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, num_variables,
  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        - num_variables: the number of variables
  46        - indices: list of constant indicies
  47        - flags: list of semantic flags
  48        - sysval: is this a system-value intrinsic
  49        """
  50        assert isinstance(name, str)
  51        assert isinstance(src_components, list)
  52        if src_components:
  53            assert isinstance(src_components[0], int)
  54        assert isinstance(dest_components, int)
  55        assert isinstance(num_variables, int)
  56        assert isinstance(indices, list)
  57        if indices:
  58            assert isinstance(indices[0], str)
  59        assert isinstance(flags, list)
  60        if flags:
  61            assert isinstance(flags[0], str)
  62        assert isinstance(sysval, bool)
  63
  64        self.name = name
  65        self.num_srcs = len(src_components)
  66        self.src_components = src_components
  67        self.has_dest = (dest_components >= 0)
  68        self.dest_components = dest_components
  69        self.num_variables = num_variables
  70        self.num_indices = len(indices)
  71        self.indices = indices
  72        self.flags = flags
  73        self.sysval = sysval
  74
  75 #
  76 # Possible indices:
  77 #
  78
  79 # A constant 'base' value that is added to an offset src:
  80 BASE = "NIR_INTRINSIC_BASE"
  81 # For store instructions, a writemask:
  82 WRMASK = "NIR_INTRINSIC_WRMASK"
  83 # The stream-id for GS emit_vertex/end_primitive intrinsics:
  84 STREAM_ID = "NIR_INTRINSIC_STREAM_ID"
  85 # The clip-plane id for load_user_clip_plane intrinsics:
  86 UCP_ID = "NIR_INTRINSIC_UCP_ID"
  87 # The amount of data, starting from BASE, that this instruction
  88 # may access.  This is used to provide bounds if the offset is
  89 # not constant.
  90 RANGE = "NIR_INTRINSIC_RANGE"
  91 # The vulkan descriptor set binding for vulkan_resource_index
  92 # intrinsic
  93 DESC_SET = "NIR_INTRINSIC_DESC_SET"
  94 # The vulkan descriptor set binding for vulkan_resource_index
  95 # intrinsic
  96 BINDING = "NIR_INTRINSIC_BINDING"
  97 # Component offset
  98 COMPONENT = "NIR_INTRINSIC_COMPONENT"
  99 # Interpolation mode (only meaningful for FS inputs)
 100 INTERP_MODE = "NIR_INTRINSIC_INTERP_MODE"
 101 # A binary nir_op to use when performing a reduction or scan operation
 102 REDUCTION_OP = "NIR_INTRINSIC_REDUCTION_OP"
 103 # Cluster size for reduction operations
 104 CLUSTER_SIZE = "NIR_INTRINSIC_CLUSTER_SIZE"
 105 # Parameter index for a load_param intrinsic
 106 PARAM_IDX = "NIR_INTRINSIC_PARAM_IDX"
 107
 108 #
 109 # Possible flags:
 110 #
 111
 112 CAN_ELIMINATE = "NIR_INTRINSIC_CAN_ELIMINATE"
 113 CAN_REORDER   = "NIR_INTRINSIC_CAN_REORDER"
 114
 115 INTR_OPCODES = {}
 116
 117 def intrinsic(name, src_comp=[], dest_comp=-1, num_vars=0, indices=[],
 118               flags=[], sysval=False):
 119     assert name not in INTR_OPCODES
 120     INTR_OPCODES[name] = Intrinsic(name, src_comp, dest_comp, num_vars,
 121                                    indices, flags, sysval)
 122
 123 intrinsic("nop", flags=[CAN_ELIMINATE])
 124
 125 intrinsic("load_param", dest_comp=0, indices=[PARAM_IDX], flags=[CAN_ELIMINATE])
 126
 127 intrinsic("load_var", dest_comp=0, num_vars=1, flags=[CAN_ELIMINATE])
 128 intrinsic("store_var", src_comp=[0], num_vars=1, indices=[WRMASK])
 129 intrinsic("copy_var", num_vars=2)
 130
 131 intrinsic("load_deref", dest_comp=0, src_comp=[1], flags=[CAN_ELIMINATE])
 132 intrinsic("store_deref", src_comp=[1, 0], indices=[WRMASK])
 133 intrinsic("copy_deref", src_comp=[1, 1])
 134
 135 # Interpolation of input.  The interp_var_at* intrinsics are similar to the
 136 # load_var intrinsic acting on a shader input except that they interpolate
 137 # the input differently.  The at_sample and at_offset intrinsics take an
 138 # additional source that is an integer sample id or a vec2 position offset
 139 # respectively.
