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 # Driver location of attribute
 128 DRIVER_LOCATION = "NIR_INTRINSIC_DRIVER_LOCATION"
 129 # Ordering and visibility of a memory operation
 130 MEMORY_SEMANTICS = "NIR_INTRINSIC_MEMORY_SEMANTICS"
 131 # Modes affected by a memory operation
 132 MEMORY_MODES = "NIR_INTRINSIC_MEMORY_MODES"
 133 # Scope of a memory operation
 134 MEMORY_SCOPE = "NIR_INTRINSIC_MEMORY_SCOPE"
 135 # Scope of a control barrier
 136 EXECUTION_SCOPE = "NIR_INTRINSIC_EXECUTION_SCOPE"
 137
 138 #
 139 # Possible flags:
 140 #
 141
 142 CAN_ELIMINATE = "NIR_INTRINSIC_CAN_ELIMINATE"
 143 CAN_REORDER   = "NIR_INTRINSIC_CAN_REORDER"
 144
 145 INTR_OPCODES = {}
 146
 147 # Defines a new NIR intrinsic.  By default, the intrinsic will have no sources
 148 # and no destination.
 149 #
 150 # You can set dest_comp=n to enable a destination for the intrinsic, in which
 151 # case it will have that many components, or =0 for "as many components as the
 152 # NIR destination value."
 153 #
 154 # Set src_comp=n to enable sources for the intruction.  It can be an array of
 155 # component counts, or (for convenience) a scalar component count if there's
 156 # only one source.  If a component count is 0, it will be as many components as
 157 # the intrinsic has based on the dest_comp.
 158 def intrinsic(name, src_comp=[], dest_comp=-1, indices=[],
 159               flags=[], sysval=False, bit_sizes=[]):
 160     assert name not in INTR_OPCODES
 161     INTR_OPCODES[name] = Intrinsic(name, src_comp, dest_comp,
 162                                    indices, flags, sysval, bit_sizes)
 163
 164 intrinsic("nop", flags=[CAN_ELIMINATE])
 165
 166 intrinsic("load_param", dest_comp=0, indices=[PARAM_IDX], flags=[CAN_ELIMINATE])
 167
 168 intrinsic("load_deref", dest_comp=0, src_comp=[-1],
 169           indices=[ACCESS], flags=[CAN_ELIMINATE])
 170 intrinsic("store_deref", src_comp=[-1, 0], indices=[WRMASK, ACCESS])
 171 intrinsic("copy_deref", src_comp=[-1, -1], indices=[DST_ACCESS, SRC_ACCESS])
 172
 173 # Interpolation of input.  The interp_deref_at* intrinsics are similar to the
 174 # load_var intrinsic acting on a shader input except that they interpolate the
 175 # input differently.  The at_sample, at_offset and at_vertex intrinsics take an
 176 # additional source that is an integer sample id, a vec2 position offset, or a
 177 # vertex ID respectively.
 178
 179 intrinsic("interp_deref_at_centroid", dest_comp=0, src_comp=[1],
 180           flags=[ CAN_ELIMINATE, CAN_REORDER])
 181 intrinsic("interp_deref_at_sample", src_comp=[1, 1], dest_comp=0,
 182           flags=[CAN_ELIMINATE, CAN_REORDER])
 183 intrinsic("interp_deref_at_offset", src_comp=[1, 2], dest_comp=0,
 184           flags=[CAN_ELIMINATE, CAN_REORDER])
 185 intrinsic("interp_deref_at_vertex", src_comp=[1, 1], dest_comp=0,
 186           flags=[CAN_ELIMINATE, CAN_REORDER])
 187
 188 # Gets the length of an unsized array at the end of a buffer
 189 intrinsic("deref_buffer_array_length", src_comp=[-1], dest_comp=1,
 190           flags=[CAN_ELIMINATE, CAN_REORDER])
 191
 192 # Ask the driver for the size of a given buffer. It takes the buffer index
 193 # as source.
 194 intrinsic("get_buffer_size", src_comp=[-1], dest_comp=1,
 195           flags=[CAN_ELIMINATE, CAN_REORDER])
 196
 197 # a barrier is an intrinsic with no inputs/outputs but which can't be moved
 198 # around/optimized in general
 199 def barrier(name):
 200     intrinsic(name)
 201
 202 barrier("discard")
 203
 204 # Demote fragment shader invocation to a helper invocation.  Any stores to
 205 # memory after this instruction are suppressed and the fragment does not write
 206 # outputs to the framebuffer.  Unlike discard, demote needs to ensure that
 207 # derivatives will still work for invocations that were not demoted.
 208 #
 209 # As specified by SPV_EXT_demote_to_helper_invocation.
 210 barrier("demote")
 211 intrinsic("is_helper_invocation", dest_comp=1, flags=[CAN_ELIMINATE])
 212
 213 # A workgroup-level control barrier.  Any thread which hits this barrier will
 214 # pause until all threads within the current workgroup have also hit the
 215 # barrier.  For compute shaders, the workgroup is defined as the local group.
 216 # For tessellation control shaders, the workgroup is defined as the current
 217 # patch.  This intrinsic does not imply any sort of memory barrier.
 218 barrier("control_barrier")
 219
 220 # Memory barrier with semantics analogous to the memoryBarrier() GLSL
 221 # intrinsic.
 222 barrier("memory_barrier")
 223
 224 # Control/Memory barrier with explicit scope.  Follows the semantics of SPIR-V
 225 # OpMemoryBarrier and OpControlBarrier, used to implement Vulkan Memory Model.
