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_queue.h"
 138
 139 #include "ac_binary.h"
 140 #include "ac_llvm_build.h"
 141 #include "si_state.h"
 142
 143 struct nir_shader;
 144
 145 #define SI_MAX_VS_OUTPUTS       40
 146
 147 /* Shader IO unique indices are supported for TGSI_SEMANTIC_GENERIC with an
 148  * index smaller than this.
 149  */
 150 #define SI_MAX_IO_GENERIC       46
 151
 152 /* SGPR user data indices */
 153 enum {
 154         SI_SGPR_RW_BUFFERS,  /* rings (& stream-out, VS only) */
 155 #if !HAVE_32BIT_POINTERS
 156         SI_SGPR_RW_BUFFERS_HI,
 157 #endif
 158         SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES,
 159 #if !HAVE_32BIT_POINTERS
 160         SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES_HI,
 161 #endif
 162         SI_SGPR_CONST_AND_SHADER_BUFFERS, /* or just a constant buffer 0 pointer */
 163 #if !HAVE_32BIT_POINTERS
 164         SI_SGPR_CONST_AND_SHADER_BUFFERS_HI,
 165 #endif
 166         SI_SGPR_SAMPLERS_AND_IMAGES,
 167 #if !HAVE_32BIT_POINTERS
 168         SI_SGPR_SAMPLERS_AND_IMAGES_HI,
 169 #endif
 170         SI_NUM_RESOURCE_SGPRS,
 171
 172         /* all VS variants */
 173         SI_SGPR_BASE_VERTEX = SI_NUM_RESOURCE_SGPRS,
 174         SI_SGPR_START_INSTANCE,
 175         SI_SGPR_DRAWID,
 176         SI_SGPR_VS_STATE_BITS,
 177         SI_VS_NUM_USER_SGPR,
 178
 179         SI_SGPR_VS_BLIT_DATA = SI_SGPR_CONST_AND_SHADER_BUFFERS,
 180
 181         /* TES */
 182         SI_SGPR_TES_OFFCHIP_LAYOUT = SI_NUM_RESOURCE_SGPRS,
 183         SI_SGPR_TES_OFFCHIP_ADDR,
 184         SI_TES_NUM_USER_SGPR,
 185
 186         /* GFX6-8: TCS only */
 187         GFX6_SGPR_TCS_OFFCHIP_LAYOUT = SI_NUM_RESOURCE_SGPRS,
 188         GFX6_SGPR_TCS_OUT_OFFSETS,
 189         GFX6_SGPR_TCS_OUT_LAYOUT,
 190         GFX6_SGPR_TCS_IN_LAYOUT,
 191         GFX6_TCS_NUM_USER_SGPR,
 192
 193         /* GFX9: Merged shaders. */
 194 #if HAVE_32BIT_POINTERS
 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 #else
 199         /* 2ND_CONST_AND_SHADER_BUFFERS is set in USER_DATA_ADDR_LO/HI (SGPR[0:1]). */
 200         GFX9_SGPR_2ND_SAMPLERS_AND_IMAGES = SI_VS_NUM_USER_SGPR,
 201         GFX9_SGPR_2ND_SAMPLERS_AND_IMAGES_HI,
 202         GFX9_MERGED_NUM_USER_SGPR,
 203 #endif
 204
 205         /* GFX9: Merged LS-HS (VS-TCS) only. */
 206         GFX9_SGPR_TCS_OFFCHIP_LAYOUT = GFX9_MERGED_NUM_USER_SGPR,
 207         GFX9_SGPR_TCS_OUT_OFFSETS,
 208         GFX9_SGPR_TCS_OUT_LAYOUT,
 209 #if !HAVE_32BIT_POINTERS
 210         GFX9_SGPR_align_for_vb_pointer,
 211 #endif
 212         GFX9_TCS_NUM_USER_SGPR,
 213
 214         /* GS limits */
 215         GFX6_GS_NUM_USER_SGPR = SI_NUM_RESOURCE_SGPRS,
 216 #if HAVE_32BIT_POINTERS
 217         GFX9_VSGS_NUM_USER_SGPR = SI_VS_NUM_USER_SGPR,
 218         GFX9_TESGS_NUM_USER_SGPR = SI_TES_NUM_USER_SGPR,
 219 #else
 220         GFX9_VSGS_NUM_USER_SGPR = GFX9_MERGED_NUM_USER_SGPR,
 221         GFX9_TESGS_NUM_USER_SGPR = GFX9_MERGED_NUM_USER_SGPR,
 222 #endif
 223         SI_GSCOPY_NUM_USER_SGPR = SI_SGPR_RW_BUFFERS + (HAVE_32BIT_POINTERS ? 1 : 2),
 224
 225         /* PS only */
 226         SI_SGPR_ALPHA_REF       = SI_NUM_RESOURCE_SGPRS,
 227         SI_PS_NUM_USER_SGPR,
 228 };
 229
 230 /* LLVM function parameter indices */
 231 enum {
 232         SI_NUM_RESOURCE_PARAMS = 4,
 233
