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 DST_ACCESS = "NIR_INTRINSIC_DST_ACCESS"
 115 SRC_ACCESS = "NIR_INTRINSIC_SRC_ACCESS"
 116 # Image format for image intrinsics
 117 FORMAT = "NIR_INTRINSIC_FORMAT"
 118 # Offset or address alignment
 119 ALIGN_MUL = "NIR_INTRINSIC_ALIGN_MUL"
 120 ALIGN_OFFSET = "NIR_INTRINSIC_ALIGN_OFFSET"
 121 # The vulkan descriptor type for vulkan_resource_index
 122 DESC_TYPE = "NIR_INTRINSIC_DESC_TYPE"
 123 # The nir_alu_type of a uniform/input/output
 124 TYPE = "NIR_INTRINSIC_TYPE"
 125 # The swizzle mask for quad_swizzle_amd & masked_swizzle_amd
 126 SWIZZLE_MASK = "NIR_INTRINSIC_SWIZZLE_MASK"
 127
 128 #
 129 # Possible flags:
 130 #
 131
 132 CAN_ELIMINATE = "NIR_INTRINSIC_CAN_ELIMINATE"
 133 CAN_REORDER   = "NIR_INTRINSIC_CAN_REORDER"
 134
 135 INTR_OPCODES = {}
 136
 137 # Defines a new NIR intrinsic.  By default, the intrinsic will have no sources
 138 # and no destination.
 139 #
 140 # You can set dest_comp=n to enable a destination for the intrinsic, in which
 141 # case it will have that many components, or =0 for "as many components as the
 142 # NIR destination value."
 143 #
 144 # Set src_comp=n to enable sources for the intruction.  It can be an array of
 145 # component counts, or (for convenience) a scalar component count if there's
 146 # only one source.  If a component count is 0, it will be as many components as
 147 # the intrinsic has based on the dest_comp.
 148 def intrinsic(name, src_comp=[], dest_comp=-1, indices=[],
 149               flags=[], sysval=False, bit_sizes=[]):
 150     assert name not in INTR_OPCODES
 151     INTR_OPCODES[name] = Intrinsic(name, src_comp, dest_comp,
 152                                    indices, flags, sysval, bit_sizes)
 153
 154 intrinsic("nop", flags=[CAN_ELIMINATE])
 155
 156 intrinsic("load_param", dest_comp=0, indices=[PARAM_IDX], flags=[CAN_ELIMINATE])
 157
 158 intrinsic("load_deref", dest_comp=0, src_comp=[-1],
 159           indices=[ACCESS], flags=[CAN_ELIMINATE])
 160 intrinsic("store_deref", src_comp=[-1, 0], indices=[WRMASK, ACCESS])
 161 intrinsic("copy_deref", src_comp=[-1, -1], indices=[DST_ACCESS, SRC_ACCESS])
 162
 163 # Interpolation of input.  The interp_deref_at* intrinsics are similar to the
 164 # load_var intrinsic acting on a shader input except that they interpolate the
 165 # input differently.  The at_sample and at_offset intrinsics take an
 166 # additional source that is an integer sample id or a vec2 position offset
 167 # respectively.
 168
 169 intrinsic("interp_deref_at_centroid", dest_comp=0, src_comp=[1],
 170           flags=[ CAN_ELIMINATE, CAN_REORDER])
 171 intrinsic("interp_deref_at_sample", src_comp=[1, 1], dest_comp=0,
 172           flags=[CAN_ELIMINATE, CAN_REORDER])
 173 intrinsic("interp_deref_at_offset", src_comp=[1, 2], dest_comp=0,
 174           flags=[CAN_ELIMINATE, CAN_REORDER])
 175
 176 # Gets the length of an unsized array at the end of a buffer
 177 intrinsic("deref_buffer_array_length", src_comp=[-1], dest_comp=1,
 178           flags=[CAN_ELIMINATE, CAN_REORDER])
 179
 180 # Ask the driver for the size of a given buffer. It takes the buffer index
 181 # as source.
 182 intrinsic("get_buffer_size", src_comp=[-1], dest_comp=1,
 183           flags=[CAN_ELIMINATE, CAN_REORDER])
 184
 185 # a barrier is an intrinsic with no inputs/outputs but which can't be moved
 186 # around/optimized in general
 187 def barrier(name):
 188     intrinsic(name)
 189
 190 barrier("barrier")
 191 barrier("discard")
 192
 193 # Demote fragment shader invocation to a helper invocation.  Any stores to
 194 # memory after this instruction are suppressed and the fragment does not write
 195 # outputs to the framebuffer.  Unlike discard, demote needs to ensure that
 196 # derivatives will still work for invocations that were not demoted.