 226 # Storage that the barrier applies is represented using NIR variable modes.
 227 # For an OpMemoryBarrier, set EXECUTION_SCOPE to NIR_SCOPE_NONE.
 228 intrinsic("scoped_barrier",
 229           indices=[EXECUTION_SCOPE, MEMORY_SEMANTICS, MEMORY_MODES, MEMORY_SCOPE])
 230
 231 # Shader clock intrinsic with semantics analogous to the clock2x32ARB()
 232 # GLSL intrinsic.
 233 # The latter can be used as code motion barrier, which is currently not
 234 # feasible with NIR.
 235 intrinsic("shader_clock", dest_comp=2, flags=[CAN_ELIMINATE],
 236           indices=[MEMORY_SCOPE])
 237
 238 # Shader ballot intrinsics with semantics analogous to the
 239 #
 240 #    ballotARB()
 241 #    readInvocationARB()
 242 #    readFirstInvocationARB()
 243 #
 244 # GLSL functions from ARB_shader_ballot.
 245 intrinsic("ballot", src_comp=[1], dest_comp=0, flags=[CAN_ELIMINATE])
 246 intrinsic("read_invocation", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 247 intrinsic("read_first_invocation", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 248
 249 # Additional SPIR-V ballot intrinsics
 250 #
 251 # These correspond to the SPIR-V opcodes
 252 #
 253 #    OpGroupUniformElect
 254 #    OpSubgroupFirstInvocationKHR
 255 intrinsic("elect", dest_comp=1, flags=[CAN_ELIMINATE])
 256 intrinsic("first_invocation", dest_comp=1, flags=[CAN_ELIMINATE])
 257
 258 # Memory barrier with semantics analogous to the compute shader
 259 # groupMemoryBarrier(), memoryBarrierAtomicCounter(), memoryBarrierBuffer(),
 260 # memoryBarrierImage() and memoryBarrierShared() GLSL intrinsics.
 261 barrier("group_memory_barrier")
 262 barrier("memory_barrier_atomic_counter")
 263 barrier("memory_barrier_buffer")
 264 barrier("memory_barrier_image")
 265 barrier("memory_barrier_shared")
 266 barrier("begin_invocation_interlock")
 267 barrier("end_invocation_interlock")
 268
 269 # Memory barrier for synchronizing TCS patch outputs
 270 barrier("memory_barrier_tcs_patch")
 271
 272 # A conditional discard/demote, with a single boolean source.
 273 intrinsic("discard_if", src_comp=[1])
 274 intrinsic("demote_if", src_comp=[1])
 275
 276 # ARB_shader_group_vote intrinsics
 277 intrinsic("vote_any", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
 278 intrinsic("vote_all", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
 279 intrinsic("vote_feq", src_comp=[0], dest_comp=1, flags=[CAN_ELIMINATE])
 280 intrinsic("vote_ieq", src_comp=[0], dest_comp=1, flags=[CAN_ELIMINATE])
 281
 282 # Ballot ALU operations from SPIR-V.
 283 #
 284 # These operations work like their ALU counterparts except that the operate
 285 # on a uvec4 which is treated as a 128bit integer.  Also, they are, in
 286 # general, free to ignore any bits which are above the subgroup size.
 287 intrinsic("ballot_bitfield_extract", src_comp=[4, 1], dest_comp=1, flags=[CAN_ELIMINATE])
 288 intrinsic("ballot_bit_count_reduce", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 289 intrinsic("ballot_bit_count_inclusive", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 290 intrinsic("ballot_bit_count_exclusive", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 291 intrinsic("ballot_find_lsb", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 292 intrinsic("ballot_find_msb", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
 293
 294 # Shuffle operations from SPIR-V.
 295 intrinsic("shuffle", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 296 intrinsic("shuffle_xor", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 297 intrinsic("shuffle_up", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 298 intrinsic("shuffle_down", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 299
 300 # Quad operations from SPIR-V.
 301 intrinsic("quad_broadcast", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 302 intrinsic("quad_swap_horizontal", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 303 intrinsic("quad_swap_vertical", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 304 intrinsic("quad_swap_diagonal", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
 305
 306 intrinsic("reduce", src_comp=[0], dest_comp=0, indices=[REDUCTION_OP, CLUSTER_SIZE],
 307           flags=[CAN_ELIMINATE])
 308 intrinsic("inclusive_scan", src_comp=[0], dest_comp=0, indices=[REDUCTION_OP],
 309           flags=[CAN_ELIMINATE])
 310 intrinsic("exclusive_scan", src_comp=[0], dest_comp=0, indices=[REDUCTION_OP],
 311           flags=[CAN_ELIMINATE])
 312
 313 # AMD shader ballot operations
 314 intrinsic("quad_swizzle_amd", src_comp=[0], dest_comp=0, indices=[SWIZZLE_MASK],
 315           flags=[CAN_ELIMINATE])
 316 intrinsic("masked_swizzle_amd", src_comp=[0], dest_comp=0, indices=[SWIZZLE_MASK],
 317           flags=[CAN_ELIMINATE])
 318 intrinsic("write_invocation_amd", src_comp=[0, 0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 319 intrinsic("mbcnt_amd", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
 320
 321 # Basic Geometry Shader intrinsics.
 322 #
 323 # emit_vertex implements GLSL's EmitStreamVertex() built-in.  It takes a single
 324 # index, which is the stream ID to write to.
 325 #
 326 # end_primitive implements GLSL's EndPrimitive() built-in.
