src/compiler/nir/nir_intrinsics.h

   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 INTRINSIC(nop, 0, ARR(0), false, 0, 0, 0, xx, xx, xx,
  45           NIR_INTRINSIC_CAN_ELIMINATE)
  46
  47 INTRINSIC(load_var, 0, ARR(0), true, 0, 1, 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
  48 INTRINSIC(store_var, 1, ARR(0), false, 0, 1, 1, WRMASK, xx, xx, 0)
  49 INTRINSIC(copy_var, 0, ARR(0), false, 0, 2, 0, xx, xx, xx, 0)
  50
  51 /*
  52  * Interpolation of input.  The interp_var_at* intrinsics are similar to the
  53  * load_var intrinsic acting on a shader input except that they interpolate
  54  * the input differently.  The at_sample and at_offset intrinsics take an
  55  * additional source that is an integer sample id or a vec2 position offset
  56  * respectively.
  57  */
  58
  59 INTRINSIC(interp_var_at_centroid, 0, ARR(0), true, 0, 1, 0, xx, xx, xx,
  60           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
  61 INTRINSIC(interp_var_at_sample, 1, ARR(1), true, 0, 1, 0, xx, xx, xx,
  62           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
  63 INTRINSIC(interp_var_at_offset, 1, ARR(2), true, 0, 1, 0, xx, xx, xx,
  64           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
  65
  66 /*
  67  * Ask the driver for the size of a given buffer. It takes the buffer index
  68  * as source.
  69  */
  70 INTRINSIC(get_buffer_size, 1, ARR(1), true, 1, 0, 0, xx, xx, xx,
  71           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
  72
  73 /*
  74  * a barrier is an intrinsic with no inputs/outputs but which can't be moved
  75  * around/optimized in general
  76  */
  77 #define BARRIER(name) INTRINSIC(name, 0, ARR(0), false, 0, 0, 0, xx, xx, xx, 0)
  78
  79 BARRIER(barrier)
  80 BARRIER(discard)
  81
  82 /*
  83  * Memory barrier with semantics analogous to the memoryBarrier() GLSL
  84  * intrinsic.
  85  */
  86 BARRIER(memory_barrier)
  87
  88 /*
  89  * Shader clock intrinsic with semantics analogous to the clock2x32ARB()
  90  * GLSL intrinsic.
  91  * The latter can be used as code motion barrier, which is currently not
  92  * feasible with NIR.
  93  */
  94 INTRINSIC(shader_clock, 0, ARR(0), true, 2, 0, 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
  95
  96 /*
  97  * Memory barrier with semantics analogous to the compute shader
  98  * groupMemoryBarrier(), memoryBarrierAtomicCounter(), memoryBarrierBuffer(),
  99  * memoryBarrierImage() and memoryBarrierShared() GLSL intrinsics.
 100  */
 101 BARRIER(group_memory_barrier)
 102 BARRIER(memory_barrier_atomic_counter)
 103 BARRIER(memory_barrier_buffer)
 104 BARRIER(memory_barrier_image)
 105 BARRIER(memory_barrier_shared)
 106
 107 /** A conditional discard, with a single boolean source. */
 108 INTRINSIC(discard_if, 1, ARR(1), false, 0, 0, 0, xx, xx, xx, 0)
 109
 110 /**
 111  * Basic Geometry Shader intrinsics.
 112  *
 113  * emit_vertex implements GLSL's EmitStreamVertex() built-in.  It takes a single
 114  * index, which is the stream ID to write to.
 115  *
 116  * end_primitive implements GLSL's EndPrimitive() built-in.
 117  */
 118 INTRINSIC(emit_vertex,   0, ARR(0), false, 0, 0, 1, STREAM_ID, xx, xx, 0)
 119 INTRINSIC(end_primitive, 0, ARR(0), false, 0, 0, 1, STREAM_ID, xx, xx, 0)
 120
 121 /**
 122  * Geometry Shader intrinsics with a vertex count.
 123  *
 124  * Alternatively, drivers may implement these intrinsics, and use
 125  * nir_lower_gs_intrinsics() to convert from the basic intrinsics.
 126  *
 127  * These maintain a count of the number of vertices emitted, as an additional
 128  * unsigned integer source.
