2 * Copyright © 2014 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Connor Abbott (cwabbott0@gmail.com)
29 * This header file defines all the available intrinsics in one place. It
30 * expands to a list of macros of the form:
32 * INTRINSIC(name, num_srcs, src_components, has_dest, dest_components,
33 * num_variables, num_indices, idx0, idx1, idx2, flags)
35 * Which should correspond one-to-one with the nir_intrinsic_info structure. It
36 * is included in both ir.h to create the nir_intrinsic enum (with members of
37 * the form nir_intrinsic_(name)) and and in opcodes.c to create
38 * nir_intrinsic_infos, which is a const array of nir_intrinsic_info structures
42 #define ARR(...) { __VA_ARGS__ }
45 INTRINSIC(load_var
, 0, ARR(), true, 0, 1, 0, xx
, xx
, xx
, NIR_INTRINSIC_CAN_ELIMINATE
)
46 INTRINSIC(store_var
, 1, ARR(0), false, 0, 1, 1, WRMASK
, xx
, xx
, 0)
47 INTRINSIC(copy_var
, 0, ARR(), false, 0, 2, 0, xx
, xx
, xx
, 0)
50 * Interpolation of input. The interp_var_at* intrinsics are similar to the
51 * load_var intrinsic acting an a shader input except that they interpolate
52 * the input differently. The at_sample and at_offset intrinsics take an
53 * aditional source that is a integer sample id or a vec2 position offset
57 INTRINSIC(interp_var_at_centroid
, 0, ARR(0), true, 0, 1, 0, xx
, xx
, xx
,
58 NIR_INTRINSIC_CAN_ELIMINATE
| NIR_INTRINSIC_CAN_REORDER
)
59 INTRINSIC(interp_var_at_sample
, 1, ARR(1), true, 0, 1, 0, xx
, xx
, xx
,
60 NIR_INTRINSIC_CAN_ELIMINATE
| NIR_INTRINSIC_CAN_REORDER
)
61 INTRINSIC(interp_var_at_offset
, 1, ARR(2), true, 0, 1, 0, xx
, xx
, xx
,
62 NIR_INTRINSIC_CAN_ELIMINATE
| NIR_INTRINSIC_CAN_REORDER
)
65 * Ask the driver for the size of a given buffer. It takes the buffer index
68 INTRINSIC(get_buffer_size
, 1, ARR(1), true, 1, 0, 0, xx
, xx
, xx
,
69 NIR_INTRINSIC_CAN_ELIMINATE
| NIR_INTRINSIC_CAN_REORDER
)
72 * a barrier is an intrinsic with no inputs/outputs but which can't be moved
73 * around/optimized in general
75 #define BARRIER(name) INTRINSIC(name, 0, ARR(), false, 0, 0, 0, xx, xx, xx, 0)
81 * Memory barrier with semantics analogous to the memoryBarrier() GLSL
84 BARRIER(memory_barrier
)
87 * Shader clock intrinsic with semantics analogous to the clock2x32ARB()
89 * The latter can be used as code motion barrier, which is currently not
92 INTRINSIC(shader_clock
, 0, ARR(), true, 1, 0, 0, xx
, xx
, xx
, NIR_INTRINSIC_CAN_ELIMINATE
)
95 * Memory barrier with semantics analogous to the compute shader
96 * groupMemoryBarrier(), memoryBarrierAtomicCounter(), memoryBarrierBuffer(),
97 * memoryBarrierImage() and memoryBarrierShared() GLSL intrinsics.
99 BARRIER(group_memory_barrier
)
100 BARRIER(memory_barrier_atomic_counter
)
101 BARRIER(memory_barrier_buffer
)
102 BARRIER(memory_barrier_image
)
103 BARRIER(memory_barrier_shared
)
105 /** A conditional discard, with a single boolean source. */
106 INTRINSIC(discard_if
, 1, ARR(1), false, 0, 0, 0, xx
, xx
, xx
, 0)
109 * Basic Geometry Shader intrinsics.
