src/gallium/drivers/radeonsi/si_shader.h

   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 <llvm-c/Core.h> /* LLVMModuleRef */
 135 #include <llvm-c/TargetMachine.h>
 136 #include "tgsi/tgsi_scan.h"
 137 #include "util/u_inlines.h"
 138 #include "util/u_queue.h"
 139
 140 #include "ac_binary.h"
 141 #include "ac_llvm_build.h"
 142 #include "ac_llvm_util.h"
 143
 144 #include <stdio.h>
 145
 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)
 149
 150 struct nir_shader;
 151 struct si_shader;
 152 struct si_context;
 153
 154 #define SI_MAX_ATTRIBS          16
 155 #define SI_MAX_VS_OUTPUTS       40
 156
 157 /* Shader IO unique indices are supported for TGSI_SEMANTIC_GENERIC with an
 158  * index smaller than this.
 159  */
 160 #define SI_MAX_IO_GENERIC       43
 161
 162 /* SGPR user data indices */
 163 enum {
 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,
 169
 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,
 173
 174         /* all VS variants */
 175         SI_SGPR_BASE_VERTEX = SI_NUM_VS_STATE_RESOURCE_SGPRS,
 176         SI_SGPR_START_INSTANCE,
 177         SI_SGPR_DRAWID,
 178         SI_VS_NUM_USER_SGPR,
 179
 180         SI_SGPR_VS_BLIT_DATA = SI_SGPR_CONST_AND_SHADER_BUFFERS,
 181
 182         /* TES */
 183         SI_SGPR_TES_OFFCHIP_LAYOUT = SI_NUM_VS_STATE_RESOURCE_SGPRS,
 184         SI_SGPR_TES_OFFCHIP_ADDR,
 185         SI_TES_NUM_USER_SGPR,
 186
 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,
 193
 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,
 198
 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,
 204
 205         /* GS limits */
 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,
 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 /* Driver-specific system values. */
 255 enum {
 256         /* Values from set_tess_state. */
 257         TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI = TGSI_SEMANTIC_COUNT,
 258         TGSI_SEMANTIC_DEFAULT_TESSINNER_SI,
 259
 260         /* Up to 4 dwords in user SGPRs for compute shaders. */
 261         TGSI_SEMANTIC_CS_USER_DATA,
 262 };
 263
 264 enum {
 265         /* Use a property enum that CS wouldn't use. */
 266         TGSI_PROPERTY_CS_LOCAL_SIZE = TGSI_PROPERTY_FS_COORD_ORIGIN,
 267
 268         /* The number of used user data dwords in the range [1, 4]. */
 269         TGSI_PROPERTY_CS_USER_DATA_DWORDS = TGSI_PROPERTY_FS_COORD_PIXEL_CENTER,
 270
 271         /* Use a property enum that VS wouldn't use. */
 272         TGSI_PROPERTY_VS_BLIT_SGPRS = TGSI_PROPERTY_FS_COORD_ORIGIN,
 273
 274         /* These represent the number of SGPRs the shader uses. */
 275         SI_VS_BLIT_SGPRS_POS = 3,
 276         SI_VS_BLIT_SGPRS_POS_COLOR = 7,
 277         SI_VS_BLIT_SGPRS_POS_TEXCOORD = 9,
 278 };
 279
 280 /**
 281  * For VS shader keys, describe any fixups required for vertex fetch.
 282  *
 283  * \ref log_size, \ref format, and the number of channels are interpreted as
 284  * by \ref ac_build_opencoded_load_format.
 285  *
 286  * Note: all bits 0 (size = 1 byte, num channels = 1, format = float) is an
 287  * impossible format and indicates that no fixup is needed (just use
 288  * buffer_load_format_xyzw).
