radeonsi/gfx10: remove incorrect ngg/pos_writes_edgeflag variables
[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 "tgsi/tgsi_scan.h"
137 #include "util/u_inlines.h"
138 #include "util/u_queue.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
216 /* LLVM function parameter indices */
217 enum {
218 SI_NUM_RESOURCE_PARAMS = 4,
219
220 /* PS only parameters */
221 SI_PARAM_ALPHA_REF = SI_NUM_RESOURCE_PARAMS,
222 SI_PARAM_PRIM_MASK,
223 SI_PARAM_PERSP_SAMPLE,
224 SI_PARAM_PERSP_CENTER,
225 SI_PARAM_PERSP_CENTROID,
226 SI_PARAM_PERSP_PULL_MODEL,
227 SI_PARAM_LINEAR_SAMPLE,
228 SI_PARAM_LINEAR_CENTER,
229 SI_PARAM_LINEAR_CENTROID,
230 SI_PARAM_LINE_STIPPLE_TEX,
231 SI_PARAM_POS_X_FLOAT,
232 SI_PARAM_POS_Y_FLOAT,
233 SI_PARAM_POS_Z_FLOAT,
234 SI_PARAM_POS_W_FLOAT,
235 SI_PARAM_FRONT_FACE,
236 SI_PARAM_ANCILLARY,
237 SI_PARAM_SAMPLE_COVERAGE,
238 SI_PARAM_POS_FIXED_PT,
239
240 SI_NUM_PARAMS = SI_PARAM_POS_FIXED_PT + 9, /* +8 for COLOR[0..1] */
241 };
242
243 /* Fields of driver-defined VS state SGPR. */
244 #define S_VS_STATE_CLAMP_VERTEX_COLOR(x) (((unsigned)(x) & 0x1) << 0)
245 #define C_VS_STATE_CLAMP_VERTEX_COLOR 0xFFFFFFFE
246 #define S_VS_STATE_INDEXED(x) (((unsigned)(x) & 0x1) << 1)
247 #define C_VS_STATE_INDEXED 0xFFFFFFFD
248 #define S_VS_STATE_OUTPRIM(x) (((unsigned)(x) & 0x3) << 2)
249 #define C_VS_STATE_OUTPRIM 0xFFFFFFF3
250 #define S_VS_STATE_PROVOKING_VTX_INDEX(x) (((unsigned)(x) & 0x3) << 4)
251 #define C_VS_STATE_PROVOKING_VTX_INDEX 0xFFFFFFCF
252 #define S_VS_STATE_STREAMOUT_QUERY_ENABLED(x) (((unsigned)(x) & 0x1) << 6)
253 #define C_VS_STATE_STREAMOUT_QUERY_ENABLED 0xFFFFFFBF
254 #define S_VS_STATE_LS_OUT_PATCH_SIZE(x) (((unsigned)(x) & 0x1FFF) << 8)
255 #define C_VS_STATE_LS_OUT_PATCH_SIZE 0xFFE000FF
256 #define S_VS_STATE_LS_OUT_VERTEX_SIZE(x) (((unsigned)(x) & 0xFF) << 24)
257 #define C_VS_STATE_LS_OUT_VERTEX_SIZE 0x00FFFFFF
258
259 enum {
260 /* Use a property enum that CS wouldn't use. */
261 TGSI_PROPERTY_CS_LOCAL_SIZE = TGSI_PROPERTY_FS_COORD_ORIGIN,
262
263 /* These represent the number of SGPRs the shader uses. */
264 SI_VS_BLIT_SGPRS_POS = 3,
265 SI_VS_BLIT_SGPRS_POS_COLOR = 7,
266 SI_VS_BLIT_SGPRS_POS_TEXCOORD = 9,
267 };
268
269 /**
270 * For VS shader keys, describe any fixups required for vertex fetch.
271 *
272 * \ref log_size, \ref format, and the number of channels are interpreted as
273 * by \ref ac_build_opencoded_load_format.
274 *
275 * Note: all bits 0 (size = 1 byte, num channels = 1, format = float) is an
276 * impossible format and indicates that no fixup is needed (just use
277 * buffer_load_format_xyzw).
