| 1 |
/* |
| 2 |
* Copyright © 2024 Imagination Technologies Ltd. |
| 3 |
* |
| 4 |
* SPDX-License-Identifier: MIT |
| 5 |
*/ |
| 6 |
|
| 7 |
/** |
| 8 |
* \file pco_nir_pvfio.c |
| 9 |
* |
| 10 |
* \brief PCO NIR per-vertex/fragment input/output passes. |
| 11 |
*/ |
| 12 |
|
| 13 |
#include "compiler/glsl_types.h" |
| 14 |
#include "compiler/shader_enums.h" |
| 15 |
#include "nir.h" |
| 16 |
#include "nir_builder.h" |
| 17 |
#include "nir_format_convert.h" |
| 18 |
#include "pco.h" |
| 19 |
#include "pco_builder.h" |
| 20 |
#include "pco_internal.h" |
| 21 |
#include "util/macros.h" |
| 22 |
#include "util/u_dynarray.h" |
| 23 |
|
| 24 |
#include <assert.h> |
| 25 |
#include <stdbool.h> |
| 26 |
#include <stdio.h> |
| 27 |
|
| 28 |
/** Per-fragment output pass state. */ |
| 29 |
struct pfo_state { |
| 30 |
struct util_dynarray loads; /** List of fragment loads. */ |
| 31 |
struct util_dynarray stores; /** List of fragment stores. */ |
| 32 |
|
| 33 |
/* Src for depth feedback (NULL if unused). */ |
| 34 |
nir_def *depth_feedback_src; |
| 35 |
|
| 36 |
nir_def *discard_cond_reg; |
| 37 |
bool has_discards; |
| 38 |
|
| 39 |
nir_intrinsic_instr *last_discard_store; |
| 40 |
|
| 41 |
bool has_sample_check; |
| 42 |
|
| 43 |
/* nir_instr *terminate; */ |
| 44 |
|
| 45 |
pco_fs_data *fs; /** Fragment-specific data. */ |
| 46 |
}; |
| 47 |
|
| 48 |
/** Per-vertex input pass state. */ |
| 49 |
struct pvi_state { |
| 50 |
nir_def *attribs[MAX_VERTEX_GENERIC_ATTRIBS]; /** Loaded vertex attribs. */ |
| 51 |
pco_vs_data *vs; /** Vertex-specific data. */ |
| 52 |
}; |
| 53 |
|
| 54 |
/** |
| 55 |
* \brief Returns the GLSL base type equivalent of a pipe format. |
| 56 |
* |
| 57 |
* \param[in] format Pipe format. |
| 58 |
* \return The GLSL base type, or GLSL_TYPE_ERROR if unsupported/invalid. |
| 59 |
*/ |
| 60 |
static inline enum glsl_base_type base_type_from_fmt(enum pipe_format format) |
| 61 |
{ |
| 62 |
const struct util_format_description *desc = util_format_description(format); |
| 63 |
int chan = util_format_get_first_non_void_channel(format); |
| 64 |
if (chan < 0) |
| 65 |
return GLSL_TYPE_ERROR; |
| 66 |
|
| 67 |
switch (desc->channel[chan].type) { |
| 68 |
case UTIL_FORMAT_TYPE_UNSIGNED: |
| 69 |
return GLSL_TYPE_UINT; |
| 70 |
|
| 71 |
case UTIL_FORMAT_TYPE_SIGNED: |
| 72 |
return GLSL_TYPE_INT; |
| 73 |
|
| 74 |
case UTIL_FORMAT_TYPE_FLOAT: |
| 75 |
return GLSL_TYPE_FLOAT; |
| 76 |
|
| 77 |
default: |
| 78 |
break; |
| 79 |
} |
| 80 |
|
| 81 |
return GLSL_TYPE_ERROR; |
| 82 |
} |
| 83 |
|
| 84 |
static enum pipe_format |
| 85 |
to_pbe_format(nir_builder *b, enum pipe_format format, nir_def **input) |
| 86 |
{ |
| 87 |
switch (format) { |
| 88 |
case PIPE_FORMAT_B5G6R5_UNORM: |
| 89 |
return PIPE_FORMAT_R8G8B8_UNORM; |
| 90 |
|
| 91 |
case PIPE_FORMAT_A4R4G4B4_UNORM: |
| 92 |
case PIPE_FORMAT_B5G5R5A1_UNORM: |
| 93 |
case PIPE_FORMAT_B8G8R8A8_UNORM: |
| 94 |
return PIPE_FORMAT_R8G8B8A8_UNORM; |
| 95 |
|
| 96 |
case PIPE_FORMAT_R8G8B8A8_SRGB: |
| 97 |
case PIPE_FORMAT_B8G8R8A8_SRGB: |
| 98 |
if (input) |
| 99 |
*input = nir_fsat(b, *input); |
| 100 |
FALLTHROUGH; |
| 101 |
|
| 102 |
case PIPE_FORMAT_R10G10B10A2_UNORM: |
| 103 |
return PIPE_FORMAT_R16G16B16A16_FLOAT; |
| 104 |
|
| 105 |
case PIPE_FORMAT_R11G11B10_FLOAT: |
| 106 |
return PIPE_FORMAT_R16G16B16_FLOAT; |
| 107 |
|
| 108 |
/* For loadops. */ |
| 109 |
case PIPE_FORMAT_Z32_FLOAT: |
| 110 |
case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT: |
| 111 |
case PIPE_FORMAT_Z24_UNORM_S8_UINT: |
| 112 |
case PIPE_FORMAT_Z16_UNORM: |
| 113 |
case PIPE_FORMAT_Z24X8_UNORM: |
| 114 |
assert(b->shader->info.internal); |
| 115 |
return PIPE_FORMAT_R32_FLOAT; |
| 116 |
|
| 117 |
default: |
| 118 |
break; |
| 119 |
} |
| 120 |
|
| 121 |
return format; |
| 122 |
} |
| 123 |
|
| 124 |
static unsigned format_chans_per_dword(enum pipe_format format) |
| 125 |
{ |
| 126 |
const struct util_format_description *desc = util_format_description(format); |
| 127 |
unsigned accum_bits = 0; |
| 128 |
|
| 129 |
for (unsigned u = 0; u < desc->nr_channels; ++u) { |
| 130 |
/* Exactly one dword, great! */ |
| 131 |
if (accum_bits == 32) |
| 132 |
return u; |
| 133 |
|
| 134 |
/* Went over, back off by one. */ |
| 135 |
if (accum_bits > 32) { |
| 136 |
/* We don't support formats with channels > 1 dword. */ |
| 137 |
assert(u > 1); |
| 138 |
return u - 1; |
| 139 |
} |
| 140 |
|
| 141 |
accum_bits += desc->channel[u].size; |
| 142 |
} |
| 143 |
|
| 144 |
/* Loop finished, all channels can fit. */ |
| 145 |
return desc->nr_channels; |
| 146 |
} |
| 147 |
|
| 148 |
static nir_def *pack_to_format(nir_builder *b, |
| 149 |
nir_def *input, |
| 150 |
nir_alu_type src_type, |
| 151 |
enum pipe_format format) |
| 152 |
{ |
| 153 |
const struct util_format_description *desc = util_format_description(format); |
| 154 |
|
| 155 |
nir_def *zero = nir_imm_int(b, 0); |
| 156 |
nir_def *one = nir_alu_type_get_base_type(src_type) == nir_type_float |
| 157 |
? nir_imm_float(b, 1.0f) |
| 158 |
: nir_imm_int(b, 1); |
| 159 |
|
| 160 |
nir_def *input_comps[4]; |
| 161 |
/* Populate any missing components .*/ |
| 162 |
for (unsigned u = 0; u < ARRAY_SIZE(input_comps); ++u) { |
| 163 |
enum pipe_swizzle s = desc->swizzle[u]; |
| 164 |
|
| 165 |
if (s <= PIPE_SWIZZLE_W) |
| 166 |
input_comps[u] = nir_channel(b, input, s); |
| 167 |
else if (s == PIPE_SWIZZLE_0) |
| 168 |
input_comps[u] = zero; |
| 169 |
else if (s == PIPE_SWIZZLE_1) |
| 170 |
input_comps[u] = one; |
| 171 |
else |
| 172 |
UNREACHABLE(""); |
| 173 |
} |
| 174 |
|
| 175 |
unsigned format_bits = util_format_get_blocksizebits(format); |
| 176 |
unsigned format_dwords = DIV_ROUND_UP(format_bits, 32); |
| 177 |
nir_def *packed_comps[] = { zero, zero, zero, zero }; |
| 178 |
|
| 179 |
/* Special case: no packing required. */ |
| 180 |
if (util_format_get_max_channel_size(format) == 32) |
| 181 |
return nir_vec(b, input_comps, format_dwords); |
| 182 |
|
| 183 |
/* Special case: can't be packed with op, need bit-packing instead. */ |
| 184 |
if (util_format_is_pure_integer(format)) { |
| 185 |
for (unsigned u = 0; u < desc->nr_channels; ++u) { |
| 186 |
unsigned dword = desc->channel[u].shift / 32; |
| 187 |
unsigned offset = desc->channel[u].shift % 32; |
| 188 |
unsigned size = desc->channel[u].size; |
| 189 |
|
| 190 |
packed_comps[dword] = nir_bitfield_insert_imm(b, |
| 191 |
packed_comps[dword], |
| 192 |
input_comps[u], |
| 193 |
offset, |
| 194 |
size); |
| 195 |
} |
| 196 |
|
| 197 |
return nir_vec(b, packed_comps, format_dwords); |
| 198 |
} |
| 199 |
|
| 200 |
unsigned chans_per_dword = format_chans_per_dword(format); |
| 201 |
unsigned chans_remaining = desc->nr_channels; |
