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
 143 #include <stdio.h>
 144
 145 struct nir_shader;
 146 struct si_shader;
 147 struct si_context;
 148
 149 #define SI_MAX_ATTRIBS          16
 150 #define SI_MAX_VS_OUTPUTS       40
 151
 152 /* Shader IO unique indices are supported for TGSI_SEMANTIC_GENERIC with an
 153  * index smaller than this.
 154  */
 155 #define SI_MAX_IO_GENERIC       43
 156
 157 /* SGPR user data indices */
 158 enum {
 159         SI_SGPR_RW_BUFFERS,  /* rings (& stream-out, VS only) */
 160 #if !HAVE_32BIT_POINTERS
 161         SI_SGPR_RW_BUFFERS_HI,
 162 #endif
 163         SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES,
 164 #if !HAVE_32BIT_POINTERS
 165         SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES_HI,
 166 #endif
 167         SI_SGPR_CONST_AND_SHADER_BUFFERS, /* or just a constant buffer 0 pointer */
 168 #if !HAVE_32BIT_POINTERS
 169         SI_SGPR_CONST_AND_SHADER_BUFFERS_HI,
 170 #endif
 171         SI_SGPR_SAMPLERS_AND_IMAGES,
 172 #if !HAVE_32BIT_POINTERS
 173         SI_SGPR_SAMPLERS_AND_IMAGES_HI,
 174 #endif
 175         SI_NUM_RESOURCE_SGPRS,
 176
 177         /* all VS variants */
 178         SI_SGPR_BASE_VERTEX = SI_NUM_RESOURCE_SGPRS,
 179         SI_SGPR_START_INSTANCE,
 180         SI_SGPR_DRAWID,
 181         SI_SGPR_VS_STATE_BITS,
 182         SI_VS_NUM_USER_SGPR,
 183
 184         SI_SGPR_VS_BLIT_DATA = SI_SGPR_CONST_AND_SHADER_BUFFERS,
 185
 186         /* TES */
 187         SI_SGPR_TES_OFFCHIP_LAYOUT = SI_NUM_RESOURCE_SGPRS,
 188         SI_SGPR_TES_OFFCHIP_ADDR,
 189         SI_TES_NUM_USER_SGPR,
 190
 191         /* GFX6-8: TCS only */
 192         GFX6_SGPR_TCS_OFFCHIP_LAYOUT = SI_NUM_RESOURCE_SGPRS,
 193         GFX6_SGPR_TCS_OUT_OFFSETS,
 194         GFX6_SGPR_TCS_OUT_LAYOUT,
 195         GFX6_SGPR_TCS_IN_LAYOUT,
 196         GFX6_TCS_NUM_USER_SGPR,
 197
 198         /* GFX9: Merged shaders. */
 199 #if HAVE_32BIT_POINTERS
 200         /* 2ND_CONST_AND_SHADER_BUFFERS is set in USER_DATA_ADDR_LO (SGPR0). */
 201         /* 2ND_SAMPLERS_AND_IMAGES is set in USER_DATA_ADDR_HI (SGPR1). */
 202         GFX9_MERGED_NUM_USER_SGPR = SI_VS_NUM_USER_SGPR,
 203 #else
 204         /* 2ND_CONST_AND_SHADER_BUFFERS is set in USER_DATA_ADDR_LO/HI (SGPR[0:1]). */
 205         GFX9_SGPR_2ND_SAMPLERS_AND_IMAGES = SI_VS_NUM_USER_SGPR,
 206         GFX9_SGPR_2ND_SAMPLERS_AND_IMAGES_HI,
 207         GFX9_MERGED_NUM_USER_SGPR,
 208 #endif
 209
 210         /* GFX9: Merged LS-HS (VS-TCS) only. */
 211         GFX9_SGPR_TCS_OFFCHIP_LAYOUT = GFX9_MERGED_NUM_USER_SGPR,
 212         GFX9_SGPR_TCS_OUT_OFFSETS,
 213         GFX9_SGPR_TCS_OUT_LAYOUT,
 214 #if !HAVE_32BIT_POINTERS
 215         GFX9_SGPR_align_for_vb_pointer,
