src/mesa/drivers/dri/i965/brw_nir_uniforms.cpp

   1 /*
   2  * Copyright © 2015 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
  24 #include "compiler/brw_nir.h"
  25 #include "compiler/glsl/ir_uniform.h"
  26 #include "compiler/nir/nir_builder.h"
  27 #include "brw_program.h"
  28
  29 static void
  30 brw_nir_setup_glsl_builtin_uniform(nir_variable *var,
  31                                    const struct gl_program *prog,
  32                                    struct brw_stage_prog_data *stage_prog_data,
  33                                    bool is_scalar)
  34 {
  35    const nir_state_slot *const slots = var->state_slots;
  36    assert(var->state_slots != NULL);
  37
  38    unsigned uniform_index = var->data.driver_location / 4;
  39    for (unsigned int i = 0; i < var->num_state_slots; i++) {
  40       /* This state reference has already been setup by ir_to_mesa, but we'll
  41        * get the same index back here.
  42        */
  43       int index = _mesa_add_state_reference(prog->Parameters,
  44                                             slots[i].tokens);
  45
  46       /* Add each of the unique swizzles of the element as a parameter.
  47        * This'll end up matching the expected layout of the
  48        * array/matrix/structure we're trying to fill in.
  49        */
  50       int last_swiz = -1;
  51       for (unsigned j = 0; j < 4; j++) {
  52          int swiz = GET_SWZ(slots[i].swizzle, j);
  53
  54          /* If we hit a pair of identical swizzles, this means we've hit the
  55           * end of the builtin variable.  In scalar mode, we should just quit
  56           * and move on to the next one.  In vec4, we need to continue and pad
  57           * it out to 4 components.
  58           */
  59          if (swiz == last_swiz && is_scalar)
  60             break;
  61
  62          last_swiz = swiz;
  63
  64          stage_prog_data->param[uniform_index++] =
  65             BRW_PARAM_PARAMETER(index, swiz);
  66       }
  67    }
  68 }
  69
  70 static void
  71 setup_vec4_image_param(uint32_t *params, uint32_t idx,
  72                        unsigned offset, unsigned n)
  73 {
  74    assert(offset % sizeof(uint32_t) == 0);
  75    for (unsigned i = 0; i < n; ++i)
  76       params[i] = BRW_PARAM_IMAGE(idx, offset / sizeof(uint32_t) + i);
  77
  78    for (unsigned i = n; i < 4; ++i)
  79       params[i] = BRW_PARAM_BUILTIN_ZERO;
  80 }
  81
  82 static void
  83 brw_setup_image_uniform_values(gl_shader_stage stage,
  84                                struct brw_stage_prog_data *stage_prog_data,
  85                                unsigned param_start_index,
  86                                const gl_uniform_storage *storage)
  87 {
  88    uint32_t *param = &stage_prog_data->param[param_start_index];
  89
  90    for (unsigned i = 0; i < MAX2(storage->array_elements, 1); i++) {
  91       const unsigned image_idx = storage->opaque[stage].index + i;
  92
  93       /* Upload the brw_image_param structure.  The order is expected to match
  94        * the BRW_IMAGE_PARAM_*_OFFSET defines.
  95        */
  96       setup_vec4_image_param(param + BRW_IMAGE_PARAM_SURFACE_IDX_OFFSET,
  97                              image_idx,
  98                              offsetof(brw_image_param, surface_idx), 1);
  99       setup_vec4_image_param(param + BRW_IMAGE_PARAM_OFFSET_OFFSET,
 100                              image_idx,
 101                              offsetof(brw_image_param, offset), 2);
 102       setup_vec4_image_param(param + BRW_IMAGE_PARAM_SIZE_OFFSET,
 103                              image_idx,
 104                              offsetof(brw_image_param, size), 3);
 105       setup_vec4_image_param(param + BRW_IMAGE_PARAM_STRIDE_OFFSET,
 106                              image_idx,
 107                              offsetof(brw_image_param, stride), 4);
 108       setup_vec4_image_param(param + BRW_IMAGE_PARAM_TILING_OFFSET,
 109                              image_idx,
 110                              offsetof(brw_image_param, tiling), 3);
 111       setup_vec4_image_param(param + BRW_IMAGE_PARAM_SWIZZLING_OFFSET,
 112                              image_idx,
 113                              offsetof(brw_image_param, swizzling), 2);
 114       param += BRW_IMAGE_PARAM_SIZE;
 115
 116       brw_mark_surface_used(
 117          stage_prog_data,
 118          stage_prog_data->binding_table.image_start + image_idx);
 119    }
 120 }
 121
 122 static unsigned
 123 count_uniform_storage_slots(const struct glsl_type *type)
 124 {
 125    /* gl_uniform_storage can cope with one level of array, so if the
 126     * type is a composite type or an array where each element occupies
 127     * more than one slot than we need to recursively process it.
