2 * Copyright 2012 Advanced Micro Devices, Inc.
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:
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
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.
25 /* The compiler middle-end architecture: Explaining (non-)monolithic shaders
26 * -------------------------------------------------------------------------
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
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
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)
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.
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".
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
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.
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
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
92 /* How linking shader inputs and outputs between vertex, tessellation, and
93 * geometry shaders works.
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.
100 * Since GS and tessellation are only possible in the OpenGL core profile,
101 * only these semantics are valid for per-vertex data:
108 * CULLDIST0..1 (not implemented)
111 * For example, a shader only writing GENERIC0 has the output stride of 5.
113 * Only these semantics are valid for per-patch data:
121 * That's how independent shaders agree on input and output locations.
122 * The si_shader_io_get_unique_index function assigns the locations.
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.
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"
140 #include "ac_binary.h"
141 #include "ac_llvm_build.h"
142 #include "ac_llvm_util.h"
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)
154 #define SI_MAX_ATTRIBS 16
155 #define SI_MAX_VS_OUTPUTS 40
157 /* Shader IO unique indices are supported for TGSI_SEMANTIC_GENERIC with an
158 * index smaller than this.
160 #define SI_MAX_IO_GENERIC 32
162 /* SGPR user data indices */
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
,
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
,
174 /* all VS variants */
175 SI_SGPR_BASE_VERTEX
= SI_NUM_VS_STATE_RESOURCE_SGPRS
,
176 SI_SGPR_START_INSTANCE
,
180 SI_SGPR_VS_BLIT_DATA
= SI_SGPR_CONST_AND_SHADER_BUFFERS
,
183 SI_SGPR_TES_OFFCHIP_LAYOUT
= SI_NUM_VS_STATE_RESOURCE_SGPRS
,
184 SI_SGPR_TES_OFFCHIP_ADDR
,
185 SI_TES_NUM_USER_SGPR
,
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
,
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
,
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
,
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
,
212 SI_SGPR_ALPHA_REF
= SI_NUM_RESOURCE_SGPRS
,
216 /* LLVM function parameter indices */
218 SI_NUM_RESOURCE_PARAMS
= 4,
220 /* PS only parameters */
221 SI_PARAM_ALPHA_REF
= SI_NUM_RESOURCE_PARAMS
,
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
,
237 SI_PARAM_SAMPLE_COVERAGE
,
238 SI_PARAM_POS_FIXED_PT
,
240 SI_NUM_PARAMS
= SI_PARAM_POS_FIXED_PT
+ 9, /* +8 for COLOR[0..1] */
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
260 /* Use a property enum that CS wouldn't use. */
261 TGSI_PROPERTY_CS_LOCAL_SIZE
= TGSI_PROPERTY_FS_COORD_ORIGIN
,
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,
270 * For VS shader keys, describe any fixups required for vertex fetch.
272 * \ref log_size, \ref format, and the number of channels are interpreted as
273 * by \ref ac_build_opencoded_load_format.
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).
279 union si_vs_fix_fetch
{
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 */
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
;
297 /* Used if thread_index == -1 or if debug.async is true. */
298 struct pipe_debug_callback debug
;
300 /* Used for creating the log string for gallium/ddebug. */
301 bool is_debug_context
;
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.
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
;
314 struct si_shader
*first_variant
; /* immutable after the first variant */
315 struct si_shader
*last_variant
; /* mutable */
317 /* The compiled TGSI shader expecting a prolog and/or epilog (not
318 * uploaded to a buffer).
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 */
326 struct si_shader
*gs_copy_shader
;
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
;
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 bool ngg_writes_edgeflag
;
340 bool pos_writes_edgeflag
;
341 unsigned pa_cl_vs_out_cntl
;
347 unsigned esgs_itemsize
; /* vertex stride */
348 unsigned lshs_vertex_stride
;
351 unsigned gs_input_verts_per_prim
;
352 unsigned gs_output_prim
;
353 unsigned gs_max_out_vertices
;
354 unsigned gs_num_invocations
;
355 unsigned max_gs_stream
; /* count - 1 */
356 unsigned gsvs_vertex_size
;
357 unsigned max_gsvs_emit_size
;
358 unsigned enabled_streamout_buffer_mask
;
359 bool tess_turns_off_ngg
;
362 unsigned color_attr_index
[2];
363 unsigned db_shader_control
;
364 /* Set 0xf or 0x0 (4 bits) per each written output.
