9da7363d99e4f875756dcbc6384c333a12cdfaf7
[mesa.git] / src / gallium / drivers / radeonsi / si_shader.h
1 /*
2 * Copyright 2012 Advanced Micro Devices, Inc.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
23 */
24
25 /* The compiler middle-end architecture: Explaining (non-)monolithic shaders
26 * -------------------------------------------------------------------------
27 *
28 * Typically, there is one-to-one correspondence between API and HW shaders,
29 * that is, for every API shader, there is exactly one shader binary in
30 * the driver.
31 *
32 * The problem with that is that we also have to emulate some API states
33 * (e.g. alpha-test, and many others) in shaders too. The two obvious ways
34 * to deal with it are:
35 * - each shader has multiple variants for each combination of emulated states,
36 * and the variants are compiled on demand, possibly relying on a shader
37 * cache for good performance
38 * - patch shaders at the binary level
39 *
40 * This driver uses something completely different. The emulated states are
41 * usually implemented at the beginning or end of shaders. Therefore, we can
42 * split the shader into 3 parts:
43 * - prolog part (shader code dependent on states)
44 * - main part (the API shader)
45 * - epilog part (shader code dependent on states)
46 *
47 * Each part is compiled as a separate shader and the final binaries are
48 * concatenated. This type of shader is called non-monolithic, because it
49 * consists of multiple independent binaries. Creating a new shader variant
50 * is therefore only a concatenation of shader parts (binaries) and doesn't
51 * involve any compilation. The main shader parts are the only parts that are
52 * compiled when applications create shader objects. The prolog and epilog
53 * parts are compiled on the first use and saved, so that their binaries can
54 * be reused by many other shaders.
55 *
56 * One of the roles of the prolog part is to compute vertex buffer addresses
57 * for vertex shaders. A few of the roles of the epilog part are color buffer
58 * format conversions in pixel shaders that we have to do manually, and write
59 * tessellation factors in tessellation control shaders. The prolog and epilog
60 * have many other important responsibilities in various shader stages.
61 * They don't just "emulate legacy stuff".
62 *
63 * Monolithic shaders are shaders where the parts are combined before LLVM
64 * compilation, and the whole thing is compiled and optimized as one unit with
65 * one binary on the output. The result is the same as the non-monolithic
66 * shader, but the final code can be better, because LLVM can optimize across
67 * all shader parts. Monolithic shaders aren't usually used except for these
68 * special cases:
69 *
70 * 1) Some rarely-used states require modification of the main shader part
71 * itself, and in such cases, only the monolithic shader variant is
72 * compiled, and that's always done on the first use.
73 *
74 * 2) When we do cross-stage optimizations for separate shader objects and
75 * e.g. eliminate unused shader varyings, the resulting optimized shader
76 * variants are always compiled as monolithic shaders, and always
77 * asynchronously (i.e. not stalling ongoing rendering). We call them
78 * "optimized monolithic" shaders. The important property here is that
79 * the non-monolithic unoptimized shader variant is always available for use
80 * when the asynchronous compilation of the optimized shader is not done
81 * yet.
82 *
83 * Starting with GFX9 chips, some shader stages are merged, and the number of
84 * shader parts per shader increased. The complete new list of shader parts is:
85 * - 1st shader: prolog part
86 * - 1st shader: main part
87 * - 2nd shader: prolog part
88 * - 2nd shader: main part
89 * - 2nd shader: epilog part
90 */
91
92 /* How linking shader inputs and outputs between vertex, tessellation, and
93 * geometry shaders works.
94 *
95 * Inputs and outputs between shaders are stored in a buffer. This buffer
96 * lives in LDS (typical case for tessellation), but it can also live
97 * in memory (ESGS). Each input or output has a fixed location within a vertex.
98 * The highest used input or output determines the stride between vertices.
99 *
100 * Since GS and tessellation are only possible in the OpenGL core profile,
101 * only these semantics are valid for per-vertex data:
102 *
103 * Name Location
104 *
105 * POSITION 0
106 * PSIZE 1
107 * CLIPDIST0..1 2..3
108 * CULLDIST0..1 (not implemented)
109 * GENERIC0..31 4..35
110 *
111 * For example, a shader only writing GENERIC0 has the output stride of 5.
112 *
113 * Only these semantics are valid for per-patch data:
114 *
115 * Name Location
116 *
117 * TESSOUTER 0
118 * TESSINNER 1
119 * PATCH0..29 2..31
120 *
121 * That's how independent shaders agree on input and output locations.
122 * The si_shader_io_get_unique_index function assigns the locations.
123 *
124 * For tessellation, other required information for calculating the input and
125 * output addresses like the vertex stride, the patch stride, and the offsets
126 * where per-vertex and per-patch data start, is passed to the shader via
127 * user data SGPRs. The offsets and strides are calculated at draw time and
128 * aren't available at compile time.
