src/broadcom/compiler/v3d_nir_lower_image_load_store.c

   1 /*
   2  * Copyright © 2018 Intel Corporation
   3  * Copyright © 2018 Broadcom
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  21  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  22  * IN THE SOFTWARE.
  23  */
  24
  25 #include "v3d_compiler.h"
  26 #include "compiler/nir/nir_builder.h"
  27 #include "compiler/nir/nir_format_convert.h"
  28
  29 /** @file v3d_nir_lower_image_load_store.c
  30  *
  31  * Performs any necessary lowering of GL_ARB_shader_image_load_store
  32  * operations.
  33  *
  34  * On V3D 4.x, we just need to do format conversion for stores such that the
  35  * GPU can effectively memcpy the arguments (in increments of 32-bit words)
  36  * into the texel.  Loads are the same as texturing, where we may need to
  37  * unpack from 16-bit ints or floats.
  38  *
  39  * On V3D 3.x, to implement image load store we would need to do manual tiling
  40  * calculations and load/store using the TMU general memory access path.
  41  */
  42
  43 bool
  44 v3d_gl_format_is_return_32(GLenum format)
  45 {
  46         switch (format) {
  47         case GL_R8:
  48         case GL_R8_SNORM:
  49         case GL_R8UI:
  50         case GL_R8I:
  51         case GL_RG8:
  52         case GL_RG8_SNORM:
  53         case GL_RG8UI:
  54         case GL_RG8I:
  55         case GL_RGBA8:
  56         case GL_RGBA8_SNORM:
  57         case GL_RGBA8UI:
  58         case GL_RGBA8I:
  59         case GL_R11F_G11F_B10F:
  60         case GL_RGB10_A2:
  61         case GL_RGB10_A2UI:
  62         case GL_R16F:
  63         case GL_R16UI:
  64         case GL_R16I:
  65         case GL_RG16F:
  66         case GL_RG16UI:
  67         case GL_RG16I:
  68         case GL_RGBA16F:
  69         case GL_RGBA16UI:
  70         case GL_RGBA16I:
  71                 return false;
  72         case GL_R16:
  73         case GL_R16_SNORM:
  74         case GL_RG16:
  75         case GL_RG16_SNORM:
  76         case GL_RGBA16:
  77         case GL_RGBA16_SNORM:
  78         case GL_R32F:
  79         case GL_R32UI:
  80         case GL_R32I:
  81         case GL_RG32F:
  82         case GL_RG32UI:
  83         case GL_RG32I:
  84         case GL_RGBA32F:
  85         case GL_RGBA32UI:
  86         case GL_RGBA32I:
  87                 return true;
  88         default:
  89                 unreachable("Invalid image format");
  90         }
  91 }
  92
  93 /* Packs a 32-bit vector of colors in the range [0, (1 << bits[i]) - 1] to a
  94  * 32-bit SSA value, with as many channels as necessary to store all the bits
  95  */
  96 static nir_ssa_def *
  97 pack_bits(nir_builder *b, nir_ssa_def *color, const unsigned *bits,
  98           int num_components, bool mask)
  99 {
 100         nir_ssa_def *results[4];
 101         int offset = 0;
 102         for (int i = 0; i < num_components; i++) {
 103                 nir_ssa_def *chan = nir_channel(b, color, i);
 104
 105                 /* Channels being stored shouldn't cross a 32-bit boundary. */
 106                 assert((offset & ~31) == ((offset + bits[i] - 1) & ~31));
 107
 108                 if (mask) {
 109                         chan = nir_iand(b, chan,
 110                                         nir_imm_int(b, (1 << bits[i]) - 1));
 111                 }
 112
 113                 if (offset % 32 == 0) {
 114                         results[offset / 32] = chan;
 115                 } else {
 116                         results[offset / 32] =
 117                                 nir_ior(b, results[offset / 32],
 118                                         nir_ishl(b, chan,
 119                                                  nir_imm_int(b, offset % 32)));
 120                 }
 121                 offset += bits[i];
 122         }
 123
 124         return nir_vec(b, results, DIV_ROUND_UP(offset, 32));
 125 }
 126
 127 static nir_ssa_def *
 128 pack_unorm(nir_builder *b, nir_ssa_def *color, const unsigned *bits,
 129            int num_components)
 130 {
 131         color = nir_channels(b, color, (1 << num_components) - 1);
 132         color = nir_format_float_to_unorm(b, color, bits);
 133         return pack_bits(b, color, bits, color->num_components, false);
 134 }
 135
 136 static nir_ssa_def *
 137 pack_snorm(nir_builder *b, nir_ssa_def *color, const unsigned *bits,
 138            int num_components)
 139 {
 140         color = nir_channels(b, color, (1 << num_components) - 1);
 141         color = nir_format_float_to_snorm(b, color, bits);
 142         return pack_bits(b, color, bits, color->num_components, true);
 143 }
 144
 145 static nir_ssa_def *
 146 pack_uint(nir_builder *b, nir_ssa_def *color, const unsigned *bits,
 147           int num_components)
 148 {
 149         color = nir_channels(b, color, (1 << num_components) - 1);
 150         color = nir_format_clamp_uint(b, color, bits);
