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, bit_sizes):
  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        - bit_sizes: allowed dest bit_sizes
  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(indices, list)
  56        if indices:
  57            assert isinstance(indices[0], str)
  58        assert isinstance(flags, list)
  59        if flags:
  60            assert isinstance(flags[0], str)
  61        assert isinstance(sysval, bool)
  62        if bit_sizes:
  63            assert isinstance(bit_sizes[0], int)
  64
  65        self.name = name
  66        self.num_srcs = len(src_components)
  67        self.src_components = src_components
  68        self.has_dest = (dest_components >= 0)
  69        self.dest_components = dest_components
  70        self.num_indices = len(indices)
  71        self.indices = indices
  72        self.flags = flags
  73        self.sysval = sysval
  74        self.bit_sizes = bit_sizes
  75
  76 #
  77 # Possible indices:
  78 #
  79
  80 # A constant 'base' value that is added to an offset src:
  81 BASE = "NIR_INTRINSIC_BASE"
  82 # For store instructions, a writemask:
  83 WRMASK = "NIR_INTRINSIC_WRMASK"
  84 # The stream-id for GS emit_vertex/end_primitive intrinsics:
  85 STREAM_ID = "NIR_INTRINSIC_STREAM_ID"
  86 # The clip-plane id for load_user_clip_plane intrinsics:
  87 UCP_ID = "NIR_INTRINSIC_UCP_ID"
  88 # The amount of data, starting from BASE, that this instruction
  89 # may access.  This is used to provide bounds if the offset is
  90 # not constant.
  91 RANGE = "NIR_INTRINSIC_RANGE"
  92 # The vulkan descriptor set binding for vulkan_resource_index
  93 # intrinsic
  94 DESC_SET = "NIR_INTRINSIC_DESC_SET"
  95 # The vulkan descriptor set binding for vulkan_resource_index
  96 # intrinsic
  97 BINDING = "NIR_INTRINSIC_BINDING"
  98 # Component offset
  99 COMPONENT = "NIR_INTRINSIC_COMPONENT"
 100 # Interpolation mode (only meaningful for FS inputs)
 101 INTERP_MODE = "NIR_INTRINSIC_INTERP_MODE"
 102 # A binary nir_op to use when performing a reduction or scan operation
 103 REDUCTION_OP = "NIR_INTRINSIC_REDUCTION_OP"
 104 # Cluster size for reduction operations
 105 CLUSTER_SIZE = "NIR_INTRINSIC_CLUSTER_SIZE"
 106 # Parameter index for a load_param intrinsic
 107 PARAM_IDX = "NIR_INTRINSIC_PARAM_IDX"
 108 # Image dimensionality for image intrinsics
 109 IMAGE_DIM = "NIR_INTRINSIC_IMAGE_DIM"
 110 # Non-zero if we are accessing an array image
 111 IMAGE_ARRAY = "NIR_INTRINSIC_IMAGE_ARRAY"
 112 # Access qualifiers for image and memory access intrinsics
 113 ACCESS = "NIR_INTRINSIC_ACCESS"
 114 # Image format for image intrinsics
 115 FORMAT = "NIR_INTRINSIC_FORMAT"
 116 # Offset or address alignment
 117 ALIGN_MUL = "NIR_INTRINSIC_ALIGN_MUL"
 118 ALIGN_OFFSET = "NIR_INTRINSIC_ALIGN_OFFSET"
 119 # The vulkan descriptor type for vulkan_resource_index
 120 DESC_TYPE = "NIR_INTRINSIC_DESC_TYPE"
 121
 122 #
 123 # Possible flags:
 124 #
 125
 126 CAN_ELIMINATE = "NIR_INTRINSIC_CAN_ELIMINATE"
 127 CAN_REORDER   = "NIR_INTRINSIC_CAN_REORDER"
 128
 129 INTR_OPCODES = {}
 130
 131 # Defines a new NIR intrinsic.  By default, the intrinsic will have no sources
 132 # and no destination.
 133 #
 134 # You can set dest_comp=n to enable a destination for the intrinsic, in which
 135 # case it will have that many components, or =0 for "as many components as the
 136 # NIR destination value."
 137 #
 138 # Set src_comp=n to enable sources for the intruction.  It can be an array of
 139 # component counts, or (for convenience) a scalar component count if there's
 140 # only one source.  If a component count is 0, it will be as many components as
 141 # the intrinsic has based on the dest_comp.