129 */
130
131 #ifndef SI_SHADER_H
132 #define SI_SHADER_H
133
134 #include "ac_binary.h"
135 #include "ac_llvm_build.h"
136 #include "ac_llvm_util.h"
137 #include "util/simple_mtx.h"
138 #include "util/u_inlines.h"
139 #include "util/u_live_shader_cache.h"
140 #include "util/u_queue.h"
141
142 #include <stdio.h>
143
144 // Use LDS symbols when supported by LLVM. Can be disabled for testing the old
145 // path on newer LLVM for now. Should be removed in the long term.
146 #define USE_LDS_SYMBOLS (true)
147
148 struct nir_shader;
149 struct si_shader;
150 struct si_context;
151
152 #define SI_MAX_ATTRIBS 16
153 #define SI_MAX_VS_OUTPUTS 40
154
155 /* Shader IO unique indices are supported for TGSI_SEMANTIC_GENERIC with an
156 * index smaller than this.
157 */
158 #define SI_MAX_IO_GENERIC 32
159
160 #define SI_NGG_PRIM_EDGE_FLAG_BITS ((1 << 9) | (1 << 19) | (1 << 29))
161
162 /* SGPR user data indices */
163 enum
164 {
165 SI_SGPR_RW_BUFFERS, /* rings (& stream-out, VS only) */
166 SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES,
167 SI_SGPR_CONST_AND_SHADER_BUFFERS, /* or just a constant buffer 0 pointer */
168 SI_SGPR_SAMPLERS_AND_IMAGES,
169 SI_NUM_RESOURCE_SGPRS,
170
171 /* API VS, TES without GS, GS copy shader */
172 SI_SGPR_VS_STATE_BITS = SI_NUM_RESOURCE_SGPRS,
173 SI_NUM_VS_STATE_RESOURCE_SGPRS,
174
175 /* all VS variants */
176 SI_SGPR_BASE_VERTEX = SI_NUM_VS_STATE_RESOURCE_SGPRS,
177 SI_SGPR_START_INSTANCE,
178 SI_SGPR_DRAWID,
179 SI_VS_NUM_USER_SGPR,
180
181 SI_SGPR_VS_BLIT_DATA = SI_SGPR_CONST_AND_SHADER_BUFFERS,
182
183 /* TES */
184 SI_SGPR_TES_OFFCHIP_LAYOUT = SI_NUM_VS_STATE_RESOURCE_SGPRS,
185 SI_SGPR_TES_OFFCHIP_ADDR,
186 SI_TES_NUM_USER_SGPR,
187
188 /* GFX6-8: TCS only */
189 GFX6_SGPR_TCS_OFFCHIP_LAYOUT = SI_NUM_RESOURCE_SGPRS,
190 GFX6_SGPR_TCS_OUT_OFFSETS,
191 GFX6_SGPR_TCS_OUT_LAYOUT,
192 GFX6_SGPR_TCS_IN_LAYOUT,
193 GFX6_TCS_NUM_USER_SGPR,
194
195 /* GFX9: Merged shaders. */
196 /* 2ND_CONST_AND_SHADER_BUFFERS is set in USER_DATA_ADDR_LO (SGPR0). */
197 /* 2ND_SAMPLERS_AND_IMAGES is set in USER_DATA_ADDR_HI (SGPR1). */
198 GFX9_MERGED_NUM_USER_SGPR = SI_VS_NUM_USER_SGPR,
199
200 /* GFX9: Merged LS-HS (VS-TCS) only. */
201 GFX9_SGPR_TCS_OFFCHIP_LAYOUT = GFX9_MERGED_NUM_USER_SGPR,
202 GFX9_SGPR_TCS_OUT_OFFSETS,
203 GFX9_SGPR_TCS_OUT_LAYOUT,
204 GFX9_TCS_NUM_USER_SGPR,
205
206 /* GS limits */
207 GFX6_GS_NUM_USER_SGPR = SI_NUM_RESOURCE_SGPRS,
208 GFX9_VSGS_NUM_USER_SGPR = SI_VS_NUM_USER_SGPR,
209 GFX9_TESGS_NUM_USER_SGPR = SI_TES_NUM_USER_SGPR,
210 SI_GSCOPY_NUM_USER_SGPR = SI_NUM_VS_STATE_RESOURCE_SGPRS,
211
212 /* PS only */
213 SI_SGPR_ALPHA_REF = SI_NUM_RESOURCE_SGPRS,
214 SI_PS_NUM_USER_SGPR,
215
216 /* The value has to be 12, because the hw requires that descriptors
217 * are aligned to 4 SGPRs.