 151         return pack_bits(b, color, bits, num_components, false);
 152 }
 153
 154 static nir_ssa_def *
 155 pack_sint(nir_builder *b, nir_ssa_def *color, const unsigned *bits,
 156           int num_components)
 157 {
 158         color = nir_channels(b, color, (1 << num_components) - 1);
 159         color = nir_format_clamp_uint(b, color, bits);
 160         return pack_bits(b, color, bits, num_components, true);
 161 }
 162
 163 static nir_ssa_def *
 164 pack_half(nir_builder *b, nir_ssa_def *color, const unsigned *bits,
 165           int num_components)
 166 {
 167         color = nir_channels(b, color, (1 << num_components) - 1);
 168         color = nir_format_float_to_half(b, color);
 169         return pack_bits(b, color, bits, color->num_components, false);
 170 }
 171
 172 static void
 173 v3d_nir_lower_image_store(nir_builder *b, nir_intrinsic_instr *instr)
 174 {
 175         nir_variable *var = nir_intrinsic_get_var(instr, 0);
 176         GLenum format = var->data.image.format;
 177         static const unsigned bits_8[4] = {8, 8, 8, 8};
 178         static const unsigned bits_16[4] = {16, 16, 16, 16};
 179         static const unsigned bits_1010102[4] = {10, 10, 10, 2};
 180
 181         b->cursor = nir_before_instr(&instr->instr);
 182
 183         nir_ssa_def *unformatted = nir_ssa_for_src(b, instr->src[3], 4);
 184         nir_ssa_def *formatted = NULL;
 185         switch (format) {
 186         case GL_RGBA32F:
 187         case GL_RGBA32UI:
 188         case GL_RGBA32I:
 189                 /* For 4-component 32-bit components, there's no packing to be
 190                  * done.
 191                  */
 192                 return;
 193
 194         case GL_R32F:
 195         case GL_R32UI:
 196         case GL_R32I:
 197                 /* For other 32-bit components, just reduce the size of
 198                  * the input vector.
 199                  */
 200                 formatted = nir_channels(b, unformatted, 1);
 201                 break;
 202         case GL_RG32F:
 203         case GL_RG32UI:
 204         case GL_RG32I:
 205                 formatted = nir_channels(b, unformatted, 2);
 206                 break;
 207
 208         case GL_R8:
 209                 formatted = pack_unorm(b, unformatted, bits_8, 1);
 210                 break;
 211         case GL_RG8:
 212                 formatted = pack_unorm(b, unformatted, bits_8, 2);
 213                 break;
 214         case GL_RGBA8:
 215                 formatted = pack_unorm(b, unformatted, bits_8, 4);
 216                 break;
 217
 218         case GL_R8_SNORM:
 219                 formatted = pack_snorm(b, unformatted, bits_8, 1);
 220                 break;
 221         case GL_RG8_SNORM:
 222                 formatted = pack_snorm(b, unformatted, bits_8, 2);
 223                 break;
 224         case GL_RGBA8_SNORM:
 225                 formatted = pack_snorm(b, unformatted, bits_8, 4);
 226                 break;
 227
 228         case GL_R16:
 229                 formatted = pack_unorm(b, unformatted, bits_16, 1);
 230                 break;
 231         case GL_RG16:
 232                 formatted = pack_unorm(b, unformatted, bits_16, 2);
 233                 break;
 234         case GL_RGBA16:
 235                 formatted = pack_unorm(b, unformatted, bits_16, 4);
 236                 break;
 237
 238         case GL_R16_SNORM:
 239                 formatted = pack_snorm(b, unformatted, bits_16, 1);
 240                 break;
 241         case GL_RG16_SNORM:
 242                 formatted = pack_snorm(b, unformatted, bits_16, 2);
 243                 break;
 244         case GL_RGBA16_SNORM:
 245                 formatted = pack_snorm(b, unformatted, bits_16, 4);
 246                 break;
 247
 248         case GL_R16F:
 249                 formatted = pack_half(b, unformatted, bits_16, 1);
 250                 break;
 251         case GL_RG16F:
 252                 formatted = pack_half(b, unformatted, bits_16, 2);
 253                 break;
 254         case GL_RGBA16F:
 255                 formatted = pack_half(b, unformatted, bits_16, 4);
 256                 break;
 257
 258         case GL_R8UI:
 259                 formatted = pack_uint(b, unformatted, bits_8, 1);
 260                 break;
 261         case GL_R8I:
 262                 formatted = pack_sint(b, unformatted, bits_8, 1);
 263                 break;
 264         case GL_RG8UI:
 265                 formatted = pack_uint(b, unformatted, bits_8, 2);
 266                 break;
 267         case GL_RG8I:
 268                 formatted = pack_sint(b, unformatted, bits_8, 2);
 269                 break;
 270         case GL_RGBA8UI:
 271                 formatted = pack_uint(b, unformatted, bits_8, 4);
 272                 break;
 273         case GL_RGBA8I:
 274                 formatted = pack_sint(b, unformatted, bits_8, 4);
 275                 break;