365 * ANDed with spi_shader_col_format.
367 unsigned colors_written_4bit
;
369 uint64_t outputs_written_before_ps
; /* "get_unique_index" bits */
370 uint64_t outputs_written
; /* "get_unique_index" bits */
371 uint32_t patch_outputs_written
; /* "get_unique_index_patch" bits */
373 uint64_t inputs_read
; /* "get_unique_index" bits */
375 /* bitmasks of used descriptor slots */
376 uint32_t active_const_and_shader_buffers
;
377 uint64_t active_samplers_and_images
;
380 /* Valid shader configurations:
382 * API shaders VS | TCS | TES | GS |pass| PS
383 * are compiled as: | | | |thru|
385 * Only VS & PS: VS | | | | | PS
386 * GFX6 - with GS: ES | | | GS | VS | PS
387 * - with tess: LS | HS | VS | | | PS
388 * - with both: LS | HS | ES | GS | VS | PS
389 * GFX9 - with GS: -> | | | GS | VS | PS
390 * - with tess: -> | HS | VS | | | PS
391 * - with both: -> | HS | -> | GS | VS | PS
393 * NGG - VS & PS: GS | | | | | PS
394 * (GFX10+) - with GS: -> | | | GS | | PS
395 * - with tess: -> | HS | GS | | | PS
396 * - with both: -> | HS | -> | GS | | PS
398 * -> = merged with the next stage
401 /* Use the byte alignment for all following structure members for optimal
402 * shader key memory footprint.
404 #pragma pack(push, 1)
406 /* Common VS bits between the shader key and the prolog key. */
407 struct si_vs_prolog_bits
{
408 /* - If neither "is_one" nor "is_fetched" has a bit set, the instance
410 * - If "is_one" has a bit set, the instance divisor is 1.
411 * - If "is_fetched" has a bit set, the instance divisor will be loaded
412 * from the constant buffer.
414 uint16_t instance_divisor_is_one
; /* bitmask of inputs */
415 uint16_t instance_divisor_is_fetched
; /* bitmask of inputs */
416 unsigned ls_vgpr_fix
:1;
417 unsigned unpack_instance_id_from_vertex_id
:1;
420 /* Common TCS bits between the shader key and the epilog key. */
421 struct si_tcs_epilog_bits
{
422 unsigned prim_mode
:3;
423 unsigned invoc0_tess_factors_are_def
:1;
424 unsigned tes_reads_tess_factors
:1;
427 struct si_gs_prolog_bits
{
428 unsigned tri_strip_adj_fix
:1;
429 unsigned gfx9_prev_is_vs
:1;
432 /* Common PS bits between the shader key and the prolog key. */
433 struct si_ps_prolog_bits
{
434 unsigned color_two_side
:1;
435 unsigned flatshade_colors
:1;
436 unsigned poly_stipple
:1;
437 unsigned force_persp_sample_interp
:1;
438 unsigned force_linear_sample_interp
:1;
439 unsigned force_persp_center_interp
:1;
440 unsigned force_linear_center_interp
:1;
441 unsigned bc_optimize_for_persp
:1;
442 unsigned bc_optimize_for_linear
:1;
443 unsigned samplemask_log_ps_iter
:3;
446 /* Common PS bits between the shader key and the epilog key. */
447 struct si_ps_epilog_bits
{
448 unsigned spi_shader_col_format
;
449 unsigned color_is_int8
:8;
450 unsigned color_is_int10
:8;
451 unsigned last_cbuf
:3;
452 unsigned alpha_func
:3;
453 unsigned alpha_to_one
:1;
454 unsigned poly_line_smoothing
:1;
455 unsigned clamp_color
:1;
458 union si_shader_part_key
{
