nir_intrinsics.h revision 60a27ad122128145d28be37e9c0b0bc86a8e5181
1/* 2 * Copyright © 2014 Intel Corporation 3 * 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: 10 * 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 13 * Software. 14 * 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 21 * IN THE SOFTWARE. 22 * 23 * Authors: 24 * Connor Abbott (cwabbott0@gmail.com) 25 * 26 */ 27 28/** 29 * This header file defines all the available intrinsics in one place. It 30 * expands to a list of macros of the form: 31 * 32 * INTRINSIC(name, num_srcs, src_components, has_dest, dest_components, 33 * num_variables, num_indices, idx0, idx1, idx2, flags) 34 * 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 39 * for each intrinsic. 40 */ 41 42#define ARR(...) { __VA_ARGS__ } 43 44 45INTRINSIC(load_var, 0, ARR(0), true, 0, 1, 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE) 46INTRINSIC(store_var, 1, ARR(0), false, 0, 1, 1, WRMASK, xx, xx, 0) 47INTRINSIC(copy_var, 0, ARR(0), false, 0, 2, 0, xx, xx, xx, 0) 48 49/* 50 * Interpolation of input. The interp_var_at* intrinsics are similar to the 51 * load_var intrinsic acting on a shader input except that they interpolate 52 * the input differently. The at_sample and at_offset intrinsics take an 53 * additional source that is an integer sample id or a vec2 position offset 54 * respectively. 55 */ 56 57INTRINSIC(interp_var_at_centroid, 0, ARR(0), true, 0, 1, 0, xx, xx, xx, 58 NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) 59INTRINSIC(interp_var_at_sample, 1, ARR(1), true, 0, 1, 0, xx, xx, xx, 60 NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) 61INTRINSIC(interp_var_at_offset, 1, ARR(2), true, 0, 1, 0, xx, xx, xx, 62 NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) 63 64/* 65 * Ask the driver for the size of a given buffer. It takes the buffer index 66 * as source. 67 */ 68INTRINSIC(get_buffer_size, 1, ARR(1), true, 1, 0, 0, xx, xx, xx, 69 NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) 70 71/* 72 * a barrier is an intrinsic with no inputs/outputs but which can't be moved 73 * around/optimized in general 74 */ 75#define BARRIER(name) INTRINSIC(name, 0, ARR(0), false, 0, 0, 0, xx, xx, xx, 0) 76 77BARRIER(barrier) 78BARRIER(discard) 79 80/* 81 * Memory barrier with semantics analogous to the memoryBarrier() GLSL 82 * intrinsic. 83 */ 84BARRIER(memory_barrier) 85 86/* 87 * Shader clock intrinsic with semantics analogous to the clock2x32ARB() 88 * GLSL intrinsic. 89 * The latter can be used as code motion barrier, which is currently not 90 * feasible with NIR. 91 */ 92INTRINSIC(shader_clock, 0, ARR(0), true, 1, 0, 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE) 93 94/* 95 * Memory barrier with semantics analogous to the compute shader 96 * groupMemoryBarrier(), memoryBarrierAtomicCounter(), memoryBarrierBuffer(), 97 * memoryBarrierImage() and memoryBarrierShared() GLSL intrinsics. 98 */ 99BARRIER(group_memory_barrier) 100BARRIER(memory_barrier_atomic_counter) 101BARRIER(memory_barrier_buffer) 102BARRIER(memory_barrier_image) 103BARRIER(memory_barrier_shared) 104 105/** A conditional discard, with a single boolean source. */ 106INTRINSIC(discard_if, 1, ARR(1), false, 0, 0, 0, xx, xx, xx, 0) 107 108/** 109 * Basic Geometry Shader intrinsics. 110 * 111 * emit_vertex implements GLSL's EmitStreamVertex() built-in. It takes a single 112 * index, which is the stream ID to write to. 113 * 114 * end_primitive implements GLSL's EndPrimitive() built-in. 115 */ 116INTRINSIC(emit_vertex, 0, ARR(0), false, 0, 0, 1, STREAM_ID, xx, xx, 0) 117INTRINSIC(end_primitive, 0, ARR(0), false, 0, 0, 1, STREAM_ID, xx, xx, 0) 118 119/** 120 * Geometry Shader intrinsics with a vertex count. 121 * 122 * Alternatively, drivers may implement these intrinsics, and use 123 * nir_lower_gs_intrinsics() to convert from the basic intrinsics. 124 * 125 * These maintain a count of the number of vertices emitted, as an additional 126 * unsigned integer source. 