1/*
2 * Copyright 2012 Advanced Micro Devices, Inc.
3 * All Rights Reserved.
4 *
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:
11 *
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
14 * Software.
15 *
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.
23 */
24
25/* The compiler middle-end architecture: Explaining (non-)monolithic shaders
26 * -------------------------------------------------------------------------
27 *
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
30 * the driver.
31 *
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
39 *
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)
46 *
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.
55 *
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".
62 *
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
68 * special cases:
69 *
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.
73 *
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
81 *    yet.
82 *
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
90 */
91
92/* How linking shader inputs and outputs between vertex, tessellation, and
93 * geometry shaders works.
94 *
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.
99 *
100 * Since GS and tessellation are only possible in the OpenGL core profile,
101 * only these semantics are valid for per-vertex data:
102 *
103 *   Name             Location
104 *
105 *   POSITION         0
106 *   PSIZE            1
107 *   CLIPDIST0..1     2..3
108 *   CULLDIST0..1     (not implemented)
109 *   GENERIC0..31     4..35
110 *
111 * For example, a shader only writing GENERIC0 has the output stride of 5.
112 *
113 * Only these semantics are valid for per-patch data:
114 *
115 *   Name             Location
116 *
117 *   TESSOUTER        0
118 *   TESSINNER        1
119 *   PATCH0..29       2..31
120 *
121 * That's how independent shaders agree on input and output locations.
122 * The si_shader_io_get_unique_index function assigns the locations.
123 *
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.
129 */
130
131#ifndef SI_SHADER_H
132#define SI_SHADER_H
133
134#include "ac_binary.h"
135#include "ac_llvm_build.h"
136#include "ac_llvm_util.h"
137#include "util/simple_mtx.h"
138#include "util/u_inlines.h"
139#include "util/u_live_shader_cache.h"
140#include "util/u_queue.h"
141#include "si_pm4.h"
142
143#include <stdio.h>
144
145#ifdef __cplusplus
146extern "C" {
147#endif
148
149// Use LDS symbols when supported by LLVM. Can be disabled for testing the old
150// path on newer LLVM for now. Should be removed in the long term.
151#define USE_LDS_SYMBOLS (true)
152
153struct nir_shader;
154struct si_shader;
155struct si_context;
156
157#define SI_MAX_ATTRIBS    16
158#define SI_MAX_VS_OUTPUTS 40
159
160#define SI_NGG_PRIM_EDGE_FLAG_BITS ((1 << 9) | (1 << 19) | (1 << 29))
161
162#define SI_PS_INPUT_CNTL_0000          (S_028644_OFFSET(0x20) | S_028644_DEFAULT_VAL(0))
163#define SI_PS_INPUT_CNTL_0001          (S_028644_OFFSET(0x20) | S_028644_DEFAULT_VAL(3))
164#define SI_PS_INPUT_CNTL_UNUSED        SI_PS_INPUT_CNTL_0000
165/* D3D9 behaviour for COLOR0 requires 0001. GL is undefined. */
166#define SI_PS_INPUT_CNTL_UNUSED_COLOR0 SI_PS_INPUT_CNTL_0001
167
168/* SGPR user data indices */
169enum
170{
171   SI_SGPR_INTERNAL_BINDINGS,
172   SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES,
173   SI_SGPR_CONST_AND_SHADER_BUFFERS, /* or just a constant buffer 0 pointer */
174   SI_SGPR_SAMPLERS_AND_IMAGES,
175   SI_NUM_RESOURCE_SGPRS,
176
177   /* API VS, TES without GS, GS copy shader */
178   SI_SGPR_VS_STATE_BITS = SI_NUM_RESOURCE_SGPRS,
179   SI_NUM_VS_STATE_RESOURCE_SGPRS,
180
181   /* all VS variants */
182   SI_SGPR_BASE_VERTEX = SI_NUM_VS_STATE_RESOURCE_SGPRS,
183   SI_SGPR_DRAWID,
184   SI_SGPR_START_INSTANCE,
185   SI_VS_NUM_USER_SGPR,
186
187   SI_SGPR_VS_BLIT_DATA = SI_SGPR_CONST_AND_SHADER_BUFFERS,
188
