c4792da4782e06190af1bb6231a2f66fd33019b7
[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 <llvm-c/Core.h> /* LLVMModuleRef */
135 #include <llvm-c/TargetMachine.h>
136 #include "util/u_inlines.h"
137 #include "util/u_queue.h"
138 #include "util/simple_mtx.h"
139
140 #include "ac_binary.h"
141 #include "ac_llvm_build.h"
142 #include "ac_llvm_util.h"
143
144 #include <stdio.h>
145
146 // Use LDS symbols when supported by LLVM. Can be disabled for testing the old
147 // path on newer LLVM for now. Should be removed in the long term.
148 #define USE_LDS_SYMBOLS (true)
149
150 struct nir_shader;
151 struct si_shader;
152 struct si_context;
153
154 #define SI_MAX_ATTRIBS 16
155 #define SI_MAX_VS_OUTPUTS 40
156
157 /* Shader IO unique indices are supported for TGSI_SEMANTIC_GENERIC with an
158 * index smaller than this.
159 */
160 #define SI_MAX_IO_GENERIC 32
161
162 /* SGPR user data indices */
163 enum {
164 SI_SGPR_RW_BUFFERS, /* rings (& stream-out, VS only) */
165 SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES,
166 SI_SGPR_CONST_AND_SHADER_BUFFERS, /* or just a constant buffer 0 pointer */
167 SI_SGPR_SAMPLERS_AND_IMAGES,
168 SI_NUM_RESOURCE_SGPRS,
169
170 /* API VS, TES without GS, GS copy shader */
171 SI_SGPR_VS_STATE_BITS = SI_NUM_RESOURCE_SGPRS,
172 SI_NUM_VS_STATE_RESOURCE_SGPRS,
173
174 /* all VS variants */
175 SI_SGPR_BASE_VERTEX = SI_NUM_VS_STATE_RESOURCE_SGPRS,
176 SI_SGPR_START_INSTANCE,
177 SI_SGPR_DRAWID,
178 SI_VS_NUM_USER_SGPR,
179
180 SI_SGPR_VS_BLIT_DATA = SI_SGPR_CONST_AND_SHADER_BUFFERS,
181
182 /* TES */
183 SI_SGPR_TES_OFFCHIP_LAYOUT = SI_NUM_VS_STATE_RESOURCE_SGPRS,
184 SI_SGPR_TES_OFFCHIP_ADDR,
185 SI_TES_NUM_USER_SGPR,
186
187 /* GFX6-8: TCS only */
188 GFX6_SGPR_TCS_OFFCHIP_LAYOUT = SI_NUM_RESOURCE_SGPRS,
189 GFX6_SGPR_TCS_OUT_OFFSETS,
190 GFX6_SGPR_TCS_OUT_LAYOUT,
191 GFX6_SGPR_TCS_IN_LAYOUT,
192 GFX6_TCS_NUM_USER_SGPR,
193
194 /* GFX9: Merged shaders. */
195 /* 2ND_CONST_AND_SHADER_BUFFERS is set in USER_DATA_ADDR_LO (SGPR0). */
196 /* 2ND_SAMPLERS_AND_IMAGES is set in USER_DATA_ADDR_HI (SGPR1). */
197 GFX9_MERGED_NUM_USER_SGPR = SI_VS_NUM_USER_SGPR,
198
199 /* GFX9: Merged LS-HS (VS-TCS) only. */
200 GFX9_SGPR_TCS_OFFCHIP_LAYOUT = GFX9_MERGED_NUM_USER_SGPR,
201 GFX9_SGPR_TCS_OUT_OFFSETS,
202 GFX9_SGPR_TCS_OUT_LAYOUT,
203 GFX9_TCS_NUM_USER_SGPR,
204
205 /* GS limits */
206 GFX6_GS_NUM_USER_SGPR = SI_NUM_RESOURCE_SGPRS,
207 GFX9_VSGS_NUM_USER_SGPR = SI_VS_NUM_USER_SGPR,
208 GFX9_TESGS_NUM_USER_SGPR = SI_TES_NUM_USER_SGPR,
209 SI_GSCOPY_NUM_USER_SGPR = SI_NUM_VS_STATE_RESOURCE_SGPRS,
210
211 /* PS only */
212 SI_SGPR_ALPHA_REF = SI_NUM_RESOURCE_SGPRS,
213 SI_PS_NUM_USER_SGPR,
214
215 /* The value has to be 12, because the hw requires that descriptors
216 * are aligned to 4 SGPRs.
