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# Offsets for load_shared2_amd/store_shared2_amd
230index("uint8_t", "offset0")
231index("uint8_t", "offset1")
232
233# If true, both offsets have an additional stride of 64 dwords (ie. they are multiplied by 256 bytes
234# in hardware, instead of 4).
235index("bool", "st64")
236
237# When set, range analysis will use it for nir_unsigned_upper_bound
238index("unsigned", "arg_upper_bound_u32_amd")
239
240# Separate source/dest access flags for copies
241index("enum gl_access_qualifier", "dst_access")
242index("enum gl_access_qualifier", "src_access")
243
244# Driver location of attribute
245index("unsigned", "driver_location")
246
247# Ordering and visibility of a memory operation
248index("nir_memory_semantics", "memory_semantics")
249
250# Modes affected by a memory operation
251index("nir_variable_mode", "memory_modes")
252
253# Scope of a memory operation
254index("nir_scope", "memory_scope")
255
256# Scope of a control barrier
257index("nir_scope", "execution_scope")
258
259# Semantics of an IO instruction
260index("struct nir_io_semantics", "io_semantics")
261
262# Transform feedback info
263index("struct nir_io_xfb", "io_xfb")
264index("struct nir_io_xfb", "io_xfb2")
265
266# Rounding mode for conversions
267index("nir_rounding_mode", "rounding_mode")
268
269# Whether or not to saturate in conversions
270index("unsigned", "saturate")
271
272# Whether or not trace_ray_intel is synchronous
273index("bool", "synchronous")
274
275intrinsic("nop", flags=[CAN_ELIMINATE])
276
277intrinsic("convert_alu_types", dest_comp=0, src_comp=[0],
278          indices=[SRC_TYPE, DEST_TYPE, ROUNDING_MODE, SATURATE],
279          flags=[CAN_ELIMINATE, CAN_REORDER])
280
281intrinsic("load_param", dest_comp=0, indices=[PARAM_IDX], flags=[CAN_ELIMINATE])
282
283intrinsic("load_deref", dest_comp=0, src_comp=[-1],
284          indices=[ACCESS], flags=[CAN_ELIMINATE])
285intrinsic("store_deref", src_comp=[-1, 0], indices=[WRITE_MASK, ACCESS])
286intrinsic("copy_deref", src_comp=[-1, -1], indices=[DST_ACCESS, SRC_ACCESS])
287intrinsic("memcpy_deref", src_comp=[-1, -1, 1], indices=[DST_ACCESS, SRC_ACCESS])
288
289# Interpolation of input.  The interp_deref_at* intrinsics are similar to the
290# load_var intrinsic acting on a shader input except that they interpolate the
291# input differently.  The at_sample, at_offset and at_vertex intrinsics take an
292# additional source that is an integer sample id, a vec2 position offset, or a
293# vertex ID respectively.
294
295intrinsic("interp_deref_at_centroid", dest_comp=0, src_comp=[1],
296          flags=[ CAN_ELIMINATE, CAN_REORDER])
297intrinsic("interp_deref_at_sample", src_comp=[1, 1], dest_comp=0,
298          flags=[CAN_ELIMINATE, CAN_REORDER])
299intrinsic("interp_deref_at_offset", src_comp=[1, 2], dest_comp=0,
300          flags=[CAN_ELIMINATE, CAN_REORDER])
301intrinsic("interp_deref_at_vertex", src_comp=[1, 1], dest_comp=0,
302          flags=[CAN_ELIMINATE, CAN_REORDER])
303
304# Gets the length of an unsized array at the end of a buffer
305intrinsic("deref_buffer_array_length", src_comp=[-1], dest_comp=1,
306          indices=[ACCESS], flags=[CAN_ELIMINATE, CAN_REORDER])
307
308# Ask the driver for the size of a given SSBO. It takes the buffer index
309# as source.
310intrinsic("get_ssbo_size", src_comp=[-1], dest_comp=1, bit_sizes=[32],
311          indices=[ACCESS], flags=[CAN_ELIMINATE, CAN_REORDER])
312intrinsic("get_ubo_size", src_comp=[-1], dest_comp=1,
313          flags=[CAN_ELIMINATE, CAN_REORDER])
314
315# Intrinsics which provide a run-time mode-check.  Unlike the compile-time
316# mode checks, a pointer can only have exactly one mode at runtime.
317intrinsic("deref_mode_is", src_comp=[-1], dest_comp=1,
318          indices=[MEMORY_MODES], flags=[CAN_ELIMINATE, CAN_REORDER])
319intrinsic("addr_mode_is", src_comp=[-1], dest_comp=1,
320          indices=[MEMORY_MODES], flags=[CAN_ELIMINATE, CAN_REORDER])
321
322intrinsic("is_sparse_texels_resident", dest_comp=1, src_comp=[1], bit_sizes=[1,32],
323          flags=[CAN_ELIMINATE, CAN_REORDER])
324# result code is resident only if both inputs are resident
325intrinsic("sparse_residency_code_and", dest_comp=1, src_comp=[1, 1], bit_sizes=[32],
326          flags=[CAN_ELIMINATE, CAN_REORDER])
327
328# a barrier is an intrinsic with no inputs/outputs but which can't be moved
329# around/optimized in general
330def barrier(name):
331    intrinsic(name)
332
333barrier("discard")
334
335# Demote fragment shader invocation to a helper invocation.  Any stores to
336# memory after this instruction are suppressed and the fragment does not write
337# outputs to the framebuffer.  Unlike discard, demote needs to ensure that
338# derivatives will still work for invocations that were not demoted.
339#
340# As specified by SPV_EXT_demote_to_helper_invocation.
341barrier("demote")
342intrinsic("is_helper_invocation", dest_comp=1, flags=[CAN_ELIMINATE])
343
344# SpvOpTerminateInvocation from SPIR-V.  Essentially a discard "for real".
345barrier("terminate")
346
347# A workgroup-level control barrier.  Any thread which hits this barrier will
348# pause until all threads within the current workgroup have also hit the
349# barrier.  For compute shaders, the workgroup is defined as the local group.
350# For tessellation control shaders, the workgroup is defined as the current
351# patch.  This intrinsic does not imply any sort of memory barrier.
352barrier("control_barrier")
353
354# Memory barrier with semantics analogous to the memoryBarrier() GLSL
355# intrinsic.
356barrier("memory_barrier")
357
358# Control/Memory barrier with explicit scope.  Follows the semantics of SPIR-V
359# OpMemoryBarrier and OpControlBarrier, used to implement Vulkan Memory Model.
360# Storage that the barrier applies is represented using NIR variable modes.
361# For an OpMemoryBarrier, set EXECUTION_SCOPE to NIR_SCOPE_NONE.
362intrinsic("scoped_barrier",
363          indices=[EXECUTION_SCOPE, MEMORY_SCOPE, MEMORY_SEMANTICS, MEMORY_MODES])
364
365# Shader clock intrinsic with semantics analogous to the clock2x32ARB()
366# GLSL intrinsic.
367# The latter can be used as code motion barrier, which is currently not
368# feasible with NIR.
369intrinsic("shader_clock", dest_comp=2, bit_sizes=[32], flags=[CAN_ELIMINATE],
370          indices=[MEMORY_SCOPE])
371
372# Shader ballot intrinsics with semantics analogous to the
373#
374#    ballotARB()
375#    readInvocationARB()
376#    readFirstInvocationARB()
377#
378# GLSL functions from ARB_shader_ballot.
379intrinsic("ballot", src_comp=[1], dest_comp=0, flags=[CAN_ELIMINATE])
380intrinsic("read_invocation", src_comp=[0, 1], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
381intrinsic("read_first_invocation", src_comp=[0], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
382
383# Returns the value of the first source for the lane where the second source is
384# true. The second source must be true for exactly one lane.
385intrinsic("read_invocation_cond_ir3", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
386
387# Additional SPIR-V ballot intrinsics
388#
389# These correspond to the SPIR-V opcodes
390#
391#    OpGroupNonUniformElect
392#    OpSubgroupFirstInvocationKHR
393intrinsic("elect", dest_comp=1, flags=[CAN_ELIMINATE])
394intrinsic("first_invocation", dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE])
395intrinsic("last_invocation", dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE])
396
397# Memory barrier with semantics analogous to the compute shader
398# groupMemoryBarrier(), memoryBarrierAtomicCounter(), memoryBarrierBuffer(),
399# memoryBarrierImage() and memoryBarrierShared() GLSL intrinsics.
400barrier("group_memory_barrier")
401barrier("memory_barrier_atomic_counter")
402barrier("memory_barrier_buffer")
403barrier("memory_barrier_image")
404barrier("memory_barrier_shared")
405barrier("begin_invocation_interlock")
406barrier("end_invocation_interlock")
407
408# Memory barrier for synchronizing TCS patch outputs
409barrier("memory_barrier_tcs_patch")
410
411# A conditional discard/demote/terminate, with a single boolean source.