460 struct si_vs_prolog_bits states
;
461 unsigned num_input_sgprs
:6;
462 /* For merged stages such as LS-HS, HS input VGPRs are first. */
463 unsigned num_merged_next_stage_vgprs
:3;
464 unsigned last_input
:4;
468 /* Prologs for monolithic shaders shouldn't set EXEC. */
469 unsigned is_monolithic
:1;
472 struct si_tcs_epilog_bits states
;
475 struct si_gs_prolog_bits states
;
476 /* Prologs of monolithic shaders shouldn't set EXEC. */
477 unsigned is_monolithic
:1;
481 struct si_ps_prolog_bits states
;
482 unsigned num_input_sgprs
:6;
483 unsigned num_input_vgprs
:5;
484 /* Color interpolation and two-side color selection. */
485 unsigned colors_read
:8; /* color input components read */
486 unsigned num_interp_inputs
:5; /* BCOLOR is at this location */
487 unsigned face_vgpr_index
:5;
488 unsigned ancillary_vgpr_index
:5;
490 char color_attr_index
[2];
491 signed char color_interp_vgpr_index
[2]; /* -1 == constant */
494 struct si_ps_epilog_bits states
;
495 unsigned colors_written
:8;
497 unsigned writes_stencil
:1;
498 unsigned writes_samplemask
:1;
502 struct si_shader_key
{
503 /* Prolog and epilog flags. */
506 struct si_vs_prolog_bits prolog
;
509 struct si_vs_prolog_bits ls_prolog
; /* for merged LS-HS */
510 struct si_shader_selector
*ls
; /* for merged LS-HS */
511 struct si_tcs_epilog_bits epilog
;
512 } tcs
; /* tessellation control shader */
514 struct si_vs_prolog_bits vs_prolog
; /* for merged ES-GS */
515 struct si_shader_selector
*es
; /* for merged ES-GS */
516 struct si_gs_prolog_bits prolog
;
519 struct si_ps_prolog_bits prolog
;
520 struct si_ps_epilog_bits epilog
;
524 /* These three are initially set according to the NEXT_SHADER property,
525 * or guessed if the property doesn't seem correct.
527 unsigned as_es
:1; /* export shader, which precedes GS */
528 unsigned as_ls
:1; /* local shader, which precedes TCS */
529 unsigned as_ngg
:1; /* VS, TES, or GS compiled as NGG primitive shader */
531 /* Flags for monolithic compilation only. */
533 /* Whether fetch should be opencoded according to vs_fix_fetch.
534 * Otherwise, if vs_fix_fetch is non-zero, buffer_load_format_xyzw
535 * with minimal fixups is used. */
536 uint16_t vs_fetch_opencode
;
537 union si_vs_fix_fetch vs_fix_fetch
[SI_MAX_ATTRIBS
];
540 uint64_t ff_tcs_inputs_to_copy
; /* for fixed-func TCS */
541 /* When PS needs PrimID and GS is disabled. */
542 unsigned vs_export_prim_id
:1;
544 unsigned interpolate_at_sample_force_center
:1;
545 unsigned fbfetch_msaa
:1;
546 unsigned fbfetch_is_1D
:1;
547 unsigned fbfetch_layered
:1;
552 /* Optimization flags for asynchronous compilation only. */
554 /* For HW VS (it can be VS, TES, GS) */
555 uint64_t kill_outputs
; /* "get_unique_index" bits */
556 unsigned clip_disable
:1;
558 /* For shaders where monolithic variants have better code.
560 * This is a flag that has no effect on code generation,
561 * but forces monolithic shaders to be used as soon as
562 * possible, because it's in the "opt" group.
