2 * Copyright 2012 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
24 * Tom Stellard <thomas.stellard@amd.com>
25 * Michel Dänzer <michel.daenzer@amd.com>
26 * Christian König <christian.koenig@amd.com>
29 /* How linking shader inputs and outputs between vertex, tessellation, and
30 * geometry shaders works.
32 * Inputs and outputs between shaders are stored in a buffer. This buffer
33 * lives in LDS (typical case for tessellation), but it can also live
34 * in memory (ESGS). Each input or output has a fixed location within a vertex.
35 * The highest used input or output determines the stride between vertices.
37 * Since GS and tessellation are only possible in the OpenGL core profile,
38 * only these semantics are valid for per-vertex data:
45 * CULLDIST0..1 (not implemented)
48 * For example, a shader only writing GENERIC0 has the output stride of 5.
50 * Only these semantics are valid for per-patch data:
58 * That's how independent shaders agree on input and output locations.
59 * The si_shader_io_get_unique_index function assigns the locations.
61 * For tessellation, other required information for calculating the input and
62 * output addresses like the vertex stride, the patch stride, and the offsets
63 * where per-vertex and per-patch data start, is passed to the shader via
64 * user data SGPRs. The offsets and strides are calculated at draw time and
65 * aren't available at compile time.
71 #include <llvm-c/Core.h> /* LLVMModuleRef */
72 #include <llvm-c/TargetMachine.h>
73 #include "tgsi/tgsi_scan.h"
74 #include "util/u_queue.h"
77 struct ac_shader_binary
;
79 #define SI_MAX_VS_OUTPUTS 40
81 /* SGPR user data indices */
83 SI_SGPR_RW_BUFFERS
, /* rings (& stream-out, VS only) */
84 SI_SGPR_RW_BUFFERS_HI
,
85 SI_SGPR_CONST_BUFFERS
,
86 SI_SGPR_CONST_BUFFERS_HI
,
87 SI_SGPR_SAMPLERS
, /* images & sampler states interleaved */
91 SI_SGPR_SHADER_BUFFERS
,
92 SI_SGPR_SHADER_BUFFERS_HI
,
93 SI_NUM_RESOURCE_SGPRS
,
96 SI_SGPR_VERTEX_BUFFERS
= SI_NUM_RESOURCE_SGPRS
,
97 SI_SGPR_VERTEX_BUFFERS_HI
,
99 SI_SGPR_START_INSTANCE
,
101 SI_SGPR_VS_STATE_BITS
,
105 SI_SGPR_TES_OFFCHIP_LAYOUT
= SI_NUM_RESOURCE_SGPRS
,
106 SI_TES_NUM_USER_SGPR
,
108 /* GFX6-8: TCS only */
109 GFX6_SGPR_TCS_OFFCHIP_LAYOUT
= SI_NUM_RESOURCE_SGPRS
,
110 GFX6_SGPR_TCS_OUT_OFFSETS
,
111 GFX6_SGPR_TCS_OUT_LAYOUT
,
112 GFX6_SGPR_TCS_IN_LAYOUT
,
113 GFX6_TCS_NUM_USER_SGPR
,
115 /* GFX9: Merged LS-HS (VS-TCS) only. */
116 GFX9_SGPR_TCS_OFFCHIP_LAYOUT
= SI_VS_NUM_USER_SGPR
,
117 GFX9_SGPR_TCS_OUT_OFFSETS
,
118 GFX9_SGPR_TCS_OUT_LAYOUT
,