 289  */
 290 union si_vs_fix_fetch {
 291         struct {
 292                 uint8_t log_size : 2; /* 1, 2, 4, 8 or bytes per channel */
 293                 uint8_t num_channels_m1 : 2; /* number of channels minus 1 */
 294                 uint8_t format : 3; /* AC_FETCH_FORMAT_xxx */
 295                 uint8_t reverse : 1; /* reverse XYZ channels */
 296         } u;
 297         uint8_t bits;
 298 };
 299
 300 struct si_shader;
 301
 302 /* State of the context creating the shader object. */
 303 struct si_compiler_ctx_state {
 304         /* Should only be used by si_init_shader_selector_async and
 305          * si_build_shader_variant if thread_index == -1 (non-threaded). */
 306         struct ac_llvm_compiler         *compiler;
 307
 308         /* Used if thread_index == -1 or if debug.async is true. */
 309         struct pipe_debug_callback      debug;
 310
 311         /* Used for creating the log string for gallium/ddebug. */
 312         bool                            is_debug_context;
 313 };
 314
 315 /* A shader selector is a gallium CSO and contains shader variants and
 316  * binaries for one TGSI program. This can be shared by multiple contexts.
 317  */
 318 struct si_shader_selector {
 319         struct pipe_reference   reference;
 320         struct si_screen        *screen;
 321         struct util_queue_fence ready;
 322         struct si_compiler_ctx_state compiler_ctx_state;
 323
 324         mtx_t           mutex;
 325         struct si_shader        *first_variant; /* immutable after the first variant */
 326         struct si_shader        *last_variant; /* mutable */
 327
 328         /* The compiled TGSI shader expecting a prolog and/or epilog (not
 329          * uploaded to a buffer).
 330          */
 331         struct si_shader        *main_shader_part;
 332         struct si_shader        *main_shader_part_ls; /* as_ls is set in the key */
 333         struct si_shader        *main_shader_part_es; /* as_es is set in the key */
 334
 335         struct si_shader        *gs_copy_shader;
 336
 337         struct tgsi_token       *tokens;
 338         struct nir_shader       *nir;
 339         struct pipe_stream_output_info  so;
 340         struct tgsi_shader_info         info;
 341         struct tgsi_tessctrl_info       tcs_info;
 342
 343         /* PIPE_SHADER_[VERTEX|FRAGMENT|...] */
 344         unsigned        type;
 345         bool            vs_needs_prolog;
 346         bool            force_correct_derivs_after_kill;
 347         bool            prim_discard_cs_allowed;
 348         unsigned        pa_cl_vs_out_cntl;
 349         ubyte           clipdist_mask;
 350         ubyte           culldist_mask;
 351
 352         /* ES parameters. */
 353         unsigned        esgs_itemsize; /* vertex stride */
 354         unsigned        lshs_vertex_stride;
 355
 356         /* GS parameters. */
 357         unsigned        gs_input_verts_per_prim;
 358         unsigned        gs_output_prim;
 359         unsigned        gs_max_out_vertices;
 360         unsigned        gs_num_invocations;
 361         unsigned        max_gs_stream; /* count - 1 */
 362         unsigned        gsvs_vertex_size;
 363         unsigned        max_gsvs_emit_size;
 364         unsigned        enabled_streamout_buffer_mask;
 365
 366         /* PS parameters. */
 367         unsigned        color_attr_index[2];
 368         unsigned        db_shader_control;
 369         /* Set 0xf or 0x0 (4 bits) per each written output.
 370          * ANDed with spi_shader_col_format.
 371          */
 372         unsigned        colors_written_4bit;
 373
 374         uint64_t        outputs_written_before_ps; /* "get_unique_index" bits */
 375         uint64_t        outputs_written;        /* "get_unique_index" bits */
 376         uint32_t        patch_outputs_written;  /* "get_unique_index_patch" bits */
 377
 378         uint64_t        inputs_read;            /* "get_unique_index" bits */
 379
 380         /* bitmasks of used descriptor slots */
 381         uint32_t        active_const_and_shader_buffers;
 382         uint64_t        active_samplers_and_images;
 383 };
 384
 385 /* Valid shader configurations:
 386  *
 387  * API shaders       VS | TCS | TES | GS |pass| PS
 388  * are compiled as:     |     |     |    |thru|
 389  *                      |     |     |    |    |
 390  * Only VS & PS:     VS |     |     |    |    | PS
 391  * GFX6 - with GS:   ES |     |     | GS | VS | PS
 392  *      - with tess: LS | HS  | VS  |    |    | PS
 393  *      - with both: LS | HS  | ES  | GS | VS | PS
 394  * GFX9 - with GS:   -> |     |     | GS | VS | PS
 395  *      - with tess: -> | HS  | VS  |    |    | PS
 396  *      - with both: -> | HS  | ->  | GS | VS | PS
 397  *
 398  * -> = merged with the next stage
 399  */
 400
 401 /* Use the byte alignment for all following structure members for optimal
 402  * shader key memory footprint.
