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
, num_variables
,
35 indices
, flags
, sysval
):
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 - num_variables: the number of variables
46 - indices: list of constant indicies
47 - flags: list of semantic flags
48 - sysval: is this a system-value intrinsic
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(num_variables
, int)
56 assert isinstance(indices
, list)
58 assert isinstance(indices
[0], str)
59 assert isinstance(flags
, list)
61 assert isinstance(flags
[0], str)
62 assert isinstance(sysval
, bool)
65 self
.num_srcs
= len(src_components
)
66 self
.src_components
= src_components
67 self
.has_dest
= (dest_components
>= 0)
68 self
.dest_components
= dest_components
69 self
.num_variables
= num_variables
70 self
.num_indices
= len(indices
)
71 self
.indices
= indices
79 # A constant 'base' value that is added to an offset src:
80 BASE
= "NIR_INTRINSIC_BASE"
81 # For store instructions, a writemask:
82 WRMASK
= "NIR_INTRINSIC_WRMASK"
83 # The stream-id for GS emit_vertex/end_primitive intrinsics:
84 STREAM_ID
= "NIR_INTRINSIC_STREAM_ID"
85 # The clip-plane id for load_user_clip_plane intrinsics:
86 UCP_ID
= "NIR_INTRINSIC_UCP_ID"
87 # The amount of data, starting from BASE, that this instruction
88 # may access. This is used to provide bounds if the offset is
90 RANGE
= "NIR_INTRINSIC_RANGE"
91 # The vulkan descriptor set binding for vulkan_resource_index
93 DESC_SET
= "NIR_INTRINSIC_DESC_SET"
94 # The vulkan descriptor set binding for vulkan_resource_index
96 BINDING
= "NIR_INTRINSIC_BINDING"
98 COMPONENT
= "NIR_INTRINSIC_COMPONENT"
99 # Interpolation mode (only meaningful for FS inputs)
100 INTERP_MODE
= "NIR_INTRINSIC_INTERP_MODE"
101 # A binary nir_op to use when performing a reduction or scan operation
102 REDUCTION_OP
= "NIR_INTRINSIC_REDUCTION_OP"
103 # Cluster size for reduction operations
104 CLUSTER_SIZE
= "NIR_INTRINSIC_CLUSTER_SIZE"
110 CAN_ELIMINATE
= "NIR_INTRINSIC_CAN_ELIMINATE"
111 CAN_REORDER
= "NIR_INTRINSIC_CAN_REORDER"
115 def intrinsic(name
, src_comp
=[], dest_comp
=-1, num_vars
=0, indices
=[],
116 flags
=[], sysval
=False):
117 assert name
not in INTR_OPCODES
118 INTR_OPCODES
[name
] = Intrinsic(name
, src_comp
, dest_comp
, num_vars
,
119 indices
, flags
, sysval
)
121 intrinsic("nop", flags
=[CAN_ELIMINATE
])
123 intrinsic("load_var", dest_comp
=0, num_vars
=1, flags
=[CAN_ELIMINATE
])
124 intrinsic("store_var", src_comp
=[0], num_vars
=1, indices
=[WRMASK
])
125 intrinsic("copy_var", num_vars
=2)
127 # Interpolation of input. The interp_var_at* intrinsics are similar to the
128 # load_var intrinsic acting on a shader input except that they interpolate
129 # the input differently. The at_sample and at_offset intrinsics take an
130 # additional source that is an integer sample id or a vec2 position offset
133 intrinsic("interp_var_at_centroid", dest_comp
=0, num_vars
=1,
134 flags
=[ CAN_ELIMINATE
, CAN_REORDER
])
135 intrinsic("interp_var_at_sample", src_comp
=[1], dest_comp
=0, num_vars
=1,
136 flags
=[CAN_ELIMINATE
, CAN_REORDER
])
137 intrinsic("interp_var_at_offset", src_comp
=[2], dest_comp
=0, num_vars
=1,
138 flags
=[CAN_ELIMINATE
, CAN_REORDER
])
140 # Ask the driver for the size of a given buffer. It takes the buffer index
142 intrinsic("get_buffer_size", src_comp
=[1], dest_comp
=1,
143 flags
=[CAN_ELIMINATE
, CAN_REORDER
])
145 # a barrier is an intrinsic with no inputs/outputs but which can't be moved
146 # around/optimized in general
153 # Memory barrier with semantics analogous to the memoryBarrier() GLSL
155 barrier("memory_barrier")
157 # Shader clock intrinsic with semantics analogous to the clock2x32ARB()
159 # The latter can be used as code motion barrier, which is currently not
161 intrinsic("shader_clock", dest_comp
=2, flags
=[CAN_ELIMINATE
])
163 # Shader ballot intrinsics with semantics analogous to the
166 # readInvocationARB()
167 # readFirstInvocationARB()
169 # GLSL functions from ARB_shader_ballot.