 142 def intrinsic(name, src_comp=[], dest_comp=-1, indices=[],
 143               flags=[], sysval=False, bit_sizes=[]):
 144     assert name not in INTR_OPCODES
 145     INTR_OPCODES[name] = Intrinsic(name, src_comp, dest_comp,
 146                                    indices, flags, sysval, bit_sizes)
 147
 148 intrinsic("nop", flags=[CAN_ELIMINATE])
 149
 150 intrinsic("load_param", dest_comp=0, indices=[PARAM_IDX], flags=[CAN_ELIMINATE])
 151
 152 intrinsic("load_deref", dest_comp=0, src_comp=[-1],
 153           indices=[ACCESS], flags=[CAN_ELIMINATE])
 154 intrinsic("store_deref", src_comp=[-1, 0], indices=[WRMASK, ACCESS])
 155 intrinsic("copy_deref", src_comp=[-1, -1])
 156
 157 # Interpolation of input.  The interp_deref_at* intrinsics are similar to the
 158 # load_var intrinsic acting on a shader input except that they interpolate the
 159 # input differently.  The at_sample and at_offset intrinsics take an
 160 # additional source that is an integer sample id or a vec2 position offset
 161 # respectively.
 162
 163 intrinsic("interp_deref_at_centroid", dest_comp=0, src_comp=[1],
 164           flags=[ CAN_ELIMINATE, CAN_REORDER])
 165 intrinsic("interp_deref_at_sample", src_comp=[1, 1], dest_comp=0,
 166           flags=[CAN_ELIMINATE, CAN_REORDER])
 167 intrinsic("interp_deref_at_offset", src_comp=[1, 2], dest_comp=0,
 168           flags=[CAN_ELIMINATE, CAN_REORDER])
 169
 170 # Gets the length of an unsized array at the end of a buffer
 171 intrinsic("deref_buffer_array_length", src_comp=[-1], dest_comp=1,
 172           flags=[CAN_ELIMINATE, CAN_REORDER])
 173
 174 # Ask the driver for the size of a given buffer. It takes the buffer index
 175 # as source.
 176 intrinsic("get_buffer_size", src_comp=[-1], dest_comp=1,
 177           flags=[CAN_ELIMINATE, CAN_REORDER])
 178
 179 # a barrier is an intrinsic with no inputs/outputs but which can't be moved
 180 # around/optimized in general
 181 def barrier(name):
 182     intrinsic(name)
 183
 184 barrier("barrier")
 185 barrier("discard")
 186
 187 # Memory barrier with semantics analogous to the memoryBarrier() GLSL
 188 # intrinsic.
 189 barrier("memory_barrier")
 190
 191 # Shader clock intrinsic with semantics analogous to the clock2x32ARB()
 192 # GLSL intrinsic.
 193 # The latter can be used as code motion barrier, which is currently not
 194 # feasible with NIR.
 195 intrinsic("shader_clock", dest_comp=2, flags=[CAN_ELIMINATE])
 196
 197 # Shader ballot intrinsics with semantics analogous to the
 198 #
 199 #    ballotARB()
 200 #    readInvocationARB()
 201 #    readFirstInvocationARB()
 202 #
 203 # GLSL functions from ARB_shader_ballot.
 204 intrinsic("ballot", src_comp=[1], dest_comp=0, flags=[CAN_ELIMINATE])
 205 intrinsic("read_invocation", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 206 intrinsic("read_first_invocation", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 207
 208 # Additional SPIR-V ballot intrinsics
 209 #
 210 # These correspond to the SPIR-V opcodes
 211 #
 212 #    OpGroupUniformElect
 213 #    OpSubgroupFirstInvocationKHR
 214 intrinsic("elect", dest_comp=1, flags=[CAN_ELIMINATE])
 215 intrinsic("first_invocation", dest_comp=1, flags=[CAN_ELIMINATE])
 216
 217 # Memory barrier with semantics analogous to the compute shader
 218 # groupMemoryBarrier(), memoryBarrierAtomicCounter(), memoryBarrierBuffer(),
 219 # memoryBarrierImage() and memoryBarrierShared() GLSL intrinsics.
 220 barrier("group_memory_barrier")
 221 barrier("memory_barrier_atomic_counter")
 222 barrier("memory_barrier_buffer")
 223 barrier("memory_barrier_image")
 224 barrier("memory_barrier_shared")
 225 barrier("begin_invocation_interlock")
 226 barrier("end_invocation_interlock")
 227
 228 # A conditional discard, with a single boolean source.