 128     */
 129    if (glsl_type_is_struct(type)) {
 130       unsigned location_count = 0;
 131
 132       for (unsigned i = 0; i < glsl_get_length(type); i++) {
 133          const struct glsl_type *field_type = glsl_get_struct_field(type, i);
 134
 135          location_count += count_uniform_storage_slots(field_type);
 136       }
 137
 138       return location_count;
 139    }
 140
 141    if (glsl_type_is_array(type)) {
 142       const struct glsl_type *element_type = glsl_get_array_element(type);
 143
 144       if (glsl_type_is_array(element_type) ||
 145           glsl_type_is_struct(element_type)) {
 146          unsigned element_count = count_uniform_storage_slots(element_type);
 147          return element_count * glsl_get_length(type);
 148       }
 149    }
 150
 151    return 1;
 152 }
 153
 154 static void
 155 brw_nir_setup_glsl_uniform(gl_shader_stage stage, nir_variable *var,
 156                            const struct gl_program *prog,
 157                            struct brw_stage_prog_data *stage_prog_data,
 158                            bool is_scalar)
 159 {
 160    /* The data for our (non-builtin) uniforms is stored in a series of
 161     * gl_uniform_storage structs for each subcomponent that
 162     * glGetUniformLocation() could name.  We know it's been set up in the same
 163     * order we'd walk the type, so walk the list of storage that matches the
 164     * range of slots covered by this variable.
 165     */
 166    unsigned uniform_index = var->data.driver_location / 4;
 167    unsigned num_slots = count_uniform_storage_slots(var->type);
 168    for (unsigned u = 0; u < num_slots; u++) {
 169       struct gl_uniform_storage *storage =
 170          &prog->sh.data->UniformStorage[var->data.location + u];
 171
 172       if (storage->builtin || storage->type->is_sampler())
 173          continue;
 174
 175       if (storage->type->is_image()) {
 176          brw_setup_image_uniform_values(stage, stage_prog_data,
 177                                         uniform_index, storage);
 178          uniform_index +=
 179             BRW_IMAGE_PARAM_SIZE * MAX2(storage->array_elements, 1);
 180       } else {
 181          gl_constant_value *components = storage->storage;
 182          unsigned vector_count = (MAX2(storage->array_elements, 1) *
 183                                   storage->type->matrix_columns);
 184          unsigned vector_size = storage->type->vector_elements;
 185          unsigned max_vector_size = 4;
 186          if (storage->type->base_type == GLSL_TYPE_DOUBLE ||
 187              storage->type->base_type == GLSL_TYPE_UINT64 ||
 188              storage->type->base_type == GLSL_TYPE_INT64) {
 189             vector_size *= 2;
 190             if (vector_size > 4)
 191                max_vector_size = 8;
 192          }
 193
 194          for (unsigned s = 0; s < vector_count; s++) {
 195             unsigned i;
 196             for (i = 0; i < vector_size; i++) {
 197                uint32_t idx = components - prog->sh.data->UniformDataSlots;
 198                stage_prog_data->param[uniform_index++] = BRW_PARAM_UNIFORM(idx);
 199                components++;
 200             }
 201
 202             if (!is_scalar) {
 203                /* Pad out with zeros if needed (only needed for vec4) */
 204                for (; i < max_vector_size; i++) {
 205                   stage_prog_data->param[uniform_index++] =
 206                      BRW_PARAM_BUILTIN_ZERO;
 207                }
 208             }
 209          }
 210       }
 211    }
 212 }
 213
 214 void
 215 brw_nir_setup_glsl_uniforms(void *mem_ctx, nir_shader *shader,