 197 #
 198 # As specified by SPV_EXT_demote_to_helper_invocation.
 199 barrier("demote")
 200 intrinsic("is_helper_invocation", dest_comp=1, flags=[CAN_ELIMINATE])
 201
 202
 203 # Memory barrier with semantics analogous to the memoryBarrier() GLSL
 204 # intrinsic.
 205 barrier("memory_barrier")
 206
 207 # Shader clock intrinsic with semantics analogous to the clock2x32ARB()
 208 # GLSL intrinsic.
 209 # The latter can be used as code motion barrier, which is currently not
 210 # feasible with NIR.
 211 intrinsic("shader_clock", dest_comp=2, flags=[CAN_ELIMINATE])
 212
 213 # Shader ballot intrinsics with semantics analogous to the
 214 #
 215 #    ballotARB()
 216 #    readInvocationARB()
 217 #    readFirstInvocationARB()
 218 #
 219 # GLSL functions from ARB_shader_ballot.
 220 intrinsic("ballot", src_comp=[1], dest_comp=0, flags=[CAN_ELIMINATE])
 221 intrinsic("read_invocation", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 222 intrinsic("read_first_invocation", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 223
 224 # Additional SPIR-V ballot intrinsics
 225 #
 226 # These correspond to the SPIR-V opcodes
 227 #
 228 #    OpGroupUniformElect
 229 #    OpSubgroupFirstInvocationKHR
 230 intrinsic("elect", dest_comp=1, flags=[CAN_ELIMINATE])
 231 intrinsic("first_invocation", dest_comp=1, flags=[CAN_ELIMINATE])
 232
 233 # Memory barrier with semantics analogous to the compute shader
 234 # groupMemoryBarrier(), memoryBarrierAtomicCounter(), memoryBarrierBuffer(),
 235 # memoryBarrierImage() and memoryBarrierShared() GLSL intrinsics.
 236 barrier("group_memory_barrier")
 237 barrier("memory_barrier_atomic_counter")
 238 barrier("memory_barrier_buffer")
 239 barrier("memory_barrier_image")
 240 barrier("memory_barrier_shared")
 241 barrier("begin_invocation_interlock")
 242 barrier("end_invocation_interlock")
 243
 244 # A conditional discard, with a single boolean source.
 245 intrinsic("discard_if", src_comp=[1])
 246
 247 # ARB_shader_group_vote intrinsics
 248 intrinsic("vote_any", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
 249 intrinsic("vote_all", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
 250 intrinsic("vote_feq", src_comp=[0], dest_comp=1, flags=[CAN_ELIMINATE])
 251 intrinsic("vote_ieq", src_comp=[0], dest_comp=1, flags=[CAN_ELIMINATE])
 252
 253 # Ballot ALU operations from SPIR-V.
 254 #
 255 # These operations work like their ALU counterparts except that the operate
 256 # on a uvec4 which is treated as a 128bit integer.  Also, they are, in
 257 # general, free to ignore any bits which are above the subgroup size.
 258 intrinsic("ballot_bitfield_extract", src_comp=[4, 1], dest_comp=1, flags=[CAN_ELIMINATE])
 259 intrinsic("ballot_bit_count_reduce", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 260 intrinsic("ballot_bit_count_inclusive", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 261 intrinsic("ballot_bit_count_exclusive", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 262 intrinsic("ballot_find_lsb", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 263 intrinsic("ballot_find_msb", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 264
 265 # Shuffle operations from SPIR-V.
 266 intrinsic("shuffle", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 267 intrinsic("shuffle_xor", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 268 intrinsic("shuffle_up", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 269 intrinsic("shuffle_down", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 270
 271 # Quad operations from SPIR-V.
