src/glsl/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, 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
  45 INTRINSIC(load_var, 0, ARR(), true, 0, 1, 0, NIR_INTRINSIC_CAN_ELIMINATE)
  46 INTRINSIC(store_var, 1, ARR(0), false, 0, 1, 0, 0)
  47 INTRINSIC(copy_var, 0, ARR(), false, 0, 2, 0, 0)
  48
  49 /*
  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
  54  * respectively.
  55  */
  56
  57 INTRINSIC(interp_var_at_centroid, 0, ARR(0), true, 0, 1, 0,
  58           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
  59 INTRINSIC(interp_var_at_sample, 1, ARR(1), true, 0, 1, 0,
  60           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
  61 INTRINSIC(interp_var_at_offset, 1, ARR(2), true, 0, 1, 0,
  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  */
  68 INTRINSIC(get_buffer_size, 1, ARR(1), true, 1, 0, 0,
  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(), false, 0, 0, 0, 0)
  76
  77 BARRIER(barrier)
  78 BARRIER(discard)
  79
  80 /*
  81  * Memory barrier with semantics analogous to the memoryBarrier() GLSL
  82  * intrinsic.
  83  */
  84 BARRIER(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  */
  92 INTRINSIC(shader_clock, 0, ARR(), true, 1, 0, 0, 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  */
  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)
 104
 105 /** A conditional discard, with a single boolean source. */
 106 INTRINSIC(discard_if, 1, ARR(1), false, 0, 0, 0, 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  */
 116 INTRINSIC(emit_vertex,   0, ARR(), false, 0, 0, 1, 0)
 117 INTRINSIC(end_primitive, 0, ARR(), false, 0, 0, 1, 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  */
 128 INTRINSIC(emit_vertex_with_counter, 1, ARR(1), false, 0, 0, 1, 0)
 129 INTRINSIC(end_primitive_with_counter, 1, ARR(1), false, 0, 0, 1, 0)
 130 INTRINSIC(set_vertex_count, 1, ARR(1), false, 0, 0, 0, 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(), true, 1, 1, 0, flags) \
 141    INTRINSIC(atomic_counter_##name, 1, ARR(1), true, 1, 0, 1, flags)
 142
 143 ATOMIC(inc, 0)
 144 ATOMIC(dec, 0)
 145 ATOMIC(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  */
 162 INTRINSIC(image_load, 2, ARR(4, 1), true, 4, 1, 0,
 163           NIR_INTRINSIC_CAN_ELIMINATE)
 164 INTRINSIC(image_store, 3, ARR(4, 1, 4), false, 0, 1, 0, 0)
 165 INTRINSIC(image_atomic_add, 3, ARR(4, 1, 1), true, 1, 1, 0, 0)
 166 INTRINSIC(image_atomic_min, 3, ARR(4, 1, 1), true, 1, 1, 0, 0)
 167 INTRINSIC(image_atomic_max, 3, ARR(4, 1, 1), true, 1, 1, 0, 0)
 168 INTRINSIC(image_atomic_and, 3, ARR(4, 1, 1), true, 1, 1, 0, 0)
 169 INTRINSIC(image_atomic_or, 3, ARR(4, 1, 1), true, 1, 1, 0, 0)
 170 INTRINSIC(image_atomic_xor, 3, ARR(4, 1, 1), true, 1, 1, 0, 0)
 171 INTRINSIC(image_atomic_exchange, 3, ARR(4, 1, 1), true, 1, 1, 0, 0)
 172 INTRINSIC(image_atomic_comp_swap, 4, ARR(4, 1, 1, 1), true, 1, 1, 0, 0)
 173 INTRINSIC(image_size, 0, ARR(), true, 4, 1, 0,
 174           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 175 INTRINSIC(image_samples, 0, ARR(), true, 1, 1, 0,
 176           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 177
 178 /*
 179  * SSBO atomic intrinsics
 180  *
 181  * All of the SSBO atomic memory operations read a value from memory,
 182  * compute a new value using one of the operations below, write the new
 183  * value to memory, and return the original value read.
 184  *
 185  * All operations take 3 sources except CompSwap that takes 4. These
 186  * sources represent:
 187  *
 188  * 0: The SSBO buffer index.
