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"
139 #include "util/simple_mtx.h"
141 #include "ac_binary.h"
142 #include "ac_llvm_build.h"
143 #include "ac_llvm_util.h"
147 // Use LDS symbols when supported by LLVM. Can be disabled for testing the old
148 // path on newer LLVM for now. Should be removed in the long term.
149 #define USE_LDS_SYMBOLS (true)
155 #define SI_MAX_ATTRIBS 16
156 #define SI_MAX_VS_OUTPUTS 40
158 /* Shader IO unique indices are supported for TGSI_SEMANTIC_GENERIC with an
159 * index smaller than this.
161 #define SI_MAX_IO_GENERIC 32
163 /* SGPR user data indices */
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
,
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
,
175 /* all VS variants */
176 SI_SGPR_BASE_VERTEX
= SI_NUM_VS_STATE_RESOURCE_SGPRS
,
177 SI_SGPR_START_INSTANCE
,
181 SI_SGPR_VS_BLIT_DATA
= SI_SGPR_CONST_AND_SHADER_BUFFERS
,
184 SI_SGPR_TES_OFFCHIP_LAYOUT
= SI_NUM_VS_STATE_RESOURCE_SGPRS
,
185 SI_SGPR_TES_OFFCHIP_ADDR
,
186 SI_TES_NUM_USER_SGPR
,
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
,
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
,
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
,
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
,
213 SI_SGPR_ALPHA_REF
= SI_NUM_RESOURCE_SGPRS
,
216 /* The value has to be 12, because the hw requires that descriptors
217 * are aligned to 4 SGPRs.
219 SI_SGPR_VS_VB_DESCRIPTOR_FIRST
= 12,
222 /* LLVM function parameter indices */
224 SI_NUM_RESOURCE_PARAMS
= 4,
226 /* PS only parameters */
227 SI_PARAM_ALPHA_REF
= SI_NUM_RESOURCE_PARAMS
,
229 SI_PARAM_PERSP_SAMPLE
,
230 SI_PARAM_PERSP_CENTER
,
231 SI_PARAM_PERSP_CENTROID
,
232 SI_PARAM_PERSP_PULL_MODEL
,
233 SI_PARAM_LINEAR_SAMPLE
,
234 SI_PARAM_LINEAR_CENTER
,
235 SI_PARAM_LINEAR_CENTROID
,
236 SI_PARAM_LINE_STIPPLE_TEX
,
237 SI_PARAM_POS_X_FLOAT
,
238 SI_PARAM_POS_Y_FLOAT
,
239 SI_PARAM_POS_Z_FLOAT
,
240 SI_PARAM_POS_W_FLOAT
,
243 SI_PARAM_SAMPLE_COVERAGE
,
244 SI_PARAM_POS_FIXED_PT
,
246 SI_NUM_PARAMS
= SI_PARAM_POS_FIXED_PT
+ 9, /* +8 for COLOR[0..1] */
249 /* Fields of driver-defined VS state SGPR. */
250 #define S_VS_STATE_CLAMP_VERTEX_COLOR(x) (((unsigned)(x) & 0x1) << 0)
251 #define C_VS_STATE_CLAMP_VERTEX_COLOR 0xFFFFFFFE
252 #define S_VS_STATE_INDEXED(x) (((unsigned)(x) & 0x1) << 1)
253 #define C_VS_STATE_INDEXED 0xFFFFFFFD
254 #define S_VS_STATE_OUTPRIM(x) (((unsigned)(x) & 0x3) << 2)
255 #define C_VS_STATE_OUTPRIM 0xFFFFFFF3
256 #define S_VS_STATE_PROVOKING_VTX_INDEX(x) (((unsigned)(x) & 0x3) << 4)
257 #define C_VS_STATE_PROVOKING_VTX_INDEX 0xFFFFFFCF
258 #define S_VS_STATE_STREAMOUT_QUERY_ENABLED(x) (((unsigned)(x) & 0x1) << 6)
259 #define C_VS_STATE_STREAMOUT_QUERY_ENABLED 0xFFFFFFBF
260 #define S_VS_STATE_LS_OUT_PATCH_SIZE(x) (((unsigned)(x) & 0x1FFF) << 8)
261 #define C_VS_STATE_LS_OUT_PATCH_SIZE 0xFFE000FF
262 #define S_VS_STATE_LS_OUT_VERTEX_SIZE(x) (((unsigned)(x) & 0xFF) << 24)
263 #define C_VS_STATE_LS_OUT_VERTEX_SIZE 0x00FFFFFF
266 /* Use a property enum that CS wouldn't use. */
267 TGSI_PROPERTY_CS_LOCAL_SIZE
= TGSI_PROPERTY_FS_COORD_ORIGIN
,
269 /* These represent the number of SGPRs the shader uses. */
270 SI_VS_BLIT_SGPRS_POS
= 3,
271 SI_VS_BLIT_SGPRS_POS_COLOR
= 7,
272 SI_VS_BLIT_SGPRS_POS_TEXCOORD
= 9,
276 * For VS shader keys, describe any fixups required for vertex fetch.