217 */
218 SI_SGPR_VS_VB_DESCRIPTOR_FIRST = 12,
219 };
220
221 /* LLVM function parameter indices */
222 enum {
223 SI_NUM_RESOURCE_PARAMS = 4,
224
225 /* PS only parameters */
226 SI_PARAM_ALPHA_REF = SI_NUM_RESOURCE_PARAMS,
227 SI_PARAM_PRIM_MASK,
228 SI_PARAM_PERSP_SAMPLE,
229 SI_PARAM_PERSP_CENTER,
230 SI_PARAM_PERSP_CENTROID,
231 SI_PARAM_PERSP_PULL_MODEL,
232 SI_PARAM_LINEAR_SAMPLE,
233 SI_PARAM_LINEAR_CENTER,
234 SI_PARAM_LINEAR_CENTROID,
235 SI_PARAM_LINE_STIPPLE_TEX,
236 SI_PARAM_POS_X_FLOAT,
237 SI_PARAM_POS_Y_FLOAT,
238 SI_PARAM_POS_Z_FLOAT,
239 SI_PARAM_POS_W_FLOAT,
240 SI_PARAM_FRONT_FACE,
241 SI_PARAM_ANCILLARY,
242 SI_PARAM_SAMPLE_COVERAGE,
243 SI_PARAM_POS_FIXED_PT,
244
245 SI_NUM_PARAMS = SI_PARAM_POS_FIXED_PT + 9, /* +8 for COLOR[0..1] */
246 };
247
248 /* Fields of driver-defined VS state SGPR. */
249 #define S_VS_STATE_CLAMP_VERTEX_COLOR(x) (((unsigned)(x) & 0x1) << 0)
250 #define C_VS_STATE_CLAMP_VERTEX_COLOR 0xFFFFFFFE
251 #define S_VS_STATE_INDEXED(x) (((unsigned)(x) & 0x1) << 1)
252 #define C_VS_STATE_INDEXED 0xFFFFFFFD
253 #define S_VS_STATE_OUTPRIM(x) (((unsigned)(x) & 0x3) << 2)
254 #define C_VS_STATE_OUTPRIM 0xFFFFFFF3
255 #define S_VS_STATE_PROVOKING_VTX_INDEX(x) (((unsigned)(x) & 0x3) << 4)
256 #define C_VS_STATE_PROVOKING_VTX_INDEX 0xFFFFFFCF
257 #define S_VS_STATE_STREAMOUT_QUERY_ENABLED(x) (((unsigned)(x) & 0x1) << 6)
258 #define C_VS_STATE_STREAMOUT_QUERY_ENABLED 0xFFFFFFBF
259 #define S_VS_STATE_LS_OUT_PATCH_SIZE(x) (((unsigned)(x) & 0x1FFF) << 8)
260 #define C_VS_STATE_LS_OUT_PATCH_SIZE 0xFFE000FF
261 #define S_VS_STATE_LS_OUT_VERTEX_SIZE(x) (((unsigned)(x) & 0xFF) << 24)
262 #define C_VS_STATE_LS_OUT_VERTEX_SIZE 0x00FFFFFF
263
264 enum {
265 /* Use a property enum that CS wouldn't use. */
266 TGSI_PROPERTY_CS_LOCAL_SIZE = TGSI_PROPERTY_FS_COORD_ORIGIN,
267
268 /* These represent the number of SGPRs the shader uses. */
269 SI_VS_BLIT_SGPRS_POS = 3,
270 SI_VS_BLIT_SGPRS_POS_COLOR = 7,
271 SI_VS_BLIT_SGPRS_POS_TEXCOORD = 9,
272 };
273
274 /**
275 * For VS shader keys, describe any fixups required for vertex fetch.
276 *
277 * \ref log_size, \ref format, and the number of channels are interpreted as
278 * by \ref ac_build_opencoded_load_format.
279 *
280 * Note: all bits 0 (size = 1 byte, num channels = 1, format = float) is an
281 * impossible format and indicates that no fixup is needed (just use
282 * buffer_load_format_xyzw).
