• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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
30src0 = ('src', 0)
31src1 = ('src', 1)
32src2 = ('src', 2)
33src3 = ('src', 3)
34src4 = ('src', 4)
35
36class Index(object):
37    def __init__(self, c_data_type, name):
38        self.c_data_type = c_data_type
39        self.name = name
40
41class Intrinsic(object):
42   """Class that represents all the information about an intrinsic opcode.
43   NOTE: this must be kept in sync with nir_intrinsic_info.
44   """
45   def __init__(self, name, src_components, dest_components,
46                indices, flags, sysval, bit_sizes):
47       """Parameters:
48
49       - name: the intrinsic name
50       - src_components: list of the number of components per src, 0 means
51         vectorized instruction with number of components given in the
52         num_components field in nir_intrinsic_instr.
53       - dest_components: number of destination components, -1 means no
54         dest, 0 means number of components given in num_components field
55         in nir_intrinsic_instr.
56       - indices: list of constant indicies
57       - flags: list of semantic flags
58       - sysval: is this a system-value intrinsic
59       - bit_sizes: allowed dest bit_sizes or the source it must match
60       """
61       assert isinstance(name, str)
62       assert isinstance(src_components, list)
63       if src_components:
64           assert isinstance(src_components[0], int)
65       assert isinstance(dest_components, int)
66       assert isinstance(indices, list)
67       if indices:
68           assert isinstance(indices[0], Index)
69       assert isinstance(flags, list)
70       if flags:
71           assert isinstance(flags[0], str)
72       assert isinstance(sysval, bool)
73       if isinstance(bit_sizes, list):
74           assert not bit_sizes or isinstance(bit_sizes[0], int)
75       else:
76           assert isinstance(bit_sizes, tuple)
77           assert bit_sizes[0] == 'src'
78           assert isinstance(bit_sizes[1], int)
79
80       self.name = name
81       self.num_srcs = len(src_components)
82       self.src_components = src_components
83       self.has_dest = (dest_components >= 0)
84       self.dest_components = dest_components
85       self.num_indices = len(indices)
86       self.indices = indices
87       self.flags = flags
88       self.sysval = sysval
89       self.bit_sizes = bit_sizes if isinstance(bit_sizes, list) else []
90       self.bit_size_src = bit_sizes[1] if isinstance(bit_sizes, tuple) else -1
91
92#
93# Possible flags:
94#
95
96CAN_ELIMINATE = "NIR_INTRINSIC_CAN_ELIMINATE"
97CAN_REORDER   = "NIR_INTRINSIC_CAN_REORDER"
98
99INTR_INDICES = []
100INTR_OPCODES = {}
101
102def index(c_data_type, name):
103    idx = Index(c_data_type, name)
104    INTR_INDICES.append(idx)
105    globals()[name.upper()] = idx
106
107# Defines a new NIR intrinsic.  By default, the intrinsic will have no sources
108# and no destination.
109#
110# You can set dest_comp=n to enable a destination for the intrinsic, in which
111# case it will have that many components, or =0 for "as many components as the
112# NIR destination value."
113#
114# Set src_comp=n to enable sources for the intruction.  It can be an array of
115# component counts, or (for convenience) a scalar component count if there's
116# only one source.  If a component count is 0, it will be as many components as
117# the intrinsic has based on the dest_comp.
118def intrinsic(name, src_comp=[], dest_comp=-1, indices=[],
119              flags=[], sysval=False, bit_sizes=[]):
120    assert name not in INTR_OPCODES
121    INTR_OPCODES[name] = Intrinsic(name, src_comp, dest_comp,
122                                   indices, flags, sysval, bit_sizes)
123
124#
125# Possible indices:
126#
127
128# Generally instructions that take a offset src argument, can encode
129# a constant 'base' value which is added to the offset.
130index("int", "base")
131
132# For store instructions, a writemask for the store.
133index("unsigned", "write_mask")
134
135# The stream-id for GS emit_vertex/end_primitive intrinsics.
136index("unsigned", "stream_id")
137
138# The clip-plane id for load_user_clip_plane intrinsic.
139index("unsigned", "ucp_id")
140
141# The offset to the start of the NIR_INTRINSIC_RANGE.  This is an alternative
142# to NIR_INTRINSIC_BASE for describing the valid range in intrinsics that don't
143# have the implicit addition of a base to the offset.
144#
145# If the [range_base, range] is [0, ~0], then we don't know the possible
146# range of the access.
147index("unsigned", "range_base")
148
149# The amount of data, starting from BASE or RANGE_BASE, that this
150# instruction may access.  This is used to provide bounds if the offset is
151# not constant.
152index("unsigned", "range")
153
154# The Vulkan descriptor set for vulkan_resource_index intrinsic.
155index("unsigned", "desc_set")
156
157# The Vulkan descriptor set binding for vulkan_resource_index intrinsic.
158index("unsigned", "binding")
159
160# Component offset
161index("unsigned", "component")
162
163# Column index for matrix system values
164index("unsigned", "column")
165
166# Interpolation mode (only meaningful for FS inputs)
167index("unsigned", "interp_mode")
168
169# A binary nir_op to use when performing a reduction or scan operation
170index("unsigned", "reduction_op")
171
172# Cluster size for reduction operations
173index("unsigned", "cluster_size")
174
175# Parameter index for a load_param intrinsic
176index("unsigned", "param_idx")
177
178# Image dimensionality for image intrinsics
179index("enum glsl_sampler_dim", "image_dim")
180
181# Non-zero if we are accessing an array image
182index("bool", "image_array")
183
184# Image format for image intrinsics
185# Vertex buffer format for load_typed_buffer_amd
186index("enum pipe_format", "format")
187
188# Access qualifiers for image and memory access intrinsics. ACCESS_RESTRICT is
189# not set at the intrinsic if the NIR was created from SPIR-V.
190index("enum gl_access_qualifier", "access")
191
192# call index for split raytracing shaders
193index("unsigned", "call_idx")
194
195# The stack size increment/decrement for split raytracing shaders
196index("unsigned", "stack_size")
197
198# Alignment for offsets and addresses
199#
200# These two parameters, specify an alignment in terms of a multiplier and
201# an offset.  The multiplier is always a power of two.  The offset or
202# address parameter X of the intrinsic is guaranteed to satisfy the
203# following:
204#
205#                (X - align_offset) % align_mul == 0
206#
207# For constant offset values, align_mul will be NIR_ALIGN_MUL_MAX and the
208# align_offset will be modulo that.
209index("unsigned", "align_mul")
210index("unsigned", "align_offset")
211
212# The Vulkan descriptor type for a vulkan_resource_[re]index intrinsic.
213index("unsigned", "desc_type")
214
215# The nir_alu_type of input data to a store or conversion
216index("nir_alu_type", "src_type")
217
218# The nir_alu_type of the data output from a load or conversion
219index("nir_alu_type", "dest_type")
220
221# The swizzle mask for quad_swizzle_amd & masked_swizzle_amd
222index("unsigned", "swizzle_mask")
223
224# Allow FI=1 for quad_swizzle_amd & masked_swizzle_amd
225index("bool", "fetch_inactive")
226
227# Offsets for load_shared2_amd/store_shared2_amd
228index("uint8_t", "offset0")
229index("uint8_t", "offset1")
230
231# If true, both offsets have an additional stride of 64 dwords (ie. they are multiplied by 256 bytes
232# in hardware, instead of 4).
233index("bool", "st64")
234
235# When set, range analysis will use it for nir_unsigned_upper_bound
236index("unsigned", "arg_upper_bound_u32_amd")
237
238# Separate source/dest access flags for copies
239index("enum gl_access_qualifier", "dst_access")
240index("enum gl_access_qualifier", "src_access")
241
242# Driver location of attribute
243index("unsigned", "driver_location")
244
245# Ordering and visibility of a memory operation
246index("nir_memory_semantics", "memory_semantics")
247
248# Modes affected by a memory operation
249index("nir_variable_mode", "memory_modes")
250
251# Scope of a memory operation
252index("mesa_scope", "memory_scope")
253
254# Scope of a control barrier
255index("mesa_scope", "execution_scope")
256
257# Semantics of an IO instruction
258index("struct nir_io_semantics", "io_semantics")
259
260# Transform feedback info
261index("struct nir_io_xfb", "io_xfb")
262index("struct nir_io_xfb", "io_xfb2")
263
264# Ray query values accessible from the RayQueryKHR object
265index("nir_ray_query_value", "ray_query_value")
266
267# Select between committed and candidate ray queriy intersections
268index("bool", "committed")
269
270# Rounding mode for conversions
271index("nir_rounding_mode", "rounding_mode")
272
273# Whether or not to saturate in conversions
274index("unsigned", "saturate")
275
276# Whether or not trace_ray_intel is synchronous
277index("bool", "synchronous")
278
279# Value ID to identify SSA value loaded/stored on the stack
280index("unsigned", "value_id")
281
282# Whether to sign-extend offsets in address arithmatic (else zero extend)
283index("bool", "sign_extend")
284
285# Instruction specific flags
286index("unsigned", "flags")
287
288# Logical operation of an atomic intrinsic
289index("nir_atomic_op", "atomic_op")
290
291# Block identifier to push promotion
292index("unsigned", "resource_block_intel")
293
294# Various flags describing the resource access
295index("nir_resource_data_intel", "resource_access_intel")
296
297# Register metadata
298# number of vector components
299index("unsigned", "num_components")
300# size of array (0 for no array)
301index("unsigned", "num_array_elems")
302# The bit-size of each channel; must be one of 1, 8, 16, 32, or 64
303index("unsigned", "bit_size")
304# True if this register may have different values in different SIMD invocations
305# of the shader.
306index("bool", "divergent")
307
308# On a register load, floating-point absolute value/negate loaded value.
309index("bool", "legacy_fabs")
310index("bool", "legacy_fneg")
311
312# On a register store, floating-point saturate the stored value.
313index("bool", "legacy_fsat")
314
315# For Cooperative Matrix intrinsics.
316index("struct glsl_cmat_description", "cmat_desc")
317index("enum glsl_matrix_layout", "matrix_layout")
318index("nir_cmat_signed", "cmat_signed_mask")
319index("nir_op", "alu_op")
320
321# For Intel DPAS instrinsic.
322index("unsigned", "systolic_depth")
323index("unsigned", "repeat_count")
324
325# For an AGX tilebuffer intrinsics, whether the coordinates are implicit or
326# explicit. Implicit coordinates are used in fragment shaders, explicit
327# coordinates in compute.
328index("bool", "explicit_coord")
329
330intrinsic("nop", flags=[CAN_ELIMINATE])
331
332# Uses a value and cannot be eliminated.
333#
334# This is helpful when writing unit tests
335intrinsic("use", src_comp=[0], flags=[])
336
337intrinsic("convert_alu_types", dest_comp=0, src_comp=[0],
338          indices=[SRC_TYPE, DEST_TYPE, ROUNDING_MODE, SATURATE],
339          flags=[CAN_ELIMINATE, CAN_REORDER])
340
341intrinsic("load_param", dest_comp=0, indices=[PARAM_IDX], flags=[CAN_ELIMINATE])
342
343intrinsic("load_deref", dest_comp=0, src_comp=[-1],
344          indices=[ACCESS], flags=[CAN_ELIMINATE])
345intrinsic("store_deref", src_comp=[-1, 0], indices=[WRITE_MASK, ACCESS])
346intrinsic("copy_deref", src_comp=[-1, -1], indices=[DST_ACCESS, SRC_ACCESS])
347intrinsic("memcpy_deref", src_comp=[-1, -1, 1], indices=[DST_ACCESS, SRC_ACCESS])
348
349# Returns an opaque handle representing a register indexed by BASE. The
350# logically def-use list of a register is given by the use list of this handle.
351# The shape of the underlying register is given by the indices, the handle
352# itself is always a 32-bit scalar.
353intrinsic("decl_reg", dest_comp=1,
354          indices=[NUM_COMPONENTS, NUM_ARRAY_ELEMS, BIT_SIZE, DIVERGENT],
355          flags=[CAN_ELIMINATE])
356
357# Load a register given as the source directly with base offset BASE.
358intrinsic("load_reg", dest_comp=0, src_comp=[1],
359          indices=[BASE, LEGACY_FABS, LEGACY_FNEG], flags=[CAN_ELIMINATE])
360
361# Load a register given as first source indirectly with base offset BASE and
362# indirect offset as second source.
363intrinsic("load_reg_indirect", dest_comp=0, src_comp=[1, 1],
364          indices=[BASE, LEGACY_FABS, LEGACY_FNEG], flags=[CAN_ELIMINATE])
365
366# Store the value in the first source to a register given as the second source
367# directly with base offset BASE.
368intrinsic("store_reg", src_comp=[0, 1],
369          indices=[BASE, WRITE_MASK, LEGACY_FSAT])
370
371# Store the value in the first source to a register given as the second
372# source indirectly with base offset BASE and indirect offset as third source.
373intrinsic("store_reg_indirect", src_comp=[0, 1, 1],
374          indices=[BASE, WRITE_MASK, LEGACY_FSAT])
375
376# Interpolation of input.  The interp_deref_at* intrinsics are similar to the
377# load_var intrinsic acting on a shader input except that they interpolate the
378# input differently.  The at_sample, at_offset and at_vertex intrinsics take an
379# additional source that is an integer sample id, a vec2 position offset, or a
380# vertex ID respectively.
381
382intrinsic("interp_deref_at_centroid", dest_comp=0, src_comp=[1],
383          flags=[ CAN_ELIMINATE, CAN_REORDER])
384intrinsic("interp_deref_at_sample", src_comp=[1, 1], dest_comp=0,
385          flags=[CAN_ELIMINATE, CAN_REORDER])
386intrinsic("interp_deref_at_offset", src_comp=[1, 2], dest_comp=0,
387          flags=[CAN_ELIMINATE, CAN_REORDER])
388intrinsic("interp_deref_at_vertex", src_comp=[1, 1], dest_comp=0,
389          flags=[CAN_ELIMINATE, CAN_REORDER])
390
391# Gets the length of an unsized array at the end of a buffer
392intrinsic("deref_buffer_array_length", src_comp=[-1], dest_comp=1,
393          indices=[ACCESS], flags=[CAN_ELIMINATE, CAN_REORDER])
394
395# Gets the length of an unsized array
396intrinsic("deref_implicit_array_length", src_comp=[-1], dest_comp=1,
397          flags=[CAN_ELIMINATE, CAN_REORDER])
398
399# Ask the driver for the size of a given SSBO. It takes the buffer index
400# as source.
401intrinsic("get_ssbo_size", src_comp=[-1], dest_comp=1, bit_sizes=[32],
402          indices=[ACCESS], flags=[CAN_ELIMINATE, CAN_REORDER])
403intrinsic("get_ubo_size", src_comp=[-1], dest_comp=1,
404          flags=[CAN_ELIMINATE, CAN_REORDER])
405
406# Intrinsics which provide a run-time mode-check.  Unlike the compile-time
407# mode checks, a pointer can only have exactly one mode at runtime.
408intrinsic("deref_mode_is", src_comp=[-1], dest_comp=1,
409          indices=[MEMORY_MODES], flags=[CAN_ELIMINATE, CAN_REORDER])
410intrinsic("addr_mode_is", src_comp=[-1], dest_comp=1,
411          indices=[MEMORY_MODES], flags=[CAN_ELIMINATE, CAN_REORDER])
412
413intrinsic("is_sparse_texels_resident", dest_comp=1, src_comp=[1], bit_sizes=[1,32],
414          flags=[CAN_ELIMINATE, CAN_REORDER])
415# result code is resident only if both inputs are resident
416intrinsic("sparse_residency_code_and", dest_comp=1, src_comp=[1, 1], bit_sizes=[32],
417          flags=[CAN_ELIMINATE, CAN_REORDER])
418
419# Unlike is_sparse_texels_resident, this intrinsic is required to consume
420# the destination of the nir_tex_instr or sparse_load intrinsic directly.