 272 intrinsic("quad_broadcast", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 273 intrinsic("quad_swap_horizontal", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 274 intrinsic("quad_swap_vertical", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 275 intrinsic("quad_swap_diagonal", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 276
 277 intrinsic("reduce", src_comp=[0], dest_comp=0, indices=[REDUCTION_OP, CLUSTER_SIZE],
 278           flags=[CAN_ELIMINATE])
 279 intrinsic("inclusive_scan", src_comp=[0], dest_comp=0, indices=[REDUCTION_OP],
 280           flags=[CAN_ELIMINATE])
 281 intrinsic("exclusive_scan", src_comp=[0], dest_comp=0, indices=[REDUCTION_OP],
 282           flags=[CAN_ELIMINATE])
 283
 284 # AMD shader ballot operations
 285 intrinsic("quad_swizzle_amd", src_comp=[0], dest_comp=0, indices=[SWIZZLE_MASK],
 286           flags=[CAN_ELIMINATE])
 287 intrinsic("masked_swizzle_amd", src_comp=[0], dest_comp=0, indices=[SWIZZLE_MASK],
 288           flags=[CAN_ELIMINATE])
 289 intrinsic("write_invocation_amd", src_comp=[0, 0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 290 intrinsic("mbcnt_amd", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
 291
 292 # Basic Geometry Shader intrinsics.
 293 #
 294 # emit_vertex implements GLSL's EmitStreamVertex() built-in.  It takes a single
 295 # index, which is the stream ID to write to.
 296 #
 297 # end_primitive implements GLSL's EndPrimitive() built-in.
 298 intrinsic("emit_vertex",   indices=[STREAM_ID])
 299 intrinsic("end_primitive", indices=[STREAM_ID])
 300
 301 # Geometry Shader intrinsics with a vertex count.
 302 #
 303 # Alternatively, drivers may implement these intrinsics, and use
 304 # nir_lower_gs_intrinsics() to convert from the basic intrinsics.
 305 #
 306 # These maintain a count of the number of vertices emitted, as an additional
 307 # unsigned integer source.
 308 intrinsic("emit_vertex_with_counter", src_comp=[1], indices=[STREAM_ID])
 309 intrinsic("end_primitive_with_counter", src_comp=[1], indices=[STREAM_ID])
 310 intrinsic("set_vertex_count", src_comp=[1])
 311
 312 # Atomic counters
 313 #
 314 # The *_var variants take an atomic_uint nir_variable, while the other,
 315 # lowered, variants take a constant buffer index and register offset.
 316
 317 def atomic(name, flags=[]):
 318     intrinsic(name + "_deref", src_comp=[-1], dest_comp=1, flags=flags)
 319     intrinsic(name, src_comp=[1], dest_comp=1, indices=[BASE], flags=flags)
 320
 321 def atomic2(name):
 322     intrinsic(name + "_deref", src_comp=[-1, 1], dest_comp=1)
 323     intrinsic(name, src_comp=[1, 1], dest_comp=1, indices=[BASE])
 324
 325 def atomic3(name):
 326     intrinsic(name + "_deref", src_comp=[-1, 1, 1], dest_comp=1)
 327     intrinsic(name, src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 328
 329 atomic("atomic_counter_inc")
 330 atomic("atomic_counter_pre_dec")
 331 atomic("atomic_counter_post_dec")
 332 atomic("atomic_counter_read", flags=[CAN_ELIMINATE])
 333 atomic2("atomic_counter_add")
 334 atomic2("atomic_counter_min")
 335 atomic2("atomic_counter_max")
 336 atomic2("atomic_counter_and")
 337 atomic2("atomic_counter_or")
 338 atomic2("atomic_counter_xor")
 339 atomic2("atomic_counter_exchange")
 340 atomic3("atomic_counter_comp_swap")
 341
 342 # Image load, store and atomic intrinsics.
 343 #
 344 # All image intrinsics come in three versions.  One which take an image target
 345 # passed as a deref chain as the first source, one which takes an index as the
 346 # first source, and one which takes a bindless handle as the first source.
 347 # In the first version, the image variable contains the memory and layout
 348 # qualifiers that influence the semantics of the intrinsic.  In the second and
 349 # third, the image format and access qualifiers are provided as constant
 350 # indices.
 351 #
 352 # All image intrinsics take a four-coordinate vector and a sample index as
 353 # 2nd and 3rd sources, determining the location within the image that will be
 354 # accessed by the intrinsic.  Components not applicable to the image target
 355 # in use are undefined.  Image store takes an additional four-component
 356 # argument with the value to be written, and image atomic operations take
 357 # either one or two additional scalar arguments with the same meaning as in
 358 # the ARB_shader_image_load_store specification.