 189  * 1: The offset into the SSBO buffer of the variable that the atomic
 190  *    operation will operate on.
 191  * 2: The data parameter to the atomic function (i.e. the value to add
 192  *    in ssbo_atomic_add, etc).
 193  * 3: For CompSwap only: the second data parameter.
 194  */
 195 INTRINSIC(ssbo_atomic_add, 3, ARR(1, 1, 1), true, 1, 0, 0, 0)
 196 INTRINSIC(ssbo_atomic_imin, 3, ARR(1, 1, 1), true, 1, 0, 0, 0)
 197 INTRINSIC(ssbo_atomic_umin, 3, ARR(1, 1, 1), true, 1, 0, 0, 0)
 198 INTRINSIC(ssbo_atomic_imax, 3, ARR(1, 1, 1), true, 1, 0, 0, 0)
 199 INTRINSIC(ssbo_atomic_umax, 3, ARR(1, 1, 1), true, 1, 0, 0, 0)
 200 INTRINSIC(ssbo_atomic_and, 3, ARR(1, 1, 1), true, 1, 0, 0, 0)
 201 INTRINSIC(ssbo_atomic_or, 3, ARR(1, 1, 1), true, 1, 0, 0, 0)
 202 INTRINSIC(ssbo_atomic_xor, 3, ARR(1, 1, 1), true, 1, 0, 0, 0)
 203 INTRINSIC(ssbo_atomic_exchange, 3, ARR(1, 1, 1), true, 1, 0, 0, 0)
 204 INTRINSIC(ssbo_atomic_comp_swap, 4, ARR(1, 1, 1, 1), true, 1, 0, 0, 0)
 205
 206 /*
 207  * CS shared variable atomic intrinsics
 208  *
 209  * All of the shared variable atomic memory operations read a value from
 210  * memory, compute a new value using one of the operations below, write the
 211  * new value to memory, and return the original value read.
 212  *
 213  * All operations take 2 sources except CompSwap that takes 3. These
 214  * sources represent:
 215  *
 216  * 0: The offset into the shared variable storage region that the atomic
 217  *    operation will operate on.
 218  * 1: The data parameter to the atomic function (i.e. the value to add
 219  *    in shared_atomic_add, etc).
 220  * 2: For CompSwap only: the second data parameter.
 221  */
 222 INTRINSIC(shared_atomic_add, 2, ARR(1, 1), true, 1, 0, 0, 0)
 223 INTRINSIC(shared_atomic_imin, 2, ARR(1, 1), true, 1, 0, 0, 0)
 224 INTRINSIC(shared_atomic_umin, 2, ARR(1, 1), true, 1, 0, 0, 0)
 225 INTRINSIC(shared_atomic_imax, 2, ARR(1, 1), true, 1, 0, 0, 0)
 226 INTRINSIC(shared_atomic_umax, 2, ARR(1, 1), true, 1, 0, 0, 0)
 227 INTRINSIC(shared_atomic_and, 2, ARR(1, 1), true, 1, 0, 0, 0)
 228 INTRINSIC(shared_atomic_or, 2, ARR(1, 1), true, 1, 0, 0, 0)
 229 INTRINSIC(shared_atomic_xor, 2, ARR(1, 1), true, 1, 0, 0, 0)
 230 INTRINSIC(shared_atomic_exchange, 2, ARR(1, 1), true, 1, 0, 0, 0)
 231 INTRINSIC(shared_atomic_comp_swap, 3, ARR(1, 1, 1), true, 1, 0, 0, 0)
 232
 233 #define SYSTEM_VALUE(name, components, num_indices) \
 234    INTRINSIC(load_##name, 0, ARR(), true, components, 0, num_indices, \
 235    NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 236
 237 SYSTEM_VALUE(front_face, 1, 0)
 238 SYSTEM_VALUE(vertex_id, 1, 0)
 239 SYSTEM_VALUE(vertex_id_zero_base, 1, 0)
 240 SYSTEM_VALUE(base_vertex, 1, 0)
 241 SYSTEM_VALUE(instance_id, 1, 0)
 242 SYSTEM_VALUE(sample_id, 1, 0)
 243 SYSTEM_VALUE(sample_pos, 2, 0)
 244 SYSTEM_VALUE(sample_mask_in, 1, 0)
 245 SYSTEM_VALUE(primitive_id, 1, 0)
 246 SYSTEM_VALUE(invocation_id, 1, 0)
 247 SYSTEM_VALUE(tess_coord, 3, 0)
 248 SYSTEM_VALUE(tess_level_outer, 4, 0)
 249 SYSTEM_VALUE(tess_level_inner, 2, 0)
 250 SYSTEM_VALUE(patch_vertices_in, 1, 0)
 251 SYSTEM_VALUE(local_invocation_id, 3, 0)
 252 SYSTEM_VALUE(work_group_id, 3, 0)
 253 SYSTEM_VALUE(user_clip_plane, 4, 1) /* const_index[0] is user_clip_plane[idx] */
 254 SYSTEM_VALUE(num_work_groups, 3, 0)
 255 SYSTEM_VALUE(helper_invocation, 1, 0)
 256
 257 /*
 258  * Load operations pull data from some piece of GPU memory.  All load
 259  * operations operate in terms of offsets into some piece of theoretical
 260  * memory.  Loads from externally visible memory (UBO and SSBO) simply take a
 261  * byte offset as a source.  Loads from opaque memory (uniforms, inputs, etc.)