564 unsigned prefer_mono
:1;
566 /* Primitive discard compute shader. */
567 unsigned vs_as_prim_discard_cs
:1;
568 unsigned cs_prim_type
:4;
569 unsigned cs_indexed
:1;
570 unsigned cs_instancing
:1;
571 unsigned cs_primitive_restart
:1;
572 unsigned cs_provoking_vertex_first
:1;
573 unsigned cs_need_correct_orientation
:1;
574 unsigned cs_cull_front
:1;
575 unsigned cs_cull_back
:1;
576 unsigned cs_cull_z
:1;
577 unsigned cs_halfz_clip_space
:1;
581 /* Restore the pack alignment to default. */
584 /* GCN-specific shader info. */
585 struct si_shader_info
{
586 ubyte vs_output_param_offset
[SI_MAX_VS_OUTPUTS
];
587 ubyte num_input_sgprs
;
588 ubyte num_input_vgprs
;
589 signed char face_vgpr_index
;
590 signed char ancillary_vgpr_index
;
591 bool uses_instanceid
;
592 ubyte nr_pos_exports
;
593 ubyte nr_param_exports
;
594 unsigned private_mem_vgprs
;
595 unsigned max_simd_waves
;
598 struct si_shader_binary
{
599 const char *elf_buffer
;
602 char *llvm_ir_string
;
605 struct gfx9_gs_info
{
606 unsigned es_verts_per_subgroup
;
607 unsigned gs_prims_per_subgroup
;
608 unsigned gs_inst_prims_in_subgroup
;
609 unsigned max_prims_per_subgroup
;
610 unsigned esgs_ring_size
; /* in bytes */
614 struct si_compiler_ctx_state compiler_ctx_state
;
616 struct si_shader_selector
*selector
;
617 struct si_shader_selector
*previous_stage_sel
; /* for refcounting */
618 struct si_shader
*next_variant
;
620 struct si_shader_part
*prolog
;
621 struct si_shader
*previous_stage
; /* for GFX9 */
622 struct si_shader_part
*prolog2
;
623 struct si_shader_part
*epilog
;
625 struct si_pm4_state
*pm4
;
626 struct si_resource
*bo
;
627 struct si_resource
*scratch_bo
;
628 struct si_shader_key key
;
629 struct util_queue_fence ready
;
630 bool compilation_failed
;
633 bool is_binary_shared
;
634 bool is_gs_copy_shader
;
636 /* The following data is all that's needed for binary shaders. */
637 struct si_shader_binary binary
;
638 struct ac_shader_config config
;
639 struct si_shader_info info
;
642 uint16_t ngg_emit_size
; /* in dwords */
643 uint16_t hw_max_esverts
;
644 uint16_t max_gsprims
;
645 uint16_t max_out_verts
;
646 uint16_t prim_amp_factor
;
647 bool max_vert_out_per_gs_instance
;
650 /* Shader key + LLVM IR + disassembly + statistics.
651 * Generated for debug contexts only.
654 size_t shader_log_size
;
656 struct gfx9_gs_info gs_info
;
658 /* For save precompute context registers values. */
661 unsigned vgt_gsvs_ring_offset_1
;
662 unsigned vgt_gsvs_ring_offset_2
;
663 unsigned vgt_gsvs_ring_offset_3
;
664 unsigned vgt_gsvs_ring_itemsize
;
665 unsigned vgt_gs_max_vert_out
;
666 unsigned vgt_gs_vert_itemsize
;
667 unsigned vgt_gs_vert_itemsize_1
;
668 unsigned vgt_gs_vert_itemsize_2
;
669 unsigned vgt_gs_vert_itemsize_3
;
670 unsigned vgt_gs_instance_cnt
;
671 unsigned vgt_gs_onchip_cntl
;
672 unsigned vgt_gs_max_prims_per_subgroup
;
673 unsigned vgt_esgs_ring_itemsize
;
677 unsigned ge_max_output_per_subgroup
;
678 unsigned ge_ngg_subgrp_cntl
;
679 unsigned vgt_primitiveid_en
;
680 unsigned vgt_gs_onchip_cntl
;
681 unsigned vgt_gs_instance_cnt
;
682 unsigned vgt_esgs_ring_itemsize
;
683 unsigned spi_vs_out_config
;
684 unsigned spi_shader_idx_format
;
685 unsigned spi_shader_pos_format
;
686 unsigned pa_cl_vte_cntl
;
687 unsigned pa_cl_ngg_cntl
;
688 unsigned vgt_gs_max_vert_out
; /* for API GS */
692 unsigned vgt_gs_mode
;
693 unsigned vgt_primitiveid_en
;
694 unsigned vgt_reuse_off
;
695 unsigned spi_vs_out_config
;
696 unsigned spi_shader_pos_format
;
697 unsigned pa_cl_vte_cntl
;
701 unsigned spi_ps_input_ena
;
