• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2012 Advanced Micro Devices, Inc.
3  *
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:
10  *
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
13  * Software.
14  *
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.
22  */
23 
24 /* The compiler middle-end architecture: Explaining (non-)monolithic shaders
25  * -------------------------------------------------------------------------
26  *
27  * Typically, there is one-to-one correspondence between API and HW shaders,
28  * that is, for every API shader, there is exactly one shader binary in
29  * the driver.
30  *
31  * The problem with that is that we also have to emulate some API states
32  * (e.g. alpha-test, and many others) in shaders too. The two obvious ways
33  * to deal with it are:
34  * - each shader has multiple variants for each combination of emulated states,
35  *   and the variants are compiled on demand, possibly relying on a shader
36  *   cache for good performance
37  * - patch shaders at the binary level
38  *
39  * This driver uses something completely different. The emulated states are
40  * usually implemented at the beginning or end of shaders. Therefore, we can
41  * split the shader into 3 parts:
42  * - prolog part (shader code dependent on states)
43  * - main part (the API shader)
44  * - epilog part (shader code dependent on states)
45  *
46  * Each part is compiled as a separate shader and the final binaries are
47  * concatenated. This type of shader is called non-monolithic, because it
48  * consists of multiple independent binaries. Creating a new shader variant
49  * is therefore only a concatenation of shader parts (binaries) and doesn't
50  * involve any compilation. The main shader parts are the only parts that are
51  * compiled when applications create shader objects. The prolog and epilog
52  * parts are compiled on the first use and saved, so that their binaries can
53  * be reused by many other shaders.
54  *
55  * One of the roles of the prolog part is to compute vertex buffer addresses
56  * for vertex shaders. A few of the roles of the epilog part are color buffer
57  * format conversions in pixel shaders that we have to do manually, and write
58  * tessellation factors in tessellation control shaders. The prolog and epilog
59  * have many other important responsibilities in various shader stages.
60  * They don't just "emulate legacy stuff".
61  *
62  * Monolithic shaders are shaders where the parts are combined before LLVM
63  * compilation, and the whole thing is compiled and optimized as one unit with
64  * one binary on the output. The result is the same as the non-monolithic
65  * shader, but the final code can be better, because LLVM can optimize across
66  * all shader parts. Monolithic shaders aren't usually used except for these
67  * special cases:
68  *
69  * 1) Some rarely-used states require modification of the main shader part
70  *    itself, and in such cases, only the monolithic shader variant is
71  *    compiled, and that's always done on the first use.
72  *
73  * 2) When we do cross-stage optimizations for separate shader objects and
74  *    e.g. eliminate unused shader varyings, the resulting optimized shader
75  *    variants are always compiled as monolithic shaders, and always
76  *    asynchronously (i.e. not stalling ongoing rendering). We call them
77  *    "optimized monolithic" shaders. The important property here is that
78  *    the non-monolithic unoptimized shader variant is always available for use
79  *    when the asynchronous compilation of the optimized shader is not done
80  *    yet.
81  *
82  * Starting with GFX9 chips, some shader stages are merged, and the number of
83  * shader parts per shader increased. The complete new list of shader parts is:
84  * - 1st shader: prolog part
85  * - 1st shader: main part
86  * - 2nd shader: prolog part
87  * - 2nd shader: main part
88  * - 2nd shader: epilog part
89  */
90 
91 /* How linking shader inputs and outputs between vertex, tessellation, and
92  * geometry shaders works.
93  *
94  * Inputs and outputs between shaders are stored in a buffer. This buffer
95  * lives in LDS (typical case for tessellation), but it can also live
96  * in memory (ESGS). Each input or output has a fixed location within a vertex.
97  * The highest used input or output determines the stride between vertices.
98  *
99  * Since GS and tessellation are only possible in the OpenGL core profile,
100  * only these semantics are valid for per-vertex data:
101  *
102  *   Name             Location
103  *
104  *   POSITION         0
105  *   PSIZE            1
106  *   CLIPDIST0..1     2..3
107  *   CULLDIST0..1     (not implemented)
108  *   GENERIC0..31     4..35
109  *
110  * For example, a shader only writing GENERIC0 has the output stride of 5.