283 */
284 union si_vs_fix_fetch {
285 struct {
286 uint8_t log_size : 2; /* 1, 2, 4, 8 or bytes per channel */
287 uint8_t num_channels_m1 : 2; /* number of channels minus 1 */
288 uint8_t format : 3; /* AC_FETCH_FORMAT_xxx */
289 uint8_t reverse : 1; /* reverse XYZ channels */
290 } u;
291 uint8_t bits;
292 };
293
294 struct si_shader;
295
296 /* State of the context creating the shader object. */
297 struct si_compiler_ctx_state {
298 /* Should only be used by si_init_shader_selector_async and
299 * si_build_shader_variant if thread_index == -1 (non-threaded). */
300 struct ac_llvm_compiler *compiler;
301
302 /* Used if thread_index == -1 or if debug.async is true. */
303 struct pipe_debug_callback debug;
304
305 /* Used for creating the log string for gallium/ddebug. */
306 bool is_debug_context;
307 };
308
309 struct si_shader_info {
310 ubyte num_inputs;
311 ubyte num_outputs;
312 ubyte input_semantic_name[PIPE_MAX_SHADER_INPUTS]; /**< TGSI_SEMANTIC_x */
313 ubyte input_semantic_index[PIPE_MAX_SHADER_INPUTS];
314 ubyte input_interpolate[PIPE_MAX_SHADER_INPUTS];
315 ubyte input_interpolate_loc[PIPE_MAX_SHADER_INPUTS];
316 ubyte input_usage_mask[PIPE_MAX_SHADER_INPUTS];
317 ubyte output_semantic_name[PIPE_MAX_SHADER_OUTPUTS]; /**< TGSI_SEMANTIC_x */
318 ubyte output_semantic_index[PIPE_MAX_SHADER_OUTPUTS];
319 ubyte output_usagemask[PIPE_MAX_SHADER_OUTPUTS];
320 ubyte output_streams[PIPE_MAX_SHADER_OUTPUTS];
321
322 ubyte processor;
323
324 int constbuf0_num_slots;
325 unsigned const_buffers_declared; /**< bitmask of declared const buffers */
326 unsigned samplers_declared; /**< bitmask of declared samplers */
327 ubyte num_stream_output_components[4];
328
329 uint num_memory_instructions; /**< sampler, buffer, and image instructions */
330
331 /**
332 * If a tessellation control shader reads outputs, this describes which ones.
333 */
334 bool reads_pervertex_outputs;
335 bool reads_perpatch_outputs;
336 bool reads_tessfactor_outputs;
337
338 ubyte colors_read; /**< which color components are read by the FS */
339 ubyte colors_written;
340 bool reads_samplemask; /**< does fragment shader read sample mask? */
341 bool reads_tess_factors; /**< If TES reads TESSINNER or TESSOUTER */
342 bool writes_z; /**< does fragment shader write Z value? */
343 bool writes_stencil; /**< does fragment shader write stencil value? */
344 bool writes_samplemask; /**< does fragment shader write sample mask? */
345 bool writes_edgeflag; /**< vertex shader outputs edgeflag */
346 bool uses_kill; /**< KILL or KILL_IF instruction used? */
347 bool uses_persp_center;
348 bool uses_persp_centroid;
349 bool uses_persp_sample;
350 bool uses_linear_center;
351 bool uses_linear_centroid;
352 bool uses_linear_sample;
353 bool uses_persp_opcode_interp_sample;
354 bool uses_linear_opcode_interp_sample;
355 bool uses_instanceid;
356 bool uses_vertexid;
357 bool uses_vertexid_nobase;
358 bool uses_basevertex;
359 bool uses_drawid;
360 bool uses_primid;
361 bool uses_frontface;
362 bool uses_invocationid;
363 bool uses_thread_id[3];
364 bool uses_block_id[3];
365 bool uses_block_size;
366 bool uses_grid_size;
367 bool uses_subgroup_info;
368 bool writes_position;
369 bool writes_psize;
370 bool writes_clipvertex;
371 bool writes_primid;
372 bool writes_viewport_index;
373 bool writes_layer;
374 bool writes_memory; /**< contains stores or atomics to buffers or images */
375 bool uses_derivatives;
376 bool uses_bindless_samplers;
377 bool uses_bindless_images;
378 bool uses_fbfetch;
379 unsigned clipdist_writemask;
380 unsigned culldist_writemask;
381 unsigned num_written_culldistance;
382 unsigned num_written_clipdistance;
383
384 unsigned images_declared; /**< bitmask of declared images */
385 unsigned msaa_images_declared; /**< bitmask of declared MSAA images */
386 unsigned shader_buffers_declared; /**< bitmask of declared shader buffers */
387
388 unsigned properties[TGSI_PROPERTY_COUNT]; /* index with TGSI_PROPERTY_ */
389
390 /** Whether all codepaths write tess factors in all invocations. */
391 bool tessfactors_are_def_in_all_invocs;
392 };
393
394 /* A shader selector is a gallium CSO and contains shader variants and
395 * binaries for one NIR program. This can be shared by multiple contexts.