421# As such it is allowed to ignore the .e component where we usually store
422# sparse information.
423intrinsic("is_sparse_resident_zink", dest_comp=1, src_comp=[0], bit_sizes=[1],
424          flags=[CAN_ELIMINATE, CAN_REORDER])
425
426# The following intrinsics calculate screen-space partial derivatives. These are
427# not CAN_REORDER as they cannot be moved across discards.
428for suffix in ["", "_fine", "_coarse"]:
429    for axis in ["x", "y"]:
430        intrinsic(f"dd{axis}{suffix}", dest_comp=0, src_comp=[0],
431                  bit_sizes=[16, 32], flags=[CAN_ELIMINATE])
432
433# a barrier is an intrinsic with no inputs/outputs but which can't be moved
434# around/optimized in general
435def barrier(name):
436    intrinsic(name)
437
438# Demote fragment shader invocation to a helper invocation.  Any stores to
439# memory after this instruction are suppressed and the fragment does not write
440# outputs to the framebuffer.  Unlike discard, demote needs to ensure that
441# derivatives will still work for invocations that were not demoted.
442#
443# As specified by SPV_EXT_demote_to_helper_invocation.
444barrier("demote")
445intrinsic("is_helper_invocation", dest_comp=1, flags=[CAN_ELIMINATE])
446
447# SpvOpTerminateInvocation from SPIR-V.  Essentially a discard "for real".
448barrier("terminate")
449
450# NonSemantic.DebugBreak from SPIR-V. Essentially used to emit breakpoints in
451# shaders.
452barrier("debug_break")
453
454# Control/Memory barrier with explicit scope.  Follows the semantics of SPIR-V
455# OpMemoryBarrier and OpControlBarrier, used to implement Vulkan Memory Model.
456# Storage that the barrier applies is represented using NIR variable modes.
457# For an OpMemoryBarrier, set EXECUTION_SCOPE to SCOPE_NONE.
458intrinsic("barrier",
459          indices=[EXECUTION_SCOPE, MEMORY_SCOPE, MEMORY_SEMANTICS, MEMORY_MODES])
460
461# Shader clock intrinsic with semantics analogous to the clock2x32ARB()
462# GLSL intrinsic.
463# The latter can be used as code motion barrier, which is currently not
464# feasible with NIR.
465intrinsic("shader_clock", dest_comp=2, bit_sizes=[32], flags=[CAN_ELIMINATE],
466          indices=[MEMORY_SCOPE])
467
468# Shader ballot intrinsics with semantics analogous to the
469#
470#    ballotARB()
471#    readInvocationARB()
472#    readFirstInvocationARB()
473#
474# GLSL functions from ARB_shader_ballot.
475intrinsic("ballot", src_comp=[1], dest_comp=0, flags=[CAN_ELIMINATE])
476intrinsic("read_invocation", src_comp=[0, 1], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
477intrinsic("read_first_invocation", src_comp=[0], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
478
479# Same as ballot, but inactive invocations contribute undefined bits.
480intrinsic("ballot_relaxed", src_comp=[1], dest_comp=0, flags=[CAN_ELIMINATE])
481
482# Allows the backend compiler to move this value to an uniform register.
483# Result is undefined if src is not uniform.
484# Unlike read_first_invocation, it may be replaced by a divergent move or CSE'd.
485intrinsic("as_uniform", src_comp=[0], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
486
487# Returns the value of the first source for the lane where the second source is
488# true. The second source must be true for exactly one lane.
489intrinsic("read_invocation_cond_ir3", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
490
491# Like read_first_invocation but using the getlast instruction instead of
492# getone. More specifically, this will read the value from the last active
493# invocation of the first cluster of 8 invocations with an active invocation.
494intrinsic("read_getlast_ir3", src_comp=[0], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
495
496# Additional SPIR-V ballot intrinsics
497#
498# These correspond to the SPIR-V opcodes
499#
500#    OpGroupNonUniformElect
501#    OpSubgroupFirstInvocationKHR
502#    OpGroupNonUniformInverseBallot
503intrinsic("elect", dest_comp=1, flags=[CAN_ELIMINATE])
504intrinsic("first_invocation", dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE])
505intrinsic("last_invocation", dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE])
506intrinsic("inverse_ballot", src_comp=[0], dest_comp=1, flags=[CAN_ELIMINATE])
507
508barrier("begin_invocation_interlock")
509barrier("end_invocation_interlock")
510
511# A conditional demote/terminate, with a single boolean source.
512intrinsic("demote_if", src_comp=[1])
513intrinsic("terminate_if", src_comp=[1])
514
515# ARB_shader_group_vote intrinsics
516intrinsic("vote_any", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
517intrinsic("vote_all", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
518intrinsic("vote_feq", src_comp=[0], dest_comp=1, flags=[CAN_ELIMINATE])
519intrinsic("vote_ieq", src_comp=[0], dest_comp=1, flags=[CAN_ELIMINATE])
520
521# Ballot ALU operations from SPIR-V.
522#
523# These operations work like their ALU counterparts except that the operate
524# on a uvec4 which is treated as a 128bit integer.  Also, they are, in
525# general, free to ignore any bits which are above the subgroup size.
526intrinsic("ballot_bitfield_extract", src_comp=[4, 1], dest_comp=1, flags=[CAN_REORDER, CAN_ELIMINATE])
527intrinsic("ballot_bit_count_reduce", src_comp=[4], dest_comp=1, flags=[CAN_REORDER, CAN_ELIMINATE])
528intrinsic("ballot_bit_count_inclusive", src_comp=[4], dest_comp=1, flags=[CAN_REORDER, CAN_ELIMINATE])
529intrinsic("ballot_bit_count_exclusive", src_comp=[4], dest_comp=1, flags=[CAN_REORDER, CAN_ELIMINATE])
530intrinsic("ballot_find_lsb", src_comp=[4], dest_comp=1, flags=[CAN_REORDER, CAN_ELIMINATE])
531intrinsic("ballot_find_msb", src_comp=[4], dest_comp=1, flags=[CAN_REORDER, CAN_ELIMINATE])
532
533# Shuffle operations from SPIR-V.
534intrinsic("shuffle", src_comp=[0, 1], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
535intrinsic("shuffle_xor", src_comp=[0, 1], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
536intrinsic("shuffle_up", src_comp=[0, 1], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
537intrinsic("shuffle_down", src_comp=[0, 1], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
538
539# Quad operations from SPIR-V.
540intrinsic("quad_broadcast", src_comp=[0, 1], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
541intrinsic("quad_swap_horizontal", src_comp=[0], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
542intrinsic("quad_swap_vertical", src_comp=[0], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
543intrinsic("quad_swap_diagonal", src_comp=[0], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
544
545# Similar to vote_any and vote_all, but per-quad instead of per-wavefront.
546# Equivalent to subgroupOr(val, 4) and subgroupAnd(val, 4) assuming val is
547# boolean.
548intrinsic("quad_vote_any", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
549intrinsic("quad_vote_all", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
550
551# Rotate operation from SPIR-V: SpvOpGroupNonUniformRotateKHR.
552intrinsic("rotate", src_comp=[0, 1], dest_comp=0, bit_sizes=src0,
553          indices=[CLUSTER_SIZE], flags=[CAN_ELIMINATE]);
554
555intrinsic("reduce", src_comp=[0], dest_comp=0, bit_sizes=src0,
556          indices=[REDUCTION_OP, CLUSTER_SIZE], flags=[CAN_ELIMINATE])
557intrinsic("inclusive_scan", src_comp=[0], dest_comp=0, bit_sizes=src0,
558          indices=[REDUCTION_OP], flags=[CAN_ELIMINATE])
559intrinsic("exclusive_scan", src_comp=[0], dest_comp=0, bit_sizes=src0,
560          indices=[REDUCTION_OP], flags=[CAN_ELIMINATE])
561
562# AMD shader ballot operations
563intrinsic("quad_swizzle_amd", src_comp=[0], dest_comp=0, bit_sizes=src0,
564          indices=[SWIZZLE_MASK, FETCH_INACTIVE], flags=[CAN_ELIMINATE])
565intrinsic("masked_swizzle_amd", src_comp=[0], dest_comp=0, bit_sizes=src0,
566          indices=[SWIZZLE_MASK, FETCH_INACTIVE], flags=[CAN_ELIMINATE])
567intrinsic("write_invocation_amd", src_comp=[0, 0, 1], dest_comp=0, bit_sizes=src0,
568          flags=[CAN_ELIMINATE])
569# src = [ mask, addition ]
570intrinsic("mbcnt_amd", src_comp=[1, 1], dest_comp=1, bit_sizes=[32], flags=[CAN_REORDER, CAN_ELIMINATE])
571# Compiled to v_permlane16_b32. src = [ value, lanesel_lo, lanesel_hi ]
572intrinsic("lane_permute_16_amd", src_comp=[1, 1, 1], dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE])
573# subgroup shuffle up/down with cluster size 16.
574# base in [-15, -1]: DPP_ROW_SR
575# base in [  1, 15]: DPP_ROW_SL, otherwise invalid.
576# Returns zero for invocations that try to read out of bounds
577intrinsic("dpp16_shift_amd", src_comp=[0], dest_comp=0, bit_sizes=src0, indices=[BASE], flags=[CAN_ELIMINATE])
578
579# Basic Geometry Shader intrinsics.
580#
581# emit_vertex implements GLSL's EmitStreamVertex() built-in.  It takes a single
582# index, which is the stream ID to write to.
583#
584# end_primitive implements GLSL's EndPrimitive() built-in.
585intrinsic("emit_vertex",   indices=[STREAM_ID])
586intrinsic("end_primitive", indices=[STREAM_ID])
587
588# Geometry Shader intrinsics with a vertex count.
589#
590# Alternatively, drivers may implement these intrinsics, and use
591# nir_lower_gs_intrinsics() to convert from the basic intrinsics.
592#
593# These contain four additional unsigned integer sources:
594# 1. The total number of vertices emitted so far.
595# 2. The number of vertices emitted for the current primitive
596#    so far if we're counting, otherwise undef.
597# 3. The total number of primitives emitted so far.
598# 4. The total number of decomposed primitives emitted so far. This counts like
599#    the PRIMITIVES_GENERATED query: a triangle strip with 5 vertices is counted
600#    as 3 primitives (not 1).
601intrinsic("emit_vertex_with_counter", src_comp=[1, 1, 1, 1], indices=[STREAM_ID])
602intrinsic("end_primitive_with_counter", src_comp=[1, 1, 1, 1], indices=[STREAM_ID])
603# Contains the final total vertex, primitive, and decomposed primitives counts
604# in the current GS thread.
605intrinsic("set_vertex_and_primitive_count", src_comp=[1, 1, 1], indices=[STREAM_ID])
606
607# Launches mesh shader workgroups from a task shader, with explicit task_payload.
608# Rules:
609# - This is a terminating instruction.
610# - May only occur in workgroup-uniform control flow.
611# - Dispatch sizes may be divergent (in which case the values
612#   from the first invocation are used).
613# Meaning of indices:
614# - BASE: address of the task_payload variable used.
615# - RANGE: size of the task_payload variable used.
616#
617# src[] = {vec(x, y, z)}
618intrinsic("launch_mesh_workgroups", src_comp=[3], indices=[BASE, RANGE])
619
620# Launches mesh shader workgroups from a task shader, with task_payload variable deref.
621# Same rules as launch_mesh_workgroups apply here as well.
622# src[] = {vec(x, y, z), payload pointer}
623intrinsic("launch_mesh_workgroups_with_payload_deref", src_comp=[3, -1], indices=[])
624
625# Trace a ray through an acceleration structure
626#
627# This instruction has a lot of parameters:
628#   0. Acceleration Structure
629#   1. Ray Flags
630#   2. Cull Mask
631#   3. SBT Offset
632#   4. SBT Stride
633#   5. Miss shader index
634#   6. Ray Origin
635#   7. Ray Tmin
636#   8. Ray Direction
637#   9. Ray Tmax
638#   10. Payload
639intrinsic("trace_ray", src_comp=[-1, 1, 1, 1, 1, 1, 3, 1, 3, 1, -1])
640# src[] = { hit_t, hit_kind }
641intrinsic("report_ray_intersection", src_comp=[1, 1], dest_comp=1)
642intrinsic("ignore_ray_intersection")
643intrinsic("accept_ray_intersection") # Not in SPIR-V; useful for lowering
644intrinsic("terminate_ray")
645# src[] = { sbt_index, payload }
646intrinsic("execute_callable", src_comp=[1, -1])
647
648# Initialize a ray query
649#
650#   0. Ray Query
651#   1. Acceleration Structure
652#   2. Ray Flags
653#   3. Cull Mask
654#   4. Ray Origin
655#   5. Ray Tmin
656#   6. Ray Direction
657#   7. Ray Tmax
658intrinsic("rq_initialize", src_comp=[-1, -1, 1, 1, 3, 1, 3, 1])
659# src[] = { query }
660intrinsic("rq_terminate", src_comp=[-1])
661# src[] = { query }
662intrinsic("rq_proceed", src_comp=[-1], dest_comp=1)
663# src[] = { query, hit }
664intrinsic("rq_generate_intersection", src_comp=[-1, 1])
665# src[] = { query }
666intrinsic("rq_confirm_intersection", src_comp=[-1])
667# src[] = { query }
668intrinsic("rq_load", src_comp=[-1], dest_comp=0, indices=[RAY_QUERY_VALUE,COMMITTED,COLUMN])
669
670# Driver independent raytracing helpers
671
672# rt_resume is a helper that that be the first instruction accesing the
673# stack/scratch in a resume shader for a raytracing pipeline. It includes the
674# resume index (for nir_lower_shader_calls_internal reasons) and the stack size
675# of the variables spilled during the call. The stack size can be use to e.g.
676# adjust a stack pointer.
677intrinsic("rt_resume", indices=[CALL_IDX, STACK_SIZE])
678
679# Lowered version of execute_callabe that includes the index of the resume
680# shader, and the amount of scratch space needed for this call (.ie. how much
681# to increase a stack pointer by).
682# src[] = { sbt_index, payload }
683intrinsic("rt_execute_callable", src_comp=[1, -1], indices=[CALL_IDX,STACK_SIZE])
684
685# Lowered version of trace_ray in a similar vein to rt_execute_callable.
686# src same as trace_ray
687intrinsic("rt_trace_ray", src_comp=[-1, 1, 1, 1, 1, 1, 3, 1, 3, 1, -1],
688          indices=[CALL_IDX, STACK_SIZE])
689
690
691# Atomic counters
692#
693# The *_deref variants take an atomic_uint nir_variable, while the other,
694# lowered, variants take a buffer index and register offset.  The buffer index
695# is always constant, as there's no way to declare an array of atomic counter
696# buffers.
697#
698# The register offset may be non-constant but must by dynamically uniform
699# ("Atomic counters aggregated into arrays within a shader can only be indexed
700# with dynamically uniform integral expressions, otherwise results are
701# undefined.")
702def atomic(name, flags=[]):
703    intrinsic(name + "_deref", src_comp=[-1], dest_comp=1, flags=flags)
704    intrinsic(name, src_comp=[1], dest_comp=1, indices=[BASE, RANGE_BASE], flags=flags)
705
706def atomic2(name):
707    intrinsic(name + "_deref", src_comp=[-1, 1], dest_comp=1)
708    intrinsic(name, src_comp=[1, 1], dest_comp=1, indices=[BASE, RANGE_BASE])
709
710def atomic3(name):
711    intrinsic(name + "_deref", src_comp=[-1, 1, 1], dest_comp=1)
712    intrinsic(name, src_comp=[1, 1, 1], dest_comp=1, indices=[BASE, RANGE_BASE])
713
714atomic("atomic_counter_inc")
715atomic("atomic_counter_pre_dec")
716atomic("atomic_counter_post_dec")
717atomic("atomic_counter_read", flags=[CAN_ELIMINATE])
718atomic2("atomic_counter_add")
719atomic2("atomic_counter_min")
720atomic2("atomic_counter_max")
721atomic2("atomic_counter_and")
722atomic2("atomic_counter_or")
723atomic2("atomic_counter_xor")
724atomic2("atomic_counter_exchange")
725atomic3("atomic_counter_comp_swap")
726
727# Image load, store and atomic intrinsics.