127 */ 128INTRINSIC(emit_vertex_with_counter, 1, ARR(1), false, 0, 0, 1, STREAM_ID, xx, xx, 0) 129INTRINSIC(end_primitive_with_counter, 1, ARR(1), false, 0, 0, 1, STREAM_ID, xx, xx, 0) 130INTRINSIC(set_vertex_count, 1, ARR(1), false, 0, 0, 0, xx, xx, xx, 0) 131 132/* 133 * Atomic counters 134 * 135 * The *_var variants take an atomic_uint nir_variable, while the other, 136 * lowered, variants take a constant buffer index and register offset. 137 */ 138 139#define ATOMIC(name, flags) \ 140 INTRINSIC(atomic_counter_##name##_var, 0, ARR(0), true, 1, 1, 0, xx, xx, xx, flags) \ 141 INTRINSIC(atomic_counter_##name, 1, ARR(1), true, 1, 0, 1, BASE, xx, xx, flags) 142 143ATOMIC(inc, 0) 144ATOMIC(dec, 0) 145ATOMIC(read, NIR_INTRINSIC_CAN_ELIMINATE) 146 147/* 148 * Image load, store and atomic intrinsics. 149 * 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. 153 * 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. 161 */ 162INTRINSIC(image_load, 2, ARR(4, 1), true, 4, 1, 0, xx, xx, xx, 163 NIR_INTRINSIC_CAN_ELIMINATE) 164INTRINSIC(image_store, 3, ARR(4, 1, 4), false, 0, 1, 0, xx, xx, xx, 0) 165INTRINSIC(image_atomic_add, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0) 166INTRINSIC(image_atomic_min, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0) 167INTRINSIC(image_atomic_max, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0) 168INTRINSIC(image_atomic_and, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0) 169INTRINSIC(image_atomic_or, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0) 170INTRINSIC(image_atomic_xor, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0) 171INTRINSIC(image_atomic_exchange, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0) 172INTRINSIC(image_atomic_comp_swap, 4, ARR(4, 1, 1, 1), true, 1, 1, 0, xx, xx, xx, 0) 173INTRINSIC(image_size, 0, ARR(0), true, 4, 1, 0, xx, xx, xx, 174 NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) 175INTRINSIC(image_samples, 0, ARR(0), true, 1, 1, 0, xx, xx, xx, 176 NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) 177 178/* 179 * Vulkan descriptor set intrinsic 180 * 181 * The Vulkan API uses a different binding model from GL. In the Vulkan 182 * API, all external resources are represented by a tuple: 183 * 184 * (descriptor set, binding, array index) 185 * 186 * where the array index is the only thing allowed to be indirect. The 187 * vulkan_surface_index intrinsic takes the descriptor set and binding as 188 * its first two indices and the array index as its source. The third 189 * index is a nir_variable_mode in case that's useful to the backend. 190 * 191 * The intended usage is that the shader will call vulkan_surface_index to 192 * get an index and then pass that as the buffer index ubo/ssbo calls. 193 */ 194INTRINSIC(vulkan_resource_index, 1, ARR(1), true, 1, 0, 2, 195 DESC_SET, BINDING, xx, 196 NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) 197 198/* 199 * variable atomic intrinsics 200 * 201 * All of these variable atomic memory operations read a value from memory, 202 * compute a new value using one of the operations below, write the new value 203 * to memory, and return the original value read. 204 * 205 * All operations take 1 source except CompSwap that takes 2. These sources 206 * represent: 207 * 208 * 0: The data parameter to the atomic function (i.e. the value to add 209 * in shared_atomic_add, etc). 210 * 1: For CompSwap only: the second data parameter. 211 * 212 * All operations take 1 variable deref. 213 */ 214INTRINSIC(var_atomic_add, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0) 215INTRINSIC(var_atomic_imin, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0) 216INTRINSIC(var_atomic_umin, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0) 217INTRINSIC(var_atomic_imax, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0) 218INTRINSIC(var_atomic_umax, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0) 219INTRINSIC(var_atomic_and, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0) 220INTRINSIC(var_atomic_or, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0) 221INTRINSIC(var_atomic_xor, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0) 222INTRINSIC(var_atomic_exchange, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0) 223INTRINSIC(var_atomic_comp_swap, 2, ARR(1, 1), true, 1, 1, 0, xx, xx, xx, 0) 224 225/* 226 * SSBO atomic intrinsics 227 * 228 * All of the SSBO atomic memory operations read a value from memory, 229 * compute a new value using one of the operations below, write the new 230 * value to memory, and return the original value read. 231 * 232 * All operations take 3 sources except CompSwap that takes 4. These 233 * sources represent: 234 * 235 * 0: The SSBO buffer index. 