396 */
397 struct si_shader_selector {
398 struct pipe_reference reference;
399 struct si_screen *screen;
400 struct util_queue_fence ready;
401 struct si_compiler_ctx_state compiler_ctx_state;
402
403 simple_mtx_t mutex;
404 struct si_shader *first_variant; /* immutable after the first variant */
405 struct si_shader *last_variant; /* mutable */
406
407 /* The compiled NIR shader without a prolog and/or epilog (not
408 * uploaded to a buffer object).
409 */
410 struct si_shader *main_shader_part;
411 struct si_shader *main_shader_part_ls; /* as_ls is set in the key */
412 struct si_shader *main_shader_part_es; /* as_es is set in the key */
413 struct si_shader *main_shader_part_ngg; /* as_ngg is set in the key */
414 struct si_shader *main_shader_part_ngg_es; /* for Wave32 TES before legacy GS */
415
416 struct si_shader *gs_copy_shader;
417
418 struct nir_shader *nir;
419 void *nir_binary;
420 unsigned nir_size;
421
422 struct pipe_stream_output_info so;
423 struct si_shader_info info;
424
425 /* PIPE_SHADER_[VERTEX|FRAGMENT|...] */
426 enum pipe_shader_type type;
427 bool vs_needs_prolog;
428 bool force_correct_derivs_after_kill;
429 bool prim_discard_cs_allowed;
430 unsigned num_vs_inputs;
431 unsigned num_vbos_in_user_sgprs;
432 unsigned pa_cl_vs_out_cntl;
433 ubyte clipdist_mask;
434 ubyte culldist_mask;
435 unsigned rast_prim;
436
437 /* ES parameters. */
438 unsigned esgs_itemsize; /* vertex stride */
439 unsigned lshs_vertex_stride;
440
441 /* GS parameters. */
442 unsigned gs_input_verts_per_prim;
443 unsigned gs_output_prim;
444 unsigned gs_max_out_vertices;
445 unsigned gs_num_invocations;
446 unsigned max_gs_stream; /* count - 1 */
447 unsigned gsvs_vertex_size;
448 unsigned max_gsvs_emit_size;
449 unsigned enabled_streamout_buffer_mask;
450 bool tess_turns_off_ngg;
451
452 /* PS parameters. */
453 unsigned color_attr_index[2];
454 unsigned db_shader_control;
455 /* Set 0xf or 0x0 (4 bits) per each written output.
456 * ANDed with spi_shader_col_format.
457 */
458 unsigned colors_written_4bit;
459
460 uint64_t outputs_written_before_ps; /* "get_unique_index" bits */
461 uint64_t outputs_written; /* "get_unique_index" bits */
462 uint32_t patch_outputs_written; /* "get_unique_index_patch" bits */
463
464 uint64_t inputs_read; /* "get_unique_index" bits */
465
466 /* bitmasks of used descriptor slots */
467 uint32_t active_const_and_shader_buffers;
468 uint64_t active_samplers_and_images;
469 };
470
471 /* Valid shader configurations:
472 *
473 * API shaders VS | TCS | TES | GS |pass| PS
474 * are compiled as: | | | |thru|
475 * | | | | |
476 * Only VS & PS: VS | | | | | PS
477 * GFX6 - with GS: ES | | | GS | VS | PS
478 * - with tess: LS | HS | VS | | | PS
479 * - with both: LS | HS | ES | GS | VS | PS
480 * GFX9 - with GS: -> | | | GS | VS | PS
481 * - with tess: -> | HS | VS | | | PS
482 * - with both: -> | HS | -> | GS | VS | PS
483 * | | | | |
484 * NGG - VS & PS: GS | | | | | PS
485 * (GFX10+) - with GS: -> | | | GS | | PS
486 * - with tess: -> | HS | GS | | | PS
487 * - with both: -> | HS | -> | GS | | PS
488 *
489 * -> = merged with the next stage
490 */
491
492 /* Use the byte alignment for all following structure members for optimal
493 * shader key memory footprint.