 229 intrinsic("discard_if", src_comp=[1])
 230
 231 # ARB_shader_group_vote intrinsics
 232 intrinsic("vote_any", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
 233 intrinsic("vote_all", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
 234 intrinsic("vote_feq", src_comp=[0], dest_comp=1, flags=[CAN_ELIMINATE])
 235 intrinsic("vote_ieq", src_comp=[0], dest_comp=1, flags=[CAN_ELIMINATE])
 236
 237 # Ballot ALU operations from SPIR-V.
 238 #
 239 # These operations work like their ALU counterparts except that the operate
 240 # on a uvec4 which is treated as a 128bit integer.  Also, they are, in
 241 # general, free to ignore any bits which are above the subgroup size.
 242 intrinsic("ballot_bitfield_extract", src_comp=[4, 1], dest_comp=1, flags=[CAN_ELIMINATE])
 243 intrinsic("ballot_bit_count_reduce", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 244 intrinsic("ballot_bit_count_inclusive", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 245 intrinsic("ballot_bit_count_exclusive", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 246 intrinsic("ballot_find_lsb", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 247 intrinsic("ballot_find_msb", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 248
 249 # Shuffle operations from SPIR-V.
 250 intrinsic("shuffle", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 251 intrinsic("shuffle_xor", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 252 intrinsic("shuffle_up", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 253 intrinsic("shuffle_down", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 254
 255 # Quad operations from SPIR-V.
 256 intrinsic("quad_broadcast", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 257 intrinsic("quad_swap_horizontal", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 258 intrinsic("quad_swap_vertical", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 259 intrinsic("quad_swap_diagonal", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 260
 261 intrinsic("reduce", src_comp=[0], dest_comp=0, indices=[REDUCTION_OP, CLUSTER_SIZE],
 262           flags=[CAN_ELIMINATE])
 263 intrinsic("inclusive_scan", src_comp=[0], dest_comp=0, indices=[REDUCTION_OP],
 264           flags=[CAN_ELIMINATE])
 265 intrinsic("exclusive_scan", src_comp=[0], dest_comp=0, indices=[REDUCTION_OP],
 266           flags=[CAN_ELIMINATE])
 267
 268 # Basic Geometry Shader intrinsics.
 269 #
 270 # emit_vertex implements GLSL's EmitStreamVertex() built-in.  It takes a single
 271 # index, which is the stream ID to write to.
 272 #
 273 # end_primitive implements GLSL's EndPrimitive() built-in.
 274 intrinsic("emit_vertex",   indices=[STREAM_ID])
 275 intrinsic("end_primitive", indices=[STREAM_ID])
 276
 277 # Geometry Shader intrinsics with a vertex count.
 278 #
 279 # Alternatively, drivers may implement these intrinsics, and use
 280 # nir_lower_gs_intrinsics() to convert from the basic intrinsics.
 281 #
 282 # These maintain a count of the number of vertices emitted, as an additional
 283 # unsigned integer source.
 284 intrinsic("emit_vertex_with_counter", src_comp=[1], indices=[STREAM_ID])
 285 intrinsic("end_primitive_with_counter", src_comp=[1], indices=[STREAM_ID])
 286 intrinsic("set_vertex_count", src_comp=[1])
 287
 288 # Atomic counters
 289 #
 290 # The *_var variants take an atomic_uint nir_variable, while the other,
 291 # lowered, variants take a constant buffer index and register offset.
 292
 293 def atomic(name, flags=[]):
 294     intrinsic(name + "_deref", src_comp=[-1], dest_comp=1, flags=flags)
 295     intrinsic(name, src_comp=[1], dest_comp=1, indices=[BASE], flags=flags)
 296
 297 def atomic2(name):
 298     intrinsic(name + "_deref", src_comp=[-1, 1], dest_comp=1)
 299     intrinsic(name, src_comp=[1, 1], dest_comp=1, indices=[BASE])
 300
 301 def atomic3(name):
 302     intrinsic(name + "_deref", src_comp=[-1, 1, 1], dest_comp=1)
 303     intrinsic(name, src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 304
 305 atomic("atomic_counter_inc")
 306 atomic("atomic_counter_pre_dec")
 307 atomic("atomic_counter_post_dec")
 308 atomic("atomic_counter_read", flags=[CAN_ELIMINATE])
 309 atomic2("atomic_counter_add")
 310 atomic2("atomic_counter_min")
 311 atomic2("atomic_counter_max")
 312 atomic2("atomic_counter_and")
 313 atomic2("atomic_counter_or")
 314 atomic2("atomic_counter_xor")
 315 atomic2("atomic_counter_exchange")
 316 atomic3("atomic_counter_comp_swap")
 317
 318 # Image load, store and atomic intrinsics.
