fb4c459e0627a1a8933881e30c6e951281d7fb4f
2 # Copyright (C) 2018 Red Hat
3 # Copyright (C) 2014 Intel Corporation
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:
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
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
25 # This file defines all the available intrinsics in one place.
27 # The Intrinsic class corresponds one-to-one with nir_intrinsic_info
30 class Intrinsic(object):
31 """Class that represents all the information about an intrinsic opcode.
32 NOTE: this must be kept in sync with nir_intrinsic_info.
34 def __init__(self
, name
, src_components
, dest_components
,
35 indices
, flags
, sysval
, bit_sizes
):
38 - name: the intrinsic name
39 - src_components: list of the number of components per src, 0 means
40 vectorized instruction with number of components given in the
41 num_components field in nir_intrinsic_instr.
42 - dest_components: number of destination components, -1 means no
43 dest, 0 means number of components given in num_components field
44 in nir_intrinsic_instr.
45 - indices: list of constant indicies
46 - flags: list of semantic flags
47 - sysval: is this a system-value intrinsic
48 - bit_sizes: allowed dest bit_sizes
50 assert isinstance(name
, str)
51 assert isinstance(src_components
, list)
53 assert isinstance(src_components
[0], int)
54 assert isinstance(dest_components
, int)
55 assert isinstance(indices
, list)
57 assert isinstance(indices
[0], str)
58 assert isinstance(flags
, list)
60 assert isinstance(flags
[0], str)
61 assert isinstance(sysval
, bool)
63 assert isinstance(bit_sizes
[0], int)
66 self
.num_srcs
= len(src_components
)
67 self
.src_components
= src_components
68 self
.has_dest
= (dest_components
>= 0)
69 self
.dest_components
= dest_components
70 self
.num_indices
= len(indices
)
71 self
.indices
= indices
74 self
.bit_sizes
= bit_sizes
80 # A constant 'base' value that is added to an offset src:
81 BASE
= "NIR_INTRINSIC_BASE"
82 # For store instructions, a writemask:
83 WRMASK
= "NIR_INTRINSIC_WRMASK"
84 # The stream-id for GS emit_vertex/end_primitive intrinsics:
85 STREAM_ID
= "NIR_INTRINSIC_STREAM_ID"
86 # The clip-plane id for load_user_clip_plane intrinsics:
87 UCP_ID
= "NIR_INTRINSIC_UCP_ID"
88 # The amount of data, starting from BASE, that this instruction
89 # may access. This is used to provide bounds if the offset is
91 RANGE
= "NIR_INTRINSIC_RANGE"
92 # The vulkan descriptor set binding for vulkan_resource_index
94 DESC_SET
= "NIR_INTRINSIC_DESC_SET"
95 # The vulkan descriptor set binding for vulkan_resource_index
97 BINDING
= "NIR_INTRINSIC_BINDING"
99 COMPONENT
= "NIR_INTRINSIC_COMPONENT"
100 # Interpolation mode (only meaningful for FS inputs)
101 INTERP_MODE
= "NIR_INTRINSIC_INTERP_MODE"
102 # A binary nir_op to use when performing a reduction or scan operation
103 REDUCTION_OP
= "NIR_INTRINSIC_REDUCTION_OP"
104 # Cluster size for reduction operations
105 CLUSTER_SIZE
= "NIR_INTRINSIC_CLUSTER_SIZE"
106 # Parameter index for a load_param intrinsic
107 PARAM_IDX
= "NIR_INTRINSIC_PARAM_IDX"
108 # Image dimensionality for image intrinsics
109 IMAGE_DIM
= "NIR_INTRINSIC_IMAGE_DIM"
110 # Non-zero if we are accessing an array image
111 IMAGE_ARRAY
= "NIR_INTRINSIC_IMAGE_ARRAY"
112 # Access qualifiers for image and memory access intrinsics
113 ACCESS
= "NIR_INTRINSIC_ACCESS"
114 DST_ACCESS
= "NIR_INTRINSIC_DST_ACCESS"
115 SRC_ACCESS
= "NIR_INTRINSIC_SRC_ACCESS"
116 # Image format for image intrinsics
117 FORMAT
= "NIR_INTRINSIC_FORMAT"
118 # Offset or address alignment
119 ALIGN_MUL
= "NIR_INTRINSIC_ALIGN_MUL"
120 ALIGN_OFFSET
= "NIR_INTRINSIC_ALIGN_OFFSET"
121 # The vulkan descriptor type for vulkan_resource_index
122 DESC_TYPE
= "NIR_INTRINSIC_DESC_TYPE"
123 # The nir_alu_type of a uniform/input/output
124 TYPE
= "NIR_INTRINSIC_TYPE"
125 # The swizzle mask for quad_swizzle_amd & masked_swizzle_amd
126 SWIZZLE_MASK
= "NIR_INTRINSIC_SWIZZLE_MASK"
127 # Driver location of attribute
128 DRIVER_LOCATION
= "NIR_INTRINSIC_DRIVER_LOCATION"
129 # Ordering and visibility of a memory operation
130 MEMORY_SEMANTICS
= "NIR_INTRINSIC_MEMORY_SEMANTICS"
131 # Modes affected by a memory operation
132 MEMORY_MODES
= "NIR_INTRINSIC_MEMORY_MODES"
133 # Scope of a memory operation
134 MEMORY_SCOPE
= "NIR_INTRINSIC_MEMORY_SCOPE"
140 CAN_ELIMINATE
= "NIR_INTRINSIC_CAN_ELIMINATE"
141 CAN_REORDER
= "NIR_INTRINSIC_CAN_REORDER"
145 # Defines a new NIR intrinsic. By default, the intrinsic will have no sources
146 # and no destination.
148 # You can set dest_comp=n to enable a destination for the intrinsic, in which
149 # case it will have that many components, or =0 for "as many components as the
150 # NIR destination value."
152 # Set src_comp=n to enable sources for the intruction. It can be an array of
153 # component counts, or (for convenience) a scalar component count if there's
154 # only one source. If a component count is 0, it will be as many components as
155 # the intrinsic has based on the dest_comp.