278 */
279 union si_vs_fix_fetch {
280 struct {
281 uint8_t log_size : 2; /* 1, 2, 4, 8 or bytes per channel */
282 uint8_t num_channels_m1 : 2; /* number of channels minus 1 */
283 uint8_t format : 3; /* AC_FETCH_FORMAT_xxx */
284 uint8_t reverse : 1; /* reverse XYZ channels */
285 } u;
286 uint8_t bits;
287 };
288
289 struct si_shader;
290
291 /* State of the context creating the shader object. */
292 struct si_compiler_ctx_state {
293 /* Should only be used by si_init_shader_selector_async and
294 * si_build_shader_variant if thread_index == -1 (non-threaded). */
295 struct ac_llvm_compiler *compiler;
296
297 /* Used if thread_index == -1 or if debug.async is true. */
298 struct pipe_debug_callback debug;
299
300 /* Used for creating the log string for gallium/ddebug. */
301 bool is_debug_context;
302 };
303
304 /* A shader selector is a gallium CSO and contains shader variants and
305 * binaries for one TGSI program. This can be shared by multiple contexts.
306 */
307 struct si_shader_selector {
308 struct pipe_reference reference;
309 struct si_screen *screen;
310 struct util_queue_fence ready;
311 struct si_compiler_ctx_state compiler_ctx_state;
312
313 mtx_t mutex;
314 struct si_shader *first_variant; /* immutable after the first variant */
315 struct si_shader *last_variant; /* mutable */
316
317 /* The compiled TGSI shader expecting a prolog and/or epilog (not
318 * uploaded to a buffer).
319 */
320 struct si_shader *main_shader_part;
321 struct si_shader *main_shader_part_ls; /* as_ls is set in the key */
322 struct si_shader *main_shader_part_es; /* as_es is set in the key */
323 struct si_shader *main_shader_part_ngg; /* as_ngg is set in the key */
324 struct si_shader *main_shader_part_ngg_es; /* for Wave32 TES before legacy GS */
325
326 struct si_shader *gs_copy_shader;
327
328 struct tgsi_token *tokens;
329 struct nir_shader *nir;
330 struct pipe_stream_output_info so;
331 struct tgsi_shader_info info;
332 struct tgsi_tessctrl_info tcs_info;
333
334 /* PIPE_SHADER_[VERTEX|FRAGMENT|...] */
335 enum pipe_shader_type type;
336 bool vs_needs_prolog;
337 bool force_correct_derivs_after_kill;
338 bool prim_discard_cs_allowed;
339 unsigned pa_cl_vs_out_cntl;
340 ubyte clipdist_mask;
341 ubyte culldist_mask;
342 unsigned rast_prim;
343
344 /* ES parameters. */
345 unsigned esgs_itemsize; /* vertex stride */
346 unsigned lshs_vertex_stride;
347
348 /* GS parameters. */
349 unsigned gs_input_verts_per_prim;
350 unsigned gs_output_prim;
351 unsigned gs_max_out_vertices;
352 unsigned gs_num_invocations;
353 unsigned max_gs_stream; /* count - 1 */
354 unsigned gsvs_vertex_size;
355 unsigned max_gsvs_emit_size;
356 unsigned enabled_streamout_buffer_mask;
357 bool tess_turns_off_ngg;
358
359 /* PS parameters. */
360 unsigned color_attr_index[2];
361 unsigned db_shader_control;
362 /* Set 0xf or 0x0 (4 bits) per each written output.
363 * ANDed with spi_shader_col_format.
364 */
365 unsigned colors_written_4bit;
366
367 uint64_t outputs_written_before_ps; /* "get_unique_index" bits */
368 uint64_t outputs_written; /* "get_unique_index" bits */
369 uint32_t patch_outputs_written; /* "get_unique_index_patch" bits */
370
371 uint64_t inputs_read; /* "get_unique_index" bits */
372
373 /* bitmasks of used descriptor slots */
374 uint32_t active_const_and_shader_buffers;
375 uint64_t active_samplers_and_images;
376 };
377
378 /* Valid shader configurations:
379 *
380 * API shaders VS | TCS | TES | GS |pass| PS
381 * are compiled as: | | | |thru|
382 * | | | | |
383 * Only VS & PS: VS | | | | | PS
384 * GFX6 - with GS: ES | | | GS | VS | PS
385 * - with tess: LS | HS | VS | | | PS
386 * - with both: LS | HS | ES | GS | VS | PS
387 * GFX9 - with GS: -> | | | GS | VS | PS
388 * - with tess: -> | HS | VS | | | PS
389 * - with both: -> | HS | -> | GS | VS | PS
390 * | | | | |
391 * NGG - VS & PS: GS | | | | | PS
392 * (GFX10+) - with GS: -> | | | GS | | PS
393 * - with tess: -> | HS | GS | | | PS
394 * - with both: -> | HS | -> | GS | | PS
395 *
396 * -> = merged with the next stage
397 */
398
399 /* Use the byte alignment for all following structure members for optimal
400 * shader key memory footprint.