236 * 1: The offset into the SSBO buffer of the variable that the atomic 237 * operation will operate on. 238 * 2: The data parameter to the atomic function (i.e. the value to add 239 * in ssbo_atomic_add, etc). 240 * 3: For CompSwap only: the second data parameter. 241 */ 242INTRINSIC(ssbo_atomic_add, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0) 243INTRINSIC(ssbo_atomic_imin, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0) 244INTRINSIC(ssbo_atomic_umin, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0) 245INTRINSIC(ssbo_atomic_imax, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0) 246INTRINSIC(ssbo_atomic_umax, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0) 247INTRINSIC(ssbo_atomic_and, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0) 248INTRINSIC(ssbo_atomic_or, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0) 249INTRINSIC(ssbo_atomic_xor, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0) 250INTRINSIC(ssbo_atomic_exchange, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0) 251INTRINSIC(ssbo_atomic_comp_swap, 4, ARR(1, 1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0) 252 253/* 254 * CS shared variable atomic intrinsics 255 * 256 * All of the shared variable atomic memory operations read a value from 257 * memory, compute a new value using one of the operations below, write the 258 * new value to memory, and return the original value read. 259 * 260 * All operations take 2 sources except CompSwap that takes 3. These 261 * sources represent: 262 * 263 * 0: The offset into the shared variable storage region that the atomic 264 * operation will operate on. 265 * 1: The data parameter to the atomic function (i.e. the value to add 266 * in shared_atomic_add, etc). 267 * 2: For CompSwap only: the second data parameter. 268 */ 269INTRINSIC(shared_atomic_add, 2, ARR(1, 1), true, 1, 0, 0, xx, xx, xx, 0) 270INTRINSIC(shared_atomic_imin, 2, ARR(1, 1), true, 1, 0, 0, xx, xx, xx, 0) 271INTRINSIC(shared_atomic_umin, 2, ARR(1, 1), true, 1, 0, 0, xx, xx, xx, 0) 272INTRINSIC(shared_atomic_imax, 2, ARR(1, 1), true, 1, 0, 0, xx, xx, xx, 0) 273INTRINSIC(shared_atomic_umax, 2, ARR(1, 1), true, 1, 0, 0, xx, xx, xx, 0) 274INTRINSIC(shared_atomic_and, 2, ARR(1, 1), true, 1, 0, 0, xx, xx, xx, 0) 275INTRINSIC(shared_atomic_or, 2, ARR(1, 1), true, 1, 0, 0, xx, xx, xx, 0) 276INTRINSIC(shared_atomic_xor, 2, ARR(1, 1), true, 1, 0, 0, xx, xx, xx, 0) 277INTRINSIC(shared_atomic_exchange, 2, ARR(1, 1), true, 1, 0, 0, xx, xx, xx, 0) 278INTRINSIC(shared_atomic_comp_swap, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0) 279 280#define SYSTEM_VALUE(name, components, num_indices, idx0, idx1, idx2) \ 281 INTRINSIC(load_##name, 0, ARR(0), true, components, 0, num_indices, \ 282 idx0, idx1, idx2, \ 283 NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) 284 285SYSTEM_VALUE(front_face, 1, 0, xx, xx, xx) 286SYSTEM_VALUE(vertex_id, 1, 0, xx, xx, xx) 287SYSTEM_VALUE(vertex_id_zero_base, 1, 0, xx, xx, xx) 288SYSTEM_VALUE(base_vertex, 1, 0, xx, xx, xx) 289SYSTEM_VALUE(instance_id, 1, 0, xx, xx, xx) 290SYSTEM_VALUE(base_instance, 1, 0, xx, xx, xx) 291SYSTEM_VALUE(draw_id, 1, 0, xx, xx, xx) 292SYSTEM_VALUE(sample_id, 1, 0, xx, xx, xx) 293SYSTEM_VALUE(sample_pos, 2, 0, xx, xx, xx) 294SYSTEM_VALUE(sample_mask_in, 1, 0, xx, xx, xx) 295SYSTEM_VALUE(primitive_id, 1, 0, xx, xx, xx) 296SYSTEM_VALUE(invocation_id, 1, 0, xx, xx, xx) 297SYSTEM_VALUE(tess_coord, 3, 0, xx, xx, xx) 298SYSTEM_VALUE(tess_level_outer, 4, 0, xx, xx, xx) 