189   /* TES */
190   SI_SGPR_TES_OFFCHIP_LAYOUT = SI_NUM_VS_STATE_RESOURCE_SGPRS,
191   SI_SGPR_TES_OFFCHIP_ADDR,
192   SI_TES_NUM_USER_SGPR,
193
194   /* GFX6-8: TCS only */
195   GFX6_SGPR_TCS_OFFCHIP_LAYOUT = SI_NUM_RESOURCE_SGPRS,
196   GFX6_SGPR_TCS_OUT_OFFSETS,
197   GFX6_SGPR_TCS_OUT_LAYOUT,
198   GFX6_SGPR_TCS_IN_LAYOUT,
199   GFX6_TCS_NUM_USER_SGPR,
200
201   /* GFX9: Merged shaders. */
202   /* 2ND_CONST_AND_SHADER_BUFFERS is set in USER_DATA_ADDR_LO (SGPR0). */
203   /* 2ND_SAMPLERS_AND_IMAGES is set in USER_DATA_ADDR_HI (SGPR1). */
204   GFX9_MERGED_NUM_USER_SGPR = SI_VS_NUM_USER_SGPR,
205
206   /* GFX9: Merged LS-HS (VS-TCS) only. */
207   GFX9_SGPR_TCS_OFFCHIP_LAYOUT = GFX9_MERGED_NUM_USER_SGPR,
208   GFX9_SGPR_TCS_OUT_OFFSETS,
209   GFX9_SGPR_TCS_OUT_LAYOUT,
210   GFX9_TCS_NUM_USER_SGPR,
211
212   /* GS limits */
213   GFX6_GS_NUM_USER_SGPR = SI_NUM_RESOURCE_SGPRS,
214   GFX9_VSGS_NUM_USER_SGPR = SI_VS_NUM_USER_SGPR,
215   GFX9_TESGS_NUM_USER_SGPR = SI_TES_NUM_USER_SGPR,
216   SI_GSCOPY_NUM_USER_SGPR = SI_NUM_VS_STATE_RESOURCE_SGPRS,
217
218   /* PS only */
219   SI_SGPR_ALPHA_REF = SI_NUM_RESOURCE_SGPRS,
220   SI_PS_NUM_USER_SGPR,
221
222   /* The value has to be 12, because the hw requires that descriptors
223    * are aligned to 4 SGPRs.
224    */
225   SI_SGPR_VS_VB_DESCRIPTOR_FIRST = 12,
226};
227
228/* LLVM function parameter indices */
229enum
230{
231   SI_NUM_RESOURCE_PARAMS = 4,
232
233   /* PS only parameters */
234   SI_PARAM_ALPHA_REF = SI_NUM_RESOURCE_PARAMS,
235   SI_PARAM_PRIM_MASK,
236   SI_PARAM_PERSP_SAMPLE,
237   SI_PARAM_PERSP_CENTER,
238   SI_PARAM_PERSP_CENTROID,
239   SI_PARAM_PERSP_PULL_MODEL,
240   SI_PARAM_LINEAR_SAMPLE,
241   SI_PARAM_LINEAR_CENTER,
242   SI_PARAM_LINEAR_CENTROID,
243   SI_PARAM_LINE_STIPPLE_TEX,
244   SI_PARAM_POS_X_FLOAT,
245   SI_PARAM_POS_Y_FLOAT,
246   SI_PARAM_POS_Z_FLOAT,
247   SI_PARAM_POS_W_FLOAT,
248   SI_PARAM_FRONT_FACE,
249   SI_PARAM_ANCILLARY,
250   SI_PARAM_SAMPLE_COVERAGE,
251   SI_PARAM_POS_FIXED_PT,
252
253   SI_NUM_PARAMS = SI_PARAM_POS_FIXED_PT + 9, /* +8 for COLOR[0..1] */
254};
255
256/* Fields of driver-defined VS state SGPR. */
257#define S_VS_STATE_CLAMP_VERTEX_COLOR(x)      (((unsigned)(x)&0x1) << 0)
258#define C_VS_STATE_CLAMP_VERTEX_COLOR         0xFFFFFFFE
259#define S_VS_STATE_INDEXED(x)                 (((unsigned)(x)&0x1) << 1)
260#define C_VS_STATE_INDEXED                    0xFFFFFFFD
261#define S_VS_STATE_OUTPRIM(x)                 (((unsigned)(x)&0x3) << 2)
262#define C_VS_STATE_OUTPRIM                    0xFFFFFFF3
263#define S_VS_STATE_PROVOKING_VTX_INDEX(x)     (((unsigned)(x)&0x3) << 4)
264#define C_VS_STATE_PROVOKING_VTX_INDEX        0xFFFFFFCF
265#define S_VS_STATE_STREAMOUT_QUERY_ENABLED(x) (((unsigned)(x)&0x1) << 6)
266#define C_VS_STATE_STREAMOUT_QUERY_ENABLED    0xFFFFFFBF
267#define S_VS_STATE_SMALL_PRIM_PRECISION(x)    (((unsigned)(x)&0xF) << 7)
268#define C_VS_STATE_SMALL_PRIM_PRECISION       0xFFFFF87F
269#define S_VS_STATE_LS_OUT_PATCH_SIZE(x)       (((unsigned)(x)&0x1FFF) << 11)
270#define C_VS_STATE_LS_OUT_PATCH_SIZE          0xFF0007FF
271#define S_VS_STATE_LS_OUT_VERTEX_SIZE(x)      (((unsigned)(x)&0xFF) << 24)
272#define C_VS_STATE_LS_OUT_VERTEX_SIZE         0x00FFFFFF
273
274enum
275{
276   /* These represent the number of SGPRs the shader uses. */
277   SI_VS_BLIT_SGPRS_POS = 3,
278   SI_VS_BLIT_SGPRS_POS_COLOR = 7,
279   SI_VS_BLIT_SGPRS_POS_TEXCOORD = 9,
280};
281
282#define SI_NGG_CULL_ENABLED                  (1 << 0)   /* this implies W, view.xy, and small prim culling */
283#define SI_NGG_CULL_BACK_FACE                (1 << 1)   /* back faces */
284#define SI_NGG_CULL_FRONT_FACE               (1 << 2)   /* front faces */
285#define SI_NGG_CULL_LINES                    (1 << 3)   /* the primitive type is lines */
286
287/**
288 * For VS shader keys, describe any fixups required for vertex fetch.