 319 #
 320 # All image intrinsics come in three versions.  One which take an image target
 321 # passed as a deref chain as the first source, one which takes an index as the
 322 # first source, and one which takes a bindless handle as the first source.
 323 # In the first version, the image variable contains the memory and layout
 324 # qualifiers that influence the semantics of the intrinsic.  In the second and
 325 # third, the image format and access qualifiers are provided as constant
 326 # indices.
 327 #
 328 # All image intrinsics take a four-coordinate vector and a sample index as
 329 # 2nd and 3rd sources, determining the location within the image that will be
 330 # accessed by the intrinsic.  Components not applicable to the image target
 331 # in use are undefined.  Image store takes an additional four-component
 332 # argument with the value to be written, and image atomic operations take
 333 # either one or two additional scalar arguments with the same meaning as in
 334 # the ARB_shader_image_load_store specification.
 335 def image(name, src_comp=[], **kwargs):
 336     intrinsic("image_deref_" + name, src_comp=[1] + src_comp, **kwargs)
 337     intrinsic("image_" + name, src_comp=[1] + src_comp,
 338               indices=[IMAGE_DIM, IMAGE_ARRAY, FORMAT, ACCESS], **kwargs)
 339     intrinsic("bindless_image_" + name, src_comp=[1] + src_comp,
 340               indices=[IMAGE_DIM, IMAGE_ARRAY, FORMAT, ACCESS], **kwargs)
 341
 342 image("load", src_comp=[4, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 343 image("store", src_comp=[4, 1, 0])
 344 image("atomic_add",  src_comp=[4, 1, 1], dest_comp=1)
 345 image("atomic_min",  src_comp=[4, 1, 1], dest_comp=1)
 346 image("atomic_max",  src_comp=[4, 1, 1], dest_comp=1)
 347 image("atomic_and",  src_comp=[4, 1, 1], dest_comp=1)
 348 image("atomic_or",   src_comp=[4, 1, 1], dest_comp=1)
 349 image("atomic_xor",  src_comp=[4, 1, 1], dest_comp=1)
 350 image("atomic_exchange",  src_comp=[4, 1, 1], dest_comp=1)
 351 image("atomic_comp_swap", src_comp=[4, 1, 1, 1], dest_comp=1)
 352 image("atomic_fadd",  src_comp=[1, 4, 1, 1], dest_comp=1)
 353 image("size",    dest_comp=0, flags=[CAN_ELIMINATE, CAN_REORDER])
 354 image("samples", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
 355
 356 # Intel-specific query for loading from the brw_image_param struct passed
 357 # into the shader as a uniform.  The variable is a deref to the image
 358 # variable. The const index specifies which of the six parameters to load.
 359 intrinsic("image_deref_load_param_intel", src_comp=[1], dest_comp=0,
 360           indices=[BASE], flags=[CAN_ELIMINATE, CAN_REORDER])
 361 image("load_raw_intel", src_comp=[1], dest_comp=0,
 362       flags=[CAN_ELIMINATE])
 363 image("store_raw_intel", src_comp=[1, 0])
 364
 365 # Vulkan descriptor set intrinsics
 366 #
 367 # The Vulkan API uses a different binding model from GL.  In the Vulkan
 368 # API, all external resources are represented by a tuple:
 369 #
 370 # (descriptor set, binding, array index)
 371 #
 372 # where the array index is the only thing allowed to be indirect.  The
 373 # vulkan_surface_index intrinsic takes the descriptor set and binding as
 374 # its first two indices and the array index as its source.  The third
 375 # index is a nir_variable_mode in case that's useful to the backend.
 376 #
 377 # The intended usage is that the shader will call vulkan_surface_index to
 378 # get an index and then pass that as the buffer index ubo/ssbo calls.
 379 #
 380 # The vulkan_resource_reindex intrinsic takes a resource index in src0
 381 # (the result of a vulkan_resource_index or vulkan_resource_reindex) which
 382 # corresponds to the tuple (set, binding, index) and computes an index
 383 # corresponding to tuple (set, binding, idx + src1).