| 202 |
input = nir_vec(b, input_comps, desc->nr_channels); |
| 203 |
for (unsigned u = 0; u < format_dwords; ++u) { |
| 204 |
unsigned chans_to_pack = |
| 205 |
chans_remaining > chans_per_dword ? chans_per_dword : chans_remaining; |
| 206 |
unsigned chans_packed = desc->nr_channels - chans_remaining; |
| 207 |
|
| 208 |
nir_def *input_chans = |
| 209 |
nir_channels(b, input, BITFIELD_RANGE(chans_packed, chans_to_pack)); |
| 210 |
packed_comps[u] = nir_pack_pco(b, input_chans, .format = format); |
| 211 |
|
| 212 |
chans_remaining -= chans_to_pack; |
| 213 |
} |
| 214 |
|
| 215 |
assert(!chans_remaining); |
| 216 |
return nir_vec(b, packed_comps, format_dwords); |
| 217 |
} |
| 218 |
|
| 219 |
static nir_def *unpack_from_format(nir_builder *b, |
| 220 |
nir_def *packed_comps[static 4], |
| 221 |
nir_alu_type dest_type, |
| 222 |
enum pipe_format format, |
| 223 |
unsigned components_needed) |
| 224 |
{ |
| 225 |
const struct util_format_description *desc = util_format_description(format); |
| 226 |
|
| 227 |
nir_def *unpacked_comps[4]; |
| 228 |
|
| 229 |
unsigned format_bits = util_format_get_blocksizebits(format); |
| 230 |
unsigned format_dwords = DIV_ROUND_UP(format_bits, 32); |
| 231 |
|
| 232 |
/* Special case: no unpacking required. */ |
| 233 |
if (util_format_get_max_channel_size(format) == 32) { |
| 234 |
for (unsigned u = 0; u < desc->nr_channels; ++u) |
| 235 |
unpacked_comps[u] = packed_comps[u]; |
| 236 |
} |
| 237 |
|
| 238 |
/* Special case: can't be unpacked with op, need bit-unpacking instead. */ |
| 239 |
else if (util_format_is_pure_integer(format)) { |
| 240 |
nir_def *(*nir_bitfield_extract_imm)(nir_builder *, |
| 241 |
nir_def *, |
| 242 |
uint32_t, |
| 243 |
uint32_t) = |
| 244 |
util_format_is_pure_uint(format) ? nir_ubitfield_extract_imm |
| 245 |
: nir_ibitfield_extract_imm; |
| 246 |
|
| 247 |
for (unsigned u = 0; u < desc->nr_channels; ++u) { |
| 248 |
unsigned dword = desc->channel[u].shift / 32; |
| 249 |
unsigned offset = desc->channel[u].shift % 32; |
| 250 |
unsigned size = desc->channel[u].size; |
| 251 |
|
| 252 |
unpacked_comps[u] = |
| 253 |
nir_bitfield_extract_imm(b, packed_comps[dword], offset, size); |
| 254 |
} |
| 255 |
} |
| 256 |
|
| 257 |
else { |
| 258 |
unsigned chans_per_dword = format_chans_per_dword(format); |
| 259 |
unsigned chans_remaining = desc->nr_channels; |
| 260 |
|
| 261 |
for (unsigned u = 0; u < format_dwords; ++u) { |
| 262 |
unsigned chans_to_unpack = chans_remaining > chans_per_dword |
| 263 |
? chans_per_dword |
| 264 |
: chans_remaining; |
| 265 |
|
| 266 |
nir_def *unpacked = nir_unpack_pco(b, |
| 267 |
chans_to_unpack, |
| 268 |
packed_comps[u], |
| 269 |
.format = format); |
| 270 |
|
| 271 |
unsigned chans_unpacked = desc->nr_channels - chans_remaining; |
| 272 |
for (unsigned v = 0; v < chans_to_unpack; ++v) |
| 273 |
unpacked_comps[chans_unpacked + v] = nir_channel(b, unpacked, v); |
| 274 |
|
| 275 |
chans_remaining -= chans_to_unpack; |
| 276 |
} |
| 277 |
|
| 278 |
assert(!chans_remaining); |
| 279 |
} |
| 280 |
|
| 281 |
nir_def *zero = nir_imm_int(b, 0); |
| 282 |
nir_def *one = nir_alu_type_get_base_type(dest_type) == nir_type_float |
| 283 |
? nir_imm_float(b, 1.0f) |
| 284 |
: nir_imm_int(b, 1); |
| 285 |
|
| 286 |
nir_def *output_comps[4]; |
| 287 |
/* Populate any missing components .*/ |
| 288 |
for (unsigned u = 0; u < ARRAY_SIZE(output_comps); ++u) { |
| 289 |
enum pipe_swizzle s = desc->swizzle[u]; |
| 290 |
|
| 291 |
if (s <= PIPE_SWIZZLE_W) |
| 292 |
output_comps[u] = unpacked_comps[s]; |
| 293 |
else if (s == PIPE_SWIZZLE_0) |
| 294 |
output_comps[u] = zero; |
| 295 |
else if (s == PIPE_SWIZZLE_1) |
| 296 |
output_comps[u] = one; |
| 297 |
else |
| 298 |
UNREACHABLE(""); |
| 299 |
} |
| 300 |
|
| 301 |
return nir_vec(b, output_comps, components_needed); |
| 302 |
} |
| 303 |
|
| 304 |
static inline bool is_processed(nir_intrinsic_instr *intr) |
| 305 |
{ |
| 306 |
nir_alu_type type; |
| 307 |
|
| 308 |
if (nir_intrinsic_has_src_type(intr)) |
| 309 |
type = nir_intrinsic_src_type(intr); |
| 310 |
else if (nir_intrinsic_has_dest_type(intr)) |
| 311 |
type = nir_intrinsic_dest_type(intr); |
| 312 |
else |
| 313 |
return true; |
| 314 |
|
| 315 |
return nir_alu_type_get_base_type(type) == nir_type_invalid; |
| 316 |
} |
| 317 |
|
| 318 |
static nir_def *lower_pfo_store(nir_builder *b, |
| 319 |
nir_intrinsic_instr *intr, |
| 320 |
struct pfo_state *state) |
| 321 |
{ |
| 322 |
/* Skip stores we've already processed. */ |
| 323 |
if (is_processed(intr)) { |
| 324 |
util_dynarray_append(&state->stores, intr); |
| 325 |
return NULL; |
| 326 |
} |
| 327 |
|
| 328 |
nir_def *input = intr->src[0].ssa; |
| 329 |
nir_src *offset = &intr->src[1]; |
| 330 |
assert(nir_src_as_uint(*offset) == 0); |
| 331 |
|
| 332 |
ASSERTED unsigned bit_size = input->bit_size; |
| 333 |
assert(bit_size == 32); |
| 334 |
|
| 335 |
unsigned component = nir_intrinsic_component(intr); |
| 336 |
assert(!component); |
| 337 |
|
| 338 |
struct nir_io_semantics io_semantics = nir_intrinsic_io_semantics(intr); |
| 339 |
gl_frag_result location = io_semantics.location; |
| 340 |
|
| 341 |
b->cursor = nir_before_instr(&intr->instr); |
| 342 |
|
| 343 |
enum pipe_format format = state->fs->output_formats[location]; |
| 344 |
if (format == PIPE_FORMAT_NONE) |
| 345 |
return NIR_LOWER_INSTR_PROGRESS_REPLACE; |
| 346 |
|
| 347 |
format = to_pbe_format(b, format, &input); |
| 348 |
|
| 349 |
nir_alu_type src_type = nir_intrinsic_src_type(intr); |
| 350 |
nir_def *output = pack_to_format(b, input, src_type, format); |
| 351 |
|
| 352 |
/* Emit and track the new store. */ |
| 353 |
nir_intrinsic_instr *store = |
| 354 |
nir_store_output(b, |
| 355 |
output, |
| 356 |
offset->ssa, |
| 357 |
.base = nir_intrinsic_base(intr), |
| 358 |
.write_mask = BITFIELD_MASK(output->num_components), |
| 359 |
.src_type = nir_type_invalid | 32, |
| 360 |
.component = component, |
| 361 |
.io_semantics = io_semantics, |
| 362 |
.io_xfb = nir_intrinsic_io_xfb(intr)); |
| 363 |
|
| 364 |
util_dynarray_append(&state->stores, store); |
| 365 |
|
| 366 |
/* Update the type of the stored variable. */ |
| 367 |
nir_variable *var = |
| 368 |
nir_find_variable_with_location(b->shader, nir_var_shader_out, location); |
| 369 |
assert(var); |
| 370 |
var->type = glsl_uvec_type(output->num_components); |
| 371 |
|
| 372 |
return NIR_LOWER_INSTR_PROGRESS_REPLACE; |
| 373 |
} |
| 374 |
|
| 375 |
static nir_def *lower_pfo_load(nir_builder *b, |
| 376 |
nir_intrinsic_instr *intr, |
| 377 |
struct pfo_state *state) |
| 378 |
{ |
| 379 |
/* Skip loads we've already processed. */ |
| 380 |
if (is_processed(intr)) { |