 359 def image(name, src_comp=[], **kwargs):
 360     intrinsic("image_deref_" + name, src_comp=[1] + src_comp,
 361               indices=[ACCESS], **kwargs)
 362     intrinsic("image_" + name, src_comp=[1] + src_comp,
 363               indices=[IMAGE_DIM, IMAGE_ARRAY, FORMAT, ACCESS], **kwargs)
 364     intrinsic("bindless_image_" + name, src_comp=[1] + src_comp,
 365               indices=[IMAGE_DIM, IMAGE_ARRAY, FORMAT, ACCESS], **kwargs)
 366
 367 image("load", src_comp=[4, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 368 image("store", src_comp=[4, 1, 0])
 369 image("atomic_add",  src_comp=[4, 1, 1], dest_comp=1)
 370 image("atomic_min",  src_comp=[4, 1, 1], dest_comp=1)
 371 image("atomic_max",  src_comp=[4, 1, 1], dest_comp=1)
 372 image("atomic_and",  src_comp=[4, 1, 1], dest_comp=1)
 373 image("atomic_or",   src_comp=[4, 1, 1], dest_comp=1)
 374 image("atomic_xor",  src_comp=[4, 1, 1], dest_comp=1)
 375 image("atomic_exchange",  src_comp=[4, 1, 1], dest_comp=1)
 376 image("atomic_comp_swap", src_comp=[4, 1, 1, 1], dest_comp=1)
 377 image("atomic_fadd",  src_comp=[1, 4, 1, 1], dest_comp=1)
 378 image("size",    dest_comp=0, flags=[CAN_ELIMINATE, CAN_REORDER])
 379 image("samples", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
 380
 381 # Intel-specific query for loading from the brw_image_param struct passed
 382 # into the shader as a uniform.  The variable is a deref to the image
 383 # variable. The const index specifies which of the six parameters to load.
 384 intrinsic("image_deref_load_param_intel", src_comp=[1], dest_comp=0,
 385           indices=[BASE], flags=[CAN_ELIMINATE, CAN_REORDER])
 386 image("load_raw_intel", src_comp=[1], dest_comp=0,
 387       flags=[CAN_ELIMINATE])
 388 image("store_raw_intel", src_comp=[1, 0])
 389
 390 # Vulkan descriptor set intrinsics
 391 #
 392 # The Vulkan API uses a different binding model from GL.  In the Vulkan
 393 # API, all external resources are represented by a tuple:
 394 #
 395 # (descriptor set, binding, array index)
 396 #
 397 # where the array index is the only thing allowed to be indirect.  The
 398 # vulkan_surface_index intrinsic takes the descriptor set and binding as
 399 # its first two indices and the array index as its source.  The third
 400 # index is a nir_variable_mode in case that's useful to the backend.
 401 #
 402 # The intended usage is that the shader will call vulkan_surface_index to
 403 # get an index and then pass that as the buffer index ubo/ssbo calls.
 404 #
 405 # The vulkan_resource_reindex intrinsic takes a resource index in src0
 406 # (the result of a vulkan_resource_index or vulkan_resource_reindex) which
 407 # corresponds to the tuple (set, binding, index) and computes an index
 408 # corresponding to tuple (set, binding, idx + src1).
 409 intrinsic("vulkan_resource_index", src_comp=[1], dest_comp=0,
 410           indices=[DESC_SET, BINDING, DESC_TYPE],
 411           flags=[CAN_ELIMINATE, CAN_REORDER])
 412 intrinsic("vulkan_resource_reindex", src_comp=[0, 1], dest_comp=0,
 413           indices=[DESC_TYPE], flags=[CAN_ELIMINATE, CAN_REORDER])
 414 intrinsic("load_vulkan_descriptor", src_comp=[-1], dest_comp=0,
 415           indices=[DESC_TYPE], flags=[CAN_ELIMINATE, CAN_REORDER])
 416
 417 # variable atomic intrinsics
 418 #
 419 # All of these variable atomic memory operations read a value from memory,
 420 # compute a new value using one of the operations below, write the new value
 421 # to memory, and return the original value read.
 422 #
 423 # All operations take 2 sources except CompSwap that takes 3. These sources
 424 # represent:
 425 #
 426 # 0: A deref to the memory on which to perform the atomic
 427 # 1: The data parameter to the atomic function (i.e. the value to add
 428 #    in shared_atomic_add, etc).
 429 # 2: For CompSwap only: the second data parameter.