278 * \ref log_size, \ref format, and the number of channels are interpreted as
279 * by \ref ac_build_opencoded_load_format.
281 * Note: all bits 0 (size = 1 byte, num channels = 1, format = float) is an
282 * impossible format and indicates that no fixup is needed (just use
283 * buffer_load_format_xyzw).
285 union si_vs_fix_fetch
{
287 uint8_t log_size
: 2; /* 1, 2, 4, 8 or bytes per channel */
288 uint8_t num_channels_m1
: 2; /* number of channels minus 1 */
289 uint8_t format
: 3; /* AC_FETCH_FORMAT_xxx */
290 uint8_t reverse
: 1; /* reverse XYZ channels */
297 /* State of the context creating the shader object. */
298 struct si_compiler_ctx_state
{
299 /* Should only be used by si_init_shader_selector_async and
300 * si_build_shader_variant if thread_index == -1 (non-threaded). */
301 struct ac_llvm_compiler
*compiler
;
303 /* Used if thread_index == -1 or if debug.async is true. */
304 struct pipe_debug_callback debug
;
306 /* Used for creating the log string for gallium/ddebug. */
307 bool is_debug_context
;
310 /* A shader selector is a gallium CSO and contains shader variants and
311 * binaries for one NIR program. This can be shared by multiple contexts.
313 struct si_shader_selector
{
314 struct pipe_reference reference
;
315 struct si_screen
*screen
;
316 struct util_queue_fence ready
;
317 struct si_compiler_ctx_state compiler_ctx_state
;
320 struct si_shader
*first_variant
; /* immutable after the first variant */
321 struct si_shader
*last_variant
; /* mutable */
323 /* The compiled NIR shader without a prolog and/or epilog (not
324 * uploaded to a buffer object).
326 struct si_shader
*main_shader_part
;
327 struct si_shader
*main_shader_part_ls
; /* as_ls is set in the key */
328 struct si_shader
*main_shader_part_es
; /* as_es is set in the key */
329 struct si_shader
*main_shader_part_ngg
; /* as_ngg is set in the key */
330 struct si_shader
*main_shader_part_ngg_es
; /* for Wave32 TES before legacy GS */
332 struct si_shader
*gs_copy_shader
;
334 struct nir_shader
*nir
;
338 struct pipe_stream_output_info so
;
339 struct tgsi_shader_info info
;
340 struct tgsi_tessctrl_info tcs_info
;
342 /* PIPE_SHADER_[VERTEX|FRAGMENT|...] */
343 enum pipe_shader_type type
;
344 bool vs_needs_prolog
;
345 bool force_correct_derivs_after_kill
;
346 bool prim_discard_cs_allowed
;
347 unsigned num_vs_inputs
;
348 unsigned num_vbos_in_user_sgprs
;
349 unsigned pa_cl_vs_out_cntl
;
355 unsigned esgs_itemsize
; /* vertex stride */
356 unsigned lshs_vertex_stride
;
359 unsigned gs_input_verts_per_prim
;
360 unsigned gs_output_prim
;
361 unsigned gs_max_out_vertices
;
362 unsigned gs_num_invocations
;
363 unsigned max_gs_stream
; /* count - 1 */
364 unsigned gsvs_vertex_size
;
365 unsigned max_gsvs_emit_size
;
366 unsigned enabled_streamout_buffer_mask
;
367 bool tess_turns_off_ngg
;
370 unsigned color_attr_index
[2];
371 unsigned db_shader_control
;
372 /* Set 0xf or 0x0 (4 bits) per each written output.
