src/compiler/nir/nir_intrinsics.py

   1 #
   2 # Copyright (C) 2018 Red Hat
   3 # Copyright (C) 2014 Intel Corporation
   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 # the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9 # and/or sell copies of the Software, and to permit persons to whom the
  10 # 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 NONINFRINGEMENT.  IN NO EVENT SHALL
  19 # THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20 # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  21 # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  22 # IN THE SOFTWARE.
  23 #
  24
  25 # This file defines all the available intrinsics in one place.
  26 #
  27 # The Intrinsic class corresponds one-to-one with nir_intrinsic_info
  28 # structure.
  29
  30 class Intrinsic(object):
  31    """Class that represents all the information about an intrinsic opcode.
  32    NOTE: this must be kept in sync with nir_intrinsic_info.
  33    """
  34    def __init__(self, name, src_components, dest_components,
  35                 indices, flags, sysval):
  36        """Parameters:
  37
  38        - name: the intrinsic name
  39        - src_components: list of the number of components per src, 0 means
  40          vectorized instruction with number of components given in the
  41          num_components field in nir_intrinsic_instr.
  42        - dest_components: number of destination components, -1 means no
  43          dest, 0 means number of components given in num_components field
  44          in nir_intrinsic_instr.
  45        - indices: list of constant indicies
  46        - flags: list of semantic flags
  47        - sysval: is this a system-value intrinsic
  48        """
  49        assert isinstance(name, str)
  50        assert isinstance(src_components, list)
  51        if src_components:
  52            assert isinstance(src_components[0], int)
  53        assert isinstance(dest_components, int)
  54        assert isinstance(indices, list)
  55        if indices:
  56            assert isinstance(indices[0], str)
  57        assert isinstance(flags, list)
  58        if flags:
  59            assert isinstance(flags[0], str)
  60        assert isinstance(sysval, bool)
  61
  62        self.name = name
  63        self.num_srcs = len(src_components)
  64        self.src_components = src_components
  65        self.has_dest = (dest_components >= 0)
  66        self.dest_components = dest_components
  67        self.num_indices = len(indices)
  68        self.indices = indices
  69        self.flags = flags
  70        self.sysval = sysval
  71
  72 #
  73 # Possible indices:
  74 #
  75
  76 # A constant 'base' value that is added to an offset src:
  77 BASE = "NIR_INTRINSIC_BASE"
  78 # For store instructions, a writemask:
  79 WRMASK = "NIR_INTRINSIC_WRMASK"
  80 # The stream-id for GS emit_vertex/end_primitive intrinsics:
  81 STREAM_ID = "NIR_INTRINSIC_STREAM_ID"
  82 # The clip-plane id for load_user_clip_plane intrinsics:
  83 UCP_ID = "NIR_INTRINSIC_UCP_ID"
  84 # The amount of data, starting from BASE, that this instruction
  85 # may access.  This is used to provide bounds if the offset is
  86 # not constant.
  87 RANGE = "NIR_INTRINSIC_RANGE"
  88 # The vulkan descriptor set binding for vulkan_resource_index
  89 # intrinsic
  90 DESC_SET = "NIR_INTRINSIC_DESC_SET"
  91 # The vulkan descriptor set binding for vulkan_resource_index
  92 # intrinsic
  93 BINDING = "NIR_INTRINSIC_BINDING"
  94 # Component offset
  95 COMPONENT = "NIR_INTRINSIC_COMPONENT"
  96 # Interpolation mode (only meaningful for FS inputs)
  97 INTERP_MODE = "NIR_INTRINSIC_INTERP_MODE"
  98 # A binary nir_op to use when performing a reduction or scan operation
  99 REDUCTION_OP = "NIR_INTRINSIC_REDUCTION_OP"
 100 # Cluster size for reduction operations
 101 CLUSTER_SIZE = "NIR_INTRINSIC_CLUSTER_SIZE"
 102 # Parameter index for a load_param intrinsic
 103 PARAM_IDX = "NIR_INTRINSIC_PARAM_IDX"
 104
 105 #
 106 # Possible flags:
 107 #
 108
 109 CAN_ELIMINATE = "NIR_INTRINSIC_CAN_ELIMINATE"
 110 CAN_REORDER   = "NIR_INTRINSIC_CAN_REORDER"
 111
 112 INTR_OPCODES = {}
 113
 114 def intrinsic(name, src_comp=[], dest_comp=-1, indices=[],
 115               flags=[], sysval=False):
 116     assert name not in INTR_OPCODES
 117     INTR_OPCODES[name] = Intrinsic(name, src_comp, dest_comp,
 118                                    indices, flags, sysval)
 119
 120 intrinsic("nop", flags=[CAN_ELIMINATE])
 121
 122 intrinsic("load_param", dest_comp=0, indices=[PARAM_IDX], flags=[CAN_ELIMINATE])
 123
 124 intrinsic("load_deref", dest_comp=0, src_comp=[1], flags=[CAN_ELIMINATE])
 125 intrinsic("store_deref", src_comp=[1, 0], indices=[WRMASK])
 126 intrinsic("copy_deref", src_comp=[1, 1])
 127
 128 # Interpolation of input.  The interp_deref_at* intrinsics are similar to the
 129 # load_var intrinsic acting on a shader input except that they interpolate the
 130 # input differently.  The at_sample and at_offset intrinsics take an
 131 # additional source that is an integer sample id or a vec2 position offset
 132 # respectively.
