• 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
185index("enum pipe_format", "format")
186
187# Access qualifiers for image and memory access intrinsics. ACCESS_RESTRICT is
188# not set at the intrinsic if the NIR was created from SPIR-V.
189index("enum gl_access_qualifier", "access")
190
191# call index for split raytracing shaders
192index("unsigned", "call_idx")
193
194# The stack size increment/decrement for split raytracing shaders
195index("unsigned", "stack_size")
196
197# Alignment for offsets and addresses
198#
199# These two parameters, specify an alignment in terms of a multiplier and
200# an offset.  The multiplier is always a power of two.  The offset or
201# address parameter X of the intrinsic is guaranteed to satisfy the
202# following:
203#
204#                (X - align_offset) % align_mul == 0
205#
206# For constant offset values, align_mul will be NIR_ALIGN_MUL_MAX and the
207# align_offset will be modulo that.
208index("unsigned", "align_mul")
209index("unsigned", "align_offset")
210
211# The Vulkan descriptor type for a vulkan_resource_[re]index intrinsic.
212index("unsigned", "desc_type")
213
214# The nir_alu_type of input data to a store or conversion
215index("nir_alu_type", "src_type")
216
217# The nir_alu_type of the data output from a load or conversion
218index("nir_alu_type", "dest_type")
219
220# The swizzle mask for quad_swizzle_amd & masked_swizzle_amd
221index("unsigned", "swizzle_mask")
222
223# Whether the load_buffer_amd/store_buffer_amd is swizzled
224index("bool", "is_swizzled")
225
226# The SLC ("system level coherent") bit of load_buffer_amd/store_buffer_amd
227index("bool", "slc_amd")
228
229# Separate source/dest access flags for copies
230index("enum gl_access_qualifier", "dst_access")
231index("enum gl_access_qualifier", "src_access")
232
233# Driver location of attribute
234index("unsigned", "driver_location")
235
236# Ordering and visibility of a memory operation
237index("nir_memory_semantics", "memory_semantics")
238
239# Modes affected by a memory operation
240index("nir_variable_mode", "memory_modes")
241
242# Scope of a memory operation
243index("nir_scope", "memory_scope")
244
245# Scope of a control barrier
246index("nir_scope", "execution_scope")
247
248# Semantics of an IO instruction
249index("struct nir_io_semantics", "io_semantics")
250
251# Rounding mode for conversions
252index("nir_rounding_mode", "rounding_mode")
253
254# Whether or not to saturate in conversions
255index("unsigned", "saturate")
256
257intrinsic("nop", flags=[CAN_ELIMINATE])
258
259intrinsic("convert_alu_types", dest_comp=0, src_comp=[0],
260          indices=[SRC_TYPE, DEST_TYPE, ROUNDING_MODE, SATURATE],
261          flags=[CAN_ELIMINATE, CAN_REORDER])
262
263intrinsic("load_param", dest_comp=0, indices=[PARAM_IDX], flags=[CAN_ELIMINATE])
264
265intrinsic("load_deref", dest_comp=0, src_comp=[-1],
266          indices=[ACCESS], flags=[CAN_ELIMINATE])
267intrinsic("store_deref", src_comp=[-1, 0], indices=[WRITE_MASK, ACCESS])
268intrinsic("copy_deref", src_comp=[-1, -1], indices=[DST_ACCESS, SRC_ACCESS])
269intrinsic("memcpy_deref", src_comp=[-1, -1, 1], indices=[DST_ACCESS, SRC_ACCESS])
270
271# Interpolation of input.  The interp_deref_at* intrinsics are similar to the
272# load_var intrinsic acting on a shader input except that they interpolate the
273# input differently.  The at_sample, at_offset and at_vertex intrinsics take an
274# additional source that is an integer sample id, a vec2 position offset, or a
275# vertex ID respectively.
276
277intrinsic("interp_deref_at_centroid", dest_comp=0, src_comp=[1],
278          flags=[ CAN_ELIMINATE, CAN_REORDER])
279intrinsic("interp_deref_at_sample", src_comp=[1, 1], dest_comp=0,
280          flags=[CAN_ELIMINATE, CAN_REORDER])
281intrinsic("interp_deref_at_offset", src_comp=[1, 2], dest_comp=0,
282          flags=[CAN_ELIMINATE, CAN_REORDER])
283intrinsic("interp_deref_at_vertex", src_comp=[1, 1], dest_comp=0,
284          flags=[CAN_ELIMINATE, CAN_REORDER])
285
286# Gets the length of an unsized array at the end of a buffer
287intrinsic("deref_buffer_array_length", src_comp=[-1], dest_comp=1,
288          indices=[ACCESS], flags=[CAN_ELIMINATE, CAN_REORDER])
289
290# Ask the driver for the size of a given SSBO. It takes the buffer index
291# as source.
292intrinsic("get_ssbo_size", src_comp=[-1], dest_comp=1, bit_sizes=[32],
293          indices=[ACCESS], flags=[CAN_ELIMINATE, CAN_REORDER])
294intrinsic("get_ubo_size", src_comp=[-1], dest_comp=1,
295          flags=[CAN_ELIMINATE, CAN_REORDER])
296
297# Intrinsics which provide a run-time mode-check.  Unlike the compile-time
298# mode checks, a pointer can only have exactly one mode at runtime.
299intrinsic("deref_mode_is", src_comp=[-1], dest_comp=1,
300          indices=[MEMORY_MODES], flags=[CAN_ELIMINATE, CAN_REORDER])
301intrinsic("addr_mode_is", src_comp=[-1], dest_comp=1,
302          indices=[MEMORY_MODES], flags=[CAN_ELIMINATE, CAN_REORDER])
303
304intrinsic("is_sparse_texels_resident", dest_comp=1, src_comp=[1], bit_sizes=[1],
305          flags=[CAN_ELIMINATE, CAN_REORDER])
306# result code is resident only if both inputs are resident
307intrinsic("sparse_residency_code_and", dest_comp=1, src_comp=[1, 1], bit_sizes=[32],
308          flags=[CAN_ELIMINATE, CAN_REORDER])
309
310# a barrier is an intrinsic with no inputs/outputs but which can't be moved
311# around/optimized in general
312def barrier(name):
313    intrinsic(name)
314
315barrier("discard")
316
317# Demote fragment shader invocation to a helper invocation.  Any stores to
318# memory after this instruction are suppressed and the fragment does not write
319# outputs to the framebuffer.  Unlike discard, demote needs to ensure that
320# derivatives will still work for invocations that were not demoted.
321#
322# As specified by SPV_EXT_demote_to_helper_invocation.
323barrier("demote")
324intrinsic("is_helper_invocation", dest_comp=1, flags=[CAN_ELIMINATE])
325
326# SpvOpTerminateInvocation from SPIR-V.  Essentially a discard "for real".
327barrier("terminate")
328
329# A workgroup-level control barrier.  Any thread which hits this barrier will
330# pause until all threads within the current workgroup have also hit the
331# barrier.  For compute shaders, the workgroup is defined as the local group.
332# For tessellation control shaders, the workgroup is defined as the current
333# patch.  This intrinsic does not imply any sort of memory barrier.
334barrier("control_barrier")
335
336# Memory barrier with semantics analogous to the memoryBarrier() GLSL
337# intrinsic.
338barrier("memory_barrier")
339
340# Control/Memory barrier with explicit scope.  Follows the semantics of SPIR-V
341# OpMemoryBarrier and OpControlBarrier, used to implement Vulkan Memory Model.
342# Storage that the barrier applies is represented using NIR variable modes.
343# For an OpMemoryBarrier, set EXECUTION_SCOPE to NIR_SCOPE_NONE.