 216 #endif
 217         GFX9_TCS_NUM_USER_SGPR,
 218
 219         /* GS limits */
 220         GFX6_GS_NUM_USER_SGPR = SI_NUM_RESOURCE_SGPRS,
 221 #if HAVE_32BIT_POINTERS
 222         GFX9_VSGS_NUM_USER_SGPR = SI_VS_NUM_USER_SGPR,
 223         GFX9_TESGS_NUM_USER_SGPR = SI_TES_NUM_USER_SGPR,
 224 #else
 225         GFX9_VSGS_NUM_USER_SGPR = GFX9_MERGED_NUM_USER_SGPR,
 226         GFX9_TESGS_NUM_USER_SGPR = GFX9_MERGED_NUM_USER_SGPR,
 227 #endif
 228         SI_GSCOPY_NUM_USER_SGPR = SI_SGPR_RW_BUFFERS + (HAVE_32BIT_POINTERS ? 1 : 2),
 229
 230         /* PS only */
 231         SI_SGPR_ALPHA_REF       = SI_NUM_RESOURCE_SGPRS,
 232         SI_PS_NUM_USER_SGPR,
 233 };
 234
 235 /* LLVM function parameter indices */
 236 enum {
 237         SI_NUM_RESOURCE_PARAMS = 4,
 238
 239         /* PS only parameters */
 240         SI_PARAM_ALPHA_REF = SI_NUM_RESOURCE_PARAMS,
 241         SI_PARAM_PRIM_MASK,
 242         SI_PARAM_PERSP_SAMPLE,
 243         SI_PARAM_PERSP_CENTER,
 244         SI_PARAM_PERSP_CENTROID,
 245         SI_PARAM_PERSP_PULL_MODEL,
 246         SI_PARAM_LINEAR_SAMPLE,
 247         SI_PARAM_LINEAR_CENTER,
 248         SI_PARAM_LINEAR_CENTROID,
 249         SI_PARAM_LINE_STIPPLE_TEX,
 250         SI_PARAM_POS_X_FLOAT,
 251         SI_PARAM_POS_Y_FLOAT,
 252         SI_PARAM_POS_Z_FLOAT,
 253         SI_PARAM_POS_W_FLOAT,
 254         SI_PARAM_FRONT_FACE,
 255         SI_PARAM_ANCILLARY,
 256         SI_PARAM_SAMPLE_COVERAGE,
 257         SI_PARAM_POS_FIXED_PT,
 258
 259         SI_NUM_PARAMS = SI_PARAM_POS_FIXED_PT + 9, /* +8 for COLOR[0..1] */
 260 };
 261
 262 /* Fields of driver-defined VS state SGPR. */
 263 /* Clamp vertex color output (only used in VS as VS). */
 264 #define S_VS_STATE_CLAMP_VERTEX_COLOR(x)        (((unsigned)(x) & 0x1) << 0)
 265 #define C_VS_STATE_CLAMP_VERTEX_COLOR           0xFFFFFFFE
 266 #define S_VS_STATE_INDEXED(x)                   (((unsigned)(x) & 0x1) << 1)
 267 #define C_VS_STATE_INDEXED                      0xFFFFFFFD
 268 #define S_VS_STATE_LS_OUT_PATCH_SIZE(x)         (((unsigned)(x) & 0x1FFF) << 8)
 269 #define C_VS_STATE_LS_OUT_PATCH_SIZE            0xFFE000FF
 270 #define S_VS_STATE_LS_OUT_VERTEX_SIZE(x)        (((unsigned)(x) & 0xFF) << 24)
 271 #define C_VS_STATE_LS_OUT_VERTEX_SIZE           0x00FFFFFF
 272
 273 /* SI-specific system values. */
 274 enum {
 275         TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI = TGSI_SEMANTIC_COUNT,
 276         TGSI_SEMANTIC_DEFAULT_TESSINNER_SI,
 277 };
 278
 279 enum {
 280         /* Use a property enum that VS wouldn't use. */
 281         TGSI_PROPERTY_VS_BLIT_SGPRS = TGSI_PROPERTY_FS_COORD_ORIGIN,
 282
 283         /* These represent the number of SGPRs the shader uses. */