373 * ANDed with spi_shader_col_format.
375 unsigned colors_written_4bit
;
377 uint64_t outputs_written_before_ps
; /* "get_unique_index" bits */
378 uint64_t outputs_written
; /* "get_unique_index" bits */
379 uint32_t patch_outputs_written
; /* "get_unique_index_patch" bits */
381 uint64_t inputs_read
; /* "get_unique_index" bits */
383 /* bitmasks of used descriptor slots */
384 uint32_t active_const_and_shader_buffers
;
385 uint64_t active_samplers_and_images
;
388 /* Valid shader configurations:
390 * API shaders VS | TCS | TES | GS |pass| PS
391 * are compiled as: | | | |thru|
393 * Only VS & PS: VS | | | | | PS
394 * GFX6 - with GS: ES | | | GS | VS | PS
395 * - with tess: LS | HS | VS | | | PS
396 * - with both: LS | HS | ES | GS | VS | PS
397 * GFX9 - with GS: -> | | | GS | VS | PS
398 * - with tess: -> | HS | VS | | | PS
399 * - with both: -> | HS | -> | GS | VS | PS
401 * NGG - VS & PS: GS | | | | | PS
402 * (GFX10+) - with GS: -> | | | GS | | PS
403 * - with tess: -> | HS | GS | | | PS
404 * - with both: -> | HS | -> | GS | | PS
406 * -> = merged with the next stage
409 /* Use the byte alignment for all following structure members for optimal
410 * shader key memory footprint.
412 #pragma pack(push, 1)
414 /* Common VS bits between the shader key and the prolog key. */
415 struct si_vs_prolog_bits
{
416 /* - If neither "is_one" nor "is_fetched" has a bit set, the instance
418 * - If "is_one" has a bit set, the instance divisor is 1.
419 * - If "is_fetched" has a bit set, the instance divisor will be loaded
420 * from the constant buffer.
422 uint16_t instance_divisor_is_one
; /* bitmask of inputs */
423 uint16_t instance_divisor_is_fetched
; /* bitmask of inputs */
424 unsigned ls_vgpr_fix
:1;
425 unsigned unpack_instance_id_from_vertex_id
:1;
428 /* Common TCS bits between the shader key and the epilog key. */
429 struct si_tcs_epilog_bits
{
430 unsigned prim_mode
:3;
431 unsigned invoc0_tess_factors_are_def
:1;
432 unsigned tes_reads_tess_factors
:1;
435 struct si_gs_prolog_bits
{
436 unsigned tri_strip_adj_fix
:1;
437 unsigned gfx9_prev_is_vs
:1;
440 /* Common PS bits between the shader key and the prolog key. */
441 struct si_ps_prolog_bits
{
442 unsigned color_two_side
:1;
443 unsigned flatshade_colors
:1;
444 unsigned poly_stipple
:1;
445 unsigned force_persp_sample_interp
:1;
446 unsigned force_linear_sample_interp
:1;
447 unsigned force_persp_center_interp
:1;
448 unsigned force_linear_center_interp
:1;
449 unsigned bc_optimize_for_persp
:1;
450 unsigned bc_optimize_for_linear
:1;
451 unsigned samplemask_log_ps_iter
:3;
454 /* Common PS bits between the shader key and the epilog key. */
455 struct si_ps_epilog_bits
{
456 unsigned spi_shader_col_format
;
457 unsigned color_is_int8
:8;
458 unsigned color_is_int10
:8;
459 unsigned last_cbuf
:3;
460 unsigned alpha_func
:3;
461 unsigned alpha_to_one
:1;
462 unsigned poly_line_smoothing
:1;
463 unsigned clamp_color
:1;
466 union si_shader_part_key
{