 327 intrinsic("emit_vertex",   indices=[STREAM_ID])
 328 intrinsic("end_primitive", indices=[STREAM_ID])
 329
 330 # Geometry Shader intrinsics with a vertex count.
 331 #
 332 # Alternatively, drivers may implement these intrinsics, and use
 333 # nir_lower_gs_intrinsics() to convert from the basic intrinsics.
 334 #
 335 # These maintain a count of the number of vertices emitted, as an additional
 336 # unsigned integer source.
 337 intrinsic("emit_vertex_with_counter", src_comp=[1], indices=[STREAM_ID])
 338 intrinsic("end_primitive_with_counter", src_comp=[1], indices=[STREAM_ID])
 339 intrinsic("set_vertex_count", src_comp=[1])
 340
 341 # Atomic counters
 342 #
 343 # The *_var variants take an atomic_uint nir_variable, while the other,
 344 # lowered, variants take a constant buffer index and register offset.
 345
 346 def atomic(name, flags=[]):
 347     intrinsic(name + "_deref", src_comp=[-1], dest_comp=1, flags=flags)
 348     intrinsic(name, src_comp=[1], dest_comp=1, indices=[BASE], flags=flags)
 349
 350 def atomic2(name):
 351     intrinsic(name + "_deref", src_comp=[-1, 1], dest_comp=1)
 352     intrinsic(name, src_comp=[1, 1], dest_comp=1, indices=[BASE])
 353
 354 def atomic3(name):
 355     intrinsic(name + "_deref", src_comp=[-1, 1, 1], dest_comp=1)
 356     intrinsic(name, src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 357
 358 atomic("atomic_counter_inc")
 359 atomic("atomic_counter_pre_dec")
 360 atomic("atomic_counter_post_dec")
 361 atomic("atomic_counter_read", flags=[CAN_ELIMINATE])
 362 atomic2("atomic_counter_add")
 363 atomic2("atomic_counter_min")
 364 atomic2("atomic_counter_max")
 365 atomic2("atomic_counter_and")
 366 atomic2("atomic_counter_or")
 367 atomic2("atomic_counter_xor")
 368 atomic2("atomic_counter_exchange")
 369 atomic3("atomic_counter_comp_swap")
 370
 371 # Image load, store and atomic intrinsics.
 372 #
 373 # All image intrinsics come in three versions.  One which take an image target
 374 # passed as a deref chain as the first source, one which takes an index as the
 375 # first source, and one which takes a bindless handle as the first source.
 376 # In the first version, the image variable contains the memory and layout
 377 # qualifiers that influence the semantics of the intrinsic.  In the second and
 378 # third, the image format and access qualifiers are provided as constant
 379 # indices.
 380 #
 381 # All image intrinsics take a four-coordinate vector and a sample index as
 382 # 2nd and 3rd sources, determining the location within the image that will be
 383 # accessed by the intrinsic.  Components not applicable to the image target
 384 # in use are undefined.  Image store takes an additional four-component
 385 # argument with the value to be written, and image atomic operations take
 386 # either one or two additional scalar arguments with the same meaning as in
 387 # the ARB_shader_image_load_store specification.
 388 def image(name, src_comp=[], **kwargs):
 389     intrinsic("image_deref_" + name, src_comp=[1] + src_comp,
 390               indices=[ACCESS], **kwargs)
 391     intrinsic("image_" + name, src_comp=[1] + src_comp,
 392               indices=[IMAGE_DIM, IMAGE_ARRAY, FORMAT, ACCESS], **kwargs)
 393     intrinsic("bindless_image_" + name, src_comp=[1] + src_comp,
 394               indices=[IMAGE_DIM, IMAGE_ARRAY, FORMAT, ACCESS], **kwargs)
 395
 396 image("load", src_comp=[4, 1, 1], dest_comp=0, flags=[CAN_ELIMINATE])
 397 image("store", src_comp=[4, 1, 0, 1])
 398 image("atomic_add",  src_comp=[4, 1, 1], dest_comp=1)
 399 image("atomic_imin",  src_comp=[4, 1, 1], dest_comp=1)
 400 image("atomic_umin",  src_comp=[4, 1, 1], dest_comp=1)
 401 image("atomic_imax",  src_comp=[4, 1, 1], dest_comp=1)
 402 image("atomic_umax",  src_comp=[4, 1, 1], dest_comp=1)
 403 image("atomic_and",  src_comp=[4, 1, 1], dest_comp=1)
 404 image("atomic_or",   src_comp=[4, 1, 1], dest_comp=1)
 405 image("atomic_xor",  src_comp=[4, 1, 1], dest_comp=1)
 406 image("atomic_exchange",  src_comp=[4, 1, 1], dest_comp=1)
 407 image("atomic_comp_swap", src_comp=[4, 1, 1, 1], dest_comp=1)
 408 image("atomic_fadd",  src_comp=[4, 1, 1], dest_comp=1)
 409 image("size",    dest_comp=0, flags=[CAN_ELIMINATE, CAN_REORDER])
 410 image("samples", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
 411 image("atomic_inc_wrap",  src_comp=[4, 1, 1], dest_comp=1)
 412 image("atomic_dec_wrap",  src_comp=[4, 1, 1], dest_comp=1)
 413
 414 # Intel-specific query for loading from the brw_image_param struct passed
 415 # into the shader as a uniform.  The variable is a deref to the image
 416 # variable. The const index specifies which of the six parameters to load.