 129  */
 130 INTRINSIC(emit_vertex_with_counter, 1, ARR(1), false, 0, 0, 1, STREAM_ID, xx, xx, 0)
 131 INTRINSIC(end_primitive_with_counter, 1, ARR(1), false, 0, 0, 1, STREAM_ID, xx, xx, 0)
 132 INTRINSIC(set_vertex_count, 1, ARR(1), false, 0, 0, 0, xx, xx, xx, 0)
 133
 134 /*
 135  * Atomic counters
 136  *
 137  * The *_var variants take an atomic_uint nir_variable, while the other,
 138  * lowered, variants take a constant buffer index and register offset.
 139  */
 140
 141 #define ATOMIC(name, flags) \
 142    INTRINSIC(name##_var, 0, ARR(0), true, 1, 1, 0, xx, xx, xx, flags) \
 143    INTRINSIC(name, 1, ARR(1), true, 1, 0, 1, BASE, xx, xx, flags)
 144 #define ATOMIC2(name) \
 145    INTRINSIC(name##_var, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0) \
 146    INTRINSIC(name, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 147 #define ATOMIC3(name) \
 148    INTRINSIC(name##_var, 2, ARR(1, 1), true, 1, 1, 0, xx, xx, xx, 0) \
 149    INTRINSIC(name, 3, ARR(1, 1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 150
 151 ATOMIC(atomic_counter_inc, 0)
 152 ATOMIC(atomic_counter_dec, 0)
 153 ATOMIC(atomic_counter_read, NIR_INTRINSIC_CAN_ELIMINATE)
 154 ATOMIC2(atomic_counter_add)
 155 ATOMIC2(atomic_counter_min)
 156 ATOMIC2(atomic_counter_max)
 157 ATOMIC2(atomic_counter_and)
 158 ATOMIC2(atomic_counter_or)
 159 ATOMIC2(atomic_counter_xor)
 160 ATOMIC2(atomic_counter_exchange)
 161 ATOMIC3(atomic_counter_comp_swap)
 162
 163 /*
 164  * Image load, store and atomic intrinsics.
 165  *
 166  * All image intrinsics take an image target passed as a nir_variable.  Image
 167  * variables contain a number of memory and layout qualifiers that influence
 168  * the semantics of the intrinsic.
 169  *
 170  * All image intrinsics take a four-coordinate vector and a sample index as
 171  * first two sources, determining the location within the image that will be
 172  * accessed by the intrinsic.  Components not applicable to the image target
 173  * in use are undefined.  Image store takes an additional four-component
 174  * argument with the value to be written, and image atomic operations take
 175  * either one or two additional scalar arguments with the same meaning as in
 176  * the ARB_shader_image_load_store specification.
 177  */
 178 INTRINSIC(image_load, 2, ARR(4, 1), true, 4, 1, 0, xx, xx, xx,
 179           NIR_INTRINSIC_CAN_ELIMINATE)
 180 INTRINSIC(image_store, 3, ARR(4, 1, 4), false, 0, 1, 0, xx, xx, xx, 0)
 181 INTRINSIC(image_atomic_add, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 182 INTRINSIC(image_atomic_min, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 183 INTRINSIC(image_atomic_max, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 184 INTRINSIC(image_atomic_and, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 185 INTRINSIC(image_atomic_or, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 186 INTRINSIC(image_atomic_xor, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 187 INTRINSIC(image_atomic_exchange, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 188 INTRINSIC(image_atomic_comp_swap, 4, ARR(4, 1, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 189 INTRINSIC(image_size, 0, ARR(0), true, 0, 1, 0, xx, xx, xx,
 190           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 191 INTRINSIC(image_samples, 0, ARR(0), true, 1, 1, 0, xx, xx, xx,
 192           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 193
 194 /*
 195  * Vulkan descriptor set intrinsic
 196  *
 197  * The Vulkan API uses a different binding model from GL.  In the Vulkan
 198  * API, all external resources are represented by a tuple:
 199  *
 200  * (descriptor set, binding, array index)
 201  *
 202  * where the array index is the only thing allowed to be indirect.  The
 203  * vulkan_surface_index intrinsic takes the descriptor set and binding as
 204  * its first two indices and the array index as its source.  The third
 205  * index is a nir_variable_mode in case that's useful to the backend.
 206  *
 207  * The intended usage is that the shader will call vulkan_surface_index to
 208  * get an index and then pass that as the buffer index ubo/ssbo calls.
 209  */
 210 INTRINSIC(vulkan_resource_index, 1, ARR(1), true, 1, 0, 2,
 211           DESC_SET, BINDING, xx,
 212           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 213
 214 /*
 215  * variable atomic intrinsics
 216  *
 217  * All of these variable atomic memory operations read a value from memory,
 218  * compute a new value using one of the operations below, write the new value
 219  * to memory, and return the original value read.