289 *
290 * \ref log_size, \ref format, and the number of channels are interpreted as
291 * by \ref ac_build_opencoded_load_format.
292 *
293 * Note: all bits 0 (size = 1 byte, num channels = 1, format = float) is an
294 * impossible format and indicates that no fixup is needed (just use
295 * buffer_load_format_xyzw).
296 */
297union si_vs_fix_fetch {
298   struct {
299      uint8_t log_size : 2;        /* 1, 2, 4, 8 or bytes per channel */
300      uint8_t num_channels_m1 : 2; /* number of channels minus 1 */
301      uint8_t format : 3;          /* AC_FETCH_FORMAT_xxx */
302      uint8_t reverse : 1;         /* reverse XYZ channels */
303   } u;
304   uint8_t bits;
305};
306
307struct si_shader;
308
309/* State of the context creating the shader object. */
310struct si_compiler_ctx_state {
311   /* Should only be used by si_init_shader_selector_async and
312    * si_build_shader_variant if thread_index == -1 (non-threaded). */
313   struct ac_llvm_compiler *compiler;
314
315   /* Used if thread_index == -1 or if debug.async is true. */
316   struct pipe_debug_callback debug;
317
318   /* Used for creating the log string for gallium/ddebug. */
319   bool is_debug_context;
320};
321
322enum si_color_output_type {
323   SI_TYPE_ANY32,
324   SI_TYPE_FLOAT16,
325   SI_TYPE_INT16,
326   SI_TYPE_UINT16,
327};
328
329union si_input_info {
330   struct {
331      ubyte semantic;
332      ubyte interpolate;
333      ubyte fp16_lo_hi_valid;
334      ubyte usage_mask;
335   };
336   uint32_t _unused; /* this just forces 4-byte alignment */
337};
338
339struct si_shader_info {
340   shader_info base;
341
342   gl_shader_stage stage;
343
344   ubyte num_inputs;
345   ubyte num_outputs;
346   union si_input_info input[PIPE_MAX_SHADER_INPUTS];
347   ubyte output_semantic[PIPE_MAX_SHADER_OUTPUTS];
348   ubyte output_usagemask[PIPE_MAX_SHADER_OUTPUTS];
349   ubyte output_readmask[PIPE_MAX_SHADER_OUTPUTS];
350   ubyte output_streams[PIPE_MAX_SHADER_OUTPUTS];
351   ubyte output_type[PIPE_MAX_SHADER_OUTPUTS]; /* enum nir_alu_type */
352
353   ubyte color_interpolate[2];
354   ubyte color_interpolate_loc[2];
355
356   int constbuf0_num_slots;
357   ubyte num_stream_output_components[4];
358
359   uint num_memory_stores;
360
361   ubyte colors_read; /**< which color components are read by the FS */
362   ubyte colors_written;
363   uint16_t output_color_types; /**< Each bit pair is enum si_color_output_type */
364   bool color0_writes_all_cbufs; /**< gl_FragColor */
365   bool reads_samplemask;   /**< does fragment shader read sample mask? */
366   bool reads_tess_factors; /**< If TES reads TESSINNER or TESSOUTER */
367   bool writes_z;           /**< does fragment shader write Z value? */
368   bool writes_stencil;     /**< does fragment shader write stencil value? */
369   bool writes_samplemask;  /**< does fragment shader write sample mask? */
370   bool writes_edgeflag;    /**< vertex shader outputs edgeflag */
371   bool uses_interp_color;
372   bool uses_persp_center_color;
373   bool uses_persp_centroid_color;
374   bool uses_persp_sample_color;
375   bool uses_persp_center;
376   bool uses_persp_centroid;
377   bool uses_persp_sample;
378   bool uses_linear_center;
379   bool uses_linear_centroid;
380   bool uses_linear_sample;
381   bool uses_interp_at_sample;
382   bool uses_instanceid;
383   bool uses_base_vertex;
384   bool uses_base_instance;
385   bool uses_drawid;
386   bool uses_primid;
387   bool uses_frontface;
388   bool uses_invocationid;
389   bool uses_thread_id[3];
390   bool uses_block_id[3];
391   bool uses_variable_block_size;
392   bool uses_grid_size;
393   bool uses_subgroup_info;
394   bool writes_position;
395   bool writes_psize;
396   bool writes_clipvertex;
397   bool writes_primid;
398   bool writes_viewport_index;
399   bool writes_layer;
400   bool uses_bindless_samplers;
401   bool uses_bindless_images;
402   bool uses_indirect_descriptor;
403
404   bool uses_vmem_return_type_sampler_or_bvh;
405   bool uses_vmem_return_type_other; /* all other VMEM loads and atomics with return */
406
407   /** Whether all codepaths write tess factors in all invocations. */
408   bool tessfactors_are_def_in_all_invocs;
409
410   /* A flag to check if vrs2x2 can be enabled to reduce number of
411    * fragment shader invocations if flat shading.
