SHADER_START,
153 AC_UD_VS_BASE_VERTEX_START_INSTANCE,
154 AC_UD_VS_MAX_UD,
155 AC_UD_PS_MAX_UD,
156 AC_UD_CS_GRID_SIZE = AC_UD_SHADER_START,
157 AC_UD_CS_MAX_UD,
158 AC_UD_GS_MAX_UD,
159 AC_UD_TCS_MAX_UD,
160 AC_UD_TES_MAX_UD,
161 AC_UD_MAX_UD = AC_UD_TCS_MAX_UD,
162 };
163
164 struct radv_stream_output {
165 uint8_t location;
166 uint8_t buffer;
167 uint16_t offset;
168 uint8_t component_mask;
169 uint8_t stream;
170 };
171
172 struct radv_streamout_info {
173 uint16_t num_outputs;
174 struct radv_stream_output outputs[MAX_SO_OUTPUTS];
175 uint16_t strides[MAX_SO_BUFFERS];
176 uint32_t enabled_stream_buffers_mask;
177 };
178
179 struct radv_userdata_info {
180 int8_t sgpr_idx;
181 uint8_t num_sgprs;
182 };
183
184 struct radv_userdata_locations {
185 struct radv_userdata_info descriptor_sets[MAX_SETS];
186 struct radv_userdata_info shader_data[AC_UD_MAX_UD];
187 uint32_t descriptor_sets_enabled;
188 };
189
190 struct radv_vs_output_info {
191 uint8_t vs_output_param_offset[VARYING_SLOT_MAX];
192 uint8_t clip_dist_mask;
193 uint8_t cull_dist_mask;
194 uint8_t param_exports;
195 bool writes_pointsize;
196 bool writes_layer;
197 bool writes_viewport_index;
198 bool export_prim_id;
199 unsigned pos_exports;
200 };
201
202 struct radv_es_output_info {
203 uint32_t esgs_itemsize;
204 };
205
206 struct gfx9_gs_info {
207 uint32_t vgt_gs_onchip_cntl;
208 uint32_t vgt_gs_max_prims_per_subgroup;
209 uint32_t vgt_esgs_ring_itemsize;
210 uint32_t lds_size;
211 };
212
213 struct gfx10_ngg_info {
214 uint16_t ngg_emit_size; /* in dwords */
215 uint32_t hw_max_esverts;
216 uint32_t max_gsprims;
217 uint32_t max_out_verts;
218 uint32_t prim_amp_factor;
219 uint32_t vgt_esgs_ring_itemsize;
220 uint32_t esgs_ring_size;
221 bool max_vert_out_per_gs_instance;
222 };
223
224 struct radv_shader_info {
225 bool loads_push_constants;
226 bool loads_dynamic_offsets;
227 uint8_t min_push_constant_used;
228 uint8_t max_push_constant_used;
229 bool has_only_32bit_push_constants;
230 bool has_indirect_push_constants;
231 uint8_t num_inline_push_consts;
232 uint8_t base_inline_push_consts;
233 uint32_t desc_set_used_mask;
234 bool needs_multiview_view_index;
235 bool uses_invocation_id;
236 bool uses_prim_id;
237 uint8_t wave_size;
238 struct radv_userdata_locations user_sgprs_locs;
239 unsigned num_user_sgprs;
240 unsigned num_input_sgprs;
241 unsigned num_input_vgprs;
242 unsigned private_mem_vgprs;
243 bool need_indirect_descriptor_sets;
244 bool is_ngg;
245 bool is_ngg_passthrough;
246 struct {
247 uint64_t ls_outputs_written;
248 uint8_t input_usage_mask[VERT_ATTRIB_MAX];
249 uint8_t output_usage_mask[VARYING_SLOT_VAR31 + 1];
250 bool has_vertex_buffers; /* needs vertex buffers and base/start */
251 bool needs_draw_id;
252 bool needs_instance_id;
253 struct radv_vs_output_info outinfo;
254 struct radv_es_output_info es_info;
255 bool as_es;
256 bool as_ls;
257 bool export_prim_id;
258 } vs;
259 struct {
260 uint8_t output_usage_mask[VARYING_SLOT_VAR31 + 1];
261 uint8_t num_stream_output_components[4];
262 uint8_t output_streams[VARYING_SLOT_VAR31 + 1];
263 uint8_t max_stream;
264 bool writes_memory;
265 unsigned gsvs_vertex_size;
266 unsigned max_gsvs_emit_size;
267 unsigned vertices_in;
268 unsigned vertices_out;
269 unsigned output_prim;
270 unsigned invocations;
271 unsigned es_type; /* GFX9: VS or TES */
272 } gs;
273 struct {
274 uint8_t output_usage_mask[VARYING_SLOT_VAR31 + 1];
275 struct radv_vs_output_info outinfo;
276 struct radv_es_output_info es_info;
277 bool as_es;
278 unsigned primitive_mode;
279 enum gl_tess_spacing spacing;
280 bool ccw;
281 bool point_mode;
282 bool export_prim_id;
283 } tes;
284 struct {
285 bool force_persample;
286 bool needs_sample_positions;
287 bool writes_memory;
288 bool writes_z;
289 bool writes_stencil;
290 bool writes_sample_mask;
291 bool has_pcoord;
292 bool prim_id_input;
293 bool layer_input;
294 uint8_t num_input_clips_culls;
295 uint32_t input_mask;
296 uint32_t flat_shaded_mask;
297 uint32_t float16_shaded_mask;
298 uint32_t num_interp;
299 bool can_discard;
300 bool early_fragment_test;
301 bool post_depth_coverage;
302 } ps;
303 struct {
304 bool uses_grid_size;
305 bool uses_block_id[3];
306 bool uses_thread_id[3];
307 bool uses_local_invocation_idx;
308 unsigned block_size[3];
309 } cs;
310 struct {
311 uint64_t outputs_written;
312 uint64_t patch_outputs_written;
313 unsigned tcs_vertices_out;
314 uint32_t num_patches;
315 uint32_t lds_size;
316 } tcs;
317
318 struct radv_streamout_info so;
319
320 struct gfx9_gs_info gs_ring_info;