156 def intrinsic(name
, src_comp
=[], dest_comp
=-1, indices
=[],
157 flags
=[], sysval
=False, bit_sizes
=[]):
158 assert name
not in INTR_OPCODES
159 INTR_OPCODES
[name
] = Intrinsic(name
, src_comp
, dest_comp
,
160 indices
, flags
, sysval
, bit_sizes
)
162 intrinsic("nop", flags
=[CAN_ELIMINATE
])
164 intrinsic("load_param", dest_comp
=0, indices
=[PARAM_IDX
], flags
=[CAN_ELIMINATE
])
166 intrinsic("load_deref", dest_comp
=0, src_comp
=[-1],
167 indices
=[ACCESS
], flags
=[CAN_ELIMINATE
])
168 intrinsic("store_deref", src_comp
=[-1, 0], indices
=[WRMASK
, ACCESS
])
169 intrinsic("copy_deref", src_comp
=[-1, -1], indices
=[DST_ACCESS
, SRC_ACCESS
])
171 # Interpolation of input. The interp_deref_at* intrinsics are similar to the
172 # load_var intrinsic acting on a shader input except that they interpolate the
173 # input differently. The at_sample, at_offset and at_vertex intrinsics take an
174 # additional source that is an integer sample id, a vec2 position offset, or a
175 # vertex ID respectively.
177 intrinsic("interp_deref_at_centroid", dest_comp
=0, src_comp
=[1],
178 flags
=[ CAN_ELIMINATE
, CAN_REORDER
])
179 intrinsic("interp_deref_at_sample", src_comp
=[1, 1], dest_comp
=0,
180 flags
=[CAN_ELIMINATE
, CAN_REORDER
])
181 intrinsic("interp_deref_at_offset", src_comp
=[1, 2], dest_comp
=0,
182 flags
=[CAN_ELIMINATE
, CAN_REORDER
])
183 intrinsic("interp_deref_at_vertex", src_comp
=[1, 1], dest_comp
=0,
184 flags
=[CAN_ELIMINATE
, CAN_REORDER
])
186 # Gets the length of an unsized array at the end of a buffer
187 intrinsic("deref_buffer_array_length", src_comp
=[-1], dest_comp
=1,
188 flags
=[CAN_ELIMINATE
, CAN_REORDER
])
190 # Ask the driver for the size of a given buffer. It takes the buffer index
192 intrinsic("get_buffer_size", src_comp
=[-1], dest_comp
=1,
193 flags
=[CAN_ELIMINATE
, CAN_REORDER
])
195 # a barrier is an intrinsic with no inputs/outputs but which can't be moved
196 # around/optimized in general
202 # Demote fragment shader invocation to a helper invocation. Any stores to
203 # memory after this instruction are suppressed and the fragment does not write
204 # outputs to the framebuffer. Unlike discard, demote needs to ensure that
205 # derivatives will still work for invocations that were not demoted.
207 # As specified by SPV_EXT_demote_to_helper_invocation.
209 intrinsic("is_helper_invocation", dest_comp
=1, flags
=[CAN_ELIMINATE
])
211 # A workgroup-level control barrier. Any thread which hits this barrier will
212 # pause until all threads within the current workgroup have also hit the
213 # barrier. For compute shaders, the workgroup is defined as the local group.
214 # For tessellation control shaders, the workgroup is defined as the current
215 # patch. This intrinsic does not imply any sort of memory barrier.
216 barrier("control_barrier")
218 # Memory barrier with semantics analogous to the memoryBarrier() GLSL
220 barrier("memory_barrier")
222 # Memory barrier with explicit scope. Follows the semantics of SPIR-V
223 # OpMemoryBarrier, used to implement Vulkan Memory Model. Storage that the
224 # barrierr applies is represented using NIR variable modes.
225 intrinsic("scoped_memory_barrier",
226 indices
=[MEMORY_SEMANTICS
, MEMORY_MODES
, MEMORY_SCOPE
])
228 # Shader clock intrinsic with semantics analogous to the clock2x32ARB()
230 # The latter can be used as code motion barrier, which is currently not
232 intrinsic("shader_clock", dest_comp
=2, flags
=[CAN_ELIMINATE
])
234 # Shader ballot intrinsics with semantics analogous to the
237 # readInvocationARB()
238 # readFirstInvocationARB()
240 # GLSL functions from ARB_shader_ballot.
241 intrinsic("ballot", src_comp
=[1], dest_comp
=0, flags
=[CAN_ELIMINATE
])
242 intrinsic("read_invocation", src_comp
=[0, 1], dest_comp
=0, flags
=[CAN_ELIMINATE
])
243 intrinsic("read_first_invocation", src_comp
=[0], dest_comp
=0, flags
=[CAN_ELIMINATE
])
245 # Additional SPIR-V ballot intrinsics
247 # These correspond to the SPIR-V opcodes
249 # OpGroupUniformElect
250 # OpSubgroupFirstInvocationKHR
251 intrinsic("elect", dest_comp
=1, flags
=[CAN_ELIMINATE
])
252 intrinsic("first_invocation", dest_comp
=1, flags
=[CAN_ELIMINATE
])
254 # Memory barrier with semantics analogous to the compute shader
255 # groupMemoryBarrier(), memoryBarrierAtomicCounter(), memoryBarrierBuffer(),
256 # memoryBarrierImage() and memoryBarrierShared() GLSL intrinsics.
257 barrier("group_memory_barrier")
258 barrier("memory_barrier_atomic_counter")
259 barrier("memory_barrier_buffer")
260 barrier("memory_barrier_image")
261 barrier("memory_barrier_shared")
262 barrier("begin_invocation_interlock")
263 barrier("end_invocation_interlock")
265 # Memory barrier for synchronizing TCS patch outputs
266 barrier("memory_barrier_tcs_patch")
268 # A conditional discard/demote, with a single boolean source.
269 intrinsic("discard_if", src_comp
=[1])
270 intrinsic("demote_if", src_comp
=[1])
272 # ARB_shader_group_vote intrinsics
273 intrinsic("vote_any", src_comp
=[1], dest_comp
=1, flags
=[CAN_ELIMINATE
])
274 intrinsic("vote_all", src_comp
=[1], dest_comp
=1, flags
=[CAN_ELIMINATE
])
275 intrinsic("vote_feq", src_comp
=[0], dest_comp
=1, flags
=[CAN_ELIMINATE
])
276 intrinsic("vote_ieq", src_comp
=[0], dest_comp
=1, flags
=[CAN_ELIMINATE
])
278 # Ballot ALU operations from SPIR-V.
280 # These operations work like their ALU counterparts except that the operate
281 # on a uvec4 which is treated as a 128bit integer. Also, they are, in
282 # general, free to ignore any bits which are above the subgroup size.
