| 1 |
/* |
| 2 |
* Copyright 2017 Red Hat Inc. |
| 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 shall be included in |
| 12 |
* all copies or substantial portions of the Software. |
| 13 |
* |
| 14 |
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 15 |
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 16 |
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
| 17 |
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR |
| 18 |
* OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, |
| 19 |
* ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR |
| 20 |
* OTHER DEALINGS IN THE SOFTWARE. |
| 21 |
* |
| 22 |
* Authors: Karol Herbst <[email protected]> |
| 23 |
*/ |
| 24 |
|
| 25 |
#include "compiler/nir/nir.h" |
| 26 |
#include "compiler/nir/nir_builder.h" |
| 27 |
|
| 28 |
#include "util/u_debug.h" |
| 29 |
#include "util/u_prim.h" |
| 30 |
|
| 31 |
#include "nv50_ir.h" |
| 32 |
#include "nv50_ir_lowering_helper.h" |
| 33 |
#include "nv50_ir_target.h" |
| 34 |
#include "nv50_ir_util.h" |
| 35 |
#include "tgsi/tgsi_from_mesa.h" |
| 36 |
|
| 37 |
#include <unordered_map> |
| 38 |
#include <cstring> |
| 39 |
#include <list> |
| 40 |
#include <vector> |
| 41 |
|
| 42 |
namespace { |
| 43 |
|
| 44 |
using namespace nv50_ir; |
| 45 |
|
| 46 |
int |
| 47 |
type_size(const struct glsl_type *type, bool bindless) |
| 48 |
{ |
| 49 |
return glsl_count_attribute_slots(type, false); |
| 50 |
} |
| 51 |
|
| 52 |
bool |
| 53 |
nv50_nir_lower_load_user_clip_plane_cb(nir_builder *b, nir_intrinsic_instr *intrin, void *params) |
| 54 |
{ |
| 55 |
struct nv50_ir_prog_info *info = (struct nv50_ir_prog_info *)params; |
| 56 |
|
| 57 |
if (intrin->intrinsic != nir_intrinsic_load_user_clip_plane) |
| 58 |
return false; |
| 59 |
|
| 60 |
uint16_t offset = info->io.ucpBase + nir_intrinsic_ucp_id(intrin) * 16; |
| 61 |
|
| 62 |
b->cursor = nir_before_instr(&intrin->instr); |
| 63 |
nir_def *replacement = |
| 64 |
nir_load_ubo(b, 4, 32, nir_imm_int(b, info->io.auxCBSlot), |
| 65 |
nir_imm_int(b, offset), .range = ~0u); |
| 66 |
|
| 67 |
nir_def_replace(&intrin->def, replacement); |
| 68 |
|
| 69 |
return true; |
| 70 |
} |
| 71 |
|
| 72 |
bool |
| 73 |
nv50_nir_lower_load_user_clip_plane(nir_shader *nir, struct nv50_ir_prog_info *info) { |
| 74 |
return nir_shader_intrinsics_pass(nir, nv50_nir_lower_load_user_clip_plane_cb, |
| 75 |
nir_metadata_control_flow, |
| 76 |
info); |
| 77 |
} |
| 78 |
|
| 79 |
class Converter : public BuildUtil |
| 80 |
{ |
| 81 |
public: |
| 82 |
Converter(Program *, nir_shader *, nv50_ir_prog_info *, nv50_ir_prog_info_out *); |
| 83 |
|
| 84 |
bool run(); |
| 85 |
private: |
| 86 |
typedef std::vector<LValue*> LValues; |
| 87 |
typedef std::unordered_map<unsigned, LValues> NirDefMap; |
| 88 |
typedef std::unordered_map<unsigned, nir_load_const_instr*> ImmediateMap; |
| 89 |
typedef std::unordered_map<unsigned, BasicBlock*> NirBlockMap; |
| 90 |
|
| 91 |
CacheMode convert(enum gl_access_qualifier); |
| 92 |
TexTarget convert(glsl_sampler_dim, bool isArray, bool isShadow); |
| 93 |
BasicBlock* convert(nir_block *); |
| 94 |
SVSemantic convert(nir_intrinsic_op); |
| 95 |
Value* convert(nir_load_const_instr*, uint8_t); |
| 96 |
LValues& convert(nir_def *); |
| 97 |
|
| 98 |
Value* getSrc(nir_alu_src *, uint8_t component = 0); |
| 99 |
Value* getSrc(nir_src *, uint8_t, bool indirect = false); |
| 100 |
Value* getSrc(nir_def *, uint8_t); |
| 101 |
|
| 102 |
// returned value is the constant part of the given source (either the |
| 103 |
// nir_src or the selected source component of an intrinsic). Even though |
| 104 |
// this is mostly an optimization to be able to skip indirects in a few |
| 105 |
// cases, sometimes we require immediate values or set some fileds on |
| 106 |
// instructions (e.g. tex) in order for codegen to consume those. |
| 107 |
// If the found value has not a constant part, the Value gets returned |
| 108 |
// through the Value parameter. |
| 109 |
uint32_t getIndirect(nir_src *, uint8_t, Value *&); |
| 110 |
// isScalar indicates that the addressing is scalar, vec4 addressing is |
| 111 |
// assumed otherwise |
| 112 |
uint32_t getIndirect(nir_intrinsic_instr *, uint8_t s, uint8_t c, Value *&, |
| 113 |
bool isScalar = false); |
| 114 |
|
| 115 |
uint32_t getSlotAddress(nir_intrinsic_instr *, uint8_t idx, uint8_t slot); |
| 116 |
|
| 117 |
uint8_t translateInterpMode(const struct nv50_ir_varying *var, operation& op); |
| 118 |
void setInterpolate(nv50_ir_varying *, |
| 119 |
uint8_t, |
| 120 |
bool centroid, |
| 121 |
unsigned semantics); |
| 122 |
|
| 123 |
Instruction *loadFrom(DataFile, uint8_t, DataType, Value *def, uint32_t base, |
| 124 |
uint8_t c, Value *indirect0 = NULL, |
| 125 |
Value *indirect1 = NULL, bool patch = false, |
| 126 |
CacheMode cache=CACHE_CA); |
| 127 |
Instruction *loadVector(nir_intrinsic_instr *insn, |
| 128 |
uint8_t buffer, Value *indirectBuffer, |
| 129 |
uint32_t offset, Value *indirectOffset); |
| 130 |
void storeTo(nir_intrinsic_instr *, DataFile, operation, DataType, |
| 131 |
Value *src, uint8_t idx, uint8_t c, Value *indirect0 = NULL, |
| 132 |
Value *indirect1 = NULL); |
| 133 |
Instruction *storeVector(nir_intrinsic_instr *insn, |
| 134 |
uint8_t buffer, Value *indirectBuffer, |
| 135 |
uint32_t offset, Value *indirectOffset); |
| 136 |
|
| 137 |
static bool memVectorizeCb(unsigned align_mul, |
| 138 |
unsigned align_offset, |
| 139 |
unsigned bit_size, |
| 140 |
unsigned num_components, |
| 141 |
int64_t hole_size, |
| 142 |
nir_intrinsic_instr *low, |
| 143 |
nir_intrinsic_instr *high, |
| 144 |
void *cb_data); |
| 145 |
static nir_mem_access_size_align |
| 146 |
getMemAccessSizeAlign(nir_intrinsic_op intrin, |
| 147 |
uint8_t bytes, |
| 148 |
uint8_t bit_size, |
| 149 |
uint32_t align_mul, |
| 150 |
uint32_t align_offset, |
| 151 |
bool offset_is_const, |
| 152 |
enum gl_access_qualifier access, |
| 153 |
const void *cb_data); |
| 154 |
|
| 155 |
bool isFloatType(nir_alu_type); |
| 156 |
bool isSignedType(nir_alu_type); |
| 157 |
bool isResultFloat(nir_op); |
| 158 |
bool isResultSigned(nir_op); |
| 159 |
|
| 160 |
DataType getDType(nir_alu_instr *); |
| 161 |
DataType getDType(nir_intrinsic_instr *); |
| 162 |
DataType getDType(nir_op, uint8_t); |
| 163 |
|
| 164 |
DataFile getFile(nir_intrinsic_op) const; |
| 165 |
|
| 166 |
std::vector<DataType> getSTypes(nir_alu_instr *); |
| 167 |
DataType getSType(nir_src &, bool isFloat, bool isSigned); |
| 168 |
|
| 169 |
operation getOperation(nir_intrinsic_op); |
| 170 |
operation getOperation(nir_op); |
| 171 |
operation getOperation(nir_texop); |
| 172 |
operation preOperationNeeded(nir_op); |
| 173 |
|
| 174 |
int getSubOp(nir_intrinsic_op); |
| 175 |
int getSubOp(nir_op); |
| 176 |
int getAtomicSubOp(nir_atomic_op); |
| 177 |
|
| 178 |
CondCode getCondCode(nir_op); |
| 179 |
|
| 180 |
bool assignSlots(); |
| 181 |
bool parseNIR(); |
| 182 |
|
| 183 |
bool visit(nir_alu_instr *); |
| 184 |
bool visit(nir_block *); |
| 185 |
bool visit(nir_cf_node *); |
| 186 |
bool visit(nir_function *); |
| 187 |
bool visit(nir_if *); |
| 188 |
bool visit(nir_instr *); |
| 189 |
bool visit(nir_intrinsic_instr *); |
| 190 |
bool visit(nir_jump_instr *); |
| 191 |
bool visit(nir_load_const_instr*); |
| 192 |
bool visit(nir_loop *); |
| 193 |
bool visit(nir_undef_instr *); |
| 194 |
bool visit(nir_tex_instr *); |
| 195 |
|
| 196 |
static unsigned lowerBitSizeCB(const nir_instr *, void *); |
| 197 |
|
| 198 |
// tex stuff |
| 199 |
unsigned int getNIRArgCount(TexInstruction::Target&); |
| 200 |
|
| 201 |
void runOptLoop(); |
| 202 |
|
| 203 |
struct nv50_ir_prog_info *info; |
| 204 |
struct nv50_ir_prog_info_out *info_out; |
| 205 |
|
| 206 |
nir_shader *nir; |
| 207 |
|
| 208 |
NirDefMap ssaDefs; |
| 209 |
NirDefMap regDefs; |
| 210 |
ImmediateMap immediates; |
| 211 |
NirBlockMap blocks; |
| 212 |
unsigned int curLoopDepth; |
| 213 |
unsigned int curIfDepth; |
| 214 |
|
| 215 |
BasicBlock *exit; |
| 216 |
Value *zero; |
| 217 |
Instruction *immInsertPos; |
| 218 |
|
| 219 |
Value *outBase; // base address of vertex out patch (for TCP) |
| 220 |
|
| 221 |
union { |
| 222 |
struct { |
| 223 |
Value *position; |
| 224 |
} fp; |
| 225 |
}; |
| 226 |
}; |
| 227 |
|
| 228 |
Converter::Converter(Program *prog, nir_shader *nir, nv50_ir_prog_info *info, |
| 229 |
nv50_ir_prog_info_out *info_out) |
| 230 |
: BuildUtil(prog), |
| 231 |
info(info), |
| 232 |
info_out(info_out), |
| 233 |
nir(nir), |
| 234 |
curLoopDepth(0), |
| 235 |
curIfDepth(0), |
| 236 |
exit(NULL), |
| 237 |
immInsertPos(NULL), |
| 238 |
outBase(nullptr) |
| 239 |
{ |
| 240 |
zero = mkImm((uint32_t)0); |
| 241 |
} |
| 242 |
|
| 243 |
BasicBlock * |
| 244 |
Converter::convert(nir_block *block) |
| 245 |
{ |
| 246 |
NirBlockMap::iterator it = blocks.find(block->index); |
| 247 |
if (it != blocks.end()) |
| 248 |
return it->second; |
| 249 |
|
| 250 |
BasicBlock *bb = new BasicBlock(func); |
| 251 |
blocks[block->index] = bb; |
| 252 |
return bb; |
| 253 |
} |
| 254 |
|
| 255 |
bool |
| 256 |
Converter::isFloatType(nir_alu_type type) |
| 257 |
{ |
| 258 |
return nir_alu_type_get_base_type(type) == nir_type_float; |
| 259 |
} |
| 260 |
|
| 261 |
bool |
| 262 |
Converter::isSignedType(nir_alu_type type) |
| 263 |
{ |
| 264 |
return nir_alu_type_get_base_type(type) == nir_type_int; |
| 265 |
} |
| 266 |
|
| 267 |
bool |
| 268 |
Converter::isResultFloat(nir_op op) |
| 269 |
{ |
| 270 |
const nir_op_info &info = nir_op_infos[op]; |
| 271 |
if (info.output_type != nir_type_invalid) |
| 272 |
return isFloatType(info.output_type); |
| 273 |
|
| 274 |
ERROR("isResultFloat not implemented for %s\n", nir_op_infos[op].name); |
| 275 |
assert(false); |
| 276 |
return true; |
| 277 |
} |
| 278 |
|
| 279 |
bool |
| 280 |
Converter::isResultSigned(nir_op op) |
| 281 |
{ |
| 282 |
switch (op) { |
| 283 |
// there is no umul and we get wrong results if we treat all muls as signed |
| 284 |
case nir_op_imul: |
| 285 |
case nir_op_inot: |
| 286 |
return false; |
| 287 |
default: |
| 288 |
const nir_op_info &info = nir_op_infos[op]; |
| 289 |
if (info.output_type != nir_type_invalid) |
| 290 |
return isSignedType(info.output_type); |
| 291 |
ERROR("isResultSigned not implemented for %s\n", nir_op_infos[op].name); |
| 292 |
assert(false); |
| 293 |
return true; |
| 294 |
} |
| 295 |
} |
| 296 |
|
| 297 |
DataType |
| 298 |
Converter::getDType(nir_alu_instr *insn) |
| 299 |
{ |
| 300 |
return getDType(insn->op, insn->def.bit_size); |
| 301 |
} |
| 302 |
|
| 303 |
DataType |
| 304 |
Converter::getDType(nir_intrinsic_instr *insn) |
| 305 |
{ |
| 306 |
bool isFloat, isSigned; |
| 307 |
switch (insn->intrinsic) { |
| 308 |
case nir_intrinsic_bindless_image_atomic: |
| 309 |
case nir_intrinsic_global_atomic: |
| 310 |
case nir_intrinsic_image_atomic: |
| 311 |
case nir_intrinsic_shared_atomic: |
| 312 |
case nir_intrinsic_ssbo_atomic: { |
| 313 |
nir_alu_type type = nir_atomic_op_type(nir_intrinsic_atomic_op(insn)); |
| 314 |
isFloat = type == nir_type_float; |
| 315 |
isSigned = type == nir_type_int; |
| 316 |
break; |
| 317 |
} |
| 318 |
default: |
| 319 |
isFloat = false; |
| 320 |
isSigned = false; |
| 321 |
break; |
| 322 |
} |
| 323 |
|
| 324 |
return typeOfSize(insn->def.bit_size / 8, isFloat, isSigned); |
| 325 |
} |
| 326 |
|
| 327 |
DataType |
| 328 |
Converter::getDType(nir_op op, uint8_t bitSize) |
| 329 |
{ |
| 330 |
DataType ty = typeOfSize(bitSize / 8, isResultFloat(op), isResultSigned(op)); |
| 331 |
if (ty == TYPE_NONE) { |
| 332 |
ERROR("couldn't get Type for op %s with bitSize %u\n", nir_op_infos[op].name, bitSize); |
| 333 |
assert(false); |
| 334 |
} |
| 335 |
return ty; |
| 336 |
} |
| 337 |
|
| 338 |
std::vector<DataType> |
| 339 |
Converter::getSTypes(nir_alu_instr *insn) |
| 340 |
{ |
| 341 |
const nir_op_info &info = nir_op_infos[insn->op]; |
| 342 |
std::vector<DataType> res(info.num_inputs); |
| 343 |
|
| 344 |
for (uint8_t i = 0; i < info.num_inputs; ++i) { |
| 345 |
if (info.input_types[i] != nir_type_invalid) { |
| 346 |
res[i] = getSType(insn->src[i].src, isFloatType(info.input_types[i]), isSignedType(info.input_types[i])); |
| 347 |
} else { |
| 348 |
ERROR("getSType not implemented for %s idx %u\n", info.name, i); |
| 349 |
assert(false); |
| 350 |
res[i] = TYPE_NONE; |
| 351 |
break; |
| 352 |
} |
| 353 |
} |
| 354 |
|
| 355 |
return res; |
| 356 |
} |
| 357 |
|
| 358 |
DataType |
| 359 |
Converter::getSType(nir_src &src, bool isFloat, bool isSigned) |
| 360 |
{ |
| 361 |
const uint8_t bitSize = src.ssa->bit_size; |
| 362 |
|
| 363 |
DataType ty = typeOfSize(bitSize / 8, isFloat, isSigned); |
| 364 |
if (ty == TYPE_NONE) { |
| 365 |
const char *str; |
| 366 |
if (isFloat) |
| 367 |
str = "float"; |
| 368 |
else if (isSigned) |
| 369 |
str = "int"; |
| 370 |
else |
| 371 |
str = "uint"; |
| 372 |
ERROR("couldn't get Type for %s with bitSize %u\n", str, bitSize); |
| 373 |
assert(false); |
| 374 |
} |
| 375 |
return ty; |
| 376 |
} |
| 377 |
|
| 378 |
DataFile |
| 379 |
Converter::getFile(nir_intrinsic_op op) const |
| 380 |
{ |
| 381 |
switch (op) { |
| 382 |
case nir_intrinsic_load_uniform: |
| 383 |
case nir_intrinsic_load_ubo: |
| 384 |
case nir_intrinsic_ldc_nv: |
| 385 |
return FILE_MEMORY_CONST; |
| 386 |
case nir_intrinsic_load_ssbo: |
| 387 |
case nir_intrinsic_store_ssbo: |
| 388 |
return FILE_MEMORY_BUFFER; |
| 389 |
case nir_intrinsic_load_global: |
| 390 |
case nir_intrinsic_store_global: |
| 391 |
case nir_intrinsic_load_global_constant: |
| 392 |
return FILE_MEMORY_GLOBAL; |
| 393 |
case nir_intrinsic_load_scratch: |
| 394 |
case nir_intrinsic_store_scratch: |
| 395 |
return FILE_MEMORY_LOCAL; |
| 396 |
case nir_intrinsic_load_shared: |
| 397 |
case nir_intrinsic_store_shared: |
| 398 |
return FILE_MEMORY_SHARED; |
| 399 |
case nir_intrinsic_load_input: |
| 400 |
case nir_intrinsic_load_interpolated_input: |
| 401 |
case nir_intrinsic_load_kernel_input: |
| 402 |
case nir_intrinsic_load_per_vertex_input: |
| 403 |
return FILE_SHADER_INPUT; |
| 404 |
case nir_intrinsic_load_output: |
| 405 |
case nir_intrinsic_load_per_vertex_output: |
| 406 |
case nir_intrinsic_store_output: |
| 407 |
case nir_intrinsic_store_per_vertex_output: |
| 408 |
return FILE_SHADER_OUTPUT; |
| 409 |
default: |
| 410 |
ERROR("couldn't get DataFile for op %s\n", nir_intrinsic_infos[op].name); |
| 411 |
assert(false); |
| 412 |
} |
| 413 |
return FILE_NULL; |
| 414 |
} |
| 415 |
|
| 416 |
operation |
| 417 |
Converter::getOperation(nir_op op) |
| 418 |
{ |
| 419 |
switch (op) { |
| 420 |
// basic ops with float and int variants |
| 421 |
case nir_op_fabs: |
| 422 |
case nir_op_iabs: |
| 423 |
return OP_ABS; |
| 424 |
case nir_op_fadd: |
| 425 |
case nir_op_iadd: |
| 426 |
return OP_ADD; |
| 427 |
case nir_op_iand: |
| 428 |
return OP_AND; |
| 429 |
case nir_op_ifind_msb: |
| 430 |
case nir_op_ufind_msb: |
| 431 |
return OP_BFIND; |
| 432 |
case nir_op_fceil: |
| 433 |
return OP_CEIL; |
| 434 |
case nir_op_fcos: |
| 435 |
return OP_COS; |
| 436 |
case nir_op_f2f32: |
| 437 |
case nir_op_f2f64: |
| 438 |
case nir_op_f2i8: |
| 439 |
case nir_op_f2i16: |
| 440 |
case nir_op_f2i32: |
| 441 |
case nir_op_f2i64: |
| 442 |
case nir_op_f2u8: |
| 443 |
case nir_op_f2u16: |
| 444 |
case nir_op_f2u32: |
| 445 |
case nir_op_f2u64: |
| 446 |
case nir_op_i2f32: |
| 447 |
case nir_op_i2f64: |
| 448 |
case nir_op_i2i8: |
| 449 |
case nir_op_i2i16: |
| 450 |
case nir_op_i2i32: |
| 451 |
case nir_op_i2i64: |
| 452 |
case nir_op_u2f32: |
| 453 |
case nir_op_u2f64: |
| 454 |
case nir_op_u2u8: |
| 455 |
case nir_op_u2u16: |
| 456 |
case nir_op_u2u32: |
| 457 |
case nir_op_u2u64: |
| 458 |
return OP_CVT; |
| 459 |
case nir_op_fdiv: |
| 460 |
case nir_op_idiv: |
| 461 |
case nir_op_udiv: |
| 462 |
return OP_DIV; |
| 463 |
case nir_op_fexp2: |
| 464 |
return OP_EX2; |
| 465 |
case nir_op_ffloor: |
| 466 |
return OP_FLOOR; |
| 467 |
case nir_op_ffma: |
| 468 |
case nir_op_ffmaz: |
| 469 |
/* No FMA op pre-nvc0 */ |
| 470 |
if (info->target < 0xc0) |
| 471 |
return OP_MAD; |
| 472 |
return OP_FMA; |
| 473 |
case nir_op_flog2: |
| 474 |
return OP_LG2; |
| 475 |
case nir_op_fmax: |
| 476 |
case nir_op_imax: |
| 477 |
case nir_op_umax: |
| 478 |
return OP_MAX; |
| 479 |
case nir_op_pack_64_2x32_split: |
| 480 |
return OP_MERGE; |
| 481 |
case nir_op_fmin: |
| 482 |
case nir_op_imin: |
| 483 |
case nir_op_umin: |
| 484 |
return OP_MIN; |
| 485 |
case nir_op_fmod: |
| 486 |
case nir_op_imod: |
| 487 |
case nir_op_umod: |
| 488 |
case nir_op_frem: |
| 489 |
case nir_op_irem: |
| 490 |
return OP_MOD; |
| 491 |
case nir_op_fmul: |
| 492 |
case nir_op_fmulz: |
| 493 |
case nir_op_amul: |
| 494 |
case nir_op_imul: |
| 495 |
case nir_op_imul_high: |
| 496 |
case nir_op_umul_high: |
| 497 |
return OP_MUL; |
| 498 |
case nir_op_fneg: |
| 499 |
case nir_op_ineg: |
| 500 |
return OP_NEG; |
| 501 |
case nir_op_inot: |
| 502 |
return OP_NOT; |
| 503 |
case nir_op_ior: |
| 504 |
return OP_OR; |
| 505 |
case nir_op_frcp: |
| 506 |
return OP_RCP; |
| 507 |
case nir_op_frsq: |
| 508 |
return OP_RSQ; |
| 509 |
case nir_op_fsat: |
| 510 |
return OP_SAT; |
| 511 |
case nir_op_ieq8: |
| 512 |
case nir_op_ige8: |
| 513 |
case nir_op_uge8: |
| 514 |
case nir_op_ilt8: |
| 515 |
case nir_op_ult8: |
| 516 |
case nir_op_ine8: |
| 517 |
case nir_op_ieq16: |
| 518 |
case nir_op_ige16: |
| 519 |
case nir_op_uge16: |
| 520 |
case nir_op_ilt16: |
| 521 |
case nir_op_ult16: |
| 522 |
case nir_op_ine16: |
| 523 |
case nir_op_feq32: |
| 524 |
case nir_op_ieq32: |
| 525 |
case nir_op_fge32: |
| 526 |