111 * emit_vertex implements GLSL's EmitStreamVertex() built-in. It takes a single
112 * index, which is the stream ID to write to.
114 * end_primitive implements GLSL's EndPrimitive() built-in.
116 INTRINSIC(emit_vertex
, 0, ARR(), false, 0, 0, 1, STREAM_ID
, xx
, xx
, 0)
117 INTRINSIC(end_primitive
, 0, ARR(), false, 0, 0, 1, STREAM_ID
, xx
, xx
, 0)
120 * Geometry Shader intrinsics with a vertex count.
122 * Alternatively, drivers may implement these intrinsics, and use
123 * nir_lower_gs_intrinsics() to convert from the basic intrinsics.
125 * These maintain a count of the number of vertices emitted, as an additional
126 * unsigned integer source.
128 INTRINSIC(emit_vertex_with_counter
, 1, ARR(1), false, 0, 0, 1, STREAM_ID
, xx
, xx
, 0)
129 INTRINSIC(end_primitive_with_counter
, 1, ARR(1), false, 0, 0, 1, STREAM_ID
, xx
, xx
, 0)
130 INTRINSIC(set_vertex_count
, 1, ARR(1), false, 0, 0, 0, xx
, xx
, xx
, 0)
135 * The *_var variants take an atomic_uint nir_variable, while the other,
136 * lowered, variants take a constant buffer index and register offset.
139 #define ATOMIC(name, flags) \
140 INTRINSIC(atomic_counter_##name##_var, 0, ARR(), true, 1, 1, 0, xx, xx, xx, flags) \
141 INTRINSIC(atomic_counter_##name, 1, ARR(1), true, 1, 0, 1, BASE, xx, xx, flags)
145 ATOMIC(read
, NIR_INTRINSIC_CAN_ELIMINATE
)
148 * Image load, store and atomic intrinsics.
150 * All image intrinsics take an image target passed as a nir_variable. Image
151 * variables contain a number of memory and layout qualifiers that influence
152 * the semantics of the intrinsic.
154 * All image intrinsics take a four-coordinate vector and a sample index as
155 * first two sources, determining the location within the image that will be
156 * accessed by the intrinsic. Components not applicable to the image target
157 * in use are undefined. Image store takes an additional four-component
158 * argument with the value to be written, and image atomic operations take
159 * either one or two additional scalar arguments with the same meaning as in
160 * the ARB_shader_image_load_store specification.
162 INTRINSIC(image_load
, 2, ARR(4, 1), true, 4, 1, 0, xx
, xx
, xx
,
163 NIR_INTRINSIC_CAN_ELIMINATE
)
164 INTRINSIC(image_store
, 3, ARR(4, 1, 4), false, 0, 1, 0, xx
, xx
, xx
, 0)
165 INTRINSIC(image_atomic_add
, 3, ARR(4, 1, 1), true, 1, 1, 0, xx
, xx
, xx
, 0)
166 INTRINSIC(image_atomic_min
, 3, ARR(4, 1, 1), true, 1, 1, 0, xx
, xx
, xx
, 0)
167 INTRINSIC(image_atomic_max
, 3, ARR(4, 1, 1), true, 1, 1, 0, xx
, xx
, xx
, 0)
168 INTRINSIC(image_atomic_and
, 3, ARR(4, 1, 1), true, 1, 1, 0, xx
, xx
, xx
, 0)
169 INTRINSIC(image_atomic_or
, 3, ARR(4, 1, 1), true, 1, 1, 0, xx
, xx
, xx
, 0)
170 INTRINSIC(image_atomic_xor
, 3, ARR(4, 1, 1), true, 1, 1, 0, xx
, xx
, xx
, 0)
171 INTRINSIC(image_atomic_exchange
, 3, ARR(4, 1, 1), true, 1, 1, 0, xx
, xx
, xx
, 0)
172 INTRINSIC(image_atomic_comp_swap
, 4, ARR(4, 1, 1, 1), true, 1, 1, 0, xx
, xx
, xx
, 0)
173 INTRINSIC(image_size
, 0, ARR(), true, 4, 1, 0, xx
, xx
, xx
,
174 NIR_INTRINSIC_CAN_ELIMINATE
| NIR_INTRINSIC_CAN_REORDER
)
175 INTRINSIC(image_samples
, 0, ARR(), true, 1, 1, 0, xx
, xx
, xx
,
176 NIR_INTRINSIC_CAN_ELIMINATE
| NIR_INTRINSIC_CAN_REORDER
)
179 * variable atomic intrinsics
181 * All of these variable atomic memory operations read a value from memory,
182 * compute a new value using one of the operations below, write the new value
183 * to memory, and return the original value read.