111  *
112  * Only these semantics are valid for per-patch data:
113  *
114  *   Name             Location
115  *
116  *   TESSOUTER        0
117  *   TESSINNER        1
118  *   PATCH0..29       2..31
119  *
120  * That's how independent shaders agree on input and output locations.
121  * The si_shader_io_get_unique_index function assigns the locations.
122  *
123  * For tessellation, other required information for calculating the input and
124  * output addresses like the vertex stride, the patch stride, and the offsets
125  * where per-vertex and per-patch data start, is passed to the shader via
126  * user data SGPRs. The offsets and strides are calculated at draw time and
127  * aren't available at compile time.
128  */
129 
130 #ifndef SI_SHADER_H
131 #define SI_SHADER_H
132 
133 #include <llvm-c/Core.h> /* LLVMModuleRef */
134 #include <llvm-c/TargetMachine.h>
135 #include "tgsi/tgsi_scan.h"
136 #include "util/u_queue.h"
137 
138 #include "ac_binary.h"
139 #include "si_state.h"
140 
141 struct nir_shader;
142 
143 #define SI_MAX_VS_OUTPUTS	40
144 
145 /* Shader IO unique indices are supported for TGSI_SEMANTIC_GENERIC with an
146  * index smaller than this.
147  */
148 #define SI_MAX_IO_GENERIC       46
149 
150 /* SGPR user data indices */
151 enum {
152 	SI_SGPR_RW_BUFFERS,  /* rings (& stream-out, VS only) */
153 	SI_SGPR_RW_BUFFERS_HI,
154 	SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES,
155 	SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES_HI,
156 	SI_SGPR_CONST_AND_SHADER_BUFFERS, /* or just a constant buffer 0 pointer */
157 	SI_SGPR_CONST_AND_SHADER_BUFFERS_HI,
158 	SI_SGPR_SAMPLERS_AND_IMAGES,
159 	SI_SGPR_SAMPLERS_AND_IMAGES_HI,
160 	SI_NUM_RESOURCE_SGPRS,
161 
162 	/* all VS variants */
163 	SI_SGPR_VERTEX_BUFFERS	= SI_NUM_RESOURCE_SGPRS,
164 	SI_SGPR_VERTEX_BUFFERS_HI,
165 	SI_SGPR_BASE_VERTEX,
166 	SI_SGPR_START_INSTANCE,
167 	SI_SGPR_DRAWID,
168 	SI_SGPR_VS_STATE_BITS,
169 	SI_VS_NUM_USER_SGPR,
170 
171 	SI_SGPR_VS_BLIT_DATA = SI_SGPR_CONST_AND_SHADER_BUFFERS,
172 
173 	/* TES */
174 	SI_SGPR_TES_OFFCHIP_LAYOUT = SI_NUM_RESOURCE_SGPRS,
175 	SI_SGPR_TES_OFFCHIP_ADDR_BASE64K,
176 	SI_TES_NUM_USER_SGPR,
177 
178 	/* GFX6-8: TCS only */
179 	GFX6_SGPR_TCS_OFFCHIP_LAYOUT = SI_NUM_RESOURCE_SGPRS,
180 	GFX6_SGPR_TCS_OUT_OFFSETS,
181 	GFX6_SGPR_TCS_OUT_LAYOUT,
182 	GFX6_SGPR_TCS_IN_LAYOUT,
183 	GFX6_SGPR_TCS_OFFCHIP_ADDR_BASE64K,