 284         SI_VS_BLIT_SGPRS_POS = 3,
 285         SI_VS_BLIT_SGPRS_POS_COLOR = 7,
 286         SI_VS_BLIT_SGPRS_POS_TEXCOORD = 9,
 287 };
 288
 289 /* For VS shader key fix_fetch. */
 290 enum {
 291         SI_FIX_FETCH_NONE = 0,
 292         SI_FIX_FETCH_A2_SNORM,
 293         SI_FIX_FETCH_A2_SSCALED,
 294         SI_FIX_FETCH_A2_SINT,
 295         SI_FIX_FETCH_RGBA_32_UNORM,
 296         SI_FIX_FETCH_RGBX_32_UNORM,
 297         SI_FIX_FETCH_RGBA_32_SNORM,
 298         SI_FIX_FETCH_RGBX_32_SNORM,
 299         SI_FIX_FETCH_RGBA_32_USCALED,
 300         SI_FIX_FETCH_RGBA_32_SSCALED,
 301         SI_FIX_FETCH_RGBA_32_FIXED,
 302         SI_FIX_FETCH_RGBX_32_FIXED,
 303         SI_FIX_FETCH_RG_64_FLOAT,
 304         SI_FIX_FETCH_RGB_64_FLOAT,
 305         SI_FIX_FETCH_RGBA_64_FLOAT,
 306         SI_FIX_FETCH_RGB_8,     /* A = 1.0 */
 307         SI_FIX_FETCH_RGB_8_INT, /* A = 1 */
 308         SI_FIX_FETCH_RGB_16,
 309         SI_FIX_FETCH_RGB_16_INT,
 310 };
 311
 312 struct si_shader;
 313
 314 /* Per-thread persistent LLVM objects. */
 315 struct si_compiler {
 316         LLVMTargetMachineRef            tm;
 317         const char                      *triple;
 318         const char                      *data_layout;
 319         LLVMTargetLibraryInfoRef        target_library_info;
 320         LLVMPassManagerRef              passmgr;
 321 };
 322
 323 /* State of the context creating the shader object. */
 324 struct si_compiler_ctx_state {
 325         /* Should only be used by si_init_shader_selector_async and
 326          * si_build_shader_variant if thread_index == -1 (non-threaded). */
 327         struct si_compiler              *compiler;
 328
 329         /* Used if thread_index == -1 or if debug.async is true. */
 330         struct pipe_debug_callback      debug;
 331
 332         /* Used for creating the log string for gallium/ddebug. */
 333         bool                            is_debug_context;
 334 };
 335
 336 /* A shader selector is a gallium CSO and contains shader variants and
 337  * binaries for one TGSI program. This can be shared by multiple contexts.
 338  */
 339 struct si_shader_selector {
 340         struct pipe_reference   reference;
 341         struct si_screen        *screen;
 342         struct util_queue_fence ready;
 343         struct si_compiler_ctx_state compiler_ctx_state;
 344
 345         mtx_t           mutex;
 346         struct si_shader        *first_variant; /* immutable after the first variant */
 347         struct si_shader        *last_variant; /* mutable */
 348
 349         /* The compiled TGSI shader expecting a prolog and/or epilog (not
 350          * uploaded to a buffer).