185 * All operations take 1 source except CompSwap that takes 2. These sources
188 * 0: The data parameter to the atomic function (i.e. the value to add
189 * in shared_atomic_add, etc).
190 * 1: For CompSwap only: the second data parameter.
192 * All operations take 1 variable deref.
194 INTRINSIC(var_atomic_add
, 1, ARR(1), true, 1, 1, 0, xx
, xx
, xx
, 0)
195 INTRINSIC(var_atomic_imin
, 1, ARR(1), true, 1, 1, 0, xx
, xx
, xx
, 0)
196 INTRINSIC(var_atomic_umin
, 1, ARR(1), true, 1, 1, 0, xx
, xx
, xx
, 0)
197 INTRINSIC(var_atomic_imax
, 1, ARR(1), true, 1, 1, 0, xx
, xx
, xx
, 0)
198 INTRINSIC(var_atomic_umax
, 1, ARR(1), true, 1, 1, 0, xx
, xx
, xx
, 0)
199 INTRINSIC(var_atomic_and
, 1, ARR(1), true, 1, 1, 0, xx
, xx
, xx
, 0)
200 INTRINSIC(var_atomic_or
, 1, ARR(1), true, 1, 1, 0, xx
, xx
, xx
, 0)
201 INTRINSIC(var_atomic_xor
, 1, ARR(1), true, 1, 1, 0, xx
, xx
, xx
, 0)
202 INTRINSIC(var_atomic_exchange
, 1, ARR(1), true, 1, 1, 0, xx
, xx
, xx
, 0)
203 INTRINSIC(var_atomic_comp_swap
, 2, ARR(1, 1), true, 1, 1, 0, xx
, xx
, xx
, 0)
206 * SSBO atomic intrinsics
208 * All of the SSBO atomic memory operations read a value from memory,
209 * compute a new value using one of the operations below, write the new
210 * value to memory, and return the original value read.
212 * All operations take 3 sources except CompSwap that takes 4. These
215 * 0: The SSBO buffer index.
216 * 1: The offset into the SSBO buffer of the variable that the atomic
217 * operation will operate on.
218 * 2: The data parameter to the atomic function (i.e. the value to add
219 * in ssbo_atomic_add, etc).
220 * 3: For CompSwap only: the second data parameter.
222 INTRINSIC(ssbo_atomic_add
, 3, ARR(1, 1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
223 INTRINSIC(ssbo_atomic_imin
, 3, ARR(1, 1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
224 INTRINSIC(ssbo_atomic_umin
, 3, ARR(1, 1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
225 INTRINSIC(ssbo_atomic_imax
, 3, ARR(1, 1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
226 INTRINSIC(ssbo_atomic_umax
, 3, ARR(1, 1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
227 INTRINSIC(ssbo_atomic_and
, 3, ARR(1, 1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
228 INTRINSIC(ssbo_atomic_or
, 3, ARR(1, 1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
229 INTRINSIC(ssbo_atomic_xor
, 3, ARR(1, 1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
230 INTRINSIC(ssbo_atomic_exchange
, 3, ARR(1, 1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
231 INTRINSIC(ssbo_atomic_comp_swap
, 4, ARR(1, 1, 1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
234 * CS shared variable atomic intrinsics
236 * All of the shared variable atomic memory operations read a value from
237 * memory, compute a new value using one of the operations below, write the
238 * new value to memory, and return the original value read.
