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 "util/u_inlines.h"
137 #include "util/u_queue.h"
138 #include "util/simple_mtx.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
,
215 /* The value has to be 12, because the hw requires that descriptors
216 * are aligned to 4 SGPRs.
218 SI_SGPR_VS_VB_DESCRIPTOR_FIRST
= 12,
221 /* LLVM function parameter indices */
223 SI_NUM_RESOURCE_PARAMS
= 4,
225 /* PS only parameters */
226 SI_PARAM_ALPHA_REF
= SI_NUM_RESOURCE_PARAMS
,
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
,
242 SI_PARAM_SAMPLE_COVERAGE
,
243 SI_PARAM_POS_FIXED_PT
,
245 SI_NUM_PARAMS
= SI_PARAM_POS_FIXED_PT
+ 9, /* +8 for COLOR[0..1] */
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
265 /* Use a property enum that CS wouldn't use. */
266 TGSI_PROPERTY_CS_LOCAL_SIZE
= TGSI_PROPERTY_FS_COORD_ORIGIN
,
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,
275 * For VS shader keys, describe any fixups required for vertex fetch.
277 * \ref log_size, \ref format, and the number of channels are interpreted as
278 * by \ref ac_build_opencoded_load_format.
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).
284 union si_vs_fix_fetch
{
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 */
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
;
302 /* Used if thread_index == -1 or if debug.async is true. */
303 struct pipe_debug_callback debug
;
305 /* Used for creating the log string for gallium/ddebug. */
306 bool is_debug_context
;
309 struct si_shader_info
{
314 ubyte input_semantic_name
[PIPE_MAX_SHADER_INPUTS
]; /**< TGSI_SEMANTIC_x */
315 ubyte input_semantic_index
[PIPE_MAX_SHADER_INPUTS
];
316 ubyte input_interpolate
[PIPE_MAX_SHADER_INPUTS
];
317 ubyte input_interpolate_loc
[PIPE_MAX_SHADER_INPUTS
];
318 ubyte input_usage_mask
[PIPE_MAX_SHADER_INPUTS
];
319 ubyte input_cylindrical_wrap
[PIPE_MAX_SHADER_INPUTS
];
320 ubyte output_semantic_name
[PIPE_MAX_SHADER_OUTPUTS
]; /**< TGSI_SEMANTIC_x */
321 ubyte output_semantic_index
[PIPE_MAX_SHADER_OUTPUTS
];
322 ubyte output_usagemask
[PIPE_MAX_SHADER_OUTPUTS
];
323 ubyte output_streams
[PIPE_MAX_SHADER_OUTPUTS
];
325 ubyte num_system_values
;
326 ubyte system_value_semantic_name
[PIPE_MAX_SHADER_INPUTS
];
330 uint file_mask
[TGSI_FILE_COUNT
]; /**< bitmask of declared registers */
331 uint file_count
[TGSI_FILE_COUNT
]; /**< number of declared registers */
332 int file_max
[TGSI_FILE_COUNT
]; /**< highest index of declared registers */
333 int const_file_max
[PIPE_MAX_CONSTANT_BUFFERS
];
334 unsigned const_buffers_declared
; /**< bitmask of declared const buffers */
335 unsigned samplers_declared
; /**< bitmask of declared samplers */
336 ubyte sampler_targets
[PIPE_MAX_SHADER_SAMPLER_VIEWS
]; /**< TGSI_TEXTURE_x values */
337 ubyte sampler_type
[PIPE_MAX_SHADER_SAMPLER_VIEWS
]; /**< TGSI_RETURN_TYPE_x */
338 ubyte num_stream_output_components
[4];
340 ubyte input_array_first
[PIPE_MAX_SHADER_INPUTS
];
341 ubyte input_array_last
[PIPE_MAX_SHADER_INPUTS
];
342 ubyte output_array_first
[PIPE_MAX_SHADER_OUTPUTS
];
343 ubyte output_array_last
[PIPE_MAX_SHADER_OUTPUTS
];
344 unsigned array_max
[TGSI_FILE_COUNT
]; /**< highest index array per register file */
346 uint immediate_count
; /**< number of immediates declared */
347 uint num_instructions
;
348 uint num_memory_instructions
; /**< sampler, buffer, and image instructions */
350 uint opcode_count
[TGSI_OPCODE_LAST
]; /**< opcode histogram */
353 * If a tessellation control shader reads outputs, this describes which ones.