| 381 |
util_dynarray_append(&state->loads, intr); |
| 382 |
return NULL; |
| 383 |
} |
| 384 |
|
| 385 |
unsigned base = nir_intrinsic_base(intr); |
| 386 |
|
| 387 |
nir_src *offset = &intr->src[0]; |
| 388 |
assert(nir_src_as_uint(*offset) == 0); |
| 389 |
|
| 390 |
struct nir_io_semantics io_semantics = nir_intrinsic_io_semantics(intr); |
| 391 |
gl_frag_result location = io_semantics.location; |
| 392 |
|
| 393 |
enum pipe_format format; |
| 394 |
/* Special case for input attachments. */ |
| 395 |
if (location == FRAG_RESULT_COLOR) { |
| 396 |
format = state->fs->ia_formats[base]; |
| 397 |
} else { |
| 398 |
assert(location >= FRAG_RESULT_DATA0); |
| 399 |
assert(!base); |
| 400 |
format = state->fs->output_formats[location]; |
| 401 |
} |
| 402 |
|
| 403 |
if (format == PIPE_FORMAT_NONE) |
| 404 |
return nir_undef(b, intr->def.num_components, intr->def.bit_size); |
| 405 |
|
| 406 |
format = to_pbe_format(b, format, NULL); |
| 407 |
|
| 408 |
nir_def *packed_comps[4]; |
| 409 |
for (unsigned c = 0; c < ARRAY_SIZE(packed_comps); ++c) { |
| 410 |
packed_comps[c] = nir_load_output(b, |
| 411 |
1, |
| 412 |
32, |
| 413 |
offset->ssa, |
| 414 |
.base = base, |
| 415 |
.component = c, |
| 416 |
.dest_type = nir_type_invalid | 32, |
| 417 |
.io_semantics = io_semantics); |
| 418 |
|
| 419 |
nir_intrinsic_instr *load = |
| 420 |
nir_def_as_intrinsic(packed_comps[c]); |
| 421 |
|
| 422 |
util_dynarray_append(&state->loads, load); |
| 423 |
} |
| 424 |
|
| 425 |
nir_alu_type dest_type = nir_intrinsic_dest_type(intr); |
| 426 |
return unpack_from_format(b, |
| 427 |
packed_comps, |
| 428 |
dest_type, |
| 429 |
format, |
| 430 |
intr->def.num_components); |
| 431 |
} |
| 432 |
|
| 433 |
/** |
| 434 |
* \brief Filters PFO-related instructions. |
| 435 |
* |
| 436 |
* \param[in] instr NIR instruction. |
| 437 |
* \param[in] cb_data User callback data. |
| 438 |
* \return True if the instruction matches the filter. |
| 439 |
*/ |
| 440 |
static bool is_pfo(const nir_instr *instr, UNUSED const void *cb_data) |
| 441 |
{ |
| 442 |
if (instr->type != nir_instr_type_intrinsic) |
| 443 |
return false; |
| 444 |
|
| 445 |
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); |
| 446 |
switch (intr->intrinsic) { |
| 447 |
case nir_intrinsic_store_output: |
| 448 |
case nir_intrinsic_load_output: |
| 449 |
case nir_intrinsic_demote: |
| 450 |
case nir_intrinsic_demote_if: |
| 451 |
return true; |
| 452 |
|
| 453 |
default: |
| 454 |
break; |
| 455 |
} |
| 456 |
|
| 457 |
return false; |
| 458 |
} |
| 459 |
|
| 460 |
/** |
| 461 |
* \brief Lowers a PFO-related instruction. |
| 462 |
* |
| 463 |
* \param[in] b NIR builder. |
| 464 |
* \param[in] instr NIR instruction. |
| 465 |
* \param[in] cb_data User callback data. |
| 466 |
* \return The replacement/lowered def. |
| 467 |
*/ |
| 468 |
static nir_def *lower_pfo(nir_builder *b, nir_instr *instr, void *cb_data) |
| 469 |
{ |
| 470 |
struct pfo_state *state = cb_data; |
| 471 |
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); |
| 472 |
|
| 473 |
switch (intr->intrinsic) { |
| 474 |
case nir_intrinsic_store_output: { |
| 475 |
nir_io_semantics sem = nir_intrinsic_io_semantics(intr); |
| 476 |
if (sem.location >= FRAG_RESULT_DATA0) |
| 477 |
return lower_pfo_store(b, intr, state); |
| 478 |
|
| 479 |
if (sem.location == FRAG_RESULT_DEPTH) { |
| 480 |
assert(!state->depth_feedback_src); |
| 481 |
state->depth_feedback_src = nir_fsat(b, intr->src[0].ssa); |
| 482 |
|
| 483 |
return NIR_LOWER_INSTR_PROGRESS_REPLACE; |
| 484 |
} |
| 485 |
|
| 486 |
if (sem.location == FRAG_RESULT_SAMPLE_MASK) { |
| 487 |
nir_def *smp_msk = |
| 488 |
nir_ishl(b, nir_imm_int(b, 1), nir_load_sample_id(b)); |
| 489 |
|
| 490 |
smp_msk = nir_iand(b, smp_msk, nir_load_sample_mask_in(b)); |
| 491 |
smp_msk = nir_iand(b, smp_msk, intr->src[0].ssa); |
| 492 |
nir_def *cond = nir_ieq_imm(b, smp_msk, 0); |
| 493 |
|
| 494 |
state->has_discards = true; |
| 495 |
state->has_sample_check = true; |
| 496 |
nir_def *val = nir_load_reg(b, state->discard_cond_reg); |
| 497 |
val = nir_ior(b, val, cond); |
| 498 |
state->last_discard_store = |
| 499 |
nir_build_store_reg(b, val, state->discard_cond_reg); |
| 500 |
return NIR_LOWER_INSTR_PROGRESS_REPLACE; |
| 501 |
} |
| 502 |
|
| 503 |
UNREACHABLE(""); |
| 504 |
} |
| 505 |
|
| 506 |
case nir_intrinsic_load_output: |
| 507 |
return lower_pfo_load(b, intr, state); |
| 508 |
|
| 509 |
case nir_intrinsic_demote: |
| 510 |
state->has_discards = true; |
| 511 |
state->last_discard_store = |
| 512 |
nir_build_store_reg(b, nir_imm_true(b), state->discard_cond_reg); |
| 513 |
return NIR_LOWER_INSTR_PROGRESS_REPLACE; |
| 514 |
|
| 515 |
case nir_intrinsic_demote_if: { |
| 516 |
state->has_discards = true; |
| 517 |
nir_def *val = nir_load_reg(b, state->discard_cond_reg); |
| 518 |
val = nir_ior(b, val, intr->src[0].ssa); |
| 519 |
state->last_discard_store = |
| 520 |
nir_build_store_reg(b, val, state->discard_cond_reg); |
| 521 |
return NIR_LOWER_INSTR_PROGRESS_REPLACE; |
| 522 |
} |
| 523 |
|
| 524 |
default: |
| 525 |
break; |
| 526 |
} |
| 527 |
|
| 528 |
return false; |
| 529 |
} |
| 530 |
|
| 531 |
static bool lower_isp_fb(nir_builder *b, struct pfo_state *state) |
| 532 |
{ |
| 533 |
if (b->shader->info.internal) |
| 534 |
return false; |
| 535 |
|
| 536 |
bool has_depth_feedback = !!state->depth_feedback_src; |
| 537 |
if (b->shader->info.writes_memory && !has_depth_feedback) { |
| 538 |
nir_variable *var_pos = nir_get_variable_with_location(b->shader, |
| 539 |
nir_var_shader_in, |
| 540 |
VARYING_SLOT_POS, |
| 541 |
glsl_vec4_type()); |
| 542 |
var_pos->data.interpolation = INTERP_MODE_NOPERSPECTIVE; |
| 543 |
|
| 544 |
b->cursor = nir_before_block( |
| 545 |
nir_start_block(nir_shader_get_entrypoint(b->shader))); |
| 546 |
|
| 547 |
state->depth_feedback_src = |
| 548 |
nir_load_input(b, |
| 549 |
1, |
| 550 |
32, |
| 551 |
nir_imm_int(b, 0), |
| 552 |
.component = 2, |
| 553 |
.dest_type = nir_type_float32, |
| 554 |
.io_semantics = (nir_io_semantics){ |
| 555 |
.location = VARYING_SLOT_POS, |
| 556 |
.num_slots = 1, |
| 557 |
}); |
| 558 |
|
| 559 |
has_depth_feedback = true; |
| 560 |
} |
| 561 |
|
| 562 |
if (!state->has_sample_check) { |
| 563 |
b->cursor = nir_after_instr(&state->last_discard_store->instr); |
| 564 |
|
| 565 |
nir_def *smp_msk = nir_ishl(b, nir_imm_int(b, 1), nir_load_sample_id(b)); |
| 566 |
smp_msk = nir_iand(b, smp_msk, nir_load_sample_mask_in(b)); |
| 567 |
nir_def *cond = nir_ieq_imm(b, smp_msk, 0); |
| 568 |
|
| 569 |
nir_def *val = nir_load_reg(b, state->discard_cond_reg); |
| 570 |
val = nir_ior(b, val, cond); |
| 571 |
state->last_discard_store = |