 430 intrinsic("deref_atomic_add",  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 431 intrinsic("deref_atomic_imin", src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 432 intrinsic("deref_atomic_umin", src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 433 intrinsic("deref_atomic_imax", src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 434 intrinsic("deref_atomic_umax", src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 435 intrinsic("deref_atomic_and",  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 436 intrinsic("deref_atomic_or",   src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 437 intrinsic("deref_atomic_xor",  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 438 intrinsic("deref_atomic_exchange", src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 439 intrinsic("deref_atomic_comp_swap", src_comp=[-1, 1, 1], dest_comp=1, indices=[ACCESS])
 440 intrinsic("deref_atomic_fadd",  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 441 intrinsic("deref_atomic_fmin",  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 442 intrinsic("deref_atomic_fmax",  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 443 intrinsic("deref_atomic_fcomp_swap", src_comp=[-1, 1, 1], dest_comp=1, indices=[ACCESS])
 444
 445 # SSBO atomic intrinsics
 446 #
 447 # All of the SSBO atomic memory operations read a value from memory,
 448 # compute a new value using one of the operations below, write the new
 449 # value to memory, and return the original value read.
 450 #
 451 # All operations take 3 sources except CompSwap that takes 4. These
 452 # sources represent:
 453 #
 454 # 0: The SSBO buffer index.
 455 # 1: The offset into the SSBO buffer of the variable that the atomic
 456 #    operation will operate on.
 457 # 2: The data parameter to the atomic function (i.e. the value to add
 458 #    in ssbo_atomic_add, etc).
 459 # 3: For CompSwap only: the second data parameter.
 460 intrinsic("ssbo_atomic_add",  src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 461 intrinsic("ssbo_atomic_imin", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 462 intrinsic("ssbo_atomic_umin", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 463 intrinsic("ssbo_atomic_imax", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 464 intrinsic("ssbo_atomic_umax", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 465 intrinsic("ssbo_atomic_and",  src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 466 intrinsic("ssbo_atomic_or",   src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 467 intrinsic("ssbo_atomic_xor",  src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 468 intrinsic("ssbo_atomic_exchange", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 469 intrinsic("ssbo_atomic_comp_swap", src_comp=[1, 1, 1, 1], dest_comp=1, indices=[ACCESS])
 470 intrinsic("ssbo_atomic_fadd", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 471 intrinsic("ssbo_atomic_fmin", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 472 intrinsic("ssbo_atomic_fmax", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 473 intrinsic("ssbo_atomic_fcomp_swap", src_comp=[1, 1, 1, 1], dest_comp=1, indices=[ACCESS])
 474
 475 # CS shared variable atomic intrinsics
 476 #
 477 # All of the shared variable atomic memory operations read a value from
 478 # memory, compute a new value using one of the operations below, write the
 479 # new value to memory, and return the original value read.
 480 #
 481 # All operations take 2 sources except CompSwap that takes 3. These
 482 # sources represent:
 483 #
 484 # 0: The offset into the shared variable storage region that the atomic
 485 #    operation will operate on.
 486 # 1: The data parameter to the atomic function (i.e. the value to add
 487 #    in shared_atomic_add, etc).
 488 # 2: For CompSwap only: the second data parameter.
 489 intrinsic("shared_atomic_add",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 490 intrinsic("shared_atomic_imin", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 491 intrinsic("shared_atomic_umin", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 492 intrinsic("shared_atomic_imax", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 493 intrinsic("shared_atomic_umax", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 494 intrinsic("shared_atomic_and",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 495 intrinsic("shared_atomic_or",   src_comp=[1, 1], dest_comp=1, indices=[BASE])
 496 intrinsic("shared_atomic_xor",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 497 intrinsic("shared_atomic_exchange", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 498 intrinsic("shared_atomic_comp_swap", src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 499 intrinsic("shared_atomic_fadd",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 500 intrinsic("shared_atomic_fmin",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 501 intrinsic("shared_atomic_fmax",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 502 intrinsic("shared_atomic_fcomp_swap", src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 503
 504 # Global atomic intrinsics
 505 #
 506 # All of the shared variable atomic memory operations read a value from
 507 # memory, compute a new value using one of the operations below, write the
 508 # new value to memory, and return the original value read.
 509 #
 510 # All operations take 2 sources except CompSwap that takes 3. These
 511 # sources represent:
 512 #
 513 # 0: The memory address that the atomic operation will operate on.
 514 # 1: The data parameter to the atomic function (i.e. the value to add
 515 #    in shared_atomic_add, etc).
 516 # 2: For CompSwap only: the second data parameter.