119 GFX9_SGPR_unused_to_align_the_next_pointer
,
120 GFX9_SGPR_TCS_CONST_BUFFERS
,
121 GFX9_SGPR_TCS_CONST_BUFFERS_HI
,
122 GFX9_SGPR_TCS_SAMPLERS
, /* images & sampler states interleaved */
123 GFX9_SGPR_TCS_SAMPLERS_HI
,
124 GFX9_SGPR_TCS_IMAGES
,
125 GFX9_SGPR_TCS_IMAGES_HI
,
126 GFX9_SGPR_TCS_SHADER_BUFFERS
,
127 GFX9_SGPR_TCS_SHADER_BUFFERS_HI
,
128 GFX9_TCS_NUM_USER_SGPR
,
131 SI_GS_NUM_USER_SGPR
= SI_NUM_RESOURCE_SGPRS
,
132 SI_GSCOPY_NUM_USER_SGPR
= SI_SGPR_RW_BUFFERS_HI
+ 1,
135 SI_SGPR_ALPHA_REF
= SI_NUM_RESOURCE_SGPRS
,
139 SI_SGPR_GRID_SIZE
= SI_NUM_RESOURCE_SGPRS
,
140 SI_SGPR_BLOCK_SIZE
= SI_SGPR_GRID_SIZE
+ 3,
141 SI_CS_NUM_USER_SGPR
= SI_SGPR_BLOCK_SIZE
+ 3
144 /* LLVM function parameter indices */
147 SI_PARAM_CONST_BUFFERS
,
150 SI_PARAM_SHADER_BUFFERS
,
151 SI_NUM_RESOURCE_PARAMS
,
153 /* VS only parameters */
154 SI_PARAM_VERTEX_BUFFERS
= SI_NUM_RESOURCE_PARAMS
,
155 SI_PARAM_BASE_VERTEX
,
156 SI_PARAM_START_INSTANCE
,
158 SI_PARAM_VS_STATE_BITS
,
160 /* Layout of TCS outputs in the offchip buffer
161 * [0:8] = the number of patches per threadgroup.
162 * [9:15] = the number of output vertices per patch.
163 * [16:31] = the offset of per patch attributes in the buffer in bytes.
165 SI_PARAM_TCS_OFFCHIP_LAYOUT
= SI_NUM_RESOURCE_PARAMS
, /* for TCS & TES */
167 /* TCS only parameters. */
169 /* Offsets where TCS outputs and TCS patch outputs live in LDS:
170 * [0:15] = TCS output patch0 offset / 16, max = NUM_PATCHES * 32 * 32
171 * [16:31] = TCS output patch0 offset for per-patch / 16, max = NUM_PATCHES*32*32* + 32*32
173 SI_PARAM_TCS_OUT_OFFSETS
,
175 /* Layout of TCS outputs / TES inputs:
176 * [0:12] = stride between output patches in dwords, num_outputs * num_vertices * 4, max = 32*32*4
177 * [13:20] = stride between output vertices in dwords = num_inputs * 4, max = 32*4
178 * [26:31] = gl_PatchVerticesIn, max = 32
180 SI_PARAM_TCS_OUT_LAYOUT
,
182 /* Layout of LS outputs / TCS inputs
183 * [8:20] = stride between patches in dwords = num_inputs * num_vertices * 4, max = 32*32*4
184 * [24:31] = stride between vertices in dwords = num_inputs * 4, max = 32*4
185 * (same layout as SI_PARAM_VS_STATE_BITS)
187 SI_PARAM_TCS_IN_LAYOUT
,
190 SI_PARAM_TESS_FACTOR_OFFSET
,
194 /* GS only parameters */
195 SI_PARAM_GS2VS_OFFSET
= SI_NUM_RESOURCE_PARAMS
,
197 SI_PARAM_VTX0_OFFSET
,
198 SI_PARAM_VTX1_OFFSET
,
199 SI_PARAM_PRIMITIVE_ID
,
200 SI_PARAM_VTX2_OFFSET
,
201 SI_PARAM_VTX3_OFFSET
,
202 SI_PARAM_VTX4_OFFSET
,
203 SI_PARAM_VTX5_OFFSET
,
204 SI_PARAM_GS_INSTANCE_ID
,
206 /* PS only parameters */
207 SI_PARAM_ALPHA_REF