| 572 |
nir_build_store_reg(b, val, state->discard_cond_reg); |
| 573 |
|
| 574 |
state->has_sample_check = true; |
| 575 |
state->has_discards = true; |
| 576 |
} |
| 577 |
|
| 578 |
/* Insert isp feedback instruction before the first store, |
| 579 |
* or if there are no stores, at the end. |
| 580 |
*/ |
| 581 |
if (state->stores.size > 0) |
| 582 |
b->cursor = nir_before_instr( |
| 583 |
&(*(nir_intrinsic_instr **)util_dynarray_begin(&state->stores))->instr); |
| 584 |
else |
| 585 |
b->cursor = nir_after_block( |
| 586 |
nir_impl_last_block(nir_shader_get_entrypoint(b->shader))); |
| 587 |
|
| 588 |
nir_def *undef = nir_undef(b, 1, 32); |
| 589 |
|
| 590 |
nir_isp_feedback_pco( |
| 591 |
b, |
| 592 |
state->has_discards ? nir_i2b(b, nir_load_reg(b, state->discard_cond_reg)) |
| 593 |
: undef, |
| 594 |
has_depth_feedback ? state->depth_feedback_src : undef); |
| 595 |
|
| 596 |
state->fs->uses.discard = state->has_discards; |
| 597 |
state->fs->uses.depth_feedback = has_depth_feedback; |
| 598 |
|
| 599 |
return true; |
| 600 |
} |
| 601 |
|
| 602 |
static bool sink_outputs(nir_shader *shader, struct pfo_state *state) |
| 603 |
{ |
| 604 |
bool progress = false; |
| 605 |
|
| 606 |
nir_instr *after_instr = nir_block_last_instr( |
| 607 |
nir_impl_last_block(nir_shader_get_entrypoint(shader))); |
| 608 |
|
| 609 |
util_dynarray_foreach (&state->stores, nir_intrinsic_instr *, store) { |
| 610 |
nir_instr *instr = &(*store)->instr; |
| 611 |
|
| 612 |
progress |= nir_instr_move(nir_after_instr(after_instr), instr); |
| 613 |
after_instr = instr; |
| 614 |
} |
| 615 |
|
| 616 |
return progress; |
| 617 |
} |
| 618 |
|
| 619 |
static bool z_replicate(nir_shader *shader, struct pfo_state *state) |
| 620 |
{ |
| 621 |
if (shader->info.internal || state->fs->z_replicate == ~0u) |
| 622 |
return false; |
| 623 |
|
| 624 |
assert(!nir_find_variable_with_location(shader, |
| 625 |
nir_var_shader_out, |
| 626 |
state->fs->z_replicate)); |
| 627 |
|
| 628 |
nir_create_variable_with_location(shader, |
| 629 |
nir_var_shader_out, |
| 630 |
state->fs->z_replicate, |
| 631 |
glsl_float_type()); |
| 632 |
|
| 633 |
if (!state->depth_feedback_src) { |
| 634 |
nir_variable *var_pos = nir_get_variable_with_location(shader, |
| 635 |
nir_var_shader_in, |
| 636 |
VARYING_SLOT_POS, |
| 637 |
glsl_vec4_type()); |
| 638 |
var_pos->data.interpolation = INTERP_MODE_NOPERSPECTIVE; |
| 639 |
|
| 640 |
nir_builder b = nir_builder_at( |
| 641 |
nir_before_block(nir_start_block(nir_shader_get_entrypoint(shader)))); |
| 642 |
|
| 643 |
state->depth_feedback_src = |
| 644 |
nir_load_input(&b, |
| 645 |
1, |
| 646 |
32, |
| 647 |
nir_imm_int(&b, 0), |
| 648 |
.component = 2, |
| 649 |
.dest_type = nir_type_float32, |
| 650 |
.io_semantics = (nir_io_semantics){ |
| 651 |
.location = VARYING_SLOT_POS, |
| 652 |
.num_slots = 1, |
| 653 |
}); |
| 654 |
} |
| 655 |
|
| 656 |
nir_builder b = nir_builder_at( |
| 657 |
nir_after_block(nir_impl_last_block(nir_shader_get_entrypoint(shader)))); |
| 658 |
nir_store_output(&b, |
| 659 |
state->depth_feedback_src, |
| 660 |
nir_imm_int(&b, 0), |
| 661 |
.write_mask = 1, |
| 662 |
.src_type = nir_type_invalid | 32, |
| 663 |
.io_semantics = (nir_io_semantics){ |
| 664 |
.location = state->fs->z_replicate, |
| 665 |
.num_slots = 1, |
| 666 |
}); |
| 667 |
|
| 668 |
return true; |
| 669 |
} |
| 670 |
|
| 671 |
static bool is_frag_color_out(const nir_instr *instr, |
| 672 |
UNUSED const void *cb_data) |
| 673 |
{ |
| 674 |
if (instr->type != nir_instr_type_intrinsic) |
| 675 |
return false; |
| 676 |
|
| 677 |
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); |
| 678 |
if (intr->intrinsic != nir_intrinsic_store_output) |
| 679 |
return false; |
| 680 |
|
| 681 |
gl_frag_result location = nir_intrinsic_io_semantics(intr).location; |
| 682 |
return location >= FRAG_RESULT_DATA0 && location < FRAG_RESULT_MAX; |
| 683 |
} |
| 684 |
|
| 685 |
static bool lower_demote_samples(nir_builder *b, |
| 686 |
nir_intrinsic_instr *intr, |
| 687 |
UNUSED void *cb_data) |
| 688 |
{ |
| 689 |
if (intr->intrinsic != nir_intrinsic_demote_samples) |
| 690 |
return false; |
| 691 |
|
| 692 |
b->cursor = nir_before_instr(&intr->instr); |
| 693 |
nir_def *to_keep = nir_u2u32(b, nir_inot(b, intr->src[0].ssa)); |
| 694 |
nir_def *sample_mask = nir_load_savmsk_vm_pco(b); |
| 695 |
nir_def *current_mask = |
| 696 |
nir_ishl(b, nir_imm_int(b, 1), nir_load_sample_id(b)); |
| 697 |
nir_def *cond = nir_iand(b, to_keep, nir_iand(b, sample_mask, current_mask)); |
| 698 |
nir_demote_if(b, nir_ieq_imm(b, cond, 0)); |
| 699 |
|
| 700 |
nir_instr_remove(&intr->instr); |
| 701 |
|
| 702 |
return true; |
| 703 |
} |
| 704 |
|
| 705 |
bool pco_nir_lower_alpha_to_coverage(nir_shader *shader) |
| 706 |
{ |
| 707 |
if (shader->info.internal) |
| 708 |
return false; |
| 709 |
|
| 710 |
nir_builder b = nir_builder_create(nir_shader_get_entrypoint(shader)); |
| 711 |
b.cursor = |
| 712 |
nir_before_block(nir_start_block(nir_shader_get_entrypoint(shader))); |
| 713 |
nir_def *a2c_enabled = nir_ine_imm( |
| 714 |
&b, |
| 715 |
nir_ubitfield_extract_imm(&b, nir_load_fs_meta_pco(&b), 25, 1), |
| 716 |
0); |
| 717 |
|
| 718 |
nir_lower_alpha_to_coverage(shader, true, a2c_enabled); |
| 719 |
|
| 720 |
nir_shader_intrinsics_pass(shader, |
| 721 |
lower_demote_samples, |
| 722 |
nir_metadata_control_flow, |
| 723 |
NULL); |
| 724 |
|
| 725 |
return true; |
| 726 |
} |
| 727 |
|
| 728 |
static nir_def * |
| 729 |
lower_alpha_to_one(nir_builder *b, nir_instr *instr, UNUSED void *cb_data) |
| 730 |
{ |
| 731 |
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); |
| 732 |
|
| 733 |
nir_src *input_src = &intr->src[0]; |
| 734 |
nir_def *input = input_src->ssa; |
| 735 |
nir_src *offset = &intr->src[1]; |
| 736 |
assert(nir_src_as_uint(*offset) == 0); |
| 737 |
|
| 738 |
/* Skip color write that don't include alpha. */ |
| 739 |
if (input->num_components != 4) |
| 740 |
return NULL; |
| 741 |
|
| 742 |
b->cursor = nir_before_instr(&intr->instr); |
| 743 |
|
| 744 |
/* TODO: define or other way of representing bit 0 of metadata... */ |
| 745 |
nir_def *alpha_to_one_enabled = |
| 746 |
nir_ine_imm(b, |
| 747 |
nir_ubitfield_extract_imm(b, nir_load_fs_meta_pco(b), 0, 1), |
| 748 |
0); |
| 749 |
|
| 750 |
nir_def *alpha = nir_bcsel(b, |
| 751 |
alpha_to_one_enabled, |
| 752 |
nir_imm_float(b, 1.0f), |
| 753 |
nir_channel(b, input, 3)); |
| 754 |
|
| 755 |
nir_src_rewrite(input_src, nir_vector_insert_imm(b, input, alpha, 3)); |
| 756 |
|
| 757 |
return NIR_LOWER_INSTR_PROGRESS; |
| 758 |
} |
| 759 |
|
| 760 |
static bool is_load_sample_mask(const nir_instr *instr, |
| 761 |
UNUSED const void *cb_data) |