case nir_op_ige32: |
| 527 |
case nir_op_uge32: |
| 528 |
case nir_op_flt32: |
| 529 |
case nir_op_ilt32: |
| 530 |
case nir_op_ult32: |
| 531 |
case nir_op_fneu32: |
| 532 |
case nir_op_ine32: |
| 533 |
return OP_SET; |
| 534 |
case nir_op_ishl: |
| 535 |
return OP_SHL; |
| 536 |
case nir_op_ishr: |
| 537 |
case nir_op_ushr: |
| 538 |
return OP_SHR; |
| 539 |
case nir_op_fsin: |
| 540 |
return OP_SIN; |
| 541 |
case nir_op_fsqrt: |
| 542 |
return OP_SQRT; |
| 543 |
case nir_op_ftrunc: |
| 544 |
return OP_TRUNC; |
| 545 |
case nir_op_ixor: |
| 546 |
return OP_XOR; |
| 547 |
default: |
| 548 |
ERROR("couldn't get operation for op %s\n", nir_op_infos[op].name); |
| 549 |
assert(false); |
| 550 |
return OP_NOP; |
| 551 |
} |
| 552 |
} |
| 553 |
|
| 554 |
operation |
| 555 |
Converter::getOperation(nir_texop op) |
| 556 |
{ |
| 557 |
switch (op) { |
| 558 |
case nir_texop_tex: |
| 559 |
return OP_TEX; |
| 560 |
case nir_texop_lod: |
| 561 |
return OP_TXLQ; |
| 562 |
case nir_texop_txb: |
| 563 |
return OP_TXB; |
| 564 |
case nir_texop_txd: |
| 565 |
return OP_TXD; |
| 566 |
case nir_texop_txf: |
| 567 |
case nir_texop_txf_ms: |
| 568 |
return OP_TXF; |
| 569 |
case nir_texop_tg4: |
| 570 |
return OP_TXG; |
| 571 |
case nir_texop_txl: |
| 572 |
return OP_TXL; |
| 573 |
case nir_texop_query_levels: |
| 574 |
case nir_texop_texture_samples: |
| 575 |
case nir_texop_txs: |
| 576 |
return OP_TXQ; |
| 577 |
default: |
| 578 |
ERROR("couldn't get operation for nir_texop %u\n", op); |
| 579 |
assert(false); |
| 580 |
return OP_NOP; |
| 581 |
} |
| 582 |
} |
| 583 |
|
| 584 |
operation |
| 585 |
Converter::getOperation(nir_intrinsic_op op) |
| 586 |
{ |
| 587 |
switch (op) { |
| 588 |
case nir_intrinsic_emit_vertex: |
| 589 |
return OP_EMIT; |
| 590 |
case nir_intrinsic_end_primitive: |
| 591 |
return OP_RESTART; |
| 592 |
case nir_intrinsic_bindless_image_atomic: |
| 593 |
case nir_intrinsic_image_atomic: |
| 594 |
case nir_intrinsic_bindless_image_atomic_swap: |
| 595 |
case nir_intrinsic_image_atomic_swap: |
| 596 |
return OP_SUREDP; |
| 597 |
case nir_intrinsic_bindless_image_load: |
| 598 |
case nir_intrinsic_image_load: |
| 599 |
return OP_SULDP; |
| 600 |
case nir_intrinsic_bindless_image_samples: |
| 601 |
case nir_intrinsic_image_samples: |
| 602 |
case nir_intrinsic_bindless_image_size: |
| 603 |
case nir_intrinsic_image_size: |
| 604 |
return OP_SUQ; |
| 605 |
case nir_intrinsic_bindless_image_store: |
| 606 |
case nir_intrinsic_image_store: |
| 607 |
return OP_SUSTP; |
| 608 |
case nir_intrinsic_ddx: |
| 609 |
case nir_intrinsic_ddx_coarse: |
| 610 |
case nir_intrinsic_ddx_fine: |
| 611 |
return OP_DFDX; |
| 612 |
case nir_intrinsic_ddy: |
| 613 |
case nir_intrinsic_ddy_coarse: |
| 614 |
case nir_intrinsic_ddy_fine: |
| 615 |
return OP_DFDY; |
| 616 |
default: |
| 617 |
ERROR("couldn't get operation for nir_intrinsic_op %u\n", op); |
| 618 |
assert(false); |
| 619 |
return OP_NOP; |
| 620 |
} |
| 621 |
} |
| 622 |
|
| 623 |
operation |
| 624 |
Converter::preOperationNeeded(nir_op op) |
| 625 |
{ |
| 626 |
switch (op) { |
| 627 |
case nir_op_fcos: |
| 628 |
case nir_op_fsin: |
| 629 |
return OP_PRESIN; |
| 630 |
default: |
| 631 |
return OP_NOP; |
| 632 |
} |
| 633 |
} |
| 634 |
|
| 635 |
int |
| 636 |
Converter::getSubOp(nir_op op) |
| 637 |
{ |
| 638 |
switch (op) { |
| 639 |
case nir_op_imul_high: |
| 640 |
case nir_op_umul_high: |
| 641 |
return NV50_IR_SUBOP_MUL_HIGH; |
| 642 |
case nir_op_ishl: |
| 643 |
case nir_op_ishr: |
| 644 |
case nir_op_ushr: |
| 645 |
return NV50_IR_SUBOP_SHIFT_WRAP; |
| 646 |
default: |
| 647 |
return 0; |
| 648 |
} |
| 649 |
} |
| 650 |
|
| 651 |
int |
| 652 |
Converter::getAtomicSubOp(nir_atomic_op op) |
| 653 |
{ |
| 654 |
switch (op) { |
| 655 |
case nir_atomic_op_fadd: |
| 656 |
case nir_atomic_op_iadd: |
| 657 |
return NV50_IR_SUBOP_ATOM_ADD; |
| 658 |
case nir_atomic_op_iand: |
| 659 |
return NV50_IR_SUBOP_ATOM_AND; |
| 660 |
case nir_atomic_op_cmpxchg: |
| 661 |
return NV50_IR_SUBOP_ATOM_CAS; |
| 662 |
case nir_atomic_op_imax: |
| 663 |
case nir_atomic_op_umax: |
| 664 |
return NV50_IR_SUBOP_ATOM_MAX; |
| 665 |
case nir_atomic_op_imin: |
| 666 |
case nir_atomic_op_umin: |
| 667 |
return NV50_IR_SUBOP_ATOM_MIN; |
| 668 |
case nir_atomic_op_xchg: |
| 669 |
return NV50_IR_SUBOP_ATOM_EXCH; |
| 670 |
case nir_atomic_op_ior: |
| 671 |
return NV50_IR_SUBOP_ATOM_OR; |
| 672 |
case nir_atomic_op_ixor: |
| 673 |
return NV50_IR_SUBOP_ATOM_XOR; |
| 674 |
case nir_atomic_op_dec_wrap: |
| 675 |
return NV50_IR_SUBOP_ATOM_DEC; |
| 676 |
case nir_atomic_op_inc_wrap: |
| 677 |
return NV50_IR_SUBOP_ATOM_INC; |
| 678 |
default: |
| 679 |
ERROR("couldn't get SubOp for atomic\n"); |
| 680 |
assert(false); |
| 681 |
return 0; |
| 682 |
} |
| 683 |
} |
| 684 |
|
| 685 |
int |
| 686 |
Converter::getSubOp(nir_intrinsic_op op) |
| 687 |
{ |
| 688 |
switch (op) { |
| 689 |
case nir_intrinsic_vote_all: |
| 690 |
return NV50_IR_SUBOP_VOTE_ALL; |
| 691 |
case nir_intrinsic_vote_any: |
| 692 |
return NV50_IR_SUBOP_VOTE_ANY; |
| 693 |
case nir_intrinsic_vote_ieq: |
| 694 |
return NV50_IR_SUBOP_VOTE_UNI; |
| 695 |
default: |
| 696 |
return 0; |
| 697 |
} |
| 698 |
} |
| 699 |
|
| 700 |
CondCode |
| 701 |
Converter::getCondCode(nir_op op) |
| 702 |
{ |
| 703 |
switch (op) { |
| 704 |
case nir_op_ieq8: |
| 705 |
case nir_op_ieq16: |
| 706 |
case nir_op_feq32: |
| 707 |
case nir_op_ieq32: |
| 708 |
return CC_EQ; |
| 709 |
case nir_op_ige8: |
| 710 |
case nir_op_uge8: |
| 711 |
case nir_op_ige16: |
| 712 |
case nir_op_uge16: |
| 713 |
case nir_op_fge32: |
| 714 |
case nir_op_ige32: |
| 715 |
case nir_op_uge32: |
| 716 |
return CC_GE; |
| 717 |
case nir_op_ilt8: |
| 718 |
case nir_op_ult8: |
| 719 |
case nir_op_ilt16: |
| 720 |
case nir_op_ult16: |
| 721 |
case nir_op_flt32: |
| 722 |
case nir_op_ilt32: |
| 723 |
case nir_op_ult32: |
| 724 |
return CC_LT; |
| 725 |
case nir_op_fneu32: |
| 726 |
return CC_NEU; |
| 727 |
case nir_op_ine8: |
| 728 |
case nir_op_ine16: |
| 729 |
case nir_op_ine32: |
| 730 |
return CC_NE; |
| 731 |
default: |
| 732 |
ERROR("couldn't get CondCode for op %s\n", nir_op_infos[op].name); |
| 733 |
assert(false); |
| 734 |
return CC_FL; |
| 735 |
} |
| 736 |
} |
| 737 |
|
| 738 |
Converter::LValues& |
| 739 |
Converter::convert(nir_def *def) |
| 740 |
{ |
| 741 |
NirDefMap::iterator it = ssaDefs.find(def->index); |
| 742 |
if (it != ssaDefs.end()) |
| 743 |
return it->second; |
| 744 |
|
| 745 |
LValues newDef(def->num_components); |
| 746 |
for (uint8_t i = 0; i < def->num_components; i++) |
| 747 |
newDef[i] = getSSA(std::max(4, def->bit_size / 8)); |
| 748 |
return ssaDefs[def->index] = newDef; |
| 749 |
} |
| 750 |
|
| 751 |
Value* |
| 752 |
Converter::getSrc(nir_alu_src *src, uint8_t component) |
| 753 |
{ |
| 754 |
return getSrc(&src->src, src->swizzle[component]); |
| 755 |
} |
| 756 |
|
| 757 |
Value* |
| 758 |
Converter::getSrc(nir_src *src, uint8_t idx, bool indirect) |
| 759 |
{ |
| 760 |
return getSrc(src->ssa, idx); |
| 761 |
} |
| 762 |
|
| 763 |
Value* |
| 764 |
Converter::getSrc(nir_def *src, uint8_t idx) |
| 765 |
{ |
| 766 |
ImmediateMap::iterator iit = immediates.find(src->index); |
| 767 |
if (iit != immediates.end()) |
| 768 |
return convert((*iit).second, idx); |
| 769 |
|
| 770 |
NirDefMap::iterator it = ssaDefs.find(src->index); |
| 771 |
if (it == ssaDefs.end()) { |
| 772 |
ERROR("SSA value %u not found\n", src->index); |
| 773 |
assert(false); |
| 774 |
return NULL; |
| 775 |
} |
| 776 |
return it->second[idx]; |
| 777 |
} |
| 778 |
|
| 779 |
uint32_t |
| 780 |
Converter::getIndirect(nir_src *src, uint8_t idx, Value *&indirect) |
| 781 |
{ |
| 782 |
nir_const_value *offset = nir_src_as_const_value(*src); |
| 783 |
|
| 784 |
if (offset) { |
| 785 |
indirect = NULL; |
| 786 |
return offset[0].u32; |
| 787 |
} |
| 788 |
|
| 789 |
indirect = getSrc(src, idx, true); |
| 790 |
return 0; |
| 791 |
} |
| 792 |
|
| 793 |
uint32_t |
| 794 |
Converter::getIndirect(nir_intrinsic_instr *insn, uint8_t s, uint8_t c, Value *&indirect, bool isScalar) |
| 795 |
{ |
| 796 |
int32_t idx = nir_intrinsic_base(insn) + getIndirect(&insn->src[s], c, indirect); |
| 797 |
|
| 798 |
if (indirect && !isScalar) |
| 799 |
indirect = mkOp2v(OP_SHL, TYPE_U32, getSSA(4, FILE_ADDRESS), indirect, loadImm(NULL, 4)); |
| 800 |
return idx; |
| 801 |
} |
| 802 |
|
| 803 |
static void |
| 804 |
vert_attrib_to_tgsi_semantic(gl_vert_attrib slot, unsigned *name, unsigned *index) |
| 805 |
{ |
| 806 |
assert(name && index); |
| 807 |
|
| 808 |
if (slot >= VERT_ATTRIB_MAX) { |
| 809 |
ERROR("invalid varying slot %u\n", slot); |
| 810 |
assert(false); |
| 811 |
return; |
| 812 |
} |
| 813 |
|
| 814 |
if (slot >= VERT_ATTRIB_GENERIC0 && |
| 815 |
slot < VERT_ATTRIB_GENERIC0 + VERT_ATTRIB_GENERIC_MAX) { |
| 816 |
*name = TGSI_SEMANTIC_GENERIC; |
| 817 |
*index = slot - VERT_ATTRIB_GENERIC0; |
| 818 |
return; |
| 819 |
} |
| 820 |
|
| 821 |
if (slot >= VERT_ATTRIB_TEX0 && |
| 822 |
slot < VERT_ATTRIB_TEX0 + VERT_ATTRIB_TEX_MAX) { |
| 823 |
*name = TGSI_SEMANTIC_TEXCOORD; |
| 824 |
*index = slot - VERT_ATTRIB_TEX0; |
| 825 |
return; |
| 826 |
} |
| 827 |
|
| 828 |
switch (slot) { |
| 829 |
case VERT_ATTRIB_COLOR0: |
| 830 |
*name = TGSI_SEMANTIC_COLOR; |
| 831 |
*index = 0; |
| 832 |
break; |
| 833 |
case VERT_ATTRIB_COLOR1: |
| 834 |
*name = TGSI_SEMANTIC_COLOR; |
| 835 |
*index = 1; |
| 836 |
break; |
| 837 |
case VERT_ATTRIB_EDGEFLAG: |
| 838 |
*name = TGSI_SEMANTIC_EDGEFLAG; |
| 839 |
*index = 0; |
| 840 |
break; |
| 841 |
case VERT_ATTRIB_FOG: |
| 842 |
*name = TGSI_SEMANTIC_FOG; |
| 843 |
*index = 0; |
| 844 |
break; |
| 845 |
case VERT_ATTRIB_NORMAL: |
| 846 |
*name = TGSI_SEMANTIC_NORMAL; |
| 847 |
*index = 0; |
| 848 |
break; |
| 849 |
case VERT_ATTRIB_POS: |
| 850 |
*name = TGSI_SEMANTIC_POSITION; |
| 851 |
*index = 0; |
| 852 |
break; |
| 853 |
case VERT_ATTRIB_POINT_SIZE: |
| 854 |
*name = TGSI_SEMANTIC_PSIZE; |
| 855 |
*index = 0; |
| 856 |
break; |
| 857 |
default: |
| 858 |
ERROR("unknown vert attrib slot %u\n", slot); |
| 859 |
assert(false); |
| 860 |
break; |
| 861 |
} |
| 862 |
} |
| 863 |
|
| 864 |
uint8_t |
| 865 |
Converter::translateInterpMode(const struct nv50_ir_varying *var, operation& op) |
| 866 |
{ |
| 867 |
uint8_t mode = NV50_IR_INTERP_PERSPECTIVE; |
| 868 |
|
| 869 |
if (var->flat) |
| 870 |
mode = NV50_IR_INTERP_FLAT; |
| 871 |
else |
| 872 |
if (var->linear) |
| 873 |
mode = NV50_IR_INTERP_LINEAR; |
| 874 |
else |
| 875 |
if (var->sc) |
| 876 |
mode = NV50_IR_INTERP_SC; |
| 877 |
|
| 878 |
op = (mode == NV50_IR_INTERP_PERSPECTIVE || mode == NV50_IR_INTERP_SC) |
| 879 |
? OP_PINTERP : OP_LINTERP; |
| 880 |
|
| 881 |
if (var->centroid) |
| 882 |
mode |= NV50_IR_INTERP_CENTROID; |
| 883 |
|
| 884 |
return mode; |
| 885 |
} |
| 886 |
|
| 887 |
void |
| 888 |
Converter::setInterpolate(nv50_ir_varying *var, |
| 889 |
uint8_t mode, |
| 890 |
bool centroid, |
| 891 |
unsigned semantic) |
| 892 |
{ |
| 893 |
switch (mode) { |
| 894 |
case INTERP_MODE_FLAT: |
| 895 |
var->flat = 1; |
| 896 |
break; |
| 897 |
case INTERP_MODE_NONE: |
| 898 |
if (semantic == TGSI_SEMANTIC_COLOR) |
| 899 |
var->sc = 1; |
| 900 |
else if (semantic == TGSI_SEMANTIC_POSITION) |
| 901 |
var->linear = 1; |
| 902 |
break; |
| 903 |
case INTERP_MODE_NOPERSPECTIVE: |
| 904 |
var->linear = 1; |
| 905 |
break; |
| 906 |
case INTERP_MODE_SMOOTH: |
| 907 |
break; |
| 908 |
} |
| 909 |
var->centroid = centroid; |
| 910 |
} |
| 911 |
|
| 912 |
static uint16_t |
| 913 |
calcSlots(const glsl_type *type, Program::Type stage, const shader_info &info, |
| 914 |
bool input, const nir_variable *var) |
| 915 |
{ |
| 916 |
if (!glsl_type_is_array(type)) |
| 917 |
return glsl_count_attribute_slots(type, false); |
| 918 |
|
| 919 |
uint16_t slots; |
| 920 |
switch (stage) { |
| 921 |
case Program::TYPE_GEOMETRY: |
| 922 |
slots = glsl_count_attribute_slots(type, false); |
| 923 |
if (input) |
| 924 |
slots /= info.gs.vertices_in; |
| 925 |
break; |
| 926 |
case Program::TYPE_TESSELLATION_CONTROL: |
| 927 |
case Program::TYPE_TESSELLATION_EVAL: |
| 928 |
// remove first dimension |
| 929 |
if (var->data.patch || (!input && stage == Program::TYPE_TESSELLATION_EVAL)) |
| 930 |
slots = glsl_count_attribute_slots(type, false); |
| 931 |
else |
| 932 |
slots = glsl_count_attribute_slots(type->fields.array, false); |
| 933 |
break; |
| 934 |
default: |
| 935 |
slots = glsl_count_attribute_slots(type, false); |
| 936 |
break; |
| 937 |
} |
| 938 |
|
| 939 |
return slots; |
| 940 |
} |
| 941 |
|
| 942 |
static uint8_t |
| 943 |
getMaskForType(const glsl_type *type, uint8_t slot) { |
| 944 |
uint16_t comp = glsl_get_components(glsl_without_array(type)); |
| 945 |
comp = comp ? comp : 4; |
| 946 |
|
| 947 |
if (glsl_base_type_is_64bit(glsl_without_array(type)->base_type)) { |
| 948 |
comp *= 2; |
| 949 |
if (comp > 4) { |
| 950 |
if (slot % 2) |
| 951 |
comp -= 4; |
| 952 |
else |
| 953 |
comp = 4; |
| 954 |
} |
| 955 |
} |
| 956 |
|
| 957 |
return (1 << comp) - 1; |
| 958 |
} |
| 959 |
|
| 960 |
bool Converter::assignSlots() { |
| 961 |
unsigned name; |
| 962 |
unsigned index; |
| 963 |
|
| 964 |
info->io.viewportId = -1; |
| 965 |
info->io.mul_zero_wins = nir->info.use_legacy_math_rules; |
| 966 |
info_out->numInputs = 0; |
| 967 |
info_out->numOutputs = 0; |
| 968 |
info_out->numSysVals = 0; |
| 969 |
|
| 970 |
uint8_t i; |
| 971 |
BITSET_FOREACH_SET(i, nir->info.system_values_read, SYSTEM_VALUE_MAX) { |
| 972 |
switch (i) { |
| 973 |
case SYSTEM_VALUE_VERTEX_ID: |
| 974 |
info_out->io.vertexId = info_out->numSysVals; |
| 975 |
break; |
| 976 |
case SYSTEM_VALUE_INSTANCE_ID: |
| 977 |
info_out->io.instanceId = info_out->numSysVals; |
| 978 |
break; |
| 979 |
default: |
| 980 |
break; |
| 981 |
} |
| 982 |
|
| 983 |
info_out->sv[info_out->numSysVals].sn = (gl_system_value)i; |
| 984 |
info_out->numSysVals += 1; |
| 985 |
} |
| 986 |
|
| 987 |
if (prog->getType() == Program::TYPE_COMPUTE) |
| 988 |
return true; |
| 989 |
|
| 990 |
nir_foreach_shader_in_variable(var, nir) { |
| 991 |
const glsl_type *type = var->type; |
| 992 |
int slot = var->data.location; |
| 993 |
uint16_t slots = calcSlots(type, prog->getType(), nir->info, true, var); |
| 994 |
uint32_t vary = var->data.driver_location; |
| 995 |
assert(vary + slots <= NV50_CODEGEN_MAX_VARYINGS); |
| 996 |
|
| 997 |
switch(prog->getType()) { |
| 998 |
case Program::TYPE_FRAGMENT: |
| 999 |
tgsi_get_gl_varying_semantic((gl_varying_slot)slot, true, |
| 1000 |
&name, &index); |
| 1001 |
for (uint16_t i = 0; i < slots; ++i) { |
| 1002 |
setInterpolate(&info_out->in[vary + i], var->data.interpolation, |
| 1003 |
var->data.centroid | var->data.sample, name); |
| 1004 |
} |
| 1005 |
break; |
| 1006 |
case Program::TYPE_GEOMETRY: |
| 1007 |
tgsi_get_gl_varying_semantic((gl_varying_slot)slot, true, |
| 1008 |
&name, &index); |
| 1009 |
break; |
| 1010 |
case Program::TYPE_TESSELLATION_CONTROL: |
| 1011 |
case Program::TYPE_TESSELLATION_EVAL: |
| 1012 |
tgsi_get_gl_varying_semantic((gl_varying_slot)slot, true, |
| 1013 |
&name, &index); |
| 1014 |
if (var->data.patch && name == TGSI_SEMANTIC_PATCH) |
| 1015 |
info_out->numPatchConstants = MAX2(info_out->numPatchConstants, index + slots); |
| 1016 |
break; |
| 1017 |
case Program::TYPE_VERTEX: |
| 1018 |
if (slot >= VERT_ATTRIB_GENERIC0 && slot < VERT_ATTRIB_GENERIC0 + VERT_ATTRIB_GENERIC_MAX) |
| 1019 |
slot = VERT_ATTRIB_GENERIC0 + vary; |
| 1020 |
vert_attrib_to_tgsi_semantic((gl_vert_attrib)slot, &name, &index); |
| 1021 |
switch (name) { |
| 1022 |
case TGSI_SEMANTIC_EDGEFLAG: |
| 1023 |
info_out->io.edgeFlagIn = vary; |
| 1024 |
break; |
| 1025 |
default: |
| 1026 |
break; |
| 1027 |
} |
| 1028 |
break; |
| 1029 |
default: |
| 1030 |
ERROR("unknown shader type %u in assignSlots\n", prog->getType()); |
| 1031 |
return false; |
| 1032 |
} |
| 1033 |
|
| 1034 |
if (var->data.compact) { |
| 1035 |
assert(!(nir->info.outputs_read & 1ull << slot)); |
| 1036 |
if (nir_is_arrayed_io(var, nir->info.stage)) { |
| 1037 |
assert(glsl_type_is_array(type->fields.array)); |
| 1038 |
assert(glsl_type_is_scalar(type->fields.array->fields.array)); |
| 1039 |
assert(slots == glsl_get_length(type->fields.array)); |
| 1040 |
} else { |
| 1041 |
assert(glsl_type_is_array(type)); |
| 1042 |
assert(glsl_type_is_scalar(type->fields.array)); |
| 1043 |
assert(slots == glsl_get_length(type)); |
| 1044 |
} |
| 1045 |
assert(!glsl_base_type_is_64bit(glsl_without_array(type)->base_type)); |
| 1046 |
|
| 1047 |
uint32_t comps = BITFIELD_RANGE(var->data.location_frac, slots); |
| 1048 |
assert(!(comps & ~0xff)); |
| 1049 |
|
| 1050 |
if (comps & 0x0f) { |
| 1051 |
nv50_ir_varying *v = &info_out->in[vary]; |
| 1052 |
v->patch = var->data.patch; |
| 1053 |
v->sn = name; |
| 1054 |
v->si = index; |
| 1055 |
v->mask |= comps & 0x0f; |
| 1056 |
info_out->numInputs = |
| 1057 |
std::max<uint8_t>(info_out->numInputs, vary + 1); |
| 1058 |
} |
| 1059 |
if (comps & 0xf0) { |
| 1060 |
nv50_ir_varying *v = &info_out->in[vary + 1]; |
| 1061 |
v->patch = var->data.patch; |
| 1062 |
v->sn = name; |
| 1063 |
v->si = index + 1; |
| 1064 |
v->mask |= (comps & 0xf0) >> 4; |
| 1065 |
info_out->numInputs = |
| 1066 |
std::max<uint8_t>(info_out->numInputs, vary + 2); |
| 1067 |
} |
| 1068 |
} else { |
| 1069 |
for (uint16_t i = 0u; i < slots; ++i, ++vary) { |
| 1070 |
nv50_ir_varying *v = &info_out->in[vary]; |
| 1071 |
|
| 1072 |
v->patch = var->data.patch; |
| 1073 |
v->sn = name; |
| 1074 |
v->si = index + i; |
| 1075 |
v->mask |= getMaskForType(type, i) << var->data.location_frac; |