 133
 134 intrinsic("interp_deref_at_centroid", dest_comp=0, src_comp=[1],
 135           flags=[ CAN_ELIMINATE, CAN_REORDER])
 136 intrinsic("interp_deref_at_sample", src_comp=[1, 1], dest_comp=0,
 137           flags=[CAN_ELIMINATE, CAN_REORDER])
 138 intrinsic("interp_deref_at_offset", src_comp=[1, 2], dest_comp=0,
 139           flags=[CAN_ELIMINATE, CAN_REORDER])
 140
 141 # Ask the driver for the size of a given buffer. It takes the buffer index
 142 # as source.
 143 intrinsic("get_buffer_size", src_comp=[1], dest_comp=1,
 144           flags=[CAN_ELIMINATE, CAN_REORDER])
 145
 146 # a barrier is an intrinsic with no inputs/outputs but which can't be moved
 147 # around/optimized in general
 148 def barrier(name):
 149     intrinsic(name)
 150
 151 barrier("barrier")
 152 barrier("discard")
 153
 154 # Memory barrier with semantics analogous to the memoryBarrier() GLSL
 155 # intrinsic.
 156 barrier("memory_barrier")
 157
 158 # Shader clock intrinsic with semantics analogous to the clock2x32ARB()
 159 # GLSL intrinsic.
 160 # The latter can be used as code motion barrier, which is currently not
 161 # feasible with NIR.
 162 intrinsic("shader_clock", dest_comp=2, flags=[CAN_ELIMINATE])
 163
 164 # Shader ballot intrinsics with semantics analogous to the
 165 #
 166 #    ballotARB()
 167 #    readInvocationARB()
 168 #    readFirstInvocationARB()
 169 #
 170 # GLSL functions from ARB_shader_ballot.
 171 intrinsic("ballot", src_comp=[1], dest_comp=0, flags=[CAN_ELIMINATE])
 172 intrinsic("read_invocation", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 173 intrinsic("read_first_invocation", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 174
 175 # Additional SPIR-V ballot intrinsics
 176 #
 177 # These correspond to the SPIR-V opcodes
 178 #
 179 #    OpGroupUniformElect
 180 #    OpSubgroupFirstInvocationKHR
 181 intrinsic("elect", dest_comp=1, flags=[CAN_ELIMINATE])
 182 intrinsic("first_invocation", dest_comp=1, flags=[CAN_ELIMINATE])
 183
 184 # Memory barrier with semantics analogous to the compute shader
 185 # groupMemoryBarrier(), memoryBarrierAtomicCounter(), memoryBarrierBuffer(),
 186 # memoryBarrierImage() and memoryBarrierShared() GLSL intrinsics.
 187 barrier("group_memory_barrier")
 188 barrier("memory_barrier_atomic_counter")
 189 barrier("memory_barrier_buffer")
 190 barrier("memory_barrier_image")
 191 barrier("memory_barrier_shared")
 192 barrier("begin_invocation_interlock")
 193 barrier("end_invocation_interlock")
 194
 195 # A conditional discard, with a single boolean source.