412intrinsic("discard_if", src_comp=[1])
413intrinsic("demote_if", src_comp=[1])
414intrinsic("terminate_if", src_comp=[1])
415
416# ARB_shader_group_vote intrinsics
417intrinsic("vote_any", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
418intrinsic("vote_all", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
419intrinsic("vote_feq", src_comp=[0], dest_comp=1, flags=[CAN_ELIMINATE])
420intrinsic("vote_ieq", src_comp=[0], dest_comp=1, flags=[CAN_ELIMINATE])
421
422# Ballot ALU operations from SPIR-V.
423#
424# These operations work like their ALU counterparts except that the operate
425# on a uvec4 which is treated as a 128bit integer.  Also, they are, in
426# general, free to ignore any bits which are above the subgroup size.
427intrinsic("ballot_bitfield_extract", src_comp=[4, 1], dest_comp=1, flags=[CAN_ELIMINATE])
428intrinsic("ballot_bit_count_reduce", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
429intrinsic("ballot_bit_count_inclusive", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
430intrinsic("ballot_bit_count_exclusive", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
431intrinsic("ballot_find_lsb", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
432intrinsic("ballot_find_msb", src_comp=[4], dest_comp=1, flags=[CAN_ELIMINATE])
433
434# Shuffle operations from SPIR-V.
435intrinsic("shuffle", src_comp=[0, 1], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
436intrinsic("shuffle_xor", src_comp=[0, 1], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
437intrinsic("shuffle_up", src_comp=[0, 1], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
438intrinsic("shuffle_down", src_comp=[0, 1], dest_comp=0, bit_sizes=src0, flags=[CAN_ELIMINATE])
439
440# Quad operations from SPIR-V.
441intrinsic("quad_broadcast", src_comp=[0, 1], dest_comp=0, flags=[CAN_ELIMINATE])
442intrinsic("quad_swap_horizontal", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
443intrinsic("quad_swap_vertical", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
444intrinsic("quad_swap_diagonal", src_comp=[0], dest_comp=0, flags=[CAN_ELIMINATE])
445
446intrinsic("reduce", src_comp=[0], dest_comp=0, bit_sizes=src0,
447          indices=[REDUCTION_OP, CLUSTER_SIZE], flags=[CAN_ELIMINATE])
448intrinsic("inclusive_scan", src_comp=[0], dest_comp=0, bit_sizes=src0,
449          indices=[REDUCTION_OP], flags=[CAN_ELIMINATE])
450intrinsic("exclusive_scan", src_comp=[0], dest_comp=0, bit_sizes=src0,
451          indices=[REDUCTION_OP], flags=[CAN_ELIMINATE])
452
453# AMD shader ballot operations
454intrinsic("quad_swizzle_amd", src_comp=[0], dest_comp=0, bit_sizes=src0,
455          indices=[SWIZZLE_MASK], flags=[CAN_ELIMINATE])
456intrinsic("masked_swizzle_amd", src_comp=[0], dest_comp=0, bit_sizes=src0,
457          indices=[SWIZZLE_MASK], flags=[CAN_ELIMINATE])
458intrinsic("write_invocation_amd", src_comp=[0, 0, 1], dest_comp=0, bit_sizes=src0,
459          flags=[CAN_ELIMINATE])
460# src = [ mask, addition ]
461intrinsic("mbcnt_amd", src_comp=[1, 1], dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE])
462# Compiled to v_perm_b32. src = [ in_bytes_hi, in_bytes_lo, selector ]
463intrinsic("byte_permute_amd", src_comp=[1, 1, 1], dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE, CAN_REORDER])
464# Compiled to v_permlane16_b32. src = [ value, lanesel_lo, lanesel_hi ]
465intrinsic("lane_permute_16_amd", src_comp=[1, 1, 1], dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE])
466
467# Basic Geometry Shader intrinsics.
468#
469# emit_vertex implements GLSL's EmitStreamVertex() built-in.  It takes a single
470# index, which is the stream ID to write to.
471#
472# end_primitive implements GLSL's EndPrimitive() built-in.
473intrinsic("emit_vertex",   indices=[STREAM_ID])
474intrinsic("end_primitive", indices=[STREAM_ID])
475
476# Geometry Shader intrinsics with a vertex count.
477#
478# Alternatively, drivers may implement these intrinsics, and use
479# nir_lower_gs_intrinsics() to convert from the basic intrinsics.
480#
481# These contain two additional unsigned integer sources:
482# 1. The total number of vertices emitted so far.
483# 2. The number of vertices emitted for the current primitive
484#    so far if we're counting, otherwise undef.
485intrinsic("emit_vertex_with_counter", src_comp=[1, 1], indices=[STREAM_ID])
486intrinsic("end_primitive_with_counter", src_comp=[1, 1], indices=[STREAM_ID])
487# Contains the final total vertex and primitive counts in the current GS thread.
488intrinsic("set_vertex_and_primitive_count", src_comp=[1, 1], indices=[STREAM_ID])
489
490# Launches mesh shader workgroups from a task shader, with explicit task_payload.
491# Rules:
492# - This is a terminating instruction.
493# - May only occur in workgroup-uniform control flow.
494# - Dispatch sizes may be divergent (in which case the values
495#   from the first invocation are used).
496# Meaning of indices:
497# - BASE: address of the task_payload variable used.
498# - RANGE: size of the task_payload variable used.
499#
500# src[] = {vec(x, y, z)}
501intrinsic("launch_mesh_workgroups", src_comp=[3], indices=[BASE, RANGE])
502
503# Trace a ray through an acceleration structure
504#
505# This instruction has a lot of parameters:
506#   0. Acceleration Structure
507#   1. Ray Flags
508#   2. Cull Mask
509#   3. SBT Offset
510#   4. SBT Stride
511#   5. Miss shader index
512#   6. Ray Origin
513#   7. Ray Tmin
514#   8. Ray Direction
515#   9. Ray Tmax
516#   10. Payload
517intrinsic("trace_ray", src_comp=[-1, 1, 1, 1, 1, 1, 3, 1, 3, 1, -1])
518# src[] = { hit_t, hit_kind }
519intrinsic("report_ray_intersection", src_comp=[1, 1], dest_comp=1)
520intrinsic("ignore_ray_intersection")
521intrinsic("accept_ray_intersection") # Not in SPIR-V; useful for lowering
522intrinsic("terminate_ray")
523# src[] = { sbt_index, payload }
524intrinsic("execute_callable", src_comp=[1, -1])
525
526# Initialize a ray query
527#
528#   0. Ray Query
529#   1. Acceleration Structure
530#   2. Ray Flags
531#   3. Cull Mask
532#   4. Ray Origin
533#   5. Ray Tmin
534#   6. Ray Direction
535#   7. Ray Tmax
536intrinsic("rq_initialize", src_comp=[-1, -1, 1, 1, 3, 1, 3, 1])
537# src[] = { query }
538intrinsic("rq_terminate", src_comp=[-1])
539# src[] = { query }
540intrinsic("rq_proceed", src_comp=[-1], dest_comp=1)
541# src[] = { query, hit }
542intrinsic("rq_generate_intersection", src_comp=[-1, 1])
543# src[] = { query }
544intrinsic("rq_confirm_intersection", src_comp=[-1])
545# src[] = { query, committed } BASE=nir_ray_query_value
546intrinsic("rq_load", src_comp=[-1, 1], dest_comp=0, indices=[BASE,COLUMN])
547
548# Driver independent raytracing helpers
549
550# rt_resume is a helper that that be the first instruction accesing the
551# stack/scratch in a resume shader for a raytracing pipeline. It includes the
552# resume index (for nir_lower_shader_calls_internal reasons) and the stack size
553# of the variables spilled during the call. The stack size can be use to e.g.
554# adjust a stack pointer.
555intrinsic("rt_resume", indices=[CALL_IDX, STACK_SIZE])
556
557# Lowered version of execute_callabe that includes the index of the resume
558# shader, and the amount of scratch space needed for this call (.ie. how much
559# to increase a stack pointer by).
560# src[] = { sbt_index, payload }
561intrinsic("rt_execute_callable", src_comp=[1, -1], indices=[CALL_IDX,STACK_SIZE])
562
563# Lowered version of trace_ray in a similar vein to rt_execute_callable.
564# src same as trace_ray
565intrinsic("rt_trace_ray", src_comp=[-1, 1, 1, 1, 1, 1, 3, 1, 3, 1, -1],
566          indices=[CALL_IDX, STACK_SIZE])
567
568
569# Atomic counters
570#
571# The *_deref variants take an atomic_uint nir_variable, while the other,
572# lowered, variants take a buffer index and register offset.  The buffer index
573# is always constant, as there's no way to declare an array of atomic counter
574# buffers.