412    */
413   bool allow_flat_shading;
414
415   /* Optimization: if the texture bound to this texunit has been cleared to 1,
416    * then the draw can be skipped (see si_draw_vbo_skip_noop). Initially the
417    * value is 0xff (undetermined) and can be later changed to 0 (= false) or
418    * texunit + 1.
419    */
420   uint8_t writes_1_if_tex_is_1;
421};
422
423/* A shader selector is a gallium CSO and contains shader variants and
424 * binaries for one NIR program. This can be shared by multiple contexts.
425 */
426struct si_shader_selector {
427   struct util_live_shader base;
428   struct si_screen *screen;
429   struct util_queue_fence ready;
430   struct si_compiler_ctx_state compiler_ctx_state;
431
432   simple_mtx_t mutex;
433   struct si_shader *first_variant; /* immutable after the first variant */
434   struct si_shader *last_variant;  /* mutable */
435
436   /* The compiled NIR shader without a prolog and/or epilog (not
437    * uploaded to a buffer object).
438    */
439   struct si_shader *main_shader_part;
440   struct si_shader *main_shader_part_ls;     /* as_ls is set in the key */
441   struct si_shader *main_shader_part_es;     /* as_es is set in the key */
442   struct si_shader *main_shader_part_ngg;    /* as_ngg is set in the key */
443   struct si_shader *main_shader_part_ngg_es; /* for Wave32 TES before legacy GS */
444
445   struct si_shader *gs_copy_shader;
446
447   struct nir_shader *nir;
448   void *nir_binary;
449   unsigned nir_size;
450
451   struct pipe_stream_output_info so;
452   struct si_shader_info info;
453
454   enum pipe_shader_type pipe_shader_type;
455   ubyte const_and_shader_buf_descriptors_index;
456   ubyte sampler_and_images_descriptors_index;
457   bool vs_needs_prolog;
458   ubyte cs_shaderbufs_sgpr_index;
459   ubyte cs_num_shaderbufs_in_user_sgprs;
460   ubyte cs_images_sgpr_index;
461   ubyte cs_images_num_sgprs;
462   ubyte cs_num_images_in_user_sgprs;
463   ubyte num_vs_inputs;
464   ubyte num_vbos_in_user_sgprs;
465   unsigned ngg_cull_vert_threshold; /* UINT32_MAX = disabled */
466   ubyte clipdist_mask;
467   ubyte culldist_mask;
468   enum pipe_prim_type rast_prim;
469
470   /* ES parameters. */
471   uint16_t esgs_itemsize; /* vertex stride */
472   uint16_t lshs_vertex_stride;
473
474   /* GS parameters. */
475   uint16_t gsvs_vertex_size;
476   ubyte gs_input_verts_per_prim;
477   unsigned max_gsvs_emit_size;
478   uint16_t enabled_streamout_buffer_mask;
479   bool tess_turns_off_ngg;
480
481   /* PS parameters. */
482   ubyte color_attr_index[2];
483   unsigned db_shader_control;
484   /* Set 0xf or 0x0 (4 bits) per each written output.
485    * ANDed with spi_shader_col_format.