 196 intrinsic("discard_if", src_comp=[1])
 197
 198 # ARB_shader_group_vote intrinsics
 199 intrinsic("vote_any", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
 200 intrinsic("vote_all", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
 201 intrinsic("vote_feq", src_comp=[0], dest_comp=1, flags=[CAN_ELIMINATE])
 202 intrinsic("vote_ieq", src_comp=[0], dest_comp=1, flags=[CAN_ELIMINATE])
 203
 204 # Ballot ALU operations from SPIR-V.
 205 #
 206 # These operations work like their ALU counterparts except that the operate
 207 # on a uvec4 which is treated as a 128bit integer.  Also, they are, in
 208 # general, free to ignore any bits which are above the subgroup size.
 209 intrinsic("ballot_bitfield_extract", src_comp=[4, 1], dest_comp=1, flags=[CAN_ELIMINATE])
 210 intrinsic("ballot_bit_count_reduce", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 211 intrinsic("ballot_bit_count_inclusive", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 212 intrinsic("ballot_bit_count_exclusive", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 213 intrinsic("ballot_find_lsb", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 214 intrinsic("ballot_find_msb", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 215
 216 # Shuffle operations from SPIR-V.
 217 intrinsic("shuffle", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 218 intrinsic("shuffle_xor", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 219 intrinsic("shuffle_up", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 220 intrinsic("shuffle_down", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 221
 222 # Quad operations from SPIR-V.
 223 intrinsic("quad_broadcast", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 224 intrinsic("quad_swap_horizontal", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 225 intrinsic("quad_swap_vertical", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 226 intrinsic("quad_swap_diagonal", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 227
 228 intrinsic("reduce", src_comp=[0], dest_comp=0, indices=[REDUCTION_OP, CLUSTER_SIZE],
 229           flags=[CAN_ELIMINATE])
 230 intrinsic("inclusive_scan", src_comp=[0], dest_comp=0, indices=[REDUCTION_OP],
 231           flags=[CAN_ELIMINATE])
 232 intrinsic("exclusive_scan", src_comp=[0], dest_comp=0, indices=[REDUCTION_OP],
 233           flags=[CAN_ELIMINATE])
 234
 235 # Basic Geometry Shader intrinsics.
 236 #
 237 # emit_vertex implements GLSL's EmitStreamVertex() built-in.  It takes a single
 238 # index, which is the stream ID to write to.
 239 #
 240 # end_primitive implements GLSL's EndPrimitive() built-in.
 241 intrinsic("emit_vertex",   indices=[STREAM_ID])
 242 intrinsic("end_primitive", indices=[STREAM_ID])
 243
 244 # Geometry Shader intrinsics with a vertex count.
 245 #
 246 # Alternatively, drivers may implement these intrinsics, and use
 247 # nir_lower_gs_intrinsics() to convert from the basic intrinsics.
 248 #
 249 # These maintain a count of the number of vertices emitted, as an additional
 250 # unsigned integer source.
 251 intrinsic("emit_vertex_with_counter", src_comp=[1], indices=[STREAM_ID])
 252 intrinsic("end_primitive_with_counter", src_comp=[1], indices=[STREAM_ID])
 253 intrinsic("set_vertex_count", src_comp=[1])
 254
 255 # Atomic counters
 256 #
 257 # The *_var variants take an atomic_uint nir_variable, while the other,
 258 # lowered, variants take a constant buffer index and register offset.
 259
 260 def atomic(name, flags=[]):
 261     intrinsic(name + "_deref", src_comp=[1], dest_comp=1, flags=flags)
 262     intrinsic(name, src_comp=[1], dest_comp=1, indices=[BASE], flags=flags)
 263
 264 def atomic2(name):
 265     intrinsic(name + "_deref", src_comp=[1, 1], dest_comp=1)
 266     intrinsic(name, src_comp=[1, 1], dest_comp=1, indices=[BASE])
 267
 268 def atomic3(name):
 269     intrinsic(name + "_deref", src_comp=[1, 1, 1], dest_comp=1)
 270     intrinsic(name, src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 271
 272 atomic("atomic_counter_inc")
 273 atomic("atomic_counter_pre_dec")
 274 atomic("atomic_counter_post_dec")
 275 atomic("atomic_counter_read", flags=[CAN_ELIMINATE])
 276 atomic2("atomic_counter_add")
 277 atomic2("atomic_counter_min")
 278 atomic2("atomic_counter_max")
 279 atomic2("atomic_counter_and")
 280 atomic2("atomic_counter_or")
 281 atomic2("atomic_counter_xor")
 282 atomic2("atomic_counter_exchange")
 283 atomic3("atomic_counter_comp_swap")
 284
 285 # Image load, store and atomic intrinsics.