283 intrinsic("ballot_bitfield_extract", src_comp
=[4, 1], dest_comp
=1, flags
=[CAN_ELIMINATE
])
284 intrinsic("ballot_bit_count_reduce", src_comp
=[4], dest_comp
=1, flags
=[CAN_ELIMINATE
])
285 intrinsic("ballot_bit_count_inclusive", src_comp
=[4], dest_comp
=1, flags
=[CAN_ELIMINATE
])
286 intrinsic("ballot_bit_count_exclusive", src_comp
=[4], dest_comp
=1, flags
=[CAN_ELIMINATE
])
287 intrinsic("ballot_find_lsb", src_comp
=[4], dest_comp
=1, flags
=[CAN_ELIMINATE
])
288 intrinsic("ballot_find_msb", src_comp
=[4], dest_comp
=1, flags
=[CAN_ELIMINATE
])
290 # Shuffle operations from SPIR-V.
291 intrinsic("shuffle", src_comp
=[0, 1], dest_comp
=0, flags
=[CAN_ELIMINATE
])
292 intrinsic("shuffle_xor", src_comp
=[0, 1], dest_comp
=0, flags
=[CAN_ELIMINATE
])
293 intrinsic("shuffle_up", src_comp
=[0, 1], dest_comp
=0, flags
=[CAN_ELIMINATE
])
294 intrinsic("shuffle_down", src_comp
=[0, 1], dest_comp
=0, flags
=[CAN_ELIMINATE
])
296 # Quad operations from SPIR-V.
297 intrinsic("quad_broadcast", src_comp
=[0, 1], dest_comp
=0, flags
=[CAN_ELIMINATE
])
298 intrinsic("quad_swap_horizontal", src_comp
=[0], dest_comp
=0, flags
=[CAN_ELIMINATE
])
299 intrinsic("quad_swap_vertical", src_comp
=[0], dest_comp
=0, flags
=[CAN_ELIMINATE
])
300 intrinsic("quad_swap_diagonal", src_comp
=[0], dest_comp
=0, flags
=[CAN_ELIMINATE
])
302 intrinsic("reduce", src_comp
=[0], dest_comp
=0, indices
=[REDUCTION_OP
, CLUSTER_SIZE
],
303 flags
=[CAN_ELIMINATE
])
304 intrinsic("inclusive_scan", src_comp
=[0], dest_comp
=0, indices
=[REDUCTION_OP
],
305 flags
=[CAN_ELIMINATE
])
306 intrinsic("exclusive_scan", src_comp
=[0], dest_comp
=0, indices
=[REDUCTION_OP
],
307 flags
=[CAN_ELIMINATE
])
309 # AMD shader ballot operations
310 intrinsic("quad_swizzle_amd", src_comp
=[0], dest_comp
=0, indices
=[SWIZZLE_MASK
],
311 flags
=[CAN_ELIMINATE
])
312 intrinsic("masked_swizzle_amd", src_comp
=[0], dest_comp
=0, indices
=[SWIZZLE_MASK
],
313 flags
=[CAN_ELIMINATE
])
314 intrinsic("write_invocation_amd", src_comp
=[0, 0, 1], dest_comp
=0, flags
=[CAN_ELIMINATE
])
315 intrinsic("mbcnt_amd", src_comp
=[1], dest_comp
=1, flags
=[CAN_ELIMINATE
])
317 # Basic Geometry Shader intrinsics.
319 # emit_vertex implements GLSL's EmitStreamVertex() built-in. It takes a single
320 # index, which is the stream ID to write to.
322 # end_primitive implements GLSL's EndPrimitive() built-in.
323 intrinsic("emit_vertex", indices
=[STREAM_ID
])
324 intrinsic("end_primitive", indices
=[STREAM_ID
])
326 # Geometry Shader intrinsics with a vertex count.
328 # Alternatively, drivers may implement these intrinsics, and use
329 # nir_lower_gs_intrinsics() to convert from the basic intrinsics.
331 # These maintain a count of the number of vertices emitted, as an additional
332 # unsigned integer source.
333 intrinsic("emit_vertex_with_counter", src_comp
=[1], indices
=[STREAM_ID
])
334 intrinsic("end_primitive_with_counter", src_comp
=[1], indices
=[STREAM_ID
])
335 intrinsic("set_vertex_count", src_comp
=[1])
339 # The *_var variants take an atomic_uint nir_variable, while the other,
340 # lowered, variants take a constant buffer index and register offset.
342 def atomic(name
, flags
=[]):
343 intrinsic(name
+ "_deref", src_comp
=[-1], dest_comp
=1, flags
=flags
)
344 intrinsic(name
, src_comp
=[1], dest_comp
=1, indices
=[BASE
], flags
=flags
)
347 intrinsic(name
+ "_deref", src_comp
=[-1, 1], dest_comp
=1)
348 intrinsic(name
, src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
351 intrinsic(name
+ "_deref", src_comp
=[-1, 1, 1], dest_comp
=1)
352 intrinsic(name
, src_comp
=[1, 1, 1], dest_comp
=1, indices
=[BASE
])
354 atomic("atomic_counter_inc")
355 atomic("atomic_counter_pre_dec")
356 atomic("atomic_counter_post_dec")
357 atomic("atomic_counter_read", flags
=[CAN_ELIMINATE
])
358 atomic2("atomic_counter_add")
359 atomic2("atomic_counter_min")
360 atomic2("atomic_counter_max")
361 atomic2("atomic_counter_and")
362 atomic2("atomic_counter_or")
363 atomic2("atomic_counter_xor")
364 atomic2("atomic_counter_exchange")
365 atomic3("atomic_counter_comp_swap")
367 # Image load, store and atomic intrinsics.
369 # All image intrinsics come in three versions. One which take an image target
370 # passed as a deref chain as the first source, one which takes an index as the
371 # first source, and one which takes a bindless handle as the first source.
372 # In the first version, the image variable contains the memory and layout
373 # qualifiers that influence the semantics of the intrinsic. In the second and
374 # third, the image format and access qualifiers are provided as constant
377 # All image intrinsics take a four-coordinate vector and a sample index as
378 # 2nd and 3rd sources, determining the location within the image that will be
379 # accessed by the intrinsic. Components not applicable to the image target
380 # in use are undefined. Image store takes an additional four-component
381 # argument with the value to be written, and image atomic operations take
382 # either one or two additional scalar arguments with the same meaning as in
383 # the ARB_shader_image_load_store specification.