486    */
487   unsigned colors_written_4bit;
488
489   uint64_t outputs_written_before_ps; /* "get_unique_index" bits */
490   uint64_t outputs_written;           /* "get_unique_index" bits */
491   uint32_t patch_outputs_written;     /* "get_unique_index_patch" bits */
492
493   uint64_t inputs_read; /* "get_unique_index" bits */
494   uint64_t tcs_vgpr_only_inputs; /* TCS inputs that are only in VGPRs, not LDS. */
495
496   /* bitmasks of used descriptor slots */
497   uint64_t active_const_and_shader_buffers;
498   uint64_t active_samplers_and_images;
499};
500
501/* Valid shader configurations:
502 *
503 * API shaders           VS | TCS | TES | GS |pass| PS
504 * are compiled as:         |     |     |    |thru|
505 *                          |     |     |    |    |
506 * Only VS & PS:         VS |     |     |    |    | PS
507 * GFX6     - with GS:   ES |     |     | GS | VS | PS
508 *          - with tess: LS | HS  | VS  |    |    | PS
509 *          - with both: LS | HS  | ES  | GS | VS | PS
510 * GFX9     - with GS:   -> |     |     | GS | VS | PS
511 *          - with tess: -> | HS  | VS  |    |    | PS
512 *          - with both: -> | HS  | ->  | GS | VS | PS
513 *                          |     |     |    |    |
514 * NGG      - VS & PS:   GS |     |     |    |    | PS
515 * (GFX10+) - with GS:   -> |     |     | GS |    | PS
516 *          - with tess: -> | HS  | GS  |    |    | PS
517 *          - with both: -> | HS  | ->  | GS |    | PS
518 *
519 * -> = merged with the next stage
520 */
521
522/* Use the byte alignment for all following structure members for optimal
523 * shader key memory footprint.
524 */
525#pragma pack(push, 1)
526
527/* Common VS bits between the shader key and the prolog key. */
528struct si_vs_prolog_bits {
529   /* - If neither "is_one" nor "is_fetched" has a bit set, the instance
530    *   divisor is 0.
531    * - If "is_one" has a bit set, the instance divisor is 1.
532    * - If "is_fetched" has a bit set, the instance divisor will be loaded
533    *   from the constant buffer.
534    */
535   uint16_t instance_divisor_is_one;     /* bitmask of inputs */
536   uint16_t instance_divisor_is_fetched; /* bitmask of inputs */
537   unsigned ls_vgpr_fix : 1;
538};
539
540/* Common TCS bits between the shader key and the epilog key. */
541struct si_tcs_epilog_bits {
542   unsigned prim_mode : 3;
543   unsigned invoc0_tess_factors_are_def : 1;
544   unsigned tes_reads_tess_factors : 1;
545};
546
547struct si_gs_prolog_bits {
548   unsigned tri_strip_adj_fix : 1;
549};
550
551/* Common PS bits between the shader key and the prolog key. */
552struct si_ps_prolog_bits {
553   unsigned color_two_side : 1;
554   unsigned flatshade_colors : 1;
555   unsigned poly_stipple : 1;
556   unsigned force_persp_sample_interp : 1;
557   unsigned force_linear_sample_interp : 1;
558   unsigned force_persp_center_interp : 1;
559   unsigned force_linear_center_interp : 1;
560   unsigned bc_optimize_for_persp : 1;
561   unsigned bc_optimize_for_linear : 1;
562   unsigned samplemask_log_ps_iter : 3;
563};
564
565/* Common PS bits between the shader key and the epilog key. */
566struct si_ps_epilog_bits {
567   unsigned spi_shader_col_format;
568   unsigned color_is_int8 : 8;
569   unsigned color_is_int10 : 8;
570   unsigned last_cbuf : 3;
571   unsigned alpha_func : 3;
572   unsigned alpha_to_one : 1;
573   unsigned poly_line_smoothing : 1;
574   unsigned clamp_color : 1;
575};
576
577union si_shader_part_key {
578   struct {
579      struct si_vs_prolog_bits states;
580      unsigned num_input_sgprs : 6;
581      /* For merged stages such as LS-HS, HS input VGPRs are first. */
582      unsigned num_merged_next_stage_vgprs : 3;
583      unsigned num_inputs : 5;
584      unsigned as_ls : 1;
585      unsigned as_es : 1;
586      unsigned as_ngg : 1;
587      unsigned load_vgprs_after_culling : 1;
588      /* Prologs for monolithic shaders shouldn't set EXEC. */
589      unsigned is_monolithic : 1;
590   } vs_prolog;
591   struct {
592      struct si_tcs_epilog_bits states;
593   } tcs_epilog;
594   struct {
595      struct si_gs_prolog_bits states;
596      unsigned as_ngg : 1;
597   } gs_prolog;
598   struct {
599      struct si_ps_prolog_bits states;
600      unsigned num_input_sgprs : 6;
601      unsigned num_input_vgprs : 5;
602      /* Color interpolation and two-side color selection. */
603      unsigned colors_read : 8;       /* color input components read */
604      unsigned num_interp_inputs : 5; /* BCOLOR is at this location */
605      unsigned face_vgpr_index : 5;
606      unsigned ancillary_vgpr_index : 5;
607      unsigned wqm : 1;
608      char color_attr_index[2];
609      signed char color_interp_vgpr_index[2]; /* -1 == constant */
610   } ps_prolog;
611   struct {
612      struct si_ps_epilog_bits states;
613      unsigned colors_written : 8;
614      unsigned color_types : 16;
615      unsigned writes_z : 1;
616      unsigned writes_stencil : 1;
617      unsigned writes_samplemask : 1;
618   } ps_epilog;
619};
620
621struct si_shader_key {
622   /* Prolog and epilog flags. */
623   union {
624      struct {
625         struct si_vs_prolog_bits prolog;
626      } vs;
627      struct {
628         struct si_vs_prolog_bits ls_prolog; /* for merged LS-HS */
629         struct si_shader_selector *ls;      /* for merged LS-HS */
630         struct si_tcs_epilog_bits epilog;
631      } tcs; /* tessellation control shader */
632      struct {
633         struct si_vs_prolog_bits vs_prolog; /* for merged ES-GS */
634         struct si_shader_selector *es;      /* for merged ES-GS */
635         struct si_gs_prolog_bits prolog;
636      } gs;
637      struct {
638         struct si_ps_prolog_bits prolog;
639         struct si_ps_epilog_bits epilog;
640      } ps;
641   } part;
642
643   /* These three are initially set according to the NEXT_SHADER property,
644    * or guessed if the property doesn't seem correct.