 286 #
 287 # All image intrinsics take an image target passed as a nir_variable.  The
 288 # variable is passed in using a chain of nir_deref_instr with as the first
 289 # source of the image intrinsic.  Image variables contain a number of memory
 290 # and layout qualifiers that influence the semantics of the intrinsic.
 291 #
 292 # All image intrinsics take a four-coordinate vector and a sample index as
 293 # 2nd and 3rd sources, determining the location within the image that will be
 294 # accessed by the intrinsic.  Components not applicable to the image target
 295 # in use are undefined.  Image store takes an additional four-component
 296 # argument with the value to be written, and image atomic operations take
 297 # either one or two additional scalar arguments with the same meaning as in
 298 # the ARB_shader_image_load_store specification.
 299 intrinsic("image_deref_load", src_comp=[1, 4, 1], dest_comp=4,
 300           flags=[CAN_ELIMINATE])
 301 intrinsic("image_deref_store", src_comp=[1, 4, 1, 4])
 302 intrinsic("image_deref_atomic_add",  src_comp=[1, 4, 1, 1], dest_comp=1)
 303 intrinsic("image_deref_atomic_min",  src_comp=[1, 4, 1, 1], dest_comp=1)
 304 intrinsic("image_deref_atomic_max",  src_comp=[1, 4, 1, 1], dest_comp=1)
 305 intrinsic("image_deref_atomic_and",  src_comp=[1, 4, 1, 1], dest_comp=1)
 306 intrinsic("image_deref_atomic_or",   src_comp=[1, 4, 1, 1], dest_comp=1)
 307 intrinsic("image_deref_atomic_xor",  src_comp=[1, 4, 1, 1], dest_comp=1)
 308 intrinsic("image_deref_atomic_exchange",  src_comp=[1, 4, 1, 1], dest_comp=1)
 309 intrinsic("image_deref_atomic_comp_swap", src_comp=[1, 4, 1, 1, 1], dest_comp=1)
 310 intrinsic("image_deref_size",    src_comp=[1], dest_comp=0, flags=[CAN_ELIMINATE, CAN_REORDER])
 311 intrinsic("image_deref_samples", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
 312
 313 # Vulkan descriptor set intrinsics
 314 #
 315 # The Vulkan API uses a different binding model from GL.  In the Vulkan
 316 # API, all external resources are represented by a tuple:
 317 #
 318 # (descriptor set, binding, array index)
 319 #
 320 # where the array index is the only thing allowed to be indirect.  The
 321 # vulkan_surface_index intrinsic takes the descriptor set and binding as
 322 # its first two indices and the array index as its source.  The third
 323 # index is a nir_variable_mode in case that's useful to the backend.
 324 #
 325 # The intended usage is that the shader will call vulkan_surface_index to
 326 # get an index and then pass that as the buffer index ubo/ssbo calls.
 327 #
 328 # The vulkan_resource_reindex intrinsic takes a resource index in src0
 329 # (the result of a vulkan_resource_index or vulkan_resource_reindex) which
 330 # corresponds to the tuple (set, binding, index) and computes an index
 331 # corresponding to tuple (set, binding, idx + src1).
 332 intrinsic("vulkan_resource_index", src_comp=[1], dest_comp=1,
 333           indices=[DESC_SET, BINDING], flags=[CAN_ELIMINATE, CAN_REORDER])
 334 intrinsic("vulkan_resource_reindex", src_comp=[1, 1], dest_comp=1,
 335           flags=[CAN_ELIMINATE, CAN_REORDER])
 336
 337 # variable atomic intrinsics
 338 #
 339 # All of these variable atomic memory operations read a value from memory,
 340 # compute a new value using one of the operations below, write the new value
 341 # to memory, and return the original value read.
