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