344intrinsic("scoped_barrier",
345          indices=[EXECUTION_SCOPE, MEMORY_SCOPE, MEMORY_SEMANTICS, MEMORY_MODES])
346
347# Shader clock intrinsic with semantics analogous to the clock2x32ARB()
348# GLSL intrinsic.
349# The latter can be used as code motion barrier, which is currently not
350# feasible with NIR.
351intrinsic("shader_clock", dest_comp=2, bit_sizes=[32], flags=[CAN_ELIMINATE],
352          indices=[MEMORY_SCOPE])
353
354# Shader ballot intrinsics with semantics analogous to the
355#
356#    ballotARB()
357#    readInvocationARB()
358#    readFirstInvocationARB()
359#
360# GLSL functions from ARB_shader_ballot.
361intrinsic("ballot", src_comp=[1], dest_comp=0, flags=[CAN_ELIMINATE])
362intrinsic("read_invocation", src_comp=[0, 1], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
363intrinsic("read_first_invocation", src_comp=[0], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
364
365# Returns the value of the first source for the lane where the second source is
366# true. The second source must be true for exactly one lane.
367intrinsic("read_invocation_cond_ir3", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
368
369# Additional SPIR-V ballot intrinsics
370#
371# These correspond to the SPIR-V opcodes
372#
373#    OpGroupNonUniformElect
374#    OpSubgroupFirstInvocationKHR
375intrinsic("elect", dest_comp=1, flags=[CAN_ELIMINATE])
376intrinsic("first_invocation", dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE])
377intrinsic("last_invocation", dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE])
378
379# Memory barrier with semantics analogous to the compute shader
380# groupMemoryBarrier(), memoryBarrierAtomicCounter(), memoryBarrierBuffer(),
381# memoryBarrierImage() and memoryBarrierShared() GLSL intrinsics.
382barrier("group_memory_barrier")
383barrier("memory_barrier_atomic_counter")
384barrier("memory_barrier_buffer")
385barrier("memory_barrier_image")
386barrier("memory_barrier_shared")
387barrier("begin_invocation_interlock")
388barrier("end_invocation_interlock")
389
390# Memory barrier for synchronizing TCS patch outputs
391barrier("memory_barrier_tcs_patch")
392
393# A conditional discard/demote/terminate, with a single boolean source.
394intrinsic("discard_if", src_comp=[1])
395intrinsic("demote_if", src_comp=[1])
396intrinsic("terminate_if", src_comp=[1])
397
398# ARB_shader_group_vote intrinsics
399intrinsic("vote_any", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
400intrinsic("vote_all", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
401intrinsic("vote_feq", src_comp=[0], dest_comp=1, flags=[CAN_ELIMINATE])
402intrinsic("vote_ieq", src_comp=[0], dest_comp=1, flags=[CAN_ELIMINATE])
403
404# Ballot ALU operations from SPIR-V.
405#
406# These operations work like their ALU counterparts except that the operate
407# on a uvec4 which is treated as a 128bit integer.  Also, they are, in
408# general, free to ignore any bits which are above the subgroup size.
409intrinsic("ballot_bitfield_extract", src_comp=[4, 1], dest_comp=1, flags=[CAN_ELIMINATE])
410intrinsic("ballot_bit_count_reduce", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
411intrinsic("ballot_bit_count_inclusive", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
412intrinsic("ballot_bit_count_exclusive", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
413intrinsic("ballot_find_lsb", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
414intrinsic("ballot_find_msb", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
415
416# Shuffle operations from SPIR-V.
417intrinsic("shuffle", src_comp=[0, 1], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
418intrinsic("shuffle_xor", src_comp=[0, 1], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
419intrinsic("shuffle_up", src_comp=[0, 1], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
420intrinsic("shuffle_down", src_comp=[0, 1], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
421
422# Quad operations from SPIR-V.
423intrinsic("quad_broadcast", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
424intrinsic("quad_swap_horizontal", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
425intrinsic("quad_swap_vertical", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
426intrinsic("quad_swap_diagonal", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
427
428intrinsic("reduce", src_comp=[0], dest_comp=0, bit_sizes=src0,
429          indices=[REDUCTION_OP, CLUSTER_SIZE], flags=[CAN_ELIMINATE])
430intrinsic("inclusive_scan", src_comp=[0], dest_comp=0, bit_sizes=src0,
431          indices=[REDUCTION_OP], flags=[CAN_ELIMINATE])
432intrinsic("exclusive_scan", src_comp=[0], dest_comp=0, bit_sizes=src0,
433          indices=[REDUCTION_OP], flags=[CAN_ELIMINATE])
434
435# AMD shader ballot operations
436intrinsic("quad_swizzle_amd", src_comp=[0], dest_comp=0, bit_sizes=src0,
437          indices=[SWIZZLE_MASK], flags=[CAN_ELIMINATE])
438intrinsic("masked_swizzle_amd", src_comp=[0], dest_comp=0, bit_sizes=src0,
439          indices=[SWIZZLE_MASK], flags=[CAN_ELIMINATE])
440intrinsic("write_invocation_amd", src_comp=[0, 0, 1], dest_comp=0, bit_sizes=src0,
441          flags=[CAN_ELIMINATE])
442# src = [ mask, addition ]
443intrinsic("mbcnt_amd", src_comp=[1, 1], dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE])
444# Compiled to v_perm_b32. src = [ in_bytes_hi, in_bytes_lo, selector ]
445intrinsic("byte_permute_amd", src_comp=[1, 1, 1], dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE, CAN_REORDER])
446# Compiled to v_permlane16_b32. src = [ value, lanesel_lo, lanesel_hi ]
447intrinsic("lane_permute_16_amd", src_comp=[1, 1, 1], dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE])
448
449# Basic Geometry Shader intrinsics.
450#
451# emit_vertex implements GLSL's EmitStreamVertex() built-in.  It takes a single
452# index, which is the stream ID to write to.
453#
454# end_primitive implements GLSL's EndPrimitive() built-in.
455intrinsic("emit_vertex",   indices=[STREAM_ID])
456intrinsic("end_primitive", indices=[STREAM_ID])
457
458# Geometry Shader intrinsics with a vertex count.
459#
460# Alternatively, drivers may implement these intrinsics, and use
461# nir_lower_gs_intrinsics() to convert from the basic intrinsics.
462#
463# These contain two additional unsigned integer sources:
464# 1. The total number of vertices emitted so far.
465# 2. The number of vertices emitted for the current primitive
466#    so far if we're counting, otherwise undef.
467intrinsic("emit_vertex_with_counter", src_comp=[1, 1], indices=[STREAM_ID])
468intrinsic("end_primitive_with_counter", src_comp=[1, 1], indices=[STREAM_ID])
469# Contains the final total vertex and primitive counts in the current GS thread.
470intrinsic("set_vertex_and_primitive_count", src_comp=[1, 1], indices=[STREAM_ID])
471
472# Trace a ray through an acceleration structure
473#
474# This instruction has a lot of parameters:
475#   0. Acceleration Structure
476#   1. Ray Flags
477#   2. Cull Mask
478#   3. SBT Offset
479#   4. SBT Stride
480#   5. Miss shader index
481#   6. Ray Origin
482#   7. Ray Tmin
483#   8. Ray Direction
484#   9. Ray Tmax
485#   10. Payload
486intrinsic("trace_ray", src_comp=[-1, 1, 1, 1, 1, 1, 3, 1, 3, 1, -1])
487# src[] = { hit_t, hit_kind }
488intrinsic("report_ray_intersection", src_comp=[1, 1], dest_comp=1)
489intrinsic("ignore_ray_intersection")
490intrinsic("accept_ray_intersection") # Not in SPIR-V; useful for lowering
491intrinsic("terminate_ray")
492# src[] = { sbt_index, payload }
493intrinsic("execute_callable", src_comp=[1, -1])
494
495# Driver independent raytracing helpers
496
497# rt_resume is a helper that that be the first instruction accesing the
498# stack/scratch in a resume shader for a raytracing pipeline. It includes the
499# resume index (for nir_lower_shader_calls_internal reasons) and the stack size
500# of the variables spilled during the call. The stack size can be use to e.g.