384 def image(name
, src_comp
=[], **kwargs
):
385 intrinsic("image_deref_" + name
, src_comp
=[1] + src_comp
,
386 indices
=[ACCESS
], **kwargs
)
387 intrinsic("image_" + name
, src_comp
=[1] + src_comp
,
388 indices
=[IMAGE_DIM
, IMAGE_ARRAY
, FORMAT
, ACCESS
], **kwargs
)
389 intrinsic("bindless_image_" + name
, src_comp
=[1] + src_comp
,
390 indices
=[IMAGE_DIM
, IMAGE_ARRAY
, FORMAT
, ACCESS
], **kwargs
)
392 image("load", src_comp
=[4, 1, 1], dest_comp
=0, flags
=[CAN_ELIMINATE
])
393 image("store", src_comp
=[4, 1, 0, 1])
394 image("atomic_add", src_comp
=[4, 1, 1], dest_comp
=1)
395 image("atomic_imin", src_comp
=[4, 1, 1], dest_comp
=1)
396 image("atomic_umin", src_comp
=[4, 1, 1], dest_comp
=1)
397 image("atomic_imax", src_comp
=[4, 1, 1], dest_comp
=1)
398 image("atomic_umax", src_comp
=[4, 1, 1], dest_comp
=1)
399 image("atomic_and", src_comp
=[4, 1, 1], dest_comp
=1)
400 image("atomic_or", src_comp
=[4, 1, 1], dest_comp
=1)
401 image("atomic_xor", src_comp
=[4, 1, 1], dest_comp
=1)
402 image("atomic_exchange", src_comp
=[4, 1, 1], dest_comp
=1)
403 image("atomic_comp_swap", src_comp
=[4, 1, 1, 1], dest_comp
=1)
404 image("atomic_fadd", src_comp
=[1, 4, 1, 1], dest_comp
=1)
405 image("size", dest_comp
=0, flags
=[CAN_ELIMINATE
, CAN_REORDER
])
406 image("samples", dest_comp
=1, flags
=[CAN_ELIMINATE
, CAN_REORDER
])
407 image("atomic_inc_wrap", src_comp
=[4, 1, 1], dest_comp
=1)
408 image("atomic_dec_wrap", src_comp
=[4, 1, 1], dest_comp
=1)
410 # Intel-specific query for loading from the brw_image_param struct passed
411 # into the shader as a uniform. The variable is a deref to the image
412 # variable. The const index specifies which of the six parameters to load.
413 intrinsic("image_deref_load_param_intel", src_comp
=[1], dest_comp
=0,
414 indices
=[BASE
], flags
=[CAN_ELIMINATE
, CAN_REORDER
])
415 image("load_raw_intel", src_comp
=[1], dest_comp
=0,
416 flags
=[CAN_ELIMINATE
])
417 image("store_raw_intel", src_comp
=[1, 0])
419 # Vulkan descriptor set intrinsics
421 # The Vulkan API uses a different binding model from GL. In the Vulkan
422 # API, all external resources are represented by a tuple:
424 # (descriptor set, binding, array index)
426 # where the array index is the only thing allowed to be indirect. The
427 # vulkan_surface_index intrinsic takes the descriptor set and binding as
428 # its first two indices and the array index as its source. The third
429 # index is a nir_variable_mode in case that's useful to the backend.
431 # The intended usage is that the shader will call vulkan_surface_index to
432 # get an index and then pass that as the buffer index ubo/ssbo calls.
434 # The vulkan_resource_reindex intrinsic takes a resource index in src0
435 # (the result of a vulkan_resource_index or vulkan_resource_reindex) which
436 # corresponds to the tuple (set, binding, index) and computes an index
437 # corresponding to tuple (set, binding, idx + src1).
438 intrinsic("vulkan_resource_index", src_comp
=[1], dest_comp
=0,
439 indices
=[DESC_SET
, BINDING
, DESC_TYPE
],
440 flags
=[CAN_ELIMINATE
, CAN_REORDER
])
441 intrinsic("vulkan_resource_reindex", src_comp
=[0, 1], dest_comp
=0,
442 indices
=[DESC_TYPE
], flags
=[CAN_ELIMINATE
, CAN_REORDER
])
443 intrinsic("load_vulkan_descriptor", src_comp
=[-1], dest_comp
=0,
444 indices
=[DESC_TYPE
], flags
=[CAN_ELIMINATE
, CAN_REORDER
])
446 # variable atomic intrinsics
448 # All of these variable atomic memory operations read a value from memory,
449 # compute a new value using one of the operations below, write the new value
450 # to memory, and return the original value read.
452 # All operations take 2 sources except CompSwap that takes 3. These sources
455 # 0: A deref to the memory on which to perform the atomic
456 # 1: The data parameter to the atomic function (i.e. the value to add
457 # in shared_atomic_add, etc).
458 # 2: For CompSwap only: the second data parameter.
459 intrinsic("deref_atomic_add", src_comp
=[-1, 1], dest_comp
=1, indices
=[ACCESS
])
460 intrinsic("deref_atomic_imin", src_comp
=[-1, 1], dest_comp
=1, indices
=[ACCESS
])
461 intrinsic("deref_atomic_umin", src_comp
=[-1, 1], dest_comp
=1, indices
=[ACCESS
])
462 intrinsic("deref_atomic_imax", src_comp
=[-1, 1], dest_comp
=1, indices
=[ACCESS
])
463 intrinsic("deref_atomic_umax", src_comp
=[-1, 1], dest_comp
=1, indices
=[ACCESS
])
464 intrinsic("deref_atomic_and", src_comp
=[-1, 1], dest_comp
=1, indices
=[ACCESS
])
465 intrinsic("deref_atomic_or", src_comp
=[-1, 1], dest_comp
=1, indices
=[ACCESS
])
466 intrinsic("deref_atomic_xor", src_comp
=[-1, 1], dest_comp
=1, indices
=[ACCESS
])
467 intrinsic("deref_atomic_exchange", src_comp
=[-1, 1], dest_comp
=1, indices
=[ACCESS
])
468 intrinsic("deref_atomic_comp_swap", src_comp
=[-1, 1, 1], dest_comp
=1, indices
=[ACCESS
])
469 intrinsic("deref_atomic_fadd", src_comp
=[-1, 1], dest_comp
=1, indices
=[ACCESS
])
470 intrinsic("deref_atomic_fmin", src_comp
=[-1, 1], dest_comp
=1, indices
=[ACCESS
])
471 intrinsic("deref_atomic_fmax", src_comp
=[-1, 1], dest_comp
=1, indices
=[ACCESS
])
472 intrinsic("deref_atomic_fcomp_swap", src_comp
=[-1, 1, 1], dest_comp
=1, indices
=[ACCESS
])
474 # SSBO atomic intrinsics
476 # All of the SSBO atomic memory operations read a value from memory,
477 # compute a new value using one of the operations below, write the new
478 # value to memory, and return the original value read.