728#
729# All image intrinsics come in three versions.  One which take an image target
730# passed as a deref chain as the first source, one which takes an index as the
731# first source, and one which takes a bindless handle as the first source.
732# In the first version, the image variable contains the memory and layout
733# qualifiers that influence the semantics of the intrinsic.  In the second and
734# third, the image format and access qualifiers are provided as constant
735# indices.  Up through GLSL ES 3.10, the image index source may only be a
736# constant array access.  GLSL ES 3.20 and GLSL 4.00 allow dynamically uniform
737# indexing.
738#
739# All image intrinsics take a four-coordinate vector and a sample index as
740# 2nd and 3rd sources, determining the location within the image that will be
741# accessed by the intrinsic.  Components not applicable to the image target
742# in use are undefined.  Image store takes an additional four-component
743# argument with the value to be written, and image atomic operations take
744# either one or two additional scalar arguments with the same meaning as in
745# the ARB_shader_image_load_store specification.
746#
747# The last source of many image intrinsics is the LOD. This source is zero
748# unless e.g. SPV_AMD_shader_image_load_store_lod is supported.
749def image(name, src_comp=[], extra_indices=[], **kwargs):
750    intrinsic("image_deref_" + name, src_comp=[-1] + src_comp,
751              indices=[IMAGE_DIM, IMAGE_ARRAY, FORMAT, ACCESS] + extra_indices, **kwargs)
752    intrinsic("image_" + name, src_comp=[1] + src_comp,
753              indices=[IMAGE_DIM, IMAGE_ARRAY, FORMAT, ACCESS, RANGE_BASE] + extra_indices, **kwargs)
754    intrinsic("bindless_image_" + name, src_comp=[-1] + src_comp,
755              indices=[IMAGE_DIM, IMAGE_ARRAY, FORMAT, ACCESS] + extra_indices, **kwargs)
756
757image("load", src_comp=[4, 1, 1], extra_indices=[DEST_TYPE], dest_comp=0, flags=[CAN_ELIMINATE])
758image("sparse_load", src_comp=[4, 1, 1], extra_indices=[DEST_TYPE], dest_comp=0, flags=[CAN_ELIMINATE])
759image("store", src_comp=[4, 1, 0, 1], extra_indices=[SRC_TYPE])
760image("atomic",  src_comp=[4, 1, 1], dest_comp=1, extra_indices=[ATOMIC_OP])
761image("atomic_swap", src_comp=[4, 1, 1, 1], dest_comp=1, extra_indices=[ATOMIC_OP])
762image("size",    dest_comp=0, src_comp=[1], flags=[CAN_ELIMINATE, CAN_REORDER])
763image("levels",  dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
764image("samples", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
765image("texel_address", dest_comp=1, src_comp=[4, 1],
766      flags=[CAN_ELIMINATE, CAN_REORDER])
767# This returns true if all samples within the pixel have equal color values.
768image("samples_identical", dest_comp=1, src_comp=[4], flags=[CAN_ELIMINATE])
769# Non-uniform access is not lowered for image_descriptor_amd.
770# dest_comp can be either 4 (buffer) or 8 (image).
771image("descriptor_amd", dest_comp=0, src_comp=[], flags=[CAN_ELIMINATE, CAN_REORDER])
772# CL-specific format queries
773image("format", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
774image("order", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
775# Multisample fragment mask load
776# src_comp[0] is same as image load src_comp[0]
777image("fragment_mask_load_amd", src_comp=[4], dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE, CAN_REORDER])
778
779# Vulkan descriptor set intrinsics
780#
781# The Vulkan API uses a different binding model from GL.  In the Vulkan
782# API, all external resources are represented by a tuple:
783#
784# (descriptor set, binding, array index)
785#
786# where the array index is the only thing allowed to be indirect.  The
787# vulkan_surface_index intrinsic takes the descriptor set and binding as
788# its first two indices and the array index as its source.  The third
789# index is a nir_variable_mode in case that's useful to the backend.
790#
791# The intended usage is that the shader will call vulkan_surface_index to
792# get an index and then pass that as the buffer index ubo/ssbo calls.
793#
794# The vulkan_resource_reindex intrinsic takes a resource index in src0
795# (the result of a vulkan_resource_index or vulkan_resource_reindex) which
796# corresponds to the tuple (set, binding, index) and computes an index
797# corresponding to tuple (set, binding, idx + src1).
798intrinsic("vulkan_resource_index", src_comp=[1], dest_comp=0,
799          indices=[DESC_SET, BINDING, DESC_TYPE],
800          flags=[CAN_ELIMINATE, CAN_REORDER])
801intrinsic("vulkan_resource_reindex", src_comp=[0, 1], dest_comp=0,
802          indices=[DESC_TYPE], flags=[CAN_ELIMINATE, CAN_REORDER])
803intrinsic("load_vulkan_descriptor", src_comp=[-1], dest_comp=0,
804          indices=[DESC_TYPE], flags=[CAN_ELIMINATE, CAN_REORDER])
805
806# atomic intrinsics
807#
808# All of these atomic memory operations read a value from memory, compute a new
809# value using one of the operations below, write the new value to memory, and
810# return the original value read.
811#
812# All variable operations take 2 sources except CompSwap that takes 3. These
813# sources represent:
814#
815# 0: A deref to the memory on which to perform the atomic
816# 1: The data parameter to the atomic function (i.e. the value to add
817#    in shared_atomic_add, etc).
818# 2: For CompSwap only: the second data parameter.
819#
820# All SSBO operations take 3 sources except CompSwap that takes 4. These
821# sources represent:
822#
823# 0: The SSBO buffer index (dynamically uniform in GLSL, possibly non-uniform
824#    with VK_EXT_descriptor_indexing).
825# 1: The offset into the SSBO buffer of the variable that the atomic
826#    operation will operate on.
827# 2: The data parameter to the atomic function (i.e. the value to add
828#    in ssbo_atomic_add, etc).
829# 3: For CompSwap only: the second data parameter.
830#
831# All shared (and task payload) variable operations take 2 sources
832# except CompSwap that takes 3.
833# These sources represent:
834#
835# 0: The offset into the shared variable storage region that the atomic
836#    operation will operate on.
837# 1: The data parameter to the atomic function (i.e. the value to add
838#    in shared_atomic_add, etc).
839# 2: For CompSwap only: the second data parameter.
840#
841# All global operations take 2 sources except CompSwap that takes 3. These
842# sources represent:
843#
844# 0: The memory address that the atomic operation will operate on.
845# 1: The data parameter to the atomic function (i.e. the value to add
846#    in shared_atomic_add, etc).
847# 2: For CompSwap only: the second data parameter.
848#
849# The 2x32 global variants use a vec2 for the memory address where component X
850# has the low 32-bit and component Y has the high 32-bit.
851#
852# IR3 global operations take 32b vec2 as memory address. IR3 doesn't support
853# float atomics.
854#
855# AGX global variants take a 64-bit base address plus a 32-bit offset in words.
856# The offset is sign-extended or zero-extended based on the SIGN_EXTEND index.
857
858intrinsic("deref_atomic",  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS, ATOMIC_OP])
859intrinsic("ssbo_atomic",  src_comp=[-1, 1, 1], dest_comp=1, indices=[ACCESS, ATOMIC_OP])
860intrinsic("shared_atomic",  src_comp=[1, 1], dest_comp=1, indices=[BASE, ATOMIC_OP])
861intrinsic("task_payload_atomic",  src_comp=[1, 1], dest_comp=1, indices=[BASE, ATOMIC_OP])
862intrinsic("global_atomic",  src_comp=[1, 1], dest_comp=1, indices=[ATOMIC_OP])
863intrinsic("global_atomic_2x32",  src_comp=[2, 1], dest_comp=1, indices=[ATOMIC_OP])
864intrinsic("global_atomic_amd",  src_comp=[1, 1, 1], dest_comp=1, indices=[BASE, ATOMIC_OP])
865intrinsic("global_atomic_ir3",  src_comp=[2, 1], dest_comp=1, indices=[BASE, ATOMIC_OP])
866intrinsic("global_atomic_agx",  src_comp=[1, 1, 1], dest_comp=1, indices=[ATOMIC_OP, SIGN_EXTEND])
867
868intrinsic("deref_atomic_swap",  src_comp=[-1, 1, 1], dest_comp=1, indices=[ACCESS, ATOMIC_OP])
869intrinsic("ssbo_atomic_swap",  src_comp=[-1, 1, 1, 1], dest_comp=1, indices=[ACCESS, ATOMIC_OP])
870intrinsic("shared_atomic_swap",  src_comp=[1, 1, 1], dest_comp=1, indices=[BASE, ATOMIC_OP])
871intrinsic("task_payload_atomic_swap",  src_comp=[1, 1, 1], dest_comp=1, indices=[BASE, ATOMIC_OP])
872intrinsic("global_atomic_swap",  src_comp=[1, 1, 1], dest_comp=1, indices=[ATOMIC_OP])
873intrinsic("global_atomic_swap_2x32",  src_comp=[2, 1, 1], dest_comp=1, indices=[ATOMIC_OP])
874intrinsic("global_atomic_swap_amd",  src_comp=[1, 1, 1, 1], dest_comp=1, indices=[BASE, ATOMIC_OP])
875intrinsic("global_atomic_swap_ir3",  src_comp=[2, 1, 1], dest_comp=1, indices=[BASE, ATOMIC_OP])
876intrinsic("global_atomic_swap_agx",  src_comp=[1, 1, 1, 1], dest_comp=1, indices=[ATOMIC_OP, SIGN_EXTEND])
877
878def system_value(name, dest_comp, indices=[], bit_sizes=[32], can_reorder=True):
879    flags = [CAN_ELIMINATE, CAN_REORDER] if can_reorder else [CAN_ELIMINATE]
880    intrinsic("load_" + name, [], dest_comp, indices,
881              flags=flags, sysval=True,
882              bit_sizes=bit_sizes)
883
884system_value("frag_coord", 4)
885# 16-bit integer vec2 of the pixel X/Y in the framebuffer.
886system_value("pixel_coord", 2, bit_sizes=[16])
887# Scalar load of frag_coord Z/W components (component=2 for Z, component=3 for
888# W). Backends can lower frag_coord to pixel_coord + frag_coord_zw, in case
889# X/Y is available as an integer but Z/W requires interpolation.
890system_value("frag_coord_zw", 1, indices=[COMPONENT])
891system_value("point_coord", 2)
892system_value("line_coord", 1)
893system_value("front_face", 1, bit_sizes=[1, 32])
894system_value("front_face_fsign", 1, bit_sizes=[32]) # front_face ? 1.0 : -1.0
895system_value("vertex_id", 1)
896system_value("vertex_id_zero_base", 1)
897system_value("first_vertex", 1)
898system_value("is_indexed_draw", 1)
899system_value("base_vertex", 1)
900system_value("instance_id", 1)
901system_value("base_instance", 1)
902system_value("draw_id", 1)
903system_value("sample_id", 1)
904# sample_id_no_per_sample is like sample_id but does not imply per-
905# sample shading.  See the lower_helper_invocation option.
906system_value("sample_id_no_per_sample", 1)
907system_value("sample_pos", 2)
908# sample_pos_or_center is like sample_pos but does not imply per-sample
909# shading.  When per-sample dispatch is not enabled, it returns (0.5, 0.5).
910system_value("sample_pos_or_center", 2)
911system_value("sample_mask_in", 1)
912system_value("primitive_id", 1)
913system_value("invocation_id", 1)
914system_value("tess_coord", 3)
915# First 2 components of tess_coord only
916system_value("tess_coord_xy", 2)
917system_value("tess_level_outer", 4)
918system_value("tess_level_inner", 2)
919system_value("tess_level_outer_default", 4)
920system_value("tess_level_inner_default", 2)
921system_value("patch_vertices_in", 1)
922system_value("local_invocation_id", 3)
923system_value("local_invocation_index", 1)
924# workgroup_id does not include the base_workgroup_id
925system_value("workgroup_id", 3)
926# The workgroup_index is intended for situations when a 3 dimensional
927# workgroup_id is not available on the HW, but a 1 dimensional index is.
928system_value("workgroup_index", 1)
929# API specific base added to the workgroup_id, e.g. baseGroup* of vkCmdDispatchBase
930system_value("base_workgroup_id", 3, bit_sizes=[32, 64])
931system_value("user_clip_plane", 4, indices=[UCP_ID])
932system_value("num_workgroups", 3)
933system_value("num_vertices", 1)
934# This can't be reordered because it's undefined after an invocation is demoted.
935system_value("helper_invocation", 1, bit_sizes=[1, 32], can_reorder=False)
936system_value("layer_id", 1)
937system_value("view_index", 1)
938system_value("subgroup_size", 1)
939system_value("subgroup_invocation", 1)
940
941# These intrinsics provide a bitmask for all invocations, with one bit per
942# invocation starting with the least significant bit, according to the
943# following table,
944#
945#    variable           equation for bit values
946#    ----------------   --------------------------------
947#    subgroup_eq_mask   bit index == subgroup_invocation
948#    subgroup_ge_mask   bit index >= subgroup_invocation
949#    subgroup_gt_mask   bit index >  subgroup_invocation
950#    subgroup_le_mask   bit index <= subgroup_invocation
951#    subgroup_lt_mask   bit index <  subgroup_invocation
952#
953# These correspond to gl_SubGroupEqMaskARB, etc. from GL_ARB_shader_ballot,
954# and the above documentation is "borrowed" from that extension spec.
955system_value("subgroup_eq_mask", 0, bit_sizes=[32, 64])
956system_value("subgroup_ge_mask", 0, bit_sizes=[32, 64])
957system_value("subgroup_gt_mask", 0, bit_sizes=[32, 64])
958system_value("subgroup_le_mask", 0, bit_sizes=[32, 64])
959system_value("subgroup_lt_mask", 0, bit_sizes=[32, 64])
960
961system_value("num_subgroups", 1)
962system_value("subgroup_id", 1)
963system_value("workgroup_size", 3)
964# note: the definition of global_invocation_id is based on
965# ((workgroup_id + base_workgroup_id) * workgroup_size) + local_invocation_id.
966system_value("global_invocation_id", 3, bit_sizes=[32, 64])
967# API specific base added to the global_invocation_id
968# e.g. global_work_offset of clEnqueueNDRangeKernel
969system_value("base_global_invocation_id", 3, bit_sizes=[32, 64])
970system_value("global_invocation_index", 1, bit_sizes=[32, 64])
971# threads per dimension in an invocation
972system_value("global_size", 3, bit_sizes=[32, 64])
973system_value("work_dim", 1)
974system_value("line_width", 1)
975system_value("aa_line_width", 1)
976# BASE=0 for global/shader, BASE=1 for local/function
977system_value("scratch_base_ptr", 0, bit_sizes=[32,64], indices=[BASE])
978system_value("constant_base_ptr", 0, bit_sizes=[32,64])
979system_value("shared_base_ptr", 0, bit_sizes=[32,64])
980system_value("global_base_ptr", 0, bit_sizes=[32,64])
981# Address and size of a transform feedback buffer, indexed by BASE
982system_value("xfb_address", 1, bit_sizes=[32,64], indices=[BASE])
983system_value("xfb_size", 1, bit_sizes=[32], indices=[BASE])
984
985# Address of the associated index buffer in a transform feedback program for an
986# indexed draw. This will be used so transform feedback can pull the gl_VertexID
987# from the index buffer.