 140
 141 intrinsic("interp_var_at_centroid", dest_comp=0, num_vars=1,
 142           flags=[ CAN_ELIMINATE, CAN_REORDER])
 143 intrinsic("interp_var_at_sample", src_comp=[1], dest_comp=0, num_vars=1,
 144           flags=[CAN_ELIMINATE, CAN_REORDER])
 145 intrinsic("interp_var_at_offset", src_comp=[2], dest_comp=0, num_vars=1,
 146           flags=[CAN_ELIMINATE, CAN_REORDER])
 147
 148 # Interpolation of input.  The interp_deref_at* intrinsics are similar to the
 149 # load_var intrinsic acting on a shader input except that they interpolate the
 150 # input differently.  The at_sample and at_offset intrinsics take an
 151 # additional source that is an integer sample id or a vec2 position offset
 152 # respectively.
 153
 154 intrinsic("interp_deref_at_centroid", dest_comp=0, src_comp=[1],
 155           flags=[ CAN_ELIMINATE, CAN_REORDER])
 156 intrinsic("interp_deref_at_sample", src_comp=[1, 1], dest_comp=0,
 157           flags=[CAN_ELIMINATE, CAN_REORDER])
 158 intrinsic("interp_deref_at_offset", src_comp=[1, 2], dest_comp=0,
 159           flags=[CAN_ELIMINATE, CAN_REORDER])
 160
 161 # Ask the driver for the size of a given buffer. It takes the buffer index
 162 # as source.
 163 intrinsic("get_buffer_size", src_comp=[1], dest_comp=1,
 164           flags=[CAN_ELIMINATE, CAN_REORDER])
 165
 166 # a barrier is an intrinsic with no inputs/outputs but which can't be moved
 167 # around/optimized in general
 168 def barrier(name):
 169     intrinsic(name)
 170
 171 barrier("barrier")
 172 barrier("discard")
 173
 174 # Memory barrier with semantics analogous to the memoryBarrier() GLSL
 175 # intrinsic.
 176 barrier("memory_barrier")
 177
 178 # Shader clock intrinsic with semantics analogous to the clock2x32ARB()
 179 # GLSL intrinsic.
 180 # The latter can be used as code motion barrier, which is currently not
 181 # feasible with NIR.
 182 intrinsic("shader_clock", dest_comp=2, flags=[CAN_ELIMINATE])
 183
 184 # Shader ballot intrinsics with semantics analogous to the
 185 #
 186 #    ballotARB()
 187 #    readInvocationARB()
 188 #    readFirstInvocationARB()
 189 #
 190 # GLSL functions from ARB_shader_ballot.
 191 intrinsic("ballot", src_comp=[1], dest_comp=0, flags=[CAN_ELIMINATE])
 192 intrinsic("read_invocation", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 193 intrinsic("read_first_invocation", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 194
 195 # Additional SPIR-V ballot intrinsics
 196 #
 197 # These correspond to the SPIR-V opcodes
 198 #
 199 #    OpGroupUniformElect
 200 #    OpSubgroupFirstInvocationKHR
 201 intrinsic("elect", dest_comp=1, flags=[CAN_ELIMINATE])
 202 intrinsic("first_invocation", dest_comp=1, flags=[CAN_ELIMINATE])
 203
 204 # Memory barrier with semantics analogous to the compute shader
 205 # groupMemoryBarrier(), memoryBarrierAtomicCounter(), memoryBarrierBuffer(),
 206 # memoryBarrierImage() and memoryBarrierShared() GLSL intrinsics.