401 */
402 #pragma pack(push, 1)
403
404 /* Common VS bits between the shader key and the prolog key. */
405 struct si_vs_prolog_bits {
406 /* - If neither "is_one" nor "is_fetched" has a bit set, the instance
407 * divisor is 0.
408 * - If "is_one" has a bit set, the instance divisor is 1.
409 * - If "is_fetched" has a bit set, the instance divisor will be loaded
410 * from the constant buffer.
411 */
412 uint16_t instance_divisor_is_one; /* bitmask of inputs */
413 uint16_t instance_divisor_is_fetched; /* bitmask of inputs */
414 unsigned ls_vgpr_fix:1;
415 unsigned unpack_instance_id_from_vertex_id:1;
416 };
417
418 /* Common TCS bits between the shader key and the epilog key. */
419 struct si_tcs_epilog_bits {
420 unsigned prim_mode:3;
421 unsigned invoc0_tess_factors_are_def:1;
422 unsigned tes_reads_tess_factors:1;
423 };
424
425 struct si_gs_prolog_bits {
426 unsigned tri_strip_adj_fix:1;
427 unsigned gfx9_prev_is_vs:1;
428 };
429
430 /* Common PS bits between the shader key and the prolog key. */
431 struct si_ps_prolog_bits {
432 unsigned color_two_side:1;
433 unsigned flatshade_colors:1;
434 unsigned poly_stipple:1;
435 unsigned force_persp_sample_interp:1;
436 unsigned force_linear_sample_interp:1;
437 unsigned force_persp_center_interp:1;
438 unsigned force_linear_center_interp:1;
439 unsigned bc_optimize_for_persp:1;
440 unsigned bc_optimize_for_linear:1;
441 unsigned samplemask_log_ps_iter:3;
442 };
443
444 /* Common PS bits between the shader key and the epilog key. */
445 struct si_ps_epilog_bits {
446 unsigned spi_shader_col_format;
447 unsigned color_is_int8:8;
448 unsigned color_is_int10:8;
449 unsigned last_cbuf:3;
450 unsigned alpha_func:3;
451 unsigned alpha_to_one:1;
452 unsigned poly_line_smoothing:1;
453 unsigned clamp_color:1;
454 };
455
456 union si_shader_part_key {
457 struct {
458 struct si_vs_prolog_bits states;
459 unsigned num_input_sgprs:6;
460 /* For merged stages such as LS-HS, HS input VGPRs are first. */
461 unsigned num_merged_next_stage_vgprs:3;
462 unsigned last_input:4;
463 unsigned as_ls:1;
464 unsigned as_es:1;
465 unsigned as_ngg:1;
466 /* Prologs for monolithic shaders shouldn't set EXEC. */
467 unsigned is_monolithic:1;
468 } vs_prolog;
469 struct {
470 struct si_tcs_epilog_bits states;
471 } tcs_epilog;
472 struct {
473 struct si_gs_prolog_bits states;
474 /* Prologs of monolithic shaders shouldn't set EXEC. */
475 unsigned is_monolithic:1;
476 unsigned as_ngg:1;
477 } gs_prolog;
478 struct {
479 struct si_ps_prolog_bits states;
480 unsigned num_input_sgprs:6;
481 unsigned num_input_vgprs:5;
482 /* Color interpolation and two-side color selection. */