| 762 |
{ |
| 763 |
if (instr->type != nir_instr_type_intrinsic) |
| 764 |
return false; |
| 765 |
|
| 766 |
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); |
| 767 |
return intr->intrinsic == nir_intrinsic_load_sample_mask_in; |
| 768 |
} |
| 769 |
|
| 770 |
static nir_def * |
| 771 |
lower_load_sample_mask(nir_builder *b, nir_instr *instr, UNUSED void *cb_data) |
| 772 |
{ |
| 773 |
b->cursor = nir_before_instr(instr); |
| 774 |
|
| 775 |
nir_def *smp_msk = |
| 776 |
nir_ubitfield_extract_imm(b, nir_load_fs_meta_pco(b), 9, 16); |
| 777 |
|
| 778 |
smp_msk = nir_iand(b, smp_msk, nir_load_savmsk_vm_pco(b)); |
| 779 |
|
| 780 |
return smp_msk; |
| 781 |
} |
| 782 |
|
| 783 |
static nir_def * |
| 784 |
lower_color_write_enable(nir_builder *b, nir_instr *instr, UNUSED void *cb_data) |
| 785 |
{ |
| 786 |
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); |
| 787 |
|
| 788 |
nir_src *input_src = &intr->src[0]; |
| 789 |
nir_def *input = input_src->ssa; |
| 790 |
nir_def *offset = intr->src[1].ssa; |
| 791 |
|
| 792 |
struct nir_io_semantics io_semantics = nir_intrinsic_io_semantics(intr); |
| 793 |
unsigned color_write_index = io_semantics.location - FRAG_RESULT_DATA0; |
| 794 |
io_semantics.fb_fetch_output = true; |
| 795 |
|
| 796 |
b->cursor = nir_before_instr(&intr->instr); |
| 797 |
|
| 798 |
/* TODO: nir op that returns bool based on whether a bit is set. */ |
| 799 |
/* TODO: define for 1 */ |
| 800 |
nir_def *color_write_enabled = |
| 801 |
nir_ine_imm(b, |
| 802 |
nir_ubitfield_extract_imm(b, |
| 803 |
nir_load_fs_meta_pco(b), |
| 804 |
1 + color_write_index, |
| 805 |
1), |
| 806 |
0); |
| 807 |
|
| 808 |
nir_def *prev_input = |
| 809 |
nir_load_output(b, |
| 810 |
input->num_components, |
| 811 |
input->bit_size, |
| 812 |
offset, |
| 813 |
.base = nir_intrinsic_base(intr), |
| 814 |
.range = nir_intrinsic_range(intr), |
| 815 |
.component = nir_intrinsic_component(intr), |
| 816 |
.dest_type = nir_intrinsic_src_type(intr), |
| 817 |
.io_semantics = io_semantics); |
| 818 |
|
| 819 |
nir_src_rewrite(input_src, |
| 820 |
nir_bcsel(b, color_write_enabled, input, prev_input)); |
| 821 |
|
| 822 |
return NIR_LOWER_INSTR_PROGRESS; |
| 823 |
} |
| 824 |
|
| 825 |
/** |
| 826 |
* \brief Per-fragment output pass. |
| 827 |
* |
| 828 |
* \param[in,out] shader NIR shader. |
| 829 |
* \param[in,out] fs Fragment shader-specific data. |
| 830 |
* \return True if the pass made progress. |
| 831 |
*/ |
| 832 |
bool pco_nir_pfo(nir_shader *shader, pco_fs_data *fs) |
| 833 |
{ |
| 834 |
assert(shader->info.stage == MESA_SHADER_FRAGMENT); |
| 835 |
|
| 836 |
nir_builder b = nir_builder_create(nir_shader_get_entrypoint(shader)); |
| 837 |
b.cursor = |
| 838 |
nir_before_block(nir_start_block(nir_shader_get_entrypoint(shader))); |
| 839 |
|
| 840 |
struct pfo_state state = { |
| 841 |
.fs = fs, |
| 842 |
.discard_cond_reg = nir_decl_reg(&b, 1, 1, 0), |
| 843 |
}; |
| 844 |
state.last_discard_store = |
| 845 |
nir_build_store_reg(&b, nir_imm_false(&b), state.discard_cond_reg); |
| 846 |
|
| 847 |
state.loads = UTIL_DYNARRAY_INIT; |
| 848 |
state.stores = UTIL_DYNARRAY_INIT; |
| 849 |
|
| 850 |
bool progress = false; |
| 851 |
|
| 852 |
/* TODO: instead of doing multiple passes, probably better to just cache all |
| 853 |
* the stores |
| 854 |
*/ |
| 855 |
if (!shader->info.internal) { |
| 856 |
progress |= nir_shader_lower_instructions(shader, |
| 857 |
is_frag_color_out, |
| 858 |
lower_alpha_to_one, |
| 859 |
&state); |
| 860 |
} |
| 861 |
|
| 862 |
if (fs->meta_present.color_write_enable) |
| 863 |
progress |= nir_shader_lower_instructions(shader, |
| 864 |
is_frag_color_out, |
| 865 |
lower_color_write_enable, |
| 866 |
NULL); |
| 867 |
|
| 868 |
progress |= nir_shader_lower_instructions(shader, is_pfo, lower_pfo, &state); |
| 869 |
progress |= lower_isp_fb(&b, &state); |
| 870 |
|
| 871 |
progress |= sink_outputs(shader, &state); |
| 872 |
progress |= z_replicate(shader, &state); |
| 873 |
|
| 874 |
progress |= nir_shader_lower_instructions(shader, |
| 875 |
is_load_sample_mask, |
| 876 |
lower_load_sample_mask, |
| 877 |
NULL); |
| 878 |
|
| 879 |
util_dynarray_fini(&state.stores); |
| 880 |
util_dynarray_fini(&state.loads); |
| 881 |
|
| 882 |
return progress; |
| 883 |
} |
| 884 |
|
| 885 |
static nir_def *lower_pvi(nir_builder *b, nir_instr *instr, void *cb_data) |
| 886 |
{ |
| 887 |
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); |
| 888 |
struct pvi_state *state = cb_data; |
| 889 |
|
| 890 |
unsigned start_comp = nir_intrinsic_component(intr); |
| 891 |
unsigned num_comps = intr->def.num_components; |
| 892 |
|
| 893 |
ASSERTED nir_src *offset = &intr->src[0]; |
| 894 |
assert(nir_src_as_uint(*offset) == 0); |
| 895 |
|
| 896 |
struct nir_io_semantics io_semantics = nir_intrinsic_io_semantics(intr); |
| 897 |
gl_vert_attrib location = io_semantics.location; |
| 898 |
nir_def *attrib = state->attribs[location - VERT_ATTRIB_GENERIC0]; |
| 899 |
assert(attrib); |
| 900 |
|
| 901 |
b->cursor = nir_before_instr(&intr->instr); |
| 902 |
return nir_channels(b, attrib, BITFIELD_RANGE(start_comp, num_comps)); |
| 903 |
} |
| 904 |
|
| 905 |
static bool is_pvi(const nir_instr *instr, const void *cb_data) |
| 906 |
{ |
| 907 |
if (instr->type != nir_instr_type_intrinsic) |
| 908 |
return false; |
| 909 |
|
| 910 |
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); |
| 911 |
if (intr->intrinsic != nir_intrinsic_load_input) |
| 912 |
return false; |
| 913 |
|
| 914 |
if (is_processed(intr)) |
| 915 |
return false; |
| 916 |
|
| 917 |
ASSERTED gl_vert_attrib location = nir_intrinsic_io_semantics(intr).location; |
| 918 |
assert(location >= VERT_ATTRIB_GENERIC0 && |
| 919 |
location <= VERT_ATTRIB_GENERIC15); |
| 920 |
|
| 921 |
return true; |
| 922 |
} |
| 923 |
|
| 924 |
/** |
| 925 |
* \brief Per-vertex input pass. |
| 926 |
* |
| 927 |
* \param[in,out] shader NIR shader. |
| 928 |
* \param[in,out] vs Vertex shader-specific data. |
| 929 |
* \return True if the pass made progress. |
| 930 |
*/ |
| 931 |
bool pco_nir_pvi(nir_shader *shader, pco_vs_data *vs) |
| 932 |
{ |
| 933 |
assert(shader->info.stage == MESA_SHADER_VERTEX); |
| 934 |
|
| 935 |
struct pvi_state state = { .vs = vs }; |
| 936 |
|
| 937 |
nir_builder b = nir_builder_at( |
| 938 |
nir_before_block(nir_start_block(nir_shader_get_entrypoint(shader)))); |
| 939 |
for (unsigned u = 0; u < ARRAY_SIZE(state.attribs); ++u) { |
| 940 |
gl_vert_attrib location = u + VERT_ATTRIB_GENERIC0; |
| 941 |
enum pipe_format format = vs->attrib_formats[location]; |
| 942 |
if (format == PIPE_FORMAT_NONE) |
| 943 |
continue; |
| 944 |
|
| 945 |