355 boolean reads_pervertex_outputs
;
356 boolean reads_perpatch_outputs
;
357 boolean reads_tessfactor_outputs
;
359 ubyte colors_read
; /**< which color components are read by the FS */
360 ubyte colors_written
;
361 boolean reads_position
; /**< does fragment shader read position? */
362 boolean reads_z
; /**< does fragment shader read depth? */
363 boolean reads_samplemask
; /**< does fragment shader read sample mask? */
364 boolean reads_tess_factors
; /**< If TES reads TESSINNER or TESSOUTER */
365 boolean writes_z
; /**< does fragment shader write Z value? */
366 boolean writes_stencil
; /**< does fragment shader write stencil value? */
367 boolean writes_samplemask
; /**< does fragment shader write sample mask? */
368 boolean writes_edgeflag
; /**< vertex shader outputs edgeflag */
369 boolean uses_kill
; /**< KILL or KILL_IF instruction used? */
370 boolean uses_persp_center
;
371 boolean uses_persp_centroid
;
372 boolean uses_persp_sample
;
373 boolean uses_linear_center
;
374 boolean uses_linear_centroid
;
375 boolean uses_linear_sample
;
376 boolean uses_persp_opcode_interp_centroid
;
377 boolean uses_persp_opcode_interp_offset
;
378 boolean uses_persp_opcode_interp_sample
;
379 boolean uses_linear_opcode_interp_centroid
;
380 boolean uses_linear_opcode_interp_offset
;
381 boolean uses_linear_opcode_interp_sample
;
382 boolean uses_instanceid
;
383 boolean uses_vertexid
;
384 boolean uses_vertexid_nobase
;
385 boolean uses_basevertex
;
388 boolean uses_frontface
;
389 boolean uses_invocationid
;
390 boolean uses_thread_id
[3];
391 boolean uses_block_id
[3];
392 boolean uses_block_size
;
393 boolean uses_grid_size
;
394 boolean uses_subgroup_info
;
395 boolean writes_position
;
396 boolean writes_psize
;
397 boolean writes_clipvertex
;
398 boolean writes_primid
;
399 boolean writes_viewport_index
;
400 boolean writes_layer
;
401 boolean writes_memory
; /**< contains stores or atomics to buffers or images */
402 boolean uses_doubles
; /**< uses any of the double instructions */
403 boolean uses_derivatives
;
404 boolean uses_bindless_samplers
;
405 boolean uses_bindless_images
;
406 boolean uses_fbfetch
;
407 unsigned clipdist_writemask
;
408 unsigned culldist_writemask
;
409 unsigned num_written_culldistance
;
410 unsigned num_written_clipdistance
;
412 unsigned images_declared
; /**< bitmask of declared images */
413 unsigned msaa_images_declared
; /**< bitmask of declared MSAA images */
416 * Bitmask indicating which declared image is a buffer.
418 unsigned images_buffers
;
419 unsigned images_load
; /**< bitmask of images using loads */
420 unsigned images_store
; /**< bitmask of images using stores */
421 unsigned images_atomic
; /**< bitmask of images using atomics */
422 unsigned shader_buffers_declared
; /**< bitmask of declared shader buffers */
423 unsigned shader_buffers_load
; /**< bitmask of shader buffers using loads */
424 unsigned shader_buffers_store
; /**< bitmask of shader buffers using stores */
425 unsigned shader_buffers_atomic
; /**< bitmask of shader buffers using atomics */
426 bool uses_bindless_buffer_load
;
427 bool uses_bindless_buffer_store
;
428 bool uses_bindless_buffer_atomic
;
429 bool uses_bindless_image_load
;
430 bool uses_bindless_image_store
;
431 bool uses_bindless_image_atomic
;
434 * Bitmask indicating which register files are accessed with
435 * indirect addressing. The bits are (1 << TGSI_FILE_x), etc.