218 */
219 SI_SGPR_VS_VB_DESCRIPTOR_FIRST = 12,
220 };
221
222 /* LLVM function parameter indices */
223 enum
224 {
225 SI_NUM_RESOURCE_PARAMS = 4,
226
227 /* PS only parameters */
228 SI_PARAM_ALPHA_REF = SI_NUM_RESOURCE_PARAMS,
229 SI_PARAM_PRIM_MASK,
230 SI_PARAM_PERSP_SAMPLE,
231 SI_PARAM_PERSP_CENTER,
232 SI_PARAM_PERSP_CENTROID,
233 SI_PARAM_PERSP_PULL_MODEL,
234 SI_PARAM_LINEAR_SAMPLE,
235 SI_PARAM_LINEAR_CENTER,
236 SI_PARAM_LINEAR_CENTROID,
237 SI_PARAM_LINE_STIPPLE_TEX,
238 SI_PARAM_POS_X_FLOAT,
239 SI_PARAM_POS_Y_FLOAT,
240 SI_PARAM_POS_Z_FLOAT,
241 SI_PARAM_POS_W_FLOAT,
242 SI_PARAM_FRONT_FACE,
243 SI_PARAM_ANCILLARY,
244 SI_PARAM_SAMPLE_COVERAGE,
245 SI_PARAM_POS_FIXED_PT,
246
247 SI_NUM_PARAMS = SI_PARAM_POS_FIXED_PT + 9, /* +8 for COLOR[0..1] */
248 };
249
250 /* Fields of driver-defined VS state SGPR. */
251 #define S_VS_STATE_CLAMP_VERTEX_COLOR(x) (((unsigned)(x)&0x1) << 0)
252 #define C_VS_STATE_CLAMP_VERTEX_COLOR 0xFFFFFFFE
253 #define S_VS_STATE_INDEXED(x) (((unsigned)(x)&0x1) << 1)
254 #define C_VS_STATE_INDEXED 0xFFFFFFFD
255 #define S_VS_STATE_OUTPRIM(x) (((unsigned)(x)&0x3) << 2)
256 #define C_VS_STATE_OUTPRIM 0xFFFFFFF3
257 #define S_VS_STATE_PROVOKING_VTX_INDEX(x) (((unsigned)(x)&0x3) << 4)
258 #define C_VS_STATE_PROVOKING_VTX_INDEX 0xFFFFFFCF
259 #define S_VS_STATE_STREAMOUT_QUERY_ENABLED(x) (((unsigned)(x)&0x1) << 6)
260 #define C_VS_STATE_STREAMOUT_QUERY_ENABLED 0xFFFFFFBF
261 #define S_VS_STATE_SMALL_PRIM_PRECISION(x) (((unsigned)(x)&0xF) << 7)
262 #define C_VS_STATE_SMALL_PRIM_PRECISION 0xFFFFF87F
263 #define S_VS_STATE_LS_OUT_PATCH_SIZE(x) (((unsigned)(x)&0x1FFF) << 11)
264 #define C_VS_STATE_LS_OUT_PATCH_SIZE 0xFF0007FF
265 #define S_VS_STATE_LS_OUT_VERTEX_SIZE(x) (((unsigned)(x)&0xFF) << 24)
266 #define C_VS_STATE_LS_OUT_VERTEX_SIZE 0x00FFFFFF
267
268 enum
269 {
270 /* Use a property enum that CS wouldn't use. */
271 TGSI_PROPERTY_CS_LOCAL_SIZE = TGSI_PROPERTY_FS_COORD_ORIGIN,
272
273 /* These represent the number of SGPRs the shader uses. */
274 SI_VS_BLIT_SGPRS_POS = 3,
275 SI_VS_BLIT_SGPRS_POS_COLOR = 7,
276 SI_VS_BLIT_SGPRS_POS_TEXCOORD = 9,
277 };
278
279 #define SI_NGG_CULL_VIEW_SMALLPRIMS (1 << 0) /* view.xy + small prims */
280 #define SI_NGG_CULL_BACK_FACE (1 << 1) /* back faces */
281 #define SI_NGG_CULL_FRONT_FACE (1 << 2) /* front faces */
282 #define SI_NGG_CULL_GS_FAST_LAUNCH_TRI_LIST (1 << 3) /* GS fast launch: triangles */
283 #define SI_NGG_CULL_GS_FAST_LAUNCH_TRI_STRIP (1 << 4) /* GS fast launch: triangle strip */
284 #define SI_NGG_CULL_GS_FAST_LAUNCH_ALL (0x3 << 3) /* GS fast launch (both prim types) */
285
286 /**
287 * For VS shader keys, describe any fixups required for vertex fetch.
288 *
289 * \ref log_size, \ref format, and the number of channels are interpreted as
290 * by \ref ac_build_opencoded_load_format.
291 *
292 * Note: all bits 0 (size = 1 byte, num channels = 1, format = float) is an
293 * impossible format and indicates that no fixup is needed (just use
294 * buffer_load_format_xyzw).