184 	GFX6_SGPR_TCS_FACTOR_ADDR_BASE64K,
185 	GFX6_TCS_NUM_USER_SGPR,
186 
187 	/* GFX9: Merged LS-HS (VS-TCS) only. */
188 	GFX9_SGPR_TCS_OFFCHIP_LAYOUT = SI_VS_NUM_USER_SGPR,
189 	GFX9_SGPR_TCS_OUT_OFFSETS,
190 	GFX9_SGPR_TCS_OUT_LAYOUT,
191 	GFX9_SGPR_TCS_OFFCHIP_ADDR_BASE64K,
192 	GFX9_SGPR_TCS_FACTOR_ADDR_BASE64K,
193 	GFX9_SGPR_unused_to_align_the_next_pointer,
194 	GFX9_SGPR_TCS_CONST_AND_SHADER_BUFFERS,
195 	GFX9_SGPR_TCS_CONST_AND_SHADER_BUFFERS_HI,
196 	GFX9_SGPR_TCS_SAMPLERS_AND_IMAGES,
197 	GFX9_SGPR_TCS_SAMPLERS_AND_IMAGES_HI,
198 	GFX9_TCS_NUM_USER_SGPR,
199 
200 	/* GFX9: Merged ES-GS (VS-GS or TES-GS). */
201 	GFX9_SGPR_GS_CONST_AND_SHADER_BUFFERS = SI_VS_NUM_USER_SGPR,
202 	GFX9_SGPR_GS_CONST_AND_SHADER_BUFFERS_HI,
203 	GFX9_SGPR_GS_SAMPLERS_AND_IMAGES,
204 	GFX9_SGPR_GS_SAMPLERS_AND_IMAGES_HI,
205 	GFX9_GS_NUM_USER_SGPR,
206 
207 	/* GS limits */
208 	GFX6_GS_NUM_USER_SGPR = SI_NUM_RESOURCE_SGPRS,
209 	SI_GSCOPY_NUM_USER_SGPR = SI_SGPR_RW_BUFFERS_HI + 1,
210 
211 	/* PS only */
212 	SI_SGPR_ALPHA_REF	= SI_NUM_RESOURCE_SGPRS,
213 	SI_PS_NUM_USER_SGPR,
214 };
215 
216 /* LLVM function parameter indices */
217 enum {
218 	SI_NUM_RESOURCE_PARAMS = 4,
219 
220 	/* PS only parameters */
221 	SI_PARAM_ALPHA_REF = SI_NUM_RESOURCE_PARAMS,
222 	SI_PARAM_PRIM_MASK,
223 	SI_PARAM_PERSP_SAMPLE,
224 	SI_PARAM_PERSP_CENTER,
225 	SI_PARAM_PERSP_CENTROID,
226 	SI_PARAM_PERSP_PULL_MODEL,
227 	SI_PARAM_LINEAR_SAMPLE,
228 	SI_PARAM_LINEAR_CENTER,
229 	SI_PARAM_LINEAR_CENTROID,
230 	SI_PARAM_LINE_STIPPLE_TEX,
231 	SI_PARAM_POS_X_FLOAT,
232 	SI_PARAM_POS_Y_FLOAT,
233 	SI_PARAM_POS_Z_FLOAT,
234 	SI_PARAM_POS_W_FLOAT,
235 	SI_PARAM_FRONT_FACE,
236 	SI_PARAM_ANCILLARY,
237 	SI_PARAM_SAMPLE_COVERAGE,
238 	SI_PARAM_POS_FIXED_PT,
239 
240 	SI_NUM_PARAMS = SI_PARAM_POS_FIXED_PT + 9, /* +8 for COLOR[0..1] */
241 };
242 
243 /* Fields of driver-defined VS state SGPR. */
244 /* Clamp vertex color output (only used in VS as VS). */
245 #define S_VS_STATE_CLAMP_VERTEX_COLOR(x)	(((unsigned)(x) & 0x1) << 0)
246 #define C_VS_STATE_CLAMP_VERTEX_COLOR		0xFFFFFFFE
247 #define S_VS_STATE_INDEXED(x)			(((unsigned)(x) & 0x1) << 1)
248 #define C_VS_STATE_INDEXED			0xFFFFFFFD
249 #define S_VS_STATE_LS_OUT_PATCH_SIZE(x)		(((unsigned)(x) & 0x1FFF) << 8)
250 #define C_VS_STATE_LS_OUT_PATCH_SIZE		0xFFE000FF
251 #define S_VS_STATE_LS_OUT_VERTEX_SIZE(x)	(((unsigned)(x) & 0xFF) << 24)