501# adjust a stack pointer.
502intrinsic("rt_resume", indices=[CALL_IDX, STACK_SIZE])
503
504# Lowered version of execute_callabe that includes the index of the resume
505# shader, and the amount of scratch space needed for this call (.ie. how much
506# to increase a stack pointer by).
507# src[] = { sbt_index, payload }
508intrinsic("rt_execute_callable", src_comp=[1, -1], indices=[CALL_IDX,STACK_SIZE])
509
510# Lowered version of trace_ray in a similar vein to rt_execute_callable.
511# src same as trace_ray
512intrinsic("rt_trace_ray", src_comp=[-1, 1, 1, 1, 1, 1, 3, 1, 3, 1, -1],
513          indices=[CALL_IDX, STACK_SIZE])
514
515
516# Atomic counters
517#
518# The *_var variants take an atomic_uint nir_variable, while the other,
519# lowered, variants take a constant buffer index and register offset.
520
521def atomic(name, flags=[]):
522    intrinsic(name + "_deref", src_comp=[-1], dest_comp=1, flags=flags)
523    intrinsic(name, src_comp=[1], dest_comp=1, indices=[BASE], flags=flags)
524
525def atomic2(name):
526    intrinsic(name + "_deref", src_comp=[-1, 1], dest_comp=1)
527    intrinsic(name, src_comp=[1, 1], dest_comp=1, indices=[BASE])
528
529def atomic3(name):
530    intrinsic(name + "_deref", src_comp=[-1, 1, 1], dest_comp=1)
531    intrinsic(name, src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
532
533atomic("atomic_counter_inc")
534atomic("atomic_counter_pre_dec")
535atomic("atomic_counter_post_dec")
536atomic("atomic_counter_read", flags=[CAN_ELIMINATE])
537atomic2("atomic_counter_add")
538atomic2("atomic_counter_min")
539atomic2("atomic_counter_max")
540atomic2("atomic_counter_and")
541atomic2("atomic_counter_or")
542atomic2("atomic_counter_xor")
543atomic2("atomic_counter_exchange")
544atomic3("atomic_counter_comp_swap")
545
546# Image load, store and atomic intrinsics.
547#
548# All image intrinsics come in three versions.  One which take an image target
549# passed as a deref chain as the first source, one which takes an index as the
550# first source, and one which takes a bindless handle as the first source.
551# In the first version, the image variable contains the memory and layout
552# qualifiers that influence the semantics of the intrinsic.  In the second and
553# third, the image format and access qualifiers are provided as constant
554# indices.
555#
556# All image intrinsics take a four-coordinate vector and a sample index as
557# 2nd and 3rd sources, determining the location within the image that will be
558# accessed by the intrinsic.  Components not applicable to the image target
559# in use are undefined.  Image store takes an additional four-component
560# argument with the value to be written, and image atomic operations take
561# either one or two additional scalar arguments with the same meaning as in
562# the ARB_shader_image_load_store specification.
563def image(name, src_comp=[], extra_indices=[], **kwargs):
564    intrinsic("image_deref_" + name, src_comp=[-1] + src_comp,
565              indices=[IMAGE_DIM, IMAGE_ARRAY, FORMAT, ACCESS] + extra_indices, **kwargs)
566    intrinsic("image_" + name, src_comp=[1] + src_comp,
567              indices=[IMAGE_DIM, IMAGE_ARRAY, FORMAT, ACCESS] + extra_indices, **kwargs)
568    intrinsic("bindless_image_" + name, src_comp=[1] + src_comp,
569              indices=[IMAGE_DIM, IMAGE_ARRAY, FORMAT, ACCESS] + extra_indices, **kwargs)
570
571image("load", src_comp=[4, 1, 1], extra_indices=[DEST_TYPE], dest_comp=0, flags=[CAN_ELIMINATE])
572image("sparse_load", src_comp=[4, 1, 1], extra_indices=[DEST_TYPE], dest_comp=0, flags=[CAN_ELIMINATE])
573image("store", src_comp=[4, 1, 0, 1], extra_indices=[SRC_TYPE])
574image("atomic_add",  src_comp=[4, 1, 1], dest_comp=1)
575image("atomic_imin",  src_comp=[4, 1, 1], dest_comp=1)
576image("atomic_umin",  src_comp=[4, 1, 1], dest_comp=1)
577image("atomic_imax",  src_comp=[4, 1, 1], dest_comp=1)
578image("atomic_umax",  src_comp=[4, 1, 1], dest_comp=1)
579image("atomic_and",  src_comp=[4, 1, 1], dest_comp=1)
580image("atomic_or",   src_comp=[4, 1, 1], dest_comp=1)
581image("atomic_xor",  src_comp=[4, 1, 1], dest_comp=1)
582image("atomic_exchange",  src_comp=[4, 1, 1], dest_comp=1)
583image("atomic_comp_swap", src_comp=[4, 1, 1, 1], dest_comp=1)
584image("atomic_fadd",  src_comp=[4, 1, 1], dest_comp=1)
585image("atomic_fmin",  src_comp=[4, 1, 1], dest_comp=1)
586image("atomic_fmax",  src_comp=[4, 1, 1], dest_comp=1)
587image("size",    dest_comp=0, src_comp=[1], flags=[CAN_ELIMINATE, CAN_REORDER])
588image("samples", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
589image("atomic_inc_wrap",  src_comp=[4, 1, 1], dest_comp=1)
590image("atomic_dec_wrap",  src_comp=[4, 1, 1], dest_comp=1)
591# CL-specific format queries
592image("format", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
593image("order", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
594
595# Vulkan descriptor set intrinsics
596#
597# The Vulkan API uses a different binding model from GL.  In the Vulkan
598# API, all external resources are represented by a tuple:
599#
600# (descriptor set, binding, array index)
601#
602# where the array index is the only thing allowed to be indirect.  The
603# vulkan_surface_index intrinsic takes the descriptor set and binding as
604# its first two indices and the array index as its source.  The third
605# index is a nir_variable_mode in case that's useful to the backend.
606#
607# The intended usage is that the shader will call vulkan_surface_index to
608# get an index and then pass that as the buffer index ubo/ssbo calls.
609#
610# The vulkan_resource_reindex intrinsic takes a resource index in src0
611# (the result of a vulkan_resource_index or vulkan_resource_reindex) which
612# corresponds to the tuple (set, binding, index) and computes an index
613# corresponding to tuple (set, binding, idx + src1).
614intrinsic("vulkan_resource_index", src_comp=[1], dest_comp=0,
615          indices=[DESC_SET, BINDING, DESC_TYPE],
616          flags=[CAN_ELIMINATE, CAN_REORDER])
617intrinsic("vulkan_resource_reindex", src_comp=[0, 1], dest_comp=0,
618          indices=[DESC_TYPE], flags=[CAN_ELIMINATE, CAN_REORDER])
619intrinsic("load_vulkan_descriptor", src_comp=[-1], dest_comp=0,
620          indices=[DESC_TYPE], flags=[CAN_ELIMINATE, CAN_REORDER])
621
622# atomic intrinsics
623#
624# All of these atomic memory operations read a value from memory, compute a new
625# value using one of the operations below, write the new value to memory, and
626# return the original value read.