= SI_NUM_RESOURCE_PARAMS
,
209 SI_PARAM_PERSP_SAMPLE
,
210 SI_PARAM_PERSP_CENTER
,
211 SI_PARAM_PERSP_CENTROID
,
212 SI_PARAM_PERSP_PULL_MODEL
,
213 SI_PARAM_LINEAR_SAMPLE
,
214 SI_PARAM_LINEAR_CENTER
,
215 SI_PARAM_LINEAR_CENTROID
,
216 SI_PARAM_LINE_STIPPLE_TEX
,
217 SI_PARAM_POS_X_FLOAT
,
218 SI_PARAM_POS_Y_FLOAT
,
219 SI_PARAM_POS_Z_FLOAT
,
220 SI_PARAM_POS_W_FLOAT
,
223 SI_PARAM_SAMPLE_COVERAGE
,
224 SI_PARAM_POS_FIXED_PT
,
226 /* CS only parameters */
227 SI_PARAM_GRID_SIZE
= SI_NUM_RESOURCE_PARAMS
,
232 SI_NUM_PARAMS
= SI_PARAM_POS_FIXED_PT
+ 9, /* +8 for COLOR[0..1] */
235 /* Fields of driver-defined VS state SGPR. */
236 /* Clamp vertex color output (only used in VS as VS). */
237 #define S_VS_STATE_CLAMP_VERTEX_COLOR(x) (((unsigned)(x) & 0x1) << 0)
238 #define C_VS_STATE_CLAMP_VERTEX_COLOR 0xFFFFFFFE
239 #define S_VS_STATE_INDEXED(x) (((unsigned)(x) & 0x1) << 1)
240 #define C_VS_STATE_INDEXED 0xFFFFFFFD
241 #define S_VS_STATE_LS_OUT_PATCH_SIZE(x) (((unsigned)(x) & 0x1FFF) << 8)
242 #define C_VS_STATE_LS_OUT_PATCH_SIZE 0xFFE000FF
243 #define S_VS_STATE_LS_OUT_VERTEX_SIZE(x) (((unsigned)(x) & 0xFF) << 24)
244 #define C_VS_STATE_LS_OUT_VERTEX_SIZE 0x00FFFFFF
246 /* SI-specific system values. */
248 TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI
= TGSI_SEMANTIC_COUNT
,
249 TGSI_SEMANTIC_DEFAULT_TESSINNER_SI
,
252 /* For VS shader key fix_fetch. */
254 SI_FIX_FETCH_NONE
= 0,
255 SI_FIX_FETCH_A2_SNORM
,
256 SI_FIX_FETCH_A2_SSCALED
,
257 SI_FIX_FETCH_A2_SINT
,
258 SI_FIX_FETCH_RGBA_32_UNORM
,
259 SI_FIX_FETCH_RGBX_32_UNORM
,
260 SI_FIX_FETCH_RGBA_32_SNORM
,
261 SI_FIX_FETCH_RGBX_32_SNORM
,
262 SI_FIX_FETCH_RGBA_32_USCALED
,
263 SI_FIX_FETCH_RGBA_32_SSCALED
,
264 SI_FIX_FETCH_RGBA_32_FIXED
,
265 SI_FIX_FETCH_RGBX_32_FIXED
,
266 SI_FIX_FETCH_RG_64_FLOAT
,
267 SI_FIX_FETCH_RGB_64_FLOAT
,
268 SI_FIX_FETCH_RGBA_64_FLOAT
,
269 SI_FIX_FETCH_RGB_8
, /* A = 1.0 */
270 SI_FIX_FETCH_RGB_8_INT
, /* A = 1 */
272 SI_FIX_FETCH_RGB_16_INT
,
277 /* State of the context creating the shader object. */
278 struct si_compiler_ctx_state
{
279 /* Should only be used by si_init_shader_selector_async and
280 * si_build_shader_variant if thread_index == -1 (non-threaded). */
281 LLVMTargetMachineRef tm
;
283 /* Used if thread_index == -1 or if debug.async is true. */
284 struct pipe_debug_callback debug
;
286 /* Used for creating the log string for gallium/ddebug. */
287 bool is_debug_context
;
290 /* A shader selector is a gallium CSO and contains shader variants and
291 * binaries for one TGSI program. This can be shared by multiple contexts.