170 intrinsic("ballot", src_comp
=[1], dest_comp
=0, flags
=[CAN_ELIMINATE
])
171 intrinsic("read_invocation", src_comp
=[0, 1], dest_comp
=0, flags
=[CAN_ELIMINATE
])
172 intrinsic("read_first_invocation", src_comp
=[0], dest_comp
=0, flags
=[CAN_ELIMINATE
])
174 # Additional SPIR-V ballot intrinsics
176 # These correspond to the SPIR-V opcodes
178 # OpGroupUniformElect
179 # OpSubgroupFirstInvocationKHR
180 intrinsic("elect", dest_comp
=1, flags
=[CAN_ELIMINATE
])
181 intrinsic("first_invocation", dest_comp
=1, flags
=[CAN_ELIMINATE
])
183 # Memory barrier with semantics analogous to the compute shader
184 # groupMemoryBarrier(), memoryBarrierAtomicCounter(), memoryBarrierBuffer(),
185 # memoryBarrierImage() and memoryBarrierShared() GLSL intrinsics.
186 barrier("group_memory_barrier")
187 barrier("memory_barrier_atomic_counter")
188 barrier("memory_barrier_buffer")
189 barrier("memory_barrier_image")
190 barrier("memory_barrier_shared")
191 barrier("begin_invocation_interlock")
192 barrier("end_invocation_interlock")
194 # A conditional discard, with a single boolean source.
195 intrinsic("discard_if", src_comp
=[1])
197 # ARB_shader_group_vote intrinsics
198 intrinsic("vote_any", src_comp
=[1], dest_comp
=1, flags
=[CAN_ELIMINATE
])
199 intrinsic("vote_all", src_comp
=[1], dest_comp
=1, flags
=[CAN_ELIMINATE
])
200 intrinsic("vote_feq", src_comp
=[0], dest_comp
=1, flags
=[CAN_ELIMINATE
])
201 intrinsic("vote_ieq", src_comp
=[0], dest_comp
=1, flags
=[CAN_ELIMINATE
])
203 # Ballot ALU operations from SPIR-V.
205 # These operations work like their ALU counterparts except that the operate
206 # on a uvec4 which is treated as a 128bit integer. Also, they are, in
207 # general, free to ignore any bits which are above the subgroup size.
208 intrinsic("ballot_bitfield_extract", src_comp
=[4, 1], dest_comp
=1, flags
=[CAN_ELIMINATE
])
209 intrinsic("ballot_bit_count_reduce", src_comp
=[4], dest_comp
=1, flags
=[CAN_ELIMINATE
])
210 intrinsic("ballot_bit_count_inclusive", src_comp
=[4], dest_comp
=1, flags
=[CAN_ELIMINATE
])
211 intrinsic("ballot_bit_count_exclusive", src_comp
=[4], dest_comp
=1, flags
=[CAN_ELIMINATE
])
212 intrinsic("ballot_find_lsb", src_comp
=[4], dest_comp
=1, flags
=[CAN_ELIMINATE
])
213 intrinsic("ballot_find_msb", src_comp
=[4], dest_comp
=1, flags
=[CAN_ELIMINATE
])
215 # Shuffle operations from SPIR-V.