627#
628# All variable operations take 2 sources except CompSwap that takes 3. These
629# sources represent:
630#
631# 0: A deref to the memory on which to perform the atomic
632# 1: The data parameter to the atomic function (i.e. the value to add
633#    in shared_atomic_add, etc).
634# 2: For CompSwap only: the second data parameter.
635#
636# All SSBO operations take 3 sources except CompSwap that takes 4. These
637# sources represent:
638#
639# 0: The SSBO buffer index.
640# 1: The offset into the SSBO buffer of the variable that the atomic
641#    operation will operate on.
642# 2: The data parameter to the atomic function (i.e. the value to add
643#    in ssbo_atomic_add, etc).
644# 3: For CompSwap only: the second data parameter.
645#
646# All shared variable operations take 2 sources except CompSwap that takes 3.
647# These sources represent:
648#
649# 0: The offset into the shared variable storage region that the atomic
650#    operation will operate on.
651# 1: The data parameter to the atomic function (i.e. the value to add
652#    in shared_atomic_add, etc).
653# 2: For CompSwap only: the second data parameter.
654#
655# All global operations take 2 sources except CompSwap that takes 3. These
656# sources represent:
657#
658# 0: The memory address that the atomic operation will operate on.
659# 1: The data parameter to the atomic function (i.e. the value to add
660#    in shared_atomic_add, etc).
661# 2: For CompSwap only: the second data parameter.
662
663def memory_atomic_data1(name):
664    intrinsic("deref_atomic_" + name,  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
665    intrinsic("ssbo_atomic_" + name,  src_comp=[-1, 1, 1], dest_comp=1, indices=[ACCESS])
666    intrinsic("shared_atomic_" + name,  src_comp=[1, 1], dest_comp=1, indices=[BASE])
667    intrinsic("global_atomic_" + name,  src_comp=[1, 1], dest_comp=1, indices=[BASE])
668
669def memory_atomic_data2(name):
670    intrinsic("deref_atomic_" + name,  src_comp=[-1, 1, 1], dest_comp=1, indices=[ACCESS])
671    intrinsic("ssbo_atomic_" + name,  src_comp=[-1, 1, 1, 1], dest_comp=1, indices=[ACCESS])
672    intrinsic("shared_atomic_" + name,  src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
673    intrinsic("global_atomic_" + name,  src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
674
675memory_atomic_data1("add")
676memory_atomic_data1("imin")
677memory_atomic_data1("umin")
678memory_atomic_data1("imax")
679memory_atomic_data1("umax")
680memory_atomic_data1("and")
681memory_atomic_data1("or")
682memory_atomic_data1("xor")
683memory_atomic_data1("exchange")
684memory_atomic_data1("fadd")
685memory_atomic_data1("fmin")
686memory_atomic_data1("fmax")
687memory_atomic_data2("comp_swap")
688memory_atomic_data2("fcomp_swap")
689
690def system_value(name, dest_comp, indices=[], bit_sizes=[32]):
691    intrinsic("load_" + name, [], dest_comp, indices,
692              flags=[CAN_ELIMINATE, CAN_REORDER], sysval=True,
693              bit_sizes=bit_sizes)
694
695system_value("frag_coord", 4)
696system_value("point_coord", 2)
697system_value("line_coord", 1)
698system_value("front_face", 1, bit_sizes=[1, 32])
699system_value("vertex_id", 1)
700system_value("vertex_id_zero_base", 1)
701system_value("first_vertex", 1)
702system_value("is_indexed_draw", 1)
703system_value("base_vertex", 1)
704system_value("instance_id", 1)
705system_value("base_instance", 1)
706system_value("draw_id", 1)
707system_value("sample_id", 1)
708# sample_id_no_per_sample is like sample_id but does not imply per-
709# sample shading.  See the lower_helper_invocation option.
710system_value("sample_id_no_per_sample", 1)
711system_value("sample_pos", 2)
712system_value("sample_mask_in", 1)
713system_value("primitive_id", 1)
714system_value("invocation_id", 1)
715system_value("tess_coord", 3)
716system_value("tess_level_outer", 4)
717system_value("tess_level_inner", 2)
718system_value("tess_level_outer_default", 4)
719system_value("tess_level_inner_default", 2)
720system_value("patch_vertices_in", 1)
721system_value("local_invocation_id", 3)
722system_value("local_invocation_index", 1)
723# zero_base indicates it starts from 0 for the current dispatch
724# non-zero_base indicates the base is included
725system_value("workgroup_id", 3, bit_sizes=[32, 64])
726system_value("workgroup_id_zero_base", 3)
727system_value("base_workgroup_id", 3, bit_sizes=[32, 64])
728system_value("user_clip_plane", 4, indices=[UCP_ID])
729system_value("num_workgroups", 3, bit_sizes=[32, 64])
730system_value("helper_invocation", 1, bit_sizes=[1, 32])
731system_value("layer_id", 1)
732system_value("view_index", 1)
733system_value("subgroup_size", 1)
734system_value("subgroup_invocation", 1)
735system_value("subgroup_eq_mask", 0, bit_sizes=[32, 64])
736system_value("subgroup_ge_mask", 0, bit_sizes=[32, 64])
737system_value("subgroup_gt_mask", 0, bit_sizes=[32, 64])
738system_value("subgroup_le_mask", 0, bit_sizes=[32, 64])
739system_value("subgroup_lt_mask", 0, bit_sizes=[32, 64])
740system_value("num_subgroups", 1)
741system_value("subgroup_id", 1)
742system_value("workgroup_size", 3)
743# note: the definition of global_invocation_id_zero_base is based on
744# (workgroup_id * workgroup_size) + local_invocation_id.
745# it is *not* based on workgroup_id_zero_base, meaning the work group
746# base is already accounted for, and the global base is additive on top of that
747system_value("global_invocation_id", 3, bit_sizes=[32, 64])
748system_value("global_invocation_id_zero_base", 3, bit_sizes=[32, 64])
749system_value("base_global_invocation_id", 3, bit_sizes=[32, 64])
750system_value("global_invocation_index", 1, bit_sizes=[32, 64])
751system_value("work_dim", 1)
752system_value("line_width", 1)
753system_value("aa_line_width", 1)
754# BASE=0 for global/shader, BASE=1 for local/function
755system_value("scratch_base_ptr", 0, bit_sizes=[32,64], indices=[BASE])
756system_value("constant_base_ptr", 0, bit_sizes=[32,64])
757system_value("shared_base_ptr", 0, bit_sizes=[32,64])
758
759# System values for ray tracing.
760system_value("ray_launch_id", 3)
761system_value("ray_launch_size", 3)
762system_value("ray_world_origin", 3)
763system_value("ray_world_direction", 3)
764system_value("ray_object_origin", 3)
765system_value("ray_object_direction", 3)
766system_value("ray_t_min", 1)
767system_value("ray_t_max", 1)
768system_value("ray_object_to_world", 3, indices=[COLUMN])
769system_value("ray_world_to_object", 3, indices=[COLUMN])
770system_value("ray_hit_kind", 1)
771system_value("ray_flags", 1)
772system_value("ray_geometry_index", 1)
773system_value("ray_instance_custom_index", 1)
774system_value("shader_record_ptr", 1, bit_sizes=[64])
775
776# Driver-specific viewport scale/offset parameters.
777#
778# VC4 and V3D need to emit a scaled version of the position in the vertex
779# shaders for binning, and having system values lets us move the math for that
780# into NIR.
781#
782# Panfrost needs to implement all coordinate transformation in the
783# vertex shader; system values allow us to share this routine in NIR.
784#
785# RADV uses these for NGG primitive culling.