575#
576# The register offset may be non-constant but must by dynamically uniform
577# ("Atomic counters aggregated into arrays within a shader can only be indexed
578# with dynamically uniform integral expressions, otherwise results are
579# undefined.")
580def atomic(name, flags=[]):
581    intrinsic(name + "_deref", src_comp=[-1], dest_comp=1, flags=flags)
582    intrinsic(name, src_comp=[1], dest_comp=1, indices=[BASE], flags=flags)
583
584def atomic2(name):
585    intrinsic(name + "_deref", src_comp=[-1, 1], dest_comp=1)
586    intrinsic(name, src_comp=[1, 1], dest_comp=1, indices=[BASE])
587
588def atomic3(name):
589    intrinsic(name + "_deref", src_comp=[-1, 1, 1], dest_comp=1)
590    intrinsic(name, src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
591
592atomic("atomic_counter_inc")
593atomic("atomic_counter_pre_dec")
594atomic("atomic_counter_post_dec")
595atomic("atomic_counter_read", flags=[CAN_ELIMINATE])
596atomic2("atomic_counter_add")
597atomic2("atomic_counter_min")
598atomic2("atomic_counter_max")
599atomic2("atomic_counter_and")
600atomic2("atomic_counter_or")
601atomic2("atomic_counter_xor")
602atomic2("atomic_counter_exchange")
603atomic3("atomic_counter_comp_swap")
604
605# Image load, store and atomic intrinsics.
606#
607# All image intrinsics come in three versions.  One which take an image target
608# passed as a deref chain as the first source, one which takes an index as the
609# first source, and one which takes a bindless handle as the first source.
610# In the first version, the image variable contains the memory and layout
611# qualifiers that influence the semantics of the intrinsic.  In the second and
612# third, the image format and access qualifiers are provided as constant
613# indices.  Up through GLSL ES 3.10, the image index source may only be a
614# constant array access.  GLSL ES 3.20 and GLSL 4.00 allow dynamically uniform
615# indexing.
616#
617# All image intrinsics take a four-coordinate vector and a sample index as
618# 2nd and 3rd sources, determining the location within the image that will be
619# accessed by the intrinsic.  Components not applicable to the image target
620# in use are undefined.  Image store takes an additional four-component
621# argument with the value to be written, and image atomic operations take
622# either one or two additional scalar arguments with the same meaning as in
623# the ARB_shader_image_load_store specification.
624def image(name, src_comp=[], extra_indices=[], **kwargs):
625    intrinsic("image_deref_" + name, src_comp=[-1] + src_comp,
626              indices=[IMAGE_DIM, IMAGE_ARRAY, FORMAT, ACCESS] + extra_indices, **kwargs)
627    intrinsic("image_" + name, src_comp=[1] + src_comp,
628              indices=[IMAGE_DIM, IMAGE_ARRAY, FORMAT, ACCESS] + extra_indices, **kwargs)
629    intrinsic("bindless_image_" + name, src_comp=[-1] + src_comp,
630              indices=[IMAGE_DIM, IMAGE_ARRAY, FORMAT, ACCESS] + extra_indices, **kwargs)
631
632image("load", src_comp=[4, 1, 1], extra_indices=[DEST_TYPE], dest_comp=0, flags=[CAN_ELIMINATE])
633image("sparse_load", src_comp=[4, 1, 1], extra_indices=[DEST_TYPE], dest_comp=0, flags=[CAN_ELIMINATE])
634image("store", src_comp=[4, 1, 0, 1], extra_indices=[SRC_TYPE])
635image("atomic_add",  src_comp=[4, 1, 1], dest_comp=1)
636image("atomic_imin",  src_comp=[4, 1, 1], dest_comp=1)
637image("atomic_umin",  src_comp=[4, 1, 1], dest_comp=1)
638image("atomic_imax",  src_comp=[4, 1, 1], dest_comp=1)
639image("atomic_umax",  src_comp=[4, 1, 1], dest_comp=1)
640image("atomic_and",  src_comp=[4, 1, 1], dest_comp=1)
641image("atomic_or",   src_comp=[4, 1, 1], dest_comp=1)
642image("atomic_xor",  src_comp=[4, 1, 1], dest_comp=1)
643image("atomic_exchange",  src_comp=[4, 1, 1], dest_comp=1)
644image("atomic_comp_swap", src_comp=[4, 1, 1, 1], dest_comp=1)
645image("atomic_fadd",  src_comp=[4, 1, 1], dest_comp=1)
646image("atomic_fmin",  src_comp=[4, 1, 1], dest_comp=1)
647image("atomic_fmax",  src_comp=[4, 1, 1], dest_comp=1)
648image("size",    dest_comp=0, src_comp=[1], flags=[CAN_ELIMINATE, CAN_REORDER])
649image("samples", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
650image("atomic_inc_wrap",  src_comp=[4, 1, 1], dest_comp=1)
651image("atomic_dec_wrap",  src_comp=[4, 1, 1], dest_comp=1)
652# CL-specific format queries
653image("format", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
654image("order", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
655
656# Vulkan descriptor set intrinsics
657#
658# The Vulkan API uses a different binding model from GL.  In the Vulkan
659# API, all external resources are represented by a tuple:
660#
661# (descriptor set, binding, array index)
662#
663# where the array index is the only thing allowed to be indirect.  The
664# vulkan_surface_index intrinsic takes the descriptor set and binding as
665# its first two indices and the array index as its source.  The third
666# index is a nir_variable_mode in case that's useful to the backend.
667#
668# The intended usage is that the shader will call vulkan_surface_index to
669# get an index and then pass that as the buffer index ubo/ssbo calls.
670#
671# The vulkan_resource_reindex intrinsic takes a resource index in src0
672# (the result of a vulkan_resource_index or vulkan_resource_reindex) which
673# corresponds to the tuple (set, binding, index) and computes an index
674# corresponding to tuple (set, binding, idx + src1).
675intrinsic("vulkan_resource_index", src_comp=[1], dest_comp=0,
676          indices=[DESC_SET, BINDING, DESC_TYPE],
677          flags=[CAN_ELIMINATE, CAN_REORDER])
678intrinsic("vulkan_resource_reindex", src_comp=[0, 1], dest_comp=0,
679          indices=[DESC_TYPE], flags=[CAN_ELIMINATE, CAN_REORDER])
680intrinsic("load_vulkan_descriptor", src_comp=[-1], dest_comp=0,
681          indices=[DESC_TYPE], flags=[CAN_ELIMINATE, CAN_REORDER])
682
683# atomic intrinsics
684#
685# All of these atomic memory operations read a value from memory, compute a new
686# value using one of the operations below, write the new value to memory, and
687# return the original value read.
688#
689# All variable operations take 2 sources except CompSwap that takes 3. These
690# sources represent:
691#
692# 0: A deref to the memory on which to perform the atomic
693# 1: The data parameter to the atomic function (i.e. the value to add
694#    in shared_atomic_add, etc).
695# 2: For CompSwap only: the second data parameter.
696#
697# All SSBO operations take 3 sources except CompSwap that takes 4. These
698# sources represent:
699#
700# 0: The SSBO buffer index (dynamically uniform in GLSL, possibly non-uniform
701#    with VK_EXT_descriptor_indexing).
702# 1: The offset into the SSBO buffer of the variable that the atomic
703#    operation will operate on.
704# 2: The data parameter to the atomic function (i.e. the value to add
705#    in ssbo_atomic_add, etc).
706# 3: For CompSwap only: the second data parameter.
707#
708# All shared (and task payload) variable operations take 2 sources
709# except CompSwap that takes 3.
710# These sources represent:
711#
712# 0: The offset into the shared variable storage region that the atomic
713#    operation will operate on.
714# 1: The data parameter to the atomic function (i.e. the value to add
715#    in shared_atomic_add, etc).
716# 2: For CompSwap only: the second data parameter.
717#
718# All global operations take 2 sources except CompSwap that takes 3. These
719# sources represent:
720#
721# 0: The memory address that the atomic operation will operate on.
722# 1: The data parameter to the atomic function (i.e. the value to add
723#    in shared_atomic_add, etc).
724# 2: For CompSwap only: the second data parameter.
725#
726# The 2x32 global variants use a vec2 for the memory address where component X
727# has the low 32-bit and component Y has the high 32-bit.
728#
729# IR3 global operations take 32b vec2 as memory address. IR3 doesn't support
730# float atomics.
