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 "ac_binary.h"
 135 #include "ac_llvm_build.h"
 136 #include "ac_llvm_util.h"
 137 #include "util/simple_mtx.h"
 138 #include "util/u_inlines.h"
 139 #include "util/u_live_shader_cache.h"
 140 #include "util/u_queue.h"
 141
 142 #include <stdio.h>
 143
 144 // Use LDS symbols when supported by LLVM. Can be disabled for testing the old
 145 // path on newer LLVM for now. Should be removed in the long term.
 146 #define USE_LDS_SYMBOLS (true)
 147
 148 struct nir_shader;
 149 struct si_shader;
 150 struct si_context;
 151
 152 #define SI_MAX_ATTRIBS    16
 153 #define SI_MAX_VS_OUTPUTS 40
 154
 155 /* Shader IO unique indices are supported for VARYING_SLOT_VARn with an
 156  * index smaller than this.
 157  */
 158 #define SI_MAX_IO_GENERIC 32
 159
 160 #define SI_NGG_PRIM_EDGE_FLAG_BITS ((1 << 9) | (1 << 19) | (1 << 29))
 161
 162 /* SGPR user data indices */
 163 enum
 164 {
 165    SI_SGPR_RW_BUFFERS, /* rings (& stream-out, VS only) */
 166    SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES,
 167    SI_SGPR_CONST_AND_SHADER_BUFFERS, /* or just a constant buffer 0 pointer */
 168    SI_SGPR_SAMPLERS_AND_IMAGES,
 169    SI_NUM_RESOURCE_SGPRS,
 170
 171    /* API VS, TES without GS, GS copy shader */
 172    SI_SGPR_VS_STATE_BITS = SI_NUM_RESOURCE_SGPRS,
 173    SI_NUM_VS_STATE_RESOURCE_SGPRS,
 174
 175    /* all VS variants */
 176    SI_SGPR_BASE_VERTEX = SI_NUM_VS_STATE_RESOURCE_SGPRS,
 177    SI_SGPR_START_INSTANCE,
 178    SI_SGPR_DRAWID,
 179    SI_VS_NUM_USER_SGPR,
 180
 181    SI_SGPR_VS_BLIT_DATA = SI_SGPR_CONST_AND_SHADER_BUFFERS,
 182
 183    /* TES */
 184    SI_SGPR_TES_OFFCHIP_LAYOUT = SI_NUM_VS_STATE_RESOURCE_SGPRS,
 185    SI_SGPR_TES_OFFCHIP_ADDR,
 186    SI_TES_NUM_USER_SGPR,
 187
 188    /* GFX6-8: TCS only */
 189    GFX6_SGPR_TCS_OFFCHIP_LAYOUT = SI_NUM_RESOURCE_SGPRS,
 190    GFX6_SGPR_TCS_OUT_OFFSETS,
 191    GFX6_SGPR_TCS_OUT_LAYOUT,
 192    GFX6_SGPR_TCS_IN_LAYOUT,
 193    GFX6_TCS_NUM_USER_SGPR,
 194
 195    /* GFX9: Merged shaders. */
 196    /* 2ND_CONST_AND_SHADER_BUFFERS is set in USER_DATA_ADDR_LO (SGPR0). */
 197    /* 2ND_SAMPLERS_AND_IMAGES is set in USER_DATA_ADDR_HI (SGPR1). */
 198    GFX9_MERGED_NUM_USER_SGPR = SI_VS_NUM_USER_SGPR,
 199
 200    /* GFX9: Merged LS-HS (VS-TCS) only. */
 201    GFX9_SGPR_TCS_OFFCHIP_LAYOUT = GFX9_MERGED_NUM_USER_SGPR,
 202    GFX9_SGPR_TCS_OUT_OFFSETS,
 203    GFX9_SGPR_TCS_OUT_LAYOUT,
 204    GFX9_TCS_NUM_USER_SGPR,
 205
 206    /* GS limits */
 207    GFX6_GS_NUM_USER_SGPR = SI_NUM_RESOURCE_SGPRS,
 208    GFX9_VSGS_NUM_USER_SGPR = SI_VS_NUM_USER_SGPR,
 209    GFX9_TESGS_NUM_USER_SGPR = SI_TES_NUM_USER_SGPR,
 210    SI_GSCOPY_NUM_USER_SGPR = SI_NUM_VS_STATE_RESOURCE_SGPRS,
 211
 212    /* PS only */
 213    SI_SGPR_ALPHA_REF = SI_NUM_RESOURCE_SGPRS,
 214    SI_PS_NUM_USER_SGPR,
 215
 216    /* The value has to be 12, because the hw requires that descriptors
 217     * are aligned to 4 SGPRs.