786system_value("viewport_x_scale", 1)
787system_value("viewport_y_scale", 1)
788system_value("viewport_z_scale", 1)
789system_value("viewport_x_offset", 1)
790system_value("viewport_y_offset", 1)
791system_value("viewport_z_offset", 1)
792system_value("viewport_scale", 3)
793system_value("viewport_offset", 3)
794
795# Blend constant color values.  Float values are clamped. Vectored versions are
796# provided as well for driver convenience
797
798system_value("blend_const_color_r_float", 1)
799system_value("blend_const_color_g_float", 1)
800system_value("blend_const_color_b_float", 1)
801system_value("blend_const_color_a_float", 1)
802system_value("blend_const_color_rgba", 4)
803system_value("blend_const_color_rgba8888_unorm", 1)
804system_value("blend_const_color_aaaa8888_unorm", 1)
805
806# System values for gl_Color, for radeonsi which interpolates these in the
807# shader prolog to handle two-sided color without recompiles and therefore
808# doesn't handle these in the main shader part like normal varyings.
809system_value("color0", 4)
810system_value("color1", 4)
811
812# System value for internal compute shaders in radeonsi.
813system_value("user_data_amd", 4)
814
815# Barycentric coordinate intrinsics.
816#
817# These set up the barycentric coordinates for a particular interpolation.
818# The first four are for the simple cases: pixel, centroid, per-sample
819# (at gl_SampleID), or pull model (1/W, 1/I, 1/J) at the pixel center. The next
820# two handle interpolating at a specified sample location, or interpolating
821# with a vec2 offset,
822#
823# The interp_mode index should be either the INTERP_MODE_SMOOTH or
824# INTERP_MODE_NOPERSPECTIVE enum values.
825#
826# The vec2 value produced by these intrinsics is intended for use as the
827# barycoord source of a load_interpolated_input intrinsic.
828
829def barycentric(name, dst_comp, src_comp=[]):
830    intrinsic("load_barycentric_" + name, src_comp=src_comp, dest_comp=dst_comp,
831              indices=[INTERP_MODE], flags=[CAN_ELIMINATE, CAN_REORDER])
832
833# no sources.
834barycentric("pixel", 2)
835barycentric("centroid", 2)
836barycentric("sample", 2)
837barycentric("model", 3)
838# src[] = { sample_id }.
839barycentric("at_sample", 2, [1])
840# src[] = { offset.xy }.
841barycentric("at_offset", 2, [2])
842
843# Load sample position:
844#
845# Takes a sample # and returns a sample position.  Used for lowering
846# interpolateAtSample() to interpolateAtOffset()
847intrinsic("load_sample_pos_from_id", src_comp=[1], dest_comp=2,
848          flags=[CAN_ELIMINATE, CAN_REORDER])
849
850# Loads what I believe is the primitive size, for scaling ij to pixel size:
851intrinsic("load_size_ir3", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
852
853# Load texture scaling values:
854#
855# Takes a sampler # and returns 1/size values for multiplying to normalize
856# texture coordinates.  Used for lowering rect textures.
857intrinsic("load_texture_rect_scaling", src_comp=[1], dest_comp=2,
858          flags=[CAN_ELIMINATE, CAN_REORDER])
859
860# Fragment shader input interpolation delta intrinsic.
861#
862# For hw where fragment shader input interpolation is handled in shader, the
863# load_fs_input_interp deltas intrinsics can be used to load the input deltas
864# used for interpolation as follows:
865#
866#    vec3 iid = load_fs_input_interp_deltas(varying_slot)
867#    vec2 bary = load_barycentric_*(...)
868#    float result = iid.x + iid.y * bary.y + iid.z * bary.x
869
870intrinsic("load_fs_input_interp_deltas", src_comp=[1], dest_comp=3,
871          indices=[BASE, COMPONENT, IO_SEMANTICS], flags=[CAN_ELIMINATE, CAN_REORDER])
872
873# Load operations pull data from some piece of GPU memory.  All load
874# operations operate in terms of offsets into some piece of theoretical
875# memory.  Loads from externally visible memory (UBO and SSBO) simply take a
876# byte offset as a source.  Loads from opaque memory (uniforms, inputs, etc.)
877# take a base+offset pair where the nir_intrinsic_base() gives the location
878# of the start of the variable being loaded and and the offset source is a
879# offset into that variable.
880#
881# Uniform load operations have a nir_intrinsic_range() index that specifies the
882# range (starting at base) of the data from which we are loading.  If
883# range == 0, then the range is unknown.
884#
885# UBO load operations have a nir_intrinsic_range_base() and
886# nir_intrinsic_range() that specify the byte range [range_base,
887# range_base+range] of the UBO that the src offset access must lie within.
888#
889# Some load operations such as UBO/SSBO load and per_vertex loads take an
890# additional source to specify which UBO/SSBO/vertex to load from.
891#
892# The exact address type depends on the lowering pass that generates the
893# load/store intrinsics.  Typically, this is vec4 units for things such as
894# varying slots and float units for fragment shader inputs.  UBO and SSBO
895# offsets are always in bytes.
896
897def load(name, src_comp, indices=[], flags=[]):
898    intrinsic("load_" + name, src_comp, dest_comp=0, indices=indices,
899              flags=flags)
900
901# src[] = { offset }.
902load("uniform", [1], [BASE, RANGE, DEST_TYPE], [CAN_ELIMINATE, CAN_REORDER])
903# src[] = { buffer_index, offset }.
904load("ubo", [-1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET, RANGE_BASE, RANGE], flags=[CAN_ELIMINATE, CAN_REORDER])
905# src[] = { buffer_index, offset in vec4 units }
906load("ubo_vec4", [-1, 1], [ACCESS, COMPONENT], flags=[CAN_ELIMINATE, CAN_REORDER])
907# src[] = { offset }.
908load("input", [1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
909# src[] = { vertex_id, offset }.
910load("input_vertex", [1, 1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
911# src[] = { vertex, offset }.
912load("per_vertex_input", [1, 1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
913# src[] = { barycoord, offset }.
914load("interpolated_input", [2, 1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
915
916# src[] = { buffer_index, offset }.
917load("ssbo", [-1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
918# src[] = { buffer_index }
919load("ssbo_address", [1], [], [CAN_ELIMINATE, CAN_REORDER])
920# src[] = { offset }.
921load("output", [1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], flags=[CAN_ELIMINATE])
922# src[] = { vertex, offset }.
923load("per_vertex_output", [1, 1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE])
924# src[] = { primitive, offset }.
925load("per_primitive_output", [1, 1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE])
926# src[] = { offset }.
927load("shared", [1], [BASE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
928# src[] = { offset }.
929load("push_constant", [1], [BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
930# src[] = { offset }.
931load("constant", [1], [BASE, RANGE, ALIGN_MUL, ALIGN_OFFSET],
932     [CAN_ELIMINATE, CAN_REORDER])
933# src[] = { address }.
934load("global", [1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
935# src[] = { address }.
936load("global_constant", [1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET],
937     [CAN_ELIMINATE, CAN_REORDER])
938# src[] = { base_address, offset }.
939load("global_constant_offset", [1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET],
940     [CAN_ELIMINATE, CAN_REORDER])
941# src[] = { base_address, offset, bound }.
942load("global_constant_bounded", [1, 1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET],
943     [CAN_ELIMINATE, CAN_REORDER])
944# src[] = { address }.
945load("kernel_input", [1], [BASE, RANGE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE, CAN_REORDER])
946# src[] = { offset }.
947load("scratch", [1], [ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
948
949# Stores work the same way as loads, except now the first source is the value
950# to store and the second (and possibly third) source specify where to store
951# the value.  SSBO and shared memory stores also have a
952# nir_intrinsic_write_mask()
953
954def store(name, srcs, indices=[], flags=[]):
955    intrinsic("store_" + name, [0] + srcs, indices=indices, flags=flags)
956
957# src[] = { value, offset }.