 351          */
 352         struct si_shader        *main_shader_part;
 353         struct si_shader        *main_shader_part_ls; /* as_ls is set in the key */
 354         struct si_shader        *main_shader_part_es; /* as_es is set in the key */
 355
 356         struct si_shader        *gs_copy_shader;
 357
 358         struct tgsi_token       *tokens;
 359         struct nir_shader       *nir;
 360         struct pipe_stream_output_info  so;
 361         struct tgsi_shader_info         info;
 362         struct tgsi_tessctrl_info       tcs_info;
 363
 364         /* PIPE_SHADER_[VERTEX|FRAGMENT|...] */
 365         unsigned        type;
 366         bool            vs_needs_prolog;
 367         bool            force_correct_derivs_after_kill;
 368         unsigned        pa_cl_vs_out_cntl;
 369         ubyte           clipdist_mask;
 370         ubyte           culldist_mask;
 371
 372         /* ES parameters. */
 373         unsigned        esgs_itemsize;
 374
 375         /* GS parameters. */
 376         unsigned        gs_input_verts_per_prim;
 377         unsigned        gs_output_prim;
 378         unsigned        gs_max_out_vertices;
 379         unsigned        gs_num_invocations;
 380         unsigned        max_gs_stream; /* count - 1 */
 381         unsigned        gsvs_vertex_size;
 382         unsigned        max_gsvs_emit_size;
 383         unsigned        enabled_streamout_buffer_mask;
 384
 385         /* PS parameters. */
 386         unsigned        color_attr_index[2];
 387         unsigned        db_shader_control;
 388         /* Set 0xf or 0x0 (4 bits) per each written output.
 389          * ANDed with spi_shader_col_format.
 390          */
 391         unsigned        colors_written_4bit;
 392
 393         /* CS parameters */
 394         unsigned local_size;
 395
 396         uint64_t        outputs_written_before_ps; /* "get_unique_index" bits */
 397         uint64_t        outputs_written;        /* "get_unique_index" bits */
 398         uint32_t        patch_outputs_written;  /* "get_unique_index_patch" bits */
 399
 400         uint64_t        inputs_read;            /* "get_unique_index" bits */
 401
 402         /* bitmasks of used descriptor slots */
 403         uint32_t        active_const_and_shader_buffers;
 404         uint64_t        active_samplers_and_images;
 405 };
 406
 407 /* Valid shader configurations:
 408  *
 409  * API shaders       VS | TCS | TES | GS |pass| PS
 410  * are compiled as:     |     |     |    |thru|
 411  *                      |     |     |    |    |
 412  * Only VS & PS:     VS |     |     |    |    | PS
 413  * GFX6 - with GS:   ES |     |     | GS | VS | PS
 414  *      - with tess: LS | HS  | VS  |    |    | PS
 415  *      - with both: LS | HS  | ES  | GS | VS | PS
 416  * GFX9 - with GS:   -> |     |     | GS | VS | PS
 417  *      - with tess: -> | HS  | VS  |    |    | PS
 418  *      - with both: -> | HS  | ->  | GS | VS | PS
 419  *
 420  * -> = merged with the next stage
 421  */
 422
 423 /* Use the byte alignment for all following structure members for optimal
 424  * shader key memory footprint.
 425  */
 426 #pragma pack(push, 1)
 427
 428 /* Common VS bits between the shader key and the prolog key. */
 429 struct si_vs_prolog_bits {
 430         /* - If neither "is_one" nor "is_fetched" has a bit set, the instance
 431          *   divisor is 0.
 432          * - If "is_one" has a bit set, the instance divisor is 1.
 433          * - If "is_fetched" has a bit set, the instance divisor will be loaded
 434          *   from the constant buffer.