988system_value("xfb_index_buffer", 1, bit_sizes=[32,64])
989
990system_value("frag_size", 2)
991system_value("frag_invocation_count", 1)
992# Whether smooth lines or polygon smoothing is enabled
993system_value("poly_line_smooth_enabled", 1, bit_sizes=[1])
994
995# System values for ray tracing.
996system_value("ray_launch_id", 3)
997system_value("ray_launch_size", 3)
998system_value("ray_world_origin", 3)
999system_value("ray_world_direction", 3)
1000system_value("ray_object_origin", 3)
1001system_value("ray_object_direction", 3)
1002system_value("ray_t_min", 1)
1003system_value("ray_t_max", 1)
1004system_value("ray_object_to_world", 3, indices=[COLUMN])
1005system_value("ray_world_to_object", 3, indices=[COLUMN])
1006system_value("ray_hit_kind", 1)
1007system_value("ray_flags", 1)
1008system_value("ray_geometry_index", 1)
1009system_value("ray_instance_custom_index", 1)
1010system_value("shader_record_ptr", 1, bit_sizes=[64])
1011system_value("cull_mask", 1)
1012system_value("ray_triangle_vertex_positions", 3, indices=[COLUMN])
1013
1014# Driver-specific viewport scale/offset parameters.
1015#
1016# VC4 and V3D need to emit a scaled version of the position in the vertex
1017# shaders for binning, and having system values lets us move the math for that
1018# into NIR.
1019#
1020# Panfrost needs to implement all coordinate transformation in the
1021# vertex shader; system values allow us to share this routine in NIR.
1022system_value("viewport_x_scale", 1)
1023system_value("viewport_y_scale", 1)
1024system_value("viewport_z_scale", 1)
1025system_value("viewport_x_offset", 1)
1026system_value("viewport_y_offset", 1)
1027system_value("viewport_z_offset", 1)
1028system_value("viewport_scale", 3)
1029system_value("viewport_offset", 3)
1030# Pack xy scale and offset into a vec4 load (used by AMD NGG primitive culling)
1031system_value("cull_triangle_viewport_xy_scale_and_offset_amd", 4)
1032system_value("cull_line_viewport_xy_scale_and_offset_amd", 4)
1033
1034# Blend constant color values.  Float values are clamped. Vectored versions are
1035# provided as well for driver convenience
1036
1037system_value("blend_const_color_r_float", 1)
1038system_value("blend_const_color_g_float", 1)
1039system_value("blend_const_color_b_float", 1)
1040system_value("blend_const_color_a_float", 1)
1041system_value("blend_const_color_rgba", 4)
1042system_value("blend_const_color_rgba8888_unorm", 1)
1043system_value("blend_const_color_aaaa8888_unorm", 1)
1044
1045# System values for gl_Color, for radeonsi which interpolates these in the
1046# shader prolog to handle two-sided color without recompiles and therefore
1047# doesn't handle these in the main shader part like normal varyings.
1048system_value("color0", 4)
1049system_value("color1", 4)
1050
1051# System value for internal compute shaders in radeonsi.
1052system_value("user_data_amd", 8)
1053
1054# In a fragment shader, the current sample mask. At the beginning of the shader,
1055# this is the same as load_sample_mask_in, but as the shader is executed, it may
1056# be affected by writes, discards, etc.
1057#
1058# No frontend generates this, but drivers may use it for internal lowerings.
1059intrinsic("load_sample_mask", [], 1, [], flags=[CAN_ELIMINATE], sysval=True,
1060          bit_sizes=[32])
1061
1062# Barycentric coordinate intrinsics.
1063#
1064# These set up the barycentric coordinates for a particular interpolation.
1065# The first four are for the simple cases: pixel, centroid, per-sample
1066# (at gl_SampleID), or pull model (1/W, 1/I, 1/J) at the pixel center. The next
1067# two handle interpolating at a specified sample location, or interpolating
1068# with a vec2 offset,
1069#
1070# The interp_mode index should be either the INTERP_MODE_SMOOTH or
1071# INTERP_MODE_NOPERSPECTIVE enum values.
1072#
1073# The vec2 value produced by these intrinsics is intended for use as the
1074# barycoord source of a load_interpolated_input intrinsic.
1075#
1076# The vec3 variants are intended to be used for input barycentric coordinates
1077# which are system values on most hardware, compared to the vec2 variants which
1078# interpolates input varyings.
1079
1080def barycentric(name, dst_comp, src_comp=[]):
1081    intrinsic("load_barycentric_" + name, src_comp=src_comp, dest_comp=dst_comp,
1082              indices=[INTERP_MODE], flags=[CAN_ELIMINATE, CAN_REORDER])
1083
1084# no sources.
1085barycentric("pixel", 2)
1086barycentric("coord_pixel", 3)
1087barycentric("centroid", 2)
1088barycentric("coord_centroid", 3)
1089barycentric("sample", 2)
1090barycentric("coord_sample", 3)
1091barycentric("model", 3)
1092# src[] = { sample_id }.
1093barycentric("at_sample", 2, [1])
1094barycentric("coord_at_sample", 3, [1])
1095# src[] = { offset.xy }.
1096barycentric("at_offset", 2, [2])
1097barycentric("at_offset_nv", 2, [1])
1098barycentric("coord_at_offset", 3, [2])
1099
1100# Load sample position:
1101#
1102# Takes a sample # and returns a sample position.  Used for lowering
1103# interpolateAtSample() to interpolateAtOffset()
1104intrinsic("load_sample_pos_from_id", src_comp=[1], dest_comp=2,
1105          flags=[CAN_ELIMINATE, CAN_REORDER])
1106
1107intrinsic("load_persp_center_rhw_ir3", dest_comp=1,
1108          flags=[CAN_ELIMINATE, CAN_REORDER])
1109
1110# Load texture scaling values:
1111#
1112# Takes a sampler # and returns 1/size values for multiplying to normalize
1113# texture coordinates.  Used for lowering rect textures.
1114intrinsic("load_texture_scale", src_comp=[1], dest_comp=2,
1115          flags=[CAN_ELIMINATE, CAN_REORDER])
1116
1117# Gets the texture src. This intrinsic will be lowered once functions have
1118# been inlined and we know if the src is bindless or not.
1119intrinsic("deref_texture_src", src_comp=[1], dest_comp=1,
1120          flags=[CAN_ELIMINATE, CAN_REORDER])
1121
1122# Fragment shader input interpolation delta intrinsic.
1123#
1124# For hw where fragment shader input interpolation is handled in shader, the
1125# load_fs_input_interp deltas intrinsics can be used to load the input deltas
1126# used for interpolation as follows:
1127#
1128#    vec3 iid = load_fs_input_interp_deltas(varying_slot)
1129#    vec2 bary = load_barycentric_*(...)
1130#    float result = iid.x + iid.y * bary.y + iid.z * bary.x
1131
1132intrinsic("load_fs_input_interp_deltas", src_comp=[1], dest_comp=3,
1133          indices=[BASE, COMPONENT, IO_SEMANTICS], flags=[CAN_ELIMINATE, CAN_REORDER])
1134
1135# Load operations pull data from some piece of GPU memory.  All load
1136# operations operate in terms of offsets into some piece of theoretical
1137# memory.  Loads from externally visible memory (UBO and SSBO) simply take a
1138# byte offset as a source.  Loads from opaque memory (uniforms, inputs, etc.)
1139# take a base+offset pair where the nir_intrinsic_base() gives the location
1140# of the start of the variable being loaded and and the offset source is a
1141# offset into that variable.
1142#
1143# Uniform load operations have a nir_intrinsic_range() index that specifies the
1144# range (starting at base) of the data from which we are loading.  If
1145# range == 0, then the range is unknown.
1146#
1147# UBO load operations have a nir_intrinsic_range_base() and
1148# nir_intrinsic_range() that specify the byte range [range_base,
1149# range_base+range] of the UBO that the src offset access must lie within.
1150#
1151# Some load operations such as UBO/SSBO load and per_vertex loads take an
1152# additional source to specify which UBO/SSBO/vertex to load from.
1153#
1154# The exact address type depends on the lowering pass that generates the
1155# load/store intrinsics.  Typically, this is vec4 units for things such as
1156# varying slots and float units for fragment shader inputs.  UBO and SSBO
1157# offsets are always in bytes.
1158
1159def load(name, src_comp, indices=[], flags=[]):
1160    intrinsic("load_" + name, src_comp, dest_comp=0, indices=indices,
1161              flags=flags)
1162
1163# src[] = { offset }.
1164load("uniform", [1], [BASE, RANGE, DEST_TYPE], [CAN_ELIMINATE, CAN_REORDER])
1165# src[] = { buffer_index, offset }.
1166load("ubo", [-1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET, RANGE_BASE, RANGE], flags=[CAN_ELIMINATE, CAN_REORDER])
1167# src[] = { buffer_index, offset in vec4 units }.  base is also in vec4 units.
1168load("ubo_vec4", [-1, 1], [ACCESS, BASE, COMPONENT], flags=[CAN_ELIMINATE, CAN_REORDER])
1169# src[] = { offset }.
1170load("input", [1], [BASE, RANGE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
1171# src[] = { vertex_id, offset }.
1172load("input_vertex", [1, 1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
1173# src[] = { vertex, offset }.
1174load("per_vertex_input", [1, 1], [BASE, RANGE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
1175# src[] = { barycoord, offset }.
1176load("interpolated_input", [2, 1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
1177# src[] = { offset }.
1178load("per_primitive_input", [1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
1179
1180# src[] = { buffer_index, offset }.
1181load("ssbo", [-1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1182# src[] = { buffer_index, offset }
1183load("ssbo_address", [1, 1], [], [CAN_ELIMINATE, CAN_REORDER])
1184# src[] = { offset }.
1185load("output", [1], [BASE, RANGE, COMPONENT, DEST_TYPE, IO_SEMANTICS], flags=[CAN_ELIMINATE])
1186# src[] = { vertex, offset }.
1187load("per_vertex_output", [1, 1], [BASE, RANGE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE])
1188# src[] = { view_index, offset }.
1189# when nir_shader_compiler_options::compact_view_index is set, the view_index
1190# src refers to the Nth enabled view, and do not correspond directly to
1191# gl_ViewIndex. See the compact_view_index docs for more details.
1192load("per_view_output", [1, 1], [BASE, RANGE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE])
1193# src[] = { primitive, offset }.
1194load("per_primitive_output", [1, 1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE])
1195# src[] = { offset }.
1196load("shared", [1], [BASE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1197# src[] = { offset }.
1198load("task_payload", [1], [BASE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1199# src[] = { offset }.
1200load("push_constant", [1], [BASE, RANGE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE, CAN_REORDER])
1201# src[] = { offset }.
1202load("constant", [1], [BASE, RANGE, ACCESS, ALIGN_MUL, ALIGN_OFFSET],
1203     [CAN_ELIMINATE, CAN_REORDER])
1204# src[] = { address }.
1205load("global", [1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1206# src[] = { address }.
1207load("global_2x32", [2], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1208# src[] = { address }.
1209load("global_constant", [1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET],
1210     [CAN_ELIMINATE, CAN_REORDER])
1211# src[] = { base_address, offset }.
1212load("global_constant_offset", [1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET],
1213     [CAN_ELIMINATE, CAN_REORDER])
1214# src[] = { base_address, offset, bound }.
1215load("global_constant_bounded", [1, 1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET],
1216     [CAN_ELIMINATE, CAN_REORDER])
1217# src[] = { address }.
1218load("kernel_input", [1], [BASE, RANGE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE, CAN_REORDER])
1219# src[] = { offset }.
1220load("scratch", [1], [ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1221
1222# Stores work the same way as loads, except now the first source is the value
1223# to store and the second (and possibly third) source specify where to store
1224# the value.  SSBO and shared memory stores also have a
1225# nir_intrinsic_write_mask()
1226
1227def store(name, srcs, indices=[], flags=[]):
1228    intrinsic("store_" + name, [0] + srcs, indices=indices, flags=flags)
1229
1230# src[] = { value, offset }.
1231store("output", [1], [BASE, RANGE, WRITE_MASK, COMPONENT, SRC_TYPE, IO_SEMANTICS, IO_XFB, IO_XFB2])
1232# src[] = { value, vertex, offset }.
1233store("per_vertex_output", [1, 1], [BASE, RANGE, WRITE_MASK, COMPONENT, SRC_TYPE, IO_SEMANTICS])
1234# src[] = { value, view_index, offset }.
1235store("per_view_output", [1, 1], [BASE, RANGE, WRITE_MASK, COMPONENT, SRC_TYPE, IO_SEMANTICS])
1236# src[] = { value, primitive, offset }.
1237store("per_primitive_output", [1, 1], [BASE, RANGE, WRITE_MASK, COMPONENT, SRC_TYPE, IO_SEMANTICS])
1238# src[] = { value, block_index, offset }
1239store("ssbo", [-1, 1], [WRITE_MASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
1240# src[] = { value, offset }.
1241store("shared", [1], [BASE, WRITE_MASK, ALIGN_MUL, ALIGN_OFFSET])
1242# src[] = { value, offset }.
1243store("task_payload", [1], [BASE, WRITE_MASK, ALIGN_MUL, ALIGN_OFFSET])
1244# src[] = { value, address }.
1245store("global", [1], [WRITE_MASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
1246# src[] = { value, address }.
1247store("global_2x32", [2], [WRITE_MASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
1248# src[] = { value, offset }.
1249store("scratch", [1], [ALIGN_MUL, ALIGN_OFFSET, WRITE_MASK])
1250
1251# Intrinsic to load/store from the call stack.
1252# BASE is the offset relative to the current position of the stack
1253# src[] = { }.
1254intrinsic("load_stack", [], dest_comp=0,
1255          indices=[BASE, ALIGN_MUL, ALIGN_OFFSET, CALL_IDX, VALUE_ID],
1256          flags=[CAN_ELIMINATE])
1257# src[] = { value }.
1258intrinsic("store_stack", [0],
1259          indices=[BASE, ALIGN_MUL, ALIGN_OFFSET, WRITE_MASK, CALL_IDX, VALUE_ID])
1260
1261
1262# A bit field to implement SPIRV FragmentShadingRateKHR
1263# bit | name              | description
1264#   0 | Vertical2Pixels   | Fragment invocation covers 2 pixels vertically
1265#   1 | Vertical4Pixels   | Fragment invocation covers 4 pixels vertically
1266#   2 | Horizontal2Pixels | Fragment invocation covers 2 pixels horizontally
1267#   3 | Horizontal4Pixels | Fragment invocation covers 4 pixels horizontally
1268intrinsic("load_frag_shading_rate", dest_comp=1, bit_sizes=[32],
1269          flags=[CAN_ELIMINATE, CAN_REORDER])
1270
1271# Whether the rasterized fragment is fully covered by the generating primitive.
1272system_value("fully_covered", dest_comp=1, bit_sizes=[1])
1273
1274# OpenCL printf instruction
1275# First source is an index to the format string (u_printf_info element of the shader)
1276# Second source is a deref to a struct containing the args
1277# Dest is success or failure
1278intrinsic("printf", src_comp=[1, 1], dest_comp=1, bit_sizes=[32])
1279# Since most drivers will want to lower to just dumping args
1280# in a buffer, nir_lower_printf will do that, but requires
1281# the driver to at least provide a base location and size
1282system_value("printf_buffer_address", 1, bit_sizes=[32,64])
1283system_value("printf_buffer_size", 1, bit_sizes=[32])
1284# If driver wants to have all printfs from various shaders merged into a
1285# single output buffer, it needs each shader to have its own base identifier
1286# from which each printf is indexed.
1287system_value("printf_base_identifier", 1, bit_sizes=[32])
1288# Abort the program, triggering device fault. The invoking thread halts
1289# immediately. Other threads eventually terminate.