 220  *
 221  * All operations take 1 source except CompSwap that takes 2. These sources
 222  * represent:
 223  *
 224  * 0: The data parameter to the atomic function (i.e. the value to add
 225  *    in shared_atomic_add, etc).
 226  * 1: For CompSwap only: the second data parameter.
 227  *
 228  * All operations take 1 variable deref.
 229  */
 230 INTRINSIC(var_atomic_add, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
 231 INTRINSIC(var_atomic_imin, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
 232 INTRINSIC(var_atomic_umin, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
 233 INTRINSIC(var_atomic_imax, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
 234 INTRINSIC(var_atomic_umax, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
 235 INTRINSIC(var_atomic_and, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
 236 INTRINSIC(var_atomic_or, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
 237 INTRINSIC(var_atomic_xor, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
 238 INTRINSIC(var_atomic_exchange, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
 239 INTRINSIC(var_atomic_comp_swap, 2, ARR(1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 240
 241 /*
 242  * SSBO atomic intrinsics
 243  *
 244  * All of the SSBO atomic memory operations read a value from memory,
 245  * compute a new value using one of the operations below, write the new
 246  * value to memory, and return the original value read.
 247  *
 248  * All operations take 3 sources except CompSwap that takes 4. These
 249  * sources represent:
 250  *
 251  * 0: The SSBO buffer index.
 252  * 1: The offset into the SSBO buffer of the variable that the atomic
 253  *    operation will operate on.
 254  * 2: The data parameter to the atomic function (i.e. the value to add
 255  *    in ssbo_atomic_add, etc).
 256  * 3: For CompSwap only: the second data parameter.
 257  */
 258 INTRINSIC(ssbo_atomic_add, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 259 INTRINSIC(ssbo_atomic_imin, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 260 INTRINSIC(ssbo_atomic_umin, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 261 INTRINSIC(ssbo_atomic_imax, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 262 INTRINSIC(ssbo_atomic_umax, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 263 INTRINSIC(ssbo_atomic_and, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 264 INTRINSIC(ssbo_atomic_or, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 265 INTRINSIC(ssbo_atomic_xor, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 266 INTRINSIC(ssbo_atomic_exchange, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 267 INTRINSIC(ssbo_atomic_comp_swap, 4, ARR(1, 1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 268
 269 /*
 270  * CS shared variable atomic intrinsics
 271  *
 272  * All of the shared variable atomic memory operations read a value from
 273  * memory, compute a new value using one of the operations below, write the
 274  * new value to memory, and return the original value read.
 275  *
 276  * All operations take 2 sources except CompSwap that takes 3. These
 277  * sources represent:
 278  *
 279  * 0: The offset into the shared variable storage region that the atomic
 280  *    operation will operate on.
 281  * 1: The data parameter to the atomic function (i.e. the value to add
 282  *    in shared_atomic_add, etc).
 283  * 2: For CompSwap only: the second data parameter.
 284  */
 285 INTRINSIC(shared_atomic_add, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 286 INTRINSIC(shared_atomic_imin, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 287 INTRINSIC(shared_atomic_umin, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 288 INTRINSIC(shared_atomic_imax, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 289 INTRINSIC(shared_atomic_umax, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 290 INTRINSIC(shared_atomic_and, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 291 INTRINSIC(shared_atomic_or, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 292 INTRINSIC(shared_atomic_xor, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 293 INTRINSIC(shared_atomic_exchange, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 294 INTRINSIC(shared_atomic_comp_swap, 3, ARR(1, 1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 295