295 */
296 union si_vs_fix_fetch {
297 struct {
298 uint8_t log_size : 2; /* 1, 2, 4, 8 or bytes per channel */
299 uint8_t num_channels_m1 : 2; /* number of channels minus 1 */
300 uint8_t format : 3; /* AC_FETCH_FORMAT_xxx */
301 uint8_t reverse : 1; /* reverse XYZ channels */
302 } u;
303 uint8_t bits;
304 };
305
306 struct si_shader;
307
308 /* State of the context creating the shader object. */
309 struct si_compiler_ctx_state {
310 /* Should only be used by si_init_shader_selector_async and
311 * si_build_shader_variant if thread_index == -1 (non-threaded). */
312 struct ac_llvm_compiler *compiler;
313
314 /* Used if thread_index == -1 or if debug.async is true. */
315 struct pipe_debug_callback debug;
316
317 /* Used for creating the log string for gallium/ddebug. */
318 bool is_debug_context;
319 };
320
321 struct si_shader_info {
322 gl_shader_stage stage;
323
324 ubyte num_inputs;
325 ubyte num_outputs;
326 ubyte input_semantic_name[PIPE_MAX_SHADER_INPUTS]; /**< TGSI_SEMANTIC_x */
327 ubyte input_semantic_index[PIPE_MAX_SHADER_INPUTS];
328 ubyte input_interpolate[PIPE_MAX_SHADER_INPUTS];
329 ubyte input_usage_mask[PIPE_MAX_SHADER_INPUTS];
330 ubyte output_semantic_name[PIPE_MAX_SHADER_OUTPUTS]; /**< TGSI_SEMANTIC_x */
331 ubyte output_semantic_index[PIPE_MAX_SHADER_OUTPUTS];
332 ubyte output_usagemask[PIPE_MAX_SHADER_OUTPUTS];
333 ubyte output_readmask[PIPE_MAX_SHADER_OUTPUTS];
334 ubyte output_streams[PIPE_MAX_SHADER_OUTPUTS];
335
336 ubyte color_interpolate[2];
337 ubyte color_interpolate_loc[2];
338
339 int constbuf0_num_slots;
340 unsigned const_buffers_declared; /**< bitmask of declared const buffers */
341 unsigned samplers_declared; /**< bitmask of declared samplers */
342 ubyte num_stream_output_components[4];
343
344 uint num_memory_instructions; /**< sampler, buffer, and image instructions */
345
346 ubyte colors_read; /**< which color components are read by the FS */
347 ubyte colors_written;
348 bool reads_samplemask; /**< does fragment shader read sample mask? */
349 bool reads_tess_factors; /**< If TES reads TESSINNER or TESSOUTER */
350 bool writes_z; /**< does fragment shader write Z value? */
351 bool writes_stencil; /**< does fragment shader write stencil value? */
352 bool writes_samplemask; /**< does fragment shader write sample mask? */
353 bool writes_edgeflag; /**< vertex shader outputs edgeflag */
354 bool uses_kill; /**< KILL or KILL_IF instruction used? */
355 bool uses_persp_center;
356 bool uses_persp_centroid;
357 bool uses_persp_sample;
358 bool uses_linear_center;
359 bool uses_linear_centroid;
360 bool uses_linear_sample;
361 bool uses_persp_opcode_interp_sample;
362 bool uses_linear_opcode_interp_sample;
363 bool uses_instanceid;
364 bool uses_vertexid;
365 bool uses_vertexid_nobase;
366 bool uses_basevertex;
367 bool uses_drawid;
368 bool uses_primid;
369 bool uses_frontface;
370 bool uses_invocationid;
371 bool uses_thread_id[3];
372 bool uses_block_id[3];
373 bool uses_block_size;
374 bool uses_grid_size;
375 bool uses_subgroup_info;
376 bool writes_position;
377 bool writes_psize;
378 bool writes_clipvertex;
379 bool writes_primid;
380 bool writes_viewport_index;
381 bool writes_layer;
382 bool writes_memory; /**< contains stores or atomics to buffers or images */
383 bool uses_derivatives;
384 bool uses_bindless_samplers;
385 bool uses_bindless_images;
386 bool uses_fbfetch;
387 unsigned clipdist_writemask;
388 unsigned culldist_writemask;
389 unsigned num_written_culldistance;
390 unsigned num_written_clipdistance;
391
392 unsigned images_declared; /**< bitmask of declared images */
393 unsigned image_buffers; /**< bitmask of images that are buffers */
394 unsigned msaa_images_declared; /**< bitmask of declared MSAA images */
395 unsigned shader_buffers_declared; /**< bitmask of declared shader buffers */
396
397 unsigned properties[TGSI_PROPERTY_COUNT]; /* index with TGSI_PROPERTY_ */
398
399 /** Whether all codepaths write tess factors in all invocations. */
400 bool tessfactors_are_def_in_all_invocs;
401 };
402
403 /* A shader selector is a gallium CSO and contains shader variants and
404 * binaries for one NIR program. This can be shared by multiple contexts.
405 */
406 struct si_shader_selector {
407 struct util_live_shader base;
408 struct si_screen *screen;
409 struct util_queue_fence ready;
410 struct si_compiler_ctx_state compiler_ctx_state;
411
412 simple_mtx_t mutex;
413 struct si_shader *first_variant; /* immutable after the first variant */
414 struct si_shader *last_variant; /* mutable */
415
416 /* The compiled NIR shader without a prolog and/or epilog (not
417 * uploaded to a buffer object).