 207 barrier("group_memory_barrier")
 208 barrier("memory_barrier_atomic_counter")
 209 barrier("memory_barrier_buffer")
 210 barrier("memory_barrier_image")
 211 barrier("memory_barrier_shared")
 212 barrier("begin_invocation_interlock")
 213 barrier("end_invocation_interlock")
 214
 215 # A conditional discard, with a single boolean source.
 216 intrinsic("discard_if", src_comp=[1])
 217
 218 # ARB_shader_group_vote intrinsics
 219 intrinsic("vote_any", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
 220 intrinsic("vote_all", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
 221 intrinsic("vote_feq", src_comp=[0], dest_comp=1, flags=[CAN_ELIMINATE])
 222 intrinsic("vote_ieq", src_comp=[0], dest_comp=1, flags=[CAN_ELIMINATE])
 223
 224 # Ballot ALU operations from SPIR-V.
 225 #
 226 # These operations work like their ALU counterparts except that the operate
 227 # on a uvec4 which is treated as a 128bit integer.  Also, they are, in
 228 # general, free to ignore any bits which are above the subgroup size.
 229 intrinsic("ballot_bitfield_extract", src_comp=[4, 1], dest_comp=1, flags=[CAN_ELIMINATE])
 230 intrinsic("ballot_bit_count_reduce", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 231 intrinsic("ballot_bit_count_inclusive", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 232 intrinsic("ballot_bit_count_exclusive", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 233 intrinsic("ballot_find_lsb", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 234 intrinsic("ballot_find_msb", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 235
 236 # Shuffle operations from SPIR-V.
 237 intrinsic("shuffle", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 238 intrinsic("shuffle_xor", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 239 intrinsic("shuffle_up", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 240 intrinsic("shuffle_down", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 241
 242 # Quad operations from SPIR-V.
 243 intrinsic("quad_broadcast", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 244 intrinsic("quad_swap_horizontal", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 245 intrinsic("quad_swap_vertical", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 246 intrinsic("quad_swap_diagonal", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 247
 248 intrinsic("reduce", src_comp=[0], dest_comp=0, indices=[REDUCTION_OP, CLUSTER_SIZE],
 249           flags=[CAN_ELIMINATE])
 250 intrinsic("inclusive_scan", src_comp=[0], dest_comp=0, indices=[REDUCTION_OP],
 251           flags=[CAN_ELIMINATE])
 252 intrinsic("exclusive_scan", src_comp=[0], dest_comp=0, indices=[REDUCTION_OP],
 253           flags=[CAN_ELIMINATE])
 254
 255 # Basic Geometry Shader intrinsics.
 256 #
 257 # emit_vertex implements GLSL's EmitStreamVertex() built-in.  It takes a single
 258 # index, which is the stream ID to write to.
 259 #
 260 # end_primitive implements GLSL's EndPrimitive() built-in.
 261 intrinsic("emit_vertex",   indices=[STREAM_ID])
 262 intrinsic("end_primitive", indices=[STREAM_ID])
 263
 264 # Geometry Shader intrinsics with a vertex count.
 265 #
 266 # Alternatively, drivers may implement these intrinsics, and use
 267 # nir_lower_gs_intrinsics() to convert from the basic intrinsics.
 268 #
 269 # These maintain a count of the number of vertices emitted, as an additional
 270 # unsigned integer source.
 271 intrinsic("emit_vertex_with_counter", src_comp=[1], indices=[STREAM_ID])
 272 intrinsic("end_primitive_with_counter", src_comp=[1], indices=[STREAM_ID])
 273 intrinsic("set_vertex_count", src_comp=[1])
 274
 275 # Atomic counters
 276 #
 277 # The *_var variants take an atomic_uint nir_variable, while the other,
 278 # lowered, variants take a constant buffer index and register offset.