 517 intrinsic("global_atomic_add",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 518 intrinsic("global_atomic_imin", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 519 intrinsic("global_atomic_umin", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 520 intrinsic("global_atomic_imax", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 521 intrinsic("global_atomic_umax", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 522 intrinsic("global_atomic_and",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 523 intrinsic("global_atomic_or",   src_comp=[1, 1], dest_comp=1, indices=[BASE])
 524 intrinsic("global_atomic_xor",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 525 intrinsic("global_atomic_exchange", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 526 intrinsic("global_atomic_comp_swap", src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 527 intrinsic("global_atomic_fadd",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 528 intrinsic("global_atomic_fmin",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 529 intrinsic("global_atomic_fmax",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 530 intrinsic("global_atomic_fcomp_swap", src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 531
 532 def system_value(name, dest_comp, indices=[], bit_sizes=[32]):
 533     intrinsic("load_" + name, [], dest_comp, indices,
 534               flags=[CAN_ELIMINATE, CAN_REORDER], sysval=True,
 535               bit_sizes=bit_sizes)
 536
 537 system_value("frag_coord", 4)
 538 system_value("front_face", 1, bit_sizes=[1, 32])
 539 system_value("vertex_id", 1)
 540 system_value("vertex_id_zero_base", 1)
 541 system_value("first_vertex", 1)
 542 system_value("is_indexed_draw", 1)
 543 system_value("base_vertex", 1)
 544 system_value("instance_id", 1)
 545 system_value("base_instance", 1)
 546 system_value("draw_id", 1)
 547 system_value("sample_id", 1)
 548 # sample_id_no_per_sample is like sample_id but does not imply per-
 549 # sample shading.  See the lower_helper_invocation option.
 550 system_value("sample_id_no_per_sample", 1)
 551 system_value("sample_pos", 2)
 552 system_value("sample_mask_in", 1)
 553 system_value("primitive_id", 1)
 554 system_value("invocation_id", 1)
 555 system_value("tess_coord", 3)
 556 system_value("tess_level_outer", 4)
 557 system_value("tess_level_inner", 2)
 558 system_value("patch_vertices_in", 1)
 559 system_value("local_invocation_id", 3)
 560 system_value("local_invocation_index", 1)
 561 system_value("work_group_id", 3)
 562 system_value("user_clip_plane", 4, indices=[UCP_ID])
 563 system_value("num_work_groups", 3)
 564 system_value("helper_invocation", 1, bit_sizes=[1, 32])
 565 system_value("alpha_ref_float", 1)
 566 system_value("layer_id", 1)
 567 system_value("view_index", 1)
 568 system_value("subgroup_size", 1)
 569 system_value("subgroup_invocation", 1)
 570 system_value("subgroup_eq_mask", 0, bit_sizes=[32, 64])
 571 system_value("subgroup_ge_mask", 0, bit_sizes=[32, 64])
 572 system_value("subgroup_gt_mask", 0, bit_sizes=[32, 64])
 573 system_value("subgroup_le_mask", 0, bit_sizes=[32, 64])
 574 system_value("subgroup_lt_mask", 0, bit_sizes=[32, 64])
 575 system_value("num_subgroups", 1)
 576 system_value("subgroup_id", 1)
 577 system_value("local_group_size", 3)
 578 system_value("global_invocation_id", 3, bit_sizes=[32, 64])
 579 system_value("global_invocation_index", 1, bit_sizes=[32, 64])
 580 system_value("work_dim", 1)
 581 # Driver-specific viewport scale/offset parameters.
 582 #
 583 # VC4 and V3D need to emit a scaled version of the position in the vertex
 584 # shaders for binning, and having system values lets us move the math for that
 585 # into NIR.
 586 #
 587 # Panfrost needs to implement all coordinate transformation in the
 588 # vertex shader; system values allow us to share this routine in NIR.
 589 system_value("viewport_x_scale", 1)
 590 system_value("viewport_y_scale", 1)
 591 system_value("viewport_z_scale", 1)
 592 system_value("viewport_z_offset", 1)
 593 system_value("viewport_scale", 3)
 594 system_value("viewport_offset", 3)
 595
 596 # Blend constant color values.  Float values are clamped. Vectored versions are
 597 # provided as well for driver convenience
 598
 599 system_value("blend_const_color_r_float", 1)
 600 system_value("blend_const_color_g_float", 1)
 601 system_value("blend_const_color_b_float", 1)
 602 system_value("blend_const_color_a_float", 1)
 603 system_value("blend_const_color_rgba", 4)
 604 system_value("blend_const_color_rgba8888_unorm", 1)
 605 system_value("blend_const_color_aaaa8888_unorm", 1)
 606
 607 # System values for gl_Color, for radeonsi which interpolates these in the
 608 # shader prolog to handle two-sided color without recompiles and therefore
 609 # doesn't handle these in the main shader part like normal varyings.