958store("output", [1], [BASE, WRITE_MASK, COMPONENT, SRC_TYPE, IO_SEMANTICS])
959# src[] = { value, vertex, offset }.
960store("per_vertex_output", [1, 1], [BASE, WRITE_MASK, COMPONENT, SRC_TYPE, IO_SEMANTICS])
961# src[] = { value, primitive, offset }.
962store("per_primitive_output", [1, 1], [BASE, WRITE_MASK, COMPONENT, SRC_TYPE, IO_SEMANTICS])
963# src[] = { value, block_index, offset }
964store("ssbo", [-1, 1], [WRITE_MASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
965# src[] = { value, offset }.
966store("shared", [1], [BASE, WRITE_MASK, ALIGN_MUL, ALIGN_OFFSET])
967# src[] = { value, address }.
968store("global", [1], [WRITE_MASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
969# src[] = { value, offset }.
970store("scratch", [1], [ALIGN_MUL, ALIGN_OFFSET, WRITE_MASK])
971
972# A bit field to implement SPIRV FragmentShadingRateKHR
973# bit | name              | description
974#   0 | Vertical2Pixels   | Fragment invocation covers 2 pixels vertically
975#   1 | Vertical4Pixels   | Fragment invocation covers 4 pixels vertically
976#   2 | Horizontal2Pixels | Fragment invocation covers 2 pixels horizontally
977#   3 | Horizontal4Pixels | Fragment invocation covers 4 pixels horizontally
978intrinsic("load_frag_shading_rate", dest_comp=1, bit_sizes=[32],
979          flags=[CAN_ELIMINATE, CAN_REORDER])
980
981# OpenCL printf instruction
982# First source is a deref to the format string
983# Second source is a deref to a struct containing the args
984# Dest is success or failure
985intrinsic("printf", src_comp=[1, 1], dest_comp=1, bit_sizes=[32])
986# Since most drivers will want to lower to just dumping args
987# in a buffer, nir_lower_printf will do that, but requires
988# the driver to at least provide a base location
989system_value("printf_buffer_address", 1, bit_sizes=[32,64])
990
991# IR3-specific version of most SSBO intrinsics. The only different
992# compare to the originals is that they add an extra source to hold
993# the dword-offset, which is needed by the backend code apart from
994# the byte-offset already provided by NIR in one of the sources.
995#
996# NIR lowering pass 'ir3_nir_lower_io_offset' will replace the
997# original SSBO intrinsics by these, placing the computed
998# dword-offset always in the last source.
999#
1000# The float versions are not handled because those are not supported
1001# by the backend.
1002store("ssbo_ir3", [1, 1, 1],
1003      indices=[WRITE_MASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
1004load("ssbo_ir3",  [1, 1, 1],
1005     indices=[ACCESS, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN_ELIMINATE])
1006intrinsic("ssbo_atomic_add_ir3",        src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
1007intrinsic("ssbo_atomic_imin_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
1008intrinsic("ssbo_atomic_umin_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
1009intrinsic("ssbo_atomic_imax_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
1010intrinsic("ssbo_atomic_umax_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
1011intrinsic("ssbo_atomic_and_ir3",        src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
1012intrinsic("ssbo_atomic_or_ir3",         src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
1013intrinsic("ssbo_atomic_xor_ir3",        src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
1014intrinsic("ssbo_atomic_exchange_ir3",   src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
1015intrinsic("ssbo_atomic_comp_swap_ir3",  src_comp=[1, 1, 1, 1, 1], dest_comp=1, indices=[ACCESS])
1016
1017# System values for freedreno geometry shaders.
1018system_value("vs_primitive_stride_ir3", 1)
1019system_value("vs_vertex_stride_ir3", 1)
1020system_value("gs_header_ir3", 1)
1021system_value("primitive_location_ir3", 1, indices=[DRIVER_LOCATION])
1022
1023# System values for freedreno tessellation shaders.
1024system_value("hs_patch_stride_ir3", 1)
1025system_value("tess_factor_base_ir3", 2)
1026system_value("tess_param_base_ir3", 2)
1027system_value("tcs_header_ir3", 1)
1028system_value("rel_patch_id_ir3", 1)
1029
1030# System values for freedreno compute shaders.
1031system_value("subgroup_id_shift_ir3", 1)
1032
1033# IR3-specific intrinsics for tessellation control shaders.  cond_end_ir3 end
1034# the shader when src0 is false and is used to narrow down the TCS shader to
1035# just thread 0 before writing out tessellation levels.
1036intrinsic("cond_end_ir3", src_comp=[1])
1037# end_patch_ir3 is used just before thread 0 exist the TCS and presumably
1038# signals the TE that the patch is complete and can be tessellated.
1039intrinsic("end_patch_ir3")
1040
1041# IR3-specific load/store intrinsics. These access a buffer used to pass data
1042# between geometry stages - perhaps it's explicit access to the vertex cache.
1043
1044# src[] = { value, offset }.
1045store("shared_ir3", [1], [BASE, ALIGN_MUL, ALIGN_OFFSET])
1046# src[] = { offset }.
1047load("shared_ir3", [1], [BASE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1048
1049# IR3-specific load/store global intrinsics. They take a 64-bit base address
1050# and a 32-bit offset.  The hardware will add the base and the offset, which
1051# saves us from doing 64-bit math on the base address.
1052
1053# src[] = { value, address(vec2 of hi+lo uint32_t), offset }.
1054# const_index[] = { write_mask, align_mul, align_offset }
1055store("global_ir3", [2, 1], indices=[ACCESS, ALIGN_MUL, ALIGN_OFFSET])
1056# src[] = { address(vec2 of hi+lo uint32_t), offset }.
1057# const_index[] = { access, align_mul, align_offset }
1058load("global_ir3", [2, 1], indices=[ACCESS, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN_ELIMINATE])
1059
1060# IR3-specific bindless handle specifier. Similar to vulkan_resource_index, but
1061# without the binding because the hardware expects a single flattened index
1062# rather than a (binding, index) pair. We may also want to use this with GL.
1063# Note that this doesn't actually turn into a HW instruction.
1064intrinsic("bindless_resource_ir3", [1], dest_comp=1, indices=[DESC_SET], flags=[CAN_ELIMINATE, CAN_REORDER])
1065
1066# DXIL specific intrinsics
1067# src[] = { value, mask, index, offset }.
1068intrinsic("store_ssbo_masked_dxil", [1, 1, 1, 1])
1069# src[] = { value, index }.
1070intrinsic("store_shared_dxil", [1, 1])
1071# src[] = { value, mask, index }.
1072intrinsic("store_shared_masked_dxil", [1, 1, 1])
1073# src[] = { value, index }.
1074intrinsic("store_scratch_dxil", [1, 1])
1075# src[] = { index }.
1076load("shared_dxil", [1], [], [CAN_ELIMINATE])
1077# src[] = { index }.
1078load("scratch_dxil", [1], [], [CAN_ELIMINATE])
1079# src[] = { deref_var, offset }
1080load("ptr_dxil", [1, 1], [], [])
1081# src[] = { index, 16-byte-based-offset }
1082load("ubo_dxil", [1, 1], [], [CAN_ELIMINATE, CAN_REORDER])
1083
1084# DXIL Shared atomic intrinsics
1085#
1086# All of the shared variable atomic memory operations read a value from
1087# memory, compute a new value using one of the operations below, write the
1088# new value to memory, and return the original value read.