/* Update the type of the stored variable, remove any fractional vars. */ |
| 946 |
nir_variable *var = NULL; |
| 947 |
nir_alu_type base_type = 0; |
| 948 |
nir_foreach_variable_with_modes_safe (iter_var, |
| 949 |
shader, |
| 950 |
nir_var_shader_in) { |
| 951 |
if (iter_var->data.location != location) |
| 952 |
continue; |
| 953 |
|
| 954 |
if (!base_type) |
| 955 |
base_type = nir_get_nir_type_for_glsl_type(iter_var->type); |
| 956 |
#ifndef NDEBUG |
| 957 |
else |
| 958 |
assert(base_type == nir_get_nir_type_for_glsl_type(iter_var->type)); |
| 959 |
#endif /* NDEBUG */ |
| 960 |
|
| 961 |
if (!iter_var->data.location_frac) { |
| 962 |
assert(!var); |
| 963 |
var = iter_var; |
| 964 |
continue; |
| 965 |
} |
| 966 |
|
| 967 |
exec_node_remove(&iter_var->node); |
| 968 |
} |
| 969 |
|
| 970 |
if (!var) { |
| 971 |
if (!base_type) |
| 972 |
continue; |
| 973 |
|
| 974 |
/* An attrib var was found but was fractional so we dropped it. */ |
| 975 |
var = nir_variable_create(shader, nir_var_shader_in, NULL, NULL); |
| 976 |
var->data.location = location; |
| 977 |
} |
| 978 |
|
| 979 |
unsigned format_dwords = |
| 980 |
DIV_ROUND_UP(util_format_get_blocksize(format), sizeof(uint32_t)); |
| 981 |
var->type = glsl_uvec_type(format_dwords); |
| 982 |
|
| 983 |
nir_def *packed_comps[4]; |
| 984 |
for (unsigned c = 0; c < ARRAY_SIZE(packed_comps); ++c) { |
| 985 |
packed_comps[c] = nir_load_input(&b, |
| 986 |
1, |
| 987 |
32, |
| 988 |
nir_imm_int(&b, 0), |
| 989 |
.range = 1, |
| 990 |
.component = c, |
| 991 |
.dest_type = nir_type_invalid | 32, |
| 992 |
.io_semantics = (nir_io_semantics){ |
| 993 |
.location = location, |
| 994 |
.num_slots = 1, |
| 995 |
}); |
| 996 |
} |
| 997 |
|
| 998 |
state.attribs[u] = |
| 999 |
unpack_from_format(&b, packed_comps, base_type, format, 4); |
| 1000 |
} |
| 1001 |
|
| 1002 |
nir_shader_lower_instructions(shader, is_pvi, lower_pvi, &state); |
| 1003 |
|
| 1004 |
return true; |
| 1005 |
} |
| 1006 |
|
| 1007 |
/** |
| 1008 |
* \brief Checks if the point size is written. |
| 1009 |
* |
| 1010 |
* \param[in] b NIR builder. |
| 1011 |
* \param[in] intr NIR intrinsic instruction. |
| 1012 |
* \param[in] cb_data User callback data. |
| 1013 |
* \return True if the instruction was lowered. |
| 1014 |
*/ |
| 1015 |
static bool |
| 1016 |
check_psiz_write(nir_builder *b, nir_intrinsic_instr *intr, void *cb_data) |
| 1017 |
{ |
| 1018 |
if (intr->intrinsic != nir_intrinsic_store_output) |
| 1019 |
return false; |
| 1020 |
|
| 1021 |
bool *writes_psiz = cb_data; |
| 1022 |
|
| 1023 |
struct nir_io_semantics io_semantics = nir_intrinsic_io_semantics(intr); |
| 1024 |
*writes_psiz |= (io_semantics.location == VARYING_SLOT_PSIZ); |
| 1025 |
|
| 1026 |
return false; |
| 1027 |
} |
| 1028 |
|
| 1029 |
/** |
| 1030 |
* \brief Vertex shader point size pass. |
| 1031 |
* |
| 1032 |
* \param[in,out] shader NIR shader. |
| 1033 |
* \return True if the pass made progress. |
| 1034 |
*/ |
| 1035 |
bool pco_nir_point_size(nir_shader *shader) |
| 1036 |
{ |
| 1037 |
assert(shader->info.stage == MESA_SHADER_VERTEX); |
| 1038 |
if (shader->info.internal) |
| 1039 |
return false; |
| 1040 |
|
| 1041 |
bool writes_psiz = false; |
| 1042 |
nir_shader_intrinsics_pass(shader, |
| 1043 |
check_psiz_write, |
| 1044 |
nir_metadata_all, |
| 1045 |
&writes_psiz); |
| 1046 |
|
| 1047 |
/* Nothing to do if the shader already writes the point size. */ |
| 1048 |
if (writes_psiz) |
| 1049 |
return false; |
| 1050 |
|
| 1051 |
/* Create a point size variable if there isn't one. */ |
| 1052 |
nir_get_variable_with_location(shader, |
| 1053 |
nir_var_shader_out, |
| 1054 |
VARYING_SLOT_PSIZ, |
| 1055 |
glsl_float_type()); |
| 1056 |
|
| 1057 |
/* Add a point size write. */ |
| 1058 |
nir_builder b = nir_builder_at( |
| 1059 |
nir_after_block(nir_impl_last_block(nir_shader_get_entrypoint(shader)))); |
| 1060 |
|
| 1061 |
nir_store_output(&b, |
| 1062 |
nir_imm_float(&b, PVR_POINT_SIZE_RANGE_MIN), |
| 1063 |
nir_imm_int(&b, 0), |
| 1064 |
.base = 0, |
| 1065 |
.range = 1, |
| 1066 |
.write_mask = 1, |
| 1067 |
.component = 0, |
| 1068 |
.src_type = nir_type_float32, |
| 1069 |
.io_semantics = (nir_io_semantics){ |
| 1070 |
.location = VARYING_SLOT_PSIZ, |
| 1071 |
.num_slots = 1, |
| 1072 |
}); |
| 1073 |
|
| 1074 |
return true; |
| 1075 |
} |
| 1076 |
|
| 1077 |
static bool is_fs_intr(const nir_instr *instr, UNUSED const void *cb_data) |
| 1078 |
{ |
| 1079 |
if (instr->type != nir_instr_type_intrinsic) |
| 1080 |
return false; |
| 1081 |
|
| 1082 |
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); |
| 1083 |
switch (intr->intrinsic) { |
| 1084 |
case nir_intrinsic_load_front_face: |
| 1085 |
return true; |
| 1086 |
|
| 1087 |
default: |
| 1088 |
break; |
| 1089 |
} |
| 1090 |
|
| 1091 |
return false; |
| 1092 |
} |
| 1093 |
|
| 1094 |
static nir_def *lower_front_face(nir_builder *b, nir_intrinsic_instr *intr) |
| 1095 |
{ |
| 1096 |
nir_def *face_ccw = nir_load_face_ccw_pco(b); |
| 1097 |
nir_def *front_face = nir_ieq_imm(b, face_ccw, 0); |
| 1098 |
|
| 1099 |
nir_def *ff_op = nir_load_front_face_op_pco(b); |
| 1100 |
nir_def *ff_elems[] = { |
| 1101 |
[PCO_FRONT_FACE_OP_NOP] = front_face, |
| 1102 |
[PCO_FRONT_FACE_OP_SWAP] = nir_inot(b, front_face), |
| 1103 |
[PCO_FRONT_FACE_OP_TRUE] = nir_imm_true(b), |
| 1104 |
}; |
| 1105 |
|
| 1106 |
return nir_select_from_ssa_def_array(b, |
| 1107 |
ff_elems, |
| 1108 |
ARRAY_SIZE(ff_elems), |
| 1109 |
ff_op); |
| 1110 |
} |
| 1111 |
|
| 1112 |
static nir_def * |
| 1113 |
lower_fs_intr(nir_builder *b, nir_instr *instr, UNUSED void *cb_data) |
| 1114 |
{ |
| 1115 |
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); |
| 1116 |
switch (intr->intrinsic) { |
| 1117 |
case nir_intrinsic_load_front_face: |
| 1118 |
return lower_front_face(b, intr); |
| 1119 |
|
| 1120 |
default: |
| 1121 |
break; |
| 1122 |
} |
| 1123 |
|
| 1124 |
UNREACHABLE(""); |
| 1125 |
} |
| 1126 |
|
| 1127 |
bool pco_nir_lower_fs_intrinsics(nir_shader *shader) |
| 1128 |
{ |
| 1129 |
assert(shader->info.stage == MESA_SHADER_FRAGMENT); |
| 1130 |
|
| 1131 |
return nir_shader_lower_instructions(shader, is_fs_intr, lower_fs_intr, NULL); |
| 1132 |
} |
| 1133 |
|
| 1134 |
static bool |
| 1135 |
lower_vs_intr(nir_builder *b, nir_intrinsic_instr *intr, UNUSED void *cb_data) |
| 1136 |
{ |
| 1137 |
b->cursor = nir_before_instr(&intr->instr); |
| 1138 |
|
| 1139 |
switch (intr->intrinsic) { |
| 1140 |
/* First vs base vertex is handled in the PDS, so they're equivalent. */ |
| 1141 |
case nir_intrinsic_load_first_vertex: |
| 1142 |
nir_def_replace(&intr->def, nir_load_base_vertex(b)); |
| 1143 |
nir_instr_free(&intr->instr); |
| 1144 |
return true; |
| 1145 |