240 * All operations take 2 sources except CompSwap that takes 3. These
243 * 0: The offset into the shared variable storage region that the atomic
244 * operation will operate on.
245 * 1: The data parameter to the atomic function (i.e. the value to add
246 * in shared_atomic_add, etc).
247 * 2: For CompSwap only: the second data parameter.
249 INTRINSIC(shared_atomic_add
, 2, ARR(1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
250 INTRINSIC(shared_atomic_imin
, 2, ARR(1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
251 INTRINSIC(shared_atomic_umin
, 2, ARR(1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
252 INTRINSIC(shared_atomic_imax
, 2, ARR(1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
253 INTRINSIC(shared_atomic_umax
, 2, ARR(1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
254 INTRINSIC(shared_atomic_and
, 2, ARR(1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
255 INTRINSIC(shared_atomic_or
, 2, ARR(1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
256 INTRINSIC(shared_atomic_xor
, 2, ARR(1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
257 INTRINSIC(shared_atomic_exchange
, 2, ARR(1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
258 INTRINSIC(shared_atomic_comp_swap
, 3, ARR(1, 1, 1), true, 1, 0, 0, xx
, xx
, xx
, 0)
260 #define SYSTEM_VALUE(name, components, num_indices, idx0, idx1, idx2) \
261 INTRINSIC(load_##name, 0, ARR(), true, components, 0, num_indices, \
263 NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
265 SYSTEM_VALUE(front_face
, 1, 0, xx
, xx
, xx
)
266 SYSTEM_VALUE(vertex_id
, 1, 0, xx
, xx
, xx
)
267 SYSTEM_VALUE(vertex_id_zero_base
, 1, 0, xx
, xx
, xx
)
268 SYSTEM_VALUE(base_vertex
, 1, 0, xx
, xx
, xx
)
269 SYSTEM_VALUE(instance_id
, 1, 0, xx
, xx
, xx
)
270 SYSTEM_VALUE(base_instance
, 1, 0, xx
, xx
, xx
)
271 SYSTEM_VALUE(draw_id
, 1, 0, xx
, xx
, xx
)
272 SYSTEM_VALUE(sample_id
, 1, 0, xx
, xx
, xx
)
273 SYSTEM_VALUE(sample_pos
, 2, 0, xx
, xx
, xx
)
274 SYSTEM_VALUE(sample_mask_in
, 1, 0, xx
, xx
, xx
)
275 SYSTEM_VALUE(primitive_id
, 1, 0, xx
, xx
, xx
)
276 SYSTEM_VALUE(invocation_id
, 1, 0, xx
, xx
, xx
)
277 SYSTEM_VALUE(tess_coord
, 3, 0, xx
, xx
, xx
)
278 SYSTEM_VALUE(tess_level_outer
, 4, 0, xx
, xx
, xx
)
279 SYSTEM_VALUE(tess_level_inner
, 2, 0, xx
, xx
, xx
)
280 SYSTEM_VALUE(patch_vertices_in
, 1, 0, xx
, xx
, xx
)
281 SYSTEM_VALUE(local_invocation_id
, 3, 0, xx
, xx
, xx
)
282 SYSTEM_VALUE(work_group_id
, 3, 0, xx
, xx
, xx
)
283 SYSTEM_VALUE(user_clip_plane
, 4, 1, UCP_ID
, xx
, xx
)
284 SYSTEM_VALUE(num_work_groups
, 3, 0, xx
, xx
, xx
)
285 SYSTEM_VALUE(helper_invocation
, 1, 0, xx
, xx
, xx
)
288 * Load operations pull data from some piece of GPU memory. All load
289 * operations operate in terms of offsets into some piece of theoretical
290 * memory. Loads from externally visible memory (UBO and SSBO) simply take a
291 * byte offset as a source. Loads from opaque memory (uniforms, inputs, etc.)