 417 intrinsic("image_deref_load_param_intel", src_comp=[1], dest_comp=0,
 418           indices=[BASE], flags=[CAN_ELIMINATE, CAN_REORDER])
 419 image("load_raw_intel", src_comp=[1], dest_comp=0,
 420       flags=[CAN_ELIMINATE])
 421 image("store_raw_intel", src_comp=[1, 0])
 422
 423 # Vulkan descriptor set intrinsics
 424 #
 425 # The Vulkan API uses a different binding model from GL.  In the Vulkan
 426 # API, all external resources are represented by a tuple:
 427 #
 428 # (descriptor set, binding, array index)
 429 #
 430 # where the array index is the only thing allowed to be indirect.  The
 431 # vulkan_surface_index intrinsic takes the descriptor set and binding as
 432 # its first two indices and the array index as its source.  The third
 433 # index is a nir_variable_mode in case that's useful to the backend.
 434 #
 435 # The intended usage is that the shader will call vulkan_surface_index to
 436 # get an index and then pass that as the buffer index ubo/ssbo calls.
 437 #
 438 # The vulkan_resource_reindex intrinsic takes a resource index in src0
 439 # (the result of a vulkan_resource_index or vulkan_resource_reindex) which
 440 # corresponds to the tuple (set, binding, index) and computes an index
 441 # corresponding to tuple (set, binding, idx + src1).
 442 intrinsic("vulkan_resource_index", src_comp=[1], dest_comp=0,
 443           indices=[DESC_SET, BINDING, DESC_TYPE],
 444           flags=[CAN_ELIMINATE, CAN_REORDER])
 445 intrinsic("vulkan_resource_reindex", src_comp=[0, 1], dest_comp=0,
 446           indices=[DESC_TYPE], flags=[CAN_ELIMINATE, CAN_REORDER])
 447 intrinsic("load_vulkan_descriptor", src_comp=[-1], dest_comp=0,
 448           indices=[DESC_TYPE], flags=[CAN_ELIMINATE, CAN_REORDER])
 449
 450 # variable atomic intrinsics
 451 #
 452 # All of these variable atomic memory operations read a value from memory,
 453 # compute a new value using one of the operations below, write the new value
 454 # to memory, and return the original value read.
 455 #
 456 # All operations take 2 sources except CompSwap that takes 3. These sources
 457 # represent:
 458 #
 459 # 0: A deref to the memory on which to perform the atomic
 460 # 1: The data parameter to the atomic function (i.e. the value to add
 461 #    in shared_atomic_add, etc).
 462 # 2: For CompSwap only: the second data parameter.
 463 intrinsic("deref_atomic_add",  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 464 intrinsic("deref_atomic_imin", src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 465 intrinsic("deref_atomic_umin", src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 466 intrinsic("deref_atomic_imax", src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 467 intrinsic("deref_atomic_umax", src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 468 intrinsic("deref_atomic_and",  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 469 intrinsic("deref_atomic_or",   src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 470 intrinsic("deref_atomic_xor",  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 471 intrinsic("deref_atomic_exchange", src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 472 intrinsic("deref_atomic_comp_swap", src_comp=[-1, 1, 1], dest_comp=1, indices=[ACCESS])
 473 intrinsic("deref_atomic_fadd",  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 474 intrinsic("deref_atomic_fmin",  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 475 intrinsic("deref_atomic_fmax",  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
 476 intrinsic("deref_atomic_fcomp_swap", src_comp=[-1, 1, 1], dest_comp=1, indices=[ACCESS])
 477
 478 # SSBO atomic intrinsics
 479 #
 480 # All of the SSBO atomic memory operations read a value from memory,
 481 # compute a new value using one of the operations below, write the new
 482 # value to memory, and return the original value read.
 483 #
 484 # All operations take 3 sources except CompSwap that takes 4. These
 485 # sources represent:
 486 #
 487 # 0: The SSBO buffer index.
 488 # 1: The offset into the SSBO buffer of the variable that the atomic
 489 #    operation will operate on.
 490 # 2: The data parameter to the atomic function (i.e. the value to add
 491 #    in ssbo_atomic_add, etc).
 492 # 3: For CompSwap only: the second data parameter.
 493 intrinsic("ssbo_atomic_add",  src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 494 intrinsic("ssbo_atomic_imin", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 495 intrinsic("ssbo_atomic_umin", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 496 intrinsic("ssbo_atomic_imax", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 497 intrinsic("ssbo_atomic_umax", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 498 intrinsic("ssbo_atomic_and",  src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 499 intrinsic("ssbo_atomic_or",   src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 500 intrinsic("ssbo_atomic_xor",  src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 501 intrinsic("ssbo_atomic_exchange", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 502 intrinsic("ssbo_atomic_comp_swap", src_comp=[1, 1, 1, 1], dest_comp=1, indices=[ACCESS])
 503 intrinsic("ssbo_atomic_fadd", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 504 intrinsic("ssbo_atomic_fmin", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 505 intrinsic("ssbo_atomic_fmax", src_comp=[1, 1, 1], dest_comp=1, indices=[ACCESS])
 506 intrinsic("ssbo_atomic_fcomp_swap", src_comp=[1, 1, 1, 1], dest_comp=1, indices=[ACCESS])
 507
 508 # CS shared variable atomic intrinsics
 509 #
 510 # All of the shared variable atomic memory operations read a value from
 511 # memory, compute a new value using one of the operations below, write the
 512 # new value to memory, and return the original value read.
 513 #
 514 # All operations take 2 sources except CompSwap that takes 3. These
 515 # sources represent:
 516 #
 517 # 0: The offset into the shared variable storage region that the atomic
 518 #    operation will operate on.