 279
 280 def atomic(name, flags=[]):
 281     intrinsic(name + "_var", dest_comp=1, num_vars=1, flags=flags)
 282     intrinsic(name + "_deref", src_comp=[1], dest_comp=1, flags=flags)
 283     intrinsic(name, src_comp=[1], dest_comp=1, indices=[BASE], flags=flags)
 284
 285 def atomic2(name):
 286     intrinsic(name + "_var", src_comp=[1], dest_comp=1, num_vars=1)
 287     intrinsic(name + "_deref", src_comp=[1, 1], dest_comp=1)
 288     intrinsic(name, src_comp=[1, 1], dest_comp=1, indices=[BASE])
 289
 290 def atomic3(name):
 291     intrinsic(name + "_var", src_comp=[1, 1], dest_comp=1, num_vars=1)
 292     intrinsic(name + "_deref", src_comp=[1, 1, 1], dest_comp=1)
 293     intrinsic(name, src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 294
 295 atomic("atomic_counter_inc")
 296 atomic("atomic_counter_dec")
 297 atomic("atomic_counter_read", flags=[CAN_ELIMINATE])
 298 atomic2("atomic_counter_add")
 299 atomic2("atomic_counter_min")
 300 atomic2("atomic_counter_max")
 301 atomic2("atomic_counter_and")
 302 atomic2("atomic_counter_or")
 303 atomic2("atomic_counter_xor")
 304 atomic2("atomic_counter_exchange")
 305 atomic3("atomic_counter_comp_swap")
 306
 307 # Image load, store and atomic intrinsics.
 308 #
 309 # All image intrinsics take an image target passed as a nir_variable.  Image
 310 # variables contain a number of memory and layout qualifiers that influence
 311 # the semantics of the intrinsic.
 312 #
 313 # All image intrinsics take a four-coordinate vector and a sample index as
 314 # first two sources, determining the location within the image that will be
 315 # accessed by the intrinsic.  Components not applicable to the image target
 316 # in use are undefined.  Image store takes an additional four-component
 317 # argument with the value to be written, and image atomic operations take
 318 # either one or two additional scalar arguments with the same meaning as in
 319 # the ARB_shader_image_load_store specification.
 320 intrinsic("image_var_load", src_comp=[4, 1], dest_comp=4, num_vars=1,
 321           flags=[CAN_ELIMINATE])
 322 intrinsic("image_var_store", src_comp=[4, 1, 4], num_vars=1)
 323 intrinsic("image_var_atomic_add",  src_comp=[4, 1, 1], dest_comp=1, num_vars=1)
 324 intrinsic("image_var_atomic_min",  src_comp=[4, 1, 1], dest_comp=1, num_vars=1)
 325 intrinsic("image_var_atomic_max",  src_comp=[4, 1, 1], dest_comp=1, num_vars=1)
 326 intrinsic("image_var_atomic_and",  src_comp=[4, 1, 1], dest_comp=1, num_vars=1)
 327 intrinsic("image_var_atomic_or",   src_comp=[4, 1, 1], dest_comp=1, num_vars=1)
 328 intrinsic("image_var_atomic_xor",  src_comp=[4, 1, 1], dest_comp=1, num_vars=1)
 329 intrinsic("image_var_atomic_exchange",  src_comp=[4, 1, 1], dest_comp=1, num_vars=1)
 330 intrinsic("image_var_atomic_comp_swap", src_comp=[4, 1, 1, 1], dest_comp=1, num_vars=1)
 331 intrinsic("image_var_size",    dest_comp=0, num_vars=1, flags=[CAN_ELIMINATE, CAN_REORDER])
 332 intrinsic("image_var_samples", dest_comp=1, num_vars=1, flags=[CAN_ELIMINATE, CAN_REORDER])
 333
 334 # Image load, store and atomic intrinsics.
 335 #
 336 # All image intrinsics take an image target passed as a nir_variable.  The
 337 # variable is passed in using a chain of nir_deref_instr with as the first
 338 # source of the image intrinsic.  Image variables contain a number of memory
 339 # and layout qualifiers that influence the semantics of the intrinsic.
 340 #
 341 # All image intrinsics take a four-coordinate vector and a sample index as
 342 # first two sources, determining the location within the image that will be
 343 # accessed by the intrinsic.  Components not applicable to the image target
 344 # in use are undefined.  Image store takes an additional four-component
 345 # argument with the value to be written, and image atomic operations take
 346 # either one or two additional scalar arguments with the same meaning as in
 347 # the ARB_shader_image_load_store specification.