| 1076 |
} |
| 1077 |
info_out->numInputs = std::max<uint8_t>(info_out->numInputs, vary); |
| 1078 |
} |
| 1079 |
} |
| 1080 |
|
| 1081 |
nir_foreach_shader_out_variable(var, nir) { |
| 1082 |
const glsl_type *type = var->type; |
| 1083 |
int slot = var->data.location; |
| 1084 |
uint16_t slots = calcSlots(type, prog->getType(), nir->info, false, var); |
| 1085 |
uint32_t vary = var->data.driver_location; |
| 1086 |
|
| 1087 |
assert(vary < NV50_CODEGEN_MAX_VARYINGS); |
| 1088 |
|
| 1089 |
switch(prog->getType()) { |
| 1090 |
case Program::TYPE_FRAGMENT: |
| 1091 |
tgsi_get_gl_frag_result_semantic((gl_frag_result)slot, &name, &index); |
| 1092 |
switch (name) { |
| 1093 |
case TGSI_SEMANTIC_COLOR: |
| 1094 |
if (!var->data.fb_fetch_output) |
| 1095 |
info_out->prop.fp.numColourResults++; |
| 1096 |
if (var->data.location == FRAG_RESULT_COLOR && |
| 1097 |
nir->info.outputs_written & BITFIELD64_BIT(var->data.location)) |
| 1098 |
info_out->prop.fp.separateFragData = true; |
| 1099 |
// sometimes we get FRAG_RESULT_DATAX with data.index 0 |
| 1100 |
// sometimes we get FRAG_RESULT_DATA0 with data.index X |
| 1101 |
index = index == 0 ? var->data.index : index; |
| 1102 |
break; |
| 1103 |
case TGSI_SEMANTIC_POSITION: |
| 1104 |
info_out->io.fragDepth = vary; |
| 1105 |
info_out->prop.fp.writesDepth = true; |
| 1106 |
break; |
| 1107 |
case TGSI_SEMANTIC_SAMPLEMASK: |
| 1108 |
info_out->io.sampleMask = vary; |
| 1109 |
break; |
| 1110 |
default: |
| 1111 |
break; |
| 1112 |
} |
| 1113 |
break; |
| 1114 |
case Program::TYPE_GEOMETRY: |
| 1115 |
case Program::TYPE_TESSELLATION_CONTROL: |
| 1116 |
case Program::TYPE_TESSELLATION_EVAL: |
| 1117 |
case Program::TYPE_VERTEX: |
| 1118 |
tgsi_get_gl_varying_semantic((gl_varying_slot)slot, true, |
| 1119 |
&name, &index); |
| 1120 |
|
| 1121 |
if (var->data.patch && name != TGSI_SEMANTIC_TESSINNER && |
| 1122 |
name != TGSI_SEMANTIC_TESSOUTER) |
| 1123 |
info_out->numPatchConstants = MAX2(info_out->numPatchConstants, index + slots); |
| 1124 |
|
| 1125 |
switch (name) { |
| 1126 |
case TGSI_SEMANTIC_EDGEFLAG: |
| 1127 |
info_out->io.edgeFlagOut = vary; |
| 1128 |
break; |
| 1129 |
default: |
| 1130 |
break; |
| 1131 |
} |
| 1132 |
break; |
| 1133 |
default: |
| 1134 |
ERROR("unknown shader type %u in assignSlots\n", prog->getType()); |
| 1135 |
return false; |
| 1136 |
} |
| 1137 |
|
| 1138 |
if (var->data.compact) { |
| 1139 |
assert(!(nir->info.outputs_read & 1ull << slot)); |
| 1140 |
if (nir_is_arrayed_io(var, nir->info.stage)) { |
| 1141 |
assert(glsl_type_is_array(type->fields.array)); |
| 1142 |
assert(glsl_type_is_scalar(type->fields.array->fields.array)); |
| 1143 |
assert(slots == glsl_get_length(type->fields.array)); |
| 1144 |
} else { |
| 1145 |
assert(glsl_type_is_array(type)); |
| 1146 |
assert(glsl_type_is_scalar(type->fields.array)); |
| 1147 |
assert(slots == glsl_get_length(type)); |
| 1148 |
} |
| 1149 |
assert(!glsl_base_type_is_64bit(glsl_without_array(type)->base_type)); |
| 1150 |
|
| 1151 |
uint32_t comps = BITFIELD_RANGE(var->data.location_frac, slots); |
| 1152 |
assert(!(comps & ~0xff)); |
| 1153 |
|
| 1154 |
if (comps & 0x0f) { |
| 1155 |
nv50_ir_varying *v = &info_out->out[vary]; |
| 1156 |
v->patch = var->data.patch; |
| 1157 |
v->sn = name; |
| 1158 |
v->si = index; |
| 1159 |
v->mask |= comps & 0x0f; |
| 1160 |
info_out->numOutputs = |
| 1161 |
std::max<uint8_t>(info_out->numOutputs, vary + 1); |
| 1162 |
} |
| 1163 |
if (comps & 0xf0) { |
| 1164 |
nv50_ir_varying *v = &info_out->out[vary + 1]; |
| 1165 |
v->patch = var->data.patch; |
| 1166 |
v->sn = name; |
| 1167 |
v->si = index + 1; |
| 1168 |
v->mask |= (comps & 0xf0) >> 4; |
| 1169 |
info_out->numOutputs = |
| 1170 |
std::max<uint8_t>(info_out->numOutputs, vary + 2); |
| 1171 |
} |
| 1172 |
} else { |
| 1173 |
for (uint16_t i = 0u; i < slots; ++i, ++vary) { |
| 1174 |
nv50_ir_varying *v = &info_out->out[vary]; |
| 1175 |
v->patch = var->data.patch; |
| 1176 |
v->sn = name; |
| 1177 |
v->si = index + i; |
| 1178 |
v->mask |= getMaskForType(type, i) << var->data.location_frac; |
| 1179 |
|
| 1180 |
if (nir->info.outputs_read & 1ull << slot) |
| 1181 |
v->oread = 1; |
| 1182 |
} |
| 1183 |
info_out->numOutputs = std::max<uint8_t>(info_out->numOutputs, vary); |
| 1184 |
} |
| 1185 |
} |
| 1186 |
|
| 1187 |
return info->assignSlots(info_out) == 0; |
| 1188 |
} |
| 1189 |
|
| 1190 |
uint32_t |
| 1191 |
Converter::getSlotAddress(nir_intrinsic_instr *insn, uint8_t idx, uint8_t slot) |
| 1192 |
{ |
| 1193 |
DataType ty; |
| 1194 |
int offset = nir_intrinsic_component(insn); |
| 1195 |
bool input; |
| 1196 |
|
| 1197 |
if (nir_intrinsic_infos[insn->intrinsic].has_dest) |
| 1198 |
ty = getDType(insn); |
| 1199 |
else |
| 1200 |
ty = getSType(insn->src[0], false, false); |
| 1201 |
|
| 1202 |
switch (insn->intrinsic) { |
| 1203 |
case nir_intrinsic_load_input: |
| 1204 |
case nir_intrinsic_load_interpolated_input: |
| 1205 |
case nir_intrinsic_load_per_vertex_input: |
| 1206 |
input = true; |
| 1207 |
break; |
| 1208 |
case nir_intrinsic_load_output: |
| 1209 |
case nir_intrinsic_load_per_vertex_output: |
| 1210 |
case nir_intrinsic_store_output: |
| 1211 |
case nir_intrinsic_store_per_vertex_output: |
| 1212 |
input = false; |
| 1213 |
break; |
| 1214 |
default: |
| 1215 |
ERROR("unknown intrinsic in getSlotAddress %s", |
| 1216 |
nir_intrinsic_infos[insn->intrinsic].name); |
| 1217 |
input = false; |
| 1218 |
assert(false); |
| 1219 |
break; |
| 1220 |
} |
| 1221 |
|
| 1222 |
if (typeSizeof(ty) == 8) { |
| 1223 |
slot *= 2; |
| 1224 |
slot += offset; |
| 1225 |
if (slot >= 4) { |
| 1226 |
idx += 1; |
| 1227 |
slot -= 4; |
| 1228 |
} |
| 1229 |
} else { |
| 1230 |
slot += offset; |
| 1231 |
} |
| 1232 |
|
| 1233 |
assert(slot < 4); |
| 1234 |
assert(!input || idx < NV50_CODEGEN_MAX_VARYINGS); |
| 1235 |
assert(input || idx < NV50_CODEGEN_MAX_VARYINGS); |
| 1236 |
|
| 1237 |
const nv50_ir_varying *vary = input ? info_out->in : info_out->out; |
| 1238 |
return vary[idx].slot[slot] * 4; |
| 1239 |
} |
| 1240 |
|
| 1241 |
Instruction * |
| 1242 |
Converter::loadFrom(DataFile file, uint8_t i, DataType ty, Value *def, |
| 1243 |
uint32_t base, uint8_t c, Value *indirect0, |
| 1244 |
Value *indirect1, bool patch, CacheMode cache) |
| 1245 |
{ |
| 1246 |
unsigned int tySize = typeSizeof(ty); |
| 1247 |
|
| 1248 |
if (tySize == 8 && |
| 1249 |
(indirect0 || !prog->getTarget()->isAccessSupported(file, TYPE_U64))) { |
| 1250 |
Value *lo = getSSA(); |
| 1251 |
Value *hi = getSSA(); |
| 1252 |
|
| 1253 |
Instruction *loi = |
| 1254 |
mkLoad(TYPE_U32, lo, |
| 1255 |
mkSymbol(file, i, TYPE_U32, base + c * tySize), |
| 1256 |
indirect0); |
| 1257 |
loi->setIndirect(0, 1, indirect1); |
| 1258 |
loi->perPatch = patch; |
| 1259 |
loi->cache = cache; |
| 1260 |
|
| 1261 |
Instruction *hii = |
| 1262 |
mkLoad(TYPE_U32, hi, |
| 1263 |
mkSymbol(file, i, TYPE_U32, base + c * tySize + 4), |
| 1264 |
indirect0); |
| 1265 |
hii->setIndirect(0, 1, indirect1); |
| 1266 |
hii->perPatch = patch; |
| 1267 |
hii->cache = cache; |
| 1268 |
|
| 1269 |
return mkOp2(OP_MERGE, ty, def, lo, hi); |
| 1270 |
} else { |
| 1271 |
Instruction *ld = |
| 1272 |
mkLoad(ty, def, mkSymbol(file, i, ty, base + c * tySize), indirect0); |
| 1273 |
ld->setIndirect(0, 1, indirect1); |
| 1274 |
ld->perPatch = patch; |
| 1275 |
ld->cache = cache; |
| 1276 |
return ld; |
| 1277 |
} |
| 1278 |
} |
| 1279 |
|
| 1280 |
Instruction * |
| 1281 |
Converter::loadVector(nir_intrinsic_instr *insn, |
| 1282 |
uint8_t buffer, Value *indirectBuffer, |
| 1283 |
uint32_t offset, Value *indirectOffset) |
| 1284 |
{ |
| 1285 |
uint32_t load_bytes = insn->def.bit_size / 8 * insn->def.num_components; |
| 1286 |
DataType ty = typeOfSize(load_bytes, false, false); |
| 1287 |
DataFile file = getFile(insn->intrinsic); |
| 1288 |
|
| 1289 |
LValues &newDefs = convert(&insn->def); |
| 1290 |
Value* def; |
| 1291 |
if (insn->def.num_components == 1) { |
| 1292 |
def = newDefs[0]; |
| 1293 |
} else { |
| 1294 |
def = getSSA(load_bytes); |
| 1295 |
} |
| 1296 |
|
| 1297 |
Instruction *ld = mkLoad(ty, def, mkSymbol(file, buffer, ty, offset), indirectOffset); |
| 1298 |
ld->setIndirect(0, 1, indirectBuffer); |
| 1299 |
|
| 1300 |
if (insn->def.num_components != 1) { |
| 1301 |
Instruction *split = mkOp1(OP_SPLIT, ty, newDefs[0], def); |
| 1302 |
for (int i = 1; i < insn->def.num_components; i++) { |
| 1303 |
split->setDef(i, newDefs[i]); |
| 1304 |
} |
| 1305 |
} |
| 1306 |
|
| 1307 |
return ld; |
| 1308 |
} |
| 1309 |
|
| 1310 |
void |
| 1311 |
Converter::storeTo(nir_intrinsic_instr *insn, DataFile file, operation op, |
| 1312 |
DataType ty, Value *src, uint8_t idx, uint8_t c, |
| 1313 |
Value *indirect0, Value *indirect1) |
| 1314 |
{ |
| 1315 |
uint8_t size = typeSizeof(ty); |
| 1316 |
uint32_t address = getSlotAddress(insn, idx, c); |
| 1317 |
|
| 1318 |
if (size == 8 && indirect0) { |
| 1319 |
Value *split[2]; |
| 1320 |
mkSplit(split, 4, src); |
| 1321 |
|
| 1322 |
if (op == OP_EXPORT) { |
| 1323 |
split[0] = mkMov(getSSA(), split[0], ty)->getDef(0); |
| 1324 |
split[1] = mkMov(getSSA(), split[1], ty)->getDef(0); |
| 1325 |
} |
| 1326 |
|
| 1327 |
mkStore(op, TYPE_U32, mkSymbol(file, 0, TYPE_U32, address), indirect0, |
| 1328 |
split[0])->perPatch = info_out->out[idx].patch; |
| 1329 |
mkStore(op, TYPE_U32, mkSymbol(file, 0, TYPE_U32, address + 4), indirect0, |
| 1330 |
split[1])->perPatch = info_out->out[idx].patch; |
| 1331 |
} else { |
| 1332 |
if (op == OP_EXPORT) |
| 1333 |
src = mkMov(getSSA(size), src, ty)->getDef(0); |
| 1334 |
mkStore(op, ty, mkSymbol(file, 0, ty, address), indirect0, |
| 1335 |
src)->perPatch = info_out->out[idx].patch; |
| 1336 |
} |
| 1337 |
} |
| 1338 |
|
| 1339 |
Instruction * |
| 1340 |
Converter::storeVector(nir_intrinsic_instr *insn, |
| 1341 |
uint8_t buffer, Value *indirectBuffer, |
| 1342 |
uint32_t offset, Value *indirectOffset) |
| 1343 |
{ |
| 1344 |
const uint8_t num_components = insn->src[0].ssa->num_components; |
| 1345 |
uint32_t bytes = insn->src[0].ssa->bit_size / 8 * num_components; |
| 1346 |
DataType ty = typeOfSize(bytes, false, false); |
| 1347 |
DataFile file = getFile(insn->intrinsic); |
| 1348 |
assert(nir_intrinsic_write_mask(insn) == nir_component_mask(num_components)); |
| 1349 |
|
| 1350 |
Value* src; |
| 1351 |
if (num_components == 1) { |
| 1352 |
src = getSrc(&insn->src[0], 0); |
| 1353 |
} else { |
| 1354 |
src = getSSA(bytes); |
| 1355 |
|
| 1356 |
Instruction *merge = mkOp(OP_MERGE, ty, src); |
| 1357 |
for (int i = 0; i < num_components; i++) { |
| 1358 |
merge->setSrc(i, getSrc(&insn->src[0], i)); |
| 1359 |
} |
| 1360 |
} |
| 1361 |
|
| 1362 |
Instruction *st = mkStore(OP_STORE, ty, mkSymbol(file, buffer, ty, offset), |
| 1363 |
indirectOffset, src); |
| 1364 |
st->setIndirect(0, 1, indirectBuffer); |
| 1365 |
|
| 1366 |
return st; |
| 1367 |
} |
| 1368 |
|
| 1369 |
bool |
| 1370 |
Converter::memVectorizeCb(unsigned align_mul, |
| 1371 |
unsigned align_offset, |
| 1372 |
unsigned bit_size, |
| 1373 |
unsigned num_components, |
| 1374 |
int64_t hole_size, |
| 1375 |
nir_intrinsic_instr *low, |
| 1376 |
nir_intrinsic_instr *high, |
| 1377 |
void *cb_data) |
| 1378 |
{ |
| 1379 |
if (hole_size > 0) |
| 1380 |
return false; |
| 1381 |
|
| 1382 |
/* |
| 1383 |
* Since we legalize these later with nir_lower_mem_access_bit_sizes, |
| 1384 |
* we can optimistically combine anything that might be profitable |
| 1385 |
*/ |
| 1386 |
const Converter* converter = (Converter*) cb_data; |
| 1387 |
const Target* target = converter->prog->getTarget(); |
| 1388 |
|
| 1389 |
assert(util_is_power_of_two_nonzero(align_mul)); |
| 1390 |
align_mul = std::min(align_mul, 128u / 8u); |
| 1391 |
|
| 1392 |
const DataFile file = converter->getFile(low->intrinsic); |
| 1393 |
if (align_mul == 128u / 8u && !target->isAccessSupported(file, TYPE_B128)) |
| 1394 |
align_mul = 64u / 8u; |
| 1395 |
|
| 1396 |
if (align_mul == 64u / 8u && !target->isAccessSupported(file, TYPE_U64)) |
| 1397 |
align_mul = 32u / 8u; |
| 1398 |
|
| 1399 |
align_offset = align_offset % align_mul; |
| 1400 |
|
| 1401 |
return align_offset + num_components * (bit_size / 8) <= align_mul; |
| 1402 |
} |
| 1403 |
|
| 1404 |
nir_mem_access_size_align |
| 1405 |
Converter::getMemAccessSizeAlign(nir_intrinsic_op intrin, |
| 1406 |
uint8_t original_bytes, |
| 1407 |
uint8_t original_bit_size, |
| 1408 |
uint32_t align_mul, |
| 1409 |
uint32_t align_offset, |
| 1410 |
bool offset_is_const, |
| 1411 |
enum gl_access_qualifier access, |
| 1412 |
const void *cb_data) |
| 1413 |
{ |
| 1414 |
const Converter* converter = (Converter*) cb_data; |
| 1415 |
const Target* target = converter->prog->getTarget(); |
| 1416 |
|
| 1417 |
const uint32_t align = nir_combined_align(align_mul, align_offset); |
| 1418 |
|
| 1419 |
assert(original_bytes != 0); |
| 1420 |
uint32_t bytes = 1 << (util_last_bit(original_bytes) - 1); |
| 1421 |
|
| 1422 |
bytes = std::min(bytes, align); |
| 1423 |
bytes = std::min(bytes, 128u / 8u); |
| 1424 |
|
| 1425 |
assert(util_is_power_of_two_nonzero(bytes)); |
| 1426 |
|
| 1427 |
const DataFile file = converter->getFile(intrin); |
| 1428 |
if (bytes == 128u / 8u && !target->isAccessSupported(file, TYPE_B128)) |
| 1429 |
bytes = 64u / 8u; |
| 1430 |
|
| 1431 |
if (bytes == 64u / 8u && !target->isAccessSupported(file, TYPE_U64)) |
| 1432 |
bytes = 32u / 8u; |
| 1433 |
|
| 1434 |
uint32_t bit_size = original_bit_size; |
| 1435 |
bit_size = std::max(bit_size, 32u); |
| 1436 |
bit_size = std::min(bit_size, bytes * 8u); |
| 1437 |
|
| 1438 |
return { |
| 1439 |
.num_components = (uint8_t) (bytes / (bit_size / 8)), |
| 1440 |
.bit_size = (uint8_t) bit_size, |
| 1441 |
.align = (uint16_t) bytes, |
| 1442 |
.shift = nir_mem_access_shift_method_scalar, |
| 1443 |
}; |
| 1444 |
} |
| 1445 |
|
| 1446 |
bool |
| 1447 |
Converter::parseNIR() |
| 1448 |
{ |
| 1449 |
info_out->bin.tlsSpace = nir->scratch_size; |
| 1450 |
info_out->io.clipDistances = nir->info.clip_distance_array_size; |
| 1451 |
info_out->io.cullDistances = nir->info.cull_distance_array_size; |
| 1452 |
info_out->io.layer_viewport_relative = nir->info.layer_viewport_relative; |
| 1453 |
|
| 1454 |
switch(prog->getType()) { |
| 1455 |
case Program::TYPE_COMPUTE: |
| 1456 |
info->prop.cp.numThreads[0] = nir->info.workgroup_size[0]; |
| 1457 |
info->prop.cp.numThreads[1] = nir->info.workgroup_size[1]; |
| 1458 |
info->prop.cp.numThreads[2] = nir->info.workgroup_size[2]; |
| 1459 |
info_out->bin.smemSize = std::max(info_out->bin.smemSize, nir->info.shared_size); |
| 1460 |
|
| 1461 |
if (info->target < NVISA_GF100_CHIPSET) { |
| 1462 |
int gmemSlot = 0; |
| 1463 |
|
| 1464 |
for (unsigned i = 0; i < nir->info.num_ssbos; i++) { |
| 1465 |
info_out->prop.cp.gmem[gmemSlot++] = {.valid = 1, .image = 0, .slot = i}; |
| 1466 |
assert(gmemSlot < 16); |
| 1467 |
} |
| 1468 |
nir_foreach_image_variable(var, nir) { |
| 1469 |
int image_count = glsl_type_get_image_count(var->type); |
| 1470 |
for (int i = 0; i < image_count; i++) { |
| 1471 |
info_out->prop.cp.gmem[gmemSlot++] = {.valid = 1, .image = 1, .slot = var->data.binding + i}; |
| 1472 |
assert(gmemSlot < 16); |
| 1473 |
} |
| 1474 |
} |
| 1475 |
} |
| 1476 |
|
| 1477 |
break; |
| 1478 |
case Program::TYPE_FRAGMENT: |
| 1479 |
info_out->prop.fp.earlyFragTests = nir->info.fs.early_fragment_tests; |
| 1480 |
prog->persampleInvocation = |
| 1481 |
BITSET_TEST(nir->info.system_values_read, SYSTEM_VALUE_SAMPLE_ID) || |
| 1482 |
BITSET_TEST(nir->info.system_values_read, SYSTEM_VALUE_SAMPLE_POS); |
| 1483 |
info_out->prop.fp.postDepthCoverage = nir->info.fs.post_depth_coverage; |
| 1484 |
info_out->prop.fp.readsSampleLocations = |
| 1485 |
BITSET_TEST(nir->info.system_values_read, SYSTEM_VALUE_SAMPLE_POS); |
| 1486 |
info_out->prop.fp.usesDiscard = nir->info.fs.uses_discard; |
| 1487 |
info_out->prop.fp.usesSampleMaskIn = |
| 1488 |
BITSET_TEST(nir->info.system_values_read, SYSTEM_VALUE_SAMPLE_MASK_IN); |
| 1489 |
break; |
| 1490 |
case Program::TYPE_GEOMETRY: |
| 1491 |
info_out->prop.gp.instanceCount = nir->info.gs.invocations; |
| 1492 |
info_out->prop.gp.maxVertices = nir->info.gs.vertices_out; |
| 1493 |
info_out->prop.gp.outputPrim = nir->info.gs.output_primitive; |
| 1494 |
break; |
| 1495 |
case Program::TYPE_TESSELLATION_CONTROL: |
| 1496 |
case Program::TYPE_TESSELLATION_EVAL: |
| 1497 |
info_out->prop.tp.domain = u_tess_prim_from_shader(nir->info.tess._primitive_mode); |
| 1498 |
info_out->prop.tp.outputPatchSize = nir->info.tess.tcs_vertices_out; |
| 1499 |
info_out->prop.tp.outputPrim = |
| 1500 |
nir->info.tess.point_mode ? MESA_PRIM_POINTS : MESA_PRIM_TRIANGLES; |
| 1501 |
info_out->prop.tp.partitioning = (nir->info.tess.spacing + 1) % 3; |
| 1502 |
info_out->prop.tp.winding = !nir->info.tess.ccw; |
| 1503 |
break; |
| 1504 |
case Program::TYPE_VERTEX: |
| 1505 |
info_out->prop.vp.usesDrawParameters = |
| 1506 |
BITSET_TEST(nir->info.system_values_read, SYSTEM_VALUE_BASE_VERTEX) || |
| 1507 |
BITSET_TEST(nir->info.system_values_read, SYSTEM_VALUE_BASE_INSTANCE) || |
| 1508 |
BITSET_TEST(nir->info.system_values_read, SYSTEM_VALUE_DRAW_ID); |
| 1509 |
break; |
| 1510 |
default: |
| 1511 |
break; |
| 1512 |
} |
| 1513 |
|
| 1514 |
return true; |
| 1515 |
} |
| 1516 |
|
| 1517 |
bool |
| 1518 |
Converter::visit(nir_function *function) |
| 1519 |
{ |
| 1520 |
assert(function->impl); |
| 1521 |
|
| 1522 |
// usually the blocks will set everything up, but main is special |
| 1523 |
BasicBlock *entry = new BasicBlock(prog->main); |
| 1524 |