 216                             const struct gl_program *prog,
 217                             struct brw_stage_prog_data *stage_prog_data,
 218                             bool is_scalar)
 219 {
 220    unsigned nr_params = shader->num_uniforms / 4;
 221    stage_prog_data->nr_params = nr_params;
 222    stage_prog_data->param = rzalloc_array(mem_ctx, uint32_t, nr_params);
 223
 224    nir_foreach_variable(var, &shader->uniforms) {
 225       /* UBO's, atomics and samplers don't take up space in the
 226          uniform file */
 227       if (var->interface_type != NULL || var->type->contains_atomic())
 228          continue;
 229
 230       if (var->num_state_slots > 0) {
 231          brw_nir_setup_glsl_builtin_uniform(var, prog, stage_prog_data,
 232                                             is_scalar);
 233       } else {
 234          brw_nir_setup_glsl_uniform(shader->info.stage, var, prog,
 235                                     stage_prog_data, is_scalar);
 236       }
 237    }
 238 }
 239
 240 void
 241 brw_nir_setup_arb_uniforms(void *mem_ctx, nir_shader *shader,
 242                            struct gl_program *prog,
 243                            struct brw_stage_prog_data *stage_prog_data)
 244 {
 245    struct gl_program_parameter_list *plist = prog->Parameters;
 246
 247    unsigned nr_params = plist->NumParameters * 4;
 248    stage_prog_data->nr_params = nr_params;
 249    stage_prog_data->param = rzalloc_array(mem_ctx, uint32_t, nr_params);
 250
 251    /* For ARB programs, prog_to_nir generates a single "parameters" variable
 252     * for all uniform data.  nir_lower_wpos_ytransform may also create an
 253     * additional variable.
 254     */
 255    assert(shader->uniforms.length() <= 2);
 256
 257    for (unsigned p = 0; p < plist->NumParameters; p++) {
 258       /* Parameters should be either vec4 uniforms or single component
 259        * constants; matrices and other larger types should have been broken
 260        * down earlier.
 261        */
 262       assert(plist->Parameters[p].Size <= 4);
 263
 264       unsigned i;
 265       for (i = 0; i < plist->Parameters[p].Size; i++)
 266          stage_prog_data->param[4 * p + i] = BRW_PARAM_PARAMETER(p, i);
 267       for (; i < 4; i++)
 268          stage_prog_data->param[4 * p + i] = BRW_PARAM_BUILTIN_ZERO;
 269    }
 270 }
 271
 272 static nir_ssa_def *
 273 get_aoa_deref_offset(nir_builder *b,
 274                      nir_deref_instr *deref,
 275                      unsigned elem_size)
 276 {
 277    unsigned array_size = elem_size;
 278    nir_ssa_def *offset = nir_imm_int(b, 0);
 279
 280    while (deref->deref_type != nir_deref_type_var) {
 281       assert(deref->deref_type == nir_deref_type_array);
 282
 283       /* This level's element size is the previous level's array size */
 284       nir_ssa_def *index = nir_ssa_for_src(b, deref->arr.index, 1);
 285       assert(deref->arr.index.ssa);
 286       offset = nir_iadd(b, offset,
 287                            nir_imul(b, index, nir_imm_int(b, array_size)));
 288
 289       deref = nir_deref_instr_parent(deref);
 290       assert(glsl_type_is_array(deref->type));
 291       array_size *= glsl_get_length(deref->type);
 292    }
 293
 294    /* Accessing an invalid surface index with the dataport can result in a
 295     * hang.  According to the spec "if the index used to select an individual
 296     * element is negative or greater than or equal to the size of the array,
 297     * the results of the operation are undefined but may not lead to
 298     * termination" -- which is one of the possible outcomes of the hang.
 299     * Clamp the index to prevent access outside of the array bounds.