731
732def memory_atomic_data1(name):
733    intrinsic("deref_atomic_" + name,  src_comp=[-1, 1], dest_comp=1, indices=[ACCESS])
734    intrinsic("ssbo_atomic_" + name,  src_comp=[-1, 1, 1], dest_comp=1, indices=[ACCESS])
735    intrinsic("shared_atomic_" + name,  src_comp=[1, 1], dest_comp=1, indices=[BASE])
736    intrinsic("task_payload_atomic_" + name,  src_comp=[1, 1], dest_comp=1, indices=[BASE])
737    intrinsic("global_atomic_" + name,  src_comp=[1, 1], dest_comp=1, indices=[])
738    intrinsic("global_atomic_" + name + "_2x32",  src_comp=[2, 1], dest_comp=1, indices=[])
739    intrinsic("global_atomic_" + name + "_amd",  src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
740    if not name.startswith('f'):
741        intrinsic("global_atomic_" + name + "_ir3",  src_comp=[2, 1], dest_comp=1, indices=[BASE])
742
743def memory_atomic_data2(name):
744    intrinsic("deref_atomic_" + name,  src_comp=[-1, 1, 1], dest_comp=1, indices=[ACCESS])
745    intrinsic("ssbo_atomic_" + name,  src_comp=[-1, 1, 1, 1], dest_comp=1, indices=[ACCESS])
746    intrinsic("shared_atomic_" + name,  src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
747    intrinsic("task_payload_atomic_" + name,  src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
748    intrinsic("global_atomic_" + name,  src_comp=[1, 1, 1], dest_comp=1, indices=[])
749    intrinsic("global_atomic_" + name + "_2x32",  src_comp=[2, 1, 1], dest_comp=1, indices=[])
750    intrinsic("global_atomic_" + name + "_amd",  src_comp=[1, 1, 1, 1], dest_comp=1, indices=[BASE])
751    if not name.startswith('f'):
752        intrinsic("global_atomic_" + name + "_ir3",  src_comp=[2, 1, 1], dest_comp=1, indices=[BASE])
753
754memory_atomic_data1("add")
755memory_atomic_data1("imin")
756memory_atomic_data1("umin")
757memory_atomic_data1("imax")
758memory_atomic_data1("umax")
759memory_atomic_data1("and")
760memory_atomic_data1("or")
761memory_atomic_data1("xor")
762memory_atomic_data1("exchange")
763memory_atomic_data1("fadd")
764memory_atomic_data1("fmin")
765memory_atomic_data1("fmax")
766memory_atomic_data2("comp_swap")
767memory_atomic_data2("fcomp_swap")
768
769def system_value(name, dest_comp, indices=[], bit_sizes=[32]):
770    intrinsic("load_" + name, [], dest_comp, indices,
771              flags=[CAN_ELIMINATE, CAN_REORDER], sysval=True,
772              bit_sizes=bit_sizes)
773
774system_value("frag_coord", 4)
775system_value("point_coord", 2)
776system_value("line_coord", 1)
777system_value("front_face", 1, bit_sizes=[1, 32])
778system_value("vertex_id", 1)
779system_value("vertex_id_zero_base", 1)
780system_value("first_vertex", 1)
781system_value("is_indexed_draw", 1)
782system_value("base_vertex", 1)
783system_value("instance_id", 1)
784system_value("base_instance", 1)
785system_value("draw_id", 1)
786system_value("sample_id", 1)
787# sample_id_no_per_sample is like sample_id but does not imply per-
788# sample shading.  See the lower_helper_invocation option.
789system_value("sample_id_no_per_sample", 1)
790system_value("sample_pos", 2)
791# sample_pos_or_center is like sample_pos but does not imply per-sample
792# shading.  When per-sample dispatch is not enabled, it returns (0.5, 0.5).
793system_value("sample_pos_or_center", 2)
794system_value("sample_mask_in", 1)
795system_value("primitive_id", 1)
796system_value("invocation_id", 1)
797system_value("tess_coord", 3)
798system_value("tess_level_outer", 4)
799system_value("tess_level_inner", 2)
800system_value("tess_level_outer_default", 4)
801system_value("tess_level_inner_default", 2)
802system_value("patch_vertices_in", 1)
803system_value("local_invocation_id", 3)
804system_value("local_invocation_index", 1)
805# zero_base indicates it starts from 0 for the current dispatch
806# non-zero_base indicates the base is included
807system_value("workgroup_id", 3, bit_sizes=[32, 64])
808system_value("workgroup_id_zero_base", 3)
809# The workgroup_index is intended for situations when a 3 dimensional
810# workgroup_id is not available on the HW, but a 1 dimensional index is.
811system_value("workgroup_index", 1)
812system_value("base_workgroup_id", 3, bit_sizes=[32, 64])
813system_value("user_clip_plane", 4, indices=[UCP_ID])
814system_value("num_workgroups", 3, bit_sizes=[32, 64])
815system_value("num_vertices", 1)
816system_value("helper_invocation", 1, bit_sizes=[1, 32])
817system_value("layer_id", 1)
818system_value("view_index", 1)
819system_value("subgroup_size", 1)
820system_value("subgroup_invocation", 1)
821system_value("subgroup_eq_mask", 0, bit_sizes=[32, 64])
822system_value("subgroup_ge_mask", 0, bit_sizes=[32, 64])
823system_value("subgroup_gt_mask", 0, bit_sizes=[32, 64])
824system_value("subgroup_le_mask", 0, bit_sizes=[32, 64])
825system_value("subgroup_lt_mask", 0, bit_sizes=[32, 64])
826system_value("num_subgroups", 1)
827system_value("subgroup_id", 1)
828system_value("workgroup_size", 3)
829# note: the definition of global_invocation_id_zero_base is based on
830# (workgroup_id * workgroup_size) + local_invocation_id.
831# it is *not* based on workgroup_id_zero_base, meaning the work group
832# base is already accounted for, and the global base is additive on top of that
833system_value("global_invocation_id", 3, bit_sizes=[32, 64])
834system_value("global_invocation_id_zero_base", 3, bit_sizes=[32, 64])
835system_value("base_global_invocation_id", 3, bit_sizes=[32, 64])
836system_value("global_invocation_index", 1, bit_sizes=[32, 64])
837system_value("work_dim", 1)
838system_value("line_width", 1)
839system_value("aa_line_width", 1)
840# BASE=0 for global/shader, BASE=1 for local/function
841system_value("scratch_base_ptr", 0, bit_sizes=[32,64], indices=[BASE])
842system_value("constant_base_ptr", 0, bit_sizes=[32,64])
843system_value("shared_base_ptr", 0, bit_sizes=[32,64])
844system_value("global_base_ptr", 0, bit_sizes=[32,64])
845# Address of a transform feedback buffer, indexed by BASE
846system_value("xfb_address", 1, bit_sizes=[32,64], indices=[BASE])
847
848# System values for ray tracing.
849system_value("ray_launch_id", 3)
850system_value("ray_launch_size", 3)
851system_value("ray_world_origin", 3)
852system_value("ray_world_direction", 3)
853system_value("ray_object_origin", 3)
854system_value("ray_object_direction", 3)
855system_value("ray_t_min", 1)
856system_value("ray_t_max", 1)
857system_value("ray_object_to_world", 3, indices=[COLUMN])
858system_value("ray_world_to_object", 3, indices=[COLUMN])
859system_value("ray_hit_kind", 1)
860system_value("ray_flags", 1)
861system_value("ray_geometry_index", 1)
862system_value("ray_instance_custom_index", 1)
863system_value("shader_record_ptr", 1, bit_sizes=[64])
864system_value("cull_mask", 1)
865
866# Driver-specific viewport scale/offset parameters.
867#
868# VC4 and V3D need to emit a scaled version of the position in the vertex
869# shaders for binning, and having system values lets us move the math for that
870# into NIR.
871#
872# Panfrost needs to implement all coordinate transformation in the
873# vertex shader; system values allow us to share this routine in NIR.
874#
875# RADV uses these for NGG primitive culling.
876system_value("viewport_x_scale", 1)
877system_value("viewport_y_scale", 1)
878system_value("viewport_z_scale", 1)
879system_value("viewport_x_offset", 1)
880system_value("viewport_y_offset", 1)
881system_value("viewport_z_offset", 1)
882system_value("viewport_scale", 3)
883system_value("viewport_offset", 3)
884
885# Blend constant color values.  Float values are clamped. Vectored versions are
886# provided as well for driver convenience
887
888system_value("blend_const_color_r_float", 1)
889system_value("blend_const_color_g_float", 1)
890system_value("blend_const_color_b_float", 1)
891system_value("blend_const_color_a_float", 1)
892system_value("blend_const_color_rgba", 4)
893system_value("blend_const_color_rgba8888_unorm", 1)
894system_value("blend_const_color_aaaa8888_unorm", 1)
895
896# System values for gl_Color, for radeonsi which interpolates these in the
897# shader prolog to handle two-sided color without recompiles and therefore
898# doesn't handle these in the main shader part like normal varyings.
899system_value("color0", 4)
900system_value("color1", 4)
901
902# System value for internal compute shaders in radeonsi.
903system_value("user_data_amd", 4)
904
905# Barycentric coordinate intrinsics.
906#
907# These set up the barycentric coordinates for a particular interpolation.