 403  */
 404 #pragma pack(push, 1)
 405
 406 /* Common VS bits between the shader key and the prolog key. */
 407 struct si_vs_prolog_bits {
 408         /* - If neither "is_one" nor "is_fetched" has a bit set, the instance
 409          *   divisor is 0.
 410          * - If "is_one" has a bit set, the instance divisor is 1.
 411          * - If "is_fetched" has a bit set, the instance divisor will be loaded
 412          *   from the constant buffer.
 413          */
 414         uint16_t        instance_divisor_is_one;     /* bitmask of inputs */
 415         uint16_t        instance_divisor_is_fetched; /* bitmask of inputs */
 416         unsigned        ls_vgpr_fix:1;
 417         unsigned        unpack_instance_id_from_vertex_id:1;
 418 };
 419
 420 /* Common TCS bits between the shader key and the epilog key. */
 421 struct si_tcs_epilog_bits {
 422         unsigned        prim_mode:3;
 423         unsigned        invoc0_tess_factors_are_def:1;
 424         unsigned        tes_reads_tess_factors:1;
 425 };
 426
 427 struct si_gs_prolog_bits {
 428         unsigned        tri_strip_adj_fix:1;
 429         unsigned        gfx9_prev_is_vs:1;
 430 };
 431
 432 /* Common PS bits between the shader key and the prolog key. */
 433 struct si_ps_prolog_bits {
 434         unsigned        color_two_side:1;
 435         unsigned        flatshade_colors:1;
 436         unsigned        poly_stipple:1;
 437         unsigned        force_persp_sample_interp:1;
 438         unsigned        force_linear_sample_interp:1;
 439         unsigned        force_persp_center_interp:1;
 440         unsigned        force_linear_center_interp:1;
 441         unsigned        bc_optimize_for_persp:1;
 442         unsigned        bc_optimize_for_linear:1;
 443         unsigned        samplemask_log_ps_iter:3;
 444 };
 445
 446 /* Common PS bits between the shader key and the epilog key. */
 447 struct si_ps_epilog_bits {
 448         unsigned        spi_shader_col_format;
 449         unsigned        color_is_int8:8;
 450         unsigned        color_is_int10:8;
 451         unsigned        last_cbuf:3;
 452         unsigned        alpha_func:3;
 453         unsigned        alpha_to_one:1;
 454         unsigned        poly_line_smoothing:1;
 455         unsigned        clamp_color:1;
 456 };
 457
 458 union si_shader_part_key {
 459         struct {
 460                 struct si_vs_prolog_bits states;
 461                 unsigned        num_input_sgprs:6;
 462                 /* For merged stages such as LS-HS, HS input VGPRs are first. */
 463                 unsigned        num_merged_next_stage_vgprs:3;
 464                 unsigned        last_input:4;
 465                 unsigned        as_ls:1;
 466                 unsigned        as_es:1;
 467                 /* Prologs for monolithic shaders shouldn't set EXEC. */
 468                 unsigned        is_monolithic:1;
 469         } vs_prolog;
 470         struct {
 471                 struct si_tcs_epilog_bits states;
 472         } tcs_epilog;
 473         struct {
 474                 struct si_gs_prolog_bits states;
 475                 /* Prologs of monolithic shaders shouldn't set EXEC. */
 476                 unsigned        is_monolithic:1;
 477         } gs_prolog;
 478         struct {
 479                 struct si_ps_prolog_bits states;
 480                 unsigned        num_input_sgprs:6;
 481                 unsigned        num_input_vgprs:5;
 482                 /* Color interpolation and two-side color selection. */
 483                 unsigned        colors_read:8; /* color input components read */
 484                 unsigned        num_interp_inputs:5; /* BCOLOR is at this location */