exit = new BasicBlock(prog->main); |
| 1525 |
blocks[nir_start_block(function->impl)->index] = entry; |
| 1526 |
prog->main->setEntry(entry); |
| 1527 |
prog->main->setExit(exit); |
| 1528 |
|
| 1529 |
setPosition(entry, true); |
| 1530 |
|
| 1531 |
switch (prog->getType()) { |
| 1532 |
case Program::TYPE_TESSELLATION_CONTROL: |
| 1533 |
outBase = mkOp2v( |
| 1534 |
OP_SUB, TYPE_U32, getSSA(), |
| 1535 |
mkOp1v(OP_RDSV, TYPE_U32, getSSA(), mkSysVal(SV_LANEID, 0)), |
| 1536 |
mkOp1v(OP_RDSV, TYPE_U32, getSSA(), mkSysVal(SV_INVOCATION_ID, 0))); |
| 1537 |
break; |
| 1538 |
case Program::TYPE_FRAGMENT: { |
| 1539 |
Symbol *sv = mkSysVal(SV_POSITION, 3); |
| 1540 |
Value *temp = mkOp1v(OP_RDSV, TYPE_F32, getSSA(), sv); |
| 1541 |
fp.position = mkOp1v(OP_RCP, TYPE_F32, temp, temp); |
| 1542 |
break; |
| 1543 |
} |
| 1544 |
default: |
| 1545 |
break; |
| 1546 |
} |
| 1547 |
|
| 1548 |
nir_index_ssa_defs(function->impl); |
| 1549 |
foreach_list_typed(nir_cf_node, node, node, &function->impl->body) { |
| 1550 |
if (!visit(node)) |
| 1551 |
return false; |
| 1552 |
} |
| 1553 |
|
| 1554 |
bb->cfg.attach(&exit->cfg, Graph::Edge::TREE); |
| 1555 |
setPosition(exit, true); |
| 1556 |
|
| 1557 |
// TODO: for non main function this needs to be a OP_RETURN |
| 1558 |
mkOp(OP_EXIT, TYPE_NONE, NULL)->terminator = 1; |
| 1559 |
return true; |
| 1560 |
} |
| 1561 |
|
| 1562 |
bool |
| 1563 |
Converter::visit(nir_cf_node *node) |
| 1564 |
{ |
| 1565 |
switch (node->type) { |
| 1566 |
case nir_cf_node_block: |
| 1567 |
return visit(nir_cf_node_as_block(node)); |
| 1568 |
case nir_cf_node_if: |
| 1569 |
return visit(nir_cf_node_as_if(node)); |
| 1570 |
case nir_cf_node_loop: |
| 1571 |
return visit(nir_cf_node_as_loop(node)); |
| 1572 |
default: |
| 1573 |
ERROR("unknown nir_cf_node type %u\n", node->type); |
| 1574 |
return false; |
| 1575 |
} |
| 1576 |
} |
| 1577 |
|
| 1578 |
bool |
| 1579 |
Converter::visit(nir_block *block) |
| 1580 |
{ |
| 1581 |
if (!block->predecessors.entries && exec_list_is_empty(&block->instr_list)) |
| 1582 |
return true; |
| 1583 |
|
| 1584 |
BasicBlock *bb = convert(block); |
| 1585 |
|
| 1586 |
setPosition(bb, true); |
| 1587 |
nir_foreach_instr(insn, block) { |
| 1588 |
if (!visit(insn)) |
| 1589 |
return false; |
| 1590 |
} |
| 1591 |
return true; |
| 1592 |
} |
| 1593 |
|
| 1594 |
bool |
| 1595 |
Converter::visit(nir_if *nif) |
| 1596 |
{ |
| 1597 |
curIfDepth++; |
| 1598 |
|
| 1599 |
DataType sType = getSType(nif->condition, false, false); |
| 1600 |
Value *src = getSrc(&nif->condition, 0); |
| 1601 |
|
| 1602 |
nir_block *lastThen = nir_if_last_then_block(nif); |
| 1603 |
nir_block *lastElse = nir_if_last_else_block(nif); |
| 1604 |
|
| 1605 |
BasicBlock *headBB = bb; |
| 1606 |
BasicBlock *ifBB = convert(nir_if_first_then_block(nif)); |
| 1607 |
BasicBlock *elseBB = convert(nir_if_first_else_block(nif)); |
| 1608 |
|
| 1609 |
bb->cfg.attach(&ifBB->cfg, Graph::Edge::TREE); |
| 1610 |
bb->cfg.attach(&elseBB->cfg, Graph::Edge::TREE); |
| 1611 |
|
| 1612 |
bool insertJoins = lastThen->successors[0] == lastElse->successors[0]; |
| 1613 |
mkFlow(OP_BRA, elseBB, CC_EQ, src)->setType(sType); |
| 1614 |
|
| 1615 |
foreach_list_typed(nir_cf_node, node, node, &nif->then_list) { |
| 1616 |
if (!visit(node)) |
| 1617 |
return false; |
| 1618 |
} |
| 1619 |
|
| 1620 |
setPosition(convert(lastThen), true); |
| 1621 |
if (!bb->isTerminated()) { |
| 1622 |
BasicBlock *tailBB = convert(lastThen->successors[0]); |
| 1623 |
mkFlow(OP_BRA, tailBB, CC_ALWAYS, NULL); |
| 1624 |
bb->cfg.attach(&tailBB->cfg, Graph::Edge::FORWARD); |
| 1625 |
} else { |
| 1626 |
insertJoins = insertJoins && bb->getExit()->op == OP_BRA; |
| 1627 |
} |
| 1628 |
|
| 1629 |
foreach_list_typed(nir_cf_node, node, node, &nif->else_list) { |
| 1630 |
if (!visit(node)) |
| 1631 |
return false; |
| 1632 |
} |
| 1633 |
|
| 1634 |
setPosition(convert(lastElse), true); |
| 1635 |
if (!bb->isTerminated()) { |
| 1636 |
BasicBlock *tailBB = convert(lastElse->successors[0]); |
| 1637 |
mkFlow(OP_BRA, tailBB, CC_ALWAYS, NULL); |
| 1638 |
bb->cfg.attach(&tailBB->cfg, Graph::Edge::FORWARD); |
| 1639 |
} else { |
| 1640 |
insertJoins = insertJoins && bb->getExit()->op == OP_BRA; |
| 1641 |
} |
| 1642 |
|
| 1643 |
if (curIfDepth > 6) { |
| 1644 |
insertJoins = false; |
| 1645 |
} |
| 1646 |
|
| 1647 |
/* we made sure that all threads would converge at the same block */ |
| 1648 |
if (insertJoins) { |
| 1649 |
BasicBlock *conv = convert(lastThen->successors[0]); |
| 1650 |
setPosition(headBB->getExit(), false); |
| 1651 |
headBB->joinAt = mkFlow(OP_JOINAT, conv, CC_ALWAYS, NULL); |
| 1652 |
setPosition(conv, false); |
| 1653 |
mkFlow(OP_JOIN, NULL, CC_ALWAYS, NULL)->fixed = 1; |
| 1654 |
} |
| 1655 |
|
| 1656 |
curIfDepth--; |
| 1657 |
|
| 1658 |
return true; |
| 1659 |
} |
| 1660 |
|
| 1661 |
// TODO: add convergency |
| 1662 |
bool |
| 1663 |
Converter::visit(nir_loop *loop) |
| 1664 |
{ |
| 1665 |
assert(!nir_loop_has_continue_construct(loop)); |
| 1666 |
curLoopDepth += 1; |
| 1667 |
func->loopNestingBound = std::max(func->loopNestingBound, curLoopDepth); |
| 1668 |
|
| 1669 |
BasicBlock *loopBB = convert(nir_loop_first_block(loop)); |
| 1670 |
BasicBlock *tailBB = convert(nir_cf_node_as_block(nir_cf_node_next(&loop->cf_node))); |
| 1671 |
|
| 1672 |
bb->cfg.attach(&loopBB->cfg, Graph::Edge::TREE); |
| 1673 |
|
| 1674 |
mkFlow(OP_PREBREAK, tailBB, CC_ALWAYS, NULL); |
| 1675 |
setPosition(loopBB, false); |
| 1676 |
mkFlow(OP_PRECONT, loopBB, CC_ALWAYS, NULL); |
| 1677 |
|
| 1678 |
foreach_list_typed(nir_cf_node, node, node, &loop->body) { |
| 1679 |
if (!visit(node)) |
| 1680 |
return false; |
| 1681 |
} |
| 1682 |
|
| 1683 |
if (!bb->isTerminated()) { |
| 1684 |
mkFlow(OP_CONT, loopBB, CC_ALWAYS, NULL); |
| 1685 |
bb->cfg.attach(&loopBB->cfg, Graph::Edge::BACK); |
| 1686 |
} |
| 1687 |
|
| 1688 |
if (tailBB->cfg.incidentCount() == 0) |
| 1689 |
loopBB->cfg.attach(&tailBB->cfg, Graph::Edge::TREE); |
| 1690 |
|
| 1691 |
curLoopDepth -= 1; |
| 1692 |
|
| 1693 |
info_out->loops++; |
| 1694 |
|
| 1695 |
return true; |
| 1696 |
} |
| 1697 |
|
| 1698 |
bool |
| 1699 |
Converter::visit(nir_instr *insn) |
| 1700 |
{ |
| 1701 |
// we need an insertion point for on the fly generated immediate loads |
| 1702 |
immInsertPos = bb->getExit(); |
| 1703 |
switch (insn->type) { |
| 1704 |
case nir_instr_type_alu: |
| 1705 |
return visit(nir_instr_as_alu(insn)); |
| 1706 |
case nir_instr_type_intrinsic: |
| 1707 |
return visit(nir_instr_as_intrinsic(insn)); |
| 1708 |
case nir_instr_type_jump: |
| 1709 |
return visit(nir_instr_as_jump(insn)); |
| 1710 |
case nir_instr_type_load_const: |
| 1711 |
return visit(nir_instr_as_load_const(insn)); |
| 1712 |
case nir_instr_type_undef: |
| 1713 |
return visit(nir_instr_as_undef(insn)); |
| 1714 |
case nir_instr_type_tex: |
| 1715 |
return visit(nir_instr_as_tex(insn)); |
| 1716 |
default: |
| 1717 |
ERROR("unknown nir_instr type %u\n", insn->type); |
| 1718 |
return false; |
| 1719 |
} |
| 1720 |
return true; |
| 1721 |
} |
| 1722 |
|
| 1723 |
SVSemantic |
| 1724 |
Converter::convert(nir_intrinsic_op intr) |
| 1725 |
{ |
| 1726 |
switch (intr) { |
| 1727 |
case nir_intrinsic_load_base_vertex: |
| 1728 |
return SV_BASEVERTEX; |
| 1729 |
case nir_intrinsic_load_base_instance: |
| 1730 |
return SV_BASEINSTANCE; |
| 1731 |
case nir_intrinsic_load_draw_id: |
| 1732 |
return SV_DRAWID; |
| 1733 |
case nir_intrinsic_load_front_face: |
| 1734 |
return SV_FACE; |
| 1735 |
case nir_intrinsic_is_helper_invocation: |
| 1736 |
case nir_intrinsic_load_helper_invocation: |
| 1737 |
return SV_THREAD_KILL; |
| 1738 |
case nir_intrinsic_load_instance_id: |
| 1739 |
return SV_INSTANCE_ID; |
| 1740 |
case nir_intrinsic_load_invocation_id: |
| 1741 |
return SV_INVOCATION_ID; |
| 1742 |
case nir_intrinsic_load_workgroup_size: |
| 1743 |
return SV_NTID; |
| 1744 |
case nir_intrinsic_load_local_invocation_id: |
| 1745 |
return SV_TID; |
| 1746 |
case nir_intrinsic_load_num_workgroups: |
| 1747 |
return SV_NCTAID; |
| 1748 |
case nir_intrinsic_load_patch_vertices_in: |
| 1749 |
return SV_VERTEX_COUNT; |
| 1750 |
case nir_intrinsic_load_primitive_id: |
| 1751 |
return SV_PRIMITIVE_ID; |
| 1752 |
case nir_intrinsic_load_sample_id: |
| 1753 |
return SV_SAMPLE_INDEX; |
| 1754 |
case nir_intrinsic_load_sample_mask_in: |
| 1755 |
return SV_SAMPLE_MASK; |
| 1756 |
case nir_intrinsic_load_sample_pos: |
| 1757 |
return SV_SAMPLE_POS; |
| 1758 |
case nir_intrinsic_load_subgroup_eq_mask: |
| 1759 |
return SV_LANEMASK_EQ; |
| 1760 |
case nir_intrinsic_load_subgroup_ge_mask: |
| 1761 |
return SV_LANEMASK_GE; |
| 1762 |
case nir_intrinsic_load_subgroup_gt_mask: |
| 1763 |
return SV_LANEMASK_GT; |
| 1764 |
case nir_intrinsic_load_subgroup_le_mask: |
| 1765 |
return SV_LANEMASK_LE; |
| 1766 |
case nir_intrinsic_load_subgroup_lt_mask: |
| 1767 |
return SV_LANEMASK_LT; |
| 1768 |
case nir_intrinsic_load_subgroup_invocation: |
| 1769 |
return SV_LANEID; |
| 1770 |
case nir_intrinsic_load_tess_coord: |
| 1771 |
return SV_TESS_COORD; |
| 1772 |
case nir_intrinsic_load_tess_level_inner: |
| 1773 |
return SV_TESS_INNER; |
| 1774 |
case nir_intrinsic_load_tess_level_outer: |
| 1775 |
return SV_TESS_OUTER; |
| 1776 |
case nir_intrinsic_load_vertex_id: |
| 1777 |
return SV_VERTEX_ID; |
| 1778 |
case nir_intrinsic_load_workgroup_id: |
| 1779 |
return SV_CTAID; |
| 1780 |
case nir_intrinsic_load_work_dim: |
| 1781 |
return SV_WORK_DIM; |
| 1782 |
default: |
| 1783 |
ERROR("unknown SVSemantic for nir_intrinsic_op %s\n", |
| 1784 |
nir_intrinsic_infos[intr].name); |
| 1785 |
assert(false); |
| 1786 |
return SV_LAST; |
| 1787 |
} |
| 1788 |
} |
| 1789 |
|
| 1790 |
bool |
| 1791 |
Converter::visit(nir_intrinsic_instr *insn) |
| 1792 |
{ |
| 1793 |
nir_intrinsic_op op = insn->intrinsic; |
| 1794 |
const nir_intrinsic_info &opInfo = nir_intrinsic_infos[op]; |
| 1795 |
unsigned dest_components = nir_intrinsic_dest_components(insn); |
| 1796 |
|
| 1797 |
switch (op) { |
| 1798 |
case nir_intrinsic_decl_reg: { |
| 1799 |
const unsigned reg_index = insn->def.index; |
| 1800 |
const unsigned bit_size = nir_intrinsic_bit_size(insn); |
| 1801 |
const unsigned num_components = nir_intrinsic_num_components(insn); |
| 1802 |
assert(nir_intrinsic_num_array_elems(insn) == 0); |
| 1803 |
|
| 1804 |
LValues newDef(num_components); |
| 1805 |
for (uint8_t c = 0; c < num_components; c++) |
| 1806 |
newDef[c] = getScratch(std::max(4u, bit_size / 8)); |
| 1807 |
|
| 1808 |
assert(regDefs.find(reg_index) == regDefs.end()); |
| 1809 |
regDefs[reg_index] = newDef; |
| 1810 |
break; |
| 1811 |
} |
| 1812 |
|
| 1813 |
case nir_intrinsic_load_reg: { |
| 1814 |
const unsigned reg_index = insn->src[0].ssa->index; |
| 1815 |
NirDefMap::iterator it = regDefs.find(reg_index); |
| 1816 |
assert(it != regDefs.end()); |
| 1817 |
LValues &src = it->second; |
| 1818 |
|
| 1819 |
DataType dType = getDType(insn); |
| 1820 |
LValues &newDefs = convert(&insn->def); |
| 1821 |
for (uint8_t c = 0; c < insn->num_components; c++) |
| 1822 |
mkMov(newDefs[c], src[c], dType); |
| 1823 |
break; |
| 1824 |
} |
| 1825 |
|
| 1826 |
case nir_intrinsic_store_reg: { |
| 1827 |
const unsigned reg_index = insn->src[1].ssa->index; |
| 1828 |
NirDefMap::iterator it = regDefs.find(reg_index); |
| 1829 |
assert(it != regDefs.end()); |
| 1830 |
LValues &dst = it->second; |
| 1831 |
|
| 1832 |
DataType dType = Converter::getSType(insn->src[0], false, false); |
| 1833 |
|
| 1834 |
const nir_component_mask_t write_mask = nir_intrinsic_write_mask(insn); |
| 1835 |
for (uint8_t c = 0u; c < insn->num_components; c++) { |
| 1836 |
if (!((1u << c) & write_mask)) |
| 1837 |
continue; |
| 1838 |
|
| 1839 |
Value *src = getSrc(&insn->src[0], c); |
| 1840 |
mkMov(dst[c], src, dType); |
| 1841 |
} |
| 1842 |
break; |
| 1843 |
} |
| 1844 |
|
| 1845 |
case nir_intrinsic_store_output: |
| 1846 |
case nir_intrinsic_store_per_vertex_output: { |
| 1847 |
Value *indirect; |
| 1848 |
DataType dType = getSType(insn->src[0], false, false); |
| 1849 |
uint32_t idx = getIndirect(insn, op == nir_intrinsic_store_output ? 1 : 2, 0, indirect); |
| 1850 |
|
| 1851 |
for (uint8_t i = 0u; i < nir_intrinsic_src_components(insn, 0); ++i) { |
| 1852 |
if (!((1u << i) & nir_intrinsic_write_mask(insn))) |
| 1853 |
continue; |
| 1854 |
|
| 1855 |
uint8_t offset = 0; |
| 1856 |
Value *src = getSrc(&insn->src[0], i); |
| 1857 |
switch (prog->getType()) { |
| 1858 |
case Program::TYPE_FRAGMENT: { |
| 1859 |
if (info_out->out[idx].sn == TGSI_SEMANTIC_POSITION) { |
| 1860 |
// TGSI uses a different interface than NIR, TGSI stores that |
| 1861 |
// value in the z component, NIR in X |
| 1862 |
offset += 2; |
| 1863 |
src = mkOp1v(OP_SAT, TYPE_F32, getScratch(), src); |
| 1864 |
} |
| 1865 |
break; |
| 1866 |
} |
| 1867 |
default: |
| 1868 |
break; |
| 1869 |
} |
| 1870 |
|
| 1871 |
storeTo(insn, getFile(op), OP_EXPORT, dType, src, idx, i + offset, indirect); |
| 1872 |
} |
| 1873 |
break; |
| 1874 |
} |
| 1875 |
case nir_intrinsic_load_input: |
| 1876 |
case nir_intrinsic_load_interpolated_input: |
| 1877 |
case nir_intrinsic_load_output: { |
| 1878 |
LValues &newDefs = convert(&insn->def); |
| 1879 |
|
| 1880 |
// FBFetch |
| 1881 |
if (prog->getType() == Program::TYPE_FRAGMENT && |
| 1882 |
op == nir_intrinsic_load_output) { |
| 1883 |
std::vector<Value*> defs, srcs; |
| 1884 |
uint8_t mask = 0; |
| 1885 |
|
| 1886 |
srcs.push_back(getSSA()); |
| 1887 |
srcs.push_back(getSSA()); |
| 1888 |
Value *x = mkOp1v(OP_RDSV, TYPE_F32, getSSA(), mkSysVal(SV_POSITION, 0)); |
| 1889 |
Value *y = mkOp1v(OP_RDSV, TYPE_F32, getSSA(), mkSysVal(SV_POSITION, 1)); |
| 1890 |
mkCvt(OP_CVT, TYPE_U32, srcs[0], TYPE_F32, x)->rnd = ROUND_Z; |
| 1891 |
mkCvt(OP_CVT, TYPE_U32, srcs[1], TYPE_F32, y)->rnd = ROUND_Z; |
| 1892 |
|
| 1893 |
srcs.push_back(mkOp1v(OP_RDSV, TYPE_U32, getSSA(), mkSysVal(SV_LAYER, 0))); |
| 1894 |
srcs.push_back(mkOp1v(OP_RDSV, TYPE_U32, getSSA(), mkSysVal(SV_SAMPLE_INDEX, 0))); |
| 1895 |
|
| 1896 |
for (uint8_t i = 0u; i < dest_components; ++i) { |
| 1897 |
defs.push_back(newDefs[i]); |
| 1898 |
mask |= 1 << i; |
| 1899 |
} |
| 1900 |
|
| 1901 |
TexInstruction *texi = mkTex(OP_TXF, TEX_TARGET_2D_MS_ARRAY, 0, 0, defs, srcs); |
| 1902 |
texi->tex.levelZero = true; |
| 1903 |
texi->tex.mask = mask; |
| 1904 |
texi->tex.useOffsets = 0; |
| 1905 |
texi->tex.r = 0xffff; |
| 1906 |
texi->tex.s = 0xffff; |
| 1907 |
|
| 1908 |
info_out->prop.fp.readsFramebuffer = true; |
| 1909 |
break; |
| 1910 |
} |
| 1911 |
|
| 1912 |
const DataType dType = getDType(insn); |
| 1913 |
Value *indirect; |
| 1914 |
bool input = op != nir_intrinsic_load_output; |
| 1915 |
operation nvirOp = OP_LAST; |
| 1916 |
uint32_t mode = 0; |
| 1917 |
|
| 1918 |
uint32_t idx = getIndirect(insn, op == nir_intrinsic_load_interpolated_input ? 1 : 0, 0, indirect); |
| 1919 |
nv50_ir_varying& vary = input ? info_out->in[idx] : info_out->out[idx]; |
| 1920 |
|
| 1921 |
// see load_barycentric_* handling |
| 1922 |
if (prog->getType() == Program::TYPE_FRAGMENT) { |
| 1923 |
if (op == nir_intrinsic_load_interpolated_input) { |
| 1924 |
ImmediateValue immMode; |
| 1925 |
if (getSrc(&insn->src[0], 1)->getUniqueInsn()->src(0).getImmediate(immMode)) |
| 1926 |
mode = immMode.reg.data.u32; |
| 1927 |
} |
| 1928 |
if (mode == NV50_IR_INTERP_DEFAULT) |
| 1929 |
mode |= translateInterpMode(&vary, nvirOp); |
| 1930 |
else { |
| 1931 |
if (vary.linear) { |
| 1932 |
nvirOp = OP_LINTERP; |
| 1933 |
mode |= NV50_IR_INTERP_LINEAR; |
| 1934 |
} else { |
| 1935 |
nvirOp = OP_PINTERP; |
| 1936 |
mode |= NV50_IR_INTERP_PERSPECTIVE; |
| 1937 |
} |
| 1938 |
} |
| 1939 |
} |
| 1940 |
|
| 1941 |
for (uint8_t i = 0u; i < dest_components; ++i) { |
| 1942 |
uint32_t address = getSlotAddress(insn, idx, i); |
| 1943 |
Symbol *sym = mkSymbol(getFile(op), 0, dType, address); |
| 1944 |
if (prog->getType() == Program::TYPE_FRAGMENT) { |
| 1945 |
int s = 1; |
| 1946 |
if (typeSizeof(dType) == 8) { |
| 1947 |
Value *lo = getSSA(); |
| 1948 |
Value *hi = getSSA(); |
| 1949 |
Instruction *interp; |
| 1950 |
|
| 1951 |
interp = mkOp1(nvirOp, TYPE_U32, lo, sym); |
| 1952 |
if (nvirOp == OP_PINTERP) |
| 1953 |
interp->setSrc(s++, fp.position); |
| 1954 |
if (mode & NV50_IR_INTERP_OFFSET) |
| 1955 |
interp->setSrc(s++, getSrc(&insn->src[0], 0)); |
| 1956 |
interp->setInterpolate(mode); |
| 1957 |
interp->setIndirect(0, 0, indirect); |
| 1958 |
|
| 1959 |
Symbol *sym1 = mkSymbol(getFile(op), 0, dType, address + 4); |
| 1960 |
interp = mkOp1(nvirOp, TYPE_U32, hi, sym1); |
| 1961 |
if (nvirOp == OP_PINTERP) |
| 1962 |
interp->setSrc(s++, fp.position); |
| 1963 |
if (mode & NV50_IR_INTERP_OFFSET) |
| 1964 |
interp->setSrc(s++, getSrc(&insn->src[0], 0)); |
| 1965 |
interp->setInterpolate(mode); |
| 1966 |
interp->setIndirect(0, 0, indirect); |
| 1967 |
|
| 1968 |
mkOp2(OP_MERGE, dType, newDefs[i], lo, hi); |
| 1969 |
} else { |
| 1970 |
Instruction *interp = mkOp1(nvirOp, dType, newDefs[i], sym); |
| 1971 |
if (nvirOp == OP_PINTERP) |
| 1972 |
interp->setSrc(s++, fp.position); |
| 1973 |
if (mode & NV50_IR_INTERP_OFFSET) |
| 1974 |
interp->setSrc(s++, getSrc(&insn->src[0], 0)); |
| 1975 |
interp->setInterpolate(mode); |
| 1976 |
interp->setIndirect(0, 0, indirect); |
| 1977 |
} |
| 1978 |
} else { |
| 1979 |
mkLoad(dType, newDefs[i], sym, indirect)->perPatch = vary.patch; |
| 1980 |
} |
| 1981 |
} |
| 1982 |
break; |
| 1983 |
} |
| 1984 |
case nir_intrinsic_load_barycentric_at_offset: |
| 1985 |
case nir_intrinsic_load_barycentric_at_sample: |
| 1986 |
case nir_intrinsic_load_barycentric_centroid: |