 234         /* PS only parameters */
 235         SI_PARAM_ALPHA_REF = SI_NUM_RESOURCE_PARAMS,
 236         SI_PARAM_PRIM_MASK,
 237         SI_PARAM_PERSP_SAMPLE,
 238         SI_PARAM_PERSP_CENTER,
 239         SI_PARAM_PERSP_CENTROID,
 240         SI_PARAM_PERSP_PULL_MODEL,
 241         SI_PARAM_LINEAR_SAMPLE,
 242         SI_PARAM_LINEAR_CENTER,
 243         SI_PARAM_LINEAR_CENTROID,
 244         SI_PARAM_LINE_STIPPLE_TEX,
 245         SI_PARAM_POS_X_FLOAT,
 246         SI_PARAM_POS_Y_FLOAT,
 247         SI_PARAM_POS_Z_FLOAT,
 248         SI_PARAM_POS_W_FLOAT,
 249         SI_PARAM_FRONT_FACE,
 250         SI_PARAM_ANCILLARY,
 251         SI_PARAM_SAMPLE_COVERAGE,
 252         SI_PARAM_POS_FIXED_PT,
 253
 254         SI_NUM_PARAMS = SI_PARAM_POS_FIXED_PT + 9, /* +8 for COLOR[0..1] */
 255 };
 256
 257 /* Fields of driver-defined VS state SGPR. */
 258 /* Clamp vertex color output (only used in VS as VS). */
 259 #define S_VS_STATE_CLAMP_VERTEX_COLOR(x)        (((unsigned)(x) & 0x1) << 0)
 260 #define C_VS_STATE_CLAMP_VERTEX_COLOR           0xFFFFFFFE
 261 #define S_VS_STATE_INDEXED(x)                   (((unsigned)(x) & 0x1) << 1)
 262 #define C_VS_STATE_INDEXED                      0xFFFFFFFD
 263 #define S_VS_STATE_LS_OUT_PATCH_SIZE(x)         (((unsigned)(x) & 0x1FFF) << 8)
 264 #define C_VS_STATE_LS_OUT_PATCH_SIZE            0xFFE000FF
 265 #define S_VS_STATE_LS_OUT_VERTEX_SIZE(x)        (((unsigned)(x) & 0xFF) << 24)
 266 #define C_VS_STATE_LS_OUT_VERTEX_SIZE           0x00FFFFFF
 267
 268 /* SI-specific system values. */
 269 enum {
 270         TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI = TGSI_SEMANTIC_COUNT,
 271         TGSI_SEMANTIC_DEFAULT_TESSINNER_SI,
 272 };
 273
 274 enum {
 275         /* Use a property enum that VS wouldn't use. */
 276         TGSI_PROPERTY_VS_BLIT_SGPRS = TGSI_PROPERTY_FS_COORD_ORIGIN,
 277
 278         /* These represent the number of SGPRs the shader uses. */
 279         SI_VS_BLIT_SGPRS_POS = 3,
 280         SI_VS_BLIT_SGPRS_POS_COLOR = 7,
 281         SI_VS_BLIT_SGPRS_POS_TEXCOORD = 9,
 282 };
 283
 284 /* For VS shader key fix_fetch. */
 285 enum {
 286         SI_FIX_FETCH_NONE = 0,
 287         SI_FIX_FETCH_A2_SNORM,
 288         SI_FIX_FETCH_A2_SSCALED,
 289         SI_FIX_FETCH_A2_SINT,
 290         SI_FIX_FETCH_RGBA_32_UNORM,
 291         SI_FIX_FETCH_RGBX_32_UNORM,
 292         SI_FIX_FETCH_RGBA_32_SNORM,
 293         SI_FIX_FETCH_RGBX_32_SNORM,
 294         SI_FIX_FETCH_RGBA_32_USCALED,
 295         SI_FIX_FETCH_RGBA_32_SSCALED,
 296         SI_FIX_FETCH_RGBA_32_FIXED,
 297         SI_FIX_FETCH_RGBX_32_FIXED,
 298         SI_FIX_FETCH_RG_64_FLOAT,
 299         SI_FIX_FETCH_RGB_64_FLOAT,
 300         SI_FIX_FETCH_RGBA_64_FLOAT,
 301         SI_FIX_FETCH_RGB_8,     /* A = 1.0 */
 302         SI_FIX_FETCH_RGB_8_INT, /* A = 1 */
 303         SI_FIX_FETCH_RGB_16,
 304         SI_FIX_FETCH_RGB_16_INT,
 305 };
 306
 307 struct si_shader;
 308
 309 /* State of the context creating the shader object. */
 310 struct 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         LLVMTargetMachineRef            tm;
 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
 322 /* A shader selector is a gallium CSO and contains shader variants and
 323  * binaries for one TGSI program. This can be shared by multiple contexts.