 485                 unsigned        face_vgpr_index:5;
 486                 unsigned        ancillary_vgpr_index:5;
 487                 unsigned        wqm:1;
 488                 char            color_attr_index[2];
 489                 signed char     color_interp_vgpr_index[2]; /* -1 == constant */
 490         } ps_prolog;
 491         struct {
 492                 struct si_ps_epilog_bits states;
 493                 unsigned        colors_written:8;
 494                 unsigned        writes_z:1;
 495                 unsigned        writes_stencil:1;
 496                 unsigned        writes_samplemask:1;
 497         } ps_epilog;
 498 };
 499
 500 struct si_shader_key {
 501         /* Prolog and epilog flags. */
 502         union {
 503                 struct {
 504                         struct si_vs_prolog_bits prolog;
 505                 } vs;
 506                 struct {
 507                         struct si_vs_prolog_bits ls_prolog; /* for merged LS-HS */
 508                         struct si_shader_selector *ls;   /* for merged LS-HS */
 509                         struct si_tcs_epilog_bits epilog;
 510                 } tcs; /* tessellation control shader */
 511                 struct {
 512                         struct si_vs_prolog_bits vs_prolog; /* for merged ES-GS */
 513                         struct si_shader_selector *es;   /* for merged ES-GS */
 514                         struct si_gs_prolog_bits prolog;
 515                 } gs;
 516                 struct {
 517                         struct si_ps_prolog_bits prolog;
 518                         struct si_ps_epilog_bits epilog;
 519                 } ps;
 520         } part;
 521
 522         /* These two are initially set according to the NEXT_SHADER property,
 523          * or guessed if the property doesn't seem correct.
 524          */
 525         unsigned as_es:1; /* export shader, which precedes GS */
 526         unsigned as_ls:1; /* local shader, which precedes TCS */
 527
 528         /* Flags for monolithic compilation only. */
 529         struct {
 530                 /* Whether fetch should be opencoded according to vs_fix_fetch.
 531                  * Otherwise, if vs_fix_fetch is non-zero, buffer_load_format_xyzw
 532                  * with minimal fixups is used. */
 533                 uint16_t vs_fetch_opencode;
 534                 union si_vs_fix_fetch vs_fix_fetch[SI_MAX_ATTRIBS];
 535
 536                 union {
 537                         uint64_t        ff_tcs_inputs_to_copy; /* for fixed-func TCS */
 538                         /* When PS needs PrimID and GS is disabled. */
 539                         unsigned        vs_export_prim_id:1;
 540                         struct {
 541                                 unsigned interpolate_at_sample_force_center:1;
 542                                 unsigned fbfetch_msaa;
 543                                 unsigned fbfetch_is_1D;
 544                                 unsigned fbfetch_layered;
 545                         } ps;
 546                 } u;
 547         } mono;
 548
 549         /* Optimization flags for asynchronous compilation only. */
 550         struct {
 551                 /* For HW VS (it can be VS, TES, GS) */
 552                 uint64_t        kill_outputs; /* "get_unique_index" bits */
 553                 unsigned        clip_disable:1;
 554
 555                 /* For shaders where monolithic variants have better code.
 556                  *
 557                  * This is a flag that has no effect on code generation,
 558                  * but forces monolithic shaders to be used as soon as
 559                  * possible, because it's in the "opt" group.