1089#
1090# All operations take 2 sources:
1091#
1092# 0: The index in the i32 array for by the shared memory region
1093# 1: The data parameter to the atomic function (i.e. the value to add
1094#    in shared_atomic_add, etc).
1095intrinsic("shared_atomic_add_dxil",  src_comp=[1, 1], dest_comp=1)
1096intrinsic("shared_atomic_imin_dxil", src_comp=[1, 1], dest_comp=1)
1097intrinsic("shared_atomic_umin_dxil", src_comp=[1, 1], dest_comp=1)
1098intrinsic("shared_atomic_imax_dxil", src_comp=[1, 1], dest_comp=1)
1099intrinsic("shared_atomic_umax_dxil", src_comp=[1, 1], dest_comp=1)
1100intrinsic("shared_atomic_and_dxil",  src_comp=[1, 1], dest_comp=1)
1101intrinsic("shared_atomic_or_dxil",   src_comp=[1, 1], dest_comp=1)
1102intrinsic("shared_atomic_xor_dxil",  src_comp=[1, 1], dest_comp=1)
1103intrinsic("shared_atomic_exchange_dxil", src_comp=[1, 1], dest_comp=1)
1104intrinsic("shared_atomic_comp_swap_dxil", src_comp=[1, 1, 1], dest_comp=1)
1105
1106# Intrinsics used by the Midgard/Bifrost blend pipeline. These are defined
1107# within a blend shader to read/write the raw value from the tile buffer,
1108# without applying any format conversion in the process. If the shader needs
1109# usable pixel values, it must apply format conversions itself.
1110#
1111# These definitions are generic, but they are explicitly vendored to prevent
1112# other drivers from using them, as their semantics is defined in terms of the
1113# Midgard/Bifrost hardware tile buffer and may not line up with anything sane.
1114# One notable divergence is sRGB, which is asymmetric: raw_input_pan requires
1115# an sRGB->linear conversion, but linear values should be written to
1116# raw_output_pan and the hardware handles linear->sRGB.
1117
1118# src[] = { value }
1119store("raw_output_pan", [], [])
1120store("combined_output_pan", [1, 1, 1], [BASE, COMPONENT, SRC_TYPE])
1121load("raw_output_pan", [1], [BASE], [CAN_ELIMINATE, CAN_REORDER])
1122
1123# Loads the sampler paramaters <min_lod, max_lod, lod_bias>
1124# src[] = { sampler_index }
1125load("sampler_lod_parameters_pan", [1], flags=[CAN_ELIMINATE, CAN_REORDER])
1126
1127# Loads the sample position array on Bifrost, in a packed Arm-specific format
1128system_value("sample_positions_pan", 1, bit_sizes=[64])
1129
1130# R600 specific instrincs
1131#
1132# location where the tesselation data is stored in LDS
1133system_value("tcs_in_param_base_r600", 4)
1134system_value("tcs_out_param_base_r600", 4)
1135system_value("tcs_rel_patch_id_r600", 1)
1136system_value("tcs_tess_factor_base_r600", 1)
1137
1138# the tess coords come as xy only, z has to be calculated
1139system_value("tess_coord_r600", 2)
1140
1141# load as many components as needed giving per-component addresses
1142intrinsic("load_local_shared_r600", src_comp=[0], dest_comp=0, indices = [], flags = [CAN_ELIMINATE])
1143
1144store("local_shared_r600", [1], [WRITE_MASK])
1145store("tf_r600", [])
1146
1147# AMD GCN/RDNA specific intrinsics
1148
1149# src[] = { descriptor, base address, scalar offset }
1150intrinsic("load_buffer_amd", src_comp=[4, 1, 1], dest_comp=0, indices=[BASE, IS_SWIZZLED, SLC_AMD, MEMORY_MODES], flags=[CAN_ELIMINATE])
1151# src[] = { store value, descriptor, base address, scalar offset }
1152intrinsic("store_buffer_amd", src_comp=[0, 4, 1, 1], indices=[BASE, WRITE_MASK, IS_SWIZZLED, SLC_AMD, MEMORY_MODES])
1153
1154# Same as shared_atomic_add, but with GDS. src[] = {store_val, gds_addr, m0}
1155intrinsic("gds_atomic_add_amd",  src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
1156
1157# Descriptor where TCS outputs are stored for TES
1158system_value("ring_tess_offchip_amd", 4)
1159system_value("ring_tess_offchip_offset_amd", 1)
1160# Descriptor where TCS outputs are stored for the HW tessellator
1161system_value("ring_tess_factors_amd", 4)
1162system_value("ring_tess_factors_offset_amd", 1)
1163# Descriptor where ES outputs are stored for GS to read on GFX6-8
1164system_value("ring_esgs_amd", 4)
1165system_value("ring_es2gs_offset_amd", 1)
1166
1167# Number of patches processed by each TCS workgroup
1168system_value("tcs_num_patches_amd", 1)
1169# Relative tessellation patch ID within the current workgroup
1170system_value("tess_rel_patch_id_amd", 1)
1171# Vertex offsets used for GS per-vertex inputs
1172system_value("gs_vertex_offset_amd", 1, [BASE])
1173
1174# AMD merged shader intrinsics
1175
1176# Whether the current invocation has an input vertex / primitive to process (also known as "ES thread" or "GS thread").
1177# Not safe to reorder because it changes after overwrite_subgroup_num_vertices_and_primitives_amd.
1178# Also, the generated code is more optimal if they are not CSE'd.
1179intrinsic("has_input_vertex_amd", src_comp=[], dest_comp=1, bit_sizes=[1], indices=[])
1180intrinsic("has_input_primitive_amd", src_comp=[], dest_comp=1, bit_sizes=[1], indices=[])
1181
1182# AMD NGG intrinsics
1183
1184# Number of initial input vertices in the current workgroup.
1185system_value("workgroup_num_input_vertices_amd", 1)
1186# Number of initial input primitives in the current workgroup.
1187system_value("workgroup_num_input_primitives_amd", 1)
1188# For NGG passthrough mode only. Pre-packed argument for export_primitive_amd.
1189system_value("packed_passthrough_primitive_amd", 1)
1190# Whether NGG GS should execute shader query.
1191system_value("shader_query_enabled_amd", dest_comp=1, bit_sizes=[1])
1192# Whether the shader should cull front facing triangles.
1193intrinsic("load_cull_front_face_enabled_amd", dest_comp=1, bit_sizes=[1], flags=[CAN_ELIMINATE])
1194# Whether the shader should cull back facing triangles.
1195intrinsic("load_cull_back_face_enabled_amd", dest_comp=1, bit_sizes=[1], flags=[CAN_ELIMINATE])
1196# True if face culling should use CCW (false if CW).
1197intrinsic("load_cull_ccw_amd", dest_comp=1, bit_sizes=[1], flags=[CAN_ELIMINATE])
1198# Whether the shader should cull small primitives that are not visible in a pixel.
1199intrinsic("load_cull_small_primitives_enabled_amd", dest_comp=1, bit_sizes=[1], flags=[CAN_ELIMINATE])
1200# Whether any culling setting is enabled in the shader.
1201intrinsic("load_cull_any_enabled_amd", dest_comp=1, bit_sizes=[1], flags=[CAN_ELIMINATE])
1202# Small primitive culling precision
1203intrinsic("load_cull_small_prim_precision_amd", dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE, CAN_REORDER])
1204# Initial edge flags in a Vertex Shader, packed into the format the HW needs for primitive export.
1205intrinsic("load_initial_edgeflags_amd", src_comp=[], dest_comp=1, bit_sizes=[32], indices=[])
1206# Exports the current invocation's vertex. This is a placeholder where all vertex attribute export instructions should be emitted.