494 */
495 #pragma pack(push, 1)
496
497 /* Common VS bits between the shader key and the prolog key. */
498 struct si_vs_prolog_bits {
499 /* - If neither "is_one" nor "is_fetched" has a bit set, the instance
500 * divisor is 0.
501 * - If "is_one" has a bit set, the instance divisor is 1.
502 * - If "is_fetched" has a bit set, the instance divisor will be loaded
503 * from the constant buffer.
504 */
505 uint16_t instance_divisor_is_one; /* bitmask of inputs */
506 uint16_t instance_divisor_is_fetched; /* bitmask of inputs */
507 unsigned ls_vgpr_fix:1;
508 unsigned unpack_instance_id_from_vertex_id:1;
509 };
510
511 /* Common TCS bits between the shader key and the epilog key. */
512 struct si_tcs_epilog_bits {
513 unsigned prim_mode:3;
514 unsigned invoc0_tess_factors_are_def:1;
515 unsigned tes_reads_tess_factors:1;
516 };
517
518 struct si_gs_prolog_bits {
519 unsigned tri_strip_adj_fix:1;
520 unsigned gfx9_prev_is_vs:1;
521 };
522
523 /* Common PS bits between the shader key and the prolog key. */
524 struct si_ps_prolog_bits {
525 unsigned color_two_side:1;
526 unsigned flatshade_colors:1;
527 unsigned poly_stipple:1;
528 unsigned force_persp_sample_interp:1;
529 unsigned force_linear_sample_interp:1;
530 unsigned force_persp_center_interp:1;
531 unsigned force_linear_center_interp:1;
532 unsigned bc_optimize_for_persp:1;
533 unsigned bc_optimize_for_linear:1;
534 unsigned samplemask_log_ps_iter:3;
535 };
536
537 /* Common PS bits between the shader key and the epilog key. */
538 struct si_ps_epilog_bits {
539 unsigned spi_shader_col_format;
540 unsigned color_is_int8:8;
541 unsigned color_is_int10:8;
542 unsigned last_cbuf:3;
543 unsigned alpha_func:3;
544 unsigned alpha_to_one:1;
545 unsigned poly_line_smoothing:1;
546 unsigned clamp_color:1;
547 };
548
549 union si_shader_part_key {
550 struct {
551 struct si_vs_prolog_bits states;
552 unsigned num_input_sgprs:6;
553 /* For merged stages such as LS-HS, HS input VGPRs are first. */
554 unsigned num_merged_next_stage_vgprs:3;
555 unsigned num_inputs:5;
556 unsigned as_ls:1;
557 unsigned as_es:1;
558 unsigned as_ngg:1;
559 /* Prologs for monolithic shaders shouldn't set EXEC. */
560 unsigned is_monolithic:1;
561 } vs_prolog;
562 struct {
563 struct si_tcs_epilog_bits states;
564 } tcs_epilog;
565 struct {
566 struct si_gs_prolog_bits states;
567 /* Prologs of monolithic shaders shouldn't set EXEC. */
568 unsigned is_monolithic:1;
569 unsigned as_ngg:1;
570 } gs_prolog;
571 struct {
572 struct si_ps_prolog_bits states;
573 unsigned num_input_sgprs:6;
574 unsigned num_input_vgprs:5;
575 /* Color interpolation and two-side color selection. */
576 unsigned colors_read:8; /* color input components read */
577 unsigned num_interp_inputs:5; /* BCOLOR is at this location */
578 unsigned face_vgpr_index:5;
579 unsigned ancillary_vgpr_index:5;
580 unsigned wqm:1;
581 char color_attr_index[2];
582 signed char color_interp_vgpr_index[2]; /* -1 == constant */
583 } ps_prolog;
584 struct {
585 struct si_ps_epilog_bits states;
586 unsigned colors_written:8;
587 unsigned writes_z:1;
588 unsigned writes_stencil:1;
589 unsigned writes_samplemask:1;
590 } ps_epilog;
591 };
592
593 struct si_shader_key {
594 /* Prolog and epilog flags. */
595 union {
596 struct {
597 struct si_vs_prolog_bits prolog;
598 } vs;
599 struct {
600 struct si_vs_prolog_bits ls_prolog; /* for merged LS-HS */
601 struct si_shader_selector *ls; /* for merged LS-HS */
602 struct si_tcs_epilog_bits epilog;
603 } tcs; /* tessellation control shader */
604 struct {
605 struct si_vs_prolog_bits vs_prolog; /* for merged ES-GS */
606 struct si_shader_selector *es; /* for merged ES-GS */
607 struct si_gs_prolog_bits prolog;
608 } gs;
609 struct {
610 struct si_ps_prolog_bits prolog;
611 struct si_ps_epilog_bits epilog;
612 } ps;
613 } part;
614
615 /* These three are initially set according to the NEXT_SHADER property,
616 * or guessed if the property doesn't seem correct.