645    */
646   unsigned as_es : 1;  /* whether it's a shader before GS */
647   unsigned as_ls : 1;  /* whether it's VS before TCS */
648   unsigned as_ngg : 1; /* whether it's the last GE stage and NGG is enabled,
649                           also set for the stage right before GS */
650
651   /* Flags for monolithic compilation only. */
652   struct {
653      /* Whether fetch should be opencoded according to vs_fix_fetch.
654       * Otherwise, if vs_fix_fetch is non-zero, buffer_load_format_xyzw
655       * with minimal fixups is used. */
656      uint16_t vs_fetch_opencode;
657      union si_vs_fix_fetch vs_fix_fetch[SI_MAX_ATTRIBS];
658
659      union {
660         uint64_t ff_tcs_inputs_to_copy; /* for fixed-func TCS */
661         /* When PS needs PrimID and GS is disabled. */
662         unsigned vs_export_prim_id : 1;
663         struct {
664            unsigned interpolate_at_sample_force_center : 1;
665            unsigned fbfetch_msaa : 1;
666            unsigned fbfetch_is_1D : 1;
667            unsigned fbfetch_layered : 1;
668         } ps;
669      } u;
670   } mono;
671
672   /* Optimization flags for asynchronous compilation only. */
673   struct {
674      /* For HW VS (it can be VS, TES, GS) */
675      uint64_t kill_outputs; /* "get_unique_index" bits */
676      unsigned kill_clip_distances : 8;
677      unsigned kill_pointsize : 1;
678
679      /* For NGG VS and TES. */
680      unsigned ngg_culling : 4; /* SI_NGG_CULL_* */
681
682      /* For shaders where monolithic variants have better code.
683       *
684       * This is a flag that has no effect on code generation,
685       * but forces monolithic shaders to be used as soon as
686       * possible, because it's in the "opt" group.
687       */
688      unsigned prefer_mono : 1;
689
690      /* VS and TCS have the same number of patch vertices. */
691      unsigned same_patch_vertices:1;
692
693      unsigned inline_uniforms:1;
694
695      /* This must be kept last to limit the number of variants
696       * depending only on the uniform values.
697       */
698      uint32_t inlined_uniform_values[MAX_INLINABLE_UNIFORMS];
699   } opt;
700};
701
702/* Restore the pack alignment to default. */
703#pragma pack(pop)
704
705/* GCN-specific shader info. */
706struct si_shader_binary_info {
707   ubyte vs_output_param_offset[SI_MAX_VS_OUTPUTS];
708   uint32_t vs_output_ps_input_cntl[NUM_TOTAL_VARYING_SLOTS];
709   ubyte num_input_sgprs;
710   ubyte num_input_vgprs;
711   signed char face_vgpr_index;
712   signed char ancillary_vgpr_index;
713   bool uses_instanceid;
714   ubyte nr_pos_exports;
715   ubyte nr_param_exports;
716   unsigned private_mem_vgprs;
717   unsigned max_simd_waves;
718};
719
720struct si_shader_binary {
721   const char *elf_buffer;
722   size_t elf_size;
723
724   char *uploaded_code;
725   size_t uploaded_code_size;
726
727   char *llvm_ir_string;
728};
729
730struct gfx9_gs_info {
731   unsigned es_verts_per_subgroup;
732   unsigned gs_prims_per_subgroup;
733   unsigned gs_inst_prims_in_subgroup;
734   unsigned max_prims_per_subgroup;
735   unsigned esgs_ring_size; /* in bytes */
736};
737
738#define SI_NUM_VGT_STAGES_KEY_BITS 5
739#define SI_NUM_VGT_STAGES_STATES   (1 << SI_NUM_VGT_STAGES_KEY_BITS)
740
741/* The VGT_SHADER_STAGES key used to index the table of precomputed values.