 218     */
 219    SI_SGPR_VS_VB_DESCRIPTOR_FIRST = 12,
 220 };
 221
 222 /* LLVM function parameter indices */
 223 enum
 224 {
 225    SI_NUM_RESOURCE_PARAMS = 4,
 226
 227    /* PS only parameters */
 228    SI_PARAM_ALPHA_REF = SI_NUM_RESOURCE_PARAMS,
 229    SI_PARAM_PRIM_MASK,
 230    SI_PARAM_PERSP_SAMPLE,
 231    SI_PARAM_PERSP_CENTER,
 232    SI_PARAM_PERSP_CENTROID,
 233    SI_PARAM_PERSP_PULL_MODEL,
 234    SI_PARAM_LINEAR_SAMPLE,
 235    SI_PARAM_LINEAR_CENTER,
 236    SI_PARAM_LINEAR_CENTROID,
 237    SI_PARAM_LINE_STIPPLE_TEX,
 238    SI_PARAM_POS_X_FLOAT,
 239    SI_PARAM_POS_Y_FLOAT,
 240    SI_PARAM_POS_Z_FLOAT,
 241    SI_PARAM_POS_W_FLOAT,
 242    SI_PARAM_FRONT_FACE,
 243    SI_PARAM_ANCILLARY,
 244    SI_PARAM_SAMPLE_COVERAGE,
 245    SI_PARAM_POS_FIXED_PT,
 246
 247    SI_NUM_PARAMS = SI_PARAM_POS_FIXED_PT + 9, /* +8 for COLOR[0..1] */
 248 };
 249
 250 /* Fields of driver-defined VS state SGPR. */
 251 #define S_VS_STATE_CLAMP_VERTEX_COLOR(x)      (((unsigned)(x)&0x1) << 0)
 252 #define C_VS_STATE_CLAMP_VERTEX_COLOR         0xFFFFFFFE
 253 #define S_VS_STATE_INDEXED(x)                 (((unsigned)(x)&0x1) << 1)
 254 #define C_VS_STATE_INDEXED                    0xFFFFFFFD
 255 #define S_VS_STATE_OUTPRIM(x)                 (((unsigned)(x)&0x3) << 2)
 256 #define C_VS_STATE_OUTPRIM                    0xFFFFFFF3
 257 #define S_VS_STATE_PROVOKING_VTX_INDEX(x)     (((unsigned)(x)&0x3) << 4)
 258 #define C_VS_STATE_PROVOKING_VTX_INDEX        0xFFFFFFCF
 259 #define S_VS_STATE_STREAMOUT_QUERY_ENABLED(x) (((unsigned)(x)&0x1) << 6)
 260 #define C_VS_STATE_STREAMOUT_QUERY_ENABLED    0xFFFFFFBF
 261 #define S_VS_STATE_SMALL_PRIM_PRECISION(x)    (((unsigned)(x)&0xF) << 7)
 262 #define C_VS_STATE_SMALL_PRIM_PRECISION       0xFFFFF87F
 263 #define S_VS_STATE_LS_OUT_PATCH_SIZE(x)       (((unsigned)(x)&0x1FFF) << 11)
 264 #define C_VS_STATE_LS_OUT_PATCH_SIZE          0xFF0007FF
 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 enum
 269 {
 270    /* Use a property enum that CS wouldn't use. */
 271    TGSI_PROPERTY_CS_LOCAL_SIZE = TGSI_PROPERTY_FS_COORD_ORIGIN,
 272
 273    /* These represent the number of SGPRs the shader uses. */
 274    SI_VS_BLIT_SGPRS_POS = 3,
 275    SI_VS_BLIT_SGPRS_POS_COLOR = 7,
 276    SI_VS_BLIT_SGPRS_POS_TEXCOORD = 9,
 277 };
 278
 279 #define SI_NGG_CULL_VIEW_SMALLPRIMS          (1 << 0)   /* view.xy + small prims */
 280 #define SI_NGG_CULL_BACK_FACE                (1 << 1)   /* back faces */
 281 #define SI_NGG_CULL_FRONT_FACE               (1 << 2)   /* front faces */
 282 #define SI_NGG_CULL_GS_FAST_LAUNCH_TRI_LIST  (1 << 3)   /* GS fast launch: triangles */
 283 #define SI_NGG_CULL_GS_FAST_LAUNCH_TRI_STRIP (1 << 4)   /* GS fast launch: triangle strip */
 284 #define SI_NGG_CULL_GS_FAST_LAUNCH_ALL       (0x3 << 3) /* GS fast launch (both prim types) */
 285
 286 /**
 287  * For VS shader keys, describe any fixups required for vertex fetch.