 435          */
 436         uint16_t        instance_divisor_is_one;     /* bitmask of inputs */
 437         uint16_t        instance_divisor_is_fetched; /* bitmask of inputs */
 438         unsigned        ls_vgpr_fix:1;
 439 };
 440
 441 /* Common TCS bits between the shader key and the epilog key. */
 442 struct si_tcs_epilog_bits {
 443         unsigned        prim_mode:3;
 444         unsigned        invoc0_tess_factors_are_def:1;
 445         unsigned        tes_reads_tess_factors:1;
 446 };
 447
 448 struct si_gs_prolog_bits {
 449         unsigned        tri_strip_adj_fix:1;
 450         unsigned        gfx9_prev_is_vs:1;
 451 };
 452
 453 /* Common PS bits between the shader key and the prolog key. */
 454 struct si_ps_prolog_bits {
 455         unsigned        color_two_side:1;
 456         unsigned        flatshade_colors:1;
 457         unsigned        poly_stipple:1;
 458         unsigned        force_persp_sample_interp:1;
 459         unsigned        force_linear_sample_interp:1;
 460         unsigned        force_persp_center_interp:1;
 461         unsigned        force_linear_center_interp:1;
 462         unsigned        bc_optimize_for_persp:1;
 463         unsigned        bc_optimize_for_linear:1;
 464         unsigned        samplemask_log_ps_iter:3;
 465 };
 466
 467 /* Common PS bits between the shader key and the epilog key. */
 468 struct si_ps_epilog_bits {
 469         unsigned        spi_shader_col_format;
 470         unsigned        color_is_int8:8;
 471         unsigned        color_is_int10:8;
 472         unsigned        last_cbuf:3;
 473         unsigned        alpha_func:3;
 474         unsigned        alpha_to_one:1;
 475         unsigned        poly_line_smoothing:1;
 476         unsigned        clamp_color:1;
 477 };
 478
 479 union si_shader_part_key {
 480         struct {
 481                 struct si_vs_prolog_bits states;
 482                 unsigned        num_input_sgprs:6;
 483                 /* For merged stages such as LS-HS, HS input VGPRs are first. */
 484                 unsigned        num_merged_next_stage_vgprs:3;
 485                 unsigned        last_input:4;
 486                 unsigned        as_ls:1;
 487                 unsigned        as_es:1;
 488                 /* Prologs for monolithic shaders shouldn't set EXEC. */
 489                 unsigned        is_monolithic:1;
 490         } vs_prolog;
 491         struct {
 492                 struct si_tcs_epilog_bits states;
 493         } tcs_epilog;
 494         struct {
 495                 struct si_gs_prolog_bits states;
 496                 /* Prologs of monolithic shaders shouldn't set EXEC. */
 497                 unsigned        is_monolithic:1;
 498         } gs_prolog;
 499         struct {
 500                 struct si_ps_prolog_bits states;
 501                 unsigned        num_input_sgprs:6;
 502                 unsigned        num_input_vgprs:5;
 503                 /* Color interpolation and two-side color selection. */
 504                 unsigned        colors_read:8; /* color input components read */
 505                 unsigned        num_interp_inputs:5; /* BCOLOR is at this location */
 506                 unsigned        face_vgpr_index:5;
 507                 unsigned        ancillary_vgpr_index:5;
 508                 unsigned        wqm:1;
 509                 char            color_attr_index[2];
 510                 char            color_interp_vgpr_index[2]; /* -1 == constant */
 511         } ps_prolog;
 512         struct {
 513                 struct si_ps_epilog_bits states;
 514                 unsigned        colors_written:8;
 515                 unsigned        writes_z:1;
 516                 unsigned        writes_stencil:1;
 517                 unsigned        writes_samplemask:1;
 518         } ps_epilog;
 519 };
 520
 521 struct si_shader_key {
 522         /* Prolog and epilog flags. */
 523         union {
 524                 struct {
 525                         struct si_vs_prolog_bits prolog;
 526                 } vs;
 527                 struct {
 528                         struct si_vs_prolog_bits ls_prolog; /* for merged LS-HS */
 529                         struct si_shader_selector *ls;   /* for merged LS-HS */
 530                         struct si_tcs_epilog_bits epilog;
 531                 } tcs; /* tessellation control shader */
 532                 struct {
 533                         struct si_vs_prolog_bits vs_prolog; /* for merged ES-GS */
 534                         struct si_shader_selector *es;   /* for merged ES-GS */
 535                         struct si_gs_prolog_bits prolog;
 536                 } gs;
 537                 struct {
 538                         struct si_ps_prolog_bits prolog;
 539                         struct si_ps_epilog_bits epilog;
 540                 } ps;
 541         } part;
 542
 543         /* These two are initially set according to the NEXT_SHADER property,
 544          * or guessed if the property doesn't seem correct.