418 */
419 struct si_shader *main_shader_part;
420 struct si_shader *main_shader_part_ls; /* as_ls is set in the key */
421 struct si_shader *main_shader_part_es; /* as_es is set in the key */
422 struct si_shader *main_shader_part_ngg; /* as_ngg is set in the key */
423 struct si_shader *main_shader_part_ngg_es; /* for Wave32 TES before legacy GS */
424
425 struct si_shader *gs_copy_shader;
426
427 struct nir_shader *nir;
428 void *nir_binary;
429 unsigned nir_size;
430
431 struct pipe_stream_output_info so;
432 struct si_shader_info info;
433
434 /* PIPE_SHADER_[VERTEX|FRAGMENT|...] */
435 enum pipe_shader_type type;
436 ubyte const_and_shader_buf_descriptors_index;
437 ubyte sampler_and_images_descriptors_index;
438 bool vs_needs_prolog;
439 bool prim_discard_cs_allowed;
440 bool ngg_culling_allowed;
441 ubyte cs_shaderbufs_sgpr_index;
442 ubyte cs_num_shaderbufs_in_user_sgprs;
443 ubyte cs_images_sgpr_index;
444 ubyte cs_images_num_sgprs;
445 ubyte cs_num_images_in_user_sgprs;
446 unsigned num_vs_inputs;
447 unsigned num_vbos_in_user_sgprs;
448 unsigned pa_cl_vs_out_cntl;
449 ubyte clipdist_mask;
450 ubyte culldist_mask;
451 unsigned rast_prim;
452
453 /* ES parameters. */
454 unsigned esgs_itemsize; /* vertex stride */
455 unsigned lshs_vertex_stride;
456
457 /* GS parameters. */
458 unsigned gs_input_verts_per_prim;
459 unsigned gs_output_prim;
460 unsigned gs_max_out_vertices;
461 unsigned gs_num_invocations;
462 unsigned max_gs_stream; /* count - 1 */
463 unsigned gsvs_vertex_size;
464 unsigned max_gsvs_emit_size;
465 unsigned enabled_streamout_buffer_mask;
466 bool tess_turns_off_ngg;
467
468 /* PS parameters. */
469 unsigned color_attr_index[2];
470 unsigned db_shader_control;
471 /* Set 0xf or 0x0 (4 bits) per each written output.
472 * ANDed with spi_shader_col_format.
473 */
474 unsigned colors_written_4bit;
475
476 uint64_t outputs_written_before_ps; /* "get_unique_index" bits */
477 uint64_t outputs_written; /* "get_unique_index" bits */
478 uint32_t patch_outputs_written; /* "get_unique_index_patch" bits */
479
480 uint64_t inputs_read; /* "get_unique_index" bits */
481
482 /* bitmasks of used descriptor slots */
483 uint64_t active_const_and_shader_buffers;
484 uint64_t active_samplers_and_images;
485 };
486
487 /* Valid shader configurations:
488 *
489 * API shaders VS | TCS | TES | GS |pass| PS
490 * are compiled as: | | | |thru|
491 * | | | | |
492 * Only VS & PS: VS | | | | | PS
493 * GFX6 - with GS: ES | | | GS | VS | PS
494 * - with tess: LS | HS | VS | | | PS
495 * - with both: LS | HS | ES | GS | VS | PS
496 * GFX9 - with GS: -> | | | GS | VS | PS
497 * - with tess: -> | HS | VS | | | PS
498 * - with both: -> | HS | -> | GS | VS | PS
499 * | | | | |
500 * NGG - VS & PS: GS | | | | | PS
501 * (GFX10+) - with GS: -> | | | GS | | PS
502 * - with tess: -> | HS | GS | | | PS
503 * - with both: -> | HS | -> | GS | | PS
504 *
505 * -> = merged with the next stage
506 */
507
508 /* Use the byte alignment for all following structure members for optimal
509 * shader key memory footprint.
510 */
511 #pragma pack(push, 1)
512
513 /* Common VS bits between the shader key and the prolog key. */
514 struct si_vs_prolog_bits {
515 /* - If neither "is_one" nor "is_fetched" has a bit set, the instance
516 * divisor is 0.
517 * - If "is_one" has a bit set, the instance divisor is 1.
518 * - If "is_fetched" has a bit set, the instance divisor will be loaded
519 * from the constant buffer.