216 intrinsic("shuffle", src_comp
=[0, 1], dest_comp
=0, flags
=[CAN_ELIMINATE
])
217 intrinsic("shuffle_xor", src_comp
=[0, 1], dest_comp
=0, flags
=[CAN_ELIMINATE
])
218 intrinsic("shuffle_up", src_comp
=[0, 1], dest_comp
=0, flags
=[CAN_ELIMINATE
])
219 intrinsic("shuffle_down", src_comp
=[0, 1], dest_comp
=0, flags
=[CAN_ELIMINATE
])
221 # Quad operations from SPIR-V.
222 intrinsic("quad_broadcast", src_comp
=[0, 1], dest_comp
=0, flags
=[CAN_ELIMINATE
])
223 intrinsic("quad_swap_horizontal", src_comp
=[0], dest_comp
=0, flags
=[CAN_ELIMINATE
])
224 intrinsic("quad_swap_vertical", src_comp
=[0], dest_comp
=0, flags
=[CAN_ELIMINATE
])
225 intrinsic("quad_swap_diagonal", src_comp
=[0], dest_comp
=0, flags
=[CAN_ELIMINATE
])
227 intrinsic("reduce", src_comp
=[0], dest_comp
=0, indices
=[REDUCTION_OP
, CLUSTER_SIZE
],
228 flags
=[CAN_ELIMINATE
])
229 intrinsic("inclusive_scan", src_comp
=[0], dest_comp
=0, indices
=[REDUCTION_OP
],
230 flags
=[CAN_ELIMINATE
])
231 intrinsic("exclusive_scan", src_comp
=[0], dest_comp
=0, indices
=[REDUCTION_OP
],
232 flags
=[CAN_ELIMINATE
])
234 # Basic Geometry Shader intrinsics.
236 # emit_vertex implements GLSL's EmitStreamVertex() built-in. It takes a single
237 # index, which is the stream ID to write to.
239 # end_primitive implements GLSL's EndPrimitive() built-in.
240 intrinsic("emit_vertex", indices
=[STREAM_ID
])
241 intrinsic("end_primitive", indices
=[STREAM_ID
])
243 # Geometry Shader intrinsics with a vertex count.
245 # Alternatively, drivers may implement these intrinsics, and use
246 # nir_lower_gs_intrinsics() to convert from the basic intrinsics.
248 # These maintain a count of the number of vertices emitted, as an additional
249 # unsigned integer source.
250 intrinsic("emit_vertex_with_counter", src_comp
=[1], indices
=[STREAM_ID
])
251 intrinsic("end_primitive_with_counter", src_comp
=[1], indices
=[STREAM_ID
])
252 intrinsic("set_vertex_count", src_comp
=[1])
256 # The *_var variants take an atomic_uint nir_variable, while the other,
257 # lowered, variants take a constant buffer index and register offset.
259 def atomic(name
, flags
=[]):
260 intrinsic(name
+ "_var", dest_comp
=1, num_vars
=1, flags
=flags
)
261 intrinsic(name
, src_comp
=[1], dest_comp
=1, indices
=[BASE
], flags
=flags
)
264 intrinsic(name
+ "_var", src_comp
=[1], dest_comp
=1, num_vars
=1)
265 intrinsic(name
, src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
268 intrinsic(name
+ "_var", src_comp
=[1, 1], dest_comp
=1, num_vars
=1)
269 intrinsic(name
, src_comp
=[1, 1, 1], dest_comp
=1, indices
=[BASE
])
271 atomic("atomic_counter_inc")
272 atomic("atomic_counter_dec")
273 atomic("atomic_counter_read", flags
=[CAN_ELIMINATE
])
274 atomic2("atomic_counter_add")
275 atomic2("atomic_counter_min")
276 atomic2("atomic_counter_max")
277 atomic2("atomic_counter_and")
278 atomic2("atomic_counter_or")
279 atomic2("atomic_counter_xor")
280 atomic2("atomic_counter_exchange")
281 atomic3("atomic_counter_comp_swap")
283 # Image load, store and atomic intrinsics.