|
| 1146 |
default: |
| 1147 |
break; |
| 1148 |
} |
| 1149 |
|
| 1150 |
return false; |
| 1151 |
} |
| 1152 |
|
| 1153 |
bool pco_nir_lower_vs_intrinsics(nir_shader *shader) |
| 1154 |
{ |
| 1155 |
assert(shader->info.stage == MESA_SHADER_VERTEX); |
| 1156 |
|
| 1157 |
return nir_shader_intrinsics_pass(shader, |
| 1158 |
lower_vs_intr, |
| 1159 |
nir_metadata_control_flow, |
| 1160 |
NULL); |
| 1161 |
} |
| 1162 |
|
| 1163 |
bool pco_nir_lower_clip_cull_vars(nir_shader *shader) |
| 1164 |
{ |
| 1165 |
if (shader->info.internal) |
| 1166 |
return false; |
| 1167 |
|
| 1168 |
unsigned clip_cull_comps = shader->info.clip_distance_array_size + |
| 1169 |
shader->info.cull_distance_array_size; |
| 1170 |
if (!clip_cull_comps) |
| 1171 |
return false; |
| 1172 |
|
| 1173 |
/* Remove the old variables. */ |
| 1174 |
const gl_varying_slot clip_cull_locations[] = { |
| 1175 |
VARYING_SLOT_CLIP_DIST0, |
| 1176 |
VARYING_SLOT_CLIP_DIST1, |
| 1177 |
}; |
| 1178 |
|
| 1179 |
nir_variable *var; |
| 1180 |
for (unsigned u = 0; u < ARRAY_SIZE(clip_cull_locations); ++u) { |
| 1181 |
gl_varying_slot location = clip_cull_locations[u]; |
| 1182 |
while ((var = nir_find_variable_with_location(shader, |
| 1183 |
nir_var_shader_out, |
| 1184 |
location))) { |
| 1185 |
exec_node_remove(&var->node); |
| 1186 |
} |
| 1187 |
} |
| 1188 |
|
| 1189 |
/* Create new variables. */ |
| 1190 |
nir_create_variable_with_location(shader, |
| 1191 |
nir_var_shader_out, |
| 1192 |
VARYING_SLOT_CLIP_DIST0, |
| 1193 |
glsl_vec_type(MIN2(clip_cull_comps, 4))); |
| 1194 |
|
| 1195 |
if (clip_cull_comps > 4) { |
| 1196 |
nir_create_variable_with_location(shader, |
| 1197 |
nir_var_shader_out, |
| 1198 |
VARYING_SLOT_CLIP_DIST1, |
| 1199 |
glsl_vec_type(clip_cull_comps - 4)); |
| 1200 |
} |
| 1201 |
|
| 1202 |
nir_metadata_invalidate(shader); |
| 1203 |
|
| 1204 |
return true; |
| 1205 |
} |
| 1206 |
|
| 1207 |
static bool |
| 1208 |
clone_clip_cull_stores(nir_builder *b, nir_intrinsic_instr *intr, void *data) |
| 1209 |
{ |
| 1210 |
if (intr->intrinsic != nir_intrinsic_store_deref) |
| 1211 |
return false; |
| 1212 |
|
| 1213 |
nir_deref_instr *deref = nir_src_as_deref(intr->src[0]); |
| 1214 |
if (deref->deref_type != nir_deref_type_array) |
| 1215 |
return false; |
| 1216 |
|
| 1217 |
nir_variable *var = nir_deref_instr_get_variable(deref); |
| 1218 |
if (var->data.location != VARYING_SLOT_CLIP_DIST0 && |
| 1219 |
var->data.location != VARYING_SLOT_CLIP_DIST1) |
| 1220 |
return false; |
| 1221 |
|
| 1222 |
b->cursor = nir_after_instr(&intr->instr); |
| 1223 |
|
| 1224 |
unsigned var_index = var->data.location - VARYING_SLOT_CLIP_DIST0; |
| 1225 |
nir_def *index = |
| 1226 |
nir_iadd_imm(b, deref->arr.index.ssa, var->data.location_frac); |
| 1227 |
index = nir_iadd_imm(b, index, var_index * 4); |
| 1228 |
|
| 1229 |
nir_variable *clone_var = data; |
| 1230 |
nir_store_array_var(b, clone_var, index, intr->src[1].ssa, 1); |
| 1231 |
|
| 1232 |
return true; |
| 1233 |
} |
| 1234 |
|
| 1235 |
static bool is_clip_cull_load(const nir_instr *instr, |
| 1236 |
UNUSED const void *cb_data) |
| 1237 |
{ |
| 1238 |
if (instr->type != nir_instr_type_intrinsic) |
| 1239 |
return false; |
| 1240 |
|
| 1241 |
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); |
| 1242 |
if (intr->intrinsic != nir_intrinsic_load_deref) |
| 1243 |
return false; |
| 1244 |
|
| 1245 |
nir_deref_instr *deref = nir_src_as_deref(intr->src[0]); |
| 1246 |
if (deref->deref_type != nir_deref_type_array) |
| 1247 |
return false; |
| 1248 |
|
| 1249 |
nir_variable *var = nir_deref_instr_get_variable(deref); |
| 1250 |
|
| 1251 |
return var->data.location == VARYING_SLOT_CLIP_DIST0 || |
| 1252 |
var->data.location == VARYING_SLOT_CLIP_DIST1; |
| 1253 |
} |
| 1254 |
|
| 1255 |
static nir_def * |
| 1256 |
swap_clip_cull_load(nir_builder *b, nir_instr *instr, void *cb_data) |
| 1257 |
{ |
| 1258 |
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); |
| 1259 |
nir_deref_instr *deref = nir_src_as_deref(intr->src[0]); |
| 1260 |
nir_variable *var = nir_deref_instr_get_variable(deref); |
| 1261 |
|
| 1262 |
unsigned var_index = var->data.location - VARYING_SLOT_CLIP_DIST0; |
| 1263 |
nir_def *index = |
| 1264 |
nir_iadd_imm(b, deref->arr.index.ssa, var->data.location_frac); |
| 1265 |
index = nir_iadd_imm(b, index, var_index * 4); |
| 1266 |
|
| 1267 |
nir_variable *clone_var = cb_data; |
| 1268 |
return nir_load_array_var(b, clone_var, index); |
| 1269 |
} |
| 1270 |
|
| 1271 |
bool pco_nir_link_clip_cull_vars(nir_shader *producer, nir_shader *consumer) |
| 1272 |
{ |
| 1273 |
if (producer->info.stage != MESA_SHADER_VERTEX || |
| 1274 |
consumer->info.stage != MESA_SHADER_FRAGMENT) { |
| 1275 |
return false; |
| 1276 |
} |
| 1277 |
|
| 1278 |
unsigned clip_cull_comps = consumer->info.clip_distance_array_size + |
| 1279 |
consumer->info.cull_distance_array_size; |
| 1280 |
/* Skip if clip/cull comps aren't actually consumed. */ |
| 1281 |
if (!clip_cull_comps) |
| 1282 |
return false; |
| 1283 |
|
| 1284 |
const glsl_type *clone_var_type = |
| 1285 |
glsl_array_type(glsl_float_type(), clip_cull_comps, 0); |
| 1286 |
|
| 1287 |
/* Find unused varying slot to use and create the variables. */ |
| 1288 |
gl_varying_slot clone_slot = VARYING_SLOT_VAR0; |
| 1289 |
nir_foreach_shader_out_variable (var, producer) { |
| 1290 |
clone_slot = MAX2(clone_slot, var->data.location + 1); |
| 1291 |
} |
| 1292 |
assert(clone_slot < VARYING_SLOT_MAX); |
| 1293 |
|
| 1294 |
nir_variable *clone_var = |
| 1295 |
nir_variable_create(producer, nir_var_shader_out, clone_var_type, NULL); |
| 1296 |
clone_var->data.location = clone_slot; |
| 1297 |
|
| 1298 |
nir_shader_intrinsics_pass(producer, |
| 1299 |
clone_clip_cull_stores, |
| 1300 |
nir_metadata_block_index | nir_metadata_dominance, |
| 1301 |
clone_var); |
| 1302 |
|
| 1303 |
clone_var = |
| 1304 |
nir_variable_create(consumer, nir_var_shader_in, clone_var_type, NULL); |
| 1305 |
clone_var->data.location = clone_slot; |
| 1306 |
|
| 1307 |
nir_shader_lower_instructions(consumer, |
| 1308 |
is_clip_cull_load, |
| 1309 |
swap_clip_cull_load, |
| 1310 |
clone_var); |
| 1311 |
|
| 1312 |
return true; |
| 1313 |
} |
| 1314 |
|
| 1315 |
static nir_def *alu_iter(nir_builder *b, |
| 1316 |
nir_def *coords, |
| 1317 |
unsigned component, |
| 1318 |
struct nir_io_semantics io_semantics) |
| 1319 |
{ |
| 1320 |
nir_def *coeffs = nir_load_fs_coeffs_pco(b, |
| 1321 |
.component = component, |
| 1322 |
.io_semantics = io_semantics); |
| 1323 |
|
| 1324 |
nir_def *result = nir_ffma(b, |
| 1325 |
nir_channel(b, coeffs, 1), |
| 1326 |
nir_channel(b, coords, 1), |
| 1327 |
nir_channel(b, coeffs, 2)); |