742 * Some fields are set by state-change calls, most are set by draw_vbo.
743 */
744union si_vgt_stages_key {
745   struct {
746#if UTIL_ARCH_LITTLE_ENDIAN
747      uint8_t tess : 1;
748      uint8_t gs : 1;
749      uint8_t ngg_passthrough : 1;
750      uint8_t ngg : 1;       /* gfx10+ */
751      uint8_t streamout : 1; /* only used with NGG */
752      uint8_t _pad : 8 - SI_NUM_VGT_STAGES_KEY_BITS;
753#else /* UTIL_ARCH_BIG_ENDIAN */
754      uint8_t _pad : 8 - SI_NUM_VGT_STAGES_KEY_BITS;
755      uint8_t streamout : 1;
756      uint8_t ngg : 1;
757      uint8_t ngg_passthrough : 1;
758      uint8_t gs : 1;
759      uint8_t tess : 1;
760#endif
761   } u;
762   uint8_t index;
763};
764
765struct si_shader {
766   struct si_pm4_state pm4; /* base class */
767   struct si_compiler_ctx_state compiler_ctx_state;
768
769   struct si_shader_selector *selector;
770   struct si_shader_selector *previous_stage_sel; /* for refcounting */
771   struct si_shader *next_variant;
772
773   struct si_shader_part *prolog;
774   struct si_shader *previous_stage; /* for GFX9 */
775   struct si_shader_part *prolog2;
776   struct si_shader_part *epilog;
777
778   struct si_resource *bo;
779   struct si_resource *scratch_bo;
780   struct si_shader_key key;
781   struct util_queue_fence ready;
782   bool compilation_failed;
783   bool is_monolithic;
784   bool is_optimized;
785   bool is_binary_shared;
786   bool is_gs_copy_shader;
787
788   /* The following data is all that's needed for binary shaders. */
789   struct si_shader_binary binary;
790   struct ac_shader_config config;
791   struct si_shader_binary_info info;
792
793   /* SI_SGPR_VS_STATE_BITS */
794   bool uses_vs_state_provoking_vertex;
795   bool uses_vs_state_outprim;
796
797   bool uses_base_instance;
798
799   struct {
800      uint16_t ngg_emit_size; /* in dwords */
801      uint16_t hw_max_esverts;
802      uint16_t max_gsprims;
803      uint16_t max_out_verts;
804      uint16_t prim_amp_factor;
805      bool max_vert_out_per_gs_instance;
806   } ngg;
807
808   /* Shader key + LLVM IR + disassembly + statistics.
809    * Generated for debug contexts only.
810    */
811   char *shader_log;
812   size_t shader_log_size;
813
814   struct gfx9_gs_info gs_info;
815
816   /* For save precompute context registers values. */
817   union {
818      struct {
819         unsigned vgt_gsvs_ring_offset_1;
820         unsigned vgt_gsvs_ring_offset_2;
821         unsigned vgt_gsvs_ring_offset_3;
822         unsigned vgt_gsvs_ring_itemsize;
823         unsigned vgt_gs_max_vert_out;
824         unsigned vgt_gs_vert_itemsize;
825         unsigned vgt_gs_vert_itemsize_1;
826         unsigned vgt_gs_vert_itemsize_2;
827         unsigned vgt_gs_vert_itemsize_3;
828         unsigned vgt_gs_instance_cnt;
829         unsigned vgt_gs_onchip_cntl;
830         unsigned vgt_gs_max_prims_per_subgroup;
831         unsigned vgt_esgs_ring_itemsize;
832         unsigned spi_shader_pgm_rsrc3_gs;
833         unsigned spi_shader_pgm_rsrc4_gs;
834      } gs;
835
836      struct {
837         unsigned ge_max_output_per_subgroup;
838         unsigned ge_ngg_subgrp_cntl;
839         unsigned vgt_primitiveid_en;
840         unsigned vgt_gs_onchip_cntl;
841         unsigned vgt_gs_instance_cnt;
842         unsigned vgt_esgs_ring_itemsize;
843         unsigned spi_vs_out_config;
844         unsigned spi_shader_idx_format;
845         unsigned spi_shader_pos_format;
846         unsigned pa_cl_vte_cntl;
847         unsigned pa_cl_ngg_cntl;
848         unsigned vgt_gs_max_vert_out; /* for API GS */
849         unsigned ge_pc_alloc;         /* uconfig register */
850         unsigned spi_shader_pgm_rsrc3_gs;
851         unsigned spi_shader_pgm_rsrc4_gs;
852         union si_vgt_stages_key vgt_stages;
853      } ngg;
854
855      struct {
856         unsigned vgt_gs_mode;
857         unsigned vgt_primitiveid_en;
858         unsigned vgt_reuse_off;
859         unsigned spi_vs_out_config;
860         unsigned spi_shader_pos_format;
861         unsigned pa_cl_vte_cntl;
862         unsigned ge_pc_alloc; /* uconfig register */
863      } vs;
864
865      struct {
866         unsigned spi_ps_input_ena;
867         unsigned spi_ps_input_addr;
868         unsigned spi_baryc_cntl;
869         unsigned spi_ps_in_control;