 288  *
 289  * \ref log_size, \ref format, and the number of channels are interpreted as
 290  * by \ref ac_build_opencoded_load_format.
 291  *
 292  * Note: all bits 0 (size = 1 byte, num channels = 1, format = float) is an
 293  * impossible format and indicates that no fixup is needed (just use
 294  * buffer_load_format_xyzw).
 295  */
 296 union si_vs_fix_fetch {
 297    struct {
 298       uint8_t log_size : 2;        /* 1, 2, 4, 8 or bytes per channel */
 299       uint8_t num_channels_m1 : 2; /* number of channels minus 1 */
 300       uint8_t format : 3;          /* AC_FETCH_FORMAT_xxx */
 301       uint8_t reverse : 1;         /* reverse XYZ channels */
 302    } u;
 303    uint8_t bits;
 304 };
 305
 306 struct si_shader;
 307
 308 /* State of the context creating the shader object. */
 309 struct si_compiler_ctx_state {
 310    /* Should only be used by si_init_shader_selector_async and
 311     * si_build_shader_variant if thread_index == -1 (non-threaded). */
 312    struct ac_llvm_compiler *compiler;
 313
 314    /* Used if thread_index == -1 or if debug.async is true. */
 315    struct pipe_debug_callback debug;
 316
 317    /* Used for creating the log string for gallium/ddebug. */
 318    bool is_debug_context;
 319 };
 320
 321 struct si_shader_info {
 322    gl_shader_stage stage;
 323
 324    ubyte num_inputs;
 325    ubyte num_outputs;
 326    ubyte input_semantic[PIPE_MAX_SHADER_INPUTS];
 327    ubyte input_interpolate[PIPE_MAX_SHADER_INPUTS];
 328    ubyte input_usage_mask[PIPE_MAX_SHADER_INPUTS];
 329    ubyte output_semantic[PIPE_MAX_SHADER_OUTPUTS];
 330    char output_semantic_to_slot[VARYING_SLOT_TESS_MAX];
 331    ubyte output_usagemask[PIPE_MAX_SHADER_OUTPUTS];
 332    ubyte output_readmask[PIPE_MAX_SHADER_OUTPUTS];
 333    ubyte output_streams[PIPE_MAX_SHADER_OUTPUTS];
 334
 335    ubyte color_interpolate[2];
 336    ubyte color_interpolate_loc[2];
 337
 338    int constbuf0_num_slots;
 339    unsigned const_buffers_declared; /**< bitmask of declared const buffers */
 340    unsigned samplers_declared;      /**< bitmask of declared samplers */
 341    ubyte num_stream_output_components[4];
 342
 343    uint num_memory_instructions; /**< sampler, buffer, and image instructions */
 344
 345    ubyte colors_read; /**< which color components are read by the FS */
 346    ubyte colors_written;
 347    bool reads_samplemask;   /**< does fragment shader read sample mask? */
 348    bool reads_tess_factors; /**< If TES reads TESSINNER or TESSOUTER */
 349    bool writes_z;           /**< does fragment shader write Z value? */
 350    bool writes_stencil;     /**< does fragment shader write stencil value? */
 351    bool writes_samplemask;  /**< does fragment shader write sample mask? */
 352    bool writes_edgeflag;    /**< vertex shader outputs edgeflag */
 353    bool uses_kill;          /**< KILL or KILL_IF instruction used? */
 354    bool uses_persp_center;
 355    bool uses_persp_centroid;
 356    bool uses_persp_sample;
 357    bool uses_linear_center;
 358    bool uses_linear_centroid;
 359    bool uses_linear_sample;
 360    bool uses_persp_opcode_interp_sample;
 361    bool uses_linear_opcode_interp_sample;
 362    bool uses_instanceid;
 363    bool uses_vertexid;
 364    bool uses_vertexid_nobase;
 365    bool uses_basevertex;
 366    bool uses_drawid;
 367    bool uses_primid;
 368    bool uses_frontface;
 369    bool uses_invocationid;
 370    bool uses_thread_id[3];
 371    bool uses_block_id[3];
 372    bool uses_block_size;
 373    bool uses_grid_size;
 374    bool uses_subgroup_info;
 375    bool writes_position;
 376    bool writes_psize;
 377    bool writes_clipvertex;
 378    bool writes_primid;
 379    bool writes_viewport_index;
 380    bool writes_layer;
 381    bool writes_memory; /**< contains stores or atomics to buffers or images */
 382    bool uses_derivatives;
 383    bool uses_bindless_samplers;
 384    bool uses_bindless_images;
 385    bool uses_fbfetch;
 386    unsigned clipdist_writemask;
 387    unsigned culldist_writemask;
 388    unsigned num_written_culldistance;
 389    unsigned num_written_clipdistance;
 390
 391    unsigned images_declared;         /**< bitmask of declared images */
 392    unsigned image_buffers;           /**< bitmask of images that are buffers */
 393    unsigned msaa_images_declared;    /**< bitmask of declared MSAA images */
 394    unsigned shader_buffers_declared; /**< bitmask of declared shader buffers */
 395
 396    unsigned properties[TGSI_PROPERTY_COUNT]; /* index with TGSI_PROPERTY_ */
 397
 398    /** Whether all codepaths write tess factors in all invocations. */
 399    bool tessfactors_are_def_in_all_invocs;
 400 };
 401
 402 /* A shader selector is a gallium CSO and contains shader variants and
 403  * binaries for one NIR program. This can be shared by multiple contexts.