 384 intrinsic("vulkan_resource_index", src_comp=[1], dest_comp=0,
 385           indices=[DESC_SET, BINDING, DESC_TYPE],
 386           flags=[CAN_ELIMINATE, CAN_REORDER])
 387 intrinsic("vulkan_resource_reindex", src_comp=[0, 1], dest_comp=0,
 388           indices=[DESC_TYPE], flags=[CAN_ELIMINATE, CAN_REORDER])
 389 intrinsic("load_vulkan_descriptor", src_comp=[-1], dest_comp=0,
 390           indices=[DESC_TYPE], flags=[CAN_ELIMINATE, CAN_REORDER])
 391
 392 # variable atomic intrinsics
 393 #
 394 # All of these variable atomic memory operations read a value from memory,
 395 # compute a new value using one of the operations below, write the new value
 396 # to memory, and return the original value read.
 397 #
 398 # All operations take 2 sources except CompSwap that takes 3. These sources
 399 # represent:
 400 #
 401 # 0: A deref to the memory on which to perform the atomic
 402 # 1: The data parameter to the atomic function (i.e. the value to add
 403 #    in shared_atomic_add, etc).
 404 # 2: For CompSwap only: the second data parameter.
 405 intrinsic("deref_atomic_add",  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 406 intrinsic("deref_atomic_imin", src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 407 intrinsic("deref_atomic_umin", src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 408 intrinsic("deref_atomic_imax", src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 409 intrinsic("deref_atomic_umax", src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 410 intrinsic("deref_atomic_and",  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 411 intrinsic("deref_atomic_or",   src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 412 intrinsic("deref_atomic_xor",  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 413 intrinsic("deref_atomic_exchange", src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 414 intrinsic("deref_atomic_comp_swap", src_comp=[-1, 1, 1], dest_comp=1, indices=[ACCESS])
 415 intrinsic("deref_atomic_fadd",  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 416 intrinsic("deref_atomic_fmin",  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 417 intrinsic("deref_atomic_fmax",  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 418 intrinsic("deref_atomic_fcomp_swap", src_comp=[-1, 1, 1], dest_comp=1, indices=[ACCESS])
 419
 420 # SSBO atomic intrinsics
 421 #
 422 # All of the SSBO atomic memory operations read a value from memory,
 423 # compute a new value using one of the operations below, write the new
 424 # value to memory, and return the original value read.
 425 #
 426 # All operations take 3 sources except CompSwap that takes 4. These
 427 # sources represent:
 428 #
 429 # 0: The SSBO buffer index.
 430 # 1: The offset into the SSBO buffer of the variable that the atomic
 431 #    operation will operate on.
 432 # 2: The data parameter to the atomic function (i.e. the value to add
 433 #    in ssbo_atomic_add, etc).
 434 # 3: For CompSwap only: the second data parameter.
 435 intrinsic("ssbo_atomic_add",  src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 436 intrinsic("ssbo_atomic_imin", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 437 intrinsic("ssbo_atomic_umin", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 438 intrinsic("ssbo_atomic_imax", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 439 intrinsic("ssbo_atomic_umax", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 440 intrinsic("ssbo_atomic_and",  src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 441 intrinsic("ssbo_atomic_or",   src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 442 intrinsic("ssbo_atomic_xor",  src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 443 intrinsic("ssbo_atomic_exchange", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 444 intrinsic("ssbo_atomic_comp_swap", src_comp=[1, 1, 1, 1], dest_comp=1, indices=[ACCESS])
 445 intrinsic("ssbo_atomic_fadd", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 446 intrinsic("ssbo_atomic_fmin", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 447 intrinsic("ssbo_atomic_fmax", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 448 intrinsic("ssbo_atomic_fcomp_swap", src_comp=[1, 1, 1, 1], dest_comp=1, indices=[ACCESS])
 449
 450 # CS shared variable atomic intrinsics
 451 #
 452 # All of the shared variable atomic memory operations read a value from
 453 # memory, compute a new value using one of the operations below, write the
 454 # new value to memory, and return the original value read.
 455 #
 456 # All operations take 2 sources except CompSwap that takes 3. These
 457 # sources represent:
 458 #
 459 # 0: The offset into the shared variable storage region that the atomic
 460 #    operation will operate on.
 461 # 1: The data parameter to the atomic function (i.e. the value to add
 462 #    in shared_atomic_add, etc).
 463 # 2: For CompSwap only: the second data parameter.