 324  */
 325 struct si_shader_selector {
 326         struct pipe_reference   reference;
 327         struct si_screen        *screen;
 328         struct util_queue_fence ready;
 329         struct si_compiler_ctx_state compiler_ctx_state;
 330
 331         mtx_t           mutex;
 332         struct si_shader        *first_variant; /* immutable after the first variant */
 333         struct si_shader        *last_variant; /* mutable */
 334
 335         /* The compiled TGSI shader expecting a prolog and/or epilog (not
 336          * uploaded to a buffer).
 337          */
 338         struct si_shader        *main_shader_part;
 339         struct si_shader        *main_shader_part_ls; /* as_ls is set in the key */
 340         struct si_shader        *main_shader_part_es; /* as_es is set in the key */
 341
 342         struct si_shader        *gs_copy_shader;
 343
 344         struct tgsi_token       *tokens;
 345         struct nir_shader       *nir;
 346         struct pipe_stream_output_info  so;
 347         struct tgsi_shader_info         info;
 348         struct tgsi_tessctrl_info       tcs_info;
 349
 350         /* PIPE_SHADER_[VERTEX|FRAGMENT|...] */
 351         unsigned        type;
 352         bool            vs_needs_prolog;
 353         bool            force_correct_derivs_after_kill;
 354         unsigned        pa_cl_vs_out_cntl;
 355         ubyte           clipdist_mask;
 356         ubyte           culldist_mask;
 357
 358         /* ES parameters. */
 359         unsigned        esgs_itemsize;
 360
 361         /* GS parameters. */
 362         unsigned        gs_input_verts_per_prim;
 363         unsigned        gs_output_prim;
 364         unsigned        gs_max_out_vertices;
 365         unsigned        gs_num_invocations;
 366         unsigned        max_gs_stream; /* count - 1 */
 367         unsigned        gsvs_vertex_size;
 368         unsigned        max_gsvs_emit_size;
 369         unsigned        enabled_streamout_buffer_mask;
 370
 371         /* PS parameters. */
 372         unsigned        color_attr_index[2];
 373         unsigned        db_shader_control;
 374         /* Set 0xf or 0x0 (4 bits) per each written output.
 375          * ANDed with spi_shader_col_format.
 376          */
 377         unsigned        colors_written_4bit;
 378
 379         /* CS parameters */
 380         unsigned local_size;
 381
 382         uint64_t        outputs_written;        /* "get_unique_index" bits */
 383         uint32_t        patch_outputs_written;  /* "get_unique_index_patch" bits */
 384
 385         uint64_t        inputs_read;            /* "get_unique_index" bits */
 386
 387         /* bitmasks of used descriptor slots */
 388         uint32_t        active_const_and_shader_buffers;
 389         uint64_t        active_samplers_and_images;
 390 };
 391
 392 /* Valid shader configurations:
 393  *
 394  * API shaders       VS | TCS | TES | GS |pass| PS
 395  * are compiled as:     |     |     |    |thru|
 396  *                      |     |     |    |    |
 397  * Only VS & PS:     VS |     |     |    |    | PS
 398  * GFX6 - with GS:   ES |     |     | GS | VS | PS
 399  *      - with tess: LS | HS  | VS  |    |    | PS
 400  *      - with both: LS | HS  | ES  | GS | VS | PS
 401  * GFX9 - with GS:   -> |     |     | GS | VS | PS
 402  *      - with tess: -> | HS  | VS  |    |    | PS
 403  *      - with both: -> | HS  | ->  | GS | VS | PS
 404  *
 405  * -> = merged with the next stage
 406  */
 407
 408 /* Use the byte alignment for all following structure members for optimal
 409  * shader key memory footprint.