 348 intrinsic("image_deref_load", src_comp=[1, 4, 1], dest_comp=4,
 349           flags=[CAN_ELIMINATE])
 350 intrinsic("image_deref_store", src_comp=[1, 4, 1, 4])
 351 intrinsic("image_deref_atomic_add",  src_comp=[1, 4, 1, 1], dest_comp=1)
 352 intrinsic("image_deref_atomic_min",  src_comp=[1, 4, 1, 1], dest_comp=1)
 353 intrinsic("image_deref_atomic_max",  src_comp=[1, 4, 1, 1], dest_comp=1)
 354 intrinsic("image_deref_atomic_and",  src_comp=[1, 4, 1, 1], dest_comp=1)
 355 intrinsic("image_deref_atomic_or",   src_comp=[1, 4, 1, 1], dest_comp=1)
 356 intrinsic("image_deref_atomic_xor",  src_comp=[1, 4, 1, 1], dest_comp=1)
 357 intrinsic("image_deref_atomic_exchange",  src_comp=[1, 4, 1, 1], dest_comp=1)
 358 intrinsic("image_deref_atomic_comp_swap", src_comp=[1, 4, 1, 1, 1], dest_comp=1)
 359 intrinsic("image_deref_size",    src_comp=[1], dest_comp=0, flags=[CAN_ELIMINATE, CAN_REORDER])
 360 intrinsic("image_deref_samples", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
 361
 362 # Vulkan descriptor set intrinsics
 363 #
 364 # The Vulkan API uses a different binding model from GL.  In the Vulkan
 365 # API, all external resources are represented by a tuple:
 366 #
 367 # (descriptor set, binding, array index)
 368 #
 369 # where the array index is the only thing allowed to be indirect.  The
 370 # vulkan_surface_index intrinsic takes the descriptor set and binding as
 371 # its first two indices and the array index as its source.  The third
 372 # index is a nir_variable_mode in case that's useful to the backend.
 373 #
 374 # The intended usage is that the shader will call vulkan_surface_index to
 375 # get an index and then pass that as the buffer index ubo/ssbo calls.
 376 #
 377 # The vulkan_resource_reindex intrinsic takes a resource index in src0
 378 # (the result of a vulkan_resource_index or vulkan_resource_reindex) which
 379 # corresponds to the tuple (set, binding, index) and computes an index
 380 # corresponding to tuple (set, binding, idx + src1).
 381 intrinsic("vulkan_resource_index", src_comp=[1], dest_comp=1,
 382           indices=[DESC_SET, BINDING], flags=[CAN_ELIMINATE, CAN_REORDER])
 383 intrinsic("vulkan_resource_reindex", src_comp=[1, 1], dest_comp=1,
 384           flags=[CAN_ELIMINATE, CAN_REORDER])
 385
 386 # variable atomic intrinsics
 387 #
 388 # All of these variable atomic memory operations read a value from memory,
 389 # compute a new value using one of the operations below, write the new value
 390 # to memory, and return the original value read.
 391 #
 392 # All operations take 1 source except CompSwap that takes 2. These sources
 393 # represent:
 394 #
 395 # 0: The data parameter to the atomic function (i.e. the value to add
 396 #    in shared_atomic_add, etc).
 397 # 1: For CompSwap only: the second data parameter.
 398 #
 399 # All operations take 1 variable deref.
 400 intrinsic("var_atomic_add",  src_comp=[1], dest_comp=1, num_vars=1)
 401 intrinsic("var_atomic_imin", src_comp=[1], dest_comp=1, num_vars=1)
 402 intrinsic("var_atomic_umin", src_comp=[1], dest_comp=1, num_vars=1)
 403 intrinsic("var_atomic_imax", src_comp=[1], dest_comp=1, num_vars=1)
 404 intrinsic("var_atomic_umax", src_comp=[1], dest_comp=1, num_vars=1)
 405 intrinsic("var_atomic_and",  src_comp=[1], dest_comp=1, num_vars=1)
 406 intrinsic("var_atomic_or",   src_comp=[1], dest_comp=1, num_vars=1)
 407 intrinsic("var_atomic_xor",  src_comp=[1], dest_comp=1, num_vars=1)
 408 intrinsic("var_atomic_exchange", src_comp=[1], dest_comp=1, num_vars=1)
 409 intrinsic("var_atomic_comp_swap", src_comp=[1, 1], dest_comp=1, num_vars=1)
 410
 411 # variable atomic intrinsics
 412 #
 413 # All of these variable atomic memory operations read a value from memory,
 414 # compute a new value using one of the operations below, write the new value
 415 # to memory, and return the original value read.