437 unsigned indirect_files
;
439 * Bitmask indicating which register files are read / written with
440 * indirect addressing. The bits are (1 << TGSI_FILE_x).
442 unsigned indirect_files_read
;
443 unsigned indirect_files_written
;
444 unsigned dim_indirect_files
; /**< shader resource indexing */
445 unsigned const_buffers_indirect
; /**< const buffers using indirect addressing */
447 unsigned properties
[TGSI_PROPERTY_COUNT
]; /* index with TGSI_PROPERTY_ */
450 * Max nesting limit of loops/if's
454 /** Whether all codepaths write tess factors in all invocations. */
455 bool tessfactors_are_def_in_all_invocs
;
458 /* A shader selector is a gallium CSO and contains shader variants and
459 * binaries for one NIR program. This can be shared by multiple contexts.
461 struct si_shader_selector
{
462 struct pipe_reference reference
;
463 struct si_screen
*screen
;
464 struct util_queue_fence ready
;
465 struct si_compiler_ctx_state compiler_ctx_state
;
468 struct si_shader
*first_variant
; /* immutable after the first variant */
469 struct si_shader
*last_variant
; /* mutable */
471 /* The compiled NIR shader without a prolog and/or epilog (not
472 * uploaded to a buffer object).
474 struct si_shader
*main_shader_part
;
475 struct si_shader
*main_shader_part_ls
; /* as_ls is set in the key */
476 struct si_shader
*main_shader_part_es
; /* as_es is set in the key */
477 struct si_shader
*main_shader_part_ngg
; /* as_ngg is set in the key */
478 struct si_shader
*main_shader_part_ngg_es
; /* for Wave32 TES before legacy GS */
480 struct si_shader
*gs_copy_shader
;
482 struct nir_shader
*nir
;
486 struct pipe_stream_output_info so
;
487 struct si_shader_info info
;
489 /* PIPE_SHADER_[VERTEX|FRAGMENT|...] */
490 enum pipe_shader_type type
;
491 bool vs_needs_prolog
;
492 bool force_correct_derivs_after_kill
;
493 bool prim_discard_cs_allowed
;
494 unsigned num_vs_inputs
;
495 unsigned num_vbos_in_user_sgprs
;
496 unsigned pa_cl_vs_out_cntl
;
502 unsigned esgs_itemsize
; /* vertex stride */
503 unsigned lshs_vertex_stride
;
506 unsigned gs_input_verts_per_prim
;
507 unsigned gs_output_prim
;
508 unsigned gs_max_out_vertices
;
509 unsigned gs_num_invocations
;
510 unsigned max_gs_stream
; /* count - 1 */
511 unsigned gsvs_vertex_size
;
512 unsigned max_gsvs_emit_size
;
513 unsigned enabled_streamout_buffer_mask
;
514 bool tess_turns_off_ngg
;
517 unsigned color_attr_index
[2];
518 unsigned db_shader_control
;
519 /* Set 0xf or 0x0 (4 bits) per each written output.
520 * ANDed with spi_shader_col_format.
522 unsigned colors_written_4bit
;
524 uint64_t outputs_written_before_ps
; /* "get_unique_index" bits */
525 uint64_t outputs_written
; /* "get_unique_index" bits */
526 uint32_t patch_outputs_written
; /* "get_unique_index_patch" bits */
528 uint64_t inputs_read
; /* "get_unique_index" bits */
530 /* bitmasks of used descriptor slots */
531 uint32_t active_const_and_shader_buffers
;
532 uint64_t active_samplers_and_images
;
535 /* Valid shader configurations:
537 * API shaders VS | TCS | TES | GS |pass| PS
538 * are compiled as: | | | |thru|
540 * Only VS & PS: VS | | | | | PS
541 * GFX6 - with GS: ES | | | GS | VS | PS
542 * - with tess: LS | HS | VS | | | PS
543 * - with both: LS | HS | ES | GS | VS | PS
544 * GFX9 - with GS: -> | | | GS | VS | PS
545 * - with tess: -> | HS | VS | | | PS
546 * - with both: -> | HS | -> | GS | VS | PS
548 * NGG - VS & PS: GS | | | | | PS
549 * (GFX10+) - with GS: -> | | | GS | | PS
550 * - with tess: -> | HS | GS | | | PS
551 * - with both: -> | HS | -> | GS | | PS
553 * -> = merged with the next stage
556 /* Use the byte alignment for all following structure members for optimal
557 * shader key memory footprint.