483 unsigned colors_read:8; /* color input components read */
484 unsigned num_interp_inputs:5; /* BCOLOR is at this location */
485 unsigned face_vgpr_index:5;
486 unsigned ancillary_vgpr_index:5;
487 unsigned wqm:1;
488 char color_attr_index[2];
489 signed char color_interp_vgpr_index[2]; /* -1 == constant */
490 } ps_prolog;
491 struct {
492 struct si_ps_epilog_bits states;
493 unsigned colors_written:8;
494 unsigned writes_z:1;
495 unsigned writes_stencil:1;
496 unsigned writes_samplemask:1;
497 } ps_epilog;
498 };
499
500 struct si_shader_key {
501 /* Prolog and epilog flags. */
502 union {
503 struct {
504 struct si_vs_prolog_bits prolog;
505 } vs;
506 struct {
507 struct si_vs_prolog_bits ls_prolog; /* for merged LS-HS */
508 struct si_shader_selector *ls; /* for merged LS-HS */
509 struct si_tcs_epilog_bits epilog;
510 } tcs; /* tessellation control shader */
511 struct {
512 struct si_vs_prolog_bits vs_prolog; /* for merged ES-GS */
513 struct si_shader_selector *es; /* for merged ES-GS */
514 struct si_gs_prolog_bits prolog;
515 } gs;
516 struct {
517 struct si_ps_prolog_bits prolog;
518 struct si_ps_epilog_bits epilog;
519 } ps;
520 } part;
521
522 /* These three are initially set according to the NEXT_SHADER property,
523 * or guessed if the property doesn't seem correct.
524 */
525 unsigned as_es:1; /* export shader, which precedes GS */
526 unsigned as_ls:1; /* local shader, which precedes TCS */
527 unsigned as_ngg:1; /* VS, TES, or GS compiled as NGG primitive shader */
528
529 /* Flags for monolithic compilation only. */
530 struct {
531 /* Whether fetch should be opencoded according to vs_fix_fetch.
532 * Otherwise, if vs_fix_fetch is non-zero, buffer_load_format_xyzw
533 * with minimal fixups is used. */
534 uint16_t vs_fetch_opencode;
535 union si_vs_fix_fetch vs_fix_fetch[SI_MAX_ATTRIBS];
536
537 union {
538 uint64_t ff_tcs_inputs_to_copy; /* for fixed-func TCS */
539 /* When PS needs PrimID and GS is disabled. */
540 unsigned vs_export_prim_id:1;
541 struct {
542 unsigned interpolate_at_sample_force_center:1;
543 unsigned fbfetch_msaa:1;
544 unsigned fbfetch_is_1D:1;
545 unsigned fbfetch_layered:1;
546 } ps;
547 } u;
548 } mono;
549
550 /* Optimization flags for asynchronous compilation only. */
551 struct {
552 /* For HW VS (it can be VS, TES, GS) */
553 uint64_t kill_outputs; /* "get_unique_index" bits */
554 unsigned clip_disable:1;
555
556 /* For shaders where monolithic variants have better code.
557 *
558 * This is a flag that has no effect on code generation,
559 * but forces monolithic shaders to be used as soon as
560 * possible, because it's in the "opt" group.