285 # All image intrinsics take an image target passed as a nir_variable. Image
286 # variables contain a number of memory and layout qualifiers that influence
287 # the semantics of the intrinsic.
289 # All image intrinsics take a four-coordinate vector and a sample index as
290 # first two sources, determining the location within the image that will be
291 # accessed by the intrinsic. Components not applicable to the image target
292 # in use are undefined. Image store takes an additional four-component
293 # argument with the value to be written, and image atomic operations take
294 # either one or two additional scalar arguments with the same meaning as in
295 # the ARB_shader_image_load_store specification.
296 intrinsic("image_var_load", src_comp
=[4, 1], dest_comp
=4, num_vars
=1,
297 flags
=[CAN_ELIMINATE
])
298 intrinsic("image_var_store", src_comp
=[4, 1, 4], num_vars
=1)
299 intrinsic("image_var_atomic_add", src_comp
=[4, 1, 1], dest_comp
=1, num_vars
=1)
300 intrinsic("image_var_atomic_min", src_comp
=[4, 1, 1], dest_comp
=1, num_vars
=1)
301 intrinsic("image_var_atomic_max", src_comp
=[4, 1, 1], dest_comp
=1, num_vars
=1)
302 intrinsic("image_var_atomic_and", src_comp
=[4, 1, 1], dest_comp
=1, num_vars
=1)
303 intrinsic("image_var_atomic_or", src_comp
=[4, 1, 1], dest_comp
=1, num_vars
=1)
304 intrinsic("image_var_atomic_xor", src_comp
=[4, 1, 1], dest_comp
=1, num_vars
=1)
305 intrinsic("image_var_atomic_exchange", src_comp
=[4, 1, 1], dest_comp
=1, num_vars
=1)
306 intrinsic("image_var_atomic_comp_swap", src_comp
=[4, 1, 1, 1], dest_comp
=1, num_vars
=1)
307 intrinsic("image_var_size", dest_comp
=0, num_vars
=1, flags
=[CAN_ELIMINATE
, CAN_REORDER
])
308 intrinsic("image_var_samples", dest_comp
=1, num_vars
=1, flags
=[CAN_ELIMINATE
, CAN_REORDER
])
310 # Vulkan descriptor set intrinsics
312 # The Vulkan API uses a different binding model from GL. In the Vulkan
313 # API, all external resources are represented by a tuple:
315 # (descriptor set, binding, array index)
317 # where the array index is the only thing allowed to be indirect. The
318 # vulkan_surface_index intrinsic takes the descriptor set and binding as
319 # its first two indices and the array index as its source. The third
320 # index is a nir_variable_mode in case that's useful to the backend.
322 # The intended usage is that the shader will call vulkan_surface_index to
323 # get an index and then pass that as the buffer index ubo/ssbo calls.
325 # The vulkan_resource_reindex intrinsic takes a resource index in src0
326 # (the result of a vulkan_resource_index or vulkan_resource_reindex) which
327 # corresponds to the tuple (set, binding, index) and computes an index
328 # corresponding to tuple (set, binding, idx + src1).
329 intrinsic("vulkan_resource_index", src_comp
=[1], dest_comp
=1,
330 indices
=[DESC_SET
, BINDING
], flags
=[CAN_ELIMINATE
, CAN_REORDER
])
331 intrinsic("vulkan_resource_reindex", src_comp
=[1, 1], dest_comp
=1,
332 flags
=[CAN_ELIMINATE
, CAN_REORDER
])
334 # variable atomic intrinsics
336 # All of these variable atomic memory operations read a value from memory,
337 # compute a new value using one of the operations below, write the new value
338 # to memory, and return the original value read.