 296 /* Used by nir_builder.h to generate loader helpers for the system values. */
 297 #ifndef DEFINE_SYSTEM_VALUE
 298 #define DEFINE_SYSTEM_VALUE(name)
 299 #endif
 300
 301 #define SYSTEM_VALUE(name, components, num_indices, idx0, idx1, idx2) \
 302    DEFINE_SYSTEM_VALUE(name) \
 303    INTRINSIC(load_##name, 0, ARR(0), true, components, 0, num_indices, \
 304    idx0, idx1, idx2, \
 305    NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 306
 307 SYSTEM_VALUE(front_face, 1, 0, xx, xx, xx)
 308 SYSTEM_VALUE(vertex_id, 1, 0, xx, xx, xx)
 309 SYSTEM_VALUE(vertex_id_zero_base, 1, 0, xx, xx, xx)
 310 SYSTEM_VALUE(base_vertex, 1, 0, xx, xx, xx)
 311 SYSTEM_VALUE(instance_id, 1, 0, xx, xx, xx)
 312 SYSTEM_VALUE(base_instance, 1, 0, xx, xx, xx)
 313 SYSTEM_VALUE(draw_id, 1, 0, xx, xx, xx)
 314 SYSTEM_VALUE(sample_id, 1, 0, xx, xx, xx)
 315 SYSTEM_VALUE(sample_pos, 2, 0, xx, xx, xx)
 316 SYSTEM_VALUE(sample_mask_in, 1, 0, xx, xx, xx)
 317 SYSTEM_VALUE(primitive_id, 1, 0, xx, xx, xx)
 318 SYSTEM_VALUE(invocation_id, 1, 0, xx, xx, xx)
 319 SYSTEM_VALUE(tess_coord, 3, 0, xx, xx, xx)
 320 SYSTEM_VALUE(tess_level_outer, 4, 0, xx, xx, xx)
 321 SYSTEM_VALUE(tess_level_inner, 2, 0, xx, xx, xx)
 322 SYSTEM_VALUE(patch_vertices_in, 1, 0, xx, xx, xx)
 323 SYSTEM_VALUE(local_invocation_id, 3, 0, xx, xx, xx)
 324 SYSTEM_VALUE(local_invocation_index, 1, 0, xx, xx, xx)
 325 SYSTEM_VALUE(work_group_id, 3, 0, xx, xx, xx)
 326 SYSTEM_VALUE(user_clip_plane, 4, 1, UCP_ID, xx, xx)
 327 SYSTEM_VALUE(num_work_groups, 3, 0, xx, xx, xx)
 328 SYSTEM_VALUE(helper_invocation, 1, 0, xx, xx, xx)
 329 SYSTEM_VALUE(channel_num, 1, 0, xx, xx, xx)
 330 SYSTEM_VALUE(alpha_ref_float, 1, 0, xx, xx, xx)
 331 SYSTEM_VALUE(layer_id, 1, 0, xx, xx, xx)
 332
 333 /* Blend constant color values.  Float values are clamped. */
 334 SYSTEM_VALUE(blend_const_color_r_float, 1, 0, xx, xx, xx)
 335 SYSTEM_VALUE(blend_const_color_g_float, 1, 0, xx, xx, xx)
 336 SYSTEM_VALUE(blend_const_color_b_float, 1, 0, xx, xx, xx)
 337 SYSTEM_VALUE(blend_const_color_a_float, 1, 0, xx, xx, xx)
 338 SYSTEM_VALUE(blend_const_color_rgba8888_unorm, 1, 0, xx, xx, xx)
 339 SYSTEM_VALUE(blend_const_color_aaaa8888_unorm, 1, 0, xx, xx, xx)
 340
 341 /**
 342  * Barycentric coordinate intrinsics.
 343  *
 344  * These set up the barycentric coordinates for a particular interpolation.
 345  * The first three are for the simple cases: pixel, centroid, or per-sample
 346  * (at gl_SampleID).  The next two handle interpolating at a specified
 347  * sample location, or interpolating with a vec2 offset,
 348  *
 349  * The interp_mode index should be either the INTERP_MODE_SMOOTH or
 350  * INTERP_MODE_NOPERSPECTIVE enum values.
 351  *
 352  * The vec2 value produced by these intrinsics is intended for use as the
 353  * barycoord source of a load_interpolated_input intrinsic.
 354  */
 355
 356 #define BARYCENTRIC(name, sources, source_components) \
 357    INTRINSIC(load_barycentric_##name, sources, ARR(source_components), \
 358              true, 2, 0, 1, INTERP_MODE, xx, xx, \
 359              NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 360
 361 /* no sources.  const_index[] = { interp_mode } */
 362 BARYCENTRIC(pixel, 0, 0)
 363 BARYCENTRIC(centroid, 0, 0)
 364 BARYCENTRIC(sample, 0, 0)
 365 /* src[] = { sample_id }.  const_index[] = { interp_mode } */
 366 BARYCENTRIC(at_sample, 1, 1)
 367 /* src[] = { offset.xy }.  const_index[] = { interp_mode } */
 368 BARYCENTRIC(at_offset, 1, 2)
 369
 370 /*
 371  * Load operations pull data from some piece of GPU memory.  All load
 372  * operations operate in terms of offsets into some piece of theoretical
 373  * memory.  Loads from externally visible memory (UBO and SSBO) simply take a
 374  * byte offset as a source.  Loads from opaque memory (uniforms, inputs, etc.)