293 struct si_shader_selector
{
294 struct si_screen
*screen
;
295 struct util_queue_fence ready
;
296 struct si_compiler_ctx_state compiler_ctx_state
;
299 struct si_shader
*first_variant
; /* immutable after the first variant */
300 struct si_shader
*last_variant
; /* mutable */
302 /* The compiled TGSI shader expecting a prolog and/or epilog (not
303 * uploaded to a buffer).
305 struct si_shader
*main_shader_part
;
306 struct si_shader
*main_shader_part_ls
; /* as_ls is set in the key */
307 struct si_shader
*main_shader_part_es
; /* as_es is set in the key */
309 struct si_shader
*gs_copy_shader
;
311 struct tgsi_token
*tokens
;
312 struct pipe_stream_output_info so
;
313 struct tgsi_shader_info info
;
315 /* PIPE_SHADER_[VERTEX|FRAGMENT|...] */
317 bool vs_needs_prolog
;
320 unsigned esgs_itemsize
;
321 unsigned gs_input_verts_per_prim
;
322 unsigned gs_output_prim
;
323 unsigned gs_max_out_vertices
;
324 unsigned gs_num_invocations
;
325 unsigned max_gs_stream
; /* count - 1 */
326 unsigned gsvs_vertex_size
;
327 unsigned max_gsvs_emit_size
;
330 unsigned color_attr_index
[2];
331 unsigned db_shader_control
;
332 /* Set 0xf or 0x0 (4 bits) per each written output.
333 * ANDed with spi_shader_col_format.
335 unsigned colors_written_4bit
;
340 uint64_t outputs_written
; /* "get_unique_index" bits */
341 uint32_t patch_outputs_written
; /* "get_unique_index" bits */
342 uint32_t outputs_written2
; /* "get_unique_index2" bits */
344 uint64_t inputs_read
; /* "get_unique_index" bits */
345 uint32_t inputs_read2
; /* "get_unique_index2" bits */
348 /* Valid shader configurations:
350 * API shaders VS | TCS | TES | GS |pass| PS
351 * are compiled as: | | | |thru|
353 * Only VS & PS: VS | | | | | PS
354 * GFX6 - with GS: ES | | | GS | VS | PS
355 * - with tess: LS | HS | VS | | | PS
356 * - with both: LS | HS | ES | GS | VS | PS
357 * GFX9 - with GS: -> | | | GS | VS | PS
358 * - with tess: -> | HS | VS | | | PS
359 * - with both: -> | HS | -> | GS | VS | PS
361 * -> = merged with the next stage
364 /* Common VS bits between the shader key and the prolog key. */
365 struct si_vs_prolog_bits
{
366 unsigned instance_divisors
[SI_MAX_ATTRIBS
];
369 /* Common VS bits between the shader key and the epilog key. */
370 struct si_vs_epilog_bits
{
371 unsigned export_prim_id
:1; /* when PS needs it and GS is disabled */
374 /* Common TCS bits between the shader key and the epilog key. */
375 struct si_tcs_epilog_bits
{
376 unsigned prim_mode
:3;
377 unsigned tes_reads_tess_factors
:1;
380 struct si_gs_prolog_bits
{
381 unsigned tri_strip_adj_fix
:1;
384 /* Common PS bits between the shader key and the prolog key. */
385 struct si_ps_prolog_bits
{
386 unsigned color_two_side
:1;
387 unsigned flatshade_colors
:1;
388 unsigned poly_stipple
:1;
389 unsigned force_persp_sample_interp
:1;
390 unsigned force_linear_sample_interp
:1;
391 unsigned force_persp_center_interp
:1;
392 unsigned force_linear_center_interp
:1;
393 unsigned bc_optimize_for_persp
:1;
394 unsigned bc_optimize_for_linear
:1;
397 /* Common PS bits between the shader key and the epilog key. */
398 struct si_ps_epilog_bits
{
399 unsigned spi_shader_col_format
;
400 unsigned color_is_int8
:8;
401 unsigned color_is_int10
:8;
402 unsigned last_cbuf
:3;
403 unsigned alpha_func
:3;
404 unsigned alpha_to_one
:1;
405 unsigned poly_line_smoothing
:1;
406 unsigned clamp_color
:1;