 410  */
 411 #pragma pack(push, 1)
 412
 413 /* Common VS bits between the shader key and the prolog key. */
 414 struct si_vs_prolog_bits {
 415         /* - If neither "is_one" nor "is_fetched" has a bit set, the instance
 416          *   divisor is 0.
 417          * - If "is_one" has a bit set, the instance divisor is 1.
 418          * - If "is_fetched" has a bit set, the instance divisor will be loaded
 419          *   from the constant buffer.
 420          */
 421         uint16_t        instance_divisor_is_one;     /* bitmask of inputs */
 422         uint16_t        instance_divisor_is_fetched; /* bitmask of inputs */
 423         unsigned        ls_vgpr_fix:1;
 424 };
 425
 426 /* Common TCS bits between the shader key and the epilog key. */
 427 struct si_tcs_epilog_bits {
 428         unsigned        prim_mode:3;
 429         unsigned        invoc0_tess_factors_are_def:1;
 430         unsigned        tes_reads_tess_factors:1;
 431 };
 432
 433 struct si_gs_prolog_bits {
 434         unsigned        tri_strip_adj_fix:1;
 435         unsigned        gfx9_prev_is_vs:1;
 436 };
 437
 438 /* Common PS bits between the shader key and the prolog key. */
 439 struct si_ps_prolog_bits {
 440         unsigned        color_two_side:1;
 441         unsigned        flatshade_colors:1;
 442         unsigned        poly_stipple:1;
 443         unsigned        force_persp_sample_interp:1;
 444         unsigned        force_linear_sample_interp:1;
 445         unsigned        force_persp_center_interp:1;
 446         unsigned        force_linear_center_interp:1;
 447         unsigned        bc_optimize_for_persp:1;
 448         unsigned        bc_optimize_for_linear:1;
 449         unsigned        samplemask_log_ps_iter:3;
 450 };
 451
 452 /* Common PS bits between the shader key and the epilog key. */
 453 struct si_ps_epilog_bits {
 454         unsigned        spi_shader_col_format;
 455         unsigned        color_is_int8:8;
 456         unsigned        color_is_int10:8;
 457         unsigned        last_cbuf:3;
 458         unsigned        alpha_func:3;
 459         unsigned        alpha_to_one:1;
 460         unsigned        poly_line_smoothing:1;
 461         unsigned        clamp_color:1;
 462 };
 463
 464 union si_shader_part_key {
 465         struct {
 466                 struct si_vs_prolog_bits states;
 467                 unsigned        num_input_sgprs:6;
 468                 /* For merged stages such as LS-HS, HS input VGPRs are first. */
 469                 unsigned        num_merged_next_stage_vgprs:3;
 470                 unsigned        last_input:4;
 471                 unsigned        as_ls:1;
 472                 unsigned        as_es:1;
 473                 /* Prologs for monolithic shaders shouldn't set EXEC. */
 474                 unsigned        is_monolithic:1;
 475         } vs_prolog;
 476         struct {
 477                 struct si_tcs_epilog_bits states;
 478         } tcs_epilog;
 479         struct {
 480                 struct si_gs_prolog_bits states;
 481                 /* Prologs of monolithic shaders shouldn't set EXEC. */
 482                 unsigned        is_monolithic:1;
 483         } gs_prolog;
 484         struct {
 485                 struct si_ps_prolog_bits states;
 486                 unsigned        num_input_sgprs:6;
 487                 unsigned        num_input_vgprs:5;
 488                 /* Color interpolation and two-side color selection. */
 489                 unsigned        colors_read:8; /* color input components read */
 490                 unsigned        num_interp_inputs:5; /* BCOLOR is at this location */
 491                 unsigned        face_vgpr_index:5;
 492                 unsigned        ancillary_vgpr_index:5;
 493                 unsigned        wqm:1;
 494                 char            color_attr_index[2];
 495                 char            color_interp_vgpr_index[2]; /* -1 == constant */
 496         } ps_prolog;
 497         struct {
 498                 struct si_ps_epilog_bits states;
 499                 unsigned        colors_written:8;
 500                 unsigned        writes_z:1;
 501                 unsigned        writes_stencil:1;
 502                 unsigned        writes_samplemask:1;
 503         } ps_epilog;
 504 };
 505
 506 struct si_shader_key {
 507         /* Prolog and epilog flags. */
 508         union {
 509                 struct {
 510                         struct si_vs_prolog_bits prolog;
 511                 } vs;
 512                 struct {
 513                         struct si_vs_prolog_bits ls_prolog; /* for merged LS-HS */
 514                         struct si_shader_selector *ls;   /* for merged LS-HS */
 515                         struct si_tcs_epilog_bits epilog;
 516                 } tcs; /* tessellation control shader */
 517                 struct {
 518                         struct si_vs_prolog_bits vs_prolog; /* for merged ES-GS */
 519                         struct si_shader_selector *es;   /* for merged ES-GS */
 520                         struct si_gs_prolog_bits prolog;
 521                 } gs;
 522                 struct {
 523                         struct si_ps_prolog_bits prolog;
 524                         struct si_ps_epilog_bits epilog;
 525                 } ps;
 526         } part;
 527
 528         /* These two are initially set according to the NEXT_SHADER property,
 529          * or guessed if the property doesn't seem correct.