 610 system_value("color0", 4)
 611 system_value("color1", 4)
 612
 613 # Barycentric coordinate intrinsics.
 614 #
 615 # These set up the barycentric coordinates for a particular interpolation.
 616 # The first three are for the simple cases: pixel, centroid, or per-sample
 617 # (at gl_SampleID).  The next two handle interpolating at a specified
 618 # sample location, or interpolating with a vec2 offset,
 619 #
 620 # The interp_mode index should be either the INTERP_MODE_SMOOTH or
 621 # INTERP_MODE_NOPERSPECTIVE enum values.
 622 #
 623 # The vec2 value produced by these intrinsics is intended for use as the
 624 # barycoord source of a load_interpolated_input intrinsic.
 625
 626 def barycentric(name, src_comp=[]):
 627     intrinsic("load_barycentric_" + name, src_comp=src_comp, dest_comp=2,
 628               indices=[INTERP_MODE], flags=[CAN_ELIMINATE, CAN_REORDER])
 629
 630 # no sources.
 631 barycentric("pixel")
 632 barycentric("centroid")
 633 barycentric("sample")
 634 # src[] = { sample_id }.
 635 barycentric("at_sample", [1])
 636 # src[] = { offset.xy }.
 637 barycentric("at_offset", [2])
 638
 639 # Load sample position:
 640 #
 641 # Takes a sample # and returns a sample position.  Used for lowering
 642 # interpolateAtSample() to interpolateAtOffset()
 643 intrinsic("load_sample_pos_from_id", src_comp=[1], dest_comp=2,
 644           flags=[CAN_ELIMINATE, CAN_REORDER])
 645
 646 # Loads what I believe is the primitive size, for scaling ij to pixel size:
 647 intrinsic("load_size_ir3", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
 648
 649 # Fragment shader input interpolation delta intrinsic.
 650 #
 651 # For hw where fragment shader input interpolation is handled in shader, the
 652 # load_fs_input_interp deltas intrinsics can be used to load the input deltas
 653 # used for interpolation as follows:
 654 #
 655 #    vec3 iid = load_fs_input_interp_deltas(varying_slot)
 656 #    vec2 bary = load_barycentric_*(...)
 657 #    float result = iid.x + iid.y * bary.y + iid.z * bary.x
 658
 659 intrinsic("load_fs_input_interp_deltas", src_comp=[1], dest_comp=3,
 660           indices=[BASE, COMPONENT], flags=[CAN_ELIMINATE, CAN_REORDER])
 661
 662 # Load operations pull data from some piece of GPU memory.  All load
 663 # operations operate in terms of offsets into some piece of theoretical
 664 # memory.  Loads from externally visible memory (UBO and SSBO) simply take a
 665 # byte offset as a source.  Loads from opaque memory (uniforms, inputs, etc.)
 666 # take a base+offset pair where the nir_intrinsic_base() gives the location
 667 # of the start of the variable being loaded and and the offset source is a
 668 # offset into that variable.
 669 #
 670 # Uniform load operations have a nir_intrinsic_range() index that specifies the
 671 # range (starting at base) of the data from which we are loading.  If
 672 # range == 0, then the range is unknown.
 673 #
 674 # Some load operations such as UBO/SSBO load and per_vertex loads take an
 675 # additional source to specify which UBO/SSBO/vertex to load from.
 676 #
 677 # The exact address type depends on the lowering pass that generates the
 678 # load/store intrinsics.  Typically, this is vec4 units for things such as
 679 # varying slots and float units for fragment shader inputs.  UBO and SSBO
 680 # offsets are always in bytes.
 681
 682 def load(name, num_srcs, indices=[], flags=[]):
 683     intrinsic("load_" + name, [1] * num_srcs, dest_comp=0, indices=indices,
 684               flags=flags)
 685
 686 # src[] = { offset }.
 687 load("uniform", 1, [BASE, RANGE, TYPE], [CAN_ELIMINATE, CAN_REORDER])
 688 # src[] = { buffer_index, offset }.
 689 load("ubo", 2, [ACCESS, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN_ELIMINATE, CAN_REORDER])
 690 # src[] = { offset }.