409 union si_shader_part_key
{
411 struct si_vs_prolog_bits states
;
412 unsigned num_input_sgprs
:6;
413 unsigned last_input
:4;
416 struct si_vs_epilog_bits states
;
417 unsigned prim_id_param_offset
:5;
420 struct si_tcs_epilog_bits states
;
423 struct si_gs_prolog_bits states
;
426 struct si_ps_prolog_bits states
;
427 unsigned num_input_sgprs
:6;
428 unsigned num_input_vgprs
:5;
429 /* Color interpolation and two-side color selection. */
430 unsigned colors_read
:8; /* color input components read */
431 unsigned num_interp_inputs
:5; /* BCOLOR is at this location */
432 unsigned face_vgpr_index
:5;
434 char color_attr_index
[2];
435 char color_interp_vgpr_index
[2]; /* -1 == constant */
438 struct si_ps_epilog_bits states
;
439 unsigned colors_written
:8;
441 unsigned writes_stencil
:1;
442 unsigned writes_samplemask
:1;
446 struct si_shader_key
{
447 /* Prolog and epilog flags. */
450 struct si_vs_prolog_bits prolog
;
451 struct si_vs_epilog_bits epilog
;
454 struct si_vs_prolog_bits ls_prolog
; /* for merged LS-HS */
455 struct si_shader_selector
*ls
; /* for merged LS-HS */
456 struct si_tcs_epilog_bits epilog
;
457 } tcs
; /* tessellation control shader */
459 struct si_vs_epilog_bits epilog
; /* same as VS */
460 } tes
; /* tessellation evaluation shader */
462 struct si_gs_prolog_bits prolog
;
465 struct si_ps_prolog_bits prolog
;
466 struct si_ps_epilog_bits epilog
;
470 /* These two are initially set according to the NEXT_SHADER property,
471 * or guessed if the property doesn't seem correct.
473 unsigned as_es
:1; /* export shader, which precedes GS */
474 unsigned as_ls
:1; /* local shader, which precedes TCS */
476 /* Flags for monolithic compilation only. */
478 /* One byte for every input: SI_FIX_FETCH_* enums. */
479 uint8_t vs_fix_fetch
[SI_MAX_ATTRIBS
];
480 uint64_t ff_tcs_inputs_to_copy
; /* for fixed-func TCS */
483 /* Optimization flags for asynchronous compilation only. */
486 uint64_t kill_outputs
; /* "get_unique_index" bits */
487 uint32_t kill_outputs2
; /* "get_unique_index2" bits */
488 unsigned clip_disable
:1;
489 } hw_vs
; /* HW VS (it can be VS, TES, GS) */
493 struct si_shader_config
{
496 unsigned spilled_sgprs
;
497 unsigned spilled_vgprs
;
498 unsigned private_mem_vgprs
;
500 unsigned spi_ps_input_ena
;
501 unsigned spi_ps_input_addr
;
503 unsigned scratch_bytes_per_wave
;
508 /* GCN-specific shader info. */
509 struct si_shader_info
{
510 ubyte vs_output_param_offset
[SI_MAX_VS_OUTPUTS
];
511 ubyte num_input_sgprs
;
512 ubyte num_input_vgprs
;
513 char face_vgpr_index
;
514 bool uses_instanceid
;
515 ubyte nr_pos_exports
;
516 ubyte nr_param_exports
;
520 struct si_compiler_ctx_state compiler_ctx_state
;
522 struct si_shader_selector
*selector
;
523 struct si_shader
*next_variant
;
525 struct si_shader_part
*prolog
;
526 struct si_shader
*previous_stage
; /* for GFX9 */
527 struct si_shader_part
*epilog
;
529 struct si_pm4_state
*pm4
;
530 struct r600_resource
*bo
;
531 struct r600_resource
*scratch_bo
;
532 struct si_shader_key key
;
533 struct util_queue_fence optimized_ready
;
534 bool compilation_failed
;
537 bool is_binary_shared
;
538 bool is_gs_copy_shader
;
540 /* The following data is all that's needed for binary shaders. */
541 struct ac_shader_binary binary
;
542 struct si_shader_config config
;
543 struct si_shader_info info
;
545 /* Shader key + LLVM IR + disassembly + statistics.
