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