559 #pragma pack(push, 1)
561 /* Common VS bits between the shader key and the prolog key. */
562 struct si_vs_prolog_bits
{
563 /* - If neither "is_one" nor "is_fetched" has a bit set, the instance
565 * - If "is_one" has a bit set, the instance divisor is 1.
566 * - If "is_fetched" has a bit set, the instance divisor will be loaded
567 * from the constant buffer.
569 uint16_t instance_divisor_is_one
; /* bitmask of inputs */
570 uint16_t instance_divisor_is_fetched
; /* bitmask of inputs */
571 unsigned ls_vgpr_fix
:1;
572 unsigned unpack_instance_id_from_vertex_id
:1;
575 /* Common TCS bits between the shader key and the epilog key. */
576 struct si_tcs_epilog_bits
{
577 unsigned prim_mode
:3;
578 unsigned invoc0_tess_factors_are_def
:1;
579 unsigned tes_reads_tess_factors
:1;
582 struct si_gs_prolog_bits
{
583 unsigned tri_strip_adj_fix
:1;
584 unsigned gfx9_prev_is_vs
:1;
587 /* Common PS bits between the shader key and the prolog key. */
588 struct si_ps_prolog_bits
{
589 unsigned color_two_side
:1;
590 unsigned flatshade_colors
:1;
591 unsigned poly_stipple
:1;
592 unsigned force_persp_sample_interp
:1;
593 unsigned force_linear_sample_interp
:1;
594 unsigned force_persp_center_interp
:1;
595 unsigned force_linear_center_interp
:1;
596 unsigned bc_optimize_for_persp
:1;
597 unsigned bc_optimize_for_linear
:1;
598 unsigned samplemask_log_ps_iter
:3;
601 /* Common PS bits between the shader key and the epilog key. */
602 struct si_ps_epilog_bits
{
603 unsigned spi_shader_col_format
;
604 unsigned color_is_int8
:8;
605 unsigned color_is_int10
:8;
606 unsigned last_cbuf
:3;
607 unsigned alpha_func
:3;
608 unsigned alpha_to_one
:1;
609 unsigned poly_line_smoothing
:1;
610 unsigned clamp_color
:1;
613 union si_shader_part_key
{
615 struct si_vs_prolog_bits states
;
616 unsigned num_input_sgprs
:6;
617 /* For merged stages such as LS-HS, HS input VGPRs are first. */
618 unsigned num_merged_next_stage_vgprs
:3;
619 unsigned num_inputs
:5;
623 /* Prologs for monolithic shaders shouldn't set EXEC. */
624 unsigned is_monolithic
:1;
627 struct si_tcs_epilog_bits states
;
630 struct si_gs_prolog_bits states
;
631 /* Prologs of monolithic shaders shouldn't set EXEC. */
632 unsigned is_monolithic
:1;
636 struct si_ps_prolog_bits states
;
637 unsigned num_input_sgprs
:6;
638 unsigned num_input_vgprs
:5;
639 /* Color interpolation and two-side color selection. */
640 unsigned colors_read
:8; /* color input components read */
641 unsigned num_interp_inputs
:5; /* BCOLOR is at this location */
642 unsigned face_vgpr_index
:5;
643 unsigned ancillary_vgpr_index
:5;
645 char color_attr_index
[2];
646 signed char color_interp_vgpr_index
[2]; /* -1 == constant */
649 struct si_ps_epilog_bits states
;
650 unsigned colors_written
:8;
652 unsigned writes_stencil
:1;
653 unsigned writes_samplemask
:1;
657 struct si_shader_key
{
658 /* Prolog and epilog flags. */
661 struct si_vs_prolog_bits prolog
;
664 struct si_vs_prolog_bits ls_prolog
; /* for merged LS-HS */
665 struct si_shader_selector
*ls
; /* for merged LS-HS */
666 struct si_tcs_epilog_bits epilog
;
667 } tcs
; /* tessellation control shader */
669 struct si_vs_prolog_bits vs_prolog
; /* for merged ES-GS */
670 struct si_shader_selector
*es
; /* for merged ES-GS */
671 struct si_gs_prolog_bits prolog
;
674 struct si_ps_prolog_bits prolog
;
675 struct si_ps_epilog_bits epilog
;
679 /* These three are initially set according to the NEXT_SHADER property,
680 * or guessed if the property doesn't seem correct.