 691 load("input", 1, [BASE, COMPONENT, TYPE], [CAN_ELIMINATE, CAN_REORDER])
 692 # src[] = { vertex, offset }.
 693 load("per_vertex_input", 2, [BASE, COMPONENT], [CAN_ELIMINATE, CAN_REORDER])
 694 # src[] = { barycoord, offset }.
 695 intrinsic("load_interpolated_input", src_comp=[2, 1], dest_comp=0,
 696           indices=[BASE, COMPONENT], flags=[CAN_ELIMINATE, CAN_REORDER])
 697
 698 # src[] = { buffer_index, offset }.
 699 load("ssbo", 2, [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 700 # src[] = { offset }.
 701 load("output", 1, [BASE, COMPONENT], flags=[CAN_ELIMINATE])
 702 # src[] = { vertex, offset }.
 703 load("per_vertex_output", 2, [BASE, COMPONENT], [CAN_ELIMINATE])
 704 # src[] = { offset }.
 705 load("shared", 1, [BASE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 706 # src[] = { offset }.
 707 load("push_constant", 1, [BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
 708 # src[] = { offset }.
 709 load("constant", 1, [BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
 710 # src[] = { address }.
 711 load("global", 1, [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 712 # src[] = { address }.
 713 load("kernel_input", 1, [BASE, RANGE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE, CAN_REORDER])
 714 # src[] = { offset }.
 715 load("scratch", 1, [ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 716
 717 # Stores work the same way as loads, except now the first source is the value
 718 # to store and the second (and possibly third) source specify where to store
 719 # the value.  SSBO and shared memory stores also have a
 720 # nir_intrinsic_write_mask()
 721
 722 def store(name, num_srcs, indices=[], flags=[]):
 723     intrinsic("store_" + name, [0] + ([1] * (num_srcs - 1)), indices=indices, flags=flags)
 724
 725 # src[] = { value, offset }.
 726 store("output", 2, [BASE, WRMASK, COMPONENT, TYPE])
 727 # src[] = { value, vertex, offset }.
 728 store("per_vertex_output", 3, [BASE, WRMASK, COMPONENT])
 729 # src[] = { value, block_index, offset }
 730 store("ssbo", 3, [WRMASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
 731 # src[] = { value, offset }.
 732 store("shared", 2, [BASE, WRMASK, ALIGN_MUL, ALIGN_OFFSET])
 733 # src[] = { value, address }.
 734 store("global", 2, [WRMASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
 735 # src[] = { value, offset }.
 736 store("scratch", 2, [ALIGN_MUL, ALIGN_OFFSET, WRMASK])
 737
 738 # IR3-specific version of most SSBO intrinsics. The only different
 739 # compare to the originals is that they add an extra source to hold
 740 # the dword-offset, which is needed by the backend code apart from
 741 # the byte-offset already provided by NIR in one of the sources.
 742 #
 743 # NIR lowering pass 'ir3_nir_lower_io_offset' will replace the
 744 # original SSBO intrinsics by these, placing the computed
 745 # dword-offset always in the last source.
 746 #
 747 # The float versions are not handled because those are not supported
 748 # by the backend.
 749 intrinsic("store_ssbo_ir3",  src_comp=[0, 1, 1, 1],
 750           indices=[WRMASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
 751 intrinsic("load_ssbo_ir3",  src_comp=[1, 1, 1], dest_comp=0,
 752           indices=[ACCESS, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN_ELIMINATE])
 753 intrinsic("ssbo_atomic_add_ir3",        src_comp=[1, 1, 1, 1],    dest_comp=1)
 754 intrinsic("ssbo_atomic_imin_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1)
 755 intrinsic("ssbo_atomic_umin_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1)
 756 intrinsic("ssbo_atomic_imax_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1)
 757 intrinsic("ssbo_atomic_umax_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1)
 758 intrinsic("ssbo_atomic_and_ir3",        src_comp=[1, 1, 1, 1],    dest_comp=1)
 759 intrinsic("ssbo_atomic_or_ir3",         src_comp=[1, 1, 1, 1],    dest_comp=1)
 760 intrinsic("ssbo_atomic_xor_ir3",        src_comp=[1, 1, 1, 1],    dest_comp=1)
 761 intrinsic("ssbo_atomic_exchange_ir3",   src_comp=[1, 1, 1, 1],    dest_comp=1)
 762 intrinsic("ssbo_atomic_comp_swap_ir3",  src_comp=[1, 1, 1, 1, 1], dest_comp=1)