520 */
521 uint16_t instance_divisor_is_one; /* bitmask of inputs */
522 uint16_t instance_divisor_is_fetched; /* bitmask of inputs */
523 unsigned ls_vgpr_fix : 1;
524 unsigned unpack_instance_id_from_vertex_id : 1;
525 };
526
527 /* Common TCS bits between the shader key and the epilog key. */
528 struct si_tcs_epilog_bits {
529 unsigned prim_mode : 3;
530 unsigned invoc0_tess_factors_are_def : 1;
531 unsigned tes_reads_tess_factors : 1;
532 };
533
534 struct si_gs_prolog_bits {
535 unsigned tri_strip_adj_fix : 1;
536 unsigned gfx9_prev_is_vs : 1;
537 };
538
539 /* Common PS bits between the shader key and the prolog key. */
540 struct si_ps_prolog_bits {
541 unsigned color_two_side : 1;
542 unsigned flatshade_colors : 1;
543 unsigned poly_stipple : 1;
544 unsigned force_persp_sample_interp : 1;
545 unsigned force_linear_sample_interp : 1;
546 unsigned force_persp_center_interp : 1;
547 unsigned force_linear_center_interp : 1;
548 unsigned bc_optimize_for_persp : 1;
549 unsigned bc_optimize_for_linear : 1;
550 unsigned samplemask_log_ps_iter : 3;
551 };
552
553 /* Common PS bits between the shader key and the epilog key. */
554 struct si_ps_epilog_bits {
555 unsigned spi_shader_col_format;
556 unsigned color_is_int8 : 8;
557 unsigned color_is_int10 : 8;
558 unsigned last_cbuf : 3;
559 unsigned alpha_func : 3;
560 unsigned alpha_to_one : 1;
561 unsigned poly_line_smoothing : 1;
562 unsigned clamp_color : 1;
563 };
564
565 union si_shader_part_key {
566 struct {
567 struct si_vs_prolog_bits states;
568 unsigned num_input_sgprs : 6;
569 /* For merged stages such as LS-HS, HS input VGPRs are first. */
570 unsigned num_merged_next_stage_vgprs : 3;
571 unsigned num_inputs : 5;
572 unsigned as_ls : 1;
573 unsigned as_es : 1;
574 unsigned as_ngg : 1;
575 unsigned as_prim_discard_cs : 1;
576 unsigned has_ngg_cull_inputs : 1; /* from the NGG cull shader */
577 unsigned gs_fast_launch_tri_list : 1; /* for NGG culling */
578 unsigned gs_fast_launch_tri_strip : 1; /* for NGG culling */
579 /* Prologs for monolithic shaders shouldn't set EXEC. */
580 unsigned is_monolithic : 1;
581 } vs_prolog;
582 struct {
583 struct si_tcs_epilog_bits states;
584 } tcs_epilog;
585 struct {
586 struct si_gs_prolog_bits states;
587 /* Prologs of monolithic shaders shouldn't set EXEC. */
588 unsigned is_monolithic : 1;
589 unsigned as_ngg : 1;
590 } gs_prolog;
591 struct {
592 struct si_ps_prolog_bits states;
593 unsigned num_input_sgprs : 6;
594 unsigned num_input_vgprs : 5;
595 /* Color interpolation and two-side color selection. */
596 unsigned colors_read : 8; /* color input components read */
597 unsigned num_interp_inputs : 5; /* BCOLOR is at this location */
598 unsigned face_vgpr_index : 5;
599 unsigned ancillary_vgpr_index : 5;
600 unsigned wqm : 1;
601 char color_attr_index[2];
602 signed char color_interp_vgpr_index[2]; /* -1 == constant */
603 } ps_prolog;
604 struct {
605 struct si_ps_epilog_bits states;
606 unsigned colors_written : 8;
607 unsigned writes_z : 1;
608 unsigned writes_stencil : 1;
609 unsigned writes_samplemask : 1;
610 } ps_epilog;
611 };
612
613 struct si_shader_key {
614 /* Prolog and epilog flags. */
615 union {
616 struct {
617 struct si_vs_prolog_bits prolog;
618 } vs;
619 struct {
620 struct si_vs_prolog_bits ls_prolog; /* for merged LS-HS */
621 struct si_shader_selector *ls; /* for merged LS-HS */
622 struct si_tcs_epilog_bits epilog;
623 } tcs; /* tessellation control shader */
624 struct {
625 struct si_vs_prolog_bits vs_prolog; /* for merged ES-GS */
626 struct si_shader_selector *es; /* for merged ES-GS */
627 struct si_gs_prolog_bits prolog;
628 } gs;
629 struct {
630 struct si_ps_prolog_bits prolog;
631 struct si_ps_epilog_bits epilog;
632 } ps;
633 } part;
634
635 /* These three are initially set according to the NEXT_SHADER property,
636 * or guessed if the property doesn't seem correct.
637 */
638 unsigned as_es : 1; /* export shader, which precedes GS */
639 unsigned as_ls : 1; /* local shader, which precedes TCS */
640 unsigned as_ngg : 1; /* VS, TES, or GS compiled as NGG primitive shader */
641
642 /* Flags for monolithic compilation only. */
643 struct {
644 /* Whether fetch should be opencoded according to vs_fix_fetch.
645 * Otherwise, if vs_fix_fetch is non-zero, buffer_load_format_xyzw
646 * with minimal fixups is used. */
647 uint16_t vs_fetch_opencode;
648 union si_vs_fix_fetch vs_fix_fetch[SI_MAX_ATTRIBS];
649
650 union {
651 uint64_t ff_tcs_inputs_to_copy; /* for fixed-func TCS */
652 /* When PS needs PrimID and GS is disabled. */
653 unsigned vs_export_prim_id : 1;
654 struct {
655 unsigned interpolate_at_sample_force_center : 1;
656 unsigned fbfetch_msaa : 1;
657 unsigned fbfetch_is_1D : 1;
658 unsigned fbfetch_layered : 1;
659 } ps;
660 } u;
661 } mono;
662
663 /* Optimization flags for asynchronous compilation only. */
664 struct {
665 /* For HW VS (it can be VS, TES, GS) */
666 uint64_t kill_outputs; /* "get_unique_index" bits */
667 unsigned clip_disable : 1;
668
669 /* For NGG VS and TES. */
670 unsigned ngg_culling : 5; /* SI_NGG_CULL_* */
671
672 /* For shaders where monolithic variants have better code.