 560                  */
 561                 unsigned        prefer_mono:1;
 562
 563                 /* Primitive discard compute shader. */
 564                 unsigned        vs_as_prim_discard_cs:1;
 565                 unsigned        cs_prim_type:4;
 566                 unsigned        cs_indexed:1;
 567                 unsigned        cs_instancing:1;
 568                 unsigned        cs_primitive_restart:1;
 569                 unsigned        cs_provoking_vertex_first:1;
 570                 unsigned        cs_need_correct_orientation:1;
 571                 unsigned        cs_cull_front:1;
 572                 unsigned        cs_cull_back:1;
 573                 unsigned        cs_cull_z:1;
 574                 unsigned        cs_halfz_clip_space:1;
 575         } opt;
 576 };
 577
 578 /* Restore the pack alignment to default. */
 579 #pragma pack(pop)
 580
 581 /* GCN-specific shader info. */
 582 struct si_shader_info {
 583         ubyte                   vs_output_param_offset[SI_MAX_VS_OUTPUTS];
 584         ubyte                   num_input_sgprs;
 585         ubyte                   num_input_vgprs;
 586         signed char             face_vgpr_index;
 587         signed char             ancillary_vgpr_index;
 588         bool                    uses_instanceid;
 589         ubyte                   nr_pos_exports;
 590         ubyte                   nr_param_exports;
 591         unsigned                private_mem_vgprs;
 592         unsigned                max_simd_waves;
 593 };
 594
 595 struct si_shader_binary {
 596         const char *elf_buffer;
 597         size_t elf_size;
 598
 599         char *llvm_ir_string;
 600 };
 601
 602 struct gfx9_gs_info {
 603         unsigned es_verts_per_subgroup;
 604         unsigned gs_prims_per_subgroup;
 605         unsigned gs_inst_prims_in_subgroup;
 606         unsigned max_prims_per_subgroup;
 607         unsigned esgs_ring_size; /* in bytes */
 608 };
 609
 610 struct si_shader {
 611         struct si_compiler_ctx_state    compiler_ctx_state;
 612
 613         struct si_shader_selector       *selector;
 614         struct si_shader_selector       *previous_stage_sel; /* for refcounting */
 615         struct si_shader                *next_variant;
 616
 617         struct si_shader_part           *prolog;
 618         struct si_shader                *previous_stage; /* for GFX9 */
 619         struct si_shader_part           *prolog2;
 620         struct si_shader_part           *epilog;
 621
 622         struct si_pm4_state             *pm4;
 623         struct si_resource              *bo;
 624         struct si_resource              *scratch_bo;
 625         struct si_shader_key            key;
 626         struct util_queue_fence         ready;
 627         bool                            compilation_failed;
 628         bool                            is_monolithic;
 629         bool                            is_optimized;
 630         bool                            is_binary_shared;
 631         bool                            is_gs_copy_shader;
 632
 633         /* The following data is all that's needed for binary shaders. */
 634         struct si_shader_binary         binary;
 635         struct ac_shader_config         config;
 636         struct si_shader_info           info;
 637
 638         /* Shader key + LLVM IR + disassembly + statistics.
 639          * Generated for debug contexts only.
 640          */
 641         char                            *shader_log;
 642         size_t                          shader_log_size;
 643
 644         struct gfx9_gs_info gs_info;
 645
 646         /* For save precompute context registers values. */
 647         union {
 648                 struct {
 649                         unsigned        vgt_gsvs_ring_offset_1;
 650                         unsigned        vgt_gsvs_ring_offset_2;
 651                         unsigned        vgt_gsvs_ring_offset_3;
 652                         unsigned        vgt_gs_out_prim_type;
 653                         unsigned        vgt_gsvs_ring_itemsize;
 654                         unsigned        vgt_gs_max_vert_out;
 655                         unsigned        vgt_gs_vert_itemsize;
 656                         unsigned        vgt_gs_vert_itemsize_1;
 657                         unsigned        vgt_gs_vert_itemsize_2;
 658                         unsigned        vgt_gs_vert_itemsize_3;
 659                         unsigned        vgt_gs_instance_cnt;
 660                         unsigned        vgt_gs_onchip_cntl;
 661                         unsigned        vgt_gs_max_prims_per_subgroup;
 662                         unsigned        vgt_esgs_ring_itemsize;
 663                 } gs;
 664
 665                 struct {
 666                         unsigned        vgt_gs_mode;
 667                         unsigned        vgt_primitiveid_en;
 668                         unsigned        vgt_reuse_off;
 669                         unsigned        spi_vs_out_config;