 375  * take a base+offset pair where the base (const_index[0]) gives the location
 376  * of the start of the variable being loaded and and the offset source is a
 377  * offset into that variable.
 378  *
 379  * Uniform load operations have a second "range" index that specifies the
 380  * range (starting at base) of the data from which we are loading.  If
 381  * const_index[1] == 0, then the range is unknown.
 382  *
 383  * Some load operations such as UBO/SSBO load and per_vertex loads take an
 384  * additional source to specify which UBO/SSBO/vertex to load from.
 385  *
 386  * The exact address type depends on the lowering pass that generates the
 387  * load/store intrinsics.  Typically, this is vec4 units for things such as
 388  * varying slots and float units for fragment shader inputs.  UBO and SSBO
 389  * offsets are always in bytes.
 390  */
 391
 392 #define LOAD(name, srcs, num_indices, idx0, idx1, idx2, flags) \
 393    INTRINSIC(load_##name, srcs, ARR(1, 1, 1, 1), true, 0, 0, num_indices, idx0, idx1, idx2, flags)
 394
 395 /* src[] = { offset }. const_index[] = { base, range } */
 396 LOAD(uniform, 1, 2, BASE, RANGE, xx, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 397 /* src[] = { buffer_index, offset }. No const_index */
 398 LOAD(ubo, 2, 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 399 /* src[] = { offset }. const_index[] = { base, component } */
 400 LOAD(input, 1, 2, BASE, COMPONENT, xx, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 401 /* src[] = { vertex, offset }. const_index[] = { base, component } */
 402 LOAD(per_vertex_input, 2, 2, BASE, COMPONENT, xx, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 403 /* src[] = { barycoord, offset }. const_index[] = { base, component } */
 404 INTRINSIC(load_interpolated_input, 2, ARR(2, 1), true, 0, 0,
 405           2, BASE, COMPONENT, xx,
 406           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 407
 408 /* src[] = { buffer_index, offset }. No const_index */
 409 LOAD(ssbo, 2, 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
 410 /* src[] = { offset }. const_index[] = { base, component } */
 411 LOAD(output, 1, 1, BASE, COMPONENT, xx, NIR_INTRINSIC_CAN_ELIMINATE)
 412 /* src[] = { vertex, offset }. const_index[] = { base, component } */
 413 LOAD(per_vertex_output, 2, 1, BASE, COMPONENT, xx, NIR_INTRINSIC_CAN_ELIMINATE)
 414 /* src[] = { offset }. const_index[] = { base } */
 415 LOAD(shared, 1, 1, BASE, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
 416 /* src[] = { offset }. const_index[] = { base, range } */
 417 LOAD(push_constant, 1, 2, BASE, RANGE, xx,
 418      NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 419
 420 /*
 421  * Stores work the same way as loads, except now the first source is the value
 422  * to store and the second (and possibly third) source specify where to store
 423  * the value.  SSBO and shared memory stores also have a write mask as
 424  * const_index[0].
 425  */
 426
 427 #define STORE(name, srcs, num_indices, idx0, idx1, idx2, flags) \
 428    INTRINSIC(store_##name, srcs, ARR(0, 1, 1, 1), false, 0, 0, num_indices, idx0, idx1, idx2, flags)
 429
 430 /* src[] = { value, offset }. const_index[] = { base, write_mask, component } */
 431 STORE(output, 2, 3, BASE, WRMASK, COMPONENT, 0)
 432 /* src[] = { value, vertex, offset }.
 433  * const_index[] = { base, write_mask, component }
 434  */
 435 STORE(per_vertex_output, 3, 3, BASE, WRMASK, COMPONENT, 0)
 436 /* src[] = { value, block_index, offset }. const_index[] = { write_mask } */
 437 STORE(ssbo, 3, 1, WRMASK, xx, xx, 0)
 438 /* src[] = { value, offset }. const_index[] = { base, write_mask } */
 439 STORE(shared, 2, 2, BASE, WRMASK, xx, 0)
 440
 441 LAST_INTRINSIC(store_shared)
 442
 443 #undef DEFINE_SYSTEM_VALUE
 444 #undef INTRINSIC
 445 #undef LAST_INTRINSIC