480 # All operations take 3 sources except CompSwap that takes 4. These
483 # 0: The SSBO buffer index.
484 # 1: The offset into the SSBO buffer of the variable that the atomic
485 # operation will operate on.
486 # 2: The data parameter to the atomic function (i.e. the value to add
487 # in ssbo_atomic_add, etc).
488 # 3: For CompSwap only: the second data parameter.
489 intrinsic("ssbo_atomic_add", src_comp
=[1, 1, 1], dest_comp
=1, indices
=[ACCESS
])
490 intrinsic("ssbo_atomic_imin", src_comp
=[1, 1, 1], dest_comp
=1, indices
=[ACCESS
])
491 intrinsic("ssbo_atomic_umin", src_comp
=[1, 1, 1], dest_comp
=1, indices
=[ACCESS
])
492 intrinsic("ssbo_atomic_imax", src_comp
=[1, 1, 1], dest_comp
=1, indices
=[ACCESS
])
493 intrinsic("ssbo_atomic_umax", src_comp
=[1, 1, 1], dest_comp
=1, indices
=[ACCESS
])
494 intrinsic("ssbo_atomic_and", src_comp
=[1, 1, 1], dest_comp
=1, indices
=[ACCESS
])
495 intrinsic("ssbo_atomic_or", src_comp
=[1, 1, 1], dest_comp
=1, indices
=[ACCESS
])
496 intrinsic("ssbo_atomic_xor", src_comp
=[1, 1, 1], dest_comp
=1, indices
=[ACCESS
])
497 intrinsic("ssbo_atomic_exchange", src_comp
=[1, 1, 1], dest_comp
=1, indices
=[ACCESS
])
498 intrinsic("ssbo_atomic_comp_swap", src_comp
=[1, 1, 1, 1], dest_comp
=1, indices
=[ACCESS
])
499 intrinsic("ssbo_atomic_fadd", src_comp
=[1, 1, 1], dest_comp
=1, indices
=[ACCESS
])
500 intrinsic("ssbo_atomic_fmin", src_comp
=[1, 1, 1], dest_comp
=1, indices
=[ACCESS
])
501 intrinsic("ssbo_atomic_fmax", src_comp
=[1, 1, 1], dest_comp
=1, indices
=[ACCESS
])
502 intrinsic("ssbo_atomic_fcomp_swap", src_comp
=[1, 1, 1, 1], dest_comp
=1, indices
=[ACCESS
])
504 # CS shared variable atomic intrinsics
506 # All of the shared variable atomic memory operations read a value from
507 # memory, compute a new value using one of the operations below, write the
508 # new value to memory, and return the original value read.
510 # All operations take 2 sources except CompSwap that takes 3. These
513 # 0: The offset into the shared variable storage region that the atomic
514 # operation will operate on.
515 # 1: The data parameter to the atomic function (i.e. the value to add
516 # in shared_atomic_add, etc).
517 # 2: For CompSwap only: the second data parameter.
518 intrinsic("shared_atomic_add", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
519 intrinsic("shared_atomic_imin", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
520 intrinsic("shared_atomic_umin", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
521 intrinsic("shared_atomic_imax", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
522 intrinsic("shared_atomic_umax", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
523 intrinsic("shared_atomic_and", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
524 intrinsic("shared_atomic_or", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
525 intrinsic("shared_atomic_xor", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
526 intrinsic("shared_atomic_exchange", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
527 intrinsic("shared_atomic_comp_swap", src_comp
=[1, 1, 1], dest_comp
=1, indices
=[BASE
])
528 intrinsic("shared_atomic_fadd", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
529 intrinsic("shared_atomic_fmin", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
530 intrinsic("shared_atomic_fmax", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
531 intrinsic("shared_atomic_fcomp_swap", src_comp
=[1, 1, 1], dest_comp
=1, indices
=[BASE
])
533 # Global atomic intrinsics
535 # All of the shared variable atomic memory operations read a value from
536 # memory, compute a new value using one of the operations below, write the
537 # new value to memory, and return the original value read.
539 # All operations take 2 sources except CompSwap that takes 3. These
542 # 0: The memory address that the atomic operation will operate on.
543 # 1: The data parameter to the atomic function (i.e. the value to add
544 # in shared_atomic_add, etc).
545 # 2: For CompSwap only: the second data parameter.
546 intrinsic("global_atomic_add", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
547 intrinsic("global_atomic_imin", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
548 intrinsic("global_atomic_umin", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
549 intrinsic("global_atomic_imax", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
550 intrinsic("global_atomic_umax", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
551 intrinsic("global_atomic_and", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
552 intrinsic("global_atomic_or", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
553 intrinsic("global_atomic_xor", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
554 intrinsic("global_atomic_exchange", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
555 intrinsic("global_atomic_comp_swap", src_comp
=[1, 1, 1], dest_comp
=1, indices
=[BASE
])
556 intrinsic("global_atomic_fadd", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
557 intrinsic("global_atomic_fmin", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
558 intrinsic("global_atomic_fmax", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
559 intrinsic("global_atomic_fcomp_swap", src_comp
=[1, 1, 1], dest_comp
=1, indices
=[BASE
])
561 def system_value(name
, dest_comp
, indices
=[], bit_sizes
=[32]):
562 intrinsic("load_" + name
, [], dest_comp
, indices
,
563 flags
=[CAN_ELIMINATE
, CAN_REORDER
], sysval
=True,
566 system_value("frag_coord", 4)
567 system_value("point_coord", 2)
568 system_value("front_face", 1, bit_sizes
=[1, 32])
569 system_value("vertex_id", 1)
570 system_value("vertex_id_zero_base", 1)
571 system_value("first_vertex", 1)
572 system_value("is_indexed_draw", 1)
573 system_value("base_vertex", 1)
574 system_value("instance_id", 1)
575 system_value("base_instance", 1)
576 system_value("draw_id", 1)
577 system_value("sample_id", 1)
578 # sample_id_no_per_sample is like sample_id but does not imply per-
579 # sample shading. See the lower_helper_invocation option.