 416 #
 417 # All operations take 2 sources except CompSwap that takes 3. These sources
 418 # represent:
 419 #
 420 # 0: A deref to the memory on which to perform the atomic
 421 # 1: The data parameter to the atomic function (i.e. the value to add
 422 #    in shared_atomic_add, etc).
 423 # 2: For CompSwap only: the second data parameter.
 424 intrinsic("deref_atomic_add",  src_comp=[1, 1], dest_comp=1)
 425 intrinsic("deref_atomic_imin", src_comp=[1, 1], dest_comp=1)
 426 intrinsic("deref_atomic_umin", src_comp=[1, 1], dest_comp=1)
 427 intrinsic("deref_atomic_imax", src_comp=[1, 1], dest_comp=1)
 428 intrinsic("deref_atomic_umax", src_comp=[1, 1], dest_comp=1)
 429 intrinsic("deref_atomic_and",  src_comp=[1, 1], dest_comp=1)
 430 intrinsic("deref_atomic_or",   src_comp=[1, 1], dest_comp=1)
 431 intrinsic("deref_atomic_xor",  src_comp=[1, 1], dest_comp=1)
 432 intrinsic("deref_atomic_exchange", src_comp=[1, 1], dest_comp=1)
 433 intrinsic("deref_atomic_comp_swap", src_comp=[1, 1, 1], dest_comp=1)
 434
 435 # SSBO atomic intrinsics
 436 #
 437 # All of the SSBO atomic memory operations read a value from memory,
 438 # compute a new value using one of the operations below, write the new
 439 # value to memory, and return the original value read.
 440 #
 441 # All operations take 3 sources except CompSwap that takes 4. These
 442 # sources represent:
 443 #
 444 # 0: The SSBO buffer index.
 445 # 1: The offset into the SSBO buffer of the variable that the atomic
 446 #    operation will operate on.
 447 # 2: The data parameter to the atomic function (i.e. the value to add
 448 #    in ssbo_atomic_add, etc).
 449 # 3: For CompSwap only: the second data parameter.
 450 intrinsic("ssbo_atomic_add",  src_comp=[1, 1, 1], dest_comp=1)
 451 intrinsic("ssbo_atomic_imin", src_comp=[1, 1, 1], dest_comp=1)
 452 intrinsic("ssbo_atomic_umin", src_comp=[1, 1, 1], dest_comp=1)
 453 intrinsic("ssbo_atomic_imax", src_comp=[1, 1, 1], dest_comp=1)
 454 intrinsic("ssbo_atomic_umax", src_comp=[1, 1, 1], dest_comp=1)
 455 intrinsic("ssbo_atomic_and",  src_comp=[1, 1, 1], dest_comp=1)
 456 intrinsic("ssbo_atomic_or",   src_comp=[1, 1, 1], dest_comp=1)
 457 intrinsic("ssbo_atomic_xor",  src_comp=[1, 1, 1], dest_comp=1)
 458 intrinsic("ssbo_atomic_exchange", src_comp=[1, 1, 1], dest_comp=1)
 459 intrinsic("ssbo_atomic_comp_swap", src_comp=[1, 1, 1, 1], dest_comp=1)
 460
 461 # CS shared variable atomic intrinsics
 462 #
 463 # All of the shared variable atomic memory operations read a value from
 464 # memory, compute a new value using one of the operations below, write the
 465 # new value to memory, and return the original value read.
 466 #
 467 # All operations take 2 sources except CompSwap that takes 3. These
 468 # sources represent:
 469 #
 470 # 0: The offset into the shared variable storage region that the atomic
 471 #    operation will operate on.
 472 # 1: The data parameter to the atomic function (i.e. the value to add
 473 #    in shared_atomic_add, etc).
 474 # 2: For CompSwap only: the second data parameter.