908# The first four are for the simple cases: pixel, centroid, per-sample
909# (at gl_SampleID), or pull model (1/W, 1/I, 1/J) at the pixel center. The next
910# two handle interpolating at a specified sample location, or interpolating
911# with a vec2 offset,
912#
913# The interp_mode index should be either the INTERP_MODE_SMOOTH or
914# INTERP_MODE_NOPERSPECTIVE enum values.
915#
916# The vec2 value produced by these intrinsics is intended for use as the
917# barycoord source of a load_interpolated_input intrinsic.
918
919def barycentric(name, dst_comp, src_comp=[]):
920    intrinsic("load_barycentric_" + name, src_comp=src_comp, dest_comp=dst_comp,
921              indices=[INTERP_MODE], flags=[CAN_ELIMINATE, CAN_REORDER])
922
923# no sources.
924barycentric("pixel", 2)
925barycentric("centroid", 2)
926barycentric("sample", 2)
927barycentric("model", 3)
928# src[] = { sample_id }.
929barycentric("at_sample", 2, [1])
930# src[] = { offset.xy }.
931barycentric("at_offset", 2, [2])
932
933# Load sample position:
934#
935# Takes a sample # and returns a sample position.  Used for lowering
936# interpolateAtSample() to interpolateAtOffset()
937intrinsic("load_sample_pos_from_id", src_comp=[1], dest_comp=2,
938          flags=[CAN_ELIMINATE, CAN_REORDER])
939
940intrinsic("load_persp_center_rhw_ir3", dest_comp=1,
941          flags=[CAN_ELIMINATE, CAN_REORDER])
942
943# Load texture scaling values:
944#
945# Takes a sampler # and returns 1/size values for multiplying to normalize
946# texture coordinates.  Used for lowering rect textures.
947intrinsic("load_texture_rect_scaling", src_comp=[1], dest_comp=2,
948          flags=[CAN_ELIMINATE, CAN_REORDER])
949
950# Fragment shader input interpolation delta intrinsic.
951#
952# For hw where fragment shader input interpolation is handled in shader, the
953# load_fs_input_interp deltas intrinsics can be used to load the input deltas
954# used for interpolation as follows:
955#
956#    vec3 iid = load_fs_input_interp_deltas(varying_slot)
957#    vec2 bary = load_barycentric_*(...)
958#    float result = iid.x + iid.y * bary.y + iid.z * bary.x
959
960intrinsic("load_fs_input_interp_deltas", src_comp=[1], dest_comp=3,
961          indices=[BASE, COMPONENT, IO_SEMANTICS], flags=[CAN_ELIMINATE, CAN_REORDER])
962
963# Load operations pull data from some piece of GPU memory.  All load
964# operations operate in terms of offsets into some piece of theoretical
965# memory.  Loads from externally visible memory (UBO and SSBO) simply take a
966# byte offset as a source.  Loads from opaque memory (uniforms, inputs, etc.)
967# take a base+offset pair where the nir_intrinsic_base() gives the location
968# of the start of the variable being loaded and and the offset source is a
969# offset into that variable.
970#
971# Uniform load operations have a nir_intrinsic_range() index that specifies the
972# range (starting at base) of the data from which we are loading.  If
973# range == 0, then the range is unknown.
974#
975# UBO load operations have a nir_intrinsic_range_base() and
976# nir_intrinsic_range() that specify the byte range [range_base,
977# range_base+range] of the UBO that the src offset access must lie within.
978#
979# Some load operations such as UBO/SSBO load and per_vertex loads take an
980# additional source to specify which UBO/SSBO/vertex to load from.
981#
982# The exact address type depends on the lowering pass that generates the
983# load/store intrinsics.  Typically, this is vec4 units for things such as
984# varying slots and float units for fragment shader inputs.  UBO and SSBO
985# offsets are always in bytes.
986
987def load(name, src_comp, indices=[], flags=[]):
988    intrinsic("load_" + name, src_comp, dest_comp=0, indices=indices,
989              flags=flags)
990
991# src[] = { offset }.
992load("uniform", [1], [BASE, RANGE, DEST_TYPE], [CAN_ELIMINATE, CAN_REORDER])
993# src[] = { buffer_index, offset }.
994load("ubo", [-1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET, RANGE_BASE, RANGE], flags=[CAN_ELIMINATE, CAN_REORDER])
995# src[] = { buffer_index, offset in vec4 units }.  base is also in vec4 units.
996load("ubo_vec4", [-1, 1], [ACCESS, BASE, COMPONENT], flags=[CAN_ELIMINATE, CAN_REORDER])
997# src[] = { offset }.
998load("input", [1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
999# src[] = { vertex_id, offset }.
1000load("input_vertex", [1, 1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
1001# src[] = { vertex, offset }.
1002load("per_vertex_input", [1, 1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
1003# src[] = { barycoord, offset }.
1004load("interpolated_input", [2, 1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
1005
1006# src[] = { buffer_index, offset }.
1007load("ssbo", [-1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1008# src[] = { buffer_index }
1009load("ssbo_address", [1], [], [CAN_ELIMINATE, CAN_REORDER])
1010# src[] = { offset }.
1011load("output", [1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], flags=[CAN_ELIMINATE])
1012# src[] = { vertex, offset }.
1013load("per_vertex_output", [1, 1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE])
1014# src[] = { primitive, offset }.
1015load("per_primitive_output", [1, 1], [BASE, COMPONENT, DEST_TYPE, IO_SEMANTICS], [CAN_ELIMINATE])
1016# src[] = { offset }.
1017load("shared", [1], [BASE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1018# src[] = { offset }.
1019load("task_payload", [1], [BASE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1020# src[] = { offset }.
1021load("push_constant", [1], [BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
1022# src[] = { offset }.
1023load("constant", [1], [BASE, RANGE, ALIGN_MUL, ALIGN_OFFSET],
1024     [CAN_ELIMINATE, CAN_REORDER])
1025# src[] = { address }.
1026load("global", [1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1027# src[] = { address }.
1028load("global_2x32", [2], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1029# src[] = { address }.
1030load("global_constant", [1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET],
1031     [CAN_ELIMINATE, CAN_REORDER])
1032# src[] = { base_address, offset }.
1033load("global_constant_offset", [1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET],
1034     [CAN_ELIMINATE, CAN_REORDER])
1035# src[] = { base_address, offset, bound }.
1036load("global_constant_bounded", [1, 1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET],
1037     [CAN_ELIMINATE, CAN_REORDER])
1038# src[] = { address }.
1039load("kernel_input", [1], [BASE, RANGE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE, CAN_REORDER])
1040# src[] = { offset }.
1041load("scratch", [1], [ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1042
1043# Stores work the same way as loads, except now the first source is the value
1044# to store and the second (and possibly third) source specify where to store
1045# the value.  SSBO and shared memory stores also have a
1046# nir_intrinsic_write_mask()
1047
1048def store(name, srcs, indices=[], flags=[]):
1049    intrinsic("store_" + name, [0] + srcs, indices=indices, flags=flags)
1050
1051# src[] = { value, offset }.
1052store("output", [1], [BASE, WRITE_MASK, COMPONENT, SRC_TYPE, IO_SEMANTICS, IO_XFB, IO_XFB2])
1053# src[] = { value, vertex, offset }.
1054store("per_vertex_output", [1, 1], [BASE, WRITE_MASK, COMPONENT, SRC_TYPE, IO_SEMANTICS])
1055# src[] = { value, primitive, offset }.
1056store("per_primitive_output", [1, 1], [BASE, WRITE_MASK, COMPONENT, SRC_TYPE, IO_SEMANTICS])
1057# src[] = { value, block_index, offset }
1058store("ssbo", [-1, 1], [WRITE_MASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
1059# src[] = { value, offset }.
1060store("shared", [1], [BASE, WRITE_MASK, ALIGN_MUL, ALIGN_OFFSET])
1061# src[] = { value, offset }.
1062store("task_payload", [1], [BASE, WRITE_MASK, ALIGN_MUL, ALIGN_OFFSET])
1063# src[] = { value, address }.
1064store("global", [1], [WRITE_MASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
1065# src[] = { value, address }.
1066store("global_2x32", [2], [WRITE_MASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
1067# src[] = { value, offset }.