292 * take a base+offset pair where the base (const_index[0]) gives the location
293 * of the start of the variable being loaded and and the offset source is a
294 * offset into that variable.
296 * Some load operations such as UBO/SSBO load and per_vertex loads take an
297 * additional source to specify which UBO/SSBO/vertex to load from.
299 * The exact address type depends on the lowering pass that generates the
300 * load/store intrinsics. Typically, this is vec4 units for things such as
301 * varying slots and float units for fragment shader inputs. UBO and SSBO
302 * offsets are always in bytes.
305 #define LOAD(name, srcs, num_indices, idx0, idx1, idx2, flags) \
306 INTRINSIC(load_##name, srcs, ARR(1, 1, 1, 1), true, 0, 0, num_indices, idx0, idx1, idx2, flags)
308 /* src[] = { offset }. const_index[] = { base } */
309 LOAD(uniform
, 1, 1, BASE
, xx
, xx
, NIR_INTRINSIC_CAN_ELIMINATE
| NIR_INTRINSIC_CAN_REORDER
)
310 /* src[] = { buffer_index, offset }. No const_index */
311 LOAD(ubo
, 2, 0, xx
, xx
, xx
, NIR_INTRINSIC_CAN_ELIMINATE
| NIR_INTRINSIC_CAN_REORDER
)
312 /* src[] = { offset }. const_index[] = { base } */
313 LOAD(input
, 1, 1, BASE
, xx
, xx
, NIR_INTRINSIC_CAN_ELIMINATE
| NIR_INTRINSIC_CAN_REORDER
)
314 /* src[] = { vertex, offset }. const_index[] = { base } */
315 LOAD(per_vertex_input
, 2, 1, BASE
, xx
, xx
, NIR_INTRINSIC_CAN_ELIMINATE
| NIR_INTRINSIC_CAN_REORDER
)
316 /* src[] = { buffer_index, offset }. No const_index */
317 LOAD(ssbo
, 2, 0, xx
, xx
, xx
, NIR_INTRINSIC_CAN_ELIMINATE
)
318 /* src[] = { offset }. const_index[] = { base } */
319 LOAD(output
, 1, 1, BASE
, xx
, xx
, NIR_INTRINSIC_CAN_ELIMINATE
)
320 /* src[] = { vertex, offset }. const_index[] = { base } */
321 LOAD(per_vertex_output
, 2, 1, BASE
, xx
, xx
, NIR_INTRINSIC_CAN_ELIMINATE
)
322 /* src[] = { offset }. const_index[] = { base } */
323 LOAD(shared
, 1, 1, BASE
, xx
, xx
, NIR_INTRINSIC_CAN_ELIMINATE
)
326 * Stores work the same way as loads, except now the first source is the value
327 * to store and the second (and possibly third) source specify where to store
328 * the value. SSBO and shared memory stores also have a write mask as
332 #define STORE(name, srcs, num_indices, idx0, idx1, idx2, flags) \
333 INTRINSIC(store_##name, srcs, ARR(0, 1, 1, 1), false, 0, 0, num_indices, idx0, idx1, idx2, flags)
335 /* src[] = { value, offset }. const_index[] = { base, write_mask } */
336 STORE(output
, 2, 2, BASE
, WRMASK
, xx
, 0)
337 /* src[] = { value, vertex, offset }. const_index[] = { base, write_mask } */
338 STORE(per_vertex_output
, 3, 2, BASE
, WRMASK
, xx
, 0)
339 /* src[] = { value, block_index, offset }. const_index[] = { write_mask } */
340 STORE(ssbo
, 3, 1, WRMASK
, xx
, xx
, 0)
341 /* src[] = { value, offset }. const_index[] = { base, write_mask } */
342 STORE(shared
, 2, 2, BASE
, WRMASK
, xx
, 0)
344 LAST_INTRINSIC(store_shared
)