 475 intrinsic("shared_atomic_add",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 476 intrinsic("shared_atomic_imin", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 477 intrinsic("shared_atomic_umin", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 478 intrinsic("shared_atomic_imax", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 479 intrinsic("shared_atomic_umax", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 480 intrinsic("shared_atomic_and",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 481 intrinsic("shared_atomic_or",   src_comp=[1, 1], dest_comp=1, indices=[BASE])
 482 intrinsic("shared_atomic_xor",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 483 intrinsic("shared_atomic_exchange", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 484 intrinsic("shared_atomic_comp_swap", src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 485
 486 def system_value(name, dest_comp, indices=[]):
 487     intrinsic("load_" + name, [], dest_comp, 0, indices,
 488               flags=[CAN_ELIMINATE, CAN_REORDER], sysval=True)
 489
 490 system_value("frag_coord", 4)
 491 system_value("front_face", 1)
 492 system_value("vertex_id", 1)
 493 system_value("vertex_id_zero_base", 1)
 494 system_value("first_vertex", 1)
 495 system_value("is_indexed_draw", 1)
 496 system_value("base_vertex", 1)
 497 system_value("instance_id", 1)
 498 system_value("base_instance", 1)
 499 system_value("draw_id", 1)
 500 system_value("sample_id", 1)
 501 system_value("sample_pos", 2)
 502 system_value("sample_mask_in", 1)
 503 system_value("primitive_id", 1)
 504 system_value("invocation_id", 1)
 505 system_value("tess_coord", 3)
 506 system_value("tess_level_outer", 4)
 507 system_value("tess_level_inner", 2)
 508 system_value("patch_vertices_in", 1)
 509 system_value("local_invocation_id", 3)
 510 system_value("local_invocation_index", 1)
 511 system_value("work_group_id", 3)
 512 system_value("user_clip_plane", 4, indices=[UCP_ID])
 513 system_value("num_work_groups", 3)
 514 system_value("helper_invocation", 1)
 515 system_value("alpha_ref_float", 1)
 516 system_value("layer_id", 1)
 517 system_value("view_index", 1)
 518 system_value("subgroup_size", 1)
 519 system_value("subgroup_invocation", 1)
 520 system_value("subgroup_eq_mask", 0)
 521 system_value("subgroup_ge_mask", 0)
 522 system_value("subgroup_gt_mask", 0)
 523 system_value("subgroup_le_mask", 0)
 524 system_value("subgroup_lt_mask", 0)
 525 system_value("num_subgroups", 1)
 526 system_value("subgroup_id", 1)
 527 system_value("local_group_size", 3)
 528 system_value("global_invocation_id", 3)
 529
 530 # Blend constant color values.  Float values are clamped.#
 531 system_value("blend_const_color_r_float", 1)
 532 system_value("blend_const_color_g_float", 1)
 533 system_value("blend_const_color_b_float", 1)
 534 system_value("blend_const_color_a_float", 1)
 535 system_value("blend_const_color_rgba8888_unorm", 1)
 536 system_value("blend_const_color_aaaa8888_unorm", 1)
 537
 538 # Barycentric coordinate intrinsics.
 539 #
 540 # These set up the barycentric coordinates for a particular interpolation.
 541 # The first three are for the simple cases: pixel, centroid, or per-sample
 542 # (at gl_SampleID).  The next two handle interpolating at a specified
 543 # sample location, or interpolating with a vec2 offset,
 544 #
 545 # The interp_mode index should be either the INTERP_MODE_SMOOTH or
 546 # INTERP_MODE_NOPERSPECTIVE enum values.
 547 #
 548 # The vec2 value produced by these intrinsics is intended for use as the
 549 # barycoord source of a load_interpolated_input intrinsic.