252 #define C_VS_STATE_LS_OUT_VERTEX_SIZE		0x00FFFFFF
253 
254 /* SI-specific system values. */
255 enum {
256 	TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI = TGSI_SEMANTIC_COUNT,
257 	TGSI_SEMANTIC_DEFAULT_TESSINNER_SI,
258 };
259 
260 enum {
261 	/* Use a property enum that VS wouldn't use. */
262 	TGSI_PROPERTY_VS_BLIT_SGPRS = TGSI_PROPERTY_FS_COORD_ORIGIN,
263 
264 	/* These represent the number of SGPRs the shader uses. */
265 	SI_VS_BLIT_SGPRS_POS = 3,
266 	SI_VS_BLIT_SGPRS_POS_COLOR = 7,
267 	SI_VS_BLIT_SGPRS_POS_TEXCOORD = 9,
268 };
269 
270 /* For VS shader key fix_fetch. */
271 enum {
272 	SI_FIX_FETCH_NONE = 0,
273 	SI_FIX_FETCH_A2_SNORM,
274 	SI_FIX_FETCH_A2_SSCALED,
275 	SI_FIX_FETCH_A2_SINT,
276 	SI_FIX_FETCH_RGBA_32_UNORM,
277 	SI_FIX_FETCH_RGBX_32_UNORM,
278 	SI_FIX_FETCH_RGBA_32_SNORM,
279 	SI_FIX_FETCH_RGBX_32_SNORM,
280 	SI_FIX_FETCH_RGBA_32_USCALED,
281 	SI_FIX_FETCH_RGBA_32_SSCALED,
282 	SI_FIX_FETCH_RGBA_32_FIXED,
283 	SI_FIX_FETCH_RGBX_32_FIXED,
284 	SI_FIX_FETCH_RG_64_FLOAT,
285 	SI_FIX_FETCH_RGB_64_FLOAT,
286 	SI_FIX_FETCH_RGBA_64_FLOAT,
287 	SI_FIX_FETCH_RGB_8,	/* A = 1.0 */
288 	SI_FIX_FETCH_RGB_8_INT,	/* A = 1 */
289 	SI_FIX_FETCH_RGB_16,
290 	SI_FIX_FETCH_RGB_16_INT,
291 };
292 
293 struct si_shader;
294 
295 /* State of the context creating the shader object. */
296 struct si_compiler_ctx_state {
297 	/* Should only be used by si_init_shader_selector_async and
298 	 * si_build_shader_variant if thread_index == -1 (non-threaded). */
299 	LLVMTargetMachineRef		tm;
300 
301 	/* Used if thread_index == -1 or if debug.async is true. */
302 	struct pipe_debug_callback	debug;
303 
304 	/* Used for creating the log string for gallium/ddebug. */
305 	bool				is_debug_context;
306 };
307 
308 /* A shader selector is a gallium CSO and contains shader variants and
309  * binaries for one TGSI program. This can be shared by multiple contexts.
310  */
311 struct si_shader_selector {
312 	struct pipe_reference	reference;
313 	struct si_screen	*screen;
314 	struct util_queue_fence ready;
315 	struct si_compiler_ctx_state compiler_ctx_state;
316 
317 	mtx_t		mutex;
318 	struct si_shader	*first_variant; /* immutable after the first variant */
319 	struct si_shader	*last_variant; /* mutable */
320 
321 	/* The compiled TGSI shader expecting a prolog and/or epilog (not
322 	 * uploaded to a buffer).
