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