 404  */
 405 struct si_shader_selector {
 406    struct util_live_shader base;
 407    struct si_screen *screen;
 408    struct util_queue_fence ready;
 409    struct si_compiler_ctx_state compiler_ctx_state;
 410
 411    simple_mtx_t mutex;
 412    struct si_shader *first_variant; /* immutable after the first variant */
 413    struct si_shader *last_variant;  /* mutable */
 414
 415    /* The compiled NIR shader without a prolog and/or epilog (not
 416     * uploaded to a buffer object).
 417     */
 418    struct si_shader *main_shader_part;
 419    struct si_shader *main_shader_part_ls;     /* as_ls is set in the key */
 420    struct si_shader *main_shader_part_es;     /* as_es is set in the key */
 421    struct si_shader *main_shader_part_ngg;    /* as_ngg is set in the key */
 422    struct si_shader *main_shader_part_ngg_es; /* for Wave32 TES before legacy GS */
 423
 424    struct si_shader *gs_copy_shader;
 425
 426    struct nir_shader *nir;
 427    void *nir_binary;
 428    unsigned nir_size;
 429
 430    struct pipe_stream_output_info so;
 431    struct si_shader_info info;
 432
 433    ubyte const_and_shader_buf_descriptors_index;
 434    ubyte sampler_and_images_descriptors_index;
 435    bool vs_needs_prolog;
 436    bool prim_discard_cs_allowed;
 437    bool ngg_culling_allowed;
 438    ubyte cs_shaderbufs_sgpr_index;
 439    ubyte cs_num_shaderbufs_in_user_sgprs;
 440    ubyte cs_images_sgpr_index;
 441    ubyte cs_images_num_sgprs;
 442    ubyte cs_num_images_in_user_sgprs;
 443    unsigned num_vs_inputs;
 444    unsigned num_vbos_in_user_sgprs;
 445    unsigned pa_cl_vs_out_cntl;
 446    ubyte clipdist_mask;
 447    ubyte culldist_mask;
 448    unsigned rast_prim;
 449
 450    /* ES parameters. */
 451    unsigned esgs_itemsize; /* vertex stride */
 452    unsigned lshs_vertex_stride;
 453
 454    /* GS parameters. */
 455    unsigned gs_input_verts_per_prim;
 456    unsigned gs_output_prim;
 457    unsigned gs_max_out_vertices;
 458    unsigned gs_num_invocations;
 459    unsigned max_gs_stream; /* count - 1 */
 460    unsigned gsvs_vertex_size;
 461    unsigned max_gsvs_emit_size;
 462    unsigned enabled_streamout_buffer_mask;
 463    bool tess_turns_off_ngg;
 464
 465    /* PS parameters. */
 466    unsigned color_attr_index[2];
 467    unsigned db_shader_control;
 468    /* Set 0xf or 0x0 (4 bits) per each written output.
 469     * ANDed with spi_shader_col_format.