673 *
674 * This is a flag that has no effect on code generation,
675 * but forces monolithic shaders to be used as soon as
676 * possible, because it's in the "opt" group.
677 */
678 unsigned prefer_mono : 1;
679
680 /* Primitive discard compute shader. */
681 unsigned vs_as_prim_discard_cs : 1;
682 unsigned cs_prim_type : 4;
683 unsigned cs_indexed : 1;
684 unsigned cs_instancing : 1;
685 unsigned cs_primitive_restart : 1;
686 unsigned cs_provoking_vertex_first : 1;
687 unsigned cs_need_correct_orientation : 1;
688 unsigned cs_cull_front : 1;
689 unsigned cs_cull_back : 1;
690 unsigned cs_cull_z : 1;
691 unsigned cs_halfz_clip_space : 1;
692 } opt;
693 };
694
695 /* Restore the pack alignment to default. */
696 #pragma pack(pop)
697
698 /* GCN-specific shader info. */
699 struct si_shader_binary_info {
700 ubyte vs_output_param_offset[SI_MAX_VS_OUTPUTS];
701 ubyte num_input_sgprs;
702 ubyte num_input_vgprs;
703 signed char face_vgpr_index;
704 signed char ancillary_vgpr_index;
705 bool uses_instanceid;
706 ubyte nr_pos_exports;
707 ubyte nr_param_exports;
708 unsigned private_mem_vgprs;
709 unsigned max_simd_waves;
710 };
711
712 struct si_shader_binary {
713 const char *elf_buffer;
714 size_t elf_size;
715
716 char *llvm_ir_string;
717 };
718
719 struct gfx9_gs_info {
720 unsigned es_verts_per_subgroup;
721 unsigned gs_prims_per_subgroup;
722 unsigned gs_inst_prims_in_subgroup;
723 unsigned max_prims_per_subgroup;
724 unsigned esgs_ring_size; /* in bytes */
725 };
726
727 struct si_shader {
728 struct si_compiler_ctx_state compiler_ctx_state;
729
730 struct si_shader_selector *selector;
731 struct si_shader_selector *previous_stage_sel; /* for refcounting */
732 struct si_shader *next_variant;
733
734 struct si_shader_part *prolog;
735 struct si_shader *previous_stage; /* for GFX9 */
736 struct si_shader_part *prolog2;
737 struct si_shader_part *epilog;
738
739 struct si_pm4_state *pm4;
740 struct si_resource *bo;
741 struct si_resource *scratch_bo;
742 struct si_shader_key key;
743 struct util_queue_fence ready;
744 bool compilation_failed;
745 bool is_monolithic;
746 bool is_optimized;
747 bool is_binary_shared;
748 bool is_gs_copy_shader;
749
750 /* The following data is all that's needed for binary shaders. */
751 struct si_shader_binary binary;
752 struct ac_shader_config config;
753 struct si_shader_binary_info info;
754
755 struct {
756 uint16_t ngg_emit_size; /* in dwords */
757 uint16_t hw_max_esverts;
758 uint16_t max_gsprims;
759 uint16_t max_out_verts;
760 uint16_t prim_amp_factor;
761 bool max_vert_out_per_gs_instance;
762 } ngg;
763
764 /* Shader key + LLVM IR + disassembly + statistics.
765 * Generated for debug contexts only.