 519 # 1: The data parameter to the atomic function (i.e. the value to add
 520 #    in shared_atomic_add, etc).
 521 # 2: For CompSwap only: the second data parameter.
 522 intrinsic("shared_atomic_add",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 523 intrinsic("shared_atomic_imin", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 524 intrinsic("shared_atomic_umin", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 525 intrinsic("shared_atomic_imax", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 526 intrinsic("shared_atomic_umax", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 527 intrinsic("shared_atomic_and",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 528 intrinsic("shared_atomic_or",   src_comp=[1, 1], dest_comp=1, indices=[BASE])
 529 intrinsic("shared_atomic_xor",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 530 intrinsic("shared_atomic_exchange", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 531 intrinsic("shared_atomic_comp_swap", src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 532 intrinsic("shared_atomic_fadd",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 533 intrinsic("shared_atomic_fmin",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 534 intrinsic("shared_atomic_fmax",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 535 intrinsic("shared_atomic_fcomp_swap", src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 536
 537 # Global atomic intrinsics
 538 #
 539 # All of the shared variable atomic memory operations read a value from
 540 # memory, compute a new value using one of the operations below, write the
 541 # new value to memory, and return the original value read.
 542 #
 543 # All operations take 2 sources except CompSwap that takes 3. These
 544 # sources represent:
 545 #
 546 # 0: The memory address that the atomic operation will operate on.
 547 # 1: The data parameter to the atomic function (i.e. the value to add
 548 #    in shared_atomic_add, etc).
 549 # 2: For CompSwap only: the second data parameter.
 550 intrinsic("global_atomic_add",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 551 intrinsic("global_atomic_imin", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 552 intrinsic("global_atomic_umin", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 553 intrinsic("global_atomic_imax", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 554 intrinsic("global_atomic_umax", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 555 intrinsic("global_atomic_and",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 556 intrinsic("global_atomic_or",   src_comp=[1, 1], dest_comp=1, indices=[BASE])
 557 intrinsic("global_atomic_xor",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 558 intrinsic("global_atomic_exchange", src_comp=[1, 1], dest_comp=1, indices=[BASE])
 559 intrinsic("global_atomic_comp_swap", src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 560 intrinsic("global_atomic_fadd",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 561 intrinsic("global_atomic_fmin",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 562 intrinsic("global_atomic_fmax",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 563 intrinsic("global_atomic_fcomp_swap", src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 564
 565 def system_value(name, dest_comp, indices=[], bit_sizes=[32]):
 566     intrinsic("load_" + name, [], dest_comp, indices,
 567               flags=[CAN_ELIMINATE, CAN_REORDER], sysval=True,
 568               bit_sizes=bit_sizes)
 569
 570 system_value("frag_coord", 4)
 571 system_value("point_coord", 2)
 572 system_value("front_face", 1, bit_sizes=[1, 32])
 573 system_value("vertex_id", 1)
 574 system_value("vertex_id_zero_base", 1)
 575 system_value("first_vertex", 1)
 576 system_value("is_indexed_draw", 1)
 577 system_value("base_vertex", 1)
 578 system_value("instance_id", 1)
 579 system_value("base_instance", 1)
 580 system_value("draw_id", 1)
 581 system_value("sample_id", 1)
 582 # sample_id_no_per_sample is like sample_id but does not imply per-
 583 # sample shading.  See the lower_helper_invocation option.
 584 system_value("sample_id_no_per_sample", 1)
 585 system_value("sample_pos", 2)
 586 system_value("sample_mask_in", 1)
 587 system_value("primitive_id", 1)
 588 system_value("invocation_id", 1)
 589 system_value("tess_coord", 3)
 590 system_value("tess_level_outer", 4)
 591 system_value("tess_level_inner", 2)
 592 system_value("tess_level_outer_default", 4)
 593 system_value("tess_level_inner_default", 2)
 594 system_value("patch_vertices_in", 1)
 595 system_value("local_invocation_id", 3)
 596 system_value("local_invocation_index", 1)
 597 system_value("work_group_id", 3)
 598 system_value("user_clip_plane", 4, indices=[UCP_ID])
 599 system_value("num_work_groups", 3)
 600 system_value("helper_invocation", 1, bit_sizes=[1, 32])
 601 system_value("alpha_ref_float", 1)
 602 system_value("layer_id", 1)
 603 system_value("view_index", 1)
 604 system_value("subgroup_size", 1)
 605 system_value("subgroup_invocation", 1)
 606 system_value("subgroup_eq_mask", 0, bit_sizes=[32, 64])
 607 system_value("subgroup_ge_mask", 0, bit_sizes=[32, 64])
 608 system_value("subgroup_gt_mask", 0, bit_sizes=[32, 64])
 609 system_value("subgroup_le_mask", 0, bit_sizes=[32, 64])
 610 system_value("subgroup_lt_mask", 0, bit_sizes=[32, 64])
 611 system_value("num_subgroups", 1)
 612 system_value("subgroup_id", 1)
 613 system_value("local_group_size", 3)
 614 system_value("global_invocation_id", 3, bit_sizes=[32, 64])
 615 system_value("global_invocation_index", 1, bit_sizes=[32, 64])
 616 system_value("work_dim", 1)
 617 # Driver-specific viewport scale/offset parameters.
 618 #
 619 # VC4 and V3D need to emit a scaled version of the position in the vertex
 620 # shaders for binning, and having system values lets us move the math for that
 621 # into NIR.
 622 #
 623 # Panfrost needs to implement all coordinate transformation in the
 624 # vertex shader; system values allow us to share this routine in NIR.