340 # All operations take 1 source except CompSwap that takes 2. These sources
343 # 0: The data parameter to the atomic function (i.e. the value to add
344 # in shared_atomic_add, etc).
345 # 1: For CompSwap only: the second data parameter.
347 # All operations take 1 variable deref.
348 intrinsic("var_atomic_add", src_comp
=[1], dest_comp
=1, num_vars
=1)
349 intrinsic("var_atomic_imin", src_comp
=[1], dest_comp
=1, num_vars
=1)
350 intrinsic("var_atomic_umin", src_comp
=[1], dest_comp
=1, num_vars
=1)
351 intrinsic("var_atomic_imax", src_comp
=[1], dest_comp
=1, num_vars
=1)
352 intrinsic("var_atomic_umax", src_comp
=[1], dest_comp
=1, num_vars
=1)
353 intrinsic("var_atomic_and", src_comp
=[1], dest_comp
=1, num_vars
=1)
354 intrinsic("var_atomic_or", src_comp
=[1], dest_comp
=1, num_vars
=1)
355 intrinsic("var_atomic_xor", src_comp
=[1], dest_comp
=1, num_vars
=1)
356 intrinsic("var_atomic_exchange", src_comp
=[1], dest_comp
=1, num_vars
=1)
357 intrinsic("var_atomic_comp_swap", src_comp
=[1, 1], dest_comp
=1, num_vars
=1)
359 # SSBO atomic intrinsics
361 # All of the SSBO atomic memory operations read a value from memory,
362 # compute a new value using one of the operations below, write the new
363 # value to memory, and return the original value read.
365 # All operations take 3 sources except CompSwap that takes 4. These
368 # 0: The SSBO buffer index.
369 # 1: The offset into the SSBO buffer of the variable that the atomic
370 # operation will operate on.
371 # 2: The data parameter to the atomic function (i.e. the value to add
372 # in ssbo_atomic_add, etc).
373 # 3: For CompSwap only: the second data parameter.
374 intrinsic("ssbo_atomic_add", src_comp
=[1, 1, 1], dest_comp
=1)
375 intrinsic("ssbo_atomic_imin", src_comp
=[1, 1, 1], dest_comp
=1)
376 intrinsic("ssbo_atomic_umin", src_comp
=[1, 1, 1], dest_comp
=1)
377 intrinsic("ssbo_atomic_imax", src_comp
=[1, 1, 1], dest_comp
=1)
378 intrinsic("ssbo_atomic_umax", src_comp
=[1, 1, 1], dest_comp
=1)
379 intrinsic("ssbo_atomic_and", src_comp
=[1, 1, 1], dest_comp
=1)
380 intrinsic("ssbo_atomic_or", src_comp
=[1, 1, 1], dest_comp
=1)
381 intrinsic("ssbo_atomic_xor", src_comp
=[1, 1, 1], dest_comp
=1)
382 intrinsic("ssbo_atomic_exchange", src_comp
=[1, 1, 1], dest_comp
=1)
383 intrinsic("ssbo_atomic_comp_swap", src_comp
=[1, 1, 1, 1], dest_comp
=1)
385 # CS shared variable atomic intrinsics
387 # All of the shared variable atomic memory operations read a value from
388 # memory, compute a new value using one of the operations below, write the
389 # new value to memory, and return the original value read.
391 # All operations take 2 sources except CompSwap that takes 3. These
394 # 0: The offset into the shared variable storage region that the atomic
395 # operation will operate on.
396 # 1: The data parameter to the atomic function (i.e. the value to add
397 # in shared_atomic_add, etc).
398 # 2: For CompSwap only: the second data parameter.