 464 intrinsic("shared_atomic_add",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 465 intrinsic("shared_atomic_imin", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 466 intrinsic("shared_atomic_umin", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 467 intrinsic("shared_atomic_imax", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 468 intrinsic("shared_atomic_umax", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 469 intrinsic("shared_atomic_and",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 470 intrinsic("shared_atomic_or",   src_comp=[1, 1], dest_comp=1, indices=[BASE])
 471 intrinsic("shared_atomic_xor",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 472 intrinsic("shared_atomic_exchange", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 473 intrinsic("shared_atomic_comp_swap", src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 474 intrinsic("shared_atomic_fadd",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 475 intrinsic("shared_atomic_fmin",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 476 intrinsic("shared_atomic_fmax",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 477 intrinsic("shared_atomic_fcomp_swap", src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 478
 479 # Global atomic intrinsics
 480 #
 481 # All of the shared variable atomic memory operations read a value from
 482 # memory, compute a new value using one of the operations below, write the
 483 # new value to memory, and return the original value read.
 484 #
 485 # All operations take 2 sources except CompSwap that takes 3. These
 486 # sources represent:
 487 #
 488 # 0: The memory address that the atomic operation will operate on.
 489 # 1: The data parameter to the atomic function (i.e. the value to add
 490 #    in shared_atomic_add, etc).
 491 # 2: For CompSwap only: the second data parameter.
 492 intrinsic("global_atomic_add",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 493 intrinsic("global_atomic_imin", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 494 intrinsic("global_atomic_umin", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 495 intrinsic("global_atomic_imax", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 496 intrinsic("global_atomic_umax", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 497 intrinsic("global_atomic_and",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 498 intrinsic("global_atomic_or",   src_comp=[1, 1], dest_comp=1, indices=[BASE])
 499 intrinsic("global_atomic_xor",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 500 intrinsic("global_atomic_exchange", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 501 intrinsic("global_atomic_comp_swap", src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 502 intrinsic("global_atomic_fadd",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 503 intrinsic("global_atomic_fmin",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 504 intrinsic("global_atomic_fmax",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 505 intrinsic("global_atomic_fcomp_swap", src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 506
 507 def system_value(name, dest_comp, indices=[], bit_sizes=[32]):
 508     intrinsic("load_" + name, [], dest_comp, indices,
 509               flags=[CAN_ELIMINATE, CAN_REORDER], sysval=True,
 510               bit_sizes=bit_sizes)
 511
 512 system_value("frag_coord", 4)
 513 system_value("front_face", 1, bit_sizes=[1, 32])
 514 system_value("vertex_id", 1)
 515 system_value("vertex_id_zero_base", 1)
 516 system_value("first_vertex", 1)
 517 system_value("is_indexed_draw", 1)
 518 system_value("base_vertex", 1)
 519 system_value("instance_id", 1)
 520 system_value("base_instance", 1)
 521 system_value("draw_id", 1)
 522 system_value("sample_id", 1)
 523 # sample_id_no_per_sample is like sample_id but does not imply per-
 524 # sample shading.  See the lower_helper_invocation option.
 525 system_value("sample_id_no_per_sample", 1)
 526 system_value("sample_pos", 2)
 527 system_value("sample_mask_in", 1)
 528 system_value("primitive_id", 1)
 529 system_value("invocation_id", 1)
 530 system_value("tess_coord", 3)
 531 system_value("tess_level_outer", 4)
 532 system_value("tess_level_inner", 2)
 533 system_value("patch_vertices_in", 1)
 534 system_value("local_invocation_id", 3)
 535 system_value("local_invocation_index", 1)
 536 system_value("work_group_id", 3)
 537 system_value("user_clip_plane", 4, indices=[UCP_ID])
 538 system_value("num_work_groups", 3)
 539 system_value("helper_invocation", 1, bit_sizes=[1, 32])
 540 system_value("alpha_ref_float", 1)
 541 system_value("layer_id", 1)
 542 system_value("view_index", 1)
 543 system_value("subgroup_size", 1)
 544 system_value("subgroup_invocation", 1)
 545 system_value("subgroup_eq_mask", 0, bit_sizes=[32, 64])
 546 system_value("subgroup_ge_mask", 0, bit_sizes=[32, 64])
 547 system_value("subgroup_gt_mask", 0, bit_sizes=[32, 64])
 548 system_value("subgroup_le_mask", 0, bit_sizes=[32, 64])
 549 system_value("subgroup_lt_mask", 0, bit_sizes=[32, 64])
 550 system_value("num_subgroups", 1)
 551 system_value("subgroup_id", 1)
 552 system_value("local_group_size", 3)
 553 system_value("global_invocation_id", 3, bit_sizes=[32, 64])
 554 system_value("global_invocation_index", 1, bit_sizes=[32, 64])
 555 system_value("work_dim", 1)
 556 # Driver-specific viewport scale/offset parameters.