| 1987 |
case nir_intrinsic_load_barycentric_pixel: |
| 1988 |
case nir_intrinsic_load_barycentric_sample: { |
| 1989 |
LValues &newDefs = convert(&insn->def); |
| 1990 |
uint32_t mode; |
| 1991 |
|
| 1992 |
if (op == nir_intrinsic_load_barycentric_centroid || |
| 1993 |
op == nir_intrinsic_load_barycentric_sample) { |
| 1994 |
mode = NV50_IR_INTERP_CENTROID; |
| 1995 |
} else if (op == nir_intrinsic_load_barycentric_at_offset) { |
| 1996 |
Value *offs[2]; |
| 1997 |
for (uint8_t c = 0; c < 2; c++) { |
| 1998 |
offs[c] = getScratch(); |
| 1999 |
mkOp2(OP_MIN, TYPE_F32, offs[c], getSrc(&insn->src[0], c), loadImm(NULL, 0.4375f)); |
| 2000 |
mkOp2(OP_MAX, TYPE_F32, offs[c], offs[c], loadImm(NULL, -0.5f)); |
| 2001 |
mkOp2(OP_MUL, TYPE_F32, offs[c], offs[c], loadImm(NULL, 4096.0f)); |
| 2002 |
mkCvt(OP_CVT, TYPE_S32, offs[c], TYPE_F32, offs[c]); |
| 2003 |
} |
| 2004 |
mkOp3v(OP_INSBF, TYPE_U32, newDefs[0], offs[1], mkImm(0x1010), offs[0]); |
| 2005 |
|
| 2006 |
mode = NV50_IR_INTERP_OFFSET; |
| 2007 |
} else if (op == nir_intrinsic_load_barycentric_pixel) { |
| 2008 |
mode = NV50_IR_INTERP_DEFAULT; |
| 2009 |
} else if (op == nir_intrinsic_load_barycentric_at_sample) { |
| 2010 |
info_out->prop.fp.readsSampleLocations = true; |
| 2011 |
Value *sample = getSSA(); |
| 2012 |
mkOp3(OP_SELP, TYPE_U32, sample, mkImm(0), getSrc(&insn->src[0], 0), mkImm(0)) |
| 2013 |
->subOp = 2; |
| 2014 |
mkOp1(OP_PIXLD, TYPE_U32, newDefs[0], sample)->subOp = NV50_IR_SUBOP_PIXLD_OFFSET; |
| 2015 |
mode = NV50_IR_INTERP_OFFSET; |
| 2016 |
} else { |
| 2017 |
UNREACHABLE("all intrinsics already handled above"); |
| 2018 |
} |
| 2019 |
|
| 2020 |
loadImm(newDefs[1], mode); |
| 2021 |
break; |
| 2022 |
} |
| 2023 |
case nir_intrinsic_demote: |
| 2024 |
mkOp(OP_DISCARD, TYPE_NONE, NULL); |
| 2025 |
break; |
| 2026 |
case nir_intrinsic_demote_if: { |
| 2027 |
Value *pred = getSSA(1, FILE_PREDICATE); |
| 2028 |
if (insn->num_components > 1) { |
| 2029 |
ERROR("nir_intrinsic_demote_if only with 1 component supported!\n"); |
| 2030 |
assert(false); |
| 2031 |
return false; |
| 2032 |
} |
| 2033 |
mkCmp(OP_SET, CC_NE, TYPE_U8, pred, TYPE_U32, getSrc(&insn->src[0], 0), zero); |
| 2034 |
mkOp(OP_DISCARD, TYPE_NONE, NULL)->setPredicate(CC_P, pred); |
| 2035 |
break; |
| 2036 |
} |
| 2037 |
case nir_intrinsic_load_base_vertex: |
| 2038 |
case nir_intrinsic_load_base_instance: |
| 2039 |
case nir_intrinsic_load_draw_id: |
| 2040 |
case nir_intrinsic_load_front_face: |
| 2041 |
case nir_intrinsic_is_helper_invocation: |
| 2042 |
case nir_intrinsic_load_helper_invocation: |
| 2043 |
case nir_intrinsic_load_instance_id: |
| 2044 |
case nir_intrinsic_load_invocation_id: |
| 2045 |
case nir_intrinsic_load_workgroup_size: |
| 2046 |
case nir_intrinsic_load_local_invocation_id: |
| 2047 |
case nir_intrinsic_load_num_workgroups: |
| 2048 |
case nir_intrinsic_load_patch_vertices_in: |
| 2049 |
case nir_intrinsic_load_primitive_id: |
| 2050 |
case nir_intrinsic_load_sample_id: |
| 2051 |
case nir_intrinsic_load_sample_mask_in: |
| 2052 |
case nir_intrinsic_load_sample_pos: |
| 2053 |
case nir_intrinsic_load_subgroup_eq_mask: |
| 2054 |
case nir_intrinsic_load_subgroup_ge_mask: |
| 2055 |
case nir_intrinsic_load_subgroup_gt_mask: |
| 2056 |
case nir_intrinsic_load_subgroup_le_mask: |
| 2057 |
case nir_intrinsic_load_subgroup_lt_mask: |
| 2058 |
case nir_intrinsic_load_subgroup_invocation: |
| 2059 |
case nir_intrinsic_load_tess_coord: |
| 2060 |
case nir_intrinsic_load_tess_level_inner: |
| 2061 |
case nir_intrinsic_load_tess_level_outer: |
| 2062 |
case nir_intrinsic_load_vertex_id: |
| 2063 |
case nir_intrinsic_load_workgroup_id: |
| 2064 |
case nir_intrinsic_load_work_dim: { |
| 2065 |
const DataType dType = getDType(insn); |
| 2066 |
SVSemantic sv = convert(op); |
| 2067 |
LValues &newDefs = convert(&insn->def); |
| 2068 |
|
| 2069 |
for (uint8_t i = 0u; i < nir_intrinsic_dest_components(insn); ++i) { |
| 2070 |
Value *def; |
| 2071 |
if (typeSizeof(dType) == 8) |
| 2072 |
def = getSSA(); |
| 2073 |
else |
| 2074 |
def = newDefs[i]; |
| 2075 |
|
| 2076 |
if (sv == SV_TID && info->prop.cp.numThreads[i] == 1) { |
| 2077 |
loadImm(def, 0u); |
| 2078 |
} else { |
| 2079 |
Symbol *sym = mkSysVal(sv, i); |
| 2080 |
Instruction *rdsv = mkOp1(OP_RDSV, TYPE_U32, def, sym); |
| 2081 |
if (sv == SV_TESS_OUTER || sv == SV_TESS_INNER) |
| 2082 |
rdsv->perPatch = 1; |
| 2083 |
} |
| 2084 |
|
| 2085 |
if (typeSizeof(dType) == 8) |
| 2086 |
mkOp2(OP_MERGE, dType, newDefs[i], def, loadImm(getSSA(), 0u)); |
| 2087 |
} |
| 2088 |
break; |
| 2089 |
} |
| 2090 |
// constants |
| 2091 |
case nir_intrinsic_load_subgroup_size: { |
| 2092 |
LValues &newDefs = convert(&insn->def); |
| 2093 |
loadImm(newDefs[0], 32u); |
| 2094 |
break; |
| 2095 |
} |
| 2096 |
case nir_intrinsic_vote_all: |
| 2097 |
case nir_intrinsic_vote_any: |
| 2098 |
case nir_intrinsic_vote_ieq: { |
| 2099 |
LValues &newDefs = convert(&insn->def); |
| 2100 |
Value *pred = getScratch(1, FILE_PREDICATE); |
| 2101 |
mkCmp(OP_SET, CC_NE, TYPE_U32, pred, TYPE_U32, getSrc(&insn->src[0], 0), zero); |
| 2102 |
mkOp1(OP_VOTE, TYPE_U32, pred, pred)->subOp = getSubOp(op); |
| 2103 |
mkCvt(OP_CVT, TYPE_U32, newDefs[0], TYPE_U8, pred); |
| 2104 |
break; |
| 2105 |
} |
| 2106 |
case nir_intrinsic_ballot: { |
| 2107 |
LValues &newDefs = convert(&insn->def); |
| 2108 |
Value *pred = getSSA(1, FILE_PREDICATE); |
| 2109 |
mkCmp(OP_SET, CC_NE, TYPE_U32, pred, TYPE_U32, getSrc(&insn->src[0], 0), zero); |
| 2110 |
mkOp1(OP_VOTE, TYPE_U32, newDefs[0], pred)->subOp = NV50_IR_SUBOP_VOTE_ANY; |
| 2111 |
break; |
| 2112 |
} |
| 2113 |
case nir_intrinsic_read_first_invocation: |
| 2114 |
case nir_intrinsic_read_invocation: { |
| 2115 |
LValues &newDefs = convert(&insn->def); |
| 2116 |
const DataType dType = getDType(insn); |
| 2117 |
Value *tmp = getScratch(); |
| 2118 |
|
| 2119 |
if (op == nir_intrinsic_read_first_invocation) { |
| 2120 |
mkOp1(OP_VOTE, TYPE_U32, tmp, mkImm(1))->subOp = NV50_IR_SUBOP_VOTE_ANY; |
| 2121 |
mkOp1(OP_BREV, TYPE_U32, tmp, tmp); |
| 2122 |
mkOp1(OP_BFIND, TYPE_U32, tmp, tmp)->subOp = NV50_IR_SUBOP_BFIND_SAMT; |
| 2123 |
} else |
| 2124 |
tmp = getSrc(&insn->src[1], 0); |
| 2125 |
|
| 2126 |
for (uint8_t i = 0; i < dest_components; ++i) { |
| 2127 |
mkOp3(OP_SHFL, dType, newDefs[i], getSrc(&insn->src[0], i), tmp, mkImm(0x1f)) |
| 2128 |
->subOp = NV50_IR_SUBOP_SHFL_IDX; |
| 2129 |
} |
| 2130 |
break; |
| 2131 |
} |
| 2132 |
case nir_intrinsic_load_per_vertex_input: { |
| 2133 |
const DataType dType = getDType(insn); |
| 2134 |
LValues &newDefs = convert(&insn->def); |
| 2135 |
Value *indirectVertex; |
| 2136 |
Value *indirectOffset; |
| 2137 |
uint32_t baseVertex = getIndirect(&insn->src[0], 0, indirectVertex); |
| 2138 |
uint32_t idx = getIndirect(insn, 1, 0, indirectOffset); |
| 2139 |
|
| 2140 |
Value *vtxBase = mkOp2v(OP_PFETCH, TYPE_U32, getSSA(4, FILE_ADDRESS), |
| 2141 |
mkImm(baseVertex), indirectVertex); |
| 2142 |
for (uint8_t i = 0u; i < dest_components; ++i) { |
| 2143 |
uint32_t address = getSlotAddress(insn, idx, i); |
| 2144 |
loadFrom(getFile(op), 0, dType, newDefs[i], address, 0, |
| 2145 |
indirectOffset, vtxBase, info_out->in[idx].patch); |
| 2146 |
} |
| 2147 |
break; |
| 2148 |
} |
| 2149 |
case nir_intrinsic_load_per_vertex_output: { |
| 2150 |
const DataType dType = getDType(insn); |
| 2151 |
LValues &newDefs = convert(&insn->def); |
| 2152 |
Value *indirectVertex; |
| 2153 |
Value *indirectOffset; |
| 2154 |
uint32_t baseVertex = getIndirect(&insn->src[0], 0, indirectVertex); |
| 2155 |
uint32_t idx = getIndirect(insn, 1, 0, indirectOffset); |
| 2156 |
Value *vtxBase = NULL; |
| 2157 |
|
| 2158 |
if (indirectVertex) |
| 2159 |
vtxBase = indirectVertex; |
| 2160 |
else |
| 2161 |
vtxBase = loadImm(NULL, baseVertex); |
| 2162 |
|
| 2163 |
vtxBase = mkOp2v(OP_ADD, TYPE_U32, getSSA(4, FILE_ADDRESS), outBase, vtxBase); |
| 2164 |
|
| 2165 |
for (uint8_t i = 0u; i < dest_components; ++i) { |
| 2166 |
uint32_t address = getSlotAddress(insn, idx, i); |
| 2167 |
loadFrom(getFile(op), 0, dType, newDefs[i], address, 0, |
| 2168 |
indirectOffset, vtxBase, info_out->in[idx].patch); |
| 2169 |
} |
| 2170 |
break; |
| 2171 |
} |
| 2172 |
case nir_intrinsic_emit_vertex: { |
| 2173 |
uint32_t idx = nir_intrinsic_stream_id(insn); |
| 2174 |
mkOp1(getOperation(op), TYPE_U32, NULL, mkImm(idx))->fixed = 1; |
| 2175 |
break; |
| 2176 |
} |
| 2177 |
case nir_intrinsic_end_primitive: { |
| 2178 |
uint32_t idx = nir_intrinsic_stream_id(insn); |
| 2179 |
if (idx) |
| 2180 |
break; |
| 2181 |
mkOp1(getOperation(op), TYPE_U32, NULL, mkImm(idx))->fixed = 1; |
| 2182 |
break; |
| 2183 |
} |
| 2184 |
case nir_intrinsic_load_ubo: |
| 2185 |
case nir_intrinsic_ldc_nv: { |
| 2186 |
const DataType dType = getDType(insn); |
| 2187 |
LValues &newDefs = convert(&insn->def); |
| 2188 |
Value *indirectIndex; |
| 2189 |
Value *indirectOffset; |
| 2190 |
uint32_t index = getIndirect(&insn->src[0], 0, indirectIndex); |
| 2191 |
uint32_t offset = getIndirect(&insn->src[1], 0, indirectOffset); |
| 2192 |
if (indirectOffset) |
| 2193 |
indirectOffset = mkOp1v(OP_MOV, TYPE_U32, getSSA(4, FILE_ADDRESS), indirectOffset); |
| 2194 |
|
| 2195 |
for (uint8_t i = 0u; i < dest_components; ++i) { |
| 2196 |
loadFrom(getFile(op), index, dType, newDefs[i], offset, i, |
| 2197 |
indirectOffset, indirectIndex); |
| 2198 |
} |
| 2199 |
break; |
| 2200 |
} |
| 2201 |
case nir_intrinsic_get_ssbo_size: { |
| 2202 |
LValues &newDefs = convert(&insn->def); |
| 2203 |
const DataType dType = getDType(insn); |
| 2204 |
Value *indirectBuffer; |
| 2205 |
uint32_t buffer = getIndirect(&insn->src[0], 0, indirectBuffer); |
| 2206 |
|
| 2207 |
Symbol *sym = mkSymbol(FILE_MEMORY_BUFFER, buffer, dType, 0); |
| 2208 |
mkOp1(OP_BUFQ, dType, newDefs[0], sym)->setIndirect(0, 0, indirectBuffer); |
| 2209 |
break; |
| 2210 |
} |
| 2211 |
case nir_intrinsic_store_ssbo: { |
| 2212 |
Value *indirectBuffer; |
| 2213 |
Value *indirectOffset; |
| 2214 |
uint32_t buffer = getIndirect(&insn->src[1], 0, indirectBuffer); |
| 2215 |
uint32_t offset = getIndirect(&insn->src[2], 0, indirectOffset); |
| 2216 |
|
| 2217 |
CacheMode cache = convert(nir_intrinsic_access(insn)); |
| 2218 |
|
| 2219 |
storeVector(insn, buffer, indirectBuffer, offset, indirectOffset)->cache = cache; |
| 2220 |
|
| 2221 |
info_out->io.globalAccess |= 0x2; |
| 2222 |
break; |
| 2223 |
} |
| 2224 |
case nir_intrinsic_load_ssbo: { |
| 2225 |
Value *indirectBuffer; |
| 2226 |
Value *indirectOffset; |
| 2227 |
uint32_t buffer = getIndirect(&insn->src[0], 0, indirectBuffer); |
| 2228 |
uint32_t offset = getIndirect(&insn->src[1], 0, indirectOffset); |
| 2229 |
|
| 2230 |
CacheMode cache = convert(nir_intrinsic_access(insn)); |
| 2231 |
|
| 2232 |
loadVector(insn, buffer, indirectBuffer, offset, indirectOffset)->cache = cache; |
| 2233 |
|
| 2234 |
info_out->io.globalAccess |= 0x1; |
| 2235 |
break; |
| 2236 |
} |
| 2237 |
case nir_intrinsic_shared_atomic: |
| 2238 |
case nir_intrinsic_shared_atomic_swap: { |
| 2239 |
const DataType dType = getDType(insn); |
| 2240 |
LValues &newDefs = convert(&insn->def); |
| 2241 |
Value *indirectOffset; |
| 2242 |
uint32_t offset = getIndirect(&insn->src[0], 0, indirectOffset); |
| 2243 |
Symbol *sym = mkSymbol(FILE_MEMORY_SHARED, 0, dType, offset); |
| 2244 |
Instruction *atom = mkOp2(OP_ATOM, dType, newDefs[0], sym, getSrc(&insn->src[1], 0)); |
| 2245 |
if (op == nir_intrinsic_shared_atomic_swap) |
| 2246 |
atom->setSrc(2, getSrc(&insn->src[2], 0)); |
| 2247 |
atom->setIndirect(0, 0, indirectOffset); |
| 2248 |
atom->subOp = getAtomicSubOp(nir_intrinsic_atomic_op(insn)); |
| 2249 |
break; |
| 2250 |
} |
| 2251 |
case nir_intrinsic_ssbo_atomic: |
| 2252 |
case nir_intrinsic_ssbo_atomic_swap: { |
| 2253 |
const DataType dType = getDType(insn); |
| 2254 |
LValues &newDefs = convert(&insn->def); |
| 2255 |
Value *indirectBuffer; |
| 2256 |
Value *indirectOffset; |
| 2257 |
uint32_t buffer = getIndirect(&insn->src[0], 0, indirectBuffer); |
| 2258 |
uint32_t offset = getIndirect(&insn->src[1], 0, indirectOffset); |
| 2259 |
|
| 2260 |
Symbol *sym = mkSymbol(FILE_MEMORY_BUFFER, buffer, dType, offset); |
| 2261 |
Instruction *atom = mkOp2(OP_ATOM, dType, newDefs[0], sym, |
| 2262 |
getSrc(&insn->src[2], 0)); |
| 2263 |
if (op == nir_intrinsic_ssbo_atomic_swap) |
| 2264 |
atom->setSrc(2, getSrc(&insn->src[3], 0)); |
| 2265 |
atom->setIndirect(0, 0, indirectOffset); |
| 2266 |
atom->setIndirect(0, 1, indirectBuffer); |
| 2267 |
atom->subOp = getAtomicSubOp(nir_intrinsic_atomic_op(insn)); |
| 2268 |
|
| 2269 |
info_out->io.globalAccess |= 0x2; |
| 2270 |
break; |
| 2271 |
} |
| 2272 |
case nir_intrinsic_global_atomic: |
| 2273 |
case nir_intrinsic_global_atomic_swap: { |
| 2274 |
const DataType dType = getDType(insn); |
| 2275 |
LValues &newDefs = convert(&insn->def); |
| 2276 |
Value *address; |
| 2277 |
uint32_t offset = getIndirect(&insn->src[0], 0, address); |
| 2278 |
|
| 2279 |
Symbol *sym = mkSymbol(FILE_MEMORY_GLOBAL, 0, dType, offset); |
| 2280 |
Instruction *atom = |
| 2281 |
mkOp2(OP_ATOM, dType, newDefs[0], sym, getSrc(&insn->src[1], 0)); |
| 2282 |
if (op == nir_intrinsic_global_atomic_swap) |
| 2283 |
atom->setSrc(2, getSrc(&insn->src[2], 0)); |
| 2284 |
atom->setIndirect(0, 0, address); |
| 2285 |
atom->subOp = getAtomicSubOp(nir_intrinsic_atomic_op(insn)); |
| 2286 |
|
| 2287 |
info_out->io.globalAccess |= 0x2; |
| 2288 |
break; |
| 2289 |
} |
| 2290 |
case nir_intrinsic_bindless_image_atomic: |
| 2291 |
case nir_intrinsic_bindless_image_atomic_swap: |
| 2292 |
case nir_intrinsic_bindless_image_load: |
| 2293 |
case nir_intrinsic_bindless_image_samples: |
| 2294 |
case nir_intrinsic_bindless_image_size: |
| 2295 |
case nir_intrinsic_bindless_image_store: |
| 2296 |
case nir_intrinsic_image_atomic: |
| 2297 |
case nir_intrinsic_image_atomic_swap: |
| 2298 |
case nir_intrinsic_image_load: |
| 2299 |
case nir_intrinsic_image_samples: |
| 2300 |
case nir_intrinsic_image_size: |
| 2301 |
case nir_intrinsic_image_store: { |
| 2302 |
std::vector<Value*> srcs, defs; |
| 2303 |
Value *indirect; |
| 2304 |
DataType ty; |
| 2305 |
int subOp = 0; |
| 2306 |
|
| 2307 |
uint32_t mask = 0; |
| 2308 |
TexInstruction::Target target = |
| 2309 |
convert(nir_intrinsic_image_dim(insn), !!nir_intrinsic_image_array(insn), false); |
| 2310 |
unsigned int argCount = getNIRArgCount(target); |
| 2311 |
uint16_t location = 0; |
| 2312 |
|
| 2313 |
if (opInfo.has_dest) { |
| 2314 |
LValues &newDefs = convert(&insn->def); |
| 2315 |
for (uint8_t i = 0u; i < newDefs.size(); ++i) { |
| 2316 |
defs.push_back(newDefs[i]); |
| 2317 |
mask |= 1 << i; |
| 2318 |
} |
| 2319 |
} |
| 2320 |
|
| 2321 |
int lod_src = -1; |
| 2322 |
bool bindless = false; |
| 2323 |
switch (op) { |
| 2324 |
case nir_intrinsic_bindless_image_atomic: |
| 2325 |
case nir_intrinsic_bindless_image_atomic_swap: |
| 2326 |
ty = getDType(insn); |
| 2327 |
bindless = true; |
| 2328 |
info_out->io.globalAccess |= 0x2; |
| 2329 |
mask = 0x1; |
| 2330 |
subOp = getAtomicSubOp(nir_intrinsic_atomic_op(insn)); |
| 2331 |
break; |
| 2332 |
case nir_intrinsic_image_atomic: |
| 2333 |
case nir_intrinsic_image_atomic_swap: |
| 2334 |
ty = getDType(insn); |
| 2335 |
bindless = false; |
| 2336 |
info_out->io.globalAccess |= 0x2; |
| 2337 |
mask = 0x1; |
| 2338 |
subOp = getAtomicSubOp(nir_intrinsic_atomic_op(insn)); |
| 2339 |
break; |
| 2340 |
case nir_intrinsic_bindless_image_load: |
| 2341 |
case nir_intrinsic_image_load: |
| 2342 |
ty = TYPE_U32; |
| 2343 |
bindless = op == nir_intrinsic_bindless_image_load; |
| 2344 |
info_out->io.globalAccess |= 0x1; |
| 2345 |
lod_src = 4; |
| 2346 |
break; |
| 2347 |
case nir_intrinsic_bindless_image_store: |
| 2348 |
case nir_intrinsic_image_store: |
| 2349 |
ty = TYPE_U32; |
| 2350 |
bindless = op == nir_intrinsic_bindless_image_store; |
| 2351 |
info_out->io.globalAccess |= 0x2; |
| 2352 |
lod_src = 5; |
| 2353 |
mask = 0xf; |
| 2354 |
break; |
| 2355 |
case nir_intrinsic_bindless_image_samples: |
| 2356 |
mask = 0x8; |
| 2357 |
FALLTHROUGH; |
| 2358 |
case nir_intrinsic_image_samples: |
| 2359 |
argCount = 0; /* No coordinates */ |
| 2360 |
ty = TYPE_U32; |
| 2361 |
bindless = op == nir_intrinsic_bindless_image_samples; |
| 2362 |
mask = 0x8; |
| 2363 |
break; |
| 2364 |
case nir_intrinsic_bindless_image_size: |
| 2365 |
case nir_intrinsic_image_size: |
| 2366 |
assert(nir_src_as_uint(insn->src[1]) == 0); |
| 2367 |
argCount = 0; /* No coordinates */ |
| 2368 |
ty = TYPE_U32; |
| 2369 |
bindless = op == nir_intrinsic_bindless_image_size; |
| 2370 |
break; |
| 2371 |
default: |
| 2372 |
UNREACHABLE("unhandled image opcode"); |
| 2373 |
break; |
| 2374 |
} |
| 2375 |
|
| 2376 |
if (bindless) |
| 2377 |
indirect = getSrc(&insn->src[0], 0); |
| 2378 |
else |
| 2379 |
location = getIndirect(&insn->src[0], 0, indirect); |
| 2380 |
|
| 2381 |
/* Pre-GF100, SSBOs and images are in the same HW file, managed by |
| 2382 |
* prop.cp.gmem. images are located after SSBOs. |
| 2383 |
*/ |
| 2384 |
if (info->target < NVISA_GF100_CHIPSET) |
| 2385 |
location += nir->info.num_ssbos; |
| 2386 |
|
| 2387 |
// coords |
| 2388 |
if (opInfo.num_srcs >= 2) |
| 2389 |
for (unsigned int i = 0u; i < argCount; ++i) |
| 2390 |
srcs.push_back(getSrc(&insn->src[1], i)); |
| 2391 |
|
| 2392 |
// the sampler is just another src added after coords |
| 2393 |
if (opInfo.num_srcs >= 3 && target.isMS()) |
| 2394 |
srcs.push_back(getSrc(&insn->src[2], 0)); |
| 2395 |
|
| 2396 |
if (opInfo.num_srcs >= 4 && lod_src != 4) { |