 300     */
 301    return nir_umin(b, offset, nir_imm_int(b, array_size - elem_size));
 302 }
 303
 304 void
 305 brw_nir_lower_glsl_images(nir_shader *shader,
 306                           const struct gl_program *prog)
 307 {
 308    /* We put image uniforms at the end */
 309    nir_foreach_variable(var, &shader->uniforms) {
 310       if (!var->type->contains_image())
 311          continue;
 312
 313       /* GL Only allows arrays of arrays of images */
 314       assert(var->type->without_array()->is_image());
 315       const unsigned num_images = MAX2(1, var->type->arrays_of_arrays_size());
 316
 317       var->data.driver_location = shader->num_uniforms;
 318       shader->num_uniforms += num_images * BRW_IMAGE_PARAM_SIZE * 4;
 319    }
 320
 321    nir_function_impl *impl = nir_shader_get_entrypoint(shader);
 322
 323    nir_builder b;
 324    nir_builder_init(&b, impl);
 325
 326    nir_foreach_block(block, impl) {
 327       nir_foreach_instr_safe(instr, block) {
 328          if (instr->type != nir_instr_type_intrinsic)
 329             continue;
 330
 331          nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
 332          switch (intrin->intrinsic) {
 333          case nir_intrinsic_image_deref_load:
 334          case nir_intrinsic_image_deref_store:
 335          case nir_intrinsic_image_deref_atomic_add:
 336          case nir_intrinsic_image_deref_atomic_min:
 337          case nir_intrinsic_image_deref_atomic_max:
 338          case nir_intrinsic_image_deref_atomic_and:
 339          case nir_intrinsic_image_deref_atomic_or:
 340          case nir_intrinsic_image_deref_atomic_xor:
 341          case nir_intrinsic_image_deref_atomic_exchange:
 342          case nir_intrinsic_image_deref_atomic_comp_swap:
 343          case nir_intrinsic_image_deref_size:
 344          case nir_intrinsic_image_deref_samples:
 345          case nir_intrinsic_image_deref_load_raw_intel:
 346          case nir_intrinsic_image_deref_store_raw_intel: {
 347             nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
 348             nir_variable *var = nir_deref_instr_get_variable(deref);
 349             const unsigned num_images =
 350                MAX2(1, var->type->arrays_of_arrays_size());
 351
 352             struct gl_uniform_storage *storage =
 353                &prog->sh.data->UniformStorage[var->data.location];
 354             const unsigned image_var_idx =
 355                storage->opaque[shader->info.stage].index;
 356
 357             b.cursor = nir_before_instr(&intrin->instr);
 358             nir_ssa_def *index = nir_iadd(&b, nir_imm_int(&b, image_var_idx),
 359                                           get_aoa_deref_offset(&b, deref, 1));
 360             brw_nir_rewrite_image_intrinsic(intrin, index);
 361             break;
 362          }
 363
 364          case nir_intrinsic_image_deref_load_param_intel: {
 365             nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
 366             nir_variable *var = nir_deref_instr_get_variable(deref);
 367             const unsigned num_images =
 368                MAX2(1, var->type->arrays_of_arrays_size());
 369
 370             b.cursor = nir_instr_remove(&intrin->instr);
 371
 372             const unsigned param = nir_intrinsic_base(intrin);
 373             nir_ssa_def *offset =
 374                get_aoa_deref_offset(&b, deref, BRW_IMAGE_PARAM_SIZE * 4);
 375             offset = nir_iadd(&b, offset, nir_imm_int(&b, param * 16));
 376
 377             nir_intrinsic_instr *load =
 378                nir_intrinsic_instr_create(b.shader,
 379                                           nir_intrinsic_load_uniform);
 380             nir_intrinsic_set_base(load, var->data.driver_location);
 381             nir_intrinsic_set_range(load, num_images * BRW_IMAGE_PARAM_SIZE * 4);
 382             load->src[0] = nir_src_for_ssa(offset);
 383             load->num_components = intrin->dest.ssa.num_components;
 384             nir_ssa_dest_init(&load->instr, &load->dest,
 385                               intrin->dest.ssa.num_components,
 386                               intrin->dest.ssa.bit_size, NULL);
 387             nir_builder_instr_insert(&b, &load->instr);
 388
 389             nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
 390                                      nir_src_for_ssa(&load->dest.ssa));
 391             break;
 392          }
 393
 394          default:
 395             break;
 396          }
 397       }
 398    }
 399 }