 470     */
 471    unsigned colors_written_4bit;
 472
 473    uint64_t outputs_written_before_ps; /* "get_unique_index" bits */
 474    uint64_t outputs_written;           /* "get_unique_index" bits */
 475    uint32_t patch_outputs_written;     /* "get_unique_index_patch" bits */
 476
 477    uint64_t inputs_read; /* "get_unique_index" bits */
 478
 479    /* bitmasks of used descriptor slots */
 480    uint64_t active_const_and_shader_buffers;
 481    uint64_t active_samplers_and_images;
 482 };
 483
 484 /* Valid shader configurations:
 485  *
 486  * API shaders           VS | TCS | TES | GS |pass| PS
 487  * are compiled as:         |     |     |    |thru|
 488  *                          |     |     |    |    |
 489  * Only VS & PS:         VS |     |     |    |    | PS
 490  * GFX6     - with GS:   ES |     |     | GS | VS | PS
 491  *          - with tess: LS | HS  | VS  |    |    | PS
 492  *          - with both: LS | HS  | ES  | GS | VS | PS
 493  * GFX9     - with GS:   -> |     |     | GS | VS | PS
 494  *          - with tess: -> | HS  | VS  |    |    | PS
 495  *          - with both: -> | HS  | ->  | GS | VS | PS
 496  *                          |     |     |    |    |
 497  * NGG      - VS & PS:   GS |     |     |    |    | PS
 498  * (GFX10+) - with GS:   -> |     |     | GS |    | PS
 499  *          - with tess: -> | HS  | GS  |    |    | PS
 500  *          - with both: -> | HS  | ->  | GS |    | PS
 501  *
 502  * -> = merged with the next stage
 503  */
 504
 505 /* Use the byte alignment for all following structure members for optimal
 506  * shader key memory footprint.
 507  */
 508 #pragma pack(push, 1)
 509
 510 /* Common VS bits between the shader key and the prolog key. */
 511 struct si_vs_prolog_bits {
 512    /* - If neither "is_one" nor "is_fetched" has a bit set, the instance
 513     *   divisor is 0.
 514     * - If "is_one" has a bit set, the instance divisor is 1.
 515     * - If "is_fetched" has a bit set, the instance divisor will be loaded
 516     *   from the constant buffer.
 517     */
 518    uint16_t instance_divisor_is_one;     /* bitmask of inputs */
 519    uint16_t instance_divisor_is_fetched; /* bitmask of inputs */
 520    unsigned ls_vgpr_fix : 1;
 521    unsigned unpack_instance_id_from_vertex_id : 1;
 522 };
 523
 524 /* Common TCS bits between the shader key and the epilog key. */
 525 struct si_tcs_epilog_bits {
 526    unsigned prim_mode : 3;
 527    unsigned invoc0_tess_factors_are_def : 1;
 528    unsigned tes_reads_tess_factors : 1;
 529 };
 530
 531 struct si_gs_prolog_bits {
 532    unsigned tri_strip_adj_fix : 1;
 533    unsigned gfx9_prev_is_vs : 1;
 534 };
 535
 536 /* Common PS bits between the shader key and the prolog key. */
 537 struct si_ps_prolog_bits {
 538    unsigned color_two_side : 1;
 539    unsigned flatshade_colors : 1;
 540    unsigned poly_stipple : 1;
 541    unsigned force_persp_sample_interp : 1;
 542    unsigned force_linear_sample_interp : 1;
 543    unsigned force_persp_center_interp : 1;
 544    unsigned force_linear_center_interp : 1;
 545    unsigned bc_optimize_for_persp : 1;
 546    unsigned bc_optimize_for_linear : 1;
 547    unsigned samplemask_log_ps_iter : 3;
 548 };
 549
 550 /* Common PS bits between the shader key and the epilog key. */
 551 struct si_ps_epilog_bits {
 552    unsigned spi_shader_col_format;
 553    unsigned color_is_int8 : 8;
 554    unsigned color_is_int10 : 8;
 555    unsigned last_cbuf : 3;
 556    unsigned alpha_func : 3;
 557    unsigned alpha_to_one : 1;
 558    unsigned poly_line_smoothing : 1;
 559    unsigned clamp_color : 1;
 560 };
 561
 562 union si_shader_part_key {
 563    struct {
 564       struct si_vs_prolog_bits states;
 565       unsigned num_input_sgprs : 6;
 566       /* For merged stages such as LS-HS, HS input VGPRs are first. */
 567       unsigned num_merged_next_stage_vgprs : 3;
 568       unsigned num_inputs : 5;
 569       unsigned as_ls : 1;
 570       unsigned as_es : 1;
 571       unsigned as_ngg : 1;
 572       unsigned as_prim_discard_cs : 1;
 573       unsigned has_ngg_cull_inputs : 1;      /* from the NGG cull shader */