| 2397 |
unsigned components = opInfo.src_components[3] ? opInfo.src_components[3] : insn->num_components; |
| 2398 |
for (uint8_t i = 0u; i < components; ++i) |
| 2399 |
srcs.push_back(getSrc(&insn->src[3], i)); |
| 2400 |
} |
| 2401 |
|
| 2402 |
if (opInfo.num_srcs >= 5 && lod_src != 5) |
| 2403 |
// 1 for aotmic swap |
| 2404 |
for (uint8_t i = 0u; i < opInfo.src_components[4]; ++i) |
| 2405 |
srcs.push_back(getSrc(&insn->src[4], i)); |
| 2406 |
|
| 2407 |
TexInstruction *texi = mkTex(getOperation(op), target.getEnum(), location, 0, defs, srcs); |
| 2408 |
texi->tex.bindless = bindless; |
| 2409 |
texi->tex.format = nv50_ir::TexInstruction::translateImgFormat(nir_intrinsic_format(insn)); |
| 2410 |
texi->tex.mask = mask; |
| 2411 |
texi->cache = convert(nir_intrinsic_access(insn)); |
| 2412 |
texi->setType(ty); |
| 2413 |
texi->subOp = subOp; |
| 2414 |
|
| 2415 |
if (indirect) |
| 2416 |
texi->setIndirectR(indirect); |
| 2417 |
|
| 2418 |
break; |
| 2419 |
} |
| 2420 |
case nir_intrinsic_store_scratch: |
| 2421 |
case nir_intrinsic_store_shared: { |
| 2422 |
Value *indirectOffset; |
| 2423 |
uint32_t offset = getIndirect(&insn->src[1], 0, indirectOffset); |
| 2424 |
if (indirectOffset) |
| 2425 |
indirectOffset = mkOp1v(OP_MOV, TYPE_U32, getSSA(4, FILE_ADDRESS), indirectOffset); |
| 2426 |
|
| 2427 |
storeVector(insn, 0, nullptr, offset, indirectOffset); |
| 2428 |
break; |
| 2429 |
} |
| 2430 |
case nir_intrinsic_load_kernel_input: { |
| 2431 |
const DataType dType = getDType(insn); |
| 2432 |
LValues &newDefs = convert(&insn->def); |
| 2433 |
Value *indirectOffset; |
| 2434 |
uint32_t offset = getIndirect(&insn->src[0], 0, indirectOffset); |
| 2435 |
if (indirectOffset) |
| 2436 |
indirectOffset = mkOp1v(OP_MOV, TYPE_U32, getSSA(4, FILE_ADDRESS), indirectOffset); |
| 2437 |
|
| 2438 |
for (uint8_t i = 0u; i < dest_components; ++i) |
| 2439 |
loadFrom(getFile(op), 0, dType, newDefs[i], offset, i, indirectOffset); |
| 2440 |
|
| 2441 |
break; |
| 2442 |
} |
| 2443 |
case nir_intrinsic_load_scratch: |
| 2444 |
case nir_intrinsic_load_shared: { |
| 2445 |
Value *indirectOffset; |
| 2446 |
uint32_t offset = getIndirect(&insn->src[0], 0, indirectOffset); |
| 2447 |
if (indirectOffset) |
| 2448 |
indirectOffset = mkOp1v(OP_MOV, TYPE_U32, getSSA(4, FILE_ADDRESS), indirectOffset); |
| 2449 |
|
| 2450 |
loadVector(insn, 0, nullptr, offset, indirectOffset); |
| 2451 |
|
| 2452 |
break; |
| 2453 |
} |
| 2454 |
case nir_intrinsic_barrier: { |
| 2455 |
mesa_scope exec_scope = nir_intrinsic_execution_scope(insn); |
| 2456 |
mesa_scope mem_scope = nir_intrinsic_memory_scope(insn); |
| 2457 |
nir_variable_mode modes = nir_intrinsic_memory_modes(insn); |
| 2458 |
nir_variable_mode valid_modes = |
| 2459 |
nir_var_mem_global | nir_var_image | nir_var_mem_ssbo | nir_var_mem_shared; |
| 2460 |
|
| 2461 |
if (mem_scope != SCOPE_NONE && (modes & valid_modes)) { |
| 2462 |
|
| 2463 |
Instruction *bar = mkOp(OP_MEMBAR, TYPE_NONE, NULL); |
| 2464 |
bar->fixed = 1; |
| 2465 |
|
| 2466 |
if (mem_scope >= SCOPE_QUEUE_FAMILY) |
| 2467 |
bar->subOp = NV50_IR_SUBOP_MEMBAR(M, GL); |
| 2468 |
else |
| 2469 |
bar->subOp = NV50_IR_SUBOP_MEMBAR(M, CTA); |
| 2470 |
} |
| 2471 |
|
| 2472 |
if (exec_scope != SCOPE_NONE && |
| 2473 |
!(exec_scope == SCOPE_WORKGROUP && nir->info.stage == MESA_SHADER_TESS_CTRL)) { |
| 2474 |
Instruction *bar = mkOp2(OP_BAR, TYPE_U32, NULL, mkImm(0), mkImm(0)); |
| 2475 |
bar->fixed = 1; |
| 2476 |
bar->subOp = NV50_IR_SUBOP_BAR_SYNC; |
| 2477 |
info_out->numBarriers = 1; |
| 2478 |
} |
| 2479 |
|
| 2480 |
break; |
| 2481 |
} |
| 2482 |
case nir_intrinsic_shader_clock: { |
| 2483 |
const DataType dType = getDType(insn); |
| 2484 |
LValues &newDefs = convert(&insn->def); |
| 2485 |
|
| 2486 |
loadImm(newDefs[0], 0u); |
| 2487 |
mkOp1(OP_RDSV, dType, newDefs[1], mkSysVal(SV_CLOCK, 0))->fixed = 1; |
| 2488 |
break; |
| 2489 |
} |
| 2490 |
case nir_intrinsic_load_global: |
| 2491 |
case nir_intrinsic_load_global_constant: { |
| 2492 |
Value *indirectOffset; |
| 2493 |
uint32_t offset = getIndirect(&insn->src[0], 0, indirectOffset); |
| 2494 |
|
| 2495 |
loadVector(insn, 0, nullptr, offset, indirectOffset); |
| 2496 |
|
| 2497 |
info_out->io.globalAccess |= 0x1; |
| 2498 |
break; |
| 2499 |
} |
| 2500 |
case nir_intrinsic_store_global: { |
| 2501 |
Value *indirectOffset; |
| 2502 |
uint32_t offset = getIndirect(&insn->src[1], 0, indirectOffset); |
| 2503 |
|
| 2504 |
storeVector(insn, 0, nullptr, offset, indirectOffset); |
| 2505 |
|
| 2506 |
info_out->io.globalAccess |= 0x2; |
| 2507 |
break; |
| 2508 |
} |
| 2509 |
case nir_intrinsic_ddx: |
| 2510 |
case nir_intrinsic_ddx_coarse: |
| 2511 |
case nir_intrinsic_ddx_fine: |
| 2512 |
case nir_intrinsic_ddy: |
| 2513 |
case nir_intrinsic_ddy_coarse: |
| 2514 |
case nir_intrinsic_ddy_fine: { |
| 2515 |
assert(insn->def.num_components == 1); |
| 2516 |
const DataType dType = getDType(insn); |
| 2517 |
LValues &newDefs = convert(&insn->def); |
| 2518 |
mkOp1(getOperation(op), dType, newDefs[0], getSrc(&insn->src[0], 0)); |
| 2519 |
break; |
| 2520 |
} |
| 2521 |
default: |
| 2522 |
ERROR("unknown nir_intrinsic_op %s\n", nir_intrinsic_infos[op].name); |
| 2523 |
return false; |
| 2524 |
} |
| 2525 |
|
| 2526 |
return true; |
| 2527 |
} |
| 2528 |
|
| 2529 |
bool |
| 2530 |
Converter::visit(nir_jump_instr *insn) |
| 2531 |
{ |
| 2532 |
switch (insn->type) { |
| 2533 |
case nir_jump_break: |
| 2534 |
case nir_jump_continue: { |
| 2535 |
bool isBreak = insn->type == nir_jump_break; |
| 2536 |
nir_block *block = insn->instr.block; |
| 2537 |
BasicBlock *target = convert(block->successors[0]); |
| 2538 |
mkFlow(isBreak ? OP_BREAK : OP_CONT, target, CC_ALWAYS, NULL); |
| 2539 |
bb->cfg.attach(&target->cfg, isBreak ? Graph::Edge::CROSS : Graph::Edge::BACK); |
| 2540 |
break; |
| 2541 |
} |
| 2542 |
default: |
| 2543 |
ERROR("unknown nir_jump_type %u\n", insn->type); |
| 2544 |
return false; |
| 2545 |
} |
| 2546 |
|
| 2547 |
return true; |
| 2548 |
} |
| 2549 |
|
| 2550 |
Value* |
| 2551 |
Converter::convert(nir_load_const_instr *insn, uint8_t idx) |
| 2552 |
{ |
| 2553 |
Value *val; |
| 2554 |
|
| 2555 |
if (immInsertPos) |
| 2556 |
setPosition(immInsertPos, true); |
| 2557 |
else |
| 2558 |
setPosition(bb, false); |
| 2559 |
|
| 2560 |
switch (insn->def.bit_size) { |
| 2561 |
case 64: |
| 2562 |
val = loadImm(getSSA(8), insn->value[idx].u64); |
| 2563 |
break; |
| 2564 |
case 32: |
| 2565 |
val = loadImm(getSSA(4), insn->value[idx].u32); |
| 2566 |
break; |
| 2567 |
case 16: |
| 2568 |
val = loadImm(getSSA(4), insn->value[idx].u16); |
| 2569 |
break; |
| 2570 |
case 8: |
| 2571 |
val = loadImm(getSSA(4), insn->value[idx].u8); |
| 2572 |
break; |
| 2573 |
default: |
| 2574 |
UNREACHABLE("unhandled bit size!\n"); |
| 2575 |
} |
| 2576 |
setPosition(bb, true); |
| 2577 |
return val; |
| 2578 |
} |
| 2579 |
|
| 2580 |
bool |
| 2581 |
Converter::visit(nir_load_const_instr *insn) |
| 2582 |
{ |
| 2583 |
assert(insn->def.bit_size <= 64); |
| 2584 |
immediates[insn->def.index] = insn; |
| 2585 |
return true; |
| 2586 |
} |
| 2587 |
|
| 2588 |
#define DEFAULT_CHECKS \ |
| 2589 |
if (insn->def.num_components > 1) { \ |
| 2590 |
ERROR("nir_alu_instr only supported with 1 component!\n"); \ |
| 2591 |
return false; \ |
| 2592 |
} |
| 2593 |
bool |
| 2594 |
Converter::visit(nir_alu_instr *insn) |
| 2595 |
{ |
| 2596 |
const nir_op op = insn->op; |
| 2597 |
const nir_op_info &info = nir_op_infos[op]; |
| 2598 |
DataType dType = getDType(insn); |
| 2599 |
const std::vector<DataType> sTypes = getSTypes(insn); |
| 2600 |
|
| 2601 |
Instruction *oldPos = this->bb->getExit(); |
| 2602 |
|
| 2603 |
switch (op) { |
| 2604 |
case nir_op_fabs: |
| 2605 |
case nir_op_iabs: |
| 2606 |
case nir_op_fadd: |
| 2607 |
case nir_op_iadd: |
| 2608 |
case nir_op_iand: |
| 2609 |
case nir_op_fceil: |
| 2610 |
case nir_op_fcos: |
| 2611 |
case nir_op_fdiv: |
| 2612 |
case nir_op_idiv: |
| 2613 |
case nir_op_udiv: |
| 2614 |
case nir_op_fexp2: |
| 2615 |
case nir_op_ffloor: |
| 2616 |
case nir_op_ffma: |
| 2617 |
case nir_op_ffmaz: |
| 2618 |
case nir_op_flog2: |
| 2619 |
case nir_op_fmax: |
| 2620 |
case nir_op_imax: |
| 2621 |
case nir_op_umax: |
| 2622 |
case nir_op_fmin: |
| 2623 |
case nir_op_imin: |
| 2624 |
case nir_op_umin: |
| 2625 |
case nir_op_fmod: |
| 2626 |
case nir_op_imod: |
| 2627 |
case nir_op_umod: |
| 2628 |
case nir_op_fmul: |
| 2629 |
case nir_op_fmulz: |
| 2630 |
case nir_op_amul: |
| 2631 |
case nir_op_imul: |
| 2632 |
case nir_op_imul_high: |
| 2633 |
case nir_op_umul_high: |
| 2634 |
case nir_op_fneg: |
| 2635 |
case nir_op_ineg: |
| 2636 |
case nir_op_inot: |
| 2637 |
case nir_op_ior: |
| 2638 |
case nir_op_pack_64_2x32_split: |
| 2639 |
case nir_op_frcp: |
| 2640 |
case nir_op_frem: |
| 2641 |
case nir_op_irem: |
| 2642 |
case nir_op_frsq: |
| 2643 |
case nir_op_fsat: |
| 2644 |
case nir_op_ishr: |
| 2645 |
case nir_op_ushr: |
| 2646 |
case nir_op_fsin: |
| 2647 |
case nir_op_fsqrt: |
| 2648 |
case nir_op_ftrunc: |
| 2649 |
case nir_op_ishl: |
| 2650 |
case nir_op_ixor: { |
| 2651 |
DEFAULT_CHECKS; |
| 2652 |
LValues &newDefs = convert(&insn->def); |
| 2653 |
operation preOp = preOperationNeeded(op); |
| 2654 |
if (preOp != OP_NOP) { |
| 2655 |
assert(info.num_inputs < 2); |
| 2656 |
Value *tmp = getSSA(typeSizeof(dType)); |
| 2657 |
Instruction *i0 = mkOp(preOp, dType, tmp); |
| 2658 |
Instruction *i1 = mkOp(getOperation(op), dType, newDefs[0]); |
| 2659 |
if (info.num_inputs) { |
| 2660 |
i0->setSrc(0, getSrc(&insn->src[0])); |
| 2661 |
i1->setSrc(0, tmp); |
| 2662 |
} |
| 2663 |
i1->subOp = getSubOp(op); |
| 2664 |
} else { |
| 2665 |
Instruction *i = mkOp(getOperation(op), dType, newDefs[0]); |
| 2666 |
for (unsigned s = 0u; s < info.num_inputs; ++s) { |
| 2667 |
i->setSrc(s, getSrc(&insn->src[s])); |
| 2668 |
|
| 2669 |
switch (op) { |
| 2670 |
case nir_op_fmul: |
| 2671 |
case nir_op_ffma: |
| 2672 |
i->dnz = this->info->io.mul_zero_wins; |
| 2673 |
break; |
| 2674 |
case nir_op_fmulz: |
| 2675 |
case nir_op_ffmaz: |
| 2676 |
i->dnz = true; |
| 2677 |
break; |
| 2678 |
default: |
| 2679 |
break; |
| 2680 |
} |
| 2681 |
} |
| 2682 |
i->subOp = getSubOp(op); |
| 2683 |
} |
| 2684 |
break; |
| 2685 |
} |
| 2686 |
case nir_op_ifind_msb: |
| 2687 |
case nir_op_ufind_msb: { |
| 2688 |
DEFAULT_CHECKS; |
| 2689 |
LValues &newDefs = convert(&insn->def); |
| 2690 |
dType = sTypes[0]; |
| 2691 |
mkOp1(getOperation(op), dType, newDefs[0], getSrc(&insn->src[0])); |
| 2692 |
break; |
| 2693 |
} |
| 2694 |
case nir_op_fround_even: { |
| 2695 |
DEFAULT_CHECKS; |
| 2696 |
LValues &newDefs = convert(&insn->def); |
| 2697 |
mkCvt(OP_CVT, dType, newDefs[0], dType, getSrc(&insn->src[0]))->rnd = ROUND_NI; |
| 2698 |
break; |
| 2699 |
} |
| 2700 |
// convert instructions |
| 2701 |
case nir_op_f2i8: |
| 2702 |
case nir_op_f2u8: |
| 2703 |
case nir_op_i2i8: |
| 2704 |
case nir_op_u2u8: |
| 2705 |
case nir_op_f2i16: |
| 2706 |
case nir_op_f2u16: |
| 2707 |
case nir_op_i2i16: |
| 2708 |
case nir_op_u2u16: |
| 2709 |
case nir_op_f2f32: |
| 2710 |
case nir_op_f2i32: |
| 2711 |
case nir_op_f2u32: |
| 2712 |
case nir_op_i2f32: |
| 2713 |
case nir_op_i2i32: |
| 2714 |
case nir_op_u2f32: |
| 2715 |
case nir_op_u2u32: |
| 2716 |
case nir_op_f2f64: |
| 2717 |
case nir_op_f2i64: |
| 2718 |
case nir_op_f2u64: |
| 2719 |
case nir_op_i2f64: |
| 2720 |
case nir_op_i2i64: |
| 2721 |
case nir_op_u2f64: |
| 2722 |
case nir_op_u2u64: { |
| 2723 |
DEFAULT_CHECKS; |
| 2724 |
LValues &newDefs = convert(&insn->def); |
| 2725 |
DataType stype = sTypes[0]; |
| 2726 |
Instruction *i = mkOp1(getOperation(op), dType, newDefs[0], getSrc(&insn->src[0])); |
| 2727 |
if (::isFloatType(stype) && isIntType(dType)) |
| 2728 |
i->rnd = ROUND_Z; |
| 2729 |
i->sType = stype; |
| 2730 |
break; |
| 2731 |
} |
| 2732 |
// compare instructions |
| 2733 |
case nir_op_ieq8: |
| 2734 |
case nir_op_ige8: |
| 2735 |
case nir_op_uge8: |
| 2736 |
case nir_op_ilt8: |
| 2737 |
case nir_op_ult8: |
| 2738 |
case nir_op_ine8: |
| 2739 |
case nir_op_ieq16: |
| 2740 |
case nir_op_ige16: |
| 2741 |
case nir_op_uge16: |
| 2742 |
case nir_op_ilt16: |
| 2743 |
case nir_op_ult16: |
| 2744 |
case nir_op_ine16: |
| 2745 |
case nir_op_feq32: |
| 2746 |
case nir_op_ieq32: |
| 2747 |
case nir_op_fge32: |
| 2748 |
case nir_op_ige32: |
| 2749 |
case nir_op_uge32: |
| 2750 |
case nir_op_flt32: |
| 2751 |
case nir_op_ilt32: |
| 2752 |
case nir_op_ult32: |
| 2753 |
case nir_op_fneu32: |
| 2754 |
case nir_op_ine32: { |
| 2755 |
DEFAULT_CHECKS; |
| 2756 |
LValues &newDefs = convert(&insn->def); |
| 2757 |
Instruction *i = mkCmp(getOperation(op), |
| 2758 |
getCondCode(op), |
| 2759 |
dType, |
| 2760 |
newDefs[0], |
| 2761 |
dType, |
| 2762 |
getSrc(&insn->src[0]), |
| 2763 |
getSrc(&insn->src[1])); |
| 2764 |
if (info.num_inputs == 3) |
| 2765 |
i->setSrc(2, getSrc(&insn->src[2])); |
| 2766 |
i->sType = sTypes[0]; |
| 2767 |
break; |
| 2768 |
} |
| 2769 |
case nir_op_mov: { |
| 2770 |
LValues &newDefs = convert(&insn->def); |
| 2771 |
for (LValues::size_type c = 0u; c < newDefs.size(); ++c) { |
| 2772 |
mkMov(newDefs[c], getSrc(&insn->src[0], c), dType); |
| 2773 |
} |
| 2774 |
break; |
| 2775 |
} |
| 2776 |
case nir_op_vec2: |
| 2777 |
case nir_op_vec3: |
| 2778 |
case nir_op_vec4: |
| 2779 |
case nir_op_vec8: |
| 2780 |
case nir_op_vec16: { |
| 2781 |
LValues &newDefs = convert(&insn->def); |
| 2782 |
for (LValues::size_type c = 0u; c < newDefs.size(); ++c) { |
| 2783 |
mkMov(newDefs[c], getSrc(&insn->src[c]), dType); |
| 2784 |
} |
| 2785 |
break; |
| 2786 |
} |
| 2787 |
// (un)pack |
| 2788 |
case nir_op_pack_64_2x32: { |
| 2789 |
LValues &newDefs = convert(&insn->def); |
| 2790 |
Instruction *merge = mkOp(OP_MERGE, dType, newDefs[0]); |
| 2791 |
merge->setSrc(0, getSrc(&insn->src[0], 0)); |
| 2792 |
merge->setSrc(1, getSrc(&insn->src[0], 1)); |
| 2793 |
break; |
| 2794 |
} |
| 2795 |
case nir_op_pack_half_2x16_split: { |
| 2796 |
LValues &newDefs = convert(&insn->def); |
| 2797 |
Value *tmpH = getSSA(); |
| 2798 |
Value *tmpL = getSSA(); |
| 2799 |
|
| 2800 |
mkCvt(OP_CVT, TYPE_F16, tmpL, TYPE_F32, getSrc(&insn->src[0])); |
| 2801 |
mkCvt(OP_CVT, TYPE_F16, tmpH, TYPE_F32, getSrc(&insn->src[1])); |
| 2802 |
mkOp3(OP_INSBF, TYPE_U32, newDefs[0], tmpH, mkImm(0x1010), tmpL); |
| 2803 |
break; |
| 2804 |
} |
| 2805 |
case nir_op_unpack_half_2x16_split_x: |
| 2806 |
case nir_op_unpack_half_2x16_split_y: { |
| 2807 |
LValues &newDefs = convert(&insn->def); |
| 2808 |
Instruction *cvt = mkCvt(OP_CVT, TYPE_F32, newDefs[0], TYPE_F16, getSrc(&insn->src[0])); |
| 2809 |
if (op == nir_op_unpack_half_2x16_split_y) |
| 2810 |
cvt->subOp = 1; |
| 2811 |
break; |
| 2812 |
} |
| 2813 |
case nir_op_unpack_64_2x32: { |
| 2814 |
LValues &newDefs = convert(&insn->def); |
| 2815 |
mkOp1(OP_SPLIT, dType, newDefs[0], getSrc(&insn->src[0]))->setDef(1, newDefs[1]); |
| 2816 |
break; |
| 2817 |
} |
| 2818 |
case nir_op_unpack_64_2x32_split_x: { |
| 2819 |
LValues &newDefs = convert(&insn->def); |
| 2820 |
mkOp1(OP_SPLIT, dType, newDefs[0], getSrc(&insn->src[0]))->setDef(1, getSSA()); |
| 2821 |
break; |
| 2822 |
} |
| 2823 |
case nir_op_unpack_64_2x32_split_y: { |
| 2824 |
LValues &newDefs = convert(&insn->def); |
| 2825 |
mkOp1(OP_SPLIT, dType, getSSA(), getSrc(&insn->src[0]))->setDef(1, newDefs[0]); |
| 2826 |
break; |
| 2827 |
} |
| 2828 |
// special instructions |
| 2829 |
case nir_op_fsign: |
| 2830 |
case nir_op_isign: { |
| 2831 |
DEFAULT_CHECKS; |
| 2832 |
DataType iType; |
| 2833 |
if (::isFloatType(dType)) |
| 2834 |
iType = TYPE_F32; |
| 2835 |
else |
| 2836 |
iType = TYPE_S32; |
| 2837 |
|
| 2838 |
LValues &newDefs = convert(&insn->def); |
| 2839 |
LValue *val0 = getScratch(); |
| 2840 |
LValue *val1 = getScratch(); |
| 2841 |
mkCmp(OP_SET, CC_GT, iType, val0, dType, getSrc(&insn->src[0]), zero); |
| 2842 |
mkCmp(OP_SET, CC_LT, iType, val1, dType, getSrc(&insn->src[0]), zero); |
| 2843 |
|
| 2844 |
if (dType == TYPE_F64) { |
| 2845 |
mkOp2(OP_SUB, iType, val0, val0, val1); |
| 2846 |
mkCvt(OP_CVT, TYPE_F64, newDefs[0], iType, val0); |
| 2847 |
} else if (dType == TYPE_S64 || dType == TYPE_U64) { |
| 2848 |
mkOp2(OP_SUB, iType, val0, val1, val0); |
| 2849 |
mkOp2(OP_SHR, iType, val1, val0, loadImm(NULL, 31)); |
| 2850 |
mkOp2(OP_MERGE, dType, newDefs[0], val0, val1); |
| 2851 |
} else if (::isFloatType(dType)) |
| 2852 |
mkOp2(OP_SUB, iType, newDefs[0], val0, val1); |
| 2853 |
else |
| 2854 |
mkOp2(OP_SUB, iType, newDefs[0], val1, val0); |
| 2855 |
break; |
| 2856 |
} |
| 2857 |
case nir_op_fcsel: |
| 2858 |
case nir_op_b32csel: { |
| 2859 |
DEFAULT_CHECKS; |
| 2860 |
LValues &newDefs = convert(&insn->def); |
| 2861 |
mkCmp(OP_SLCT, CC_NE, dType, newDefs[0], sTypes[0], getSrc(&insn->src[1]), getSrc(&insn->src[2]), getSrc(&insn->src[0])); |
| 2862 |
break; |
| 2863 |
} |
| 2864 |
case nir_op_ibitfield_extract: |
| 2865 |
case nir_op_ubitfield_extract: { |
| 2866 |
DEFAULT_CHECKS; |
| 2867 |
Value *tmp = getSSA(); |
| 2868 |
LValues &newDefs = convert(&insn->def); |
| 2869 |
mkOp3(OP_INSBF, dType, tmp, getSrc(&insn->src[2]), loadImm(NULL, 0x808), getSrc(&insn->src[1])); |
| 2870 |
mkOp2(OP_EXTBF, dType, newDefs[0], getSrc(&insn->src[0]), tmp); |
| 2871 |
break; |
| 2872 |
} |
| 2873 |
case nir_op_bfm: { |
| 2874 |
DEFAULT_CHECKS; |
| 2875 |
LValues &newDefs = convert(&insn->def); |
| 2876 |
mkOp2(OP_BMSK, dType, newDefs[0], getSrc(&insn->src[1]), getSrc(&insn->src[0]))->subOp = NV50_IR_SUBOP_BMSK_W; |
| 2877 |
break; |
| 2878 |
} |
| 2879 |
case nir_op_bitfield_insert: { |