299SYSTEM_VALUE(tess_level_inner, 2, 0, xx, xx, xx) 300SYSTEM_VALUE(patch_vertices_in, 1, 0, xx, xx, xx) 301SYSTEM_VALUE(local_invocation_id, 3, 0, xx, xx, xx) 302SYSTEM_VALUE(local_invocation_index, 1, 0, xx, xx, xx) 303SYSTEM_VALUE(work_group_id, 3, 0, xx, xx, xx) 304SYSTEM_VALUE(user_clip_plane, 4, 1, UCP_ID, xx, xx) 305SYSTEM_VALUE(num_work_groups, 3, 0, xx, xx, xx) 306SYSTEM_VALUE(helper_invocation, 1, 0, xx, xx, xx) 307SYSTEM_VALUE(channel_num, 1, 0, xx, xx, xx) 308 309/* 310 * Load operations pull data from some piece of GPU memory. All load 311 * operations operate in terms of offsets into some piece of theoretical 312 * memory. Loads from externally visible memory (UBO and SSBO) simply take a 313 * byte offset as a source. Loads from opaque memory (uniforms, inputs, etc.) 314 * take a base+offset pair where the base (const_index[0]) gives the location 315 * of the start of the variable being loaded and and the offset source is a 316 * offset into that variable. 317 * 318 * Uniform load operations have a second "range" index that specifies the 319 * range (starting at base) of the data from which we are loading. If 320 * const_index[1] == 0, then the range is unknown. 321 * 322 * Some load operations such as UBO/SSBO load and per_vertex loads take an 323 * additional source to specify which UBO/SSBO/vertex to load from. 324 * 325 * The exact address type depends on the lowering pass that generates the 326 * load/store intrinsics. Typically, this is vec4 units for things such as 327 * varying slots and float units for fragment shader inputs. UBO and SSBO 328 * offsets are always in bytes. 329 */ 330 331#define LOAD(name, srcs, num_indices, idx0, idx1, idx2, flags) \ 332 INTRINSIC(load_##name, srcs, ARR(1, 1, 1, 1), true, 0, 0, num_indices, idx0, idx1, idx2, flags) 333 334/* src[] = { offset }. const_index[] = { base } */ 335LOAD(uniform, 1, 2, BASE, RANGE, xx, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) 336/* src[] = { buffer_index, offset }. No const_index */ 337LOAD(ubo, 2, 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) 338/* src[] = { offset }. const_index[] = { base } */ 339LOAD(input, 1, 1, BASE, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) 340/* src[] = { vertex, offset }. const_index[] = { base } */ 341LOAD(per_vertex_input, 2, 1, BASE, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) 342/* src[] = { buffer_index, offset }. No const_index */ 343LOAD(ssbo, 2, 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE) 344/* src[] = { offset }. const_index[] = { base } */ 345LOAD(output, 1, 1, BASE, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE) 346/* src[] = { vertex, offset }. const_index[] = { base } */ 347LOAD(per_vertex_output, 2, 1, BASE, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE) 348/* src[] = { offset }. const_index[] = { base } */ 349LOAD(shared, 1, 1, BASE, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE) 350/* src[] = { offset }. const_index[] = { base, range } */ 351LOAD(push_constant, 1, 2, BASE, RANGE, xx, 352 NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) 353 354/* 355 * Stores work the same way as loads, except now the first source is the value 356 * to store and the second (and possibly third) source specify where to store 357 * the value. SSBO and shared memory stores also have a write mask as 358 * const_index[0]. 359 */ 360 361#define STORE(name, srcs, num_indices, idx0, idx1, idx2, flags) \ 362 INTRINSIC(store_##name, srcs, ARR(0, 1, 1, 1), false, 0, 0, num_indices, idx0, idx1, idx2, flags) 363 364/* src[] = { value, offset }. const_index[] = { base, write_mask } */ 365STORE(output, 2, 2, BASE, WRMASK, xx, 0) 366/* src[] = { value, vertex, offset }. const_index[] = { base, write_mask } */ 367STORE(per_vertex_output, 3, 2, BASE, WRMASK, xx, 0) 368/* src[] = { value, block_index, offset }. const_index[] = { write_mask } */ 369STORE(ssbo, 3, 1, WRMASK, xx, xx, 0) 370/* src[] = { value, offset }. const_index[] = { base, write_mask } */ 371STORE(shared, 2, 2, BASE, WRMASK, xx, 0) 372 373LAST_INTRINSIC(store_shared) 374