323 	 */
324 	struct si_shader	*main_shader_part;
325 	struct si_shader	*main_shader_part_ls; /* as_ls is set in the key */
326 	struct si_shader	*main_shader_part_es; /* as_es is set in the key */
327 
328 	struct si_shader	*gs_copy_shader;
329 
330 	struct tgsi_token       *tokens;
331 	struct nir_shader       *nir;
332 	struct pipe_stream_output_info  so;
333 	struct tgsi_shader_info		info;
334 	struct tgsi_tessctrl_info	tcs_info;
335 
336 	/* PIPE_SHADER_[VERTEX|FRAGMENT|...] */
337 	unsigned	type;
338 	bool		vs_needs_prolog;
339 	bool		force_correct_derivs_after_kill;
340 	unsigned	pa_cl_vs_out_cntl;
341 	ubyte		clipdist_mask;
342 	ubyte		culldist_mask;
343 
344 	/* ES parameters. */
345 	unsigned	esgs_itemsize;
346 
347 	/* GS parameters. */
348 	unsigned	gs_input_verts_per_prim;
349 	unsigned	gs_output_prim;
350 	unsigned	gs_max_out_vertices;
351 	unsigned	gs_num_invocations;
352 	unsigned	max_gs_stream; /* count - 1 */
353 	unsigned	gsvs_vertex_size;
354 	unsigned	max_gsvs_emit_size;
355 	unsigned	enabled_streamout_buffer_mask;
356 
357 	/* PS parameters. */
358 	unsigned	color_attr_index[2];
359 	unsigned	db_shader_control;
360 	/* Set 0xf or 0x0 (4 bits) per each written output.
361 	 * ANDed with spi_shader_col_format.
362 	 */
363 	unsigned	colors_written_4bit;
364 
365 	/* CS parameters */
366 	unsigned local_size;
367 
368 	uint64_t	outputs_written;	/* "get_unique_index" bits */
369 	uint32_t	patch_outputs_written;	/* "get_unique_index_patch" bits */
370 
371 	uint64_t	inputs_read;		/* "get_unique_index" bits */
372 
373 	/* bitmasks of used descriptor slots */
374 	uint32_t	active_const_and_shader_buffers;
375 	uint64_t	active_samplers_and_images;
376 };
377 
378 /* Valid shader configurations:
379  *
380  * API shaders       VS | TCS | TES | GS |pass| PS
381  * are compiled as:     |     |     |    |thru|
382  *                      |     |     |    |    |
383  * Only VS & PS:     VS |     |     |    |    | PS
384  * GFX6 - with GS:   ES |     |     | GS | VS | PS
385  *      - with tess: LS | HS  | VS  |    |    | PS
386  *      - with both: LS | HS  | ES  | GS | VS | PS
387  * GFX9 - with GS:   -> |     |     | GS | VS | PS
388  *      - with tess: -> | HS  | VS  |    |    | PS
389  *      - with both: -> | HS  | ->  | GS | VS | PS
390  *
391  * -> = merged with the next stage
392  */
393 
394 /* Use the byte alignment for all following structure members for optimal
395  * shader key memory footprint.
396  */
397 #pragma pack(push, 1)
398 
399 /* Common VS bits between the shader key and the prolog key. */
400 struct si_vs_prolog_bits {
401 	/* - If neither "is_one" nor "is_fetched" has a bit set, the instance
402 	 *   divisor is 0.
403 	 * - If "is_one" has a bit set, the instance divisor is 1.
404 	 * - If "is_fetched" has a bit set, the instance divisor will be loaded
405 	 *   from the constant buffer.