1068store("scratch", [1], [ALIGN_MUL, ALIGN_OFFSET, WRITE_MASK])
1069
1070# A bit field to implement SPIRV FragmentShadingRateKHR
1071# bit | name              | description
1072#   0 | Vertical2Pixels   | Fragment invocation covers 2 pixels vertically
1073#   1 | Vertical4Pixels   | Fragment invocation covers 4 pixels vertically
1074#   2 | Horizontal2Pixels | Fragment invocation covers 2 pixels horizontally
1075#   3 | Horizontal4Pixels | Fragment invocation covers 4 pixels horizontally
1076intrinsic("load_frag_shading_rate", dest_comp=1, bit_sizes=[32],
1077          flags=[CAN_ELIMINATE, CAN_REORDER])
1078
1079# OpenCL printf instruction
1080# First source is a deref to the format string
1081# Second source is a deref to a struct containing the args
1082# Dest is success or failure
1083intrinsic("printf", src_comp=[1, 1], dest_comp=1, bit_sizes=[32])
1084# Since most drivers will want to lower to just dumping args
1085# in a buffer, nir_lower_printf will do that, but requires
1086# the driver to at least provide a base location
1087system_value("printf_buffer_address", 1, bit_sizes=[32,64])
1088
1089# Mesh shading MultiView intrinsics
1090system_value("mesh_view_count", 1)
1091load("mesh_view_indices", [1], [BASE, RANGE], [CAN_ELIMINATE, CAN_REORDER])
1092
1093# Used to pass values from the preamble to the main shader.
1094# This should use something similar to Vulkan push constants and load_preamble
1095# should be relatively cheap.
1096# For now we only support accesses with a constant offset.
1097load("preamble", [], indices=[BASE], flags=[CAN_ELIMINATE, CAN_REORDER])
1098store("preamble", [], indices=[BASE])
1099
1100# IR3-specific version of most SSBO intrinsics. The only different
1101# compare to the originals is that they add an extra source to hold
1102# the dword-offset, which is needed by the backend code apart from
1103# the byte-offset already provided by NIR in one of the sources.
1104#
1105# NIR lowering pass 'ir3_nir_lower_io_offset' will replace the
1106# original SSBO intrinsics by these, placing the computed
1107# dword-offset always in the last source.
1108#
1109# The float versions are not handled because those are not supported
1110# by the backend.
1111store("ssbo_ir3", [1, 1, 1],
1112      indices=[WRITE_MASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
1113load("ssbo_ir3",  [1, 1, 1],
1114     indices=[ACCESS, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN_ELIMINATE])
1115intrinsic("ssbo_atomic_add_ir3",        src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
1116intrinsic("ssbo_atomic_imin_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
1117intrinsic("ssbo_atomic_umin_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
1118intrinsic("ssbo_atomic_imax_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
1119intrinsic("ssbo_atomic_umax_ir3",       src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
1120intrinsic("ssbo_atomic_and_ir3",        src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
1121intrinsic("ssbo_atomic_or_ir3",         src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
1122intrinsic("ssbo_atomic_xor_ir3",        src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
1123intrinsic("ssbo_atomic_exchange_ir3",   src_comp=[1, 1, 1, 1],    dest_comp=1, indices=[ACCESS])
1124intrinsic("ssbo_atomic_comp_swap_ir3",  src_comp=[1, 1, 1, 1, 1], dest_comp=1, indices=[ACCESS])
1125
1126# System values for freedreno geometry shaders.
1127system_value("vs_primitive_stride_ir3", 1)
1128system_value("vs_vertex_stride_ir3", 1)
1129system_value("gs_header_ir3", 1)
1130system_value("primitive_location_ir3", 1, indices=[DRIVER_LOCATION])
1131
1132# System values for freedreno tessellation shaders.
1133system_value("hs_patch_stride_ir3", 1)
1134system_value("tess_factor_base_ir3", 2)
1135system_value("tess_param_base_ir3", 2)
1136system_value("tcs_header_ir3", 1)
1137system_value("rel_patch_id_ir3", 1)
1138
1139# System values for freedreno compute shaders.
1140system_value("subgroup_id_shift_ir3", 1)
1141
1142# IR3-specific intrinsics for tessellation control shaders.  cond_end_ir3 end
1143# the shader when src0 is false and is used to narrow down the TCS shader to
1144# just thread 0 before writing out tessellation levels.
1145intrinsic("cond_end_ir3", src_comp=[1])
1146# end_patch_ir3 is used just before thread 0 exist the TCS and presumably
1147# signals the TE that the patch is complete and can be tessellated.
1148intrinsic("end_patch_ir3")
1149
1150# IR3-specific load/store intrinsics. These access a buffer used to pass data
1151# between geometry stages - perhaps it's explicit access to the vertex cache.
1152
1153# src[] = { value, offset }.
1154store("shared_ir3", [1], [BASE, ALIGN_MUL, ALIGN_OFFSET])
1155# src[] = { offset }.
1156load("shared_ir3", [1], [BASE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1157
1158# IR3-specific load/store global intrinsics. They take a 64-bit base address
1159# and a 32-bit offset.  The hardware will add the base and the offset, which
1160# saves us from doing 64-bit math on the base address.
1161
1162# src[] = { value, address(vec2 of hi+lo uint32_t), offset }.
1163# const_index[] = { write_mask, align_mul, align_offset }
1164store("global_ir3", [2, 1], indices=[ACCESS, ALIGN_MUL, ALIGN_OFFSET])
1165# src[] = { address(vec2 of hi+lo uint32_t), offset }.
1166# const_index[] = { access, align_mul, align_offset }
1167load("global_ir3", [2, 1], indices=[ACCESS, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN_ELIMINATE])
1168
1169# IR3-specific bindless handle specifier. Similar to vulkan_resource_index, but
1170# without the binding because the hardware expects a single flattened index
1171# rather than a (binding, index) pair. We may also want to use this with GL.
1172# Note that this doesn't actually turn into a HW instruction.
1173intrinsic("bindless_resource_ir3", [1], dest_comp=1, indices=[DESC_SET], flags=[CAN_ELIMINATE, CAN_REORDER])
1174
1175# IR3-specific intrinsics for shader preamble. These are meant to be used like
1176# this:
1177#
1178# if (preamble_start()) {
1179#    if (subgroupElect()) {
1180#       // preamble
1181#       ...
1182#       preamble_end();
1183#    }
1184# }
1185# // main shader
1186# ...
1187
1188intrinsic("preamble_start_ir3", [], dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
1189
1190barrier("preamble_end_ir3")
1191
1192# IR3-specific intrinsic for stc. Should be used in the shader preamble.
1193store("uniform_ir3", [], indices=[BASE])
1194
1195# IR3-specific intrinsic for ldc.k. Copies UBO to constant file.
1196# base is the const file base in components, range is the amount to copy in
1197# vec4's.
1198intrinsic("copy_ubo_to_uniform_ir3", [1, 1], indices=[BASE, RANGE])
1199
1200# DXIL specific intrinsics
1201# src[] = { value, mask, index, offset }.
1202intrinsic("store_ssbo_masked_dxil", [1, 1, 1, 1])
1203# src[] = { value, index }.
1204intrinsic("store_shared_dxil", [1, 1])
1205# src[] = { value, mask, index }.
1206intrinsic("store_shared_masked_dxil", [1, 1, 1])
1207# src[] = { value, index }.
1208intrinsic("store_scratch_dxil", [1, 1])
1209# src[] = { index }.
1210load("shared_dxil", [1], [], [CAN_ELIMINATE])
1211# src[] = { index }.
1212load("scratch_dxil", [1], [], [CAN_ELIMINATE])
1213# src[] = { deref_var, offset }
1214load("ptr_dxil", [1, 1], [], [])
1215# src[] = { index, 16-byte-based-offset }
1216load("ubo_dxil", [1, 1], [], [CAN_ELIMINATE, CAN_REORDER])
1217
1218# DXIL Shared atomic intrinsics
1219#
1220# All of the shared variable atomic memory operations read a value from
1221# memory, compute a new value using one of the operations below, write the
1222# new value to memory, and return the original value read.
1223#
1224# All operations take 2 sources:
1225#
1226# 0: The index in the i32 array for by the shared memory region
1227# 1: The data parameter to the atomic function (i.e. the value to add
1228#    in shared_atomic_add, etc).
1229intrinsic("shared_atomic_add_dxil",  src_comp=[1, 1], dest_comp=1)
1230intrinsic("shared_atomic_imin_dxil", src_comp=[1, 1], dest_comp=1)
1231intrinsic("shared_atomic_umin_dxil", src_comp=[1, 1], dest_comp=1)
1232intrinsic("shared_atomic_imax_dxil", src_comp=[1, 1], dest_comp=1)
1233intrinsic("shared_atomic_umax_dxil", src_comp=[1, 1], dest_comp=1)
1234intrinsic("shared_atomic_and_dxil",  src_comp=[1, 1], dest_comp=1)
1235intrinsic("shared_atomic_or_dxil",   src_comp=[1, 1], dest_comp=1)
1236intrinsic("shared_atomic_xor_dxil",  src_comp=[1, 1], dest_comp=1)
1237intrinsic("shared_atomic_exchange_dxil", src_comp=[1, 1], dest_comp=1)
1238intrinsic("shared_atomic_comp_swap_dxil", src_comp=[1, 1, 1], dest_comp=1)
1239
1240# Intrinsics used by the Midgard/Bifrost blend pipeline. These are defined
1241# within a blend shader to read/write the raw value from the tile buffer,
1242# without applying any format conversion in the process. If the shader needs
1243# usable pixel values, it must apply format conversions itself.