 625 system_value("viewport_x_scale", 1)
 626 system_value("viewport_y_scale", 1)
 627 system_value("viewport_z_scale", 1)
 628 system_value("viewport_z_offset", 1)
 629 system_value("viewport_scale", 3)
 630 system_value("viewport_offset", 3)
 631
 632 # Blend constant color values.  Float values are clamped. Vectored versions are
 633 # provided as well for driver convenience
 634
 635 system_value("blend_const_color_r_float", 1)
 636 system_value("blend_const_color_g_float", 1)
 637 system_value("blend_const_color_b_float", 1)
 638 system_value("blend_const_color_a_float", 1)
 639 system_value("blend_const_color_rgba", 4)
 640 system_value("blend_const_color_rgba8888_unorm", 1)
 641 system_value("blend_const_color_aaaa8888_unorm", 1)
 642
 643 # System values for gl_Color, for radeonsi which interpolates these in the
 644 # shader prolog to handle two-sided color without recompiles and therefore
 645 # doesn't handle these in the main shader part like normal varyings.
 646 system_value("color0", 4)
 647 system_value("color1", 4)
 648
 649 # System value for internal compute shaders in radeonsi.
 650 system_value("user_data_amd", 4)
 651
 652 # Number of data items being operated on for a SIMD program.
 653 system_value("simd_width_intel", 1)
 654
 655 # Barycentric coordinate intrinsics.
 656 #
 657 # These set up the barycentric coordinates for a particular interpolation.
 658 # The first four are for the simple cases: pixel, centroid, per-sample
 659 # (at gl_SampleID), or pull model (1/W, 1/I, 1/J) at the pixel center. The next
 660 # three two handle interpolating at a specified sample location, or
 661 # interpolating with a vec2 offset,
 662 #
 663 # The interp_mode index should be either the INTERP_MODE_SMOOTH or
 664 # INTERP_MODE_NOPERSPECTIVE enum values.
 665 #
 666 # The vec2 value produced by these intrinsics is intended for use as the
 667 # barycoord source of a load_interpolated_input intrinsic.
 668
 669 def barycentric(name, dst_comp, src_comp=[]):
 670     intrinsic("load_barycentric_" + name, src_comp=src_comp, dest_comp=dst_comp,
 671               indices=[INTERP_MODE], flags=[CAN_ELIMINATE, CAN_REORDER])
 672
 673 # no sources.
 674 barycentric("pixel", 2)
 675 barycentric("centroid", 2)
 676 barycentric("sample", 2)
 677 barycentric("model", 3)
 678 # src[] = { sample_id }.
 679 barycentric("at_sample", 2, [1])
 680 # src[] = { offset.xy }.
 681 barycentric("at_offset", 2, [2])
 682
 683 # Load sample position:
 684 #
 685 # Takes a sample # and returns a sample position.  Used for lowering
 686 # interpolateAtSample() to interpolateAtOffset()
 687 intrinsic("load_sample_pos_from_id", src_comp=[1], dest_comp=2,
 688           flags=[CAN_ELIMINATE, CAN_REORDER])
 689
 690 # Loads what I believe is the primitive size, for scaling ij to pixel size:
 691 intrinsic("load_size_ir3", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
 692
 693 # Fragment shader input interpolation delta intrinsic.
 694 #
 695 # For hw where fragment shader input interpolation is handled in shader, the
 696 # load_fs_input_interp deltas intrinsics can be used to load the input deltas
 697 # used for interpolation as follows:
 698 #
 699 #    vec3 iid = load_fs_input_interp_deltas(varying_slot)
 700 #    vec2 bary = load_barycentric_*(...)
 701 #    float result = iid.x + iid.y * bary.y + iid.z * bary.x
 702
 703 intrinsic("load_fs_input_interp_deltas", src_comp=[1], dest_comp=3,
 704           indices=[BASE, COMPONENT], flags=[CAN_ELIMINATE, CAN_REORDER])
 705
 706 # Load operations pull data from some piece of GPU memory.  All load
 707 # operations operate in terms of offsets into some piece of theoretical
 708 # memory.  Loads from externally visible memory (UBO and SSBO) simply take a
 709 # byte offset as a source.  Loads from opaque memory (uniforms, inputs, etc.)
 710 # take a base+offset pair where the nir_intrinsic_base() gives the location
 711 # of the start of the variable being loaded and and the offset source is a
 712 # offset into that variable.
 713 #
 714 # Uniform load operations have a nir_intrinsic_range() index that specifies the
 715 # range (starting at base) of the data from which we are loading.  If
 716 # range == 0, then the range is unknown.
 717 #
 718 # Some load operations such as UBO/SSBO load and per_vertex loads take an
 719 # additional source to specify which UBO/SSBO/vertex to load from.
 720 #
 721 # The exact address type depends on the lowering pass that generates the
 722 # load/store intrinsics.  Typically, this is vec4 units for things such as
 723 # varying slots and float units for fragment shader inputs.  UBO and SSBO
 724 # offsets are always in bytes.
 725
 726 def load(name, num_srcs, indices=[], flags=[]):
 727     intrinsic("load_" + name, [1] * num_srcs, dest_comp=0, indices=indices,
 728               flags=flags)
 729
 730 # src[] = { offset }.
 731 load("uniform", 1, [BASE, RANGE, TYPE], [CAN_ELIMINATE, CAN_REORDER])
 732 # src[] = { buffer_index, offset }.