399 intrinsic("shared_atomic_add", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
400 intrinsic("shared_atomic_imin", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
401 intrinsic("shared_atomic_umin", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
402 intrinsic("shared_atomic_imax", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
403 intrinsic("shared_atomic_umax", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
404 intrinsic("shared_atomic_and", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
405 intrinsic("shared_atomic_or", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
406 intrinsic("shared_atomic_xor", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
407 intrinsic("shared_atomic_exchange", src_comp
=[1, 1], dest_comp
=1, indices
=[BASE
])
408 intrinsic("shared_atomic_comp_swap", src_comp
=[1, 1, 1], dest_comp
=1, indices
=[BASE
])
410 def system_value(name
, dest_comp
, indices
=[]):
411 intrinsic("load_" + name
, [], dest_comp
, 0, indices
,
412 flags
=[CAN_ELIMINATE
, CAN_REORDER
], sysval
=True)
414 system_value("frag_coord", 4)
415 system_value("front_face", 1)
416 system_value("vertex_id", 1)
417 system_value("vertex_id_zero_base", 1)
418 system_value("first_vertex", 1)
419 system_value("is_indexed_draw", 1)
420 system_value("base_vertex", 1)
421 system_value("instance_id", 1)
422 system_value("base_instance", 1)
423 system_value("draw_id", 1)
424 system_value("sample_id", 1)
425 system_value("sample_pos", 2)
426 system_value("sample_mask_in", 1)
427 system_value("primitive_id", 1)
428 system_value("invocation_id", 1)
429 system_value("tess_coord", 3)
430 system_value("tess_level_outer", 4)
431 system_value("tess_level_inner", 2)
432 system_value("patch_vertices_in", 1)
433 system_value("local_invocation_id", 3)
434 system_value("local_invocation_index", 1)
435 system_value("work_group_id", 3)
436 system_value("user_clip_plane", 4, indices
=[UCP_ID
])
437 system_value("num_work_groups", 3)
438 system_value("helper_invocation", 1)
439 system_value("alpha_ref_float", 1)
440 system_value("layer_id", 1)
441 system_value("view_index", 1)
442 system_value("subgroup_size", 1)
443 system_value("subgroup_invocation", 1)
444 system_value("subgroup_eq_mask", 0)
445 system_value("subgroup_ge_mask", 0)
446 system_value("subgroup_gt_mask", 0)
447 system_value("subgroup_le_mask", 0)
448 system_value("subgroup_lt_mask", 0)
449 system_value("num_subgroups", 1)
450 system_value("subgroup_id", 1)
451 system_value("local_group_size", 3)
453 # Blend constant color values. Float values are clamped.#
454 system_value("blend_const_color_r_float", 1)
455 system_value("blend_const_color_g_float", 1)
456 system_value("blend_const_color_b_float", 1)
457 system_value("blend_const_color_a_float", 1)
458 system_value("blend_const_color_rgba8888_unorm", 1)
459 system_value("blend_const_color_aaaa8888_unorm", 1)
461 # Barycentric coordinate intrinsics.
463 # These set up the barycentric coordinates for a particular interpolation.
464 # The first three are for the simple cases: pixel, centroid, or per-sample
465 # (at gl_SampleID). The next two handle interpolating at a specified
466 # sample location, or interpolating with a vec2 offset,
468 # The interp_mode index should be either the INTERP_MODE_SMOOTH or
469 # INTERP_MODE_NOPERSPECTIVE enum values.
471 # The vec2 value produced by these intrinsics is intended for use as the
472 # barycoord source of a load_interpolated_input intrinsic.
474 def barycentric(name
, src_comp
=[]):
475 intrinsic("load_barycentric_" + name
, src_comp
=src_comp
, dest_comp
=2,
476 indices
=[INTERP_MODE
], flags
=[CAN_ELIMINATE
, CAN_REORDER
])
478 # no sources. const_index[] = { interp_mode }
480 barycentric("centroid")
481 barycentric("sample")
482 # src[] = { sample_id }. const_index[] = { interp_mode }
483 barycentric("at_sample", [1])
484 # src[] = { offset.xy }. const_index[] = { interp_mode }
485 barycentric("at_offset", [2])
487 # Load operations pull data from some piece of GPU memory. All load
488 # operations operate in terms of offsets into some piece of theoretical
489 # memory. Loads from externally visible memory (UBO and SSBO) simply take a
490 # byte offset as a source. Loads from opaque memory (uniforms, inputs, etc.)