1244#
1245# These definitions are generic, but they are explicitly vendored to prevent
1246# other drivers from using them, as their semantics is defined in terms of the
1247# Midgard/Bifrost hardware tile buffer and may not line up with anything sane.
1248# One notable divergence is sRGB, which is asymmetric: raw_input_pan requires
1249# an sRGB->linear conversion, but linear values should be written to
1250# raw_output_pan and the hardware handles linear->sRGB.
1251
1252# src[] = { value }
1253store("raw_output_pan", [], [])
1254store("combined_output_pan", [1, 1, 1, 4], [BASE, COMPONENT, SRC_TYPE, DEST_TYPE])
1255load("raw_output_pan", [1], [BASE], [CAN_ELIMINATE, CAN_REORDER])
1256
1257# Loads the sampler paramaters <min_lod, max_lod, lod_bias>
1258# src[] = { sampler_index }
1259load("sampler_lod_parameters_pan", [1], flags=[CAN_ELIMINATE, CAN_REORDER])
1260
1261# Loads the sample position array on Bifrost, in a packed Arm-specific format
1262system_value("sample_positions_pan", 1, bit_sizes=[64])
1263
1264# R600 specific instrincs
1265#
1266# location where the tesselation data is stored in LDS
1267system_value("tcs_in_param_base_r600", 4)
1268system_value("tcs_out_param_base_r600", 4)
1269system_value("tcs_rel_patch_id_r600", 1)
1270system_value("tcs_tess_factor_base_r600", 1)
1271
1272# the tess coords come as xy only, z has to be calculated
1273system_value("tess_coord_r600", 2)
1274
1275# load as many components as needed giving per-component addresses
1276intrinsic("load_local_shared_r600", src_comp=[0], dest_comp=0, indices = [], flags = [CAN_ELIMINATE])
1277
1278store("local_shared_r600", [1], [WRITE_MASK])
1279store("tf_r600", [])
1280
1281# AMD GCN/RDNA specific intrinsics
1282
1283# src[] = { descriptor, base address, scalar offset }
1284intrinsic("load_buffer_amd", src_comp=[4, 1, 1], dest_comp=0, indices=[BASE, IS_SWIZZLED, SLC_AMD, MEMORY_MODES], flags=[CAN_ELIMINATE])
1285# src[] = { store value, descriptor, base address, scalar offset }
1286intrinsic("store_buffer_amd", src_comp=[0, 4, 1, 1], indices=[BASE, WRITE_MASK, IS_SWIZZLED, SLC_AMD, MEMORY_MODES])
1287
1288# src[] = { address, unsigned 32-bit offset }.
1289load("global_amd", [1, 1], indices=[BASE, ACCESS, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN_ELIMINATE])
1290# src[] = { value, address, unsigned 32-bit offset }.
1291store("global_amd", [1, 1], indices=[BASE, ACCESS, ALIGN_MUL, ALIGN_OFFSET, WRITE_MASK])
1292
1293# Same as shared_atomic_add, but with GDS. src[] = {store_val, gds_addr, m0}
1294intrinsic("gds_atomic_add_amd",  src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
1295
1296# Descriptor where TCS outputs are stored for TES
1297system_value("ring_tess_offchip_amd", 4)
1298system_value("ring_tess_offchip_offset_amd", 1)
1299# Descriptor where TCS outputs are stored for the HW tessellator
1300system_value("ring_tess_factors_amd", 4)
1301system_value("ring_tess_factors_offset_amd", 1)
1302# Descriptor where ES outputs are stored for GS to read on GFX6-8
1303system_value("ring_esgs_amd", 4)
1304system_value("ring_es2gs_offset_amd", 1)
1305# Address of the task shader draw ring (used for VARYING_SLOT_TASK_COUNT)
1306system_value("ring_task_draw_amd", 4)
1307# Address of the task shader payload ring (used for all other outputs)
1308system_value("ring_task_payload_amd", 4)
1309# Address of the mesh shader scratch ring (used for excess mesh shader outputs)
1310system_value("ring_mesh_scratch_amd", 4)
1311system_value("ring_mesh_scratch_offset_amd", 1)
1312# Pointer into the draw and payload rings
1313system_value("task_ring_entry_amd", 1)
1314# Pointer into the draw and payload rings
1315system_value("task_ib_addr", 2)
1316system_value("task_ib_stride", 1)
1317
1318# Number of patches processed by each TCS workgroup
1319system_value("tcs_num_patches_amd", 1)
1320# Relative tessellation patch ID within the current workgroup
1321system_value("tess_rel_patch_id_amd", 1)
1322# Vertex offsets used for GS per-vertex inputs
1323system_value("gs_vertex_offset_amd", 1, [BASE])
1324
1325# AMD merged shader intrinsics
1326
1327# Whether the current invocation has an input vertex / primitive to process (also known as "ES thread" or "GS thread").
1328# Not safe to reorder because it changes after overwrite_subgroup_num_vertices_and_primitives_amd.
1329# Also, the generated code is more optimal if they are not CSE'd.
1330intrinsic("has_input_vertex_amd", src_comp=[], dest_comp=1, bit_sizes=[1], indices=[])
1331intrinsic("has_input_primitive_amd", src_comp=[], dest_comp=1, bit_sizes=[1], indices=[])
1332
1333# AMD NGG intrinsics
1334
1335# Number of initial input vertices in the current workgroup.
1336system_value("workgroup_num_input_vertices_amd", 1)
1337# Number of initial input primitives in the current workgroup.
1338system_value("workgroup_num_input_primitives_amd", 1)
1339# For NGG passthrough mode only. Pre-packed argument for export_primitive_amd.
1340system_value("packed_passthrough_primitive_amd", 1)
1341# Whether NGG GS should execute shader query.
1342system_value("shader_query_enabled_amd", dest_comp=1, bit_sizes=[1])
1343# Whether the shader should cull front facing triangles.
1344intrinsic("load_cull_front_face_enabled_amd", dest_comp=1, bit_sizes=[1], flags=[CAN_ELIMINATE])
1345# Whether the shader should cull back facing triangles.
1346intrinsic("load_cull_back_face_enabled_amd", dest_comp=1, bit_sizes=[1], flags=[CAN_ELIMINATE])
1347# True if face culling should use CCW (false if CW).
1348intrinsic("load_cull_ccw_amd", dest_comp=1, bit_sizes=[1], flags=[CAN_ELIMINATE])
1349# Whether the shader should cull small primitives that are not visible in a pixel.
1350intrinsic("load_cull_small_primitives_enabled_amd", dest_comp=1, bit_sizes=[1], flags=[CAN_ELIMINATE])
1351# Whether any culling setting is enabled in the shader.
1352intrinsic("load_cull_any_enabled_amd", dest_comp=1, bit_sizes=[1], flags=[CAN_ELIMINATE])
1353# Small primitive culling precision
1354intrinsic("load_cull_small_prim_precision_amd", dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE, CAN_REORDER])
1355# Initial edge flags in a Vertex Shader, packed into the format the HW needs for primitive export.
1356intrinsic("load_initial_edgeflags_amd", src_comp=[], dest_comp=1, bit_sizes=[32], indices=[])
1357# Exports the current invocation's vertex. This is a placeholder where all vertex attribute export instructions should be emitted.
1358intrinsic("export_vertex_amd", src_comp=[], indices=[])
1359# Exports the current invocation's primitive. src[] = {packed_primitive_data}.
1360intrinsic("export_primitive_amd", src_comp=[1], indices=[])
1361# Allocates export space for vertices and primitives. src[] = {num_vertices, num_primitives}.
1362intrinsic("alloc_vertices_and_primitives_amd", src_comp=[1, 1], indices=[])
1363# Overwrites VS input registers, for use with vertex compaction after culling. src = {vertex_id, instance_id}.
1364intrinsic("overwrite_vs_arguments_amd", src_comp=[1, 1], indices=[])
1365# Overwrites TES input registers, for use with vertex compaction after culling. src = {tes_u, tes_v, rel_patch_id, patch_id}.
1366intrinsic("overwrite_tes_arguments_amd", src_comp=[1, 1, 1, 1], indices=[])
1367
1368# The address of the sbt descriptors.
1369system_value("sbt_base_amd", 1, bit_sizes=[64])
1370
1371# 1. HW descriptor
1372# 2. BVH node(64-bit pointer as 2x32 ...)
1373# 3. ray extent
1374# 4. ray origin
1375# 5. ray direction
1376# 6. inverse ray direction (componentwise 1.0/ray direction)
1377intrinsic("bvh64_intersect_ray_amd", [4, 2, 1, 3, 3, 3], 4, flags=[CAN_ELIMINATE, CAN_REORDER])
1378
1379# Return of a callable in raytracing pipelines
1380intrinsic("rt_return_amd")
1381
1382# offset into scratch for the input callable data in a raytracing pipeline.