546 * Generated for debug contexts only.
549 size_t shader_log_size
;
552 struct si_shader_part
{
553 struct si_shader_part
*next
;
554 union si_shader_part_key key
;
555 struct ac_shader_binary binary
;
556 struct si_shader_config config
;
561 si_generate_gs_copy_shader(struct si_screen
*sscreen
,
562 LLVMTargetMachineRef tm
,
563 struct si_shader_selector
*gs_selector
,
564 struct pipe_debug_callback
*debug
);
565 int si_compile_tgsi_shader(struct si_screen
*sscreen
,
566 LLVMTargetMachineRef tm
,
567 struct si_shader
*shader
,
569 struct pipe_debug_callback
*debug
);
570 int si_shader_create(struct si_screen
*sscreen
, LLVMTargetMachineRef tm
,
571 struct si_shader
*shader
,
572 struct pipe_debug_callback
*debug
);
573 int si_compile_llvm(struct si_screen
*sscreen
,
574 struct ac_shader_binary
*binary
,
575 struct si_shader_config
*conf
,
576 LLVMTargetMachineRef tm
,
578 struct pipe_debug_callback
*debug
,
581 void si_shader_destroy(struct si_shader
*shader
);
582 unsigned si_shader_io_get_unique_index(unsigned semantic_name
, unsigned index
);
583 unsigned si_shader_io_get_unique_index2(unsigned name
, unsigned index
);
584 int si_shader_binary_upload(struct si_screen
*sscreen
, struct si_shader
*shader
);
585 void si_shader_dump(struct si_screen
*sscreen
, struct si_shader
*shader
,
586 struct pipe_debug_callback
*debug
, unsigned processor
,
587 FILE *f
, bool check_debug_option
);
588 void si_multiwave_lds_size_workaround(struct si_screen
*sscreen
,
590 void si_shader_apply_scratch_relocs(struct si_context
*sctx
,
591 struct si_shader
*shader
,
592 struct si_shader_config
*config
,
593 uint64_t scratch_va
);
594 void si_shader_binary_read_config(struct ac_shader_binary
*binary
,
595 struct si_shader_config
*conf
,
596 unsigned symbol_offset
);
597 unsigned si_get_spi_shader_z_format(bool writes_z
, bool writes_stencil
,
598 bool writes_samplemask
);
599 const char *si_get_shader_name(struct si_shader
*shader
, unsigned processor
);
601 /* Inline helpers. */
603 /* Return the pointer to the main shader part's pointer. */
604 static inline struct si_shader
**
605 si_get_main_shader_part(struct si_shader_selector
*sel
,
606 struct si_shader_key
*key
)
609 return &sel
->main_shader_part_ls
;
611 return &sel
->main_shader_part_es
;
612 return &sel
->main_shader_part
;