406 	 */
407 	uint16_t	instance_divisor_is_one;     /* bitmask of inputs */
408 	uint16_t	instance_divisor_is_fetched; /* bitmask of inputs */
409 	unsigned	ls_vgpr_fix:1;
410 };
411 
412 /* Common TCS bits between the shader key and the epilog key. */
413 struct si_tcs_epilog_bits {
414 	unsigned	prim_mode:3;
415 	unsigned	invoc0_tess_factors_are_def:1;
416 	unsigned	tes_reads_tess_factors:1;
417 };
418 
419 struct si_gs_prolog_bits {
420 	unsigned	tri_strip_adj_fix:1;
421 };
422 
423 /* Common PS bits between the shader key and the prolog key. */
424 struct si_ps_prolog_bits {
425 	unsigned	color_two_side:1;
426 	unsigned	flatshade_colors:1;
427 	unsigned	poly_stipple:1;
428 	unsigned	force_persp_sample_interp:1;
429 	unsigned	force_linear_sample_interp:1;
430 	unsigned	force_persp_center_interp:1;
431 	unsigned	force_linear_center_interp:1;
432 	unsigned	bc_optimize_for_persp:1;
433 	unsigned	bc_optimize_for_linear:1;
434 	unsigned	samplemask_log_ps_iter:3;
435 };
436 
437 /* Common PS bits between the shader key and the epilog key. */
438 struct si_ps_epilog_bits {
439 	unsigned	spi_shader_col_format;
440 	unsigned	color_is_int8:8;
441 	unsigned	color_is_int10:8;
442 	unsigned	last_cbuf:3;
443 	unsigned	alpha_func:3;
444 	unsigned	alpha_to_one:1;
445 	unsigned	poly_line_smoothing:1;
446 	unsigned	clamp_color:1;
447 };
448 
449 union si_shader_part_key {
450 	struct {
451 		struct si_vs_prolog_bits states;
452 		unsigned	num_input_sgprs:6;
453 		/* For merged stages such as LS-HS, HS input VGPRs are first. */
454 		unsigned	num_merged_next_stage_vgprs:3;
455 		unsigned	last_input:4;
456 		unsigned	as_ls:1;
457 		unsigned	as_es:1;
458 		/* Prologs for monolithic shaders shouldn't set EXEC. */
459 		unsigned	is_monolithic:1;
460 	} vs_prolog;
461 	struct {
462 		struct si_tcs_epilog_bits states;
463 	} tcs_epilog;
464 	struct {
465 		struct si_gs_prolog_bits states;
466 		/* Prologs of monolithic shaders shouldn't set EXEC. */
467 		unsigned	is_monolithic:1;
468 	} gs_prolog;
469 	struct {
470 		struct si_ps_prolog_bits states;
471 		unsigned	num_input_sgprs:6;
472 		unsigned	num_input_vgprs:5;
473 		/* Color interpolation and two-side color selection. */
474 		unsigned	colors_read:8; /* color input components read */
475 		unsigned	num_interp_inputs:5; /* BCOLOR is at this location */
476 		unsigned	face_vgpr_index:5;
477 		unsigned	ancillary_vgpr_index:5;
478 		unsigned	wqm:1;
479 		char		color_attr_index[2];
480 		char		color_interp_vgpr_index[2]; /* -1 == constant */
481 	} ps_prolog;
482 	struct {
483 		struct si_ps_epilog_bits states;
484 		unsigned	colors_written:8;
485 		unsigned	writes_z:1;
486 		unsigned	writes_stencil:1;
487 		unsigned	writes_samplemask:1;
488 	} ps_epilog;
489 };
490 
491 struct si_shader_key {
492 	/* Prolog and epilog flags. */
493 	union {
494 		struct {
495 			struct si_vs_prolog_bits prolog;
496 		} vs;
497 		struct {
498 			struct si_vs_prolog_bits ls_prolog; /* for merged LS-HS */
499 			struct si_shader_selector *ls;   /* for merged LS-HS */
500 			struct si_tcs_epilog_bits epilog;
501 		} tcs; /* tessellation control shader */
502 		struct {
503 			struct si_vs_prolog_bits vs_prolog; /* for merged ES-GS */
504 			struct si_shader_selector *es;   /* for merged ES-GS */
505 			struct si_gs_prolog_bits prolog;
506 		} gs;
507 		struct {
508 			struct si_ps_prolog_bits prolog;
509 			struct si_ps_epilog_bits epilog;
510 		} ps;
511 	} part;
512 
513 	/* These two are initially set according to the NEXT_SHADER property,
514 	 * or guessed if the property doesn't seem correct.