468 struct si_vs_prolog_bits states
;
469 unsigned num_input_sgprs
:6;
470 /* For merged stages such as LS-HS, HS input VGPRs are first. */
471 unsigned num_merged_next_stage_vgprs
:3;
472 unsigned num_inputs
:5;
476 /* Prologs for monolithic shaders shouldn't set EXEC. */
477 unsigned is_monolithic
:1;
480 struct si_tcs_epilog_bits states
;
483 struct si_gs_prolog_bits states
;
484 /* Prologs of monolithic shaders shouldn't set EXEC. */
485 unsigned is_monolithic
:1;
489 struct si_ps_prolog_bits states
;
490 unsigned num_input_sgprs
:6;
491 unsigned num_input_vgprs
:5;
492 /* Color interpolation and two-side color selection. */
493 unsigned colors_read
:8; /* color input components read */
494 unsigned num_interp_inputs
:5; /* BCOLOR is at this location */
495 unsigned face_vgpr_index
:5;
496 unsigned ancillary_vgpr_index
:5;
498 char color_attr_index
[2];
499 signed char color_interp_vgpr_index
[2]; /* -1 == constant */
502 struct si_ps_epilog_bits states
;
503 unsigned colors_written
:8;
505 unsigned writes_stencil
:1;
506 unsigned writes_samplemask
:1;
510 struct si_shader_key
{
511 /* Prolog and epilog flags. */
514 struct si_vs_prolog_bits prolog
;
517 struct si_vs_prolog_bits ls_prolog
; /* for merged LS-HS */
518 struct si_shader_selector
*ls
; /* for merged LS-HS */
519 struct si_tcs_epilog_bits epilog
;
520 } tcs
; /* tessellation control shader */
522 struct si_vs_prolog_bits vs_prolog
; /* for merged ES-GS */
523 struct si_shader_selector
*es
; /* for merged ES-GS */
524 struct si_gs_prolog_bits prolog
;
527 struct si_ps_prolog_bits prolog
;
528 struct si_ps_epilog_bits epilog
;
532 /* These three are initially set according to the NEXT_SHADER property,
533 * or guessed if the property doesn't seem correct.
535 unsigned as_es
:1; /* export shader, which precedes GS */
536 unsigned as_ls
:1; /* local shader, which precedes TCS */
537 unsigned as_ngg
:1; /* VS, TES, or GS compiled as NGG primitive shader */
539 /* Flags for monolithic compilation only. */
541 /* Whether fetch should be opencoded according to vs_fix_fetch.
542 * Otherwise, if vs_fix_fetch is non-zero, buffer_load_format_xyzw
543 * with minimal fixups is used. */
544 uint16_t vs_fetch_opencode
;
545 union si_vs_fix_fetch vs_fix_fetch
[SI_MAX_ATTRIBS
];
548 uint64_t ff_tcs_inputs_to_copy
; /* for fixed-func TCS */
549 /* When PS needs PrimID and GS is disabled. */
550 unsigned vs_export_prim_id
:1;
552 unsigned interpolate_at_sample_force_center
:1;
553 unsigned fbfetch_msaa
:1;
554 unsigned fbfetch_is_1D
:1;
555 unsigned fbfetch_layered
:1;
560 /* Optimization flags for asynchronous compilation only. */
562 /* For HW VS (it can be VS, TES, GS) */
563 uint64_t kill_outputs
; /* "get_unique_index" bits */
564 unsigned clip_disable
:1;
566 /* For shaders where monolithic variants have better code.
568 * This is a flag that has no effect on code generation,
569 * but forces monolithic shaders to be used as soon as
570 * possible, because it's in the "opt" group.