682 unsigned as_es
:1; /* export shader, which precedes GS */
683 unsigned as_ls
:1; /* local shader, which precedes TCS */
684 unsigned as_ngg
:1; /* VS, TES, or GS compiled as NGG primitive shader */
686 /* Flags for monolithic compilation only. */
688 /* Whether fetch should be opencoded according to vs_fix_fetch.
689 * Otherwise, if vs_fix_fetch is non-zero, buffer_load_format_xyzw
690 * with minimal fixups is used. */
691 uint16_t vs_fetch_opencode
;
692 union si_vs_fix_fetch vs_fix_fetch
[SI_MAX_ATTRIBS
];
695 uint64_t ff_tcs_inputs_to_copy
; /* for fixed-func TCS */
696 /* When PS needs PrimID and GS is disabled. */
697 unsigned vs_export_prim_id
:1;
699 unsigned interpolate_at_sample_force_center
:1;
700 unsigned fbfetch_msaa
:1;
701 unsigned fbfetch_is_1D
:1;
702 unsigned fbfetch_layered
:1;
707 /* Optimization flags for asynchronous compilation only. */
709 /* For HW VS (it can be VS, TES, GS) */
710 uint64_t kill_outputs
; /* "get_unique_index" bits */
711 unsigned clip_disable
:1;
713 /* For shaders where monolithic variants have better code.
715 * This is a flag that has no effect on code generation,
716 * but forces monolithic shaders to be used as soon as
717 * possible, because it's in the "opt" group.
719 unsigned prefer_mono
:1;
721 /* Primitive discard compute shader. */
722 unsigned vs_as_prim_discard_cs
:1;
723 unsigned cs_prim_type
:4;
724 unsigned cs_indexed
:1;
725 unsigned cs_instancing
:1;
726 unsigned cs_primitive_restart
:1;
727 unsigned cs_provoking_vertex_first
:1;
728 unsigned cs_need_correct_orientation
:1;
729 unsigned cs_cull_front
:1;
730 unsigned cs_cull_back
:1;
731 unsigned cs_cull_z
:1;
732 unsigned cs_halfz_clip_space
:1;
736 /* Restore the pack alignment to default. */
739 /* GCN-specific shader info. */
740 struct si_shader_binary_info
{
741 ubyte vs_output_param_offset
[SI_MAX_VS_OUTPUTS
];
742 ubyte num_input_sgprs
;
743 ubyte num_input_vgprs
;
744 signed char face_vgpr_index
;
745 signed char ancillary_vgpr_index
;
746 bool uses_instanceid
;
747 ubyte nr_pos_exports
;
748 ubyte nr_param_exports
;
749 unsigned private_mem_vgprs
;
750 unsigned max_simd_waves
;
753 struct si_shader_binary
{
754 const char *elf_buffer
;
757 char *llvm_ir_string
;
760 struct gfx9_gs_info
{
761 unsigned es_verts_per_subgroup
;
762 unsigned gs_prims_per_subgroup
;
763 unsigned gs_inst_prims_in_subgroup
;
764 unsigned max_prims_per_subgroup
;
765 unsigned esgs_ring_size
; /* in bytes */
769 struct si_compiler_ctx_state compiler_ctx_state
;
771 struct si_shader_selector
*selector
;
772 struct si_shader_selector
*previous_stage_sel
; /* for refcounting */
773 struct si_shader
*next_variant
;
775 struct si_shader_part
*prolog
;
776 struct si_shader
*previous_stage
; /* for GFX9 */
777 struct si_shader_part
*prolog2
;
778 struct si_shader_part
*epilog
;
780 struct si_pm4_state
*pm4
;
781 struct si_resource