1207intrinsic("export_vertex_amd", src_comp=[], indices=[])
1208# Exports the current invocation's primitive. src[] = {packed_primitive_data}.
1209intrinsic("export_primitive_amd", src_comp=[1], indices=[])
1210# Allocates export space for vertices and primitives. src[] = {num_vertices, num_primitives}.
1211intrinsic("alloc_vertices_and_primitives_amd", src_comp=[1, 1], indices=[])
1212# Overwrites VS input registers, for use with vertex compaction after culling. src = {vertex_id, instance_id}.
1213intrinsic("overwrite_vs_arguments_amd", src_comp=[1, 1], indices=[])
1214# Overwrites TES input registers, for use with vertex compaction after culling. src = {tes_u, tes_v, rel_patch_id, patch_id}.
1215intrinsic("overwrite_tes_arguments_amd", src_comp=[1, 1, 1, 1], indices=[])
1216
1217# loads a descriptor for an sbt.
1218# src = [index] BINDING = which table
1219intrinsic("load_sbt_amd", dest_comp=4, bit_sizes=[32], indices=[BINDING],
1220          flags=[CAN_ELIMINATE, CAN_REORDER])
1221
1222# 1. HW descriptor
1223# 2. BVH node(64-bit pointer as 2x32 ...)
1224# 3. ray extent
1225# 4. ray origin
1226# 5. ray direction
1227# 6. inverse ray direction (componentwise 1.0/ray direction)
1228intrinsic("bvh64_intersect_ray_amd", [4, 2, 1, 3, 3, 3], 4, flags=[CAN_ELIMINATE, CAN_REORDER])
1229
1230# Return of a callable in raytracing pipelines
1231intrinsic("rt_return_amd")
1232
1233# offset into scratch for the input callable data in a raytracing pipeline.
1234system_value("rt_arg_scratch_offset_amd", 1)
1235
1236# Whether to call the anyhit shader for an intersection in an intersection shader.
1237system_value("intersection_opaque_amd", 1, bit_sizes=[1])
1238
1239# V3D-specific instrinc for tile buffer color reads.
1240#
1241# The hardware requires that we read the samples and components of a pixel
1242# in order, so we cannot eliminate or remove any loads in a sequence.
1243#
1244# src[] = { render_target }
1245# BASE = sample index
1246load("tlb_color_v3d", [1], [BASE, COMPONENT], [])
1247
1248# V3D-specific instrinc for per-sample tile buffer color writes.
1249#
1250# The driver backend needs to identify per-sample color writes and emit
1251# specific code for them.
1252#
1253# src[] = { value, render_target }
1254# BASE = sample index
1255store("tlb_sample_color_v3d", [1], [BASE, COMPONENT, SRC_TYPE], [])
1256
1257# V3D-specific intrinsic to load the number of layers attached to
1258# the target framebuffer
1259intrinsic("load_fb_layers_v3d", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
1260
1261# Logical complement of load_front_face, mapping to an AGX system value
1262system_value("back_face_agx", 1, bit_sizes=[1, 32])
1263
1264# Intel-specific query for loading from the brw_image_param struct passed
1265# into the shader as a uniform.  The variable is a deref to the image
1266# variable. The const index specifies which of the six parameters to load.
1267intrinsic("image_deref_load_param_intel", src_comp=[1], dest_comp=0,
1268          indices=[BASE], flags=[CAN_ELIMINATE, CAN_REORDER])
1269image("load_raw_intel", src_comp=[1], dest_comp=0,
1270      flags=[CAN_ELIMINATE])
1271image("store_raw_intel", src_comp=[1, 0])
1272
1273# Intrinsic to load a block of at least 32B of constant data from a 64-bit
1274# global memory address.  The memory address must be uniform and 32B-aligned.
1275# The second source is a predicate which indicates whether or not to actually
1276# do the load.
1277# src[] = { address, predicate }.
1278intrinsic("load_global_const_block_intel", src_comp=[1, 1], dest_comp=0,
1279          bit_sizes=[32], indices=[BASE], flags=[CAN_ELIMINATE, CAN_REORDER])
1280
1281# Number of data items being operated on for a SIMD program.
1282system_value("simd_width_intel", 1)
1283
1284# Load a relocatable 32-bit value
1285intrinsic("load_reloc_const_intel", dest_comp=1, bit_sizes=[32],
1286          indices=[PARAM_IDX], flags=[CAN_ELIMINATE, CAN_REORDER])
1287
1288# 64-bit global address for a Vulkan descriptor set
1289# src[0] = { set }
1290intrinsic("load_desc_set_address_intel", dest_comp=1, bit_sizes=[64],
1291          src_comp=[1], flags=[CAN_ELIMINATE, CAN_REORDER])
1292
1293# OpSubgroupBlockReadINTEL and OpSubgroupBlockWriteINTEL from SPV_INTEL_subgroups.
1294intrinsic("load_deref_block_intel", dest_comp=0, src_comp=[-1],
1295          indices=[ACCESS], flags=[CAN_ELIMINATE])
1296intrinsic("store_deref_block_intel", src_comp=[-1, 0], indices=[WRITE_MASK, ACCESS])
1297
1298# src[] = { address }.
1299load("global_block_intel", [1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1300
1301# src[] = { buffer_index, offset }.
1302load("ssbo_block_intel", [-1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1303
1304# src[] = { offset }.
1305load("shared_block_intel", [1], [BASE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1306
1307# src[] = { value, address }.
1308store("global_block_intel", [1], [WRITE_MASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
1309
1310# src[] = { value, block_index, offset }
1311store("ssbo_block_intel", [-1, 1], [WRITE_MASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
1312
1313# src[] = { value, offset }.
1314store("shared_block_intel", [1], [BASE, WRITE_MASK, ALIGN_MUL, ALIGN_OFFSET])
1315
1316# Intrinsics for Intel bindless thread dispatch
1317system_value("btd_dss_id_intel", 1)
1318system_value("btd_stack_id_intel", 1)
1319system_value("btd_global_arg_addr_intel", 1, bit_sizes=[64])
1320system_value("btd_local_arg_addr_intel", 1, bit_sizes=[64])
1321system_value("btd_resume_sbt_addr_intel", 1, bit_sizes=[64])
1322# src[] = { global_arg_addr, btd_record }
1323intrinsic("btd_spawn_intel", src_comp=[1, 1])
1324# RANGE=stack_size
1325intrinsic("btd_stack_push_intel", indices=[STACK_SIZE])
1326# src[] = { }
1327intrinsic("btd_retire_intel")
1328
1329# Intel-specific ray-tracing intrinsics
1330intrinsic("trace_ray_initial_intel")
1331intrinsic("trace_ray_commit_intel")
1332intrinsic("trace_ray_continue_intel")
1333
1334# System values used for ray-tracing on Intel
1335system_value("ray_base_mem_addr_intel", 1, bit_sizes=[64])
1336system_value("ray_hw_stack_size_intel", 1)
1337system_value("ray_sw_stack_size_intel", 1)
1338system_value("ray_num_dss_rt_stacks_intel", 1)
1339system_value("ray_hit_sbt_addr_intel", 1, bit_sizes=[64])
1340system_value("ray_hit_sbt_stride_intel", 1, bit_sizes=[16])
1341system_value("ray_miss_sbt_addr_intel", 1, bit_sizes=[64])
1342system_value("ray_miss_sbt_stride_intel", 1, bit_sizes=[16])
1343system_value("callable_sbt_addr_intel", 1, bit_sizes=[64])
1344system_value("callable_sbt_stride_intel", 1, bit_sizes=[16])
1345system_value("leaf_opaque_intel", 1, bit_sizes=[1])
1346system_value("leaf_procedural_intel", 1, bit_sizes=[1])
1347