617 */
618 unsigned as_es:1; /* export shader, which precedes GS */
619 unsigned as_ls:1; /* local shader, which precedes TCS */
620 unsigned as_ngg:1; /* VS, TES, or GS compiled as NGG primitive shader */
621
622 /* Flags for monolithic compilation only. */
623 struct {
624 /* Whether fetch should be opencoded according to vs_fix_fetch.
625 * Otherwise, if vs_fix_fetch is non-zero, buffer_load_format_xyzw
626 * with minimal fixups is used. */
627 uint16_t vs_fetch_opencode;
628 union si_vs_fix_fetch vs_fix_fetch[SI_MAX_ATTRIBS];
629
630 union {
631 uint64_t ff_tcs_inputs_to_copy; /* for fixed-func TCS */
632 /* When PS needs PrimID and GS is disabled. */
633 unsigned vs_export_prim_id:1;
634 struct {
635 unsigned interpolate_at_sample_force_center:1;
636 unsigned fbfetch_msaa:1;
637 unsigned fbfetch_is_1D:1;
638 unsigned fbfetch_layered:1;
639 } ps;
640 } u;
641 } mono;
642
643 /* Optimization flags for asynchronous compilation only. */
644 struct {
645 /* For HW VS (it can be VS, TES, GS) */
646 uint64_t kill_outputs; /* "get_unique_index" bits */
647 unsigned clip_disable:1;
648
649 /* For shaders where monolithic variants have better code.
650 *
651 * This is a flag that has no effect on code generation,
652 * but forces monolithic shaders to be used as soon as
653 * possible, because it's in the "opt" group.
654 */
655 unsigned prefer_mono:1;
656
657 /* Primitive discard compute shader. */
658 unsigned vs_as_prim_discard_cs:1;
659 unsigned cs_prim_type:4;
660 unsigned cs_indexed:1;
661 unsigned cs_instancing:1;
662 unsigned cs_primitive_restart:1;
663 unsigned cs_provoking_vertex_first:1;
664 unsigned cs_need_correct_orientation:1;
665 unsigned cs_cull_front:1;
666 unsigned cs_cull_back:1;
667 unsigned cs_cull_z:1;
668 unsigned cs_halfz_clip_space:1;
669 } opt;
670 };
671
672 /* Restore the pack alignment to default. */
673 #pragma pack(pop)
674
675 /* GCN-specific shader info. */
676 struct si_shader_binary_info {
677 ubyte vs_output_param_offset[SI_MAX_VS_OUTPUTS];
678 ubyte num_input_sgprs;
679 ubyte num_input_vgprs;
680 signed char face_vgpr_index;
681 signed char ancillary_vgpr_index;
682 bool uses_instanceid;
683 ubyte nr_pos_exports;
684 ubyte nr_param_exports;
685 unsigned private_mem_vgprs;
686 unsigned max_simd_waves;
687 };
688
689 struct si_shader_binary {
690 const char *elf_buffer;
691 size_t elf_size;
692
693 char *llvm_ir_string;
694 };
695
696 struct gfx9_gs_info {
697 unsigned es_verts_per_subgroup;
698 unsigned gs_prims_per_subgroup;
699 unsigned gs_inst_prims_in_subgroup;
700 unsigned max_prims_per_subgroup;
701 unsigned esgs_ring_size; /* in bytes */
702 };
703
704 struct si_shader {
705 struct si_compiler_ctx_state compiler_ctx_state;
706
707 struct si_shader_selector *selector;
708 struct si_shader_selector *previous_stage_sel; /* for refcounting */
709 struct si_shader *next_variant;
710
711 struct si_shader_part *prolog;
712 struct si_shader *previous_stage; /* for GFX9 */
713 struct si_shader_part *prolog2;
714 struct si_shader_part *epilog;
715
716 struct si_pm4_state *pm4;
717 struct si_resource *bo;
718 struct si_resource *scratch_bo;