321 struct gfx10_ngg_info ngg_info;
322
323 unsigned float_controls_mode;
324 };
325
326 enum radv_shader_binary_type {
327 RADV_BINARY_TYPE_LEGACY,
328 RADV_BINARY_TYPE_RTLD
329 };
330
331 struct radv_shader_binary {
332 enum radv_shader_binary_type type;
333 gl_shader_stage stage;
334 bool is_gs_copy_shader;
335
336 struct radv_shader_info info;
337
338 /* Self-referential size so we avoid consistency issues. */
339 uint32_t total_size;
340 };
341
342 struct radv_shader_binary_legacy {
343 struct radv_shader_binary base;
344 struct ac_shader_config config;
345 unsigned code_size;
346 unsigned exec_size;
347 unsigned ir_size;
348 unsigned disasm_size;
349
350 /* data has size of code_size + ir_size + disasm_size + 2, where
351 * the +2 is for 0 of the ir strings. */
352 uint8_t data[0];
353 };
354
355 struct radv_shader_binary_rtld {
356 struct radv_shader_binary base;
357 unsigned elf_size;
358 unsigned llvm_ir_size;
359 uint8_t data[0];
360 };
361
362 struct radv_shader_variant {
363 uint32_t ref_count;
364
365 struct radeon_winsys_bo *bo;
366 uint64_t bo_offset;
367 struct ac_shader_config config;
368 uint32_t code_size;
369 uint32_t exec_size;
370 struct radv_shader_info info;
371
372 /* debug only */
373 bool aco_used;
374 char *spirv;
375 uint32_t spirv_size;
376 char *nir_string;
377 char *disasm_string;
378 char *ir_string;
379
380 struct list_head slab_list;
381 };
382
383 struct radv_shader_slab {
384 struct list_head slabs;
385 struct list_head shaders;
386 struct radeon_winsys_bo *bo;
387 uint64_t size;
388 char *ptr;
389 };
390
391 void
392 radv_optimize_nir(struct nir_shader *shader, bool optimize_conservatively,
393 bool allow_copies);
394 bool
395 radv_nir_lower_ycbcr_textures(nir_shader *shader,
396 const struct radv_pipeline_layout *layout);
397
398 nir_shader *
399 radv_shader_compile_to_nir(struct radv_device *device,
400 struct radv_shader_module *module,
401 const char *entrypoint_name,
402 gl_shader_stage stage,
403 const VkSpecializationInfo *spec_info,
404 const VkPipelineCreateFlags flags,
405 const struct radv_pipeline_layout *layout,
406 bool use_aco);
407
408 void *
409 radv_alloc_shader_memory(struct radv_device *device,
410 struct radv_shader_variant *shader);
411
412 void
413 radv_destroy_shader_slabs(struct radv_device *device);
414
415 void
416 radv_create_shaders(struct radv_pipeline *pipeline,
417 struct radv_device *device,
418 struct radv_pipeline_cache *cache,
419 const struct radv_pipeline_key *key,
420 const VkPipelineShaderStageCreateInfo **pStages,
421 const VkPipelineCreateFlags flags,
422 VkPipelineCreationFeedbackEXT *pipeline_feedback,
423 VkPipelineCreationFeedbackEXT **stage_feedbacks);
424
425 struct radv_shader_variant *
426 radv_shader_variant_create(struct radv_device *device,
427 const struct radv_shader_binary *binary,
428 bool keep_shader_info);
429 struct radv_shader_variant *
430 radv_shader_variant_compile(struct radv_device *device,
431 struct radv_shader_module *module,
432 struct nir_shader *const *shaders,
433 int shader_count,
434 struct radv_pipeline_layout *layout,
435 const struct radv_shader_variant_key *key,
436 struct radv_shader_info *info,
437 bool keep_shader_info,
438 bool use_aco,
439 struct radv_shader_binary **binary_out);
440
441 struct radv_shader_variant *
442 radv_create_gs_copy_shader(struct radv_device *device, struct nir_shader *nir,
443 struct radv_shader_info *info,
444 struct radv_shader_binary **binary_out,
445 bool multiview, bool keep_shader_info);
446
447 void
448 radv_shader_variant_destroy(struct radv_device *device,
449 struct radv_shader_variant *variant);
450
451
452 unsigned
453 radv_get_max_waves(struct radv_device *device,
454 struct radv_shader_variant *variant,
455 gl_shader_stage stage);
456
457 unsigned
458 radv_get_max_workgroup_size(enum chip_class chip_class,
459 gl_shader_stage stage,
460 const unsigned *sizes);
461
462 const char *
463 radv_get_shader_name(struct radv_shader_info *info,
464 gl_shader_stage stage);
465
466 void
467 radv_shader_dump_stats(struct radv_device *device,
468 struct radv_shader_variant *variant,
469 gl_shader_stage stage,
470 FILE *file);
471
472 bool
473 radv_can_dump_shader(struct radv_device *device,
474 struct radv_shader_module *module,
475 bool is_gs_copy_shader);
476
477 bool
478 radv_can_dump_shader_stats(struct radv_device *device,
479 struct radv_shader_module *module);
480
481 unsigned
482 shader_io_get_unique_index(gl_varying_slot slot);
483
484 void
485 radv_lower_fs_io(nir_shader *nir);
486
487 #endif