 574       unsigned gs_fast_launch_tri_list : 1;  /* for NGG culling */
 575       unsigned gs_fast_launch_tri_strip : 1; /* for NGG culling */
 576       /* Prologs for monolithic shaders shouldn't set EXEC. */
 577       unsigned is_monolithic : 1;
 578    } vs_prolog;
 579    struct {
 580       struct si_tcs_epilog_bits states;
 581    } tcs_epilog;
 582    struct {
 583       struct si_gs_prolog_bits states;
 584       /* Prologs of monolithic shaders shouldn't set EXEC. */
 585       unsigned is_monolithic : 1;
 586       unsigned as_ngg : 1;
 587    } gs_prolog;
 588    struct {
 589       struct si_ps_prolog_bits states;
 590       unsigned num_input_sgprs : 6;
 591       unsigned num_input_vgprs : 5;
 592       /* Color interpolation and two-side color selection. */
 593       unsigned colors_read : 8;       /* color input components read */
 594       unsigned num_interp_inputs : 5; /* BCOLOR is at this location */
 595       unsigned face_vgpr_index : 5;
 596       unsigned ancillary_vgpr_index : 5;
 597       unsigned wqm : 1;
 598       char color_attr_index[2];
 599       signed char color_interp_vgpr_index[2]; /* -1 == constant */
 600    } ps_prolog;
 601    struct {
 602       struct si_ps_epilog_bits states;
 603       unsigned colors_written : 8;
 604       unsigned writes_z : 1;
 605       unsigned writes_stencil : 1;
 606       unsigned writes_samplemask : 1;
 607    } ps_epilog;
 608 };
 609
 610 struct si_shader_key {
 611    /* Prolog and epilog flags. */
 612    union {
 613       struct {
 614          struct si_vs_prolog_bits prolog;
 615       } vs;
 616       struct {
 617          struct si_vs_prolog_bits ls_prolog; /* for merged LS-HS */
 618          struct si_shader_selector *ls;      /* for merged LS-HS */
 619          struct si_tcs_epilog_bits epilog;
 620       } tcs; /* tessellation control shader */
 621       struct {
 622          struct si_vs_prolog_bits vs_prolog; /* for merged ES-GS */
 623          struct si_shader_selector *es;      /* for merged ES-GS */
 624          struct si_gs_prolog_bits prolog;
 625       } gs;
 626       struct {
 627          struct si_ps_prolog_bits prolog;
 628          struct si_ps_epilog_bits epilog;
 629       } ps;
 630    } part;
 631
 632    /* These three are initially set according to the NEXT_SHADER property,
 633     * or guessed if the property doesn't seem correct.
 634     */
 635    unsigned as_es : 1;  /* export shader, which precedes GS */
 636    unsigned as_ls : 1;  /* local shader, which precedes TCS */
 637    unsigned as_ngg : 1; /* VS, TES, or GS compiled as NGG primitive shader */
 638
 639    /* Flags for monolithic compilation only. */
 640    struct {
 641       /* Whether fetch should be opencoded according to vs_fix_fetch.
 642        * Otherwise, if vs_fix_fetch is non-zero, buffer_load_format_xyzw
 643        * with minimal fixups is used. */
 644       uint16_t vs_fetch_opencode;
 645       union si_vs_fix_fetch vs_fix_fetch[SI_MAX_ATTRIBS];
 646
 647       union {
 648          uint64_t ff_tcs_inputs_to_copy; /* for fixed-func TCS */
 649          /* When PS needs PrimID and GS is disabled. */
 650          unsigned vs_export_prim_id : 1;
 651          struct {
 652             unsigned interpolate_at_sample_force_center : 1;
 653             unsigned fbfetch_msaa : 1;
 654             unsigned fbfetch_is_1D : 1;
 655             unsigned fbfetch_layered : 1;
 656          } ps;
 657       } u;
 658    } mono;
 659
 660    /* Optimization flags for asynchronous compilation only. */
 661    struct {
 662       /* For HW VS (it can be VS, TES, GS) */
 663       uint64_t kill_outputs; /* "get_unique_index" bits */
 664       unsigned clip_disable : 1;
 665
 666       /* For NGG VS and TES. */
 667       unsigned ngg_culling : 5; /* SI_NGG_CULL_* */
 668
 669       /* For shaders where monolithic variants have better code.
 670        *
 671        * This is a flag that has no effect on code generation,
 672        * but forces monolithic shaders to be used as soon as
 673        * possible, because it's in the "opt" group.