| 2880 |
DEFAULT_CHECKS; |
| 2881 |
LValues &newDefs = convert(&insn->def); |
| 2882 |
LValue *temp = getSSA(); |
| 2883 |
mkOp3(OP_INSBF, TYPE_U32, temp, getSrc(&insn->src[3]), mkImm(0x808), getSrc(&insn->src[2])); |
| 2884 |
mkOp3(OP_INSBF, dType, newDefs[0], getSrc(&insn->src[1]), temp, getSrc(&insn->src[0])); |
| 2885 |
break; |
| 2886 |
} |
| 2887 |
case nir_op_bit_count: { |
| 2888 |
DEFAULT_CHECKS; |
| 2889 |
LValues &newDefs = convert(&insn->def); |
| 2890 |
mkOp2(OP_POPCNT, dType, newDefs[0], getSrc(&insn->src[0]), getSrc(&insn->src[0])); |
| 2891 |
break; |
| 2892 |
} |
| 2893 |
case nir_op_bitfield_reverse: { |
| 2894 |
DEFAULT_CHECKS; |
| 2895 |
LValues &newDefs = convert(&insn->def); |
| 2896 |
mkOp1(OP_BREV, TYPE_U32, newDefs[0], getSrc(&insn->src[0])); |
| 2897 |
break; |
| 2898 |
} |
| 2899 |
case nir_op_find_lsb: { |
| 2900 |
DEFAULT_CHECKS; |
| 2901 |
LValues &newDefs = convert(&insn->def); |
| 2902 |
Value *tmp = getSSA(); |
| 2903 |
mkOp1(OP_BREV, TYPE_U32, tmp, getSrc(&insn->src[0])); |
| 2904 |
mkOp1(OP_BFIND, TYPE_U32, newDefs[0], tmp)->subOp = NV50_IR_SUBOP_BFIND_SAMT; |
| 2905 |
break; |
| 2906 |
} |
| 2907 |
case nir_op_extract_u8: { |
| 2908 |
DEFAULT_CHECKS; |
| 2909 |
LValues &newDefs = convert(&insn->def); |
| 2910 |
Value *prmt = getSSA(); |
| 2911 |
mkOp2(OP_OR, TYPE_U32, prmt, getSrc(&insn->src[1]), loadImm(NULL, 0x4440)); |
| 2912 |
mkOp3(OP_PERMT, TYPE_U32, newDefs[0], getSrc(&insn->src[0]), prmt, loadImm(NULL, 0)); |
| 2913 |
break; |
| 2914 |
} |
| 2915 |
case nir_op_extract_i8: { |
| 2916 |
DEFAULT_CHECKS; |
| 2917 |
LValues &newDefs = convert(&insn->def); |
| 2918 |
Value *prmt = getSSA(); |
| 2919 |
mkOp3(OP_MAD, TYPE_U32, prmt, getSrc(&insn->src[1]), loadImm(NULL, 0x1111), loadImm(NULL, 0x8880)); |
| 2920 |
mkOp3(OP_PERMT, TYPE_U32, newDefs[0], getSrc(&insn->src[0]), prmt, loadImm(NULL, 0)); |
| 2921 |
break; |
| 2922 |
} |
| 2923 |
case nir_op_extract_u16: { |
| 2924 |
DEFAULT_CHECKS; |
| 2925 |
LValues &newDefs = convert(&insn->def); |
| 2926 |
Value *prmt = getSSA(); |
| 2927 |
mkOp3(OP_MAD, TYPE_U32, prmt, getSrc(&insn->src[1]), loadImm(NULL, 0x22), loadImm(NULL, 0x4410)); |
| 2928 |
mkOp3(OP_PERMT, TYPE_U32, newDefs[0], getSrc(&insn->src[0]), prmt, loadImm(NULL, 0)); |
| 2929 |
break; |
| 2930 |
} |
| 2931 |
case nir_op_extract_i16: { |
| 2932 |
DEFAULT_CHECKS; |
| 2933 |
LValues &newDefs = convert(&insn->def); |
| 2934 |
Value *prmt = getSSA(); |
| 2935 |
mkOp3(OP_MAD, TYPE_U32, prmt, getSrc(&insn->src[1]), loadImm(NULL, 0x2222), loadImm(NULL, 0x9910)); |
| 2936 |
mkOp3(OP_PERMT, TYPE_U32, newDefs[0], getSrc(&insn->src[0]), prmt, loadImm(NULL, 0)); |
| 2937 |
break; |
| 2938 |
} |
| 2939 |
case nir_op_fquantize2f16: { |
| 2940 |
DEFAULT_CHECKS; |
| 2941 |
LValues &newDefs = convert(&insn->def); |
| 2942 |
Value *tmp = getSSA(); |
| 2943 |
mkCvt(OP_CVT, TYPE_F16, tmp, TYPE_F32, getSrc(&insn->src[0]))->ftz = 1; |
| 2944 |
mkCvt(OP_CVT, TYPE_F32, newDefs[0], TYPE_F16, tmp); |
| 2945 |
break; |
| 2946 |
} |
| 2947 |
case nir_op_urol: { |
| 2948 |
DEFAULT_CHECKS; |
| 2949 |
LValues &newDefs = convert(&insn->def); |
| 2950 |
mkOp3(OP_SHF, TYPE_U32, newDefs[0], getSrc(&insn->src[0]), |
| 2951 |
getSrc(&insn->src[1]), getSrc(&insn->src[0])) |
| 2952 |
->subOp = NV50_IR_SUBOP_SHF_L | |
| 2953 |
NV50_IR_SUBOP_SHF_W | |
| 2954 |
NV50_IR_SUBOP_SHF_HI; |
| 2955 |
break; |
| 2956 |
} |
| 2957 |
case nir_op_uror: { |
| 2958 |
DEFAULT_CHECKS; |
| 2959 |
LValues &newDefs = convert(&insn->def); |
| 2960 |
mkOp3(OP_SHF, TYPE_U32, newDefs[0], getSrc(&insn->src[0]), |
| 2961 |
getSrc(&insn->src[1]), getSrc(&insn->src[0])) |
| 2962 |
->subOp = NV50_IR_SUBOP_SHF_R | |
| 2963 |
NV50_IR_SUBOP_SHF_W | |
| 2964 |
NV50_IR_SUBOP_SHF_LO; |
| 2965 |
break; |
| 2966 |
} |
| 2967 |
// boolean conversions |
| 2968 |
case nir_op_b2f32: { |
| 2969 |
DEFAULT_CHECKS; |
| 2970 |
LValues &newDefs = convert(&insn->def); |
| 2971 |
mkOp2(OP_AND, TYPE_U32, newDefs[0], getSrc(&insn->src[0]), loadImm(NULL, 1.0f)); |
| 2972 |
break; |
| 2973 |
} |
| 2974 |
case nir_op_b2f64: { |
| 2975 |
DEFAULT_CHECKS; |
| 2976 |
LValues &newDefs = convert(&insn->def); |
| 2977 |
Value *tmp = getSSA(4); |
| 2978 |
mkOp2(OP_AND, TYPE_U32, tmp, getSrc(&insn->src[0]), loadImm(NULL, 0x3ff00000)); |
| 2979 |
mkOp2(OP_MERGE, TYPE_U64, newDefs[0], loadImm(NULL, 0), tmp); |
| 2980 |
break; |
| 2981 |
} |
| 2982 |
case nir_op_b2i8: |
| 2983 |
case nir_op_b2i16: |
| 2984 |
case nir_op_b2i32: { |
| 2985 |
DEFAULT_CHECKS; |
| 2986 |
LValues &newDefs = convert(&insn->def); |
| 2987 |
mkOp2(OP_AND, TYPE_U32, newDefs[0], getSrc(&insn->src[0]), loadImm(NULL, 1)); |
| 2988 |
break; |
| 2989 |
} |
| 2990 |
case nir_op_b2i64: { |
| 2991 |
DEFAULT_CHECKS; |
| 2992 |
LValues &newDefs = convert(&insn->def); |
| 2993 |
LValue *def = getScratch(); |
| 2994 |
mkOp2(OP_AND, TYPE_U32, def, getSrc(&insn->src[0]), loadImm(NULL, 1)); |
| 2995 |
mkOp2(OP_MERGE, TYPE_S64, newDefs[0], def, loadImm(NULL, 0)); |
| 2996 |
break; |
| 2997 |
} |
| 2998 |
default: |
| 2999 |
ERROR("unknown nir_op %s\n", info.name); |
| 3000 |
assert(false); |
| 3001 |
return false; |
| 3002 |
} |
| 3003 |
|
| 3004 |
if (!oldPos) { |
| 3005 |
oldPos = this->bb->getEntry(); |
| 3006 |
oldPos->precise = nir_alu_instr_is_exact(insn); |
| 3007 |
} |
| 3008 |
|
| 3009 |
if (unlikely(!oldPos)) |
| 3010 |
return true; |
| 3011 |
|
| 3012 |
while (oldPos->next) { |
| 3013 |
oldPos = oldPos->next; |
| 3014 |
oldPos->precise = nir_alu_instr_is_exact(insn); |
| 3015 |
} |
| 3016 |
|
| 3017 |
return true; |
| 3018 |
} |
| 3019 |
#undef DEFAULT_CHECKS |
| 3020 |
|
| 3021 |
bool |
| 3022 |
Converter::visit(nir_undef_instr *insn) |
| 3023 |
{ |
| 3024 |
LValues &newDefs = convert(&insn->def); |
| 3025 |
for (uint8_t i = 0u; i < insn->def.num_components; ++i) { |
| 3026 |
mkOp(OP_NOP, TYPE_NONE, newDefs[i]); |
| 3027 |
} |
| 3028 |
return true; |
| 3029 |
} |
| 3030 |
|
| 3031 |
#define CASE_SAMPLER(ty) \ |
| 3032 |
case GLSL_SAMPLER_DIM_ ## ty : \ |
| 3033 |
if (isArray && !isShadow) \ |
| 3034 |
return TEX_TARGET_ ## ty ## _ARRAY; \ |
| 3035 |
else if (!isArray && isShadow) \ |
| 3036 |
return TEX_TARGET_## ty ## _SHADOW; \ |
| 3037 |
else if (isArray && isShadow) \ |
| 3038 |
return TEX_TARGET_## ty ## _ARRAY_SHADOW; \ |
| 3039 |
else \ |
| 3040 |
return TEX_TARGET_ ## ty |
| 3041 |
|
| 3042 |
TexTarget |
| 3043 |
Converter::convert(glsl_sampler_dim dim, bool isArray, bool isShadow) |
| 3044 |
{ |
| 3045 |
switch (dim) { |
| 3046 |
CASE_SAMPLER(1D); |
| 3047 |
case GLSL_SAMPLER_DIM_SUBPASS: |
| 3048 |
CASE_SAMPLER(2D); |
| 3049 |
CASE_SAMPLER(CUBE); |
| 3050 |
case GLSL_SAMPLER_DIM_3D: |
| 3051 |
return TEX_TARGET_3D; |
| 3052 |
case GLSL_SAMPLER_DIM_MS: |
| 3053 |
case GLSL_SAMPLER_DIM_SUBPASS_MS: |
| 3054 |
if (isArray) |
| 3055 |
return TEX_TARGET_2D_MS_ARRAY; |
| 3056 |
return TEX_TARGET_2D_MS; |
| 3057 |
case GLSL_SAMPLER_DIM_RECT: |
| 3058 |
if (isShadow) |
| 3059 |
return TEX_TARGET_RECT_SHADOW; |
| 3060 |
return TEX_TARGET_RECT; |
| 3061 |
case GLSL_SAMPLER_DIM_BUF: |
| 3062 |
return TEX_TARGET_BUFFER; |
| 3063 |
case GLSL_SAMPLER_DIM_EXTERNAL: |
| 3064 |
return TEX_TARGET_2D; |
| 3065 |
default: |
| 3066 |
ERROR("unknown glsl_sampler_dim %u\n", dim); |
| 3067 |
assert(false); |
| 3068 |
return TEX_TARGET_COUNT; |
| 3069 |
} |
| 3070 |
} |
| 3071 |
#undef CASE_SAMPLER |
| 3072 |
|
| 3073 |
unsigned int |
| 3074 |
Converter::getNIRArgCount(TexInstruction::Target& target) |
| 3075 |
{ |
| 3076 |
unsigned int result = target.getArgCount(); |
| 3077 |
if (target.isCube() && target.isArray()) |
| 3078 |
result--; |
| 3079 |
if (target.isMS()) |
| 3080 |
result--; |
| 3081 |
return result; |
| 3082 |
} |
| 3083 |
|
| 3084 |
CacheMode |
| 3085 |
Converter::convert(enum gl_access_qualifier access) |
| 3086 |
{ |
| 3087 |
if (access & ACCESS_VOLATILE) |
| 3088 |
return CACHE_CV; |
| 3089 |
if (access & ACCESS_COHERENT) |
| 3090 |
return CACHE_CG; |
| 3091 |
return CACHE_CA; |
| 3092 |
} |
| 3093 |
|
| 3094 |
bool |
| 3095 |
Converter::visit(nir_tex_instr *insn) |
| 3096 |
{ |
| 3097 |
switch (insn->op) { |
| 3098 |
case nir_texop_lod: |
| 3099 |
case nir_texop_query_levels: |
| 3100 |
case nir_texop_tex: |
| 3101 |
case nir_texop_texture_samples: |
| 3102 |
case nir_texop_tg4: |
| 3103 |
case nir_texop_txb: |
| 3104 |
case nir_texop_txd: |
| 3105 |
case nir_texop_txf: |
| 3106 |
case nir_texop_txf_ms: |
| 3107 |
case nir_texop_txl: |
| 3108 |
case nir_texop_txs: { |
| 3109 |
LValues &newDefs = convert(&insn->def); |
| 3110 |
std::vector<Value*> srcs; |
| 3111 |
std::vector<Value*> defs; |
| 3112 |
std::vector<nir_src*> offsets; |
| 3113 |
uint8_t mask = 0; |
| 3114 |
bool lz = false; |
| 3115 |
TexInstruction::Target target = convert(insn->sampler_dim, insn->is_array, insn->is_shadow); |
| 3116 |
operation op = getOperation(insn->op); |
| 3117 |
|
| 3118 |
int r, s; |
| 3119 |
int biasIdx = nir_tex_instr_src_index(insn, nir_tex_src_bias); |
| 3120 |
int compIdx = nir_tex_instr_src_index(insn, nir_tex_src_comparator); |
| 3121 |
int coordsIdx = nir_tex_instr_src_index(insn, nir_tex_src_coord); |
| 3122 |
int ddxIdx = nir_tex_instr_src_index(insn, nir_tex_src_ddx); |
| 3123 |
int ddyIdx = nir_tex_instr_src_index(insn, nir_tex_src_ddy); |
| 3124 |
int msIdx = nir_tex_instr_src_index(insn, nir_tex_src_ms_index); |
| 3125 |
int lodIdx = nir_tex_instr_src_index(insn, nir_tex_src_lod); |
| 3126 |
int offsetIdx = nir_tex_instr_src_index(insn, nir_tex_src_offset); |
| 3127 |
int sampOffIdx = nir_tex_instr_src_index(insn, nir_tex_src_sampler_offset); |
| 3128 |
int texOffIdx = nir_tex_instr_src_index(insn, nir_tex_src_texture_offset); |
| 3129 |
int sampHandleIdx = nir_tex_instr_src_index(insn, nir_tex_src_sampler_handle); |
| 3130 |
int texHandleIdx = nir_tex_instr_src_index(insn, nir_tex_src_texture_handle); |
| 3131 |
|
| 3132 |
bool bindless = sampHandleIdx != -1 || texHandleIdx != -1; |
| 3133 |
assert((sampHandleIdx != -1) == (texHandleIdx != -1)); |
| 3134 |
|
| 3135 |
srcs.resize(insn->coord_components); |
| 3136 |
for (uint8_t i = 0u; i < insn->coord_components; ++i) |
| 3137 |
srcs[i] = getSrc(&insn->src[coordsIdx].src, i); |
| 3138 |
|
| 3139 |
// sometimes we get less args than target.getArgCount, but codegen expects the latter |
| 3140 |
if (insn->coord_components) { |
| 3141 |
uint32_t argCount = target.getArgCount(); |
| 3142 |
|
| 3143 |
if (target.isMS()) |
| 3144 |
argCount -= 1; |
| 3145 |
|
| 3146 |
for (uint32_t i = 0u; i < (argCount - insn->coord_components); ++i) |
| 3147 |
srcs.push_back(getSSA()); |
| 3148 |
} |
| 3149 |
|
| 3150 |
if (biasIdx != -1) |
| 3151 |
srcs.push_back(getSrc(&insn->src[biasIdx].src, 0)); |
| 3152 |
// TXQ requires a lod argument for all queries we care about here. |
| 3153 |
// For other ops on MS textures we skip it. |
| 3154 |
if (lodIdx != -1 && !target.isMS()) |
| 3155 |
srcs.push_back(getSrc(&insn->src[lodIdx].src, 0)); |
| 3156 |
else if (op == OP_TXQ) |
| 3157 |
srcs.push_back(loadImm(NULL, 0)); // TXQ always needs an LOD |
| 3158 |
else if (op == OP_TXF) |
| 3159 |
lz = true; |
| 3160 |
if (msIdx != -1) |
| 3161 |
srcs.push_back(getSrc(&insn->src[msIdx].src, 0)); |
| 3162 |
if (offsetIdx != -1) |
| 3163 |
offsets.push_back(&insn->src[offsetIdx].src); |
| 3164 |
if (compIdx != -1) |
| 3165 |
srcs.push_back(getSrc(&insn->src[compIdx].src, 0)); |
| 3166 |
if (texOffIdx != -1) { |
| 3167 |
srcs.push_back(getSrc(&insn->src[texOffIdx].src, 0)); |
| 3168 |
texOffIdx = srcs.size() - 1; |
| 3169 |
} |
| 3170 |
if (sampOffIdx != -1) { |
| 3171 |
srcs.push_back(getSrc(&insn->src[sampOffIdx].src, 0)); |
| 3172 |
sampOffIdx = srcs.size() - 1; |
| 3173 |
} |
| 3174 |
if (bindless) { |
| 3175 |
// currently we use the lower bits |
| 3176 |
Value *split[2]; |
| 3177 |
Value *handle = getSrc(&insn->src[sampHandleIdx].src, 0); |
| 3178 |
|
| 3179 |
mkSplit(split, 4, handle); |
| 3180 |
|
| 3181 |
srcs.push_back(split[0]); |
| 3182 |
texOffIdx = srcs.size() - 1; |
| 3183 |
} |
| 3184 |
|
| 3185 |
r = bindless ? 0xff : insn->texture_index; |
| 3186 |
s = bindless ? 0x1f : insn->sampler_index; |
| 3187 |
if (op == OP_TXF || op == OP_TXQ) |
| 3188 |
s = 0; |
| 3189 |
|
| 3190 |
defs.resize(newDefs.size()); |
| 3191 |
for (uint8_t d = 0u; d < newDefs.size(); ++d) { |
| 3192 |
defs[d] = newDefs[d]; |
| 3193 |
mask |= 1 << d; |
| 3194 |
} |
| 3195 |
if (target.isMS() || (op == OP_TEX && prog->getType() != Program::TYPE_FRAGMENT)) |
| 3196 |
lz = true; |
| 3197 |
|
| 3198 |
// TODO figure out which instructions still need this. |
| 3199 |
if (srcs.empty()) |
| 3200 |
srcs.push_back(loadImm(NULL, 0)); |
| 3201 |
|
| 3202 |
TexInstruction *texi = mkTex(op, target.getEnum(), r, s, defs, srcs); |
| 3203 |
texi->tex.levelZero = lz; |
| 3204 |
texi->tex.mask = mask; |
| 3205 |
texi->tex.bindless = bindless; |
| 3206 |
|
| 3207 |
if (texOffIdx != -1) |
| 3208 |
texi->tex.rIndirectSrc = texOffIdx; |
| 3209 |
if (sampOffIdx != -1) |
| 3210 |
texi->tex.sIndirectSrc = sampOffIdx; |
| 3211 |
|
| 3212 |
switch (insn->op) { |
| 3213 |
case nir_texop_tg4: |
| 3214 |
if (!target.isShadow()) |
| 3215 |
texi->tex.gatherComp = insn->component; |
| 3216 |
break; |
| 3217 |
case nir_texop_txs: |
| 3218 |
texi->tex.query = TXQ_DIMS; |
| 3219 |
break; |
| 3220 |
case nir_texop_texture_samples: |
| 3221 |
texi->tex.mask = 0x4; |
| 3222 |
texi->tex.query = TXQ_TYPE; |
| 3223 |
break; |
| 3224 |
case nir_texop_query_levels: |
| 3225 |
texi->tex.mask = 0x8; |
| 3226 |
texi->tex.query = TXQ_DIMS; |
| 3227 |
break; |
| 3228 |
// TODO: TXQ_SAMPLE_POSITION needs the sample id instead of the LOD emited further up. |
| 3229 |
default: |
| 3230 |
break; |
| 3231 |
} |
| 3232 |
|
| 3233 |
texi->tex.useOffsets = offsets.size(); |
| 3234 |
if (texi->tex.useOffsets) { |
| 3235 |
for (uint8_t s = 0; s < texi->tex.useOffsets; ++s) { |
| 3236 |
for (uint32_t c = 0u; c < 3; ++c) { |
| 3237 |
uint8_t s2 = std::min(c, target.getDim() - 1); |
| 3238 |
texi->offset[s][c].set(getSrc(offsets[s], s2)); |
| 3239 |
texi->offset[s][c].setInsn(texi); |
| 3240 |
} |
| 3241 |
} |
| 3242 |
} |
| 3243 |
|
| 3244 |
if (op == OP_TXG && offsetIdx == -1) { |
| 3245 |
if (nir_tex_instr_has_explicit_tg4_offsets(insn)) { |
| 3246 |
texi->tex.useOffsets = 4; |
| 3247 |
setPosition(texi, false); |
| 3248 |
for (uint8_t i = 0; i < 4; ++i) { |
| 3249 |
for (uint8_t j = 0; j < 2; ++j) { |
| 3250 |
texi->offset[i][j].set(loadImm(NULL, insn->tg4_offsets[i][j])); |
| 3251 |
texi->offset[i][j].setInsn(texi); |
| 3252 |
} |
| 3253 |
} |
| 3254 |
setPosition(texi, true); |
| 3255 |
} |
| 3256 |
} |
| 3257 |
|
| 3258 |
if (ddxIdx != -1 && ddyIdx != -1) { |
| 3259 |
for (uint8_t c = 0u; c < target.getDim() + target.isCube(); ++c) { |
| 3260 |
texi->dPdx[c].set(getSrc(&insn->src[ddxIdx].src, c)); |
| 3261 |
texi->dPdy[c].set(getSrc(&insn->src[ddyIdx].src, c)); |
| 3262 |
} |
| 3263 |
} |
| 3264 |
|
| 3265 |
break; |
| 3266 |
} |
| 3267 |
default: |
| 3268 |
ERROR("unknown nir_texop %u\n", insn->op); |
| 3269 |
return false; |
| 3270 |
} |
| 3271 |
return true; |
| 3272 |
} |
| 3273 |
|
| 3274 |
/* nouveau's RA doesn't track the liveness of exported registers in the fragment |
| 3275 |
* shader, so we need all the store_outputs to appear at the end of the shader |
| 3276 |
* with no other instructions that might generate a temp value in between them. |
| 3277 |
*/ |
| 3278 |
static bool |
| 3279 |
nv_nir_move_stores_to_end(nir_shader *s) |
| 3280 |
{ |
| 3281 |
nir_function_impl *impl = nir_shader_get_entrypoint(s); |
| 3282 |
nir_block *block = nir_impl_last_block(impl); |
| 3283 |
nir_instr *first_store = NULL; |
| 3284 |
bool progress = false; |
| 3285 |
|
| 3286 |
nir_foreach_instr_safe(instr, block) { |
| 3287 |
if (instr == first_store) |
| 3288 |
break; |
| 3289 |
if (instr->type != nir_instr_type_intrinsic) |
| 3290 |
continue; |
| 3291 |
nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); |
| 3292 |
if (intrin->intrinsic == nir_intrinsic_store_output) { |
| 3293 |
nir_instr_remove(instr); |
| 3294 |
nir_instr_insert(nir_after_block(block), instr); |
| 3295 |
|
| 3296 |
if (!first_store) |
| 3297 |
first_store = instr; |
| 3298 |
|
| 3299 |
progress = true; |
| 3300 |
} |
| 3301 |
} |
| 3302 |
return nir_progress(progress, impl, nir_metadata_control_flow); |
| 3303 |
} |
| 3304 |
|
| 3305 |
unsigned |
| 3306 |
Converter::lowerBitSizeCB(const nir_instr *instr, void *data) |
| 3307 |
{ |
| 3308 |
Converter *instance = static_cast<Converter *>(data); |
| 3309 |
nir_alu_instr *alu; |
| 3310 |
|
| 3311 |
if (instr->type != nir_instr_type_alu) |
| 3312 |
return 0; |
| 3313 |
|
| 3314 |
alu = nir_instr_as_alu(instr); |
| 3315 |
|
| 3316 |
switch (alu->op) { |
| 3317 |
/* TODO: Check for operation OP_SET instead of all listed nir opcodes |
| 3318 |
* individually. |
| 3319 |
* |
| 3320 |
* Currently, we can't call getOperation(nir_op), since not all nir opcodes |
| 3321 |
* are handled within getOperation() and we'd run into an assert(). |
| 3322 |
* |
| 3323 |
* Adding all nir opcodes to getOperation() isn't trivial, since the |
| 3324 |
* enum operation of some of the nir opcodes isn't distinct (e.g. depends |
| 3325 |
* on the data type). |
| 3326 |
*/ |
| 3327 |
case nir_op_ieq8: |
| 3328 |
case nir_op_ige8: |
| 3329 |
case nir_op_uge8: |
| 3330 |
case nir_op_ilt8: |
| 3331 |
case nir_op_ult8: |
| 3332 |
case nir_op_ine8: |
| 3333 |
case nir_op_ieq16: |
| 3334 |
case nir_op_ige16: |
| 3335 |
case nir_op_uge16: |
| 3336 |
case nir_op_ilt16: |
| 3337 |
case nir_op_ult16: |
| 3338 |
case nir_op_ine16: |
| 3339 |
case nir_op_feq32: |