*bo
;
782 struct si_resource
*scratch_bo
;
783 struct si_shader_key key
;
784 struct util_queue_fence ready
;
785 bool compilation_failed
;
788 bool is_binary_shared
;
789 bool is_gs_copy_shader
;
791 /* The following data is all that's needed for binary shaders. */
792 struct si_shader_binary binary
;
793 struct ac_shader_config config
;
794 struct si_shader_binary_info info
;
797 uint16_t ngg_emit_size
; /* in dwords */
798 uint16_t hw_max_esverts
;
799 uint16_t max_gsprims
;
800 uint16_t max_out_verts
;
801 uint16_t prim_amp_factor
;
802 bool max_vert_out_per_gs_instance
;
805 /* Shader key + LLVM IR + disassembly + statistics.
806 * Generated for debug contexts only.
809 size_t shader_log_size
;
811 struct gfx9_gs_info gs_info
;
813 /* For save precompute context registers values. */
816 unsigned vgt_gsvs_ring_offset_1
;
817 unsigned vgt_gsvs_ring_offset_2
;
818 unsigned vgt_gsvs_ring_offset_3
;
819 unsigned vgt_gsvs_ring_itemsize
;
820 unsigned vgt_gs_max_vert_out
;
821 unsigned vgt_gs_vert_itemsize
;
822 unsigned vgt_gs_vert_itemsize_1
;
823 unsigned vgt_gs_vert_itemsize_2
;
824 unsigned vgt_gs_vert_itemsize_3
;
825 unsigned vgt_gs_instance_cnt
;
826 unsigned vgt_gs_onchip_cntl
;
827 unsigned vgt_gs_max_prims_per_subgroup
;
828 unsigned vgt_esgs_ring_itemsize
;
832 unsigned ge_max_output_per_subgroup
;
833 unsigned ge_ngg_subgrp_cntl
;
834 unsigned vgt_primitiveid_en
;
835 unsigned vgt_gs_onchip_cntl
;
836 unsigned vgt_gs_instance_cnt
;
837 unsigned vgt_esgs_ring_itemsize
;
838 unsigned spi_vs_out_config
;
839 unsigned spi_shader_idx_format
;
840 unsigned spi_shader_pos_format
;
841 unsigned pa_cl_vte_cntl
;
842 unsigned pa_cl_ngg_cntl
;
843 unsigned vgt_gs_max_vert_out
; /* for API GS */
847 unsigned vgt_gs_mode
;
848 unsigned vgt_primitiveid_en
;
849 unsigned vgt_reuse_off
;
850 unsigned spi_vs_out_config
;
851 unsigned spi_shader_pos_format
;
852 unsigned pa_cl_vte_cntl
;
856 unsigned spi_ps_input_ena
;
857 unsigned spi_ps_input_addr
;
858 unsigned spi_baryc_cntl
;
859 unsigned spi_ps_in_control
;
860 unsigned spi_shader_z_format
;
861 unsigned spi_shader_col_format
;
862 unsigned cb_shader_mask
;
866 /*For save precompute registers value */
867 unsigned vgt_tf_param
; /* VGT_TF_PARAM */
868 unsigned vgt_vertex_reuse_block_cntl
; /* VGT_VERTEX_REUSE_BLOCK_CNTL */
869 unsigned pa_cl_vs_out_cntl
;
873 struct si_shader_part
{
874 struct si_shader_part
*next
;
875 union si_shader_part_key key
;
876 struct si_shader_binary binary
;
877 struct ac_shader_config config
;
882 si_generate_gs_copy_shader(struct si_screen
*sscreen
,
883 struct ac_llvm_compiler
*compiler
,
884 struct si_shader_selector
*gs_selector
,
885 struct pipe_debug_callback
*debug
);
886 int si_compile_shader(struct si_screen
*sscreen
,
887 struct ac_llvm_compiler
*compiler
,
888 struct si_shader
*shader
,
889 struct pipe_debug_callback
*debug
);