1290#
1291# This does not take a payload, payloads should be specified with a preceding
1292# printf. After lowering, the intrinsic will set an aborted? bit in the printf
1293# buffer. This avoids a separate abort buffer.
1294intrinsic("printf_abort")
1295
1296# Mesh shading MultiView intrinsics
1297system_value("mesh_view_count", 1)
1298load("mesh_view_indices", [1], [BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
1299
1300# Used to pass values from the preamble to the main shader.
1301# This should use something similar to Vulkan push constants and load_preamble
1302# should be relatively cheap.
1303# For now we only support accesses with a constant offset.
1304load("preamble", [], indices=[BASE], flags=[CAN_ELIMINATE, CAN_REORDER])
1305store("preamble", [], indices=[BASE])
1306
1307# A 64-bit bitfield indexed by I/O location storing 1 in bits corresponding to
1308# varyings that have the flat interpolation specifier in the fragment shader and
1309# 0 otherwise
1310system_value("flat_mask", 1, bit_sizes=[64])
1311
1312# Whether provoking vertex mode is last
1313system_value("provoking_last", 1)
1314
1315# SPV_KHR_cooperative_matrix.
1316#
1317# Cooperative matrices are referred through derefs to variables,
1318# the destination of the operations appears as the first source,
1319# ordering follows SPIR-V operation.
1320#
1321# Load/Store include an extra source for stride, since that
1322# can be a _dynamically_ uniform value.
1323#
1324# Length takes a type not a value, that's encoded as a MATRIX_DESC.
1325intrinsic("cmat_construct", src_comp=[-1, 1])
1326intrinsic("cmat_load", src_comp=[-1, -1, 1], indices=[MATRIX_LAYOUT])
1327intrinsic("cmat_store", src_comp=[-1, -1, 1], indices=[MATRIX_LAYOUT])
1328intrinsic("cmat_length", src_comp=[], dest_comp=1, indices=[CMAT_DESC], bit_sizes=[32])
1329intrinsic("cmat_muladd", src_comp=[-1, -1, -1, -1], indices=[SATURATE, CMAT_SIGNED_MASK])
1330intrinsic("cmat_unary_op", src_comp=[-1, -1], indices=[ALU_OP])
1331intrinsic("cmat_binary_op", src_comp=[-1, -1, -1], indices=[ALU_OP])
1332intrinsic("cmat_scalar_op", src_comp=[-1, -1, -1], indices=[ALU_OP])
1333intrinsic("cmat_bitcast", src_comp=[-1, -1])
1334intrinsic("cmat_extract", src_comp=[-1, 1], dest_comp=1)
1335intrinsic("cmat_insert", src_comp=[-1, 1, -1, 1])
1336intrinsic("cmat_copy", src_comp=[-1, -1])
1337
1338# IR3-specific version of most SSBO intrinsics. The only different
1339# compare to the originals is that they add an extra source to hold
1340# the dword-offset, which is needed by the backend code apart from
1341# the byte-offset already provided by NIR in one of the sources.
1342#
1343# NIR lowering pass 'ir3_nir_lower_io_offset' will replace the
1344# original SSBO intrinsics by these, placing the computed
1345# dword-offset always in the last source.
1346#
1347# The float versions are not handled because those are not supported
1348# by the backend.
1349store("ssbo_ir3", [1, 1, 1],
1350      indices=[BASE, WRITE_MASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
1351load("ssbo_ir3",  [1, 1, 1],
1352     indices=[BASE, ACCESS, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN_ELIMINATE])
1353intrinsic("ssbo_atomic_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1,
1354          indices=[ACCESS, ATOMIC_OP])
1355intrinsic("ssbo_atomic_swap_ir3",  src_comp=[1, 1, 1, 1, 1], dest_comp=1,
1356          indices=[ACCESS, ATOMIC_OP])
1357
1358# System values for freedreno geometry shaders.
1359system_value("vs_primitive_stride_ir3", 1)
1360system_value("vs_vertex_stride_ir3", 1)
1361system_value("gs_header_ir3", 1)
1362system_value("primitive_location_ir3", 1, indices=[DRIVER_LOCATION])
1363
1364# System values for freedreno tessellation shaders.
1365system_value("hs_patch_stride_ir3", 1)
1366system_value("tess_factor_base_ir3", 2)
1367system_value("tess_param_base_ir3", 2)
1368system_value("tcs_header_ir3", 1)
1369system_value("rel_patch_id_ir3", 1)
1370
1371# System values for freedreno compute shaders.
1372system_value("subgroup_id_shift_ir3", 1)
1373
1374# System values for freedreno fragment shaders.
1375intrinsic("load_frag_coord_unscaled_ir3", dest_comp=4,
1376          flags=[CAN_ELIMINATE, CAN_REORDER], bit_sizes=[32])
1377
1378# Per-view gl_FragSizeEXT and gl_FragCoord offset.
1379intrinsic("load_frag_size_ir3", src_comp=[1], dest_comp=2, indices=[RANGE],
1380        flags=[CAN_ELIMINATE, CAN_REORDER], bit_sizes=[32])
1381intrinsic("load_frag_offset_ir3", src_comp=[1], dest_comp=2, indices=[RANGE],
1382        flags=[CAN_ELIMINATE, CAN_REORDER], bit_sizes=[32])
1383
1384# IR3-specific load/store intrinsics. These access a buffer used to pass data
1385# between geometry stages - perhaps it's explicit access to the vertex cache.
1386
1387# src[] = { value, offset }.
1388store("shared_ir3", [1], [BASE, ALIGN_MUL, ALIGN_OFFSET])
1389# src[] = { offset }.
1390load("shared_ir3", [1], [BASE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1391
1392# IR3-specific load/store global intrinsics. They take a 64-bit base address
1393# and a 32-bit offset.  The hardware will add the base and the offset, which
1394# saves us from doing 64-bit math on the base address.
1395
1396# src[] = { value, address(vec2 of hi+lo uint32_t), offset }.
1397# const_index[] = { write_mask, align_mul, align_offset }
1398store("global_ir3", [2, 1], indices=[ACCESS, ALIGN_MUL, ALIGN_OFFSET])
1399# src[] = { address(vec2 of hi+lo uint32_t), offset }.
1400# const_index[] = { access, align_mul, align_offset }
1401# the alignment applies to the base address
1402load("global_ir3", [2, 1], indices=[ACCESS, ALIGN_MUL, ALIGN_OFFSET, RANGE_BASE, RANGE], flags=[CAN_ELIMINATE])
1403
1404# Etnaviv-specific load/glboal intrinsics. They take a 32-bit base address and
1405# a 32-bit offset, which doesn't need to be an immediate.
1406# src[] = { value, address, 32-bit offset }.
1407store("global_etna", [1, 1], [WRITE_MASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
1408# src[] = { address, 32-bit offset }.
1409load("global_etna", [1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1410
1411# IR3-specific bindless handle specifier. Similar to vulkan_resource_index, but
1412# without the binding because the hardware expects a single flattened index
1413# rather than a (binding, index) pair. We may also want to use this with GL.
1414# Note that this doesn't actually turn into a HW instruction.
1415intrinsic("bindless_resource_ir3", [1], dest_comp=1, indices=[DESC_SET], flags=[CAN_ELIMINATE, CAN_REORDER])
1416
1417# IR3-specific intrinsics for shader preamble. These are meant to be used like
1418# this:
1419#
1420# if (preamble_start()) {
1421#    if (subgroupElect()) {
1422#       // preamble
1423#       ...
1424#       preamble_end();
1425#    }
1426# }
1427# // main shader
1428# ...
1429
1430intrinsic("preamble_start_ir3", [], dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
1431
1432barrier("preamble_end_ir3")
1433
1434# IR3-specific intrinsic to choose any invocation. This is implemented the same
1435# as elect, except that it doesn't require helper invocations. Used by preambles.
1436intrinsic("elect_any_ir3", dest_comp=1, flags=[CAN_ELIMINATE])
1437
1438# IR3-specific intrinsic for stc. Should be used in the shader preamble.
1439store("const_ir3", [], indices=[BASE])
1440
1441# IR3-specific intrinsic for loading from a const reg.
1442load("const_ir3", [1], indices=[BASE], flags=[CAN_ELIMINATE, CAN_REORDER])
1443
1444# IR3-specific intrinsic for ldc.k. Copies UBO to constant file.
1445# base is the const file base in components, range is the amount to copy in
1446# vec4's.
1447intrinsic("copy_ubo_to_uniform_ir3", [1, 1], indices=[BASE, RANGE])
1448
1449# IR3-specific intrinsic for ldg.k.
1450# base is an offset to apply to the address in bytes, range_base is the
1451# const file base in components, range is the amount to copy in vec4's.
1452intrinsic("copy_global_to_uniform_ir3", [2], indices=[BASE, RANGE_BASE, RANGE])
1453
1454# IR3-specific intrinsic for stsc. Loads from push consts to constant file
1455# Should be used in the shader preamble.
1456intrinsic("copy_push_const_to_uniform_ir3", [1], indices=[BASE, RANGE])
1457
1458intrinsic("brcst_active_ir3", dest_comp=1, src_comp=[1, 1], bit_sizes=src0,
1459          indices=[CLUSTER_SIZE])
1460intrinsic("reduce_clusters_ir3", dest_comp=1, src_comp=[1], bit_sizes=src0,
1461          indices=[REDUCTION_OP])
1462intrinsic("inclusive_scan_clusters_ir3", dest_comp=1, src_comp=[1],
1463          bit_sizes=src0, indices=[REDUCTION_OP])
1464intrinsic("exclusive_scan_clusters_ir3", dest_comp=1, src_comp=[1, 1],
1465          bit_sizes=src0, indices=[REDUCTION_OP])
1466
1467# Like shuffle_{xor,up,down} except with a uniform index. Necessary since the
1468# ir3 shfl instruction doesn't work with divergent indices.
1469intrinsic("shuffle_xor_uniform_ir3", src_comp=[0, 1], dest_comp=0,
1470          bit_sizes=src0, flags=[CAN_ELIMINATE])
1471intrinsic("shuffle_up_uniform_ir3", src_comp=[0, 1], dest_comp=0,
1472          bit_sizes=src0, flags=[CAN_ELIMINATE])
1473intrinsic("shuffle_down_uniform_ir3", src_comp=[0, 1], dest_comp=0,
1474          bit_sizes=src0, flags=[CAN_ELIMINATE])
1475
1476# IR3-specific intrinsics for prefetching descriptors in preambles.
1477intrinsic("prefetch_sam_ir3", [1, 1], flags=[CAN_REORDER])
1478intrinsic("prefetch_tex_ir3", [1], flags=[CAN_REORDER])
1479intrinsic("prefetch_ubo_ir3", [1], flags=[CAN_REORDER])
1480
1481# Panfrost-specific intrinsic for loading vertex attributes. Takes explicit
1482# vertex and instance IDs which we need in order to implement vertex attribute
1483# divisor with non-zero base instance on v9+.
1484# src[] = { vertex_id, instance_id, offset }
1485load("attribute_pan", [1, 1, 1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
1486
1487# Panfrost-specific intrinsics for accessing the raw vertex ID and the
1488# associated offset such that
1489#   vertex_id = raw_vertex_id_pan + raw_vertex_offset_pan
1490# The raw vertex ID differs from the zero-based vertex ID in that, in an index
1491# draw, it is offset by the minimum vertex ID in the index buffer range
1492# covered by the draw
1493system_value("raw_vertex_id_pan", 1)
1494system_value("raw_vertex_offset_pan", 1)
1495
1496# Intrinsics used by the Midgard/Bifrost blend pipeline. These are defined
1497# within a blend shader to read/write the raw value from the tile buffer,
1498# without applying any format conversion in the process. If the shader needs
1499# usable pixel values, it must apply format conversions itself.
1500#
1501# These definitions are generic, but they are explicitly vendored to prevent
1502# other drivers from using them, as their semantics is defined in terms of the
1503# Midgard/Bifrost hardware tile buffer and may not line up with anything sane.
1504# One notable divergence is sRGB, which is asymmetric: raw_input_pan requires
1505# an sRGB->linear conversion, but linear values should be written to
1506# raw_output_pan and the hardware handles linear->sRGB.
1507#
1508# store_raw_output_pan is used only for blend shaders, and writes out only a
1509# single 128-bit chunk. To support multisampling, the BASE index specifies the
1510# bas sample index written out.
1511
1512# src[] = { value }
1513store("raw_output_pan", [], [IO_SEMANTICS, BASE])
1514store("combined_output_pan", [1, 1, 1, 4], [IO_SEMANTICS, COMPONENT, SRC_TYPE, DEST_TYPE])
1515load("raw_output_pan", [1], [IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
1516
1517# Like the frag_coord_zw intrinsic, but takes a barycentric. This is needed for
1518# noperspective lowering.
1519# src[] = { barycoord }
1520intrinsic("load_frag_coord_zw_pan", [2], dest_comp=1, indices=[COMPONENT], flags=[CAN_ELIMINATE, CAN_REORDER], bit_sizes=[32])
1521
1522# Loads the sampler paramaters <min_lod, max_lod, lod_bias>
1523# src[] = { sampler_index }
1524load("sampler_lod_parameters_pan", [1], flags=[CAN_ELIMINATE, CAN_REORDER])
1525
1526# Like load_output but using a specified render target conversion descriptor
1527load("converted_output_pan", [1], indices=[DEST_TYPE, IO_SEMANTICS], flags=[CAN_ELIMINATE])
1528
1529# Load the render target conversion descriptor for a given render target given
1530# in the BASE index. Converts to a type with size given by the source type.
1531# Valid in fragment and blend stages.
1532system_value("rt_conversion_pan", 1, indices=[BASE, SRC_TYPE], bit_sizes=[32])
1533
1534# Loads the sample position array on Bifrost, in a packed Arm-specific format
1535system_value("sample_positions_pan", 1, bit_sizes=[64])
1536
1537# In a fragment shader, is the framebuffer single-sampled? 0/~0 bool
1538system_value("multisampled_pan", 1, bit_sizes=[32])
1539
1540# In a vertex shader, a bitfield of varying slots that use noperspective
1541# interpolation in the linked fragment shader. Since special slots cannot be
1542# noperspective, this is 32 bits and starts from VARYING_SLOT_VAR0.
1543system_value("noperspective_varyings_pan", 1, bit_sizes=[32])
1544
1545# R600 specific instrincs
1546#
1547# location where the tesselation data is stored in LDS
1548system_value("tcs_in_param_base_r600", 4)
1549system_value("tcs_out_param_base_r600", 4)
1550system_value("tcs_rel_patch_id_r600", 1)
1551system_value("tcs_tess_factor_base_r600", 1)
1552
1553# load as many components as needed giving per-component addresses
1554intrinsic("load_local_shared_r600", src_comp=[0], dest_comp=0, indices = [], flags = [CAN_ELIMINATE])
1555
1556store("local_shared_r600", [1], [WRITE_MASK])
1557store("tf_r600", [])
1558
1559# AMD GCN/RDNA specific intrinsics
1560
1561# This barrier is a hint that prevents moving the instruction that computes
1562# src after this barrier. It's a constraint for the instruction scheduler.
1563# Otherwise it's identical to a move instruction.
1564# The VGPR version forces the src value to be stored in a VGPR, while the SGPR
1565# version enforces an SGPR.
1566intrinsic("optimization_barrier_vgpr_amd", dest_comp=0, src_comp=[0],
1567          flags=[CAN_ELIMINATE])
1568intrinsic("optimization_barrier_sgpr_amd", dest_comp=0, src_comp=[0],
1569          flags=[CAN_ELIMINATE])
1570
1571# These are no-op intrinsics used as a simple source and user of SSA defs for testing.
1572intrinsic("unit_test_amd", src_comp=[0], indices=[BASE])
1573intrinsic("unit_test_uniform_amd", dest_comp=0, indices=[BASE])
1574intrinsic("unit_test_divergent_amd", dest_comp=0, indices=[BASE])
1575
1576# Untyped buffer load/store instructions of arbitrary length.