 545          */
 546         unsigned as_es:1; /* export shader, which precedes GS */
 547         unsigned as_ls:1; /* local shader, which precedes TCS */
 548
 549         /* Flags for monolithic compilation only. */
 550         struct {
 551                 /* One byte for every input: SI_FIX_FETCH_* enums. */
 552                 uint8_t         vs_fix_fetch[SI_MAX_ATTRIBS];
 553
 554                 union {
 555                         uint64_t        ff_tcs_inputs_to_copy; /* for fixed-func TCS */
 556                         /* When PS needs PrimID and GS is disabled. */
 557                         unsigned        vs_export_prim_id:1;
 558                         struct {
 559                                 unsigned interpolate_at_sample_force_center:1;
 560                                 unsigned fbfetch_msaa;
 561                                 unsigned fbfetch_is_1D;
 562                                 unsigned fbfetch_layered;
 563                         } ps;
 564                 } u;
 565         } mono;
 566
 567         /* Optimization flags for asynchronous compilation only. */
 568         struct {
 569                 /* For HW VS (it can be VS, TES, GS) */
 570                 uint64_t        kill_outputs; /* "get_unique_index" bits */
 571                 unsigned        clip_disable:1;
 572
 573                 /* For shaders where monolithic variants have better code.
 574                  *
 575                  * This is a flag that has no effect on code generation,
 576                  * but forces monolithic shaders to be used as soon as
 577                  * possible, because it's in the "opt" group.
 578                  */
 579                 unsigned        prefer_mono:1;
 580         } opt;
 581 };
 582
 583 /* Restore the pack alignment to default. */
 584 #pragma pack(pop)
 585
 586 struct si_shader_config {
 587         unsigned                        num_sgprs;
 588         unsigned                        num_vgprs;
 589         unsigned                        spilled_sgprs;
 590         unsigned                        spilled_vgprs;
 591         unsigned                        private_mem_vgprs;
 592         unsigned                        lds_size;
 593         unsigned                        max_simd_waves;
 594         unsigned                        spi_ps_input_ena;
 595         unsigned                        spi_ps_input_addr;
 596         unsigned                        float_mode;
 597         unsigned                        scratch_bytes_per_wave;
 598         unsigned                        rsrc1;
 599         unsigned                        rsrc2;
 600 };
 601
 602 /* GCN-specific shader info. */
 603 struct si_shader_info {
 604         ubyte                   vs_output_param_offset[SI_MAX_VS_OUTPUTS];
 605         ubyte                   num_input_sgprs;
 606         ubyte                   num_input_vgprs;
 607         signed char             face_vgpr_index;
 608         signed char             ancillary_vgpr_index;
 609         bool                    uses_instanceid;
 610         ubyte                   nr_pos_exports;
 611         ubyte                   nr_param_exports;
 612 };
 613
 614 struct si_shader {
 615         struct si_compiler_ctx_state    compiler_ctx_state;
 616
 617         struct si_shader_selector       *selector;
 618         struct si_shader_selector       *previous_stage_sel; /* for refcounting */
 619         struct si_shader                *next_variant;
 620
 621         struct si_shader_part           *prolog;
 622         struct si_shader                *previous_stage; /* for GFX9 */
 623         struct si_shader_part           *prolog2;
 624         struct si_shader_part           *epilog;
 625
 626         struct si_pm4_state             *pm4;
 627         struct r600_resource            *bo;
 628         struct r600_resource            *scratch_bo;
 629         struct si_shader_key            key;
 630         struct util_queue_fence         ready;
 631         bool                            compilation_failed;
 632         bool                            is_monolithic;
 633         bool                            is_optimized;
 634         bool                            is_binary_shared;
 635         bool                            is_gs_copy_shader;
 636
 637         /* The following data is all that's needed for binary shaders. */
 638         struct ac_shader_binary binary;
 639         struct si_shader_config         config;
 640         struct si_shader_info           info;
 641
 642         /* Shader key + LLVM IR + disassembly + statistics.
 643          * Generated for debug contexts only.