580 system_value("sample_id_no_per_sample", 1)
581 system_value("sample_pos", 2)
582 system_value("sample_mask_in", 1)
583 system_value("primitive_id", 1)
584 system_value("invocation_id", 1)
585 system_value("tess_coord", 3)
586 system_value("tess_level_outer", 4)
587 system_value("tess_level_inner", 2)
588 system_value("tess_level_outer_default", 4)
589 system_value("tess_level_inner_default", 2)
590 system_value("patch_vertices_in", 1)
591 system_value("local_invocation_id", 3)
592 system_value("local_invocation_index", 1)
593 system_value("work_group_id", 3)
594 system_value("user_clip_plane", 4, indices
=[UCP_ID
])
595 system_value("num_work_groups", 3)
596 system_value("helper_invocation", 1, bit_sizes
=[1, 32])
597 system_value("alpha_ref_float", 1)
598 system_value("layer_id", 1)
599 system_value("view_index", 1)
600 system_value("subgroup_size", 1)
601 system_value("subgroup_invocation", 1)
602 system_value("subgroup_eq_mask", 0, bit_sizes
=[32, 64])
603 system_value("subgroup_ge_mask", 0, bit_sizes
=[32, 64])
604 system_value("subgroup_gt_mask", 0, bit_sizes
=[32, 64])
605 system_value("subgroup_le_mask", 0, bit_sizes
=[32, 64])
606 system_value("subgroup_lt_mask", 0, bit_sizes
=[32, 64])
607 system_value("num_subgroups", 1)
608 system_value("subgroup_id", 1)
609 system_value("local_group_size", 3)
610 system_value("global_invocation_id", 3, bit_sizes
=[32, 64])
611 system_value("global_invocation_index", 1, bit_sizes
=[32, 64])
612 system_value("work_dim", 1)
613 # Driver-specific viewport scale/offset parameters.
615 # VC4 and V3D need to emit a scaled version of the position in the vertex
616 # shaders for binning, and having system values lets us move the math for that
619 # Panfrost needs to implement all coordinate transformation in the
620 # vertex shader; system values allow us to share this routine in NIR.
621 system_value("viewport_x_scale", 1)
622 system_value("viewport_y_scale", 1)
623 system_value("viewport_z_scale", 1)
624 system_value("viewport_z_offset", 1)
625 system_value("viewport_scale", 3)
626 system_value("viewport_offset", 3)
628 # Blend constant color values. Float values are clamped. Vectored versions are
629 # provided as well for driver convenience
631 system_value("blend_const_color_r_float", 1)
632 system_value("blend_const_color_g_float", 1)
633 system_value("blend_const_color_b_float", 1)
634 system_value("blend_const_color_a_float", 1)
635 system_value("blend_const_color_rgba", 4)
636 system_value("blend_const_color_rgba8888_unorm", 1)
637 system_value("blend_const_color_aaaa8888_unorm", 1)
639 # System values for gl_Color, for radeonsi which interpolates these in the
640 # shader prolog to handle two-sided color without recompiles and therefore
641 # doesn't handle these in the main shader part like normal varyings.
642 system_value("color0", 4)
643 system_value("color1", 4)
645 # System value for internal compute shaders in radeonsi.
646 system_value("user_data_amd", 4)
648 # Barycentric coordinate intrinsics.
650 # These set up the barycentric coordinates for a particular interpolation.
651 # The first four are for the simple cases: pixel, centroid, per-sample
652 # (at gl_SampleID), or pull model (1/W, 1/I, 1/J) at the pixel center. The next
653 # three two handle interpolating at a specified sample location, or
654 # interpolating with a vec2 offset,
656 # The interp_mode index should be either the INTERP_MODE_SMOOTH or
657 # INTERP_MODE_NOPERSPECTIVE enum values.
659 # The vec2 value produced by these intrinsics is intended for use as the
660 # barycoord source of a load_interpolated_input intrinsic.
662 def barycentric(name
, dst_comp
, src_comp
=[]):
663 intrinsic("load_barycentric_" + name
, src_comp
=src_comp
, dest_comp
=dst_comp
,
664 indices
=[INTERP_MODE
], flags
=[CAN_ELIMINATE
, CAN_REORDER
])
667 barycentric("pixel", 2)
668 barycentric("centroid", 2)
669 barycentric("sample", 2)
670 barycentric("model", 3)
671 # src[] = { sample_id }.
672 barycentric("at_sample", 2, [1])
673 # src[] = { offset.xy }.
674 barycentric("at_offset", 2, [2])
676 # Load sample position:
678 # Takes a sample # and returns a sample position. Used for lowering
679 # interpolateAtSample() to interpolateAtOffset()
680 intrinsic("load_sample_pos_from_id", src_comp
=[1], dest_comp
=2,
681 flags
=[CAN_ELIMINATE
, CAN_REORDER
])
683 # Loads what I believe is the primitive size, for scaling ij to pixel size:
684 intrinsic("load_size_ir3", dest_comp
=1, flags
=[CAN_ELIMINATE
, CAN_REORDER
])
686 # Fragment shader input interpolation delta intrinsic.
688 # For hw where fragment shader input interpolation is handled in shader, the
689 # load_fs_input_interp deltas intrinsics can be used to load the input deltas
690 # used for interpolation as follows:
692 # vec3 iid = load_fs_input_interp_deltas(varying_slot)
693 # vec2 bary = load_barycentric_*(...)
694 # float result = iid.x + iid.y * bary.y + iid.z * bary.x
696 intrinsic("load_fs_input_interp_deltas", src_comp
=[1], dest_comp
=3,
697 indices
=[BASE
, COMPONENT
], flags
=[CAN_ELIMINATE
, CAN_REORDER
])
699 # Load operations pull data from some piece of GPU memory. All load
700 # operations operate in terms of offsets into some piece of theoretical
701 # memory. Loads from externally visible memory (UBO and SSBO) simply take a
702 # byte offset as a source. Loads from opaque memory (uniforms, inputs, etc.)
703 # take a base+offset pair where the nir_intrinsic_base() gives the location
704 # of the start of the variable being loaded and and the offset source is a
705 # offset into that variable.
707 # Uniform load operations have a nir_intrinsic_range() index that specifies the
708 # range (starting at base) of the data from which we are loading. If
709 # range == 0, then the range is unknown.
711 # Some load operations such as UBO/SSBO load and per_vertex loads take an
712 # additional source to specify which UBO/SSBO/vertex to load from.