1577# src[] = { descriptor, vector byte offset, scalar byte offset, index offset }
1578# The index offset is multiplied by the stride in the descriptor.
1579# The vector/scalar offsets are in bytes, BASE is a constant byte offset.
1580intrinsic("load_buffer_amd", src_comp=[4, 1, 1, 1], dest_comp=0, indices=[BASE, MEMORY_MODES, ACCESS], flags=[CAN_ELIMINATE])
1581# src[] = { store value, descriptor, vector byte offset, scalar byte offset, index offset }
1582intrinsic("store_buffer_amd", src_comp=[0, 4, 1, 1, 1], indices=[BASE, WRITE_MASK, MEMORY_MODES, ACCESS])
1583
1584# Typed buffer load of arbitrary length, using a specified format.
1585# src[] = { descriptor, vector byte offset, scalar byte offset, index offset }
1586#
1587# The compiler backend is responsible for emitting correct HW instructions according to alignment, range etc.
1588# Users of this intrinsic must ensure that the first component being loaded is really the first component
1589# of the specified format, because range analysis assumes this.
1590# The size of the specified format also determines the memory range that this instruction is allowed to access.
1591#
1592# The index offset is multiplied by the stride in the descriptor, if any.
1593# The vector/scalar offsets are in bytes, BASE is a constant byte offset.
1594intrinsic("load_typed_buffer_amd", src_comp=[4, 1, 1, 1], dest_comp=0, indices=[BASE, MEMORY_MODES, ACCESS, FORMAT, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN_ELIMINATE])
1595
1596# src[] = { address, unsigned 32-bit offset }.
1597load("global_amd", [1, 1], indices=[BASE, ACCESS, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN_ELIMINATE])
1598# src[] = { value, address, unsigned 32-bit offset }.
1599store("global_amd", [1, 1], indices=[BASE, ACCESS, ALIGN_MUL, ALIGN_OFFSET, WRITE_MASK])
1600
1601# Same as shared_atomic_add, but with GDS. src[] = {store_val, gds_addr, m0}
1602intrinsic("gds_atomic_add_amd",  src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
1603
1604# Optimized shared_atomic_add (1/-1) with constant address
1605# returning the uniform pre-op value for all invocations.
1606intrinsic("shared_append_amd",  src_comp=[], dest_comp=1, bit_sizes=[32], indices=[BASE])
1607intrinsic("shared_consume_amd",  src_comp=[], dest_comp=1, bit_sizes=[32], indices=[BASE])
1608
1609# src[] = { sample_id, num_samples }
1610intrinsic("load_sample_positions_amd", src_comp=[1, 1], dest_comp=2, flags=[CAN_ELIMINATE, CAN_REORDER])
1611
1612# Descriptor where TCS outputs are stored for TES
1613system_value("ring_tess_offchip_amd", 4)
1614system_value("ring_tess_offchip_offset_amd", 1)
1615# Descriptor where TCS outputs are stored for the HW tessellator
1616system_value("ring_tess_factors_amd", 4)
1617system_value("ring_tess_factors_offset_amd", 1)
1618# Descriptor where ES outputs are stored for GS to read on GFX6-8
1619system_value("ring_esgs_amd", 4)
1620system_value("ring_es2gs_offset_amd", 1)
1621# Address of the task shader draw ring (used for VARYING_SLOT_TASK_COUNT)
1622system_value("ring_task_draw_amd", 4)
1623# Address of the task shader payload ring (used for all other outputs)
1624system_value("ring_task_payload_amd", 4)
1625# Address of the mesh shader scratch ring (used for excess mesh shader outputs)
1626system_value("ring_mesh_scratch_amd", 4)
1627system_value("ring_mesh_scratch_offset_amd", 1)
1628# Pointer into the draw and payload rings
1629system_value("task_ring_entry_amd", 1)
1630# Descriptor where NGG attributes are stored on GFX11.
1631system_value("ring_attr_amd", 4)
1632system_value("ring_attr_offset_amd", 1)
1633
1634# Load provoking vertex info
1635system_value("provoking_vtx_amd", 1)
1636
1637# Load rasterization primitive
1638system_value("rasterization_primitive_amd", 1);
1639
1640# Number of patches processed by each TCS workgroup
1641system_value("tcs_num_patches_amd", 1)
1642# Whether TCS should store tessellation level outputs for TES to read
1643system_value("tcs_tess_levels_to_tes_amd", dest_comp=1, bit_sizes=[1])
1644# Tessellation primitive mode for TCS
1645system_value("tcs_primitive_mode_amd", 1)
1646# Relative tessellation patch ID within the current workgroup
1647system_value("tess_rel_patch_id_amd", 1)
1648# Vertex offsets used for GS per-vertex inputs
1649system_value("gs_vertex_offset_amd", 1, [BASE])
1650# Number of rasterization samples
1651system_value("rasterization_samples_amd", 1)
1652
1653# Descriptor where GS outputs are stored for GS copy shader to read on GFX6-9
1654system_value("ring_gsvs_amd", 4, indices=[STREAM_ID])
1655# Write offset in gsvs ring for legacy GS shader
1656system_value("ring_gs2vs_offset_amd", 1)
1657
1658# Streamout configuration
1659system_value("streamout_config_amd", 1)
1660# Position to write within the streamout buffers
1661system_value("streamout_write_index_amd", 1)
1662# Offset to write within a streamout buffer
1663system_value("streamout_offset_amd", 1, indices=[BASE])
1664
1665# AMD merged shader intrinsics
1666
1667# Whether the current invocation index in the subgroup is less than the source. The source must be
1668# subgroup uniform and the 8 bits starting at the base bit must be less than or equal to the wave size.
1669intrinsic("is_subgroup_invocation_lt_amd", src_comp=[1], dest_comp=1, bit_sizes=[1], indices=[BASE], flags=[CAN_ELIMINATE])
1670
1671# AMD NGG intrinsics
1672
1673# Number of initial input vertices in the current workgroup.
1674system_value("workgroup_num_input_vertices_amd", 1)
1675# Number of initial input primitives in the current workgroup.
1676system_value("workgroup_num_input_primitives_amd", 1)
1677# For NGG passthrough mode only. Pre-packed argument for export_primitive_amd.
1678system_value("packed_passthrough_primitive_amd", 1)
1679# Whether NGG should execute shader query for pipeline statistics.
1680system_value("pipeline_stat_query_enabled_amd", dest_comp=1, bit_sizes=[1])
1681# Whether NGG should execute shader query for primitive generated.
1682system_value("prim_gen_query_enabled_amd", dest_comp=1, bit_sizes=[1])
1683# Whether NGG should execute shader query for primitive streamouted.
1684system_value("prim_xfb_query_enabled_amd", dest_comp=1, bit_sizes=[1])
1685# 64-bit memory address to struct {uint32_t ordered_id; uint32_t dwords_written;}[4]
1686system_value("xfb_state_address_gfx12_amd", dest_comp=1, bit_sizes=[64])
1687# Merged wave info. Bits 0-7 are the ES thread count, 8-15 are the GS thread count, 16-24 is the
1688# GS Wave ID, 24-27 is the wave index in the workgroup, and 28-31 is the workgroup size in waves.
1689system_value("merged_wave_info_amd", dest_comp=1)
1690# Global ID for GS waves on GCN/RDNA legacy GS.
1691system_value("gs_wave_id_amd", dest_comp=1)
1692# Whether the shader should clamp vertex color outputs to [0, 1].
1693system_value("clamp_vertex_color_amd", dest_comp=1, bit_sizes=[1])
1694# Whether the shader should cull front facing triangles.
1695intrinsic("load_cull_front_face_enabled_amd", dest_comp=1, bit_sizes=[1], flags=[CAN_ELIMINATE])
1696# Whether the shader should cull back facing triangles.
1697intrinsic("load_cull_back_face_enabled_amd", dest_comp=1, bit_sizes=[1], flags=[CAN_ELIMINATE])
1698# True if face culling should use CCW (false if CW).
1699intrinsic("load_cull_ccw_amd", dest_comp=1, bit_sizes=[1], flags=[CAN_ELIMINATE])
1700# Whether the shader should cull small triangles that are not visible in a pixel.
1701intrinsic("load_cull_small_triangles_enabled_amd", dest_comp=1, bit_sizes=[1], flags=[CAN_ELIMINATE])
1702# Whether the shader should cull small lines that are not visible in a pixel.
1703intrinsic("load_cull_small_lines_enabled_amd", dest_comp=1, bit_sizes=[1], flags=[CAN_ELIMINATE])
1704# Whether any culling setting is enabled in the shader.
1705intrinsic("load_cull_any_enabled_amd", dest_comp=1, bit_sizes=[1], flags=[CAN_ELIMINATE])
1706# Small triangle culling precision
1707intrinsic("load_cull_small_triangle_precision_amd", dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE, CAN_REORDER])
1708# Small line culling precision
1709intrinsic("load_cull_small_line_precision_amd", dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE, CAN_REORDER])
1710# Initial edge flags in a Vertex Shader, packed into the format the HW needs for primitive export.
1711intrinsic("load_initial_edgeflags_amd", src_comp=[], dest_comp=1, bit_sizes=[32], indices=[])
1712# Corresponds to s_sendmsg in the GCN/RDNA ISA, src[] = { m0_content }, BASE = imm
1713intrinsic("sendmsg_amd", src_comp=[1], indices=[BASE])
1714# Overwrites VS input registers, for use with vertex compaction after culling. src = {vertex_id, instance_id}.
1715intrinsic("overwrite_vs_arguments_amd", src_comp=[1, 1], indices=[])
1716# Overwrites TES input registers, for use with vertex compaction after culling. src = {tes_u, tes_v, rel_patch_id, patch_id}.
1717intrinsic("overwrite_tes_arguments_amd", src_comp=[1, 1, 1, 1], indices=[])
1718
1719# The address of the sbt descriptors.
1720system_value("sbt_base_amd", 1, bit_sizes=[64])
1721
1722# 1. HW descriptor
1723# 2. BVH node(64-bit pointer as 2x32 ...)
1724# 3. ray extent
1725# 4. ray origin
1726# 5. ray direction
1727# 6. inverse ray direction (componentwise 1.0/ray direction)
1728intrinsic("bvh64_intersect_ray_amd", [4, 2, 1, 3, 3, 3], 4, flags=[CAN_ELIMINATE, CAN_REORDER])
1729
1730# Return of a callable in raytracing pipelines
1731intrinsic("rt_return_amd")
1732
1733# offset into scratch for the input callable data in a raytracing pipeline.
1734system_value("rt_arg_scratch_offset_amd", 1)
1735
1736# Whether to call the anyhit shader for an intersection in an intersection shader.
1737system_value("intersection_opaque_amd", 1, bit_sizes=[1])
1738
1739# pointer to the next resume shader
1740system_value("resume_shader_address_amd", 1, bit_sizes=[64], indices=[CALL_IDX])
1741
1742# Ray Tracing Traversal inputs
1743system_value("sbt_offset_amd", 1)
1744system_value("sbt_stride_amd", 1)
1745system_value("accel_struct_amd", 1, bit_sizes=[64])
1746system_value("cull_mask_and_flags_amd", 1)
1747
1748#   0. SBT Index
1749#   1. Ray Tmax
1750#   2. Primitive Id
1751#   3. Instance Addr
1752#   4. Geometry Id and Flags
1753#   5. Hit Kind
1754intrinsic("execute_closest_hit_amd", src_comp=[1, 1, 1, 1, 1, 1])
1755
1756#   0. Ray Tmax
1757intrinsic("execute_miss_amd", src_comp=[1])
1758
1759# Used for saving and restoring hit attribute variables.
1760# BASE=dword index
1761intrinsic("load_hit_attrib_amd", dest_comp=1, bit_sizes=[32], indices=[BASE])
1762intrinsic("store_hit_attrib_amd", src_comp=[1], indices=[BASE])
1763
1764# Load forced VRS rates.
1765intrinsic("load_force_vrs_rates_amd", dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE, CAN_REORDER])
1766
1767intrinsic("load_scalar_arg_amd", dest_comp=0, bit_sizes=[32],
1768          indices=[BASE, ARG_UPPER_BOUND_U32_AMD],
1769          flags=[CAN_ELIMINATE, CAN_REORDER])
1770intrinsic("load_vector_arg_amd", dest_comp=0, bit_sizes=[32],
1771          indices=[BASE, ARG_UPPER_BOUND_U32_AMD, FLAGS],
1772          flags=[CAN_ELIMINATE, CAN_REORDER])
1773store("scalar_arg_amd", [], [BASE])
1774store("vector_arg_amd", [], [BASE])
1775
1776# src[] = { 32/64-bit base address, 32-bit offset }.
1777#
1778# Similar to load_global_constant, the memory accessed must be read-only. This
1779# restriction justifies the CAN_REORDER flag. Additionally, the base/offset must
1780# be subgroup uniform.
1781intrinsic("load_smem_amd", src_comp=[1, 1], dest_comp=0, bit_sizes=[32],
1782                           indices=[ALIGN_MUL, ALIGN_OFFSET],
1783                           flags=[CAN_ELIMINATE, CAN_REORDER])
1784
1785# src[] = { offset }.
1786intrinsic("load_shared2_amd", [1], dest_comp=2, indices=[OFFSET0, OFFSET1, ST64], flags=[CAN_ELIMINATE])
1787
1788# src[] = { value, offset }.
1789intrinsic("store_shared2_amd", [2, 1], indices=[OFFSET0, OFFSET1, ST64])
1790
1791# Vertex stride in LS-HS buffer
1792system_value("lshs_vertex_stride_amd", 1)
1793
1794# Vertex stride in ES-GS buffer
1795system_value("esgs_vertex_stride_amd", 1)
1796
1797# Per patch data offset in HS VRAM output buffer
1798system_value("hs_out_patch_data_offset_amd", 1)
1799
1800# line_width * 0.5 / abs(viewport_scale[2])
1801system_value("clip_half_line_width_amd", 2)
1802
1803# Number of vertices in a primitive
1804system_value("num_vertices_per_primitive_amd", 1)
1805
1806# Load streamout buffer desc
1807# BASE = buffer index
1808intrinsic("load_streamout_buffer_amd", dest_comp=4, indices=[BASE], bit_sizes=[32], flags=[CAN_ELIMINATE, CAN_REORDER])
1809
1810# An ID for each workgroup ordered by primitve sequence
1811system_value("ordered_id_amd", 1)
1812
1813# Add src1 to global streamout buffer offsets in the specified order.
1814# Only 1 lane must be active.
1815# src[] = { ordered_id, counter }
1816# WRITE_MASK = mask for counter channel to update
1817intrinsic("ordered_xfb_counter_add_gfx11_amd", dest_comp=0, src_comp=[1, 0], indices=[WRITE_MASK], bit_sizes=[32])
1818
1819# Execute the atomic ordered add loop. This does what ds_ordered_count did in previous generations.
1820# This is implemented with inline assembly to get the most optimal code.
1821#
1822# Inputs:
1823#   exec = one lane per counter (use nir_push_if, streamout should always enable 4 lanes)
1824#   src[0] = 64-bit SGPR atomic base address (streamout should use nir_load_xfb_state_address_gfx12_amd)
1825#   src[1] = 32-bit VGPR voffset (streamout should set local_invocation_index * 8)
1826#   src[2] = 32-bit SGPR ordered_id (use nir_load_ordered_id_amd for streamout, compute shaders
1827#            should generated it manually)
1828#   src[3] = 64-bit VGPR atomic src, use pack_64_2x32_split(ordered_id, value), streamout should do:
1829#            pack_64_2x32_split(ordered_id, "dwords written per workgroup" for each buffer)
1830#
1831# dst = 32-bit VGPR of the previous value of 32-bit value in memory, returned for all enabled lanes
1832
1833# Example - streamout: It's used to add dwords_written[] to global streamout offsets.