572 unsigned prefer_mono
:1;
574 /* Primitive discard compute shader. */
575 unsigned vs_as_prim_discard_cs
:1;
576 unsigned cs_prim_type
:4;
577 unsigned cs_indexed
:1;
578 unsigned cs_instancing
:1;
579 unsigned cs_primitive_restart
:1;
580 unsigned cs_provoking_vertex_first
:1;
581 unsigned cs_need_correct_orientation
:1;
582 unsigned cs_cull_front
:1;
583 unsigned cs_cull_back
:1;
584 unsigned cs_cull_z
:1;
585 unsigned cs_halfz_clip_space
:1;
589 /* Restore the pack alignment to default. */
592 /* GCN-specific shader info. */
593 struct si_shader_info
{
594 ubyte vs_output_param_offset
[SI_MAX_VS_OUTPUTS
];
595 ubyte num_input_sgprs
;
596 ubyte num_input_vgprs
;
597 signed char face_vgpr_index
;
598 signed char ancillary_vgpr_index
;
599 bool uses_instanceid
;
600 ubyte nr_pos_exports
;
601 ubyte nr_param_exports
;
602 unsigned private_mem_vgprs
;
603 unsigned max_simd_waves
;
606 struct si_shader_binary
{
607 const char *elf_buffer
;
610 char *llvm_ir_string
;
613 struct gfx9_gs_info
{
614 unsigned es_verts_per_subgroup
;
615 unsigned gs_prims_per_subgroup
;
616 unsigned gs_inst_prims_in_subgroup
;
617 unsigned max_prims_per_subgroup
;
618 unsigned esgs_ring_size
; /* in bytes */
622 struct si_compiler_ctx_state compiler_ctx_state
;
624 struct si_shader_selector
*selector
;
625 struct si_shader_selector
*previous_stage_sel
; /* for refcounting */
626 struct si_shader
*next_variant
;
628 struct si_shader_part
*prolog
;
629 struct si_shader
*previous_stage
; /* for GFX9 */
630 struct si_shader_part
*prolog2
;
631 struct si_shader_part
*epilog
;
633 struct si_pm4_state
*pm4
;
634 struct si_resource
*bo
;
635 struct si_resource
*scratch_bo
;
636 struct si_shader_key key
;
637 struct util_queue_fence ready
;
638 bool compilation_failed
;
641 bool is_binary_shared
;
642 bool is_gs_copy_shader
;
644 /* The following data is all that's needed for binary shaders. */
645 struct si_shader_binary binary
;
646 struct ac_shader_config config
;
647 struct si_shader_info info
;
650 uint16_t ngg_emit_size
; /* in dwords */
651 uint16_t hw_max_esverts
;
652 uint16_t max_gsprims
;
653 uint16_t max_out_verts
;
654 uint16_t prim_amp_factor
;
655 bool max_vert_out_per_gs_instance
;
658 /* Shader key + LLVM IR + disassembly + statistics.
659 * Generated for debug contexts only.
662 size_t shader_log_size
;
664 struct gfx9_gs_info gs_info
;
666 /* For save precompute context registers values. */
669 unsigned vgt_gsvs_ring_offset_1
;
670 unsigned vgt_gsvs_ring_offset_2
;
671 unsigned vgt_gsvs_ring_offset_3
;
672 unsigned vgt_gsvs_ring_itemsize
;
673 unsigned vgt_gs_max_vert_out
;
674 unsigned vgt_gs_vert_itemsize
;
675 unsigned vgt_gs_vert_itemsize_1
;
676 unsigned vgt_gs_vert_itemsize_2
;
677 unsigned vgt_gs_vert_itemsize_3
;
678 unsigned vgt_gs_instance_cnt
;
679 unsigned vgt_gs_onchip_cntl
;
680 unsigned vgt_gs_max_prims_per_subgroup
;
681 unsigned vgt_esgs_ring_itemsize
;
685 unsigned ge_max_output_per_subgroup
;
686 unsigned ge_ngg_subgrp_cntl
;
687 unsigned vgt_primitiveid_en
;
688 unsigned vgt_gs_onchip_cntl
;
689 unsigned vgt_gs_instance_cnt
;
690 unsigned vgt_esgs_ring_itemsize
;
691 unsigned spi_vs_out_config
;
692 unsigned spi_shader_idx_format
;
693 unsigned spi_shader_pos_format
;
694 unsigned pa_cl_vte_cntl
;
695 unsigned pa_cl_ngg_cntl
;
696 unsigned vgt_gs_max_vert_out
; /* for API GS */
700 unsigned vgt_gs_mode
;
701 unsigned vgt_primitiveid_en
;
702 unsigned vgt_reuse_off
;
703 unsigned spi_vs_out_config
;
704 unsigned spi_shader_pos_format
;
705 unsigned pa_cl_vte_cntl
;
709 unsigned spi_ps_input_ena
;
710 unsigned spi_ps_input_addr
;
711 unsigned spi_baryc_cntl
;
712 unsigned spi_ps_in_control
;
713 unsigned spi_shader_z_format
;
714 unsigned spi_shader_col_format
;