714 # The exact address type depends on the lowering pass that generates the
715 # load/store intrinsics. Typically, this is vec4 units for things such as
716 # varying slots and float units for fragment shader inputs. UBO and SSBO
717 # offsets are always in bytes.
719 def load(name
, num_srcs
, indices
=[], flags
=[]):
720 intrinsic("load_" + name
, [1] * num_srcs
, dest_comp
=0, indices
=indices
,
723 # src[] = { offset }.
724 load("uniform", 1, [BASE
, RANGE
, TYPE
], [CAN_ELIMINATE
, CAN_REORDER
])
725 # src[] = { buffer_index, offset }.
726 load("ubo", 2, [ACCESS
, ALIGN_MUL
, ALIGN_OFFSET
], flags
=[CAN_ELIMINATE
, CAN_REORDER
])
727 # src[] = { offset }.
728 load("input", 1, [BASE
, COMPONENT
, TYPE
], [CAN_ELIMINATE
, CAN_REORDER
])
729 # src[] = { vertex_id, offset }.
730 load("input_vertex", 2, [BASE
, COMPONENT
, TYPE
], [CAN_ELIMINATE
, CAN_REORDER
])
731 # src[] = { vertex, offset }.
732 load("per_vertex_input", 2, [BASE
, COMPONENT
], [CAN_ELIMINATE
, CAN_REORDER
])
733 # src[] = { barycoord, offset }.
734 intrinsic("load_interpolated_input", src_comp
=[2, 1], dest_comp
=0,
735 indices
=[BASE
, COMPONENT
], flags
=[CAN_ELIMINATE
, CAN_REORDER
])
737 # src[] = { buffer_index, offset }.
738 load("ssbo", 2, [ACCESS
, ALIGN_MUL
, ALIGN_OFFSET
], [CAN_ELIMINATE
])
739 # src[] = { buffer_index }
740 load("ssbo_address", 1, [], [CAN_ELIMINATE
, CAN_REORDER
])
741 # src[] = { offset }.
742 load("output", 1, [BASE
, COMPONENT
], flags
=[CAN_ELIMINATE
])
743 # src[] = { vertex, offset }.
744 load("per_vertex_output", 2, [BASE
, COMPONENT
], [CAN_ELIMINATE
])
745 # src[] = { offset }.
746 load("shared", 1, [BASE
, ALIGN_MUL
, ALIGN_OFFSET
], [CAN_ELIMINATE
])
747 # src[] = { offset }.
748 load("push_constant", 1, [BASE
, RANGE
], [CAN_ELIMINATE
, CAN_REORDER
])
749 # src[] = { offset }.
750 load("constant", 1, [BASE
, RANGE
, ALIGN_MUL
, ALIGN_OFFSET
],
751 [CAN_ELIMINATE
, CAN_REORDER
])
752 # src[] = { address }.
753 load("global", 1, [ACCESS
, ALIGN_MUL
, ALIGN_OFFSET
], [CAN_ELIMINATE
])
754 # src[] = { address }.
755 load("kernel_input", 1, [BASE
, RANGE
, ALIGN_MUL
, ALIGN_OFFSET
], [CAN_ELIMINATE
, CAN_REORDER
])
756 # src[] = { offset }.
757 load("scratch", 1, [ALIGN_MUL
, ALIGN_OFFSET
], [CAN_ELIMINATE
])
759 # Stores work the same way as loads, except now the first source is the value
760 # to store and the second (and possibly third) source specify where to store
761 # the value. SSBO and shared memory stores also have a
762 # nir_intrinsic_write_mask()
764 def store(name
, num_srcs
, indices
=[], flags
=[]):
765 intrinsic("store_" + name
, [0] + ([1] * (num_srcs
- 1)), indices
=indices
, flags
=flags
)
767 # src[] = { value, offset }.
768 store("output", 2, [BASE
, WRMASK
, COMPONENT
, TYPE
])
769 # src[] = { value, vertex, offset }.
770 store("per_vertex_output", 3, [BASE
, WRMASK
, COMPONENT
])
771 # src[] = { value, block_index, offset }
772 store("ssbo", 3, [WRMASK
, ACCESS
, ALIGN_MUL
, ALIGN_OFFSET
])
773 # src[] = { value, offset }.
774 store("shared", 2, [BASE
, WRMASK
, ALIGN_MUL
, ALIGN_OFFSET
])
775 # src[] = { value, address }.
776 store("global", 2, [WRMASK
, ACCESS
, ALIGN_MUL
, ALIGN_OFFSET
])
777 # src[] = { value, offset }.
778 store("scratch", 2, [ALIGN_MUL
, ALIGN_OFFSET
, WRMASK
])
780 # IR3-specific version of most SSBO intrinsics. The only different
781 # compare to the originals is that they add an extra source to hold
782 # the dword-offset, which is needed by the backend code apart from
783 # the byte-offset already provided by NIR in one of the sources.
785 # NIR lowering pass 'ir3_nir_lower_io_offset' will replace the
786 # original SSBO intrinsics by these, placing the computed
787 # dword-offset always in the last source.
789 # The float versions are not handled because those are not supported
791 intrinsic("store_ssbo_ir3", src_comp
=[0, 1, 1, 1],
792 indices
=[WRMASK
, ACCESS
, ALIGN_MUL
, ALIGN_OFFSET
])
793 intrinsic("load_ssbo_ir3", src_comp
=[1, 1, 1], dest_comp
=0,
794 indices
=[ACCESS
, ALIGN_MUL
, ALIGN_OFFSET
], flags
=[CAN_ELIMINATE
])
795 intrinsic("ssbo_atomic_add_ir3", src_comp
=[1, 1, 1, 1], dest_comp
=1)
796 intrinsic("ssbo_atomic_imin_ir3", src_comp
=[1, 1, 1, 1], dest_comp
=1)
797 intrinsic("ssbo_atomic_umin_ir3", src_comp
=[1, 1, 1, 1], dest_comp
=1)
798 intrinsic("ssbo_atomic_imax_ir3", src_comp
=[1, 1, 1, 1], dest_comp
=1)
799 intrinsic("ssbo_atomic_umax_ir3", src_comp
=[1, 1, 1, 1], dest_comp
=1)
800 intrinsic("ssbo_atomic_and_ir3", src_comp
=[1, 1, 1, 1], dest_comp
=1)
801 intrinsic("ssbo_atomic_or_ir3", src_comp
=[1, 1, 1, 1], dest_comp
=1)
802 intrinsic("ssbo_atomic_xor_ir3", src_comp
=[1, 1, 1, 1], dest_comp
=1)
803 intrinsic("ssbo_atomic_exchange_ir3", src_comp
=[1, 1, 1, 1], dest_comp
=1)
804 intrinsic("ssbo_atomic_comp_swap_ir3", src_comp
=[1, 1, 1, 1, 1], dest_comp
=1)
806 # IR3-specific instruction for UBO loads using the ldc instruction. The second
807 # source is the indirect offset, in units of four dwords. The base is a
808 # component offset, in dword units.