 674        */
 675       unsigned prefer_mono : 1;
 676
 677       /* Primitive discard compute shader. */
 678       unsigned vs_as_prim_discard_cs : 1;
 679       unsigned cs_prim_type : 4;
 680       unsigned cs_indexed : 1;
 681       unsigned cs_instancing : 1;
 682       unsigned cs_primitive_restart : 1;
 683       unsigned cs_provoking_vertex_first : 1;
 684       unsigned cs_need_correct_orientation : 1;
 685       unsigned cs_cull_front : 1;
 686       unsigned cs_cull_back : 1;
 687       unsigned cs_cull_z : 1;
 688       unsigned cs_halfz_clip_space : 1;
 689    } opt;
 690 };
 691
 692 /* Restore the pack alignment to default. */
 693 #pragma pack(pop)
 694
 695 /* GCN-specific shader info. */
 696 struct si_shader_binary_info {
 697    ubyte vs_output_param_offset[SI_MAX_VS_OUTPUTS];
 698    ubyte num_input_sgprs;
 699    ubyte num_input_vgprs;
 700    signed char face_vgpr_index;
 701    signed char ancillary_vgpr_index;
 702    bool uses_instanceid;
 703    ubyte nr_pos_exports;
 704    ubyte nr_param_exports;
 705    unsigned private_mem_vgprs;
 706    unsigned max_simd_waves;
 707 };
 708
 709 struct si_shader_binary {
 710    const char *elf_buffer;
 711    size_t elf_size;
 712
 713    char *llvm_ir_string;
 714 };
 715
 716 struct gfx9_gs_info {
 717    unsigned es_verts_per_subgroup;
 718    unsigned gs_prims_per_subgroup;
 719    unsigned gs_inst_prims_in_subgroup;
 720    unsigned max_prims_per_subgroup;
 721    unsigned esgs_ring_size; /* in bytes */
 722 };
 723
 724 struct si_shader {
 725    struct si_compiler_ctx_state compiler_ctx_state;
 726
 727    struct si_shader_selector *selector;
 728    struct si_shader_selector *previous_stage_sel; /* for refcounting */
 729    struct si_shader *next_variant;
 730
 731    struct si_shader_part *prolog;
 732    struct si_shader *previous_stage; /* for GFX9 */
 733    struct si_shader_part *prolog2;
 734    struct si_shader_part *epilog;
 735
 736    struct si_pm4_state *pm4;
 737    struct si_resource *bo;
 738    struct si_resource *scratch_bo;
 739    struct si_shader_key key;
 740    struct util_queue_fence ready;
 741    bool compilation_failed;
 742    bool is_monolithic;
 743    bool is_optimized;
 744    bool is_binary_shared;
 745    bool is_gs_copy_shader;
 746
 747    /* The following data is all that's needed for binary shaders. */
 748    struct si_shader_binary binary;
 749    struct ac_shader_config config;
 750    struct si_shader_binary_info info;
 751
 752    struct {
 753       uint16_t ngg_emit_size; /* in dwords */
 754       uint16_t hw_max_esverts;
 755       uint16_t max_gsprims;
 756       uint16_t max_out_verts;
 757       uint16_t prim_amp_factor;
 758       bool max_vert_out_per_gs_instance;
 759    } ngg;
 760
 761    /* Shader key + LLVM IR + disassembly + statistics.
 762     * Generated for debug contexts only.
 763     */
 764    char *shader_log;
 765    size_t shader_log_size;
 766
 767    struct gfx9_gs_info gs_info;
 768
 769    /* For save precompute context registers values. */
 770    union {
 771       struct {
 772          unsigned vgt_gsvs_ring_offset_1;
 773          unsigned vgt_gsvs_ring_offset_2;
 774          unsigned vgt_gsvs_ring_offset_3;
 775          unsigned vgt_gsvs_ring_itemsize;
 776          unsigned vgt_gs_max_vert_out;
 777          unsigned vgt_gs_vert_itemsize;
 778          unsigned vgt_gs_vert_itemsize_1;
 779          unsigned vgt_gs_vert_itemsize_2;
 780          unsigned vgt_gs_vert_itemsize_3;
 781          unsigned vgt_gs_instance_cnt;
 782          unsigned vgt_gs_onchip_cntl;
 783          unsigned vgt_gs_max_prims_per_subgroup;
 784          unsigned vgt_esgs_ring_itemsize;
 785       } gs;
 786
 787       struct {
 788          unsigned ge_max_output_per_subgroup;
 789          unsigned ge_ngg_subgrp_cntl;
 790          unsigned vgt_primitiveid_en;
 791          unsigned vgt_gs_onchip_cntl;
 792          unsigned vgt_gs_instance_cnt;
 793          unsigned vgt_esgs_ring_itemsize;
 794          unsigned spi_vs_out_config;
 795          unsigned spi_shader_idx_format;
 796          unsigned spi_shader_pos_format;
 797          unsigned pa_cl_vte_cntl;
 798          unsigned pa_cl_ngg_cntl;
 799          unsigned vgt_gs_max_vert_out; /* for API GS */
 800          unsigned ge_pc_alloc;         /* uconfig register */
 801       } ngg;
 802
 803       struct {
 804          unsigned vgt_gs_mode;
 805          unsigned vgt_primitiveid_en;
 806          unsigned vgt_reuse_off;