 733 load("ubo", 2, [ACCESS, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN_ELIMINATE, CAN_REORDER])
 734 # src[] = { offset }.
 735 load("input", 1, [BASE, COMPONENT, TYPE], [CAN_ELIMINATE, CAN_REORDER])
 736 # src[] = { vertex_id, offset }.
 737 load("input_vertex", 2, [BASE, COMPONENT, TYPE], [CAN_ELIMINATE, CAN_REORDER])
 738 # src[] = { vertex, offset }.
 739 load("per_vertex_input", 2, [BASE, COMPONENT], [CAN_ELIMINATE, CAN_REORDER])
 740 # src[] = { barycoord, offset }.
 741 intrinsic("load_interpolated_input", src_comp=[2, 1], dest_comp=0,
 742           indices=[BASE, COMPONENT], flags=[CAN_ELIMINATE, CAN_REORDER])
 743
 744 # src[] = { buffer_index, offset }.
 745 load("ssbo", 2, [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 746 # src[] = { buffer_index }
 747 load("ssbo_address", 1, [], [CAN_ELIMINATE, CAN_REORDER])
 748 # src[] = { offset }.
 749 load("output", 1, [BASE, COMPONENT], flags=[CAN_ELIMINATE])
 750 # src[] = { vertex, offset }.
 751 load("per_vertex_output", 2, [BASE, COMPONENT], [CAN_ELIMINATE])
 752 # src[] = { offset }.
 753 load("shared", 1, [BASE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 754 # src[] = { offset }.
 755 load("push_constant", 1, [BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
 756 # src[] = { offset }.
 757 load("constant", 1, [BASE, RANGE, ALIGN_MUL, ALIGN_OFFSET],
 758      [CAN_ELIMINATE, CAN_REORDER])
 759 # src[] = { address }.
 760 load("global", 1, [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 761 # src[] = { address }.
 762 load("kernel_input", 1, [BASE, RANGE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE, CAN_REORDER])
 763 # src[] = { offset }.
 764 load("scratch", 1, [ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 765
 766 # Stores work the same way as loads, except now the first source is the value
 767 # to store and the second (and possibly third) source specify where to store
 768 # the value.  SSBO and shared memory stores also have a
 769 # nir_intrinsic_write_mask()
 770
 771 def store(name, num_srcs, indices=[], flags=[]):
 772     intrinsic("store_" + name, [0] + ([1] * (num_srcs - 1)), indices=indices, flags=flags)
 773
 774 # src[] = { value, offset }.
 775 store("output", 2, [BASE, WRMASK, COMPONENT, TYPE])
 776 # src[] = { value, vertex, offset }.
 777 store("per_vertex_output", 3, [BASE, WRMASK, COMPONENT])
 778 # src[] = { value, block_index, offset }
 779 store("ssbo", 3, [WRMASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
 780 # src[] = { value, offset }.
 781 store("shared", 2, [BASE, WRMASK, ALIGN_MUL, ALIGN_OFFSET])
 782 # src[] = { value, address }.
 783 store("global", 2, [WRMASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
 784 # src[] = { value, offset }.
 785 store("scratch", 2, [ALIGN_MUL, ALIGN_OFFSET, WRMASK])
 786
 787 # IR3-specific version of most SSBO intrinsics. The only different
 788 # compare to the originals is that they add an extra source to hold
 789 # the dword-offset, which is needed by the backend code apart from
 790 # the byte-offset already provided by NIR in one of the sources.
 791 #
 792 # NIR lowering pass 'ir3_nir_lower_io_offset' will replace the
 793 # original SSBO intrinsics by these, placing the computed
 794 # dword-offset always in the last source.
 795 #
 796 # The float versions are not handled because those are not supported
 797 # by the backend.
 798 intrinsic("store_ssbo_ir3",  src_comp=[0, 1, 1, 1],
 799           indices=[WRMASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
 800 intrinsic("load_ssbo_ir3",  src_comp=[1, 1, 1], dest_comp=0,
 801           indices=[ACCESS, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN_ELIMINATE])
 802 intrinsic("ssbo_atomic_add_ir3",        src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
 803 intrinsic("ssbo_atomic_imin_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
 804 intrinsic("ssbo_atomic_umin_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
 805 intrinsic("ssbo_atomic_imax_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
 806 intrinsic("ssbo_atomic_umax_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
 807 intrinsic("ssbo_atomic_and_ir3",        src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
 808 intrinsic("ssbo_atomic_or_ir3",         src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
 809 intrinsic("ssbo_atomic_xor_ir3",        src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
 810 intrinsic("ssbo_atomic_exchange_ir3",   src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
 811 intrinsic("ssbo_atomic_comp_swap_ir3",  src_comp=[1, 1, 1, 1, 1], dest_comp=1, indices=[ACCESS])
 812
 813 # IR3-specific instruction for UBO loads using the ldc instruction. The second
 814 # source is the indirect offset, in units of four dwords. The base is a
 815 # component offset, in dword units.
 816 intrinsic("load_ubo_ir3", src_comp=[1, 1], bit_sizes=[32], dest_comp=0, indices=[BASE],
 817           flags=[CAN_REORDER, CAN_ELIMINATE])
 818
 819 # System values for freedreno geometry shaders.
 820 system_value("vs_primitive_stride_ir3", 1)
 821 system_value("vs_vertex_stride_ir3", 1)
 822 system_value("gs_header_ir3", 1)
 823 system_value("primitive_location_ir3", 1, indices=[DRIVER_LOCATION])
 824
 825 # System values for freedreno tessellation shaders.