719 struct si_shader_key key;
720 struct util_queue_fence ready;
721 bool compilation_failed;
722 bool is_monolithic;
723 bool is_optimized;
724 bool is_binary_shared;
725 bool is_gs_copy_shader;
726
727 /* The following data is all that's needed for binary shaders. */
728 struct si_shader_binary binary;
729 struct ac_shader_config config;
730 struct si_shader_binary_info info;
731
732 struct {
733 uint16_t ngg_emit_size; /* in dwords */
734 uint16_t hw_max_esverts;
735 uint16_t max_gsprims;
736 uint16_t max_out_verts;
737 uint16_t prim_amp_factor;
738 bool max_vert_out_per_gs_instance;
739 } ngg;
740
741 /* Shader key + LLVM IR + disassembly + statistics.
742 * Generated for debug contexts only.
743 */
744 char *shader_log;
745 size_t shader_log_size;
746
747 struct gfx9_gs_info gs_info;
748
749 /* For save precompute context registers values. */
750 union {
751 struct {
752 unsigned vgt_gsvs_ring_offset_1;
753 unsigned vgt_gsvs_ring_offset_2;
754 unsigned vgt_gsvs_ring_offset_3;
755 unsigned vgt_gsvs_ring_itemsize;
756 unsigned vgt_gs_max_vert_out;
757 unsigned vgt_gs_vert_itemsize;
758 unsigned vgt_gs_vert_itemsize_1;
759 unsigned vgt_gs_vert_itemsize_2;
760 unsigned vgt_gs_vert_itemsize_3;
761 unsigned vgt_gs_instance_cnt;
762 unsigned vgt_gs_onchip_cntl;
763 unsigned vgt_gs_max_prims_per_subgroup;
764 unsigned vgt_esgs_ring_itemsize;
765 } gs;
766
767 struct {
768 unsigned ge_max_output_per_subgroup;
769 unsigned ge_ngg_subgrp_cntl;
770 unsigned vgt_primitiveid_en;
771 unsigned vgt_gs_onchip_cntl;
772 unsigned vgt_gs_instance_cnt;
773 unsigned vgt_esgs_ring_itemsize;
774 unsigned spi_vs_out_config;
775 unsigned spi_shader_idx_format;
776 unsigned spi_shader_pos_format;
777 unsigned pa_cl_vte_cntl;
778 unsigned pa_cl_ngg_cntl;
779 unsigned vgt_gs_max_vert_out; /* for API GS */
780 } ngg;
781
782 struct {
783 unsigned vgt_gs_mode;
784 unsigned vgt_primitiveid_en;
785 unsigned vgt_reuse_off;
786 unsigned spi_vs_out_config;
787 unsigned spi_shader_pos_format;
788 unsigned pa_cl_vte_cntl;
789 } vs;
790
791 struct {
792 unsigned spi_ps_input_ena;
793 unsigned spi_ps_input_addr;
794 unsigned spi_baryc_cntl;
795 unsigned spi_ps_in_control;
796 unsigned spi_shader_z_format;
797 unsigned spi_shader_col_format;
798 unsigned cb_shader_mask;
799 } ps;
800 } ctx_reg;
801
802 /*For save precompute registers value */
803 unsigned vgt_tf_param; /* VGT_TF_PARAM */
804 unsigned vgt_vertex_reuse_block_cntl; /* VGT_VERTEX_REUSE_BLOCK_CNTL */
805 unsigned pa_cl_vs_out_cntl;
806 unsigned ge_cntl;
807 };
808
809 struct si_shader_part {
810 struct si_shader_part *next;
811 union si_shader_part_key key;
812 struct si_shader_binary binary;
813 struct ac_shader_config config;
814 };
815
816 /* si_shader.c */
817 struct si_shader *
818 si_generate_gs_copy_shader(struct si_screen *sscreen,
819 struct ac_llvm_compiler *compiler,
820 struct si_shader_selector *gs_selector,
821 struct pipe_debug_callback *debug);
822 int si_compile_shader(struct si_screen *sscreen,
823 struct ac_llvm_compiler *compiler,
824 struct si_shader *shader,
825 struct pipe_debug_callback *debug);