561 */
562 unsigned prefer_mono:1;
563
564 /* Primitive discard compute shader. */
565 unsigned vs_as_prim_discard_cs:1;
566 unsigned cs_prim_type:4;
567 unsigned cs_indexed:1;
568 unsigned cs_instancing:1;
569 unsigned cs_primitive_restart:1;
570 unsigned cs_provoking_vertex_first:1;
571 unsigned cs_need_correct_orientation:1;
572 unsigned cs_cull_front:1;
573 unsigned cs_cull_back:1;
574 unsigned cs_cull_z:1;
575 unsigned cs_halfz_clip_space:1;
576 } opt;
577 };
578
579 /* Restore the pack alignment to default. */
580 #pragma pack(pop)
581
582 /* GCN-specific shader info. */
583 struct si_shader_info {
584 ubyte vs_output_param_offset[SI_MAX_VS_OUTPUTS];
585 ubyte num_input_sgprs;
586 ubyte num_input_vgprs;
587 signed char face_vgpr_index;
588 signed char ancillary_vgpr_index;
589 bool uses_instanceid;
590 ubyte nr_pos_exports;
591 ubyte nr_param_exports;
592 unsigned private_mem_vgprs;
593 unsigned max_simd_waves;
594 };
595
596 struct si_shader_binary {
597 const char *elf_buffer;
598 size_t elf_size;
599
600 char *llvm_ir_string;
601 };
602
603 struct gfx9_gs_info {
604 unsigned es_verts_per_subgroup;
605 unsigned gs_prims_per_subgroup;
606 unsigned gs_inst_prims_in_subgroup;
607 unsigned max_prims_per_subgroup;
608 unsigned esgs_ring_size; /* in bytes */
609 };
610
611 struct si_shader {
612 struct si_compiler_ctx_state compiler_ctx_state;
613
614 struct si_shader_selector *selector;
615 struct si_shader_selector *previous_stage_sel; /* for refcounting */
616 struct si_shader *next_variant;
617
618 struct si_shader_part *prolog;
619 struct si_shader *previous_stage; /* for GFX9 */
620 struct si_shader_part *prolog2;
621 struct si_shader_part *epilog;
622
623 struct si_pm4_state *pm4;
624 struct si_resource *bo;
625 struct si_resource *scratch_bo;
626 struct si_shader_key key;
627 struct util_queue_fence ready;
628 bool compilation_failed;
629 bool is_monolithic;
630 bool is_optimized;
631 bool is_binary_shared;
632 bool is_gs_copy_shader;
633
634 /* The following data is all that's needed for binary shaders. */
635 struct si_shader_binary binary;
636 struct ac_shader_config config;
637 struct si_shader_info info;
638
639 struct {
640 uint16_t ngg_emit_size; /* in dwords */
641 uint16_t hw_max_esverts;
642 uint16_t max_gsprims;
643 uint16_t max_out_verts;
644 uint16_t prim_amp_factor;
645 bool max_vert_out_per_gs_instance;
646 } ngg;
647
648 /* Shader key + LLVM IR + disassembly + statistics.
649 * Generated for debug contexts only.
650 */
651 char *shader_log;
652 size_t shader_log_size;
653
654 struct gfx9_gs_info gs_info;
655
656 /* For save precompute context registers values. */
657 union {
658 struct {
659 unsigned vgt_gsvs_ring_offset_1;
660 unsigned vgt_gsvs_ring_offset_2;
661 unsigned vgt_gsvs_ring_offset_3;
662 unsigned vgt_gsvs_ring_itemsize;
663 unsigned vgt_gs_max_vert_out;
664 unsigned vgt_gs_vert_itemsize;
665 unsigned vgt_gs_vert_itemsize_1;
666 unsigned vgt_gs_vert_itemsize_2;
667 unsigned vgt_gs_vert_itemsize_3;
668 unsigned vgt_gs_instance_cnt;
669 unsigned vgt_gs_onchip_cntl;
670 unsigned vgt_gs_max_prims_per_subgroup;
671 unsigned vgt_esgs_ring_itemsize;
672 } gs;
673
674 struct {
675 unsigned ge_max_output_per_subgroup;
676 unsigned ge_ngg_subgrp_cntl;
677 unsigned vgt_primitiveid_en;
678 unsigned vgt_gs_onchip_cntl;
679 unsigned vgt_gs_instance_cnt;
680 unsigned vgt_esgs_ring_itemsize;
681 unsigned spi_vs_out_config;
682 unsigned spi_shader_idx_format;
683 unsigned spi_shader_pos_format;
684 unsigned pa_cl_vte_cntl;
685 unsigned pa_cl_ngg_cntl;
686 unsigned vgt_gs_max_vert_out; /* for API GS */
687 } ngg;
688
689 struct {
690 unsigned vgt_gs_mode;
691 unsigned vgt_primitiveid_en;
692 unsigned vgt_reuse_off;
693 unsigned spi_vs_out_config;
694 unsigned spi_shader_pos_format;
695 unsigned pa_cl_vte_cntl;
696 } vs;
697
698 struct {
699 unsigned spi_ps_input_ena;