715 unsigned cb_shader_mask
;
719 /*For save precompute registers value */
720 unsigned vgt_tf_param
; /* VGT_TF_PARAM */
721 unsigned vgt_vertex_reuse_block_cntl
; /* VGT_VERTEX_REUSE_BLOCK_CNTL */
722 unsigned pa_cl_vs_out_cntl
;
726 struct si_shader_part
{
727 struct si_shader_part
*next
;
728 union si_shader_part_key key
;
729 struct si_shader_binary binary
;
730 struct ac_shader_config config
;
735 si_generate_gs_copy_shader(struct si_screen
*sscreen
,
736 struct ac_llvm_compiler
*compiler
,
737 struct si_shader_selector
*gs_selector
,
738 struct pipe_debug_callback
*debug
);
739 int si_compile_shader(struct si_screen
*sscreen
,
740 struct ac_llvm_compiler
*compiler
,
741 struct si_shader
*shader
,
742 struct pipe_debug_callback
*debug
);
743 bool si_shader_create(struct si_screen
*sscreen
, struct ac_llvm_compiler
*compiler
,
744 struct si_shader
*shader
,
745 struct pipe_debug_callback
*debug
);
746 void si_shader_destroy(struct si_shader
*shader
);
747 unsigned si_shader_io_get_unique_index_patch(unsigned semantic_name
, unsigned index
);
748 unsigned si_shader_io_get_unique_index(unsigned semantic_name
, unsigned index
,
749 unsigned is_varying
);
750 bool si_shader_binary_upload(struct si_screen
*sscreen
, struct si_shader
*shader
,
751 uint64_t scratch_va
);
752 void si_shader_dump(struct si_screen
*sscreen
, struct si_shader
*shader
,
753 struct pipe_debug_callback
*debug
,
754 FILE *f
, bool check_debug_option
);
755 void si_shader_dump_stats_for_shader_db(struct si_screen
*screen
,
756 struct si_shader
*shader
,
757 struct pipe_debug_callback
*debug
);
758 void si_multiwave_lds_size_workaround(struct si_screen
*sscreen
,
760 const char *si_get_shader_name(const struct si_shader
*shader
);
761 void si_shader_binary_clean(struct si_shader_binary
*binary
);
763 /* si_shader_nir.c */
764 void si_nir_scan_shader(const struct nir_shader
*nir
,
765 struct tgsi_shader_info
*info
);
766 void si_nir_scan_tess_ctrl(const struct nir_shader
*nir
,
767 struct tgsi_tessctrl_info
*out
);
768 void si_nir_adjust_driver_locations(struct nir_shader
*nir
);
769 void si_finalize_nir(struct pipe_screen
*screen
, void *nirptr
, bool optimize
);
771 /* si_state_shaders.c */
772 void gfx9_get_gs_info(struct si_shader_selector
*es
,
773 struct si_shader_selector
*gs
,
774 struct gfx9_gs_info
*out
);
776 /* Inline helpers. */
778 /* Return the pointer to the main shader part's pointer. */
779 static inline struct si_shader
**
780 si_get_main_shader_part(struct si_shader_selector
*sel
,
781 struct si_shader_key
*key
)
784 return &sel
->main_shader_part_ls
;
785 if (key
->as_es
&& key
->as_ngg
)
786 return &sel
->main_shader_part_ngg_es
;
788 return &sel
->main_shader_part_es
;
790 return &sel
->main_shader_part_ngg
;
791 return &sel
->main_shader_part
;
795 gfx10_is_ngg_passthrough(struct si_shader
*shader
)
797 struct si_shader_selector
*sel
= shader
->selector
;
799 return sel
->type
!= PIPE_SHADER_GEOMETRY
&&
800 !sel
->so
.num_outputs
&&
801 !sel
->info
.writes_edgeflag
&&
802 (sel
->type
!= PIPE_SHADER_VERTEX
||
803 !shader
->key
.mono
.u
.vs_export_prim_id
);
807 si_shader_uses_bindless_samplers(struct si_shader_selector
*selector
)
809 return selector
? selector
->info
.uses_bindless_samplers
: false;
813 si_shader_uses_bindless_images(struct si_shader_selector
*selector
)
815 return selector
? selector
->info
.uses_bindless_images
: false;
818 void si_destroy_shader_selector(struct si_context
*sctx
,
819 struct si_shader_selector
*sel
);
822 si_shader_selector_reference(struct si_context
*sctx
,
823 struct si_shader_selector
**dst
,
824 struct si_shader_selector
*src
)
826 if (pipe_reference(&(*dst
)->reference
, &src
->reference
))
827 si_destroy_shader_selector(sctx
, *dst
);