 262  * take a base+offset pair where the base (const_index[0]) gives the location
 263  * of the start of the variable being loaded and and the offset source is a
 264  * offset into that variable.
 265  *
 266  * Some load operations such as UBO/SSBO load and per_vertex loads take an
 267  * additional source to specify which UBO/SSBO/vertex to load from.
 268  *
 269  * The exact address type depends on the lowering pass that generates the
 270  * load/store intrinsics.  Typically, this is vec4 units for things such as
 271  * varying slots and float units for fragment shader inputs.  UBO and SSBO
 272  * offsets are always in bytes.
 273  */
 274
 275 #define LOAD(name, srcs, indices, flags) \
 276    INTRINSIC(load_##name, srcs, ARR(1, 1, 1, 1), true, 0, 0, indices, flags)
 277
 278 /* src[] = { offset }. const_index[] = { base } */
 279 LOAD(uniform, 1, 1, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 280 /* src[] = { buffer_index, offset }. No const_index */
 281 LOAD(ubo, 2, 0, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 282 /* src[] = { offset }. const_index[] = { base } */
 283 LOAD(input, 1, 1, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 284 /* src[] = { vertex, offset }. const_index[] = { base } */
 285 LOAD(per_vertex_input, 2, 1, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 286 /* src[] = { buffer_index, offset }. No const_index */
 287 LOAD(ssbo, 2, 0, NIR_INTRINSIC_CAN_ELIMINATE)
 288 /* src[] = { offset }. const_index[] = { base } */
 289 LOAD(output, 1, 1, NIR_INTRINSIC_CAN_ELIMINATE)
 290 /* src[] = { vertex, offset }. const_index[] = { base } */
 291 LOAD(per_vertex_output, 2, 1, NIR_INTRINSIC_CAN_ELIMINATE)
 292 /* src[] = { offset }. const_index[] = { base } */
 293 LOAD(shared, 1, 1, NIR_INTRINSIC_CAN_ELIMINATE)
 294
 295 /*
 296  * Stores work the same way as loads, except now the first source is the value
 297  * to store and the second (and possibly third) source specify where to store
 298  * the value.  SSBO and shared memory stores also have a write mask as
 299  * const_index[0].
 300  */
 301
 302 #define STORE(name, srcs, indices, flags) \
 303    INTRINSIC(store_##name, srcs, ARR(0, 1, 1, 1), false, 0, 0, indices, flags)
 304
 305 /* src[] = { value, offset }. const_index[] = { base } */
 306 STORE(output, 2, 1, 0)
 307 /* src[] = { value, vertex, offset }. const_index[] = { base } */
 308 STORE(per_vertex_output, 3, 1, 0)
 309 /* src[] = { value, block_index, offset }. const_index[] = { write_mask } */
 310 STORE(ssbo, 3, 1, 0)
 311 /* src[] = { value, offset }. const_index[] = { base, write_mask } */
 312 STORE(shared, 2, 1, 0)
 313
 314 LAST_INTRINSIC(store_shared)