700 unsigned spi_ps_input_addr;
701 unsigned spi_baryc_cntl;
702 unsigned spi_ps_in_control;
703 unsigned spi_shader_z_format;
704 unsigned spi_shader_col_format;
705 unsigned cb_shader_mask;
706 } ps;
707 } ctx_reg;
708
709 /*For save precompute registers value */
710 unsigned vgt_tf_param; /* VGT_TF_PARAM */
711 unsigned vgt_vertex_reuse_block_cntl; /* VGT_VERTEX_REUSE_BLOCK_CNTL */
712 unsigned ge_cntl;
713 };
714
715 struct si_shader_part {
716 struct si_shader_part *next;
717 union si_shader_part_key key;
718 struct si_shader_binary binary;
719 struct ac_shader_config config;
720 };
721
722 /* si_shader.c */
723 struct si_shader *
724 si_generate_gs_copy_shader(struct si_screen *sscreen,
725 struct ac_llvm_compiler *compiler,
726 struct si_shader_selector *gs_selector,
727 struct pipe_debug_callback *debug);
728 int si_compile_tgsi_shader(struct si_screen *sscreen,
729 struct ac_llvm_compiler *compiler,
730 struct si_shader *shader,
731 struct pipe_debug_callback *debug);
732 bool si_shader_create(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
733 struct si_shader *shader,
734 struct pipe_debug_callback *debug);
735 void si_shader_destroy(struct si_shader *shader);
736 unsigned si_shader_io_get_unique_index_patch(unsigned semantic_name, unsigned index);
737 unsigned si_shader_io_get_unique_index(unsigned semantic_name, unsigned index,
738 unsigned is_varying);
739 bool si_shader_binary_upload(struct si_screen *sscreen, struct si_shader *shader,
740 uint64_t scratch_va);
741 void si_shader_dump(struct si_screen *sscreen, struct si_shader *shader,
742 struct pipe_debug_callback *debug,
743 FILE *f, bool check_debug_option);
744 void si_shader_dump_stats_for_shader_db(struct si_screen *screen,
745 struct si_shader *shader,
746 struct pipe_debug_callback *debug);
747 void si_multiwave_lds_size_workaround(struct si_screen *sscreen,
748 unsigned *lds_size);
749 const char *si_get_shader_name(const struct si_shader *shader);
750 void si_shader_binary_clean(struct si_shader_binary *binary);
751
752 /* si_shader_nir.c */
753 void si_nir_scan_shader(const struct nir_shader *nir,
754 struct tgsi_shader_info *info);
755 void si_nir_scan_tess_ctrl(const struct nir_shader *nir,
756 struct tgsi_tessctrl_info *out);
757 void si_nir_lower_ps_inputs(struct nir_shader *nir);
758 void si_lower_nir(struct si_shader_selector *sel);
759 void si_nir_opts(struct nir_shader *nir);
760
761 /* si_state_shaders.c */
762 void gfx9_get_gs_info(struct si_shader_selector *es,
763 struct si_shader_selector *gs,
764 struct gfx9_gs_info *out);
765
766 /* Inline helpers. */
767
768 /* Return the pointer to the main shader part's pointer. */
769 static inline struct si_shader **
770 si_get_main_shader_part(struct si_shader_selector *sel,
771 struct si_shader_key *key)
772 {
773 if (key->as_ls)
774 return &sel->main_shader_part_ls;
775 if (key->as_es && key->as_ngg)
776 return &sel->main_shader_part_ngg_es;
777 if (key->as_es)
778 return &sel->main_shader_part_es;
779 if (key->as_ngg)
780 return &sel->main_shader_part_ngg;
781 return &sel->main_shader_part;
782 }
783
784 static inline bool
785 si_shader_uses_bindless_samplers(struct si_shader_selector *selector)
786 {
787 return selector ? selector->info.uses_bindless_samplers : false;
788 }
789
790 static inline bool
791 si_shader_uses_bindless_images(struct si_shader_selector *selector)
792 {
793 return selector ? selector->info.uses_bindless_images : false;
794 }
795
796 void si_destroy_shader_selector(struct si_context *sctx,
797 struct si_shader_selector *sel);
798
799 static inline void
800 si_shader_selector_reference(struct si_context *sctx,
801 struct si_shader_selector **dst,
802 struct si_shader_selector *src)
803 {
804 if (pipe_reference(&(*dst)->reference, &src->reference))
805 si_destroy_shader_selector(sctx, *dst);
806
807 *dst = src;
808 }
809
810 #endif