 342 #
 343 # All operations take 2 sources except CompSwap that takes 3. These sources
 344 # represent:
 345 #
 346 # 0: A deref to the memory on which to perform the atomic
 347 # 1: The data parameter to the atomic function (i.e. the value to add
 348 #    in shared_atomic_add, etc).
 349 # 2: For CompSwap only: the second data parameter.
 350 intrinsic("deref_atomic_add",  src_comp=[1, 1], dest_comp=1)
 351 intrinsic("deref_atomic_imin", src_comp=[1, 1], dest_comp=1)
 352 intrinsic("deref_atomic_umin", src_comp=[1, 1], dest_comp=1)
 353 intrinsic("deref_atomic_imax", src_comp=[1, 1], dest_comp=1)
 354 intrinsic("deref_atomic_umax", src_comp=[1, 1], dest_comp=1)
 355 intrinsic("deref_atomic_and",  src_comp=[1, 1], dest_comp=1)
 356 intrinsic("deref_atomic_or",   src_comp=[1, 1], dest_comp=1)
 357 intrinsic("deref_atomic_xor",  src_comp=[1, 1], dest_comp=1)
 358 intrinsic("deref_atomic_exchange", src_comp=[1, 1], dest_comp=1)
 359 intrinsic("deref_atomic_comp_swap", src_comp=[1, 1, 1], dest_comp=1)
 360
 361 # SSBO atomic intrinsics
 362 #
 363 # All of the SSBO atomic memory operations read a value from memory,
 364 # compute a new value using one of the operations below, write the new
 365 # value to memory, and return the original value read.
 366 #
 367 # All operations take 3 sources except CompSwap that takes 4. These
 368 # sources represent:
 369 #
 370 # 0: The SSBO buffer index.
 371 # 1: The offset into the SSBO buffer of the variable that the atomic
 372 #    operation will operate on.
 373 # 2: The data parameter to the atomic function (i.e. the value to add
 374 #    in ssbo_atomic_add, etc).
 375 # 3: For CompSwap only: the second data parameter.
 376 intrinsic("ssbo_atomic_add",  src_comp=[1, 1, 1], dest_comp=1)
 377 intrinsic("ssbo_atomic_imin", src_comp=[1, 1, 1], dest_comp=1)
 378 intrinsic("ssbo_atomic_umin", src_comp=[1, 1, 1], dest_comp=1)
 379 intrinsic("ssbo_atomic_imax", src_comp=[1, 1, 1], dest_comp=1)
 380 intrinsic("ssbo_atomic_umax", src_comp=[1, 1, 1], dest_comp=1)
 381 intrinsic("ssbo_atomic_and",  src_comp=[1, 1, 1], dest_comp=1)
 382 intrinsic("ssbo_atomic_or",   src_comp=[1, 1, 1], dest_comp=1)
 383 intrinsic("ssbo_atomic_xor",  src_comp=[1, 1, 1], dest_comp=1)
 384 intrinsic("ssbo_atomic_exchange", src_comp=[1, 1, 1], dest_comp=1)
 385 intrinsic("ssbo_atomic_comp_swap", src_comp=[1, 1, 1, 1], dest_comp=1)
 386
 387 # CS shared variable atomic intrinsics
 388 #
 389 # All of the shared variable atomic memory operations read a value from
 390 # memory, compute a new value using one of the operations below, write the
 391 # new value to memory, and return the original value read.
 392 #
 393 # All operations take 2 sources except CompSwap that takes 3. These
 394 # sources represent:
 395 #
 396 # 0: The offset into the shared variable storage region that the atomic
 397 #    operation will operate on.
 398 # 1: The data parameter to the atomic function (i.e. the value to add
 399 #    in shared_atomic_add, etc).
 400 # 2: For CompSwap only: the second data parameter.