 807          unsigned spi_vs_out_config;
 808          unsigned spi_shader_pos_format;
 809          unsigned pa_cl_vte_cntl;
 810          unsigned ge_pc_alloc; /* uconfig register */
 811       } vs;
 812
 813       struct {
 814          unsigned spi_ps_input_ena;
 815          unsigned spi_ps_input_addr;
 816          unsigned spi_baryc_cntl;
 817          unsigned spi_ps_in_control;
 818          unsigned spi_shader_z_format;
 819          unsigned spi_shader_col_format;
 820          unsigned cb_shader_mask;
 821       } ps;
 822    } ctx_reg;
 823
 824    /*For save precompute registers value */
 825    unsigned vgt_tf_param;                /* VGT_TF_PARAM */
 826    unsigned vgt_vertex_reuse_block_cntl; /* VGT_VERTEX_REUSE_BLOCK_CNTL */
 827    unsigned pa_cl_vs_out_cntl;
 828    unsigned ge_cntl;
 829 };
 830
 831 struct si_shader_part {
 832    struct si_shader_part *next;
 833    union si_shader_part_key key;
 834    struct si_shader_binary binary;
 835    struct ac_shader_config config;
 836 };
 837
 838 /* si_shader.c */
 839 bool si_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
 840                        struct si_shader *shader, struct pipe_debug_callback *debug);
 841 bool si_create_shader_variant(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
 842                               struct si_shader *shader, struct pipe_debug_callback *debug);
 843 void si_shader_destroy(struct si_shader *shader);
 844 unsigned si_shader_io_get_unique_index_patch(unsigned semantic);
 845 unsigned si_shader_io_get_unique_index(unsigned semantic, bool is_varying);
 846 bool si_shader_binary_upload(struct si_screen *sscreen, struct si_shader *shader,
 847                              uint64_t scratch_va);
 848 void si_shader_dump(struct si_screen *sscreen, struct si_shader *shader,
 849                     struct pipe_debug_callback *debug, FILE *f, bool check_debug_option);
 850 void si_shader_dump_stats_for_shader_db(struct si_screen *screen, struct si_shader *shader,
 851                                         struct pipe_debug_callback *debug);
 852 void si_multiwave_lds_size_workaround(struct si_screen *sscreen, unsigned *lds_size);
 853 const char *si_get_shader_name(const struct si_shader *shader);
 854 void si_shader_binary_clean(struct si_shader_binary *binary);
 855
 856 /* si_shader_llvm_gs.c */
 857 struct si_shader *si_generate_gs_copy_shader(struct si_screen *sscreen,
 858                                              struct ac_llvm_compiler *compiler,
 859                                              struct si_shader_selector *gs_selector,
 860                                              struct pipe_debug_callback *debug);
 861
 862 /* si_shader_nir.c */
 863 void si_nir_scan_shader(const struct nir_shader *nir, struct si_shader_info *info);
 864 void si_finalize_nir(struct pipe_screen *screen, void *nirptr, bool optimize);
 865
 866 /* si_state_shaders.c */
 867 void gfx9_get_gs_info(struct si_shader_selector *es, struct si_shader_selector *gs,
 868                       struct gfx9_gs_info *out);
 869
 870 /* Inline helpers. */
 871
 872 /* Return the pointer to the main shader part's pointer. */
 873 static inline struct si_shader **si_get_main_shader_part(struct si_shader_selector *sel,
 874                                                          struct si_shader_key *key)
 875 {
 876    if (key->as_ls)
 877       return &sel->main_shader_part_ls;
 878    if (key->as_es && key->as_ngg)
 879       return &sel->main_shader_part_ngg_es;
 880    if (key->as_es)
 881       return &sel->main_shader_part_es;
 882    if (key->as_ngg)
 883       return &sel->main_shader_part_ngg;
 884    return &sel->main_shader_part;
 885 }
 886
 887 static inline bool gfx10_is_ngg_passthrough(struct si_shader *shader)
 888 {
 889    struct si_shader_selector *sel = shader->selector;
 890
 891    return sel->info.stage != MESA_SHADER_GEOMETRY && !sel->so.num_outputs && !sel->info.writes_edgeflag &&
 892           !shader->key.opt.ngg_culling &&
 893           (sel->info.stage != MESA_SHADER_VERTEX || !shader->key.mono.u.vs_export_prim_id);
 894 }
 895
 896 static inline bool si_shader_uses_bindless_samplers(struct si_shader_selector *selector)
 897 {
 898    return selector ? selector->info.uses_bindless_samplers : false;
 899 }
 900
 901 static inline bool si_shader_uses_bindless_images(struct si_shader_selector *selector)
 902 {
 903    return selector ? selector->info.uses_bindless_images : false;
 904 }
 905
 906 #endif