1383system_value("rt_arg_scratch_offset_amd", 1)
1384
1385# Whether to call the anyhit shader for an intersection in an intersection shader.
1386system_value("intersection_opaque_amd", 1, bit_sizes=[1])
1387
1388# Used for indirect ray tracing.
1389system_value("ray_launch_size_addr_amd", 1, bit_sizes=[64])
1390
1391# Load forced VRS rates.
1392intrinsic("load_force_vrs_rates_amd", dest_comp=1, bit_sizes=[32], flags=[CAN_ELIMINATE, CAN_REORDER])
1393
1394intrinsic("load_scalar_arg_amd", dest_comp=0, bit_sizes=[32], indices=[BASE, ARG_UPPER_BOUND_U32_AMD], flags=[CAN_ELIMINATE, CAN_REORDER])
1395intrinsic("load_vector_arg_amd", dest_comp=0, bit_sizes=[32], indices=[BASE, ARG_UPPER_BOUND_U32_AMD], flags=[CAN_ELIMINATE, CAN_REORDER])
1396
1397# src[] = { 64-bit base address, 32-bit offset }.
1398intrinsic("load_smem_amd", src_comp=[1, 1], dest_comp=0, bit_sizes=[32],
1399                           indices=[ALIGN_MUL, ALIGN_OFFSET],
1400                           flags=[CAN_ELIMINATE, CAN_REORDER])
1401
1402# src[] = { offset }.
1403intrinsic("load_shared2_amd", [1], dest_comp=2, indices=[OFFSET0, OFFSET1, ST64], flags=[CAN_ELIMINATE])
1404
1405# src[] = { value, offset }.
1406intrinsic("store_shared2_amd", [2, 1], indices=[OFFSET0, OFFSET1, ST64])
1407
1408# Vertex stride in LS-HS buffer
1409system_value("lshs_vertex_stride_amd", 1)
1410
1411# Per patch data offset in HS VRAM output buffer
1412system_value("hs_out_patch_data_offset_amd", 1)
1413
1414# V3D-specific instrinc for tile buffer color reads.
1415#
1416# The hardware requires that we read the samples and components of a pixel
1417# in order, so we cannot eliminate or remove any loads in a sequence.
1418#
1419# src[] = { render_target }
1420# BASE = sample index
1421load("tlb_color_v3d", [1], [BASE, COMPONENT], [])
1422
1423# V3D-specific instrinc for per-sample tile buffer color writes.
1424#
1425# The driver backend needs to identify per-sample color writes and emit
1426# specific code for them.
1427#
1428# src[] = { value, render_target }
1429# BASE = sample index
1430store("tlb_sample_color_v3d", [1], [BASE, COMPONENT, SRC_TYPE], [])
1431
1432# V3D-specific intrinsic to load the number of layers attached to
1433# the target framebuffer
1434intrinsic("load_fb_layers_v3d", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
1435
1436# Logical complement of load_front_face, mapping to an AGX system value
1437system_value("back_face_agx", 1, bit_sizes=[1, 32])
1438
1439# Intel-specific query for loading from the brw_image_param struct passed
1440# into the shader as a uniform.  The variable is a deref to the image
1441# variable. The const index specifies which of the six parameters to load.
1442intrinsic("image_deref_load_param_intel", src_comp=[1], dest_comp=0,
1443          indices=[BASE], flags=[CAN_ELIMINATE, CAN_REORDER])
1444image("load_raw_intel", src_comp=[1], dest_comp=0,
1445      flags=[CAN_ELIMINATE])
1446image("store_raw_intel", src_comp=[1, 0])
1447
1448# Intrinsic to load a block of at least 32B of constant data from a 64-bit
1449# global memory address.  The memory address must be uniform and 32B-aligned.
1450# The second source is a predicate which indicates whether or not to actually
1451# do the load.
1452# src[] = { address, predicate }.
1453intrinsic("load_global_const_block_intel", src_comp=[1, 1], dest_comp=0,
1454          bit_sizes=[32], indices=[BASE], flags=[CAN_ELIMINATE, CAN_REORDER])
1455
1456# Number of data items being operated on for a SIMD program.
1457system_value("simd_width_intel", 1)
1458
1459# Load a relocatable 32-bit value
1460intrinsic("load_reloc_const_intel", dest_comp=1, bit_sizes=[32],
1461          indices=[PARAM_IDX], flags=[CAN_ELIMINATE, CAN_REORDER])
1462
1463# 64-bit global address for a Vulkan descriptor set
1464# src[0] = { set }
1465intrinsic("load_desc_set_address_intel", dest_comp=1, bit_sizes=[64],
1466          src_comp=[1], flags=[CAN_ELIMINATE, CAN_REORDER])
1467
1468# OpSubgroupBlockReadINTEL and OpSubgroupBlockWriteINTEL from SPV_INTEL_subgroups.
1469intrinsic("load_deref_block_intel", dest_comp=0, src_comp=[-1],
1470          indices=[ACCESS], flags=[CAN_ELIMINATE])
1471intrinsic("store_deref_block_intel", src_comp=[-1, 0], indices=[WRITE_MASK, ACCESS])
1472
1473# src[] = { address }.
1474load("global_block_intel", [1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1475
1476# src[] = { buffer_index, offset }.
1477load("ssbo_block_intel", [-1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1478
1479# src[] = { offset }.
1480load("shared_block_intel", [1], [BASE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
1481
1482# src[] = { value, address }.
1483store("global_block_intel", [1], [WRITE_MASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
1484
1485# src[] = { value, block_index, offset }
1486store("ssbo_block_intel", [-1, 1], [WRITE_MASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
1487
1488# src[] = { value, offset }.
1489store("shared_block_intel", [1], [BASE, WRITE_MASK, ALIGN_MUL, ALIGN_OFFSET])
1490
1491# Intrinsics for Intel mesh shading
1492system_value("mesh_inline_data_intel", 1, [ALIGN_OFFSET], bit_sizes=[32, 64])
1493
1494# Intrinsics for Intel bindless thread dispatch
1495# BASE=brw_topoloy_id
1496system_value("topology_id_intel", 1, indices=[BASE])
1497system_value("btd_stack_id_intel", 1)
1498system_value("btd_global_arg_addr_intel", 1, bit_sizes=[64])
1499system_value("btd_local_arg_addr_intel", 1, bit_sizes=[64])
1500system_value("btd_resume_sbt_addr_intel", 1, bit_sizes=[64])
1501# src[] = { global_arg_addr, btd_record }
1502intrinsic("btd_spawn_intel", src_comp=[1, 1])
1503# RANGE=stack_size
1504intrinsic("btd_stack_push_intel", indices=[STACK_SIZE])
1505# src[] = { }
1506intrinsic("btd_retire_intel")
1507
1508# Intel-specific ray-tracing intrinsic
1509# src[] = { globals, level, operation } SYNCHRONOUS=synchronous
1510intrinsic("trace_ray_intel", src_comp=[1, 1, 1], indices=[SYNCHRONOUS])
1511
1512# System values used for ray-tracing on Intel
1513system_value("ray_base_mem_addr_intel", 1, bit_sizes=[64])
1514system_value("ray_hw_stack_size_intel", 1)
1515system_value("ray_sw_stack_size_intel", 1)
1516system_value("ray_num_dss_rt_stacks_intel", 1)
1517system_value("ray_hit_sbt_addr_intel", 1, bit_sizes=[64])
1518system_value("ray_hit_sbt_stride_intel", 1, bit_sizes=[16])
1519system_value("ray_miss_sbt_addr_intel", 1, bit_sizes=[64])
1520system_value("ray_miss_sbt_stride_intel", 1, bit_sizes=[16])
1521system_value("callable_sbt_addr_intel", 1, bit_sizes=[64])
1522system_value("callable_sbt_stride_intel", 1, bit_sizes=[16])
1523system_value("leaf_opaque_intel", 1, bit_sizes=[1])
1524system_value("leaf_procedural_intel", 1, bit_sizes=[1])
1525# Values :
1526#  0: AnyHit
1527#  1: ClosestHit
1528#  2: Miss
1529#  3: Intersection
1530system_value("btd_shader_type_intel", 1)
1531system_value("ray_query_global_intel", 1, bit_sizes=[64])
1532
1533# In order to deal with flipped render targets, gl_PointCoord may be flipped
1534# in the shader requiring a shader key or extra instructions or it may be
1535# flipped in hardware based on a state bit.  This version of gl_PointCoord
1536# is defined to be whatever thing the hardware can easily give you, so long as
1537# it's in normalized coordinates in the range [0, 1] across the point.
1538intrinsic("load_point_coord_maybe_flipped", dest_comp=2, bit_sizes=[32])
1539