766 */
767 char *shader_log;
768 size_t shader_log_size;
769
770 struct gfx9_gs_info gs_info;
771
772 /* For save precompute context registers values. */
773 union {
774 struct {
775 unsigned vgt_gsvs_ring_offset_1;
776 unsigned vgt_gsvs_ring_offset_2;
777 unsigned vgt_gsvs_ring_offset_3;
778 unsigned vgt_gsvs_ring_itemsize;
779 unsigned vgt_gs_max_vert_out;
780 unsigned vgt_gs_vert_itemsize;
781 unsigned vgt_gs_vert_itemsize_1;
782 unsigned vgt_gs_vert_itemsize_2;
783 unsigned vgt_gs_vert_itemsize_3;
784 unsigned vgt_gs_instance_cnt;
785 unsigned vgt_gs_onchip_cntl;
786 unsigned vgt_gs_max_prims_per_subgroup;
787 unsigned vgt_esgs_ring_itemsize;
788 } gs;
789
790 struct {
791 unsigned ge_max_output_per_subgroup;
792 unsigned ge_ngg_subgrp_cntl;
793 unsigned vgt_primitiveid_en;
794 unsigned vgt_gs_onchip_cntl;
795 unsigned vgt_gs_instance_cnt;
796 unsigned vgt_esgs_ring_itemsize;
797 unsigned spi_vs_out_config;
798 unsigned spi_shader_idx_format;
799 unsigned spi_shader_pos_format;
800 unsigned pa_cl_vte_cntl;
801 unsigned pa_cl_ngg_cntl;
802 unsigned vgt_gs_max_vert_out; /* for API GS */
803 unsigned ge_pc_alloc; /* uconfig register */
804 } ngg;
805
806 struct {
807 unsigned vgt_gs_mode;
808 unsigned vgt_primitiveid_en;
809 unsigned vgt_reuse_off;
810 unsigned spi_vs_out_config;
811 unsigned spi_shader_pos_format;
812 unsigned pa_cl_vte_cntl;
813 unsigned ge_pc_alloc; /* uconfig register */
814 } vs;
815
816 struct {
817 unsigned spi_ps_input_ena;
818 unsigned spi_ps_input_addr;
819 unsigned spi_baryc_cntl;
820 unsigned spi_ps_in_control;
821 unsigned spi_shader_z_format;
822 unsigned spi_shader_col_format;
823 unsigned cb_shader_mask;
824 } ps;
825 } ctx_reg;
826
827 /*For save precompute registers value */
828 unsigned vgt_tf_param; /* VGT_TF_PARAM */
829 unsigned vgt_vertex_reuse_block_cntl; /* VGT_VERTEX_REUSE_BLOCK_CNTL */
830 unsigned pa_cl_vs_out_cntl;
831 unsigned ge_cntl;
832 };
833
834 struct si_shader_part {
835 struct si_shader_part *next;
836 union si_shader_part_key key;
837 struct si_shader_binary binary;
838 struct ac_shader_config config;
839 };
840
841 /* si_shader.c */
842 bool si_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
843 struct si_shader *shader, struct pipe_debug_callback *debug);
844 bool si_create_shader_variant(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
845 struct si_shader *shader, struct pipe_debug_callback *debug);
846 void si_shader_destroy(struct si_shader *shader);
847 unsigned si_shader_io_get_unique_index_patch(unsigned semantic_name, unsigned index);
848 unsigned si_shader_io_get_unique_index(unsigned semantic_name, unsigned index, unsigned is_varying);
849 bool si_shader_binary_upload(struct si_screen *sscreen, struct si_shader *shader,
850 uint64_t scratch_va);
851 void si_shader_dump(struct si_screen *sscreen, struct si_shader *shader,
852 struct pipe_debug_callback *debug, FILE *f, bool check_debug_option);
853 void si_shader_dump_stats_for_shader_db(struct si_screen *screen, struct si_shader *shader,
854 struct pipe_debug_callback *debug);
855 void si_multiwave_lds_size_workaround(struct si_screen *sscreen, unsigned *lds_size);
856 const char *si_get_shader_name(const struct si_shader *shader);
857 void si_shader_binary_clean(struct si_shader_binary *binary);
858
859 /* si_shader_llvm_gs.c */
860 struct si_shader *si_generate_gs_copy_shader(struct si_screen *sscreen,
861 struct ac_llvm_compiler *compiler,
862 struct si_shader_selector *gs_selector,
863 struct pipe_debug_callback *debug);
864
865 /* si_shader_nir.c */
866 void si_nir_scan_shader(const struct nir_shader *nir, struct si_shader_info *info);
867 void si_finalize_nir(struct pipe_screen *screen, void *nirptr, bool optimize);
868
869 /* si_state_shaders.c */
870 void gfx9_get_gs_info(struct si_shader_selector *es, struct si_shader_selector *gs,
871 struct gfx9_gs_info *out);
872
873 /* Inline helpers. */
874
875 /* Return the pointer to the main shader part's pointer. */
876 static inline struct si_shader **si_get_main_shader_part(struct si_shader_selector *sel,
877 struct si_shader_key *key)
878 {
879 if (key->as_ls)
880 return &sel->main_shader_part_ls;
881 if (key->as_es && key->as_ngg)
882 return &sel->main_shader_part_ngg_es;
883 if (key->as_es)
884 return &sel->main_shader_part_es;
885 if (key->as_ngg)
886 return &sel->main_shader_part_ngg;
887 return &sel->main_shader_part;
888 }
889
890 static inline bool gfx10_is_ngg_passthrough(struct si_shader *shader)
891 {
892 struct si_shader_selector *sel = shader->selector;
893
894 return sel->info.stage != MESA_SHADER_GEOMETRY && !sel->so.num_outputs && !sel->info.writes_edgeflag &&
895 !shader->key.opt.ngg_culling &&
896 (sel->info.stage != MESA_SHADER_VERTEX || !shader->key.mono.u.vs_export_prim_id);
897 }
898
899 static inline bool si_shader_uses_bindless_samplers(struct si_shader_selector *selector)
900 {
901 return selector ? selector->info.uses_bindless_samplers : false;
902 }
903
904 static inline bool si_shader_uses_bindless_images(struct si_shader_selector *selector)
905 {
906 return selector ? selector->info.uses_bindless_images : false;
907 }
908
909 #endif