 401 intrinsic("shared_atomic_add",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 402 intrinsic("shared_atomic_imin", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 403 intrinsic("shared_atomic_umin", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 404 intrinsic("shared_atomic_imax", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 405 intrinsic("shared_atomic_umax", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 406 intrinsic("shared_atomic_and",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 407 intrinsic("shared_atomic_or",   src_comp=[1, 1], dest_comp=1, indices=[BASE])
 408 intrinsic("shared_atomic_xor",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 409 intrinsic("shared_atomic_exchange", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 410 intrinsic("shared_atomic_comp_swap", src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 411
 412 def system_value(name, dest_comp, indices=[]):
 413     intrinsic("load_" + name, [], dest_comp, indices,
 414               flags=[CAN_ELIMINATE, CAN_REORDER], sysval=True)
 415
 416 system_value("frag_coord", 4)
 417 system_value("front_face", 1)
 418 system_value("vertex_id", 1)
 419 system_value("vertex_id_zero_base", 1)
 420 system_value("first_vertex", 1)
 421 system_value("is_indexed_draw", 1)
 422 system_value("base_vertex", 1)
 423 system_value("instance_id", 1)
 424 system_value("base_instance", 1)
 425 system_value("draw_id", 1)
 426 system_value("sample_id", 1)
 427 # sample_id_no_per_sample is like sample_id but does not imply per-
 428 # sample shading.  See the lower_helper_invocation option.
 429 system_value("sample_id_no_per_sample", 1)
 430 system_value("sample_pos", 2)
 431 system_value("sample_mask_in", 1)
 432 system_value("primitive_id", 1)
 433 system_value("invocation_id", 1)
 434 system_value("tess_coord", 3)
 435 system_value("tess_level_outer", 4)
 436 system_value("tess_level_inner", 2)
 437 system_value("patch_vertices_in", 1)
 438 system_value("local_invocation_id", 3)
 439 system_value("local_invocation_index", 1)
 440 system_value("work_group_id", 3)
 441 system_value("user_clip_plane", 4, indices=[UCP_ID])
 442 system_value("num_work_groups", 3)
 443 system_value("helper_invocation", 1)
 444 system_value("alpha_ref_float", 1)
 445 system_value("layer_id", 1)
 446 system_value("view_index", 1)
 447 system_value("subgroup_size", 1)
 448 system_value("subgroup_invocation", 1)
 449 system_value("subgroup_eq_mask", 0)
 450 system_value("subgroup_ge_mask", 0)
 451 system_value("subgroup_gt_mask", 0)
 452 system_value("subgroup_le_mask", 0)
 453 system_value("subgroup_lt_mask", 0)
 454 system_value("num_subgroups", 1)
 455 system_value("subgroup_id", 1)
 456 system_value("local_group_size", 3)
 457 system_value("global_invocation_id", 3)
 458 system_value("work_dim", 1)
 459
 460 # Blend constant color values.  Float values are clamped.#
 461 system_value("blend_const_color_r_float", 1)
 462 system_value("blend_const_color_g_float", 1)
 463 system_value("blend_const_color_b_float", 1)
 464 system_value("blend_const_color_a_float", 1)
 465 system_value("blend_const_color_rgba8888_unorm", 1)
 466 system_value("blend_const_color_aaaa8888_unorm", 1)
 467
 468 # Barycentric coordinate intrinsics.
 469 #
 470 # These set up the barycentric coordinates for a particular interpolation.
 471 # The first three are for the simple cases: pixel, centroid, or per-sample
 472 # (at gl_SampleID).  The next two handle interpolating at a specified
 473 # sample location, or interpolating with a vec2 offset,
 474 #
 475 # The interp_mode index should be either the INTERP_MODE_SMOOTH or
 476 # INTERP_MODE_NOPERSPECTIVE enum values.
 477 #
 478 # The vec2 value produced by these intrinsics is intended for use as the
 479 # barycoord source of a load_interpolated_input intrinsic.
 480
 481 def barycentric(name, src_comp=[]):
 482     intrinsic("load_barycentric_" + name, src_comp=src_comp, dest_comp=2,
 483               indices=[INTERP_MODE], flags=[CAN_ELIMINATE, CAN_REORDER])
 484
 485 # no sources.  const_index[] = { interp_mode }
 486 barycentric("pixel")
 487 barycentric("centroid")
 488 barycentric("sample")
 489 # src[] = { sample_id }.  const_index[] = { interp_mode }
 490 barycentric("at_sample", [1])
 491 # src[] = { offset.xy }.  const_index[] = { interp_mode }
 492 barycentric("at_offset", [2])
 493
 494 # Load operations pull data from some piece of GPU memory.  All load
 495 # operations operate in terms of offsets into some piece of theoretical
 496 # memory.  Loads from externally visible memory (UBO and SSBO) simply take a
 497 # byte offset as a source.  Loads from opaque memory (uniforms, inputs, etc.)