 557 #
 558 # VC4 and V3D need to emit a scaled version of the position in the vertex
 559 # shaders for binning, and having system values lets us move the math for that
 560 # into NIR.
 561 #
 562 # Panfrost needs to implement all coordinate transformation in the
 563 # vertex shader; system values allow us to share this routine in NIR.
 564 system_value("viewport_x_scale", 1)
 565 system_value("viewport_y_scale", 1)
 566 system_value("viewport_z_scale", 1)
 567 system_value("viewport_z_offset", 1)
 568 system_value("viewport_scale", 3)
 569 system_value("viewport_offset", 3)
 570
 571 # Blend constant color values.  Float values are clamped. Vectored versions are
 572 # provided as well for driver convenience
 573
 574 system_value("blend_const_color_r_float", 1)
 575 system_value("blend_const_color_g_float", 1)
 576 system_value("blend_const_color_b_float", 1)
 577 system_value("blend_const_color_a_float", 1)
 578 system_value("blend_const_color_rgba", 4)
 579 system_value("blend_const_color_rgba8888_unorm", 1)
 580 system_value("blend_const_color_aaaa8888_unorm", 1)
 581
 582 # Barycentric coordinate intrinsics.
 583 #
 584 # These set up the barycentric coordinates for a particular interpolation.
 585 # The first three are for the simple cases: pixel, centroid, or per-sample
 586 # (at gl_SampleID).  The next two handle interpolating at a specified
 587 # sample location, or interpolating with a vec2 offset,
 588 #
 589 # The interp_mode index should be either the INTERP_MODE_SMOOTH or
 590 # INTERP_MODE_NOPERSPECTIVE enum values.
 591 #
 592 # The vec2 value produced by these intrinsics is intended for use as the
 593 # barycoord source of a load_interpolated_input intrinsic.
 594
 595 def barycentric(name, src_comp=[]):
 596     intrinsic("load_barycentric_" + name, src_comp=src_comp, dest_comp=2,
 597               indices=[INTERP_MODE], flags=[CAN_ELIMINATE, CAN_REORDER])
 598
 599 # no sources.
 600 barycentric("pixel")
 601 barycentric("centroid")
 602 barycentric("sample")
 603 # src[] = { sample_id }.
 604 barycentric("at_sample", [1])
 605 # src[] = { offset.xy }.
 606 barycentric("at_offset", [2])
 607
 608 # Load sample position:
 609 #
 610 # Takes a sample # and returns a sample position.  Used for lowering
 611 # interpolateAtSample() to interpolateAtOffset()
 612 intrinsic("load_sample_pos_from_id", src_comp=[1], dest_comp=2,
 613           flags=[CAN_ELIMINATE, CAN_REORDER])
 614
 615 # Loads what I believe is the primitive size, for scaling ij to pixel size:
 616 intrinsic("load_size_ir3", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
 617
 618 # Load operations pull data from some piece of GPU memory.  All load
 619 # operations operate in terms of offsets into some piece of theoretical
 620 # memory.  Loads from externally visible memory (UBO and SSBO) simply take a
 621 # byte offset as a source.  Loads from opaque memory (uniforms, inputs, etc.)
 622 # take a base+offset pair where the nir_intrinsic_base() gives the location
 623 # of the start of the variable being loaded and and the offset source is a
 624 # offset into that variable.
 625 #
 626 # Uniform load operations have a nir_intrinsic_range() index that specifies the
 627 # range (starting at base) of the data from which we are loading.  If
 628 # range == 0, then the range is unknown.
 629 #
 630 # Some load operations such as UBO/SSBO load and per_vertex loads take an
 631 # additional source to specify which UBO/SSBO/vertex to load from.
 632 #
 633 # The exact address type depends on the lowering pass that generates the
 634 # load/store intrinsics.  Typically, this is vec4 units for things such as
 635 # varying slots and float units for fragment shader inputs.  UBO and SSBO
 636 # offsets are always in bytes.
 637
 638 def load(name, num_srcs, indices=[], flags=[]):
 639     intrinsic("load_" + name, [1] * num_srcs, dest_comp=0, indices=indices,
 640               flags=flags)
 641
 642 # src[] = { offset }.
 643 load("uniform", 1, [BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
 644 # src[] = { buffer_index, offset }.
 645 load("ubo", 2, [ACCESS, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN_ELIMINATE, CAN_REORDER])
 646 # src[] = { offset }.