1834# * Exactly 4 lanes must be active, one for each buffer binding.
1835# * Disabled buffers must set dwords_written=0 for their lane, but the lane
1836#   must be enabled.
1837#
1838intrinsic("ordered_add_loop_gfx12_amd", dest_comp=1, src_comp=[1, 1, 1, 1], bit_sizes=[32])
1839
1840# Subtract from global streamout buffer offsets. Used to fix up the offsets
1841# when we overflow streamout buffers.
1842# src[] = { offsets }
1843# WRITE_MASK = mask of offsets to subtract
1844intrinsic("xfb_counter_sub_gfx11_amd", src_comp=[0], indices=[WRITE_MASK], bit_sizes=[32])
1845
1846# Provoking vertex index in a primitive
1847system_value("provoking_vtx_in_prim_amd", 1)
1848
1849# Atomically add current wave's primitive count to query result
1850#   * GS emitted primitive is primitive emitted by any GS stream
1851#   * generated primitive is primitive that has been produced for that stream by VS/TES/GS
1852#   * streamout primitve is primitve that has been written to xfb buffer, may be different
1853#     than generated primitive when xfb buffer is too small to hold more primitives
1854# src[] = { primitive_count }.
1855intrinsic("atomic_add_gs_emit_prim_count_amd", [1])
1856intrinsic("atomic_add_gen_prim_count_amd", [1], indices=[STREAM_ID])
1857intrinsic("atomic_add_xfb_prim_count_amd", [1], indices=[STREAM_ID])
1858
1859# Atomically add current shader's invocation count to query result
1860# src[] = { invocation_count }.
1861intrinsic("atomic_add_shader_invocation_count_amd", [1])
1862
1863# LDS offset for scratch section in NGG shader
1864system_value("lds_ngg_scratch_base_amd", 1)
1865# LDS offset for NGG GS shader vertex emit
1866system_value("lds_ngg_gs_out_vertex_base_amd", 1)
1867
1868# AMD GPU shader output export instruction
1869# src[] = { export_value, row }
1870# BASE = export target
1871# FLAGS = AC_EXP_FLAG_*
1872intrinsic("export_amd", [0], indices=[BASE, WRITE_MASK, FLAGS])
1873intrinsic("export_row_amd", [0, 1], indices=[BASE, WRITE_MASK, FLAGS])
1874
1875# Export dual source blend outputs with swizzle operation
1876# src[] = { mrt0, mrt1 }
1877intrinsic("export_dual_src_blend_amd", [0, 0], indices=[WRITE_MASK])
1878
1879# Alpha test reference value
1880system_value("alpha_reference_amd", 1)
1881
1882# Whether to enable barycentric optimization
1883system_value("barycentric_optimize_amd", dest_comp=1, bit_sizes=[1])
1884
1885# Copy the input into a register which will remain valid for entire quads, even in control flow.
1886# This should only be used directly for texture sources.
1887intrinsic("strict_wqm_coord_amd", src_comp=[0], dest_comp=0, bit_sizes=[32], indices=[BASE],
1888          flags=[CAN_ELIMINATE])
1889
1890intrinsic("cmat_muladd_amd", src_comp=[16, 16, 0], dest_comp=0, bit_sizes=src2,
1891          indices=[SATURATE, CMAT_SIGNED_MASK], flags=[CAN_ELIMINATE])
1892
1893# Get the debug log buffer descriptor.
1894intrinsic("load_debug_log_desc_amd", bit_sizes=[32], dest_comp=4, flags=[CAN_ELIMINATE, CAN_REORDER])
1895
1896# s_sleep BASE (sleep for 64*BASE cycles). BASE must be in [0, 0xffff].
1897# BASE=0 is valid but isn't useful.
1898# GFX12+: If BASE & 0x8000, sleep forever (until wakeup, trap, or kill).
1899intrinsic("sleep_amd", indices=[BASE])
1900
1901# s_nop BASE (sleep for BASE+1 cycles, BASE must be in [0, 15]).
1902intrinsic("nop_amd", indices=[BASE])
1903
1904system_value("ray_tracing_stack_base_lvp", 1)
1905
1906system_value("shader_call_data_offset_lvp", 1)
1907
1908# Broadcom-specific instrinc for tile buffer color reads.
1909#
1910# The hardware requires that we read the samples and components of a pixel
1911# in order, so we cannot eliminate or remove any loads in a sequence.
1912#
1913# src[] = { render_target }
1914# BASE = sample index
1915load("tlb_color_brcm", [1], [BASE, COMPONENT], [])
1916
1917# V3D-specific instrinc for per-sample tile buffer color writes.
1918#
1919# The driver backend needs to identify per-sample color writes and emit
1920# specific code for them.
1921#
1922# src[] = { value, render_target }
1923# BASE = sample index
1924store("tlb_sample_color_v3d", [1], [BASE, COMPONENT, SRC_TYPE], [])
1925
1926# V3D-specific intrinsic to load the number of layers attached to
1927# the target framebuffer
1928intrinsic("load_fb_layers_v3d", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
1929
1930# V3D-specific intrinsic to load W coordinate from the fragment shader payload
1931intrinsic("load_fep_w_v3d", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
1932
1933# Active invocation index within the subgroup.
1934# Equivalent to popcount(ballot(true) & ((1 << subgroup_invocation) - 1))
1935intrinsic("load_active_subgroup_invocation_agx", dest_comp=1, flags=[CAN_ELIMINATE])
1936
1937# Total active invocations within the subgroup.
1938# Equivalent to popcount(ballot(true))
1939intrinsic("load_active_subgroup_count_agx", dest_comp=1, flags=[CAN_ELIMINATE])
1940
1941# Like ballot() but only within a quad.
1942intrinsic("quad_ballot_agx", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
1943
1944# With [0, 1] clipping, no transform is needed on the output z' = z. But with [-1,
1945# 1] clipping, we need to transform z' = (z + w) / 2. We express both cases as a
1946# lerp between z and w, where this is the lerp coefficient: 0 for [0, 1] and 0.5
1947# for [-1, 1].
1948system_value("clip_z_coeff_agx", 1)
1949
1950# True if drawing triangle fans with first vertex provoking, false otherwise.
1951# This affects flatshading, which is defined weirdly for fans with first.
1952system_value("is_first_fan_agx", 1, bit_sizes=[1])
1953
1954# mesa_prim for the input topology (in a geometry shader)
1955system_value("input_topology_agx", 1)
1956
1957# Load a bindless sampler handle mapping a binding table sampler.
1958intrinsic("load_sampler_handle_agx", [1], 1, [],
1959          flags=[CAN_ELIMINATE, CAN_REORDER],
1960          bit_sizes=[16])
1961
1962# Load a bindless texture handle mapping a binding table texture.
1963intrinsic("load_texture_handle_agx", [1], 2, [],
1964          flags=[CAN_ELIMINATE, CAN_REORDER],
1965          bit_sizes=[32])
1966
1967# Given a vec2 bindless texture handle, load the address of the texture
1968# descriptor described by that vec2. This allows inspecting the descriptor from
1969# the shader. This does not actually load the content of the descriptor, only
1970# the content of the handle (which is the address of the descriptor).
1971intrinsic("load_from_texture_handle_agx", [2], 1, [],
1972          flags=[CAN_ELIMINATE, CAN_REORDER],
1973          bit_sizes=[64])
1974
1975# Load the coefficient register corresponding to a given fragment shader input.
1976# Coefficient registers are vec3s that are dotted with <x, y, 1> to interpolate
1977# the input, where x and y are relative to the 32x32 supertile.
1978intrinsic("load_coefficients_agx", [1],
1979          bit_sizes = [32],
1980          dest_comp = 3,
1981          indices=[COMPONENT, IO_SEMANTICS, INTERP_MODE],
1982          flags=[CAN_ELIMINATE, CAN_REORDER])
1983
1984# src[] = { value, index }
1985# Store a vertex shader output to the Unified Vertex Store (UVS). Indexed by UVS
1986# index, which must be assigned by the driver based on the linked fragment
1987# shader's interpolation qualifiers. This corresponds to the native instruction.
1988store("uvs_agx", [1], [], [CAN_REORDER])
1989
1990# Driver intrinsic to map a location to a UVS index. This is generated when
1991# lowering store_output to store_uvs_agx, and must be lowered by the driver.
1992intrinsic("load_uvs_index_agx", dest_comp = 1, bit_sizes=[16],
1993          indices=[IO_SEMANTICS], flags=[CAN_ELIMINATE, CAN_REORDER])
1994
1995# Load/store a pixel in local memory. This operation is formatted, with
1996# conversion between the specified format and the implied register format of the
1997# source/destination (for store/loads respectively). This mostly matters for
1998# converting between floating-point registers and normalized memory formats.
1999#
2000# The format is the pipe_format of the local memory (the source), see
2001# ail for the supported list.
2002#
2003# Logically, this loads/stores a single sample. The sample to load is
2004# specified by the bitfield sample mask source. However, for stores multiple
2005# bits of the sample mask may be set, which will replicate the value. For
2006# pixel rate shading, use 0xFF as the mask to store to all samples regardless of
2007# the sample count.
2008#
2009# All calculations are relative to an immediate byte offset into local
2010# memory, which acts relative to the start of the sample. These instructions
2011# logically access:
2012#
2013#   (((((y * tile_width) + x) * nr_samples) + sample) * sample_stride) + offset
2014#
2015# src[] = { sample mask }
2016# base = offset
2017load("local_pixel_agx", [1], [BASE, FORMAT], [CAN_REORDER, CAN_ELIMINATE])
2018# src[] = { value, sample mask, coordinates }
2019# base = offset
2020store("local_pixel_agx", [1, -1], [BASE, WRITE_MASK, FORMAT, EXPLICIT_COORD], [CAN_REORDER])
2021
2022# Combined depth/stencil emit, applying to a mask of samples. base indicates
2023# which to write (1 = depth, 2 = stencil, 3 = both).
2024#
2025# src[] = { sample mask, depth, stencil }
2026intrinsic("store_zs_agx", [1, 1, 1], indices=[BASE], flags=[])
2027
2028# Store a block from local memory into a bound image. Used to write out render
2029# targets within the end-of-tile shader, although it is valid in general compute
2030# kernels.
2031#
2032# The format is the pipe_format of the local memory (the source), see
2033# ail for the supported list. The image format is
2034# specified in the PBE descriptor.
2035#
2036# The image dimension is used to distinguish multisampled images from
2037# non-multisampled images. It must be 2D or MS.
2038#
2039# extra src[] = { logical offset within shared memory, coordinates/layer }
2040image("store_block_agx", [1, -1], extra_indices=[EXPLICIT_COORD])
2041
2042# Formatted load/store. The format is the pipe_format in memory (see ail for the
2043# supported list). This accesses:
2044#
2045#     address + extend(index) << (format shift + shift)
2046#
2047# The nir_intrinsic_base() index encodes the shift. The sign_extend index
2048# determines whether sign- or zero-extension is used for the index.
2049#
2050# All loads and stores on AGX uses these hardware instructions, so while these are
2051# logically load_global_agx/load_global_constant_agx/store_global_agx, the
2052# _global is omitted as it adds nothing.
2053#
2054# src[] = { address, index }.
2055load("agx", [1, 1], [ACCESS, BASE, FORMAT, SIGN_EXTEND], [CAN_ELIMINATE])
2056load("constant_agx", [1, 1], [ACCESS, BASE, FORMAT, SIGN_EXTEND],
2057     [CAN_ELIMINATE, CAN_REORDER])
2058# src[] = { value, address, index }.
2059store("agx", [1, 1], [ACCESS, BASE, FORMAT, SIGN_EXTEND])
2060
2061# Logical complement of load_front_face, mapping to an AGX system value
2062system_value("back_face_agx", 1, bit_sizes=[1, 32])
2063
2064# Load the base address of an indexed vertex attribute (for lowering).
2065intrinsic("load_vbo_base_agx", src_comp=[1], dest_comp=1, bit_sizes=[64],
2066          flags=[CAN_ELIMINATE, CAN_REORDER])
2067
2068# When vertex robustness is enabled, loads the maximum valid attribute index for
2069# a given attribute. This is unsigned: the driver ensures that at least one
2070# vertex is always valid to load, directing loads to a zero sink if necessary.
2071intrinsic("load_attrib_clamp_agx", src_comp=[1], dest_comp=1,
2072          bit_sizes=[32], flags=[CAN_ELIMINATE, CAN_REORDER])
2073
2074# Load a driver-internal system value from a given system value set at a given
2075# binding within the set. This is used for correctness when lowering things like
2076# UBOs with merged shaders.
2077#
2078# The FLAGS are used internally for loading the index of the uniform itself,
2079# rather than the contents, used for lowering bindless handles (which encode
2080# uniform indices as immediates in the NIR for technical reasons).
2081load("sysval_agx", [], [DESC_SET, BINDING, FLAGS], [CAN_REORDER, CAN_ELIMINATE])
2082
2083# Write out a sample mask for a targeted subset of samples, specified in the two
2084# masks. Maps to the corresponding AGX instruction, the actual workings are
2085# documented elsewhere as they are too complicated for this comment.
2086intrinsic("sample_mask_agx", src_comp=[1, 1])
2087
2088# Discard a subset of samples given by a specified sample mask. This acts like a
2089# per-sample discard, or an inverted accumulating gl_SampleMask write. The
2090# compiler will lower to sample_mask_agx, but that lowering is nontrivial as
2091# sample_mask_agx also triggers depth/stencil testing.
2092intrinsic("discard_agx", src_comp=[1])
2093
2094# For a given row of the polygon stipple given as an integer source in [0, 31],
2095# load the 32-bit stipple pattern for that row.
2096intrinsic("load_polygon_stipple_agx", src_comp=[1], dest_comp=1, bit_sizes=[32],
2097          flags=[CAN_ELIMINATE, CAN_ELIMINATE])
2098
2099# The fixed-function sample mask specified in the API (e.g. glSampleMask)
2100system_value("api_sample_mask_agx", 1, bit_sizes=[16])
2101
2102# Bit mask of samples currently being shaded. For API-level sample shading, this
2103# will usually equal (1 << sample_id). Multiple bits can be set when sample
2104# shading is only enabled due to framebuffer fetch, and the framebuffer has
2105# multiple samples with the same value.
2106#
2107# Used as a loop variable with dynamic sample shading.
2108system_value("active_samples_agx", 1, bit_sizes=[16])
2109
2110# Loads the sample position array as fixed point packed into a 32-bit word
2111system_value("sample_positions_agx", 1, bit_sizes=[32])
2112
2113# In a non-monolithic fragment shader part, returns whether this shader part is
2114# responsible for Z/S testing after its final discard. ~0/0 boolean.
2115system_value("shader_part_tests_zs_agx", 1, bit_sizes=[16])
2116
2117# Returns whether the API depth test is NEVER. We emulate this in shader when
2118# fragment side effects are used to ensure the fragment shader executes.
2119system_value("depth_never_agx", 1, bit_sizes=[16])
2120
2121# In a fragment shader, returns the log2 of the number of samples in the
2122# tilebuffer. This is the unprocessed value written in the corresponding USC
2123# word. Used to determine whether sample mask writes have any effect when sample
2124# count is dynamic.
2125system_value("samples_log2_agx", 1, bit_sizes=[16])
2126
2127# Loads the fixed-function glPointSize() value, or zero if the
2128# shader-supplied value should be used.
2129system_value("fixed_point_size_agx", 1, bit_sizes=[32])
2130
2131# Bit mask of TEX locations that are replaced with point sprites
2132system_value("tex_sprite_mask_agx", 1, bit_sizes=[16])
2133
2134# Image loads go through the texture cache, which is not coherent with the PBE
2135# or memory access, so fencing is necessary for writes to become visible.
2136
2137# Make writes via main memory (image atomics) visible for texturing.