702 unsigned spi_ps_input_addr
;
703 unsigned spi_baryc_cntl
;
704 unsigned spi_ps_in_control
;
705 unsigned spi_shader_z_format
;
706 unsigned spi_shader_col_format
;
707 unsigned cb_shader_mask
;
711 /*For save precompute registers value */
712 unsigned vgt_tf_param
; /* VGT_TF_PARAM */
713 unsigned vgt_vertex_reuse_block_cntl
; /* VGT_VERTEX_REUSE_BLOCK_CNTL */
717 struct si_shader_part
{
718 struct si_shader_part
*next
;
719 union si_shader_part_key key
;
720 struct si_shader_binary binary
;
721 struct ac_shader_config config
;
726 si_generate_gs_copy_shader(struct si_screen
*sscreen
,
727 struct ac_llvm_compiler
*compiler
,
728 struct si_shader_selector
*gs_selector
,
729 struct pipe_debug_callback
*debug
);
730 int si_compile_tgsi_shader(struct si_screen
*sscreen
,
731 struct ac_llvm_compiler
*compiler
,
732 struct si_shader
*shader
,
733 struct pipe_debug_callback
*debug
);
734 bool si_shader_create(struct si_screen
*sscreen
, struct ac_llvm_compiler
*compiler
,
735 struct si_shader
*shader
,
736 struct pipe_debug_callback
*debug
);
737 void si_shader_destroy(struct si_shader
*shader
);
738 unsigned si_shader_io_get_unique_index_patch(unsigned semantic_name
, unsigned index
);
739 unsigned si_shader_io_get_unique_index(unsigned semantic_name
, unsigned index
,
740 unsigned is_varying
);
741 bool si_shader_binary_upload(struct si_screen
*sscreen
, struct si_shader
*shader
,
742 uint64_t scratch_va
);
743 void si_shader_dump(struct si_screen
*sscreen
, struct si_shader
*shader
,
744 struct pipe_debug_callback
*debug
,
745 FILE *f
, bool check_debug_option
);
746 void si_shader_dump_stats_for_shader_db(struct si_screen
*screen
,
747 struct si_shader
*shader
,
748 struct pipe_debug_callback
*debug
);
749 void si_multiwave_lds_size_workaround(struct si_screen
*sscreen
,
751 const char *si_get_shader_name(const struct si_shader
*shader
);
752 void si_shader_binary_clean(struct si_shader_binary
*binary
);
754 /* si_shader_nir.c */
755 void si_nir_scan_shader(const struct nir_shader
*nir
,
756 struct tgsi_shader_info
*info
);
757 void si_nir_scan_tess_ctrl(const struct nir_shader
*nir
,
758 struct tgsi_tessctrl_info
*out
);
759 void si_nir_lower_ps_inputs(struct nir_shader
*nir
);
760 void si_lower_nir(struct si_shader_selector
*sel
, unsigned wave_size
);
761 void si_nir_opts(struct nir_shader
*nir
);
763 /* si_state_shaders.c */
764 void gfx9_get_gs_info(struct si_shader_selector
*es
,
765 struct si_shader_selector
*gs
,
766 struct gfx9_gs_info
*out
);
768 /* Inline helpers. */
770 /* Return the pointer to the main shader part's pointer. */
771 static inline struct si_shader
**
772 si_get_main_shader_part(struct si_shader_selector
*sel
,
773 struct si_shader_key
*key
)
776 return &sel
->main_shader_part_ls
;
777 if (key
->as_es
&& key
->as_ngg
)
778 return &sel
->main_shader_part_ngg_es
;
780 return &sel
->main_shader_part_es
;
782 return &sel
->main_shader_part_ngg
;
783 return &sel
->main_shader_part
;
787 si_shader_uses_bindless_samplers(struct si_shader_selector
*selector
)
789 return selector
? selector
->info
.uses_bindless_samplers
: false;
793 si_shader_uses_bindless_images(struct si_shader_selector
*selector
)
795 return selector
? selector
->info
.uses_bindless_images
: false;
798 void si_destroy_shader_selector(struct si_context
*sctx
,
799 struct si_shader_selector
*sel
);
802 si_shader_selector_reference(struct si_context
*sctx
,
803 struct si_shader_selector
**dst
,
804 struct si_shader_selector
*src
)
806 if (pipe_reference(&(*dst
)->reference
, &src
->reference
))
807 si_destroy_shader_selector(sctx
, *dst
);