515 	 */
516 	unsigned as_es:1; /* export shader, which precedes GS */
517 	unsigned as_ls:1; /* local shader, which precedes TCS */
518 
519 	/* Flags for monolithic compilation only. */
520 	struct {
521 		/* One byte for every input: SI_FIX_FETCH_* enums. */
522 		uint8_t		vs_fix_fetch[SI_MAX_ATTRIBS];
523 
524 		union {
525 			uint64_t	ff_tcs_inputs_to_copy; /* for fixed-func TCS */
526 			/* When PS needs PrimID and GS is disabled. */
527 			unsigned	vs_export_prim_id:1;
528 			struct {
529 				unsigned interpolate_at_sample_force_center:1;
530 			} ps;
531 		} u;
532 	} mono;
533 
534 	/* Optimization flags for asynchronous compilation only. */
535 	struct {
536 		/* For HW VS (it can be VS, TES, GS) */
537 		uint64_t	kill_outputs; /* "get_unique_index" bits */
538 		unsigned	clip_disable:1;
539 
540 		/* For shaders where monolithic variants have better code.
541 		 *
542 		 * This is a flag that has no effect on code generation,
543 		 * but forces monolithic shaders to be used as soon as
544 		 * possible, because it's in the "opt" group.
545 		 */
546 		unsigned	prefer_mono:1;
547 	} opt;
548 };
549 
550 /* Restore the pack alignment to default. */
551 #pragma pack(pop)
552 
553 struct si_shader_config {
554 	unsigned			num_sgprs;
555 	unsigned			num_vgprs;
556 	unsigned			spilled_sgprs;
557 	unsigned			spilled_vgprs;
558 	unsigned			private_mem_vgprs;
559 	unsigned			lds_size;
560 	unsigned			spi_ps_input_ena;
561 	unsigned			spi_ps_input_addr;
562 	unsigned			float_mode;
563 	unsigned			scratch_bytes_per_wave;
564 	unsigned			rsrc1;
565 	unsigned			rsrc2;
566 };
567 
568 /* GCN-specific shader info. */
569 struct si_shader_info {
570 	ubyte			vs_output_param_offset[SI_MAX_VS_OUTPUTS];
571 	ubyte			num_input_sgprs;
572 	ubyte			num_input_vgprs;
573 	signed char		face_vgpr_index;
574 	signed char		ancillary_vgpr_index;
575 	bool			uses_instanceid;
576 	ubyte			nr_pos_exports;
577 	ubyte			nr_param_exports;
578 };
579 
580 struct si_shader {
581 	struct si_compiler_ctx_state	compiler_ctx_state;
582 
583 	struct si_shader_selector	*selector;
584 	struct si_shader_selector	*previous_stage_sel; /* for refcounting */
585 	struct si_shader		*next_variant;
586 
587 	struct si_shader_part		*prolog;
588 	struct si_shader		*previous_stage; /* for GFX9 */
589 	struct si_shader_part		*prolog2;
590 	struct si_shader_part		*epilog;
591 
592 	struct si_pm4_state		*pm4;
593 	struct r600_resource		*bo;
594 	struct r600_resource		*scratch_bo;
595 	struct si_shader_key		key;
596 	struct util_queue_fence		ready;
597 	bool				compilation_failed;
598 	bool				is_monolithic;
599 	bool				is_optimized;
600 	bool				is_binary_shared;
601 	bool				is_gs_copy_shader;
602 
603 	/* The following data is all that's needed for binary shaders. */
604 	struct ac_shader_binary	binary;
605 	struct si_shader_config		config;
606 	struct si_shader_info		info;
607 
608 	/* Shader key + LLVM IR + disassembly + statistics.
609 	 * Generated for debug contexts only.