870         unsigned spi_shader_z_format;
871         unsigned spi_shader_col_format;
872         unsigned cb_shader_mask;
873         unsigned num_interp;
874      } ps;
875   } ctx_reg;
876
877   /*For save precompute registers value */
878   unsigned vgt_tf_param;                /* VGT_TF_PARAM */
879   unsigned vgt_vertex_reuse_block_cntl; /* VGT_VERTEX_REUSE_BLOCK_CNTL */
880   unsigned pa_cl_vs_out_cntl;
881   unsigned ge_cntl;
882};
883
884struct si_shader_part {
885   struct si_shader_part *next;
886   union si_shader_part_key key;
887   struct si_shader_binary binary;
888   struct ac_shader_config config;
889};
890
891/* si_shader.c */
892bool si_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
893                       struct si_shader *shader, struct pipe_debug_callback *debug);
894bool si_create_shader_variant(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
895                              struct si_shader *shader, struct pipe_debug_callback *debug);
896void si_shader_destroy(struct si_shader *shader);
897unsigned si_shader_io_get_unique_index_patch(unsigned semantic);
898unsigned si_shader_io_get_unique_index(unsigned semantic, bool is_varying);
899bool si_shader_binary_upload(struct si_screen *sscreen, struct si_shader *shader,
900                             uint64_t scratch_va);
901void si_shader_dump(struct si_screen *sscreen, struct si_shader *shader,
902                    struct pipe_debug_callback *debug, FILE *f, bool check_debug_option);
903void si_shader_dump_stats_for_shader_db(struct si_screen *screen, struct si_shader *shader,
904                                        struct pipe_debug_callback *debug);
905void si_multiwave_lds_size_workaround(struct si_screen *sscreen, unsigned *lds_size);
906const char *si_get_shader_name(const struct si_shader *shader);
907void si_shader_binary_clean(struct si_shader_binary *binary);
908
909/* si_shader_llvm_gs.c */
910struct si_shader *si_generate_gs_copy_shader(struct si_screen *sscreen,
911                                             struct ac_llvm_compiler *compiler,
912                                             struct si_shader_selector *gs_selector,
913                                             struct pipe_debug_callback *debug);
914
915/* si_shader_nir.c */
916void si_nir_scan_shader(const struct nir_shader *nir, struct si_shader_info *info);
917void si_nir_opts(struct si_screen *sscreen, struct nir_shader *nir, bool first);
918void si_nir_late_opts(nir_shader *nir);
919char *si_finalize_nir(struct pipe_screen *screen, void *nirptr);
920
921/* si_state_shaders.c */
922void gfx9_get_gs_info(struct si_shader_selector *es, struct si_shader_selector *gs,
923                      struct gfx9_gs_info *out);
924bool gfx10_is_ngg_passthrough(struct si_shader *shader);
925
926/* Inline helpers. */
927
928/* Return the pointer to the main shader part's pointer. */
929static inline struct si_shader **si_get_main_shader_part(struct si_shader_selector *sel,
930                                                         const struct si_shader_key *key)
931{
932   if (key->as_ls)
933      return &sel->main_shader_part_ls;
934   if (key->as_es && key->as_ngg)
935      return &sel->main_shader_part_ngg_es;
936   if (key->as_es)
937      return &sel->main_shader_part_es;
938   if (key->as_ngg)
939      return &sel->main_shader_part_ngg;
940   return &sel->main_shader_part;
941}
942
943static inline bool si_shader_uses_bindless_samplers(struct si_shader_selector *selector)
944{
945   return selector ? selector->info.uses_bindless_samplers : false;
946}
947
948static inline bool si_shader_uses_bindless_images(struct si_shader_selector *selector)
949{
950   return selector ? selector->info.uses_bindless_images : false;
951}
952
953static inline bool gfx10_edgeflags_have_effect(struct si_shader *shader)
954{
955   if (shader->selector->info.stage == MESA_SHADER_VERTEX &&
956       !shader->selector->info.base.vs.blit_sgprs_amd &&
957       !(shader->key.opt.ngg_culling & SI_NGG_CULL_LINES))
958      return true;
959
960   return false;
961}
962
963static inline bool gfx10_ngg_writes_user_edgeflags(struct si_shader *shader)
964{
965   return gfx10_edgeflags_have_effect(shader) &&
966          shader->selector->info.writes_edgeflag;
967}
968
969#ifdef __cplusplus
970}
971#endif
972
973#endif
974