 498 # take a base+offset pair where the base (const_index[0]) gives the location
 499 # of the start of the variable being loaded and and the offset source is a
 500 # offset into that variable.
 501 #
 502 # Uniform load operations have a second "range" index that specifies the
 503 # range (starting at base) of the data from which we are loading.  If
 504 # const_index[1] == 0, then the range is unknown.
 505 #
 506 # Some load operations such as UBO/SSBO load and per_vertex loads take an
 507 # additional source to specify which UBO/SSBO/vertex to load from.
 508 #
 509 # The exact address type depends on the lowering pass that generates the
 510 # load/store intrinsics.  Typically, this is vec4 units for things such as
 511 # varying slots and float units for fragment shader inputs.  UBO and SSBO
 512 # offsets are always in bytes.
 513
 514 def load(name, num_srcs, indices=[], flags=[]):
 515     intrinsic("load_" + name, [1] * num_srcs, dest_comp=0, indices=indices,
 516               flags=flags)
 517
 518 # src[] = { offset }. const_index[] = { base, range }
 519 load("uniform", 1, [BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
 520 # src[] = { buffer_index, offset }. No const_index
 521 load("ubo", 2, flags=[CAN_ELIMINATE, CAN_REORDER])
 522 # src[] = { offset }. const_index[] = { base, component }
 523 load("input", 1, [BASE, COMPONENT], [CAN_ELIMINATE, CAN_REORDER])
 524 # src[] = { vertex, offset }. const_index[] = { base, component }
 525 load("per_vertex_input", 2, [BASE, COMPONENT], [CAN_ELIMINATE, CAN_REORDER])
 526 # src[] = { barycoord, offset }. const_index[] = { base, component }
 527 intrinsic("load_interpolated_input", src_comp=[2, 1], dest_comp=0,
 528           indices=[BASE, COMPONENT], flags=[CAN_ELIMINATE, CAN_REORDER])
 529
 530 # src[] = { buffer_index, offset }. No const_index
 531 load("ssbo", 2, flags=[CAN_ELIMINATE])
 532 # src[] = { offset }. const_index[] = { base, component }
 533 load("output", 1, [BASE, COMPONENT], flags=[CAN_ELIMINATE])
 534 # src[] = { vertex, offset }. const_index[] = { base }
 535 load("per_vertex_output", 2, [BASE, COMPONENT], [CAN_ELIMINATE])
 536 # src[] = { offset }. const_index[] = { base }
 537 load("shared", 1, [BASE], [CAN_ELIMINATE])
 538 # src[] = { offset }. const_index[] = { base, range }
 539 load("push_constant", 1, [BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
 540 # src[] = { offset }. const_index[] = { base, range }
 541 load("constant", 1, [BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
 542
 543 # Stores work the same way as loads, except now the first source is the value
 544 # to store and the second (and possibly third) source specify where to store
 545 # the value.  SSBO and shared memory stores also have a write mask as
 546 # const_index[0].
 547
 548 def store(name, num_srcs, indices=[], flags=[]):
 549     intrinsic("store_" + name, [0] + ([1] * (num_srcs - 1)), indices=indices, flags=flags)
 550
 551 # src[] = { value, offset }. const_index[] = { base, write_mask, component }
 552 store("output", 2, [BASE, WRMASK, COMPONENT])
 553 # src[] = { value, vertex, offset }.
 554 # const_index[] = { base, write_mask, component }
 555 store("per_vertex_output", 3, [BASE, WRMASK, COMPONENT])
 556 # src[] = { value, block_index, offset }. const_index[] = { write_mask }
 557 store("ssbo", 3, [WRMASK])
 558 # src[] = { value, offset }. const_index[] = { base, write_mask }
 559 store("shared", 2, [BASE, WRMASK])