610 	 */
611 	char				*shader_log;
612 	size_t				shader_log_size;
613 };
614 
615 struct si_shader_part {
616 	struct si_shader_part *next;
617 	union si_shader_part_key key;
618 	struct ac_shader_binary binary;
619 	struct si_shader_config config;
620 };
621 
622 /* si_shader.c */
623 struct si_shader *
624 si_generate_gs_copy_shader(struct si_screen *sscreen,
625 			   LLVMTargetMachineRef tm,
626 			   struct si_shader_selector *gs_selector,
627 			   struct pipe_debug_callback *debug);
628 int si_compile_tgsi_shader(struct si_screen *sscreen,
629 			   LLVMTargetMachineRef tm,
630 			   struct si_shader *shader,
631 			   bool is_monolithic,
632 			   struct pipe_debug_callback *debug);
633 int si_shader_create(struct si_screen *sscreen, LLVMTargetMachineRef tm,
634 		     struct si_shader *shader,
635 		     struct pipe_debug_callback *debug);
636 void si_shader_destroy(struct si_shader *shader);
637 unsigned si_shader_io_get_unique_index_patch(unsigned semantic_name, unsigned index);
638 unsigned si_shader_io_get_unique_index(unsigned semantic_name, unsigned index);
639 int si_shader_binary_upload(struct si_screen *sscreen, struct si_shader *shader);
640 void si_shader_dump(struct si_screen *sscreen, const struct si_shader *shader,
641 		    struct pipe_debug_callback *debug, unsigned processor,
642 		    FILE *f, bool check_debug_option);
643 void si_multiwave_lds_size_workaround(struct si_screen *sscreen,
644 				      unsigned *lds_size);
645 void si_shader_apply_scratch_relocs(struct si_shader *shader,
646 				    uint64_t scratch_va);
647 void si_shader_binary_read_config(struct ac_shader_binary *binary,
648 				  struct si_shader_config *conf,
649 				  unsigned symbol_offset);
650 const char *si_get_shader_name(const struct si_shader *shader, unsigned processor);
651 
652 /* si_shader_nir.c */
653 void si_nir_scan_shader(const struct nir_shader *nir,
654 			struct tgsi_shader_info *info);
655 void si_nir_scan_tess_ctrl(const struct nir_shader *nir,
656 			   const struct tgsi_shader_info *info,
657 			   struct tgsi_tessctrl_info *out);
658 void si_lower_nir(struct si_shader_selector *sel);
659 
660 /* Inline helpers. */
661 
662 /* Return the pointer to the main shader part's pointer. */
663 static inline struct si_shader **
si_get_main_shader_part(struct si_shader_selector * sel,struct si_shader_key * key)664 si_get_main_shader_part(struct si_shader_selector *sel,
665 			struct si_shader_key *key)
666 {
667 	if (key->as_ls)
668 		return &sel->main_shader_part_ls;
669 	if (key->as_es)
670 		return &sel->main_shader_part_es;
671 	return &sel->main_shader_part;
672 }
673 
674 static inline bool
si_shader_uses_bindless_samplers(struct si_shader_selector * selector)675 si_shader_uses_bindless_samplers(struct si_shader_selector *selector)
676 {
677 	return selector ? selector->info.uses_bindless_samplers : false;
678 }
679 
680 static inline bool
si_shader_uses_bindless_images(struct si_shader_selector * selector)681 si_shader_uses_bindless_images(struct si_shader_selector *selector)
682 {
683 	return selector ? selector->info.uses_bindless_images : false;
684 }
685 
686 void si_destroy_shader_selector(struct si_context *sctx,
687 			        struct si_shader_selector *sel);
688 
689 static inline void
si_shader_selector_reference(struct si_context * sctx,struct si_shader_selector ** dst,struct si_shader_selector * src)690 si_shader_selector_reference(struct si_context *sctx,
691 			     struct si_shader_selector **dst,
692 			     struct si_shader_selector *src)
693 {
694 	if (pipe_reference(&(*dst)->reference, &src->reference))
695 		si_destroy_shader_selector(sctx, *dst);
696 
697 	*dst = src;
698 }
699 
700 #endif
701