 530          */
 531         unsigned as_es:1; /* export shader, which precedes GS */
 532         unsigned as_ls:1; /* local shader, which precedes TCS */
 533
 534         /* Flags for monolithic compilation only. */
 535         struct {
 536                 /* One byte for every input: SI_FIX_FETCH_* enums. */
 537                 uint8_t         vs_fix_fetch[SI_MAX_ATTRIBS];
 538
 539                 union {
 540                         uint64_t        ff_tcs_inputs_to_copy; /* for fixed-func TCS */
 541                         /* When PS needs PrimID and GS is disabled. */
 542                         unsigned        vs_export_prim_id:1;
 543                         struct {
 544                                 unsigned interpolate_at_sample_force_center:1;
 545                                 unsigned fbfetch_msaa;
 546                                 unsigned fbfetch_is_1D;
 547                                 unsigned fbfetch_layered;
 548                         } ps;
 549                 } u;
 550         } mono;
 551
 552         /* Optimization flags for asynchronous compilation only. */
 553         struct {
 554                 /* For HW VS (it can be VS, TES, GS) */
 555                 uint64_t        kill_outputs; /* "get_unique_index" bits */
 556                 unsigned        clip_disable:1;
 557
 558                 /* For shaders where monolithic variants have better code.
 559                  *
 560                  * This is a flag that has no effect on code generation,
 561                  * but forces monolithic shaders to be used as soon as
 562                  * possible, because it's in the "opt" group.
 563                  */
 564                 unsigned        prefer_mono:1;
 565         } opt;
 566 };
 567
 568 /* Restore the pack alignment to default. */
 569 #pragma pack(pop)
 570
 571 struct si_shader_config {
 572         unsigned                        num_sgprs;
 573         unsigned                        num_vgprs;
 574         unsigned                        spilled_sgprs;
 575         unsigned                        spilled_vgprs;
 576         unsigned                        private_mem_vgprs;
 577         unsigned                        lds_size;
 578         unsigned                        max_simd_waves;
 579         unsigned                        spi_ps_input_ena;
 580         unsigned                        spi_ps_input_addr;
 581         unsigned                        float_mode;
 582         unsigned                        scratch_bytes_per_wave;
 583         unsigned                        rsrc1;
 584         unsigned                        rsrc2;
 585 };
 586
 587 /* GCN-specific shader info. */
 588 struct si_shader_info {
 589         ubyte                   vs_output_param_offset[SI_MAX_VS_OUTPUTS];
 590         ubyte                   num_input_sgprs;
 591         ubyte                   num_input_vgprs;
 592         signed char             face_vgpr_index;
 593         signed char             ancillary_vgpr_index;
 594         bool                    uses_instanceid;
 595         ubyte                   nr_pos_exports;
 596         ubyte                   nr_param_exports;
 597 };
 598
 599 struct si_shader {
 600         struct si_compiler_ctx_state    compiler_ctx_state;
 601
 602         struct si_shader_selector       *selector;
 603         struct si_shader_selector       *previous_stage_sel; /* for refcounting */
 604         struct si_shader                *next_variant;
 605
 606         struct si_shader_part           *prolog;
 607         struct si_shader                *previous_stage; /* for GFX9 */
 608         struct si_shader_part           *prolog2;
 609         struct si_shader_part           *epilog;
 610
 611         struct si_pm4_state             *pm4;
 612         struct r600_resource            *bo;
 613         struct r600_resource            *scratch_bo;
 614         struct si_shader_key            key;
 615         struct util_queue_fence         ready;
 616         bool                            compilation_failed;
 617         bool                            is_monolithic;
 618         bool                            is_optimized;
 619         bool                            is_binary_shared;
 620         bool                            is_gs_copy_shader;
 621
 622         /* The following data is all that's needed for binary shaders. */
 623         struct ac_shader_binary binary;
 624         struct si_shader_config         config;
 625         struct si_shader_info           info;
 626
 627         /* Shader key + LLVM IR + disassembly + statistics.