890 bool si_shader_create(struct si_screen
*sscreen
, struct ac_llvm_compiler
*compiler
,
891 struct si_shader
*shader
,
892 struct pipe_debug_callback
*debug
);
893 void si_shader_destroy(struct si_shader
*shader
);
894 unsigned si_shader_io_get_unique_index_patch(unsigned semantic_name
, unsigned index
);
895 unsigned si_shader_io_get_unique_index(unsigned semantic_name
, unsigned index
,
896 unsigned is_varying
);
897 bool si_shader_binary_upload(struct si_screen
*sscreen
, struct si_shader
*shader
,
898 uint64_t scratch_va
);
899 void si_shader_dump(struct si_screen
*sscreen
, struct si_shader
*shader
,
900 struct pipe_debug_callback
*debug
,
901 FILE *f
, bool check_debug_option
);
902 void si_shader_dump_stats_for_shader_db(struct si_screen
*screen
,
903 struct si_shader
*shader
,
904 struct pipe_debug_callback
*debug
);
905 void si_multiwave_lds_size_workaround(struct si_screen
*sscreen
,
907 const char *si_get_shader_name(const struct si_shader
*shader
);
908 void si_shader_binary_clean(struct si_shader_binary
*binary
);
910 /* si_shader_nir.c */
911 void si_nir_scan_shader(const struct nir_shader
*nir
,
912 struct si_shader_info
*info
);
913 void si_nir_adjust_driver_locations(struct nir_shader
*nir
);
914 void si_finalize_nir(struct pipe_screen
*screen
, void *nirptr
, bool optimize
);
916 /* si_state_shaders.c */
917 void gfx9_get_gs_info(struct si_shader_selector
*es
,
918 struct si_shader_selector
*gs
,
919 struct gfx9_gs_info
*out
);
921 /* Inline helpers. */
923 /* Return the pointer to the main shader part's pointer. */
924 static inline struct si_shader
**
925 si_get_main_shader_part(struct si_shader_selector
*sel
,
926 struct si_shader_key
*key
)
929 return &sel
->main_shader_part_ls
;
930 if (key
->as_es
&& key
->as_ngg
)
931 return &sel
->main_shader_part_ngg_es
;
933 return &sel
->main_shader_part_es
;
935 return &sel
->main_shader_part_ngg
;
936 return &sel
->main_shader_part
;
940 gfx10_is_ngg_passthrough(struct si_shader
*shader
)
942 struct si_shader_selector
*sel
= shader
->selector
;
944 return sel
->type
!= PIPE_SHADER_GEOMETRY
&&
945 !sel
->so
.num_outputs
&&
946 !sel
->info
.writes_edgeflag
&&
947 (sel
->type
!= PIPE_SHADER_VERTEX
||
948 !shader
->key
.mono
.u
.vs_export_prim_id
);
952 si_shader_uses_bindless_samplers(struct si_shader_selector
*selector
)
954 return selector
? selector
->info
.uses_bindless_samplers
: false;
958 si_shader_uses_bindless_images(struct si_shader_selector
*selector
)
960 return selector
? selector
->info
.uses_bindless_images
: false;
963 void si_destroy_shader_selector(struct si_context
*sctx
,
964 struct si_shader_selector
*sel
);
967 si_shader_selector_reference(struct si_context
*sctx
,
968 struct si_shader_selector
**dst
,
969 struct si_shader_selector
*src
)
971 if (pipe_reference(&(*dst
)->reference
, &src
->reference
))
972 si_destroy_shader_selector(sctx
, *dst
);