 670                         unsigned        spi_shader_pos_format;
 671                         unsigned        pa_cl_vte_cntl;
 672                 } vs;
 673
 674                 struct {
 675                         unsigned        spi_ps_input_ena;
 676                         unsigned        spi_ps_input_addr;
 677                         unsigned        spi_baryc_cntl;
 678                         unsigned        spi_ps_in_control;
 679                         unsigned        spi_shader_z_format;
 680                         unsigned        spi_shader_col_format;
 681                         unsigned        cb_shader_mask;
 682                 } ps;
 683         } ctx_reg;
 684
 685         /*For save precompute registers value */
 686         unsigned vgt_tf_param; /* VGT_TF_PARAM */
 687         unsigned vgt_vertex_reuse_block_cntl; /* VGT_VERTEX_REUSE_BLOCK_CNTL */
 688 };
 689
 690 struct si_shader_part {
 691         struct si_shader_part *next;
 692         union si_shader_part_key key;
 693         struct si_shader_binary binary;
 694         struct ac_shader_config config;
 695 };
 696
 697 /* si_shader.c */
 698 struct si_shader *
 699 si_generate_gs_copy_shader(struct si_screen *sscreen,
 700                            struct ac_llvm_compiler *compiler,
 701                            struct si_shader_selector *gs_selector,
 702                            struct pipe_debug_callback *debug);
 703 int si_compile_tgsi_shader(struct si_screen *sscreen,
 704                            struct ac_llvm_compiler *compiler,
 705                            struct si_shader *shader,
 706                            struct pipe_debug_callback *debug);
 707 bool si_shader_create(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
 708                      struct si_shader *shader,
 709                      struct pipe_debug_callback *debug);
 710 void si_shader_destroy(struct si_shader *shader);
 711 unsigned si_shader_io_get_unique_index_patch(unsigned semantic_name, unsigned index);
 712 unsigned si_shader_io_get_unique_index(unsigned semantic_name, unsigned index,
 713                                        unsigned is_varying);
 714 bool si_shader_binary_upload(struct si_screen *sscreen, struct si_shader *shader,
 715                              uint64_t scratch_va);
 716 void si_shader_dump(struct si_screen *sscreen, struct si_shader *shader,
 717                     struct pipe_debug_callback *debug, unsigned processor,
 718                     FILE *f, bool check_debug_option);
 719 void si_shader_dump_stats_for_shader_db(struct si_screen *screen,
 720                                         struct si_shader *shader,
 721                                         struct pipe_debug_callback *debug);
 722 void si_multiwave_lds_size_workaround(struct si_screen *sscreen,
 723                                       unsigned *lds_size);
 724 const char *si_get_shader_name(const struct si_shader *shader, unsigned processor);
 725 void si_shader_binary_clean(struct si_shader_binary *binary);
 726
 727 /* si_shader_nir.c */
 728 void si_nir_scan_shader(const struct nir_shader *nir,
 729                         struct tgsi_shader_info *info);
 730 void si_nir_scan_tess_ctrl(const struct nir_shader *nir,
 731                            struct tgsi_tessctrl_info *out);
 732 void si_lower_nir(struct si_shader_selector *sel);
 733 void si_nir_opts(struct nir_shader *nir);
 734
 735 /* si_state_shaders.c */
 736 void gfx9_get_gs_info(struct si_shader_selector *es,
 737                       struct si_shader_selector *gs,
 738                       struct gfx9_gs_info *out);
 739
 740 /* Inline helpers. */
 741
 742 /* Return the pointer to the main shader part's pointer. */
 743 static inline struct si_shader **
 744 si_get_main_shader_part(struct si_shader_selector *sel,
 745                         struct si_shader_key *key)
 746 {
 747         if (key->as_ls)
 748                 return &sel->main_shader_part_ls;
 749         if (key->as_es)
 750                 return &sel->main_shader_part_es;
 751         return &sel->main_shader_part;
 752 }
 753
 754 static inline bool
 755 si_shader_uses_bindless_samplers(struct si_shader_selector *selector)
 756 {
 757         return selector ? selector->info.uses_bindless_samplers : false;
 758 }
 759
 760 static inline bool
 761 si_shader_uses_bindless_images(struct si_shader_selector *selector)
 762 {
 763         return selector ? selector->info.uses_bindless_images : false;
 764 }
 765
 766 void si_destroy_shader_selector(struct si_context *sctx,
 767                                 struct si_shader_selector *sel);
 768
 769 static inline void
 770 si_shader_selector_reference(struct si_context *sctx,
 771                              struct si_shader_selector **dst,
 772                              struct si_shader_selector *src)
 773 {
 774         if (pipe_reference(&(*dst)->reference, &src->reference))
 775                 si_destroy_shader_selector(sctx, *dst);
 776
 777         *dst = src;
 778 }
 779
 780 #endif