826 bool si_create_shader_variant(struct si_screen *sscreen,
827 struct ac_llvm_compiler *compiler,
828 struct si_shader *shader,
829 struct pipe_debug_callback *debug);
830 void si_shader_destroy(struct si_shader *shader);
831 unsigned si_shader_io_get_unique_index_patch(unsigned semantic_name, unsigned index);
832 unsigned si_shader_io_get_unique_index(unsigned semantic_name, unsigned index,
833 unsigned is_varying);
834 bool si_shader_binary_upload(struct si_screen *sscreen, struct si_shader *shader,
835 uint64_t scratch_va);
836 void si_shader_dump(struct si_screen *sscreen, struct si_shader *shader,
837 struct pipe_debug_callback *debug,
838 FILE *f, bool check_debug_option);
839 void si_shader_dump_stats_for_shader_db(struct si_screen *screen,
840 struct si_shader *shader,
841 struct pipe_debug_callback *debug);
842 void si_multiwave_lds_size_workaround(struct si_screen *sscreen,
843 unsigned *lds_size);
844 const char *si_get_shader_name(const struct si_shader *shader);
845 void si_shader_binary_clean(struct si_shader_binary *binary);
846
847 /* si_shader_nir.c */
848 void si_nir_scan_shader(const struct nir_shader *nir,
849 struct si_shader_info *info);
850 void si_nir_adjust_driver_locations(struct nir_shader *nir);
851 void si_finalize_nir(struct pipe_screen *screen, void *nirptr, bool optimize);
852
853 /* si_state_shaders.c */
854 void gfx9_get_gs_info(struct si_shader_selector *es,
855 struct si_shader_selector *gs,
856 struct gfx9_gs_info *out);
857
858 /* Inline helpers. */
859
860 /* Return the pointer to the main shader part's pointer. */
861 static inline struct si_shader **
862 si_get_main_shader_part(struct si_shader_selector *sel,
863 struct si_shader_key *key)
864 {
865 if (key->as_ls)
866 return &sel->main_shader_part_ls;
867 if (key->as_es && key->as_ngg)
868 return &sel->main_shader_part_ngg_es;
869 if (key->as_es)
870 return &sel->main_shader_part_es;
871 if (key->as_ngg)
872 return &sel->main_shader_part_ngg;
873 return &sel->main_shader_part;
874 }
875
876 static inline bool
877 gfx10_is_ngg_passthrough(struct si_shader *shader)
878 {
879 struct si_shader_selector *sel = shader->selector;
880
881 return sel->type != PIPE_SHADER_GEOMETRY &&
882 !sel->so.num_outputs &&
883 !sel->info.writes_edgeflag &&
884 (sel->type != PIPE_SHADER_VERTEX ||
885 !shader->key.mono.u.vs_export_prim_id);
886 }
887
888 static inline bool
889 si_shader_uses_bindless_samplers(struct si_shader_selector *selector)
890 {
891 return selector ? selector->info.uses_bindless_samplers : false;
892 }
893
894 static inline bool
895 si_shader_uses_bindless_images(struct si_shader_selector *selector)
896 {
897 return selector ? selector->info.uses_bindless_images : false;
898 }
899
900 void si_destroy_shader_selector(struct si_context *sctx,
901 struct si_shader_selector *sel);
902
903 static inline void
904 si_shader_selector_reference(struct si_context *sctx,
905 struct si_shader_selector **dst,
906 struct si_shader_selector *src)
907 {
908 if (pipe_reference(&(*dst)->reference, &src->reference))
909 si_destroy_shader_selector(sctx, *dst);
910
911 *dst = src;
912 }
913
914 #endif