809 intrinsic("load_ubo_ir3", src_comp
=[1, 1], bit_sizes
=[32], dest_comp
=0, indices
=[BASE
],
810 flags
=[CAN_REORDER
, CAN_ELIMINATE
])
812 # System values for freedreno geometry shaders.
813 system_value("vs_primitive_stride_ir3", 1)
814 system_value("vs_vertex_stride_ir3", 1)
815 system_value("gs_header_ir3", 1)
816 system_value("primitive_location_ir3", 1, indices
=[DRIVER_LOCATION
])
818 # System values for freedreno tessellation shaders.
819 system_value("hs_patch_stride_ir3", 1)
820 system_value("tess_factor_base_ir3", 2)
821 system_value("tess_param_base_ir3", 2)
822 system_value("tcs_header_ir3", 1)
824 # IR3-specific intrinsics for tessellation control shaders. cond_end_ir3 end
825 # the shader when src0 is false and is used to narrow down the TCS shader to
826 # just thread 0 before writing out tessellation levels.
827 intrinsic("cond_end_ir3", src_comp
=[1])
828 # end_patch_ir3 is used just before thread 0 exist the TCS and presumably
829 # signals the TE that the patch is complete and can be tessellated.
830 intrinsic("end_patch_ir3")
832 # IR3-specific load/store intrinsics. These access a buffer used to pass data
833 # between geometry stages - perhaps it's explicit access to the vertex cache.
835 # src[] = { value, offset }.
836 store("shared_ir3", 2, [BASE
, WRMASK
, ALIGN_MUL
, ALIGN_OFFSET
])
837 # src[] = { offset }.
838 load("shared_ir3", 1, [BASE
, ALIGN_MUL
, ALIGN_OFFSET
], [CAN_ELIMINATE
])
840 # IR3-specific load/store global intrinsics. They take a 64-bit base address
841 # and a 32-bit offset. The hardware will add the base and the offset, which
842 # saves us from doing 64-bit math on the base address.
844 # src[] = { value, address(vec2 of hi+lo uint32_t), offset }.
845 # const_index[] = { write_mask, align_mul, align_offset }
846 intrinsic("store_global_ir3", [0, 2, 1], indices
=[WRMASK
, ACCESS
, ALIGN_MUL
, ALIGN_OFFSET
])
847 # src[] = { address(vec2 of hi+lo uint32_t), offset }.
848 # const_index[] = { access, align_mul, align_offset }
849 intrinsic("load_global_ir3", [2, 1], dest_comp
=0, indices
=[ACCESS
, ALIGN_MUL
, ALIGN_OFFSET
], flags
=[CAN_ELIMINATE
])
851 # IR3-specific bindless handle specifier. Similar to vulkan_resource_index, but
852 # without the binding because the hardware expects a single flattened index
853 # rather than a (binding, index) pair. We may also want to use this with GL.
854 # Note that this doesn't actually turn into a HW instruction.
855 intrinsic("bindless_resource_ir3", [1], dest_comp
=1, indices
=[DESC_SET
], flags
=[CAN_ELIMINATE
, CAN_REORDER
])
857 # Intrinsics used by the Midgard/Bifrost blend pipeline. These are defined
858 # within a blend shader to read/write the raw value from the tile buffer,
859 # without applying any format conversion in the process. If the shader needs
860 # usable pixel values, it must apply format conversions itself.
862 # These definitions are generic, but they are explicitly vendored to prevent
863 # other drivers from using them, as their semantics is defined in terms of the
864 # Midgard/Bifrost hardware tile buffer and may not line up with anything sane.
865 # One notable divergence is sRGB, which is asymmetric: raw_input_pan requires
866 # an sRGB->linear conversion, but linear values should be written to
867 # raw_output_pan and the hardware handles linear->sRGB.
869 # We also have format-specific Midgard intrinsics. There are rather
870 # here-be-dragons. load_output_u8_as_fp16_pan does the equivalent of
871 # load_raw_out_pan on an RGBA8 UNORM framebuffer followed by u2u16 -> fp16 ->
875 store("raw_output_pan", 1, [])
876 store("zs_output_pan", 1, [COMPONENT
])
877 load("raw_output_pan", 0, [], [CAN_ELIMINATE
, CAN_REORDER
])
878 load("output_u8_as_fp16_pan", 0, [], [CAN_ELIMINATE
, CAN_REORDER
])
880 # Loads the sampler paramaters <min_lod, max_lod, lod_bias>
881 # src[] = { sampler_index }
882 load("sampler_lod_parameters_pan", 1, [CAN_ELIMINATE
, CAN_REORDER
])
884 # R600 specific instrincs
886 # R600 can only fetch 16 byte aligned data from an UBO, and the actual offset
887 # is given in vec4 units, so we have to fetch the a vec4 and get the component
889 # src[] = { buffer_index, offset }.
890 load("ubo_r600", 2, [ACCESS
, ALIGN_MUL
, ALIGN_OFFSET
], flags
=[CAN_ELIMINATE
, CAN_REORDER
])
892 # V3D-specific instrinc for tile buffer color reads.
894 # The hardware requires that we read the samples and components of a pixel
895 # in order, so we cannot eliminate or remove any loads in a sequence.
897 # src[] = { render_target }
898 # BASE = sample index
899 load("tlb_color_v3d", 1, [BASE
, COMPONENT
], [])
901 # V3D-specific instrinc for per-sample tile buffer color writes.
903 # The driver backend needs to identify per-sample color writes and emit
904 # specific code for them.
906 # src[] = { value, render_target }
907 # BASE = sample index
908 store("tlb_sample_color_v3d", 2, [BASE
, COMPONENT
, TYPE
], [])
910 # V3D-specific intrinsic to load the number of layers attached to
911 # the target framebuffer
912 intrinsic("load_fb_layers_v3d", dest_comp
=1, flags
=[CAN_ELIMINATE
, CAN_REORDER
])