 647 load("input", 1, [BASE, COMPONENT], [CAN_ELIMINATE, CAN_REORDER])
 648 # src[] = { vertex, offset }.
 649 load("per_vertex_input", 2, [BASE, COMPONENT], [CAN_ELIMINATE, CAN_REORDER])
 650 # src[] = { barycoord, offset }.
 651 intrinsic("load_interpolated_input", src_comp=[2, 1], dest_comp=0,
 652           indices=[BASE, COMPONENT], flags=[CAN_ELIMINATE, CAN_REORDER])
 653
 654 # src[] = { buffer_index, offset }.
 655 load("ssbo", 2, [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 656 # src[] = { offset }.
 657 load("output", 1, [BASE, COMPONENT], flags=[CAN_ELIMINATE])
 658 # src[] = { vertex, offset }.
 659 load("per_vertex_output", 2, [BASE, COMPONENT], [CAN_ELIMINATE])
 660 # src[] = { offset }.
 661 load("shared", 1, [BASE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 662 # src[] = { offset }.
 663 load("push_constant", 1, [BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
 664 # src[] = { offset }.
 665 load("constant", 1, [BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
 666 # src[] = { address }.
 667 load("global", 1, [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 668 # src[] = { address }.
 669 load("kernel_input", 1, [BASE, RANGE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE, CAN_REORDER])
 670 # src[] = { offset }.
 671 load("scratch", 1, [ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 672
 673 # Stores work the same way as loads, except now the first source is the value
 674 # to store and the second (and possibly third) source specify where to store
 675 # the value.  SSBO and shared memory stores also have a
 676 # nir_intrinsic_write_mask()
 677
 678 def store(name, num_srcs, indices=[], flags=[]):
 679     intrinsic("store_" + name, [0] + ([1] * (num_srcs - 1)), indices=indices, flags=flags)
 680
 681 # src[] = { value, offset }.
 682 store("output", 2, [BASE, WRMASK, COMPONENT])
 683 # src[] = { value, vertex, offset }.
 684 store("per_vertex_output", 3, [BASE, WRMASK, COMPONENT])
 685 # src[] = { value, block_index, offset }
 686 store("ssbo", 3, [WRMASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
 687 # src[] = { value, offset }.
 688 store("shared", 2, [BASE, WRMASK, ALIGN_MUL, ALIGN_OFFSET])
 689 # src[] = { value, address }.
 690 store("global", 2, [WRMASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
 691 # src[] = { value, offset }.
 692 store("scratch", 2, [ALIGN_MUL, ALIGN_OFFSET, WRMASK])
 693
 694 # IR3-specific version of most SSBO intrinsics. The only different
 695 # compare to the originals is that they add an extra source to hold
 696 # the dword-offset, which is needed by the backend code apart from
 697 # the byte-offset already provided by NIR in one of the sources.
 698 #
 699 # NIR lowering pass 'ir3_nir_lower_io_offset' will replace the
 700 # original SSBO intrinsics by these, placing the computed
 701 # dword-offset always in the last source.
 702 #
 703 # The float versions are not handled because those are not supported
 704 # by the backend.
 705 intrinsic("store_ssbo_ir3",  src_comp=[0, 1, 1, 1],
 706           indices=[WRMASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
 707 intrinsic("load_ssbo_ir3",  src_comp=[1, 1, 1], dest_comp=0,
 708           indices=[ACCESS, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN_ELIMINATE])
 709 intrinsic("ssbo_atomic_add_ir3",        src_comp=[1, 1, 1, 1],    dest_comp=1)
 710 intrinsic("ssbo_atomic_imin_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1)
 711 intrinsic("ssbo_atomic_umin_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1)
 712 intrinsic("ssbo_atomic_imax_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1)
 713 intrinsic("ssbo_atomic_umax_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1)
 714 intrinsic("ssbo_atomic_and_ir3",        src_comp=[1, 1, 1, 1],    dest_comp=1)
 715 intrinsic("ssbo_atomic_or_ir3",         src_comp=[1, 1, 1, 1],    dest_comp=1)
 716 intrinsic("ssbo_atomic_xor_ir3",        src_comp=[1, 1, 1, 1],    dest_comp=1)
 717 intrinsic("ssbo_atomic_exchange_ir3",   src_comp=[1, 1, 1, 1],    dest_comp=1)
 718 intrinsic("ssbo_atomic_comp_swap_ir3",  src_comp=[1, 1, 1, 1, 1], dest_comp=1)