 550
 551 def barycentric(name, src_comp=[]):
 552     intrinsic("load_barycentric_" + name, src_comp=src_comp, dest_comp=2,
 553               indices=[INTERP_MODE], flags=[CAN_ELIMINATE, CAN_REORDER])
 554
 555 # no sources.  const_index[] = { interp_mode }
 556 barycentric("pixel")
 557 barycentric("centroid")
 558 barycentric("sample")
 559 # src[] = { sample_id }.  const_index[] = { interp_mode }
 560 barycentric("at_sample", [1])
 561 # src[] = { offset.xy }.  const_index[] = { interp_mode }
 562 barycentric("at_offset", [2])
 563
 564 # Load operations pull data from some piece of GPU memory.  All load
 565 # operations operate in terms of offsets into some piece of theoretical
 566 # memory.  Loads from externally visible memory (UBO and SSBO) simply take a
 567 # byte offset as a source.  Loads from opaque memory (uniforms, inputs, etc.)
 568 # take a base+offset pair where the base (const_index[0]) gives the location
 569 # of the start of the variable being loaded and and the offset source is a
 570 # offset into that variable.
 571 #
 572 # Uniform load operations have a second "range" index that specifies the
 573 # range (starting at base) of the data from which we are loading.  If
 574 # const_index[1] == 0, then the range is unknown.
 575 #
 576 # Some load operations such as UBO/SSBO load and per_vertex loads take an
 577 # additional source to specify which UBO/SSBO/vertex to load from.
 578 #
 579 # The exact address type depends on the lowering pass that generates the
 580 # load/store intrinsics.  Typically, this is vec4 units for things such as
 581 # varying slots and float units for fragment shader inputs.  UBO and SSBO
 582 # offsets are always in bytes.
 583
 584 def load(name, num_srcs, indices=[], flags=[]):
 585     intrinsic("load_" + name, [1] * num_srcs, dest_comp=0, indices=indices,
 586               flags=flags)
 587
 588 # src[] = { offset }. const_index[] = { base, range }
 589 load("uniform", 1, [BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
 590 # src[] = { buffer_index, offset }. No const_index
 591 load("ubo", 2, flags=[CAN_ELIMINATE, CAN_REORDER])
 592 # src[] = { offset }. const_index[] = { base, component }
 593 load("input", 1, [BASE, COMPONENT], [CAN_ELIMINATE, CAN_REORDER])
 594 # src[] = { vertex, offset }. const_index[] = { base, component }
 595 load("per_vertex_input", 2, [BASE, COMPONENT], [CAN_ELIMINATE, CAN_REORDER])
 596 # src[] = { barycoord, offset }. const_index[] = { base, component }
 597 intrinsic("load_interpolated_input", src_comp=[2, 1], dest_comp=0,
 598           indices=[BASE, COMPONENT], flags=[CAN_ELIMINATE, CAN_REORDER])
 599
 600 # src[] = { buffer_index, offset }. No const_index
 601 load("ssbo", 2, flags=[CAN_ELIMINATE])
 602 # src[] = { offset }. const_index[] = { base, component }
 603 load("output", 1, [BASE, COMPONENT], flags=[CAN_ELIMINATE])
 604 # src[] = { vertex, offset }. const_index[] = { base }
 605 load("per_vertex_output", 2, [BASE, COMPONENT], [CAN_ELIMINATE])
 606 # src[] = { offset }. const_index[] = { base }
 607 load("shared", 1, [BASE], [CAN_ELIMINATE])
 608 # src[] = { offset }. const_index[] = { base, range }
 609 load("push_constant", 1, [BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
 610
 611 # Stores work the same way as loads, except now the first source is the value
 612 # to store and the second (and possibly third) source specify where to store
 613 # the value.  SSBO and shared memory stores also have a write mask as
 614 # const_index[0].
 615
 616 def store(name, num_srcs, indices=[], flags=[]):
 617     intrinsic("store_" + name, [0] + ([1] * (num_srcs - 1)), indices=indices, flags=flags)
 618
 619 # src[] = { value, offset }. const_index[] = { base, write_mask, component }
 620 store("output", 2, [BASE, WRMASK, COMPONENT])
 621 # src[] = { value, vertex, offset }.
 622 # const_index[] = { base, write_mask, component }
 623 store("per_vertex_output", 3, [BASE, WRMASK, COMPONENT])
 624 # src[] = { value, block_index, offset }. const_index[] = { write_mask }
 625 store("ssbo", 3, [WRMASK])
 626 # src[] = { value, offset }. const_index[] = { base, write_mask }
 627 store("shared", 2, [BASE, WRMASK])