 276
 277         case GL_R16UI:
 278                 formatted = pack_uint(b, unformatted, bits_16, 1);
 279                 break;
 280         case GL_R16I:
 281                 formatted = pack_sint(b, unformatted, bits_16, 1);
 282                 break;
 283         case GL_RG16UI:
 284                 formatted = pack_uint(b, unformatted, bits_16, 2);
 285                 break;
 286         case GL_RG16I:
 287                 formatted = pack_sint(b, unformatted, bits_16, 2);
 288                 break;
 289         case GL_RGBA16UI:
 290                 formatted = pack_uint(b, unformatted, bits_16, 4);
 291                 break;
 292         case GL_RGBA16I:
 293                 formatted = pack_sint(b, unformatted, bits_16, 4);
 294                 break;
 295
 296         case GL_R11F_G11F_B10F:
 297                 formatted = nir_format_pack_11f11f10f(b, unformatted);
 298                 break;
 299         case GL_RGB9_E5:
 300                 formatted = nir_format_pack_r9g9b9e5(b, unformatted);
 301                 break;
 302
 303         case GL_RGB10_A2:
 304                 formatted = pack_unorm(b, unformatted, bits_1010102, 4);
 305                 break;
 306
 307         case GL_RGB10_A2UI:
 308                 formatted = pack_uint(b, unformatted, bits_1010102, 4);
 309                 break;
 310
 311         default:
 312                 unreachable("bad format");
 313         }
 314
 315         nir_instr_rewrite_src(&instr->instr, &instr->src[3],
 316                               nir_src_for_ssa(formatted));
 317         instr->num_components = formatted->num_components;
 318 }
 319
 320 static void
 321 v3d_nir_lower_image_load(nir_builder *b, nir_intrinsic_instr *instr)
 322 {
 323         static const unsigned bits16[] = {16, 16, 16, 16};
 324         nir_variable *var = nir_intrinsic_get_var(instr, 0);
 325         const struct glsl_type *sampler_type = glsl_without_array(var->type);
 326         enum glsl_base_type base_type =
 327                 glsl_get_sampler_result_type(sampler_type);
 328
 329         if (v3d_gl_format_is_return_32(var->data.image.format))
 330                 return;
 331
 332         b->cursor = nir_after_instr(&instr->instr);
 333
 334         assert(instr->dest.is_ssa);
 335         nir_ssa_def *result = &instr->dest.ssa;
 336         if (base_type == GLSL_TYPE_FLOAT) {
 337             nir_ssa_def *rg = nir_channel(b, result, 0);
 338             nir_ssa_def *ba = nir_channel(b, result, 1);
 339             result = nir_vec4(b,
 340                               nir_unpack_half_2x16_split_x(b, rg),
 341                               nir_unpack_half_2x16_split_y(b, rg),
 342                               nir_unpack_half_2x16_split_x(b, ba),
 343                               nir_unpack_half_2x16_split_y(b, ba));
 344         } else if (base_type == GLSL_TYPE_INT) {
 345                 result = nir_format_unpack_sint(b, result, bits16, 4);
 346         } else {
 347                 assert(base_type == GLSL_TYPE_UINT);
 348                 result = nir_format_unpack_uint(b, result, bits16, 4);
 349         }
 350
 351         nir_ssa_def_rewrite_uses_after(&instr->dest.ssa, nir_src_for_ssa(result),
 352                                        result->parent_instr);
 353 }
 354
 355 void
 356 v3d_nir_lower_image_load_store(nir_shader *s)
 357 {
 358         nir_foreach_function(function, s) {
 359                 if (!function->impl)
 360                         continue;
 361
 362                 nir_builder b;
 363                 nir_builder_init(&b, function->impl);
 364
 365                 nir_foreach_block(block, function->impl) {
 366                         nir_foreach_instr_safe(instr, block) {
 367                                 if (instr->type != nir_instr_type_intrinsic)
 368                                         continue;
 369
 370                                 nir_intrinsic_instr *intr =
 371                                         nir_instr_as_intrinsic(instr);
 372
 373                                 switch (intr->intrinsic) {
 374                                 case nir_intrinsic_image_deref_load:
 375                                         v3d_nir_lower_image_load(&b, intr);
 376                                         break;
 377                                 case nir_intrinsic_image_deref_store:
 378                                         v3d_nir_lower_image_store(&b, intr);
 379                                         break;
 380                                 default:
 381                                         break;
 382                                 }
 383                         }
 384                 }
 385
 386                 nir_metadata_preserve(function->impl,
 387                                       nir_metadata_block_index |
 388                                       nir_metadata_dominance);
 389         }
 390 }