 628          * Generated for debug contexts only.
 629          */
 630         char                            *shader_log;
 631         size_t                          shader_log_size;
 632 };
 633
 634 struct si_shader_part {
 635         struct si_shader_part *next;
 636         union si_shader_part_key key;
 637         struct ac_shader_binary binary;
 638         struct si_shader_config config;
 639 };
 640
 641 /* si_shader.c */
 642 struct si_shader *
 643 si_generate_gs_copy_shader(struct si_screen *sscreen,
 644                            LLVMTargetMachineRef tm,
 645                            struct si_shader_selector *gs_selector,
 646                            struct pipe_debug_callback *debug);
 647 int si_compile_tgsi_shader(struct si_screen *sscreen,
 648                            LLVMTargetMachineRef tm,
 649                            struct si_shader *shader,
 650                            bool is_monolithic,
 651                            struct pipe_debug_callback *debug);
 652 int si_shader_create(struct si_screen *sscreen, LLVMTargetMachineRef tm,
 653                      struct si_shader *shader,
 654                      struct pipe_debug_callback *debug);
 655 void si_shader_destroy(struct si_shader *shader);
 656 unsigned si_shader_io_get_unique_index_patch(unsigned semantic_name, unsigned index);
 657 unsigned si_shader_io_get_unique_index(unsigned semantic_name, unsigned index);
 658 int si_shader_binary_upload(struct si_screen *sscreen, struct si_shader *shader);
 659 void si_shader_dump(struct si_screen *sscreen, const struct si_shader *shader,
 660                     struct pipe_debug_callback *debug, unsigned processor,
 661                     FILE *f, bool check_debug_option);
 662 void si_shader_dump_stats_for_shader_db(const struct si_shader *shader,
 663                                         struct pipe_debug_callback *debug);
 664 void si_multiwave_lds_size_workaround(struct si_screen *sscreen,
 665                                       unsigned *lds_size);
 666 void si_shader_apply_scratch_relocs(struct si_shader *shader,
 667                                     uint64_t scratch_va);
 668 void si_shader_binary_read_config(struct ac_shader_binary *binary,
 669                                   struct si_shader_config *conf,
 670                                   unsigned symbol_offset);
 671 const char *si_get_shader_name(const struct si_shader *shader, unsigned processor);
 672
 673 /* si_shader_nir.c */
 674 void si_nir_scan_shader(const struct nir_shader *nir,
 675                         struct tgsi_shader_info *info);
 676 void si_nir_scan_tess_ctrl(const struct nir_shader *nir,
 677                            const struct tgsi_shader_info *info,
 678                            struct tgsi_tessctrl_info *out);
 679 void si_lower_nir(struct si_shader_selector *sel);
 680
 681 /* Inline helpers. */
 682
 683 /* Return the pointer to the main shader part's pointer. */
 684 static inline struct si_shader **
 685 si_get_main_shader_part(struct si_shader_selector *sel,
 686                         struct si_shader_key *key)
 687 {
 688         if (key->as_ls)
 689                 return &sel->main_shader_part_ls;
 690         if (key->as_es)
 691                 return &sel->main_shader_part_es;
 692         return &sel->main_shader_part;
 693 }
 694
 695 static inline bool
 696 si_shader_uses_bindless_samplers(struct si_shader_selector *selector)
 697 {
 698         return selector ? selector->info.uses_bindless_samplers : false;
 699 }
 700
 701 static inline bool
 702 si_shader_uses_bindless_images(struct si_shader_selector *selector)
 703 {
 704         return selector ? selector->info.uses_bindless_images : false;
 705 }
 706
 707 void si_destroy_shader_selector(struct si_context *sctx,
 708                                 struct si_shader_selector *sel);
 709
 710 static inline void
 711 si_shader_selector_reference(struct si_context *sctx,
 712                              struct si_shader_selector **dst,
 713                              struct si_shader_selector *src)
 714 {
 715         if (pipe_reference(&(*dst)->reference, &src->reference))
 716                 si_destroy_shader_selector(sctx, *dst);
 717
 718         *dst = src;
 719 }
 720
 721 #endif