| 3340 |
case nir_op_ieq32: |
| 3341 |
case nir_op_fge32: |
| 3342 |
case nir_op_ige32: |
| 3343 |
case nir_op_uge32: |
| 3344 |
case nir_op_flt32: |
| 3345 |
case nir_op_ilt32: |
| 3346 |
case nir_op_ult32: |
| 3347 |
case nir_op_fneu32: |
| 3348 |
case nir_op_ine32: { |
| 3349 |
DataType stype = instance->getSTypes(alu)[0]; |
| 3350 |
|
| 3351 |
if (isSignedIntType(stype) && typeSizeof(stype) < 4) |
| 3352 |
return 32; |
| 3353 |
|
| 3354 |
return 0; |
| 3355 |
} |
| 3356 |
case nir_op_i2f64: |
| 3357 |
case nir_op_u2f64: { |
| 3358 |
DataType stype = instance->getSTypes(alu)[0]; |
| 3359 |
|
| 3360 |
if (isIntType(stype) && (typeSizeof(stype) <= 2)) |
| 3361 |
return 32; |
| 3362 |
|
| 3363 |
return 0; |
| 3364 |
} |
| 3365 |
default: |
| 3366 |
return 0; |
| 3367 |
} |
| 3368 |
} |
| 3369 |
|
| 3370 |
void |
| 3371 |
Converter::runOptLoop() |
| 3372 |
{ |
| 3373 |
bool progress; |
| 3374 |
do { |
| 3375 |
progress = false; |
| 3376 |
NIR_PASS(progress, nir, nir_opt_copy_prop); |
| 3377 |
NIR_PASS(progress, nir, nir_opt_remove_phis); |
| 3378 |
NIR_PASS(progress, nir, nir_opt_loop); |
| 3379 |
NIR_PASS(progress, nir, nir_opt_cse); |
| 3380 |
NIR_PASS(progress, nir, nir_opt_algebraic); |
| 3381 |
NIR_PASS(progress, nir, nir_opt_constant_folding); |
| 3382 |
NIR_PASS(progress, nir, nir_opt_copy_prop); |
| 3383 |
NIR_PASS(progress, nir, nir_opt_dce); |
| 3384 |
NIR_PASS(progress, nir, nir_opt_dead_cf); |
| 3385 |
NIR_PASS(progress, nir, nir_lower_64bit_phis); |
| 3386 |
} while (progress); |
| 3387 |
} |
| 3388 |
|
| 3389 |
bool |
| 3390 |
Converter::run() |
| 3391 |
{ |
| 3392 |
if (prog->dbgFlags & NV50_IR_DEBUG_VERBOSE) |
| 3393 |
nir_print_shader(nir, stderr); |
| 3394 |
|
| 3395 |
struct nir_lower_subgroups_options subgroup_options = {}; |
| 3396 |
subgroup_options.subgroup_size = 32; |
| 3397 |
subgroup_options.ballot_bit_size = 32; |
| 3398 |
subgroup_options.ballot_components = 1; |
| 3399 |
subgroup_options.lower_elect = true; |
| 3400 |
subgroup_options.lower_inverse_ballot = true; |
| 3401 |
|
| 3402 |
unsigned lower_flrp = (nir->options->lower_flrp16 ? 16 : 0) | |
| 3403 |
(nir->options->lower_flrp32 ? 32 : 0) | |
| 3404 |
(nir->options->lower_flrp64 ? 64 : 0); |
| 3405 |
assert(lower_flrp); |
| 3406 |
|
| 3407 |
info_out->io.genUserClip = info->io.genUserClip; |
| 3408 |
if (info->io.genUserClip > 0) { |
| 3409 |
bool lowered = false; |
| 3410 |
|
| 3411 |
if (nir->info.stage == MESA_SHADER_VERTEX || |
| 3412 |
nir->info.stage == MESA_SHADER_TESS_EVAL) |
| 3413 |
NIR_PASS(lowered, nir, nir_lower_clip_vs, |
| 3414 |
(1 << info->io.genUserClip) - 1, true, false, NULL); |
| 3415 |
else if (nir->info.stage == MESA_SHADER_GEOMETRY) |
| 3416 |
NIR_PASS(lowered, nir, nir_lower_clip_gs, |
| 3417 |
(1 << info->io.genUserClip) - 1, false, NULL); |
| 3418 |
|
| 3419 |
if (lowered) { |
| 3420 |
nir_function_impl *impl = nir_shader_get_entrypoint(nir); |
| 3421 |
NIR_PASS(_, nir, nir_lower_io_vars_to_temporaries, impl, nir_var_shader_out); |
| 3422 |
NIR_PASS(_, nir, nir_lower_global_vars_to_local); |
| 3423 |
NIR_PASS(_, nir, nv50_nir_lower_load_user_clip_plane, info); |
| 3424 |
} else { |
| 3425 |
info_out->io.genUserClip = -1; |
| 3426 |
} |
| 3427 |
} |
| 3428 |
|
| 3429 |
/* prepare for IO lowering */ |
| 3430 |
NIR_PASS(_, nir, nir_lower_flrp, lower_flrp, false); |
| 3431 |
NIR_PASS(_, nir, nir_opt_deref); |
| 3432 |
NIR_PASS(_, nir, nir_lower_vars_to_ssa); |
| 3433 |
|
| 3434 |
NIR_PASS(_, nir, nir_lower_io, nir_var_shader_in | nir_var_shader_out, |
| 3435 |
type_size, nir_lower_io_use_interpolated_input_intrinsics); |
| 3436 |
|
| 3437 |
NIR_PASS(_, nir, nir_lower_subgroups, &subgroup_options); |
| 3438 |
|
| 3439 |
struct nir_lower_tex_options tex_options = {}; |
| 3440 |
tex_options.lower_txp = ~0; |
| 3441 |
|
| 3442 |
NIR_PASS(_, nir, nir_lower_tex, &tex_options); |
| 3443 |
|
| 3444 |
NIR_PASS(_, nir, nir_lower_load_const_to_scalar); |
| 3445 |
NIR_PASS(_, nir, nir_lower_alu_to_scalar, NULL, NULL); |
| 3446 |
NIR_PASS(_, nir, nir_lower_phis_to_scalar, NULL, NULL); |
| 3447 |
|
| 3448 |
NIR_PASS(_, nir, nir_lower_frexp); |
| 3449 |
|
| 3450 |
/*TODO: improve this lowering/optimisation loop so that we can use |
| 3451 |
* nir_opt_idiv_const effectively before this. |
| 3452 |
*/ |
| 3453 |
nir_lower_idiv_options idiv_options = { |
| 3454 |
.allow_fp16 = true, |
| 3455 |
}; |
| 3456 |
NIR_PASS(_, nir, nir_lower_idiv, &idiv_options); |
| 3457 |
|
| 3458 |
runOptLoop(); |
| 3459 |
|
| 3460 |
NIR_PASS(_, nir, nir_remove_dead_variables, nir_var_function_temp, NULL); |
| 3461 |
NIR_PASS(_, nir, nir_lower_vars_to_explicit_types, nir_var_function_temp, |
| 3462 |
glsl_get_natural_size_align_bytes); |
| 3463 |
NIR_PASS(_, nir, nir_lower_explicit_io, nir_var_function_temp, nir_address_format_32bit_offset); |
| 3464 |
NIR_PASS(_, nir, nir_remove_dead_variables, nir_var_function_temp, NULL); |
| 3465 |
|
| 3466 |
NIR_PASS(_, nir, nir_opt_constant_folding); // Improves aliasing information |
| 3467 |
|
| 3468 |
nir_load_store_vectorize_options vectorize_opts = {}; |
| 3469 |
vectorize_opts.modes = nir_var_mem_global | |
| 3470 |
nir_var_mem_ssbo | |
| 3471 |
nir_var_mem_shared | |
| 3472 |
nir_var_shader_temp; |
| 3473 |
vectorize_opts.callback = Converter::memVectorizeCb; |
| 3474 |
vectorize_opts.cb_data = this; |
| 3475 |
NIR_PASS(_, nir, nir_opt_load_store_vectorize, &vectorize_opts); |
| 3476 |
|
| 3477 |
nir_lower_mem_access_bit_sizes_options mem_bit_sizes = {}; |
| 3478 |
mem_bit_sizes.modes = nir_var_mem_global | |
| 3479 |
nir_var_mem_constant | |
| 3480 |
nir_var_mem_ssbo | |
| 3481 |
nir_var_mem_shared | |
| 3482 |
nir_var_shader_temp | nir_var_function_temp; |
| 3483 |
mem_bit_sizes.callback = Converter::getMemAccessSizeAlign; |
| 3484 |
mem_bit_sizes.cb_data = this; |
| 3485 |
NIR_PASS(_, nir, nir_lower_mem_access_bit_sizes, &mem_bit_sizes); |
| 3486 |
|
| 3487 |
NIR_PASS(_, nir, nir_lower_alu_to_scalar, NULL, NULL); |
| 3488 |
NIR_PASS(_, nir, nir_lower_pack); |
| 3489 |
|
| 3490 |
runOptLoop(); |
| 3491 |
|
| 3492 |
NIR_PASS(_, nir, nir_opt_combine_barriers, NULL, NULL); |
| 3493 |
|
| 3494 |
nir_move_options move_options = |
| 3495 |
(nir_move_options)(nir_move_const_undef | |
| 3496 |
nir_move_load_ubo | |
| 3497 |
nir_move_load_uniform | |
| 3498 |
nir_move_load_input | nir_move_load_frag_coord); |
| 3499 |
NIR_PASS(_, nir, nir_opt_sink, move_options); |
| 3500 |
NIR_PASS(_, nir, nir_opt_move, move_options); |
| 3501 |
|
| 3502 |
if (nir->info.stage == MESA_SHADER_FRAGMENT) |
| 3503 |
NIR_PASS(_, nir, nv_nir_move_stores_to_end); |
| 3504 |
|
| 3505 |
NIR_PASS(_, nir, nir_opt_algebraic_late); |
| 3506 |
NIR_PASS(_, nir, nir_opt_dce); |
| 3507 |
|
| 3508 |
NIR_PASS(_, nir, nir_lower_bool_to_int32); |
| 3509 |
NIR_PASS(_, nir, nir_lower_bit_size, Converter::lowerBitSizeCB, this); |
| 3510 |
|
| 3511 |
nir_divergence_analysis(nir); |
| 3512 |
NIR_PASS(_, nir, nir_convert_from_ssa, true, true); |
| 3513 |
|
| 3514 |
// Garbage collect dead instructions |
| 3515 |
nir_sweep(nir); |
| 3516 |
|
| 3517 |
nir_shader_gather_info(nir, nir_shader_get_entrypoint(nir)); |
| 3518 |
|
| 3519 |
if (!parseNIR()) { |
| 3520 |
ERROR("Couldn't prase NIR!\n"); |
| 3521 |
return false; |
| 3522 |
} |
| 3523 |
|
| 3524 |
if (!assignSlots()) { |
| 3525 |
ERROR("Couldn't assign slots!\n"); |
| 3526 |
return false; |
| 3527 |
} |
| 3528 |
|
| 3529 |
if (prog->dbgFlags & NV50_IR_DEBUG_BASIC) |
| 3530 |
nir_print_shader(nir, stderr); |
| 3531 |
|
| 3532 |
nir_foreach_function(function, nir) { |
| 3533 |
if (!visit(function)) |
| 3534 |
return false; |
| 3535 |
} |
| 3536 |
|
| 3537 |
return true; |
| 3538 |
} |
| 3539 |
|
| 3540 |
} // unnamed namespace |
| 3541 |
|
| 3542 |
namespace nv50_ir { |
| 3543 |
|
| 3544 |
bool |
| 3545 |
Program::makeFromNIR(struct nv50_ir_prog_info *info, |
| 3546 |
struct nv50_ir_prog_info_out *info_out) |
| 3547 |
{ |
| 3548 |
nir_shader *nir = info->bin.nir; |
| 3549 |
Converter converter(this, nir, info, info_out); |
| 3550 |
bool result = converter.run(); |
| 3551 |
if (!result) |
| 3552 |
return result; |
| 3553 |
LoweringHelper lowering; |
| 3554 |
lowering.run(this); |
| 3555 |
tlsSize = info_out->bin.tlsSpace; |
| 3556 |
return result; |
| 3557 |
} |
| 3558 |
|
| 3559 |
} // namespace nv50_ir |
| 3560 |
|
| 3561 |
static nir_shader_compiler_options |
| 3562 |
nvir_nir_shader_compiler_options(int chipset, uint8_t shader_type) |
| 3563 |
{ |
| 3564 |
nir_shader_compiler_options op = {}; |
| 3565 |
op.lower_fdiv = (chipset >= NVISA_GV100_CHIPSET); |
| 3566 |
op.lower_ffma16 = false; |
| 3567 |
op.lower_ffma32 = false; |
| 3568 |
op.lower_ffma64 = false; |
| 3569 |
op.fuse_ffma16 = false; /* nir doesn't track mad vs fma */ |
| 3570 |
op.fuse_ffma32 = false; /* nir doesn't track mad vs fma */ |
| 3571 |
op.fuse_ffma64 = false; /* nir doesn't track mad vs fma */ |
| 3572 |
op.lower_flrp16 = (chipset >= NVISA_GV100_CHIPSET); |
| 3573 |
op.lower_flrp32 = true; |
| 3574 |
op.lower_flrp64 = true; |
| 3575 |
op.lower_fpow = true; |
| 3576 |
op.lower_fsat = false; |
| 3577 |
op.lower_fsqrt = false; // TODO: only before gm200 |
| 3578 |
op.lower_sincos = false; |
| 3579 |
op.lower_fmod = true; |
| 3580 |
op.lower_bitfield_extract = (chipset >= NVISA_GV100_CHIPSET || chipset < NVISA_GF100_CHIPSET); |
| 3581 |
op.lower_bitfield_insert = (chipset >= NVISA_GV100_CHIPSET || chipset < NVISA_GF100_CHIPSET); |
| 3582 |
op.lower_bitfield_reverse = (chipset < NVISA_GF100_CHIPSET); |
| 3583 |
op.lower_bit_count = (chipset < NVISA_GF100_CHIPSET); |
| 3584 |
op.lower_ifind_msb = (chipset < NVISA_GF100_CHIPSET); |
| 3585 |
op.lower_find_lsb = (chipset < NVISA_GF100_CHIPSET); |
| 3586 |
op.lower_uadd_carry = true; // TODO |
| 3587 |
op.lower_usub_borrow = true; // TODO |
| 3588 |
op.lower_mul_high = false; |
| 3589 |
op.lower_fneg = false; |
| 3590 |
op.lower_ineg = false; |
| 3591 |
op.lower_scmp = true; // TODO: not implemented yet |
| 3592 |
op.lower_vector_cmp = false; |
| 3593 |
op.lower_bitops = false; |
| 3594 |
op.lower_isign = (chipset >= NVISA_GV100_CHIPSET); |
| 3595 |
op.lower_fsign = (chipset >= NVISA_GV100_CHIPSET); |
| 3596 |
op.lower_fdph = false; |
| 3597 |
op.fdot_replicates = false; // TODO |
| 3598 |
op.lower_ffloor = false; // TODO |
| 3599 |
op.lower_ffract = true; |
| 3600 |
op.lower_fceil = false; // TODO |
| 3601 |
op.lower_ftrunc = false; |
| 3602 |
op.lower_ldexp = true; |
| 3603 |
op.lower_pack_half_2x16 = true; |
| 3604 |
op.lower_pack_unorm_2x16 = true; |
| 3605 |
op.lower_pack_snorm_2x16 = true; |
| 3606 |
op.lower_pack_unorm_4x8 = true; |
| 3607 |
op.lower_pack_snorm_4x8 = true; |
| 3608 |
op.lower_unpack_half_2x16 = true; |
| 3609 |
op.lower_unpack_unorm_2x16 = true; |
| 3610 |
op.lower_unpack_snorm_2x16 = true; |
| 3611 |
op.lower_unpack_unorm_4x8 = true; |
| 3612 |
op.lower_unpack_snorm_4x8 = true; |
| 3613 |
op.lower_pack_split = false; |
| 3614 |
op.lower_extract_byte = (chipset < NVISA_GM107_CHIPSET); |
| 3615 |
op.lower_extract_word = (chipset < NVISA_GM107_CHIPSET); |
| 3616 |
op.lower_insert_byte = true; |
| 3617 |
op.lower_insert_word = true; |
| 3618 |
op.vertex_id_zero_based = false; |
| 3619 |
op.lower_base_vertex = false; |
| 3620 |
op.lower_helper_invocation = false; |
| 3621 |
op.optimize_sample_mask_in = false; |
| 3622 |
op.lower_cs_local_index_to_id = true; |
| 3623 |
op.lower_cs_local_id_to_index = false; |
| 3624 |
op.lower_device_index_to_zero = true; |
| 3625 |
op.lower_wpos_pntc = false; // TODO |
| 3626 |
op.lower_hadd = true; // TODO |
| 3627 |
op.lower_uadd_sat = true; // TODO |
| 3628 |
op.lower_usub_sat = true; // TODO |
| 3629 |
op.lower_iadd_sat = true; // TODO |
| 3630 |
op.lower_to_scalar = false; |
| 3631 |
op.unify_interfaces = false; |
| 3632 |
op.lower_mul_2x32_64 = true; // TODO |
| 3633 |
op.has_rotate32 = (chipset >= NVISA_GV100_CHIPSET); |
| 3634 |
op.has_imul24 = false; |
| 3635 |
op.has_fmulz = (chipset > NVISA_G80_CHIPSET); |
| 3636 |
op.intel_vec4 = false; |
| 3637 |
op.lower_uniforms_to_ubo = true; |
| 3638 |
op.force_indirect_unrolling = (nir_variable_mode) ( |
| 3639 |
((shader_type == MESA_SHADER_FRAGMENT) ? nir_var_shader_out : 0) | |
| 3640 |
/* HW doesn't support indirect addressing of fragment program inputs |
| 3641 |
* on Volta. The binary driver generates a function to handle every |
| 3642 |
* possible indirection, and indirectly calls the function to handle |
| 3643 |
* this instead. |
| 3644 |
*/ |
| 3645 |
((chipset >= NVISA_GV100_CHIPSET && shader_type == MESA_SHADER_FRAGMENT) ? nir_var_shader_in : 0) |
| 3646 |
); |
| 3647 |
op.force_indirect_unrolling_sampler = (chipset < NVISA_GF100_CHIPSET); |
| 3648 |
op.max_unroll_iterations = 32; |
| 3649 |
op.lower_int64_options = (nir_lower_int64_options) ( |
| 3650 |
((chipset >= NVISA_GV100_CHIPSET) ? nir_lower_imul64 : 0) | |
| 3651 |
((chipset >= NVISA_GV100_CHIPSET) ? nir_lower_isign64 : 0) | |
| 3652 |
nir_lower_divmod64 | |
| 3653 |
((chipset >= NVISA_GV100_CHIPSET) ? nir_lower_imul_high64 : 0) | |
| 3654 |
((chipset >= NVISA_GV100_CHIPSET) ? nir_lower_bcsel64 : 0) | |
| 3655 |
((chipset >= NVISA_GV100_CHIPSET) ? nir_lower_icmp64 : 0) | |
| 3656 |
((chipset >= NVISA_GV100_CHIPSET) ? nir_lower_iabs64 : 0) | |
| 3657 |
((chipset >= NVISA_GV100_CHIPSET) ? nir_lower_ineg64 : 0) | |
| 3658 |
((chipset >= NVISA_GV100_CHIPSET) ? nir_lower_logic64 : 0) | |
| 3659 |
((chipset >= NVISA_GV100_CHIPSET) ? nir_lower_minmax64 : 0) | |
| 3660 |
((chipset >= NVISA_GV100_CHIPSET) ? nir_lower_shift64 : 0) | |
| 3661 |
nir_lower_imul_2x32_64 | |
| 3662 |
((chipset >= NVISA_GM107_CHIPSET) ? nir_lower_extract64 : 0) | |
| 3663 |
nir_lower_ufind_msb64 | |
| 3664 |
((chipset >= NVISA_GV100_CHIPSET) ? nir_lower_conv64 : 0) |
| 3665 |
); |
| 3666 |
op.lower_doubles_options = (nir_lower_doubles_options) ( |
| 3667 |
((chipset >= NVISA_GV100_CHIPSET) ? nir_lower_drcp : 0) | |
| 3668 |
((chipset >= NVISA_GV100_CHIPSET) ? nir_lower_dsqrt : 0) | |
| 3669 |
((chipset >= NVISA_GV100_CHIPSET) ? nir_lower_drsq : 0) | |
| 3670 |
((chipset >= NVISA_GV100_CHIPSET) ? nir_lower_dfract : 0) | |
| 3671 |
nir_lower_dmod | |
| 3672 |
((chipset >= NVISA_GV100_CHIPSET) ? nir_lower_dsub : 0) | |
| 3673 |
((chipset >= NVISA_GV100_CHIPSET) ? nir_lower_ddiv : 0) |
| 3674 |
); |
| 3675 |
op.discard_is_demote = true; |
| 3676 |
op.scalarize_ddx = true; |
| 3677 |
op.support_indirect_inputs = (uint8_t)BITFIELD_MASK(MESA_SHADER_GEOMETRY + 1); |
| 3678 |
op.support_indirect_outputs = (uint8_t)BITFIELD_MASK(MESA_SHADER_GEOMETRY + 1); |
| 3679 |
|
| 3680 |
/* HW doesn't support indirect addressing of fragment program inputs |
| 3681 |
* on Volta. The binary driver generates a function to handle every |
| 3682 |
* possible indirection, and indirectly calls the function to handle |
| 3683 |
* this instead. |
| 3684 |
*/ |
| 3685 |
if (chipset < NVISA_GV100_CHIPSET) |
| 3686 |
op.support_indirect_outputs |= BITFIELD_BIT(MESA_SHADER_FRAGMENT); |
| 3687 |
|
| 3688 |
return op; |
| 3689 |
} |
| 3690 |
|
| 3691 |
static const nir_shader_compiler_options g80_nir_shader_compiler_options = |
| 3692 |
nvir_nir_shader_compiler_options(NVISA_G80_CHIPSET, MESA_SHADER_STAGES); |
| 3693 |
static const nir_shader_compiler_options g80_fs_nir_shader_compiler_options = |
| 3694 |
nvir_nir_shader_compiler_options(NVISA_G80_CHIPSET, MESA_SHADER_FRAGMENT); |
| 3695 |
static const nir_shader_compiler_options gf100_nir_shader_compiler_options = |
| 3696 |
nvir_nir_shader_compiler_options(NVISA_GF100_CHIPSET, MESA_SHADER_STAGES); |
| 3697 |
static const nir_shader_compiler_options gf100_fs_nir_shader_compiler_options = |
| 3698 |
nvir_nir_shader_compiler_options(NVISA_GF100_CHIPSET, MESA_SHADER_FRAGMENT); |
| 3699 |
static const nir_shader_compiler_options gm107_nir_shader_compiler_options = |
| 3700 |
nvir_nir_shader_compiler_options(NVISA_GM107_CHIPSET, MESA_SHADER_STAGES); |
| 3701 |
static const nir_shader_compiler_options gm107_fs_nir_shader_compiler_options = |
| 3702 |
nvir_nir_shader_compiler_options(NVISA_GM107_CHIPSET, MESA_SHADER_FRAGMENT); |
| 3703 |
static const nir_shader_compiler_options gv100_nir_shader_compiler_options = |
| 3704 |
nvir_nir_shader_compiler_options(NVISA_GV100_CHIPSET, MESA_SHADER_STAGES); |
| 3705 |
static const nir_shader_compiler_options gv100_fs_nir_shader_compiler_options = |
| 3706 |
nvir_nir_shader_compiler_options(NVISA_GV100_CHIPSET, MESA_SHADER_FRAGMENT); |
| 3707 |
|
| 3708 |
const nir_shader_compiler_options * |
| 3709 |
nv50_ir_nir_shader_compiler_options(int chipset, uint8_t shader_type) |
| 3710 |
{ |
| 3711 |
if (chipset >= NVISA_GV100_CHIPSET) { |
| 3712 |
if (shader_type == MESA_SHADER_FRAGMENT) { |
| 3713 |
return &gv100_fs_nir_shader_compiler_options; |
| 3714 |
} else { |
| 3715 |
return &gv100_nir_shader_compiler_options; |
| 3716 |
} |
| 3717 |
} |
| 3718 |
|
| 3719 |
if (chipset >= NVISA_GM107_CHIPSET) { |
| 3720 |
if (shader_type == MESA_SHADER_FRAGMENT) { |
| 3721 |
return &gm107_fs_nir_shader_compiler_options; |
| 3722 |
} else { |
| 3723 |
return &gm107_nir_shader_compiler_options; |
| 3724 |
} |
| 3725 |
} |
| 3726 |
|
| 3727 |
if (chipset >= NVISA_GF100_CHIPSET) { |
| 3728 |
if (shader_type == MESA_SHADER_FRAGMENT) { |
| 3729 |
return &gf100_fs_nir_shader_compiler_options; |
| 3730 |
} else { |
| 3731 |
return &gf100_nir_shader_compiler_options; |
| 3732 |
} |
| 3733 |
} |
| 3734 |
|
| 3735 |
if (shader_type == MESA_SHADER_FRAGMENT) { |
| 3736 |
return &g80_fs_nir_shader_compiler_options; |
| 3737 |
} else { |
| 3738 |
return &g80_nir_shader_compiler_options; |
| 3739 |
} |
| 3740 |
} |