 644          */
 645         char                            *shader_log;
 646         size_t                          shader_log_size;
 647 };
 648
 649 struct si_shader_part {
 650         struct si_shader_part *next;
 651         union si_shader_part_key key;
 652         struct ac_shader_binary binary;
 653         struct si_shader_config config;
 654 };
 655
 656 /* si_shader.c */
 657 struct si_shader *
 658 si_generate_gs_copy_shader(struct si_screen *sscreen,
 659                            struct si_compiler *compiler,
 660                            struct si_shader_selector *gs_selector,
 661                            struct pipe_debug_callback *debug);
 662 int si_compile_tgsi_shader(struct si_screen *sscreen,
 663                            struct si_compiler *compiler,
 664                            struct si_shader *shader,
 665                            bool is_monolithic,
 666                            struct pipe_debug_callback *debug);
 667 int si_shader_create(struct si_screen *sscreen, struct si_compiler *compiler,
 668                      struct si_shader *shader,
 669                      struct pipe_debug_callback *debug);
 670 void si_shader_destroy(struct si_shader *shader);
 671 unsigned si_shader_io_get_unique_index_patch(unsigned semantic_name, unsigned index);
 672 unsigned si_shader_io_get_unique_index(unsigned semantic_name, unsigned index,
 673                                        unsigned is_varying);
 674 int si_shader_binary_upload(struct si_screen *sscreen, struct si_shader *shader);
 675 void si_shader_dump(struct si_screen *sscreen, const struct si_shader *shader,
 676                     struct pipe_debug_callback *debug, unsigned processor,
 677                     FILE *f, bool check_debug_option);
 678 void si_shader_dump_stats_for_shader_db(const struct si_shader *shader,
 679                                         struct pipe_debug_callback *debug);
 680 void si_multiwave_lds_size_workaround(struct si_screen *sscreen,
 681                                       unsigned *lds_size);
 682 void si_shader_apply_scratch_relocs(struct si_shader *shader,
 683                                     uint64_t scratch_va);
 684 void si_shader_binary_read_config(struct ac_shader_binary *binary,
 685                                   struct si_shader_config *conf,
 686                                   unsigned symbol_offset);
 687 const char *si_get_shader_name(const struct si_shader *shader, unsigned processor);
 688
 689 /* si_shader_nir.c */
 690 void si_nir_scan_shader(const struct nir_shader *nir,
 691                         struct tgsi_shader_info *info);
 692 void si_nir_scan_tess_ctrl(const struct nir_shader *nir,
 693                            const struct tgsi_shader_info *info,
 694                            struct tgsi_tessctrl_info *out);
 695 void si_lower_nir(struct si_shader_selector *sel);
 696
 697 /* Inline helpers. */
 698
 699 /* Return the pointer to the main shader part's pointer. */
 700 static inline struct si_shader **
 701 si_get_main_shader_part(struct si_shader_selector *sel,
 702                         struct si_shader_key *key)
 703 {
 704         if (key->as_ls)
 705                 return &sel->main_shader_part_ls;
 706         if (key->as_es)
 707                 return &sel->main_shader_part_es;
 708         return &sel->main_shader_part;
 709 }
 710
 711 static inline bool
 712 si_shader_uses_bindless_samplers(struct si_shader_selector *selector)
 713 {
 714         return selector ? selector->info.uses_bindless_samplers : false;
 715 }
 716
 717 static inline bool
 718 si_shader_uses_bindless_images(struct si_shader_selector *selector)
 719 {
 720         return selector ? selector->info.uses_bindless_images : false;
 721 }
 722
 723 void si_destroy_shader_selector(struct si_context *sctx,
 724                                 struct si_shader_selector *sel);
 725
 726 static inline void
 727 si_shader_selector_reference(struct si_context *sctx,
 728                              struct si_shader_selector **dst,
 729                              struct si_shader_selector *src)
 730 {
 731         if (pipe_reference(&(*dst)->reference, &src->reference))
 732                 si_destroy_shader_selector(sctx, *dst);
 733
 734         *dst = src;
 735 }
 736
 737 #endif