491 # take a base+offset pair where the base (const_index[0]) gives the location
492 # of the start of the variable being loaded and and the offset source is a
493 # offset into that variable.
495 # Uniform load operations have a second "range" index that specifies the
496 # range (starting at base) of the data from which we are loading. If
497 # const_index[1] == 0, then the range is unknown.
499 # Some load operations such as UBO/SSBO load and per_vertex loads take an
500 # additional source to specify which UBO/SSBO/vertex to load from.
502 # The exact address type depends on the lowering pass that generates the
503 # load/store intrinsics. Typically, this is vec4 units for things such as
504 # varying slots and float units for fragment shader inputs. UBO and SSBO
505 # offsets are always in bytes.
507 def load(name
, num_srcs
, indices
=[], flags
=[]):
508 intrinsic("load_" + name
, [1] * num_srcs
, dest_comp
=0, indices
=indices
,
511 # src[] = { offset }. const_index[] = { base, range }
512 load("uniform", 1, [BASE
, RANGE
], [CAN_ELIMINATE
, CAN_REORDER
])
513 # src[] = { buffer_index, offset }. No const_index
514 load("ubo", 2, flags
=[CAN_ELIMINATE
, CAN_REORDER
])
515 # src[] = { offset }. const_index[] = { base, component }
516 load("input", 1, [BASE
, COMPONENT
], [CAN_ELIMINATE
, CAN_REORDER
])
517 # src[] = { vertex, offset }. const_index[] = { base, component }
518 load("per_vertex_input", 2, [BASE
, COMPONENT
], [CAN_ELIMINATE
, CAN_REORDER
])
519 # src[] = { barycoord, offset }. const_index[] = { base, component }
520 intrinsic("load_interpolated_input", src_comp
=[2, 1], dest_comp
=0,
521 indices
=[BASE
, COMPONENT
], flags
=[CAN_ELIMINATE
, CAN_REORDER
])
523 # src[] = { buffer_index, offset }. No const_index
524 load("ssbo", 2, flags
=[CAN_ELIMINATE
])
525 # src[] = { offset }. const_index[] = { base, component }
526 load("output", 1, [BASE
, COMPONENT
], flags
=[CAN_ELIMINATE
])
527 # src[] = { vertex, offset }. const_index[] = { base }
528 load("per_vertex_output", 2, [BASE
, COMPONENT
], [CAN_ELIMINATE
])
529 # src[] = { offset }. const_index[] = { base }
530 load("shared", 1, [BASE
], [CAN_ELIMINATE
])
531 # src[] = { offset }. const_index[] = { base, range }
532 load("push_constant", 1, [BASE
, RANGE
], [CAN_ELIMINATE
, CAN_REORDER
])
534 # Stores work the same way as loads, except now the first source is the value
535 # to store and the second (and possibly third) source specify where to store
536 # the value. SSBO and shared memory stores also have a write mask as
539 def store(name
, num_srcs
, indices
=[], flags
=[]):
540 intrinsic("store_" + name
, [0] + ([1] * (num_srcs
- 1)), indices
=indices
, flags
=flags
)
542 # src[] = { value, offset }. const_index[] = { base, write_mask, component }
543 store("output", 2, [BASE
, WRMASK
, COMPONENT
])
544 # src[] = { value, vertex, offset }.
545 # const_index[] = { base, write_mask, component }
546 store("per_vertex_output", 3, [BASE
, WRMASK
, COMPONENT
])
547 # src[] = { value, block_index, offset }. const_index[] = { write_mask }
548 store("ssbo", 3, [WRMASK
])
549 # src[] = { value, offset }. const_index[] = { base, write_mask }
550 store("shared", 2, [BASE
, WRMASK
])