| 1328 |
result = |
| 1329 |
nir_ffma(b, nir_channel(b, coeffs, 0), nir_channel(b, coords, 0), result); |
| 1330 |
|
| 1331 |
return result; |
| 1332 |
} |
| 1333 |
|
| 1334 |
static bool |
| 1335 |
lower_sample_pos(nir_builder *b, nir_intrinsic_instr *intr, pco_fs_data *fs) |
| 1336 |
{ |
| 1337 |
b->cursor = nir_before_instr(&intr->instr); |
| 1338 |
|
| 1339 |
nir_def *msaa_samples = nir_bit_count( |
| 1340 |
b, |
| 1341 |
nir_u2u32(b, nir_alpha_to_coverage(b, nir_imm_float(b, 1.0f)))); |
| 1342 |
|
| 1343 |
nir_def *sample_id = intr->intrinsic == nir_intrinsic_load_sample_pos |
| 1344 |
? nir_load_sample_id(b) |
| 1345 |
: intr->src[0].ssa; |
| 1346 |
|
| 1347 |
nir_def *dword_index = |
| 1348 |
nir_ishr_imm(b, nir_iadd(b, msaa_samples, sample_id), 2); |
| 1349 |
|
| 1350 |
nir_def *packed_sample_location = |
| 1351 |
nir_load_packed_sample_location_pco(b, dword_index); |
| 1352 |
fs->uses.sample_locations = true; |
| 1353 |
|
| 1354 |
nir_def *byte_index = nir_iand_imm(b, sample_id, 0b11); |
| 1355 |
|
| 1356 |
packed_sample_location = |
| 1357 |
nir_extract_u8(b, packed_sample_location, byte_index); |
| 1358 |
|
| 1359 |
nir_def *sample_location = |
| 1360 |
nir_vec2(b, |
| 1361 |
nir_ubitfield_extract_imm(b, packed_sample_location, 0, 4), |
| 1362 |
nir_ubitfield_extract_imm(b, packed_sample_location, 4, 4)); |
| 1363 |
|
| 1364 |
sample_location = nir_u2f32(b, sample_location); |
| 1365 |
sample_location = nir_fdiv_imm(b, sample_location, 16.0f); |
| 1366 |
sample_location = nir_bcsel(b, |
| 1367 |
nir_ieq_imm(b, msaa_samples, 1), |
| 1368 |
nir_imm_vec2(b, 0.5f, 0.5f), |
| 1369 |
sample_location); |
| 1370 |
|
| 1371 |
nir_def_replace(&intr->def, sample_location); |
| 1372 |
nir_instr_free(&intr->instr); |
| 1373 |
|
| 1374 |
return true; |
| 1375 |
} |
| 1376 |
|
| 1377 |
static bool |
| 1378 |
lower_interp(nir_builder *b, nir_intrinsic_instr *intr, void *cb_data) |
| 1379 |
{ |
| 1380 |
pco_fs_data *fs = cb_data; |
| 1381 |
b->cursor = nir_before_instr(&intr->instr); |
| 1382 |
|
| 1383 |
switch (intr->intrinsic) { |
| 1384 |
case nir_intrinsic_load_sample_pos: |
| 1385 |
case nir_intrinsic_load_sample_pos_from_id: |
| 1386 |
return lower_sample_pos(b, intr, fs); |
| 1387 |
|
| 1388 |
case nir_intrinsic_load_interpolated_input: |
| 1389 |
break; |
| 1390 |
|
| 1391 |
default: |
| 1392 |
return false; |
| 1393 |
} |
| 1394 |
|
| 1395 |
nir_intrinsic_instr *bary = nir_src_as_intrinsic(intr->src[0]); |
| 1396 |
assert(bary); |
| 1397 |
|
| 1398 |
/* Skip cases that don't need handling. */ |
| 1399 |
if (bary->intrinsic != nir_intrinsic_load_barycentric_at_offset) |
| 1400 |
return false; |
| 1401 |
|
| 1402 |
assert(nir_src_as_uint(intr->src[1]) == 0); |
| 1403 |
|
| 1404 |
nir_def *coords = nir_load_tile_coord_pco(b, 2); |
| 1405 |
coords = nir_fadd(b, coords, bary->src[0].ssa); |
| 1406 |
|
| 1407 |
enum glsl_interp_mode interp_mode = nir_intrinsic_interp_mode(bary); |
| 1408 |
nir_def *rhw = alu_iter(b, |
| 1409 |
coords, |
| 1410 |
3, |
| 1411 |
(struct nir_io_semantics){ |
| 1412 |
.location = VARYING_SLOT_POS, |
| 1413 |
.num_slots = 1, |
| 1414 |
}); |
| 1415 |
|
| 1416 |
nir_def *comps[4]; |
| 1417 |
for (unsigned u = 0; u < intr->def.num_components; ++u) { |
| 1418 |
comps[u] = alu_iter(b, coords, u, nir_intrinsic_io_semantics(intr)); |
| 1419 |
if (interp_mode != INTERP_MODE_NOPERSPECTIVE) |
| 1420 |
comps[u] = nir_fdiv(b, comps[u], rhw); |
| 1421 |
} |
| 1422 |
|
| 1423 |
nir_def *repl = nir_vec(b, comps, intr->def.num_components); |
| 1424 |
nir_def_replace(&intr->def, repl); |
| 1425 |
nir_instr_free(&intr->instr); |
| 1426 |
|
| 1427 |
return true; |
| 1428 |
} |
| 1429 |
|
| 1430 |
bool pco_nir_lower_interpolation(nir_shader *shader, pco_fs_data *fs) |
| 1431 |
{ |
| 1432 |
bool progress = false; |
| 1433 |
|
| 1434 |
progress |= nir_opt_barycentric(shader, true); |
| 1435 |
progress |= nir_shader_intrinsics_pass(shader, |
| 1436 |
lower_interp, |
| 1437 |
nir_metadata_control_flow, |
| 1438 |
fs); |
| 1439 |
|
| 1440 |
return progress; |
| 1441 |
} |
| 1442 |
|
| 1443 |
static bool lower_load_view_index_fs(struct nir_builder *b, |
| 1444 |
nir_intrinsic_instr *intr, |
| 1445 |
void *cb_data) |
| 1446 |
{ |
| 1447 |
if (intr->intrinsic != nir_intrinsic_load_view_index) |
| 1448 |
return false; |
| 1449 |
|
| 1450 |
nir_variable *view_index_var = cb_data; |
| 1451 |
b->cursor = nir_before_instr(&intr->instr); |
| 1452 |
nir_def_replace(&intr->def, nir_load_var(b, view_index_var)); |
| 1453 |
nir_instr_free(&intr->instr); |
| 1454 |
|
| 1455 |
return true; |
| 1456 |
} |
| 1457 |
|
| 1458 |
bool pco_nir_link_multiview(nir_shader *producer, |
| 1459 |
nir_shader *consumer, |
| 1460 |
pco_data *consumer_data) |
| 1461 |
{ |
| 1462 |
if (producer->info.stage != MESA_SHADER_VERTEX || |
| 1463 |
consumer->info.stage != MESA_SHADER_FRAGMENT || |
| 1464 |
!consumer_data->common.multiview) { |
| 1465 |
return false; |
| 1466 |
} |
| 1467 |
|
| 1468 |
/* Find unused varying slot for the view index. */ |
| 1469 |
gl_varying_slot view_index_slot = VARYING_SLOT_VAR0; |
| 1470 |
nir_foreach_shader_out_variable (var, producer) { |
| 1471 |
view_index_slot = MAX2(view_index_slot, var->data.location + 1); |
| 1472 |
} |
| 1473 |
assert(view_index_slot < VARYING_SLOT_MAX); |
| 1474 |
consumer_data->fs.view_index_slot = view_index_slot; |
| 1475 |
|
| 1476 |
/* Create output variable in the producer. */ |
| 1477 |
nir_variable *view_index_var = nir_variable_create(producer, |
| 1478 |
nir_var_shader_out, |
| 1479 |
glsl_uint_type(), |
| 1480 |
"view_index"); |
| 1481 |
view_index_var->data.location = view_index_slot; |
| 1482 |
view_index_var->data.interpolation = INTERP_MODE_FLAT; |
| 1483 |
view_index_var->data.always_active_io = true; |
| 1484 |
|
| 1485 |
/* Store view index in the producer. */ |
| 1486 |
nir_builder b = nir_builder_at(nir_after_block( |
| 1487 |
nir_impl_last_block(nir_shader_get_entrypoint(producer)))); |
| 1488 |
nir_store_var(&b, view_index_var, nir_load_view_index(&b), 1); |
| 1489 |
|
| 1490 |
/* Create input variable in the consumer. */ |
| 1491 |
view_index_var = nir_variable_create(consumer, |
| 1492 |
nir_var_shader_in, |
| 1493 |
glsl_uint_type(), |
| 1494 |
"view_index"); |
| 1495 |
view_index_var->data.location = view_index_slot; |
| 1496 |
view_index_var->data.interpolation = INTERP_MODE_FLAT; |
| 1497 |
view_index_var->data.always_active_io = true; |
| 1498 |
|
| 1499 |
/* Lower view index loads in the consumer. */ |
| 1500 |
nir_shader_intrinsics_pass(consumer, |
| 1501 |
lower_load_view_index_fs, |
| 1502 |
nir_metadata_all, |
| 1503 |
view_index_var); |
| 1504 |
|
| 1505 |
return true; |
| 1506 |
} |