2138barrier("fence_pbe_to_tex_agx")
2139
2140# Make writes from global memory instructions (atomics) visible for texturing.
2141barrier("fence_mem_to_tex_agx")
2142
2143# Variant of fence_pbe_to_tex_agx specialized to stores in pixel shaders that
2144# act like render target writes, in conjunction with fragment interlock.
2145barrier("fence_pbe_to_tex_pixel_agx")
2146
2147# Unknown fence used in the helper program on exit.
2148barrier("fence_helper_exit_agx")
2149
2150# Pointer to the buffer passing outputs VS->TCS, VS->GS, or TES->GS linkage.
2151system_value("vs_output_buffer_agx", 1, bit_sizes=[64])
2152
2153# Mask of VS->TCS, VS->GS, or TES->GS outputs. This is modelled as a sysval
2154# directly so it can be dynamic with shader objects or constant folded with
2155# pipelines (including GPL)
2156system_value("vs_outputs_agx", 1, bit_sizes=[64])
2157
2158# Address of state for AGX input assembly lowering for geometry/tessellation
2159system_value("input_assembly_buffer_agx", 1, bit_sizes=[64])
2160
2161# Address of the parameter buffer for AGX geometry shaders
2162system_value("geometry_param_buffer_agx", 1, bit_sizes=[64])
2163
2164# Address of the parameter buffer for AGX tessellation shaders
2165system_value("tess_param_buffer_agx", 1, bit_sizes=[64])
2166
2167# Address of the pipeline statistic query result indexed by BASE
2168system_value("stat_query_address_agx", 1, bit_sizes=[64], indices=[BASE])
2169
2170# Helper shader intrinsics
2171# src[] = { value }.
2172intrinsic("doorbell_agx", src_comp=[1])
2173
2174# src[] = { index, stack_address }.
2175intrinsic("stack_map_agx", src_comp=[1, 1])
2176
2177# src[] = { index }.
2178# dst[] = { stack_address }.
2179intrinsic("stack_unmap_agx", src_comp=[1], dest_comp=1, bit_sizes=[32])
2180
2181# dst[] = { GPU core ID }.
2182system_value("core_id_agx", 1, bit_sizes=[32])
2183
2184# dst[] = { Helper operation type }.
2185load("helper_op_id_agx", [], [], [CAN_ELIMINATE])
2186
2187# dst[] = { Helper argument low 32 bits }.
2188load("helper_arg_lo_agx", [], [], [CAN_ELIMINATE])
2189
2190# dst[] = { Helper argument high 32 bits }.
2191load("helper_arg_hi_agx", [], [], [CAN_ELIMINATE])
2192
2193# Export a vector. At the end of the shader part, the source is copied to the
2194# indexed GPRs starting at BASE. Exports must not overlap within a shader part.
2195# Must only appear in the last block of the shader part.
2196intrinsic("export_agx", [0], indices=[BASE])
2197
2198# Load an exported vector at the beginning of the shader part from GPRs starting
2199# at BASE. Must only appear in the first block of the shader part.
2200load("exported_agx", [], [BASE], [CAN_ELIMINATE])
2201
2202# Intel-specific query for loading from the isl_image_param struct passed
2203# into the shader as a uniform.  The variable is a deref to the image
2204# variable. The const index specifies which of the six parameters to load.
2205intrinsic("image_deref_load_param_intel", src_comp=[1], dest_comp=0,
2206          indices=[BASE], flags=[CAN_ELIMINATE, CAN_REORDER])
2207image("load_raw_intel", src_comp=[1], dest_comp=0,
2208      flags=[CAN_ELIMINATE])
2209image("store_raw_intel", src_comp=[1, 0])
2210
2211# Number of data items being operated on for a SIMD program.
2212system_value("simd_width_intel", 1)
2213
2214# Load a relocatable 32-bit value
2215intrinsic("load_reloc_const_intel", dest_comp=1, bit_sizes=[32],
2216          indices=[PARAM_IDX, BASE], flags=[CAN_ELIMINATE, CAN_REORDER])
2217
2218# 1 component 32bit surface index that can be used for bindless or BTI heaps
2219#
2220# This intrinsic is used to figure out what UBOs accesses could be promoted to
2221# push constants. To allow promoting a load_ubo to push constants, we need to
2222# know that the surface & offset are constants. If we want to use the bindless
2223# heap for this we have to build the surface index with a pushed constant for
2224# the descriptor set which prevents us from doing a nir_src_is_const() check.
2225# With this intrinsic, we can just check the surface_index src with
2226# nir_src_is_const() and ignore set_offset.
2227#
2228# src[] = { set_offset, surface_index, array_index, bindless_base_offset }
2229intrinsic("resource_intel", dest_comp=1, bit_sizes=[32],
2230          src_comp=[1, 1, 1, 1],
2231          indices=[DESC_SET, BINDING, RESOURCE_ACCESS_INTEL, RESOURCE_BLOCK_INTEL],
2232          flags=[CAN_ELIMINATE, CAN_REORDER])
2233
2234# OpSubgroupBlockReadINTEL and OpSubgroupBlockWriteINTEL from SPV_INTEL_subgroups.
2235intrinsic("load_deref_block_intel", dest_comp=0, src_comp=[-1],
2236          indices=[ACCESS], flags=[CAN_ELIMINATE])
2237intrinsic("store_deref_block_intel", src_comp=[-1, 0], indices=[WRITE_MASK, ACCESS])
2238
2239# src[] = { address }.
2240load("global_block_intel", [1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
2241
2242# src[] = { buffer_index, offset }.
2243load("ssbo_block_intel", [-1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
2244
2245# src[] = { offset }.
2246load("shared_block_intel", [1], [BASE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
2247
2248# src[] = { value, address }.
2249store("global_block_intel", [1], [WRITE_MASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
2250
2251# src[] = { value, block_index, offset }
2252store("ssbo_block_intel", [-1, 1], [WRITE_MASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
2253
2254# src[] = { value, offset }.
2255store("shared_block_intel", [1], [BASE, WRITE_MASK, ALIGN_MUL, ALIGN_OFFSET])
2256
2257# src[] = { address }.
2258load("global_constant_uniform_block_intel", [1],
2259     [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE, CAN_REORDER])
2260
2261# Similar to load_global_const_block_intel but for UBOs
2262# offset should be uniform
2263# src[] = { buffer_index, offset }.
2264load("ubo_uniform_block_intel", [-1, 1],
2265     [ACCESS, ALIGN_MUL, ALIGN_OFFSET, RANGE_BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
2266
2267# Similar to load_global_const_block_intel but for SSBOs
2268# offset should be uniform
2269# src[] = { buffer_index, offset }.
2270load("ssbo_uniform_block_intel", [-1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
2271
2272# Similar to load_global_const_block_intel but for shared memory
2273# src[] = { offset }.
2274load("shared_uniform_block_intel", [1], [BASE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
2275
2276# Inline register delivery (available on Gfx12.5+ for CS/Mesh/Task stages)
2277intrinsic("load_inline_data_intel", [], dest_comp=0,
2278          indices=[BASE],
2279          flags=[CAN_ELIMINATE, CAN_REORDER])
2280
2281# Intrinsics for Intel bindless thread dispatch
2282# BASE=brw_topoloy_id
2283system_value("topology_id_intel", 1, indices=[BASE])
2284system_value("btd_stack_id_intel", 1)
2285system_value("btd_global_arg_addr_intel", 1, bit_sizes=[64])
2286system_value("btd_local_arg_addr_intel", 1, bit_sizes=[64])
2287system_value("btd_resume_sbt_addr_intel", 1, bit_sizes=[64])
2288# src[] = { global_arg_addr, btd_record }
2289intrinsic("btd_spawn_intel", src_comp=[1, 1])
2290# RANGE=stack_size
2291intrinsic("btd_stack_push_intel", indices=[STACK_SIZE])
2292# src[] = { }
2293intrinsic("btd_retire_intel")
2294
2295# Intel-specific ray-tracing intrinsic
2296# src[] = { globals, level, operation } SYNCHRONOUS=synchronous
2297intrinsic("trace_ray_intel", src_comp=[1, 1, 1], indices=[SYNCHRONOUS])
2298
2299# System values used for ray-tracing on Intel
2300system_value("ray_base_mem_addr_intel", 1, bit_sizes=[64])
2301system_value("ray_hw_stack_size_intel", 1)
2302system_value("ray_sw_stack_size_intel", 1)
2303system_value("ray_num_dss_rt_stacks_intel", 1)
2304system_value("ray_hit_sbt_addr_intel", 1, bit_sizes=[64])
2305system_value("ray_hit_sbt_stride_intel", 1, bit_sizes=[16])
2306system_value("ray_miss_sbt_addr_intel", 1, bit_sizes=[64])
2307system_value("ray_miss_sbt_stride_intel", 1, bit_sizes=[16])
2308system_value("callable_sbt_addr_intel", 1, bit_sizes=[64])
2309system_value("callable_sbt_stride_intel", 1, bit_sizes=[16])
2310system_value("leaf_opaque_intel", 1, bit_sizes=[1])
2311system_value("leaf_procedural_intel", 1, bit_sizes=[1])
2312# Values :
2313#  0: AnyHit
2314#  1: ClosestHit
2315#  2: Miss
2316#  3: Intersection
2317system_value("btd_shader_type_intel", 1)
2318system_value("ray_query_global_intel", 1, bit_sizes=[64])
2319
2320# Source 0: Accumulator matrix (type specified by DEST_TYPE)
2321# Source 1: A matrix (type specified by SRC_TYPE)
2322# Source 2: B matrix (type specified by SRC_TYPE)
2323#
2324# The matrix parameters are the slices owned by the invocation.
2325#
2326# The accumulator is source 0 because that is the source the intrinsic
2327# infrastructure in NIR uses to determine the number of components in the
2328# result.
2329#
2330# The number of components for the second source is -1 to avoid validation of
2331# its value. Some supported configurations will have the component count of
2332# that matrix different than the others.
2333intrinsic("dpas_intel", dest_comp=0, src_comp=[0, -1, 0],
2334          indices=[DEST_TYPE, SRC_TYPE, SATURATE, SYSTOLIC_DEPTH, REPEAT_COUNT],
2335          flags=[CAN_ELIMINATE])
2336
2337# NVIDIA-specific intrinsics
2338# src[] = { index, offset }.
2339intrinsic("ldc_nv", dest_comp=0, src_comp=[1, 1],
2340          indices=[ACCESS, ALIGN_MUL, ALIGN_OFFSET],
2341          flags=[CAN_ELIMINATE, CAN_REORDER])
2342# [Un]pins an LDCX handle around non-uniform control-flow sections
2343# src[] = { handle }.
2344intrinsic("pin_cx_handle_nv", src_comp=[1])
2345intrinsic("unpin_cx_handle_nv", src_comp=[1])
2346# src[] = { handle, offset }.
2347intrinsic("ldcx_nv", dest_comp=0, src_comp=[1, 1],
2348          indices=[ACCESS, ALIGN_MUL, ALIGN_OFFSET],
2349          flags=[CAN_ELIMINATE, CAN_REORDER])
2350intrinsic("load_sysval_nv", dest_comp=1, src_comp=[], bit_sizes=[32, 64],
2351          indices=[ACCESS, BASE], flags=[CAN_ELIMINATE])
2352intrinsic("isberd_nv", dest_comp=1, src_comp=[1], bit_sizes=[32],
2353          flags=[CAN_ELIMINATE, CAN_REORDER])
2354intrinsic("al2p_nv", dest_comp=1, src_comp=[1], bit_sizes=[32],
2355          indices=[BASE, FLAGS], flags=[CAN_ELIMINATE, CAN_REORDER])
2356# src[] = { vtx, offset }.
2357# FLAGS is struct nak_nir_attr_io_flags
2358intrinsic("ald_nv", dest_comp=0, src_comp=[1, 1], bit_sizes=[32],
2359          indices=[BASE, RANGE_BASE, RANGE, FLAGS, ACCESS],
2360          flags=[CAN_ELIMINATE])
2361# src[] = { data, vtx, offset }.
2362# FLAGS is struct nak_nir_attr_io_flags
2363intrinsic("ast_nv", src_comp=[0, 1, 1],
2364          indices=[BASE, RANGE_BASE, RANGE, FLAGS], flags=[])
2365# src[] = { inv_w, offset }.
2366intrinsic("ipa_nv", dest_comp=1, src_comp=[1, 1], bit_sizes=[32],
2367          indices=[BASE, FLAGS], flags=[CAN_ELIMINATE, CAN_REORDER])
2368# FLAGS indicate if we load vertex_id == 2
2369intrinsic("ldtram_nv", dest_comp=2, bit_sizes=[32],
2370          indices=[BASE, FLAGS], flags=[CAN_ELIMINATE, CAN_REORDER])
2371
2372# NVIDIA-specific Geometry Shader intrinsics.
2373# These contain an additional integer source and destination with the primitive handle input/output.
2374intrinsic("emit_vertex_nv", dest_comp=1, src_comp=[1], indices=[STREAM_ID])
2375intrinsic("end_primitive_nv", dest_comp=1, src_comp=[1], indices=[STREAM_ID])
2376# Contains the final primitive handle and indicate the end of emission.
2377intrinsic("final_primitive_nv", src_comp=[1])
2378
2379# src[] = { data }.
2380intrinsic("fs_out_nv", src_comp=[1], indices=[BASE], flags=[])
2381barrier("copy_fs_outputs_nv")
2382
2383intrinsic("bar_set_nv", dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE])
2384intrinsic("bar_break_nv", dest_comp=1, bit_sizes=[32], src_comp=[1, 1])
2385# src[] = { bar, bar_set }
2386intrinsic("bar_sync_nv", src_comp=[1, 1])
2387
2388# Stall until the given SSA value is available
2389intrinsic("ssa_bar_nv", src_comp=[1])
2390
2391# NVIDIA-specific system values
2392system_value("warps_per_sm_nv", 1, bit_sizes=[32])
2393system_value("sm_count_nv", 1, bit_sizes=[32])
2394system_value("warp_id_nv", 1, bit_sizes=[32])
2395system_value("sm_id_nv", 1, bit_sizes=[32])
2396
2397# In order to deal with flipped render targets, gl_PointCoord may be flipped
2398# in the shader requiring a shader key or extra instructions or it may be
2399# flipped in hardware based on a state bit.  This version of gl_PointCoord
2400# is defined to be whatever thing the hardware can easily give you, so long as
2401# it's in normalized coordinates in the range [0, 1] across the point.
2402#
2403# src0 contains barycentrics for interpolation.
2404intrinsic("load_point_coord_maybe_flipped", dest_comp=2, bit_sizes=[32], src_comp=[2])
2405
2406
2407# Load texture size values:
2408#
2409# Takes a sampler # and returns width, height and depth.  If texture is a array
2410# texture it returns width, height and array size.  Used for txs lowering.
2411intrinsic("load_texture_size_etna", src_comp=[1], dest_comp=3,
2412          flags=[CAN_ELIMINATE, CAN_REORDER])
2413
2414# Zink specific intrinsics
2415
2416# src[] = { field }.
2417load("push_constant_zink", [1], [COMPONENT], [CAN_ELIMINATE, CAN_REORDER])
2418
2419system_value("shader_index", 1, bit_sizes=[32])
2420
2421system_value("coalesced_input_count", 1, bit_sizes=[32])
2422
2423# Initialize a payload array per scope
2424#
2425#   0. Payloads deref
2426#   1. Payload count
2427#   2. Node index
2428intrinsic("initialize_node_payloads", src_comp=[-1, 1, 1], indices=[EXECUTION_SCOPE])
2429
2430# Optionally enqueue payloads after shader finished writing to them
2431intrinsic("enqueue_node_payloads", src_comp=[-1])
2432
2433# Returns true if it has been called for every payload.
2434intrinsic("finalize_incoming_node_payload", src_comp=[-1], dest_comp=1)
2435