 826 system_value("hs_patch_stride_ir3", 1)
 827 system_value("tess_factor_base_ir3", 2)
 828 system_value("tess_param_base_ir3", 2)
 829 system_value("tcs_header_ir3", 1)
 830
 831 # IR3-specific intrinsics for tessellation control shaders.  cond_end_ir3 end
 832 # the shader when src0 is false and is used to narrow down the TCS shader to
 833 # just thread 0 before writing out tessellation levels.
 834 intrinsic("cond_end_ir3", src_comp=[1])
 835 # end_patch_ir3 is used just before thread 0 exist the TCS and presumably
 836 # signals the TE that the patch is complete and can be tessellated.
 837 intrinsic("end_patch_ir3")
 838
 839 # IR3-specific load/store intrinsics. These access a buffer used to pass data
 840 # between geometry stages - perhaps it's explicit access to the vertex cache.
 841
 842 # src[] = { value, offset }.
 843 store("shared_ir3", 2, [BASE, ALIGN_MUL, ALIGN_OFFSET])
 844 # src[] = { offset }.
 845 load("shared_ir3", 1, [BASE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 846
 847 # IR3-specific load/store global intrinsics. They take a 64-bit base address
 848 # and a 32-bit offset.  The hardware will add the base and the offset, which
 849 # saves us from doing 64-bit math on the base address.
 850
 851 # src[] = { value, address(vec2 of hi+lo uint32_t), offset }.
 852 # const_index[] = { write_mask, align_mul, align_offset }
 853 intrinsic("store_global_ir3", [0, 2, 1], indices=[ACCESS, ALIGN_MUL, ALIGN_OFFSET])
 854 # src[] = { address(vec2 of hi+lo uint32_t), offset }.
 855 # const_index[] = { access, align_mul, align_offset }
 856 intrinsic("load_global_ir3", [2, 1], dest_comp=0, indices=[ACCESS, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN_ELIMINATE])
 857
 858 # IR3-specific bindless handle specifier. Similar to vulkan_resource_index, but
 859 # without the binding because the hardware expects a single flattened index
 860 # rather than a (binding, index) pair. We may also want to use this with GL.
 861 # Note that this doesn't actually turn into a HW instruction.
 862 intrinsic("bindless_resource_ir3", [1], dest_comp=1, indices=[DESC_SET], flags=[CAN_ELIMINATE, CAN_REORDER])
 863
 864 # Intrinsics used by the Midgard/Bifrost blend pipeline. These are defined
 865 # within a blend shader to read/write the raw value from the tile buffer,
 866 # without applying any format conversion in the process. If the shader needs
 867 # usable pixel values, it must apply format conversions itself.
 868 #
 869 # These definitions are generic, but they are explicitly vendored to prevent
 870 # other drivers from using them, as their semantics is defined in terms of the
 871 # Midgard/Bifrost hardware tile buffer and may not line up with anything sane.
 872 # One notable divergence is sRGB, which is asymmetric: raw_input_pan requires
 873 # an sRGB->linear conversion, but linear values should be written to
 874 # raw_output_pan and the hardware handles linear->sRGB.
 875
 876 # src[] = { value }
 877 store("raw_output_pan", 1, [])
 878 store("combined_output_pan", 4, [BASE, COMPONENT])
 879 load("raw_output_pan", 0, [], [CAN_ELIMINATE, CAN_REORDER])
 880
 881 # Loads the sampler paramaters <min_lod, max_lod, lod_bias>
 882 # src[] = { sampler_index }
 883 load("sampler_lod_parameters_pan", 1, [CAN_ELIMINATE, CAN_REORDER])
 884
 885 # R600 specific instrincs
 886 #
 887 # R600 can only fetch 16 byte aligned data from an UBO, and the actual offset
 888 # is given in vec4 units, so we have to fetch the a vec4 and get the component
 889 # later
 890 # src[] = { buffer_index, offset }.
 891 load("ubo_r600", 2, [ACCESS, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN_ELIMINATE, CAN_REORDER])
 892
 893 # location where the tesselation data is stored in LDS
 894 system_value("tcs_in_param_base_r600", 4)
 895 system_value("tcs_out_param_base_r600", 4)
 896 system_value("tcs_rel_patch_id_r600", 1)
 897 system_value("tcs_tess_factor_base_r600", 1)
 898
 899 # load as many components as needed giving per-component addresses
 900 intrinsic("load_local_shared_r600", src_comp=[0], dest_comp=0, indices = [COMPONENT], flags = [CAN_ELIMINATE, CAN_REORDER])
 901
 902 store("local_shared_r600", 2, [WRMASK])
 903 store("tf_r600", 1)
 904
 905 # V3D-specific instrinc for tile buffer color reads.
 906 #
 907 # The hardware requires that we read the samples and components of a pixel
 908 # in order, so we cannot eliminate or remove any loads in a sequence.
 909 #
 910 # src[] = { render_target }
 911 # BASE = sample index
 912 load("tlb_color_v3d", 1, [BASE, COMPONENT], [])
 913
 914 # V3D-specific instrinc for per-sample tile buffer color writes.
 915 #
 916 # The driver backend needs to identify per-sample color writes and emit
 917 # specific code for them.
 918 #
 919 # src[] = { value, render_target }
 920 # BASE = sample index
 921 store("tlb_sample_color_v3d", 2, [BASE, COMPONENT, TYPE], [])
 922
 923 # V3D-specific intrinsic to load the number of layers attached to
 924 # the target framebuffer
 925 intrinsic("load_fb_layers_v3d", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])