| 1 |
/* |
| 2 |
* Copyright © 2018 Jonathan Marek <[email protected]> |
| 3 |
* SPDX-License-Identifier: MIT |
| 4 |
* |
| 5 |
* Authors: |
| 6 |
* Jonathan Marek <[email protected]> |
| 7 |
*/ |
| 8 |
|
| 9 |
#include <assert.h> |
| 10 |
#include <stdint.h> |
| 11 |
#include <stdio.h> |
| 12 |
#include <stdlib.h> |
| 13 |
#include <string.h> |
| 14 |
|
| 15 |
#include "ir2/instr-a2xx.h" |
| 16 |
#include "fd2_program.h" |
| 17 |
#include "ir2.h" |
| 18 |
|
| 19 |
enum ir2_src_type { |
| 20 |
IR2_SRC_SSA, |
| 21 |
IR2_SRC_REG, |
| 22 |
IR2_SRC_INPUT, |
| 23 |
IR2_SRC_CONST, |
| 24 |
}; |
| 25 |
|
| 26 |
struct ir2_src { |
| 27 |
/* num can mean different things |
| 28 |
* ssa: index of instruction |
| 29 |
* reg: index in ctx->reg array |
| 30 |
* input: index in ctx->input array |
| 31 |
* const: constant index (C0, C1, etc) |
| 32 |
*/ |
| 33 |
uint16_t num; |
| 34 |
uint8_t swizzle; |
| 35 |
enum ir2_src_type type : 2; |
| 36 |
uint8_t abs : 1; |
| 37 |
uint8_t negate : 1; |
| 38 |
uint8_t : 4; |
| 39 |
}; |
| 40 |
|
| 41 |
struct ir2_reg_component { |
| 42 |
uint8_t c : 3; /* assigned x/y/z/w (7=dont write, for fetch instr) */ |
| 43 |
bool alloc : 1; /* is it currently allocated */ |
| 44 |
uint8_t ref_count; /* for ra */ |
| 45 |
}; |
| 46 |
|
| 47 |
struct ir2_reg { |
| 48 |
uint8_t idx; /* assigned hardware register */ |
| 49 |
uint8_t ncomp; |
| 50 |
|
| 51 |
uint8_t loop_depth; |
| 52 |
bool initialized; |
| 53 |
/* block_idx to free on (-1 = free on ref_count==0) */ |
| 54 |
int block_idx_free; |
| 55 |
struct ir2_reg_component comp[4]; |
| 56 |
}; |
| 57 |
|
| 58 |
struct ir2_instr { |
| 59 |
unsigned idx; |
| 60 |
|
| 61 |
unsigned block_idx; |
| 62 |
|
| 63 |
enum { |
| 64 |
IR2_NONE, |
| 65 |
IR2_FETCH, |
| 66 |
IR2_ALU, |
| 67 |
IR2_CF, |
| 68 |
} type : 2; |
| 69 |
|
| 70 |
/* instruction needs to be emitted (for scheduling) */ |
| 71 |
bool need_emit : 1; |
| 72 |
|
| 73 |
/* predicate value - (usually) same for entire block */ |
| 74 |
uint8_t pred : 2; |
| 75 |
|
| 76 |
/* src */ |
| 77 |
uint8_t src_count; |
| 78 |
struct ir2_src src[4]; |
| 79 |
|
| 80 |
/* dst */ |
| 81 |
bool is_ssa; |
| 82 |
union { |
| 83 |
struct ir2_reg ssa; |
| 84 |
struct ir2_reg *reg; |
| 85 |
}; |
| 86 |
|
| 87 |
/* type-specific */ |
| 88 |
union { |
| 89 |
struct { |
| 90 |
instr_fetch_opc_t opc : 5; |
| 91 |
union { |
| 92 |
struct { |
| 93 |
uint8_t const_idx; |
| 94 |
uint8_t const_idx_sel; |
| 95 |
} vtx; |
| 96 |
struct { |
| 97 |
bool is_cube : 1; |
| 98 |
bool is_rect : 1; |
| 99 |
uint8_t samp_id; |
| 100 |
} tex; |
| 101 |
}; |
| 102 |
} fetch; |
| 103 |
struct { |
| 104 |
/* store possible opcs, then we can choose vector/scalar instr */ |
| 105 |
instr_scalar_opc_t scalar_opc : 6; |
| 106 |
instr_vector_opc_t vector_opc : 5; |
| 107 |
/* same as nir */ |
| 108 |
uint8_t write_mask : 4; |
| 109 |
bool saturate : 1; |
| 110 |
|
| 111 |
/* export idx (-1 no export) */ |
| 112 |
int8_t export; |
| 113 |
|
| 114 |
/* for scalarized 2 src instruction */ |
| 115 |
uint8_t src1_swizzle; |
| 116 |
} alu; |
| 117 |
struct { |
| 118 |
/* jmp dst block_idx */ |
| 119 |
uint8_t block_idx; |
| 120 |
} cf; |
| 121 |
}; |
| 122 |
}; |
| 123 |
|
| 124 |
struct ir2_sched_instr { |
| 125 |
uint32_t reg_state[8]; |
| 126 |
struct ir2_instr *instr, *instr_s; |
| 127 |
}; |
| 128 |
|
| 129 |
struct ir2_context { |
| 130 |
struct fd2_shader_stateobj *so; |
| 131 |
|
| 132 |
unsigned block_idx, pred_idx; |
| 133 |
uint8_t pred; |
| 134 |
bool block_has_jump[64]; |
| 135 |
|
| 136 |
unsigned loop_last_block[64]; |
| 137 |
unsigned loop_depth; |
| 138 |
|
| 139 |
nir_shader *nir; |
| 140 |
|
| 141 |
/* ssa index of position output */ |
| 142 |
struct ir2_src position; |
| 143 |
|
| 144 |
/* to translate SSA ids to instruction ids */ |
| 145 |
int16_t ssa_map[1024]; |
| 146 |
|
| 147 |
struct ir2_shader_info *info; |
| 148 |
struct ir2_frag_linkage *f; |
| 149 |
|
| 150 |
int prev_export; |
| 151 |
|
| 152 |
/* RA state */ |
| 153 |
struct ir2_reg *live_regs[64]; |
| 154 |
uint32_t reg_state[256 / 32]; /* 64*4 bits */ |
| 155 |
|
| 156 |
/* inputs */ |
| 157 |
struct ir2_reg input[16 + 1]; /* 16 + param */ |
| 158 |
|
| 159 |
/* non-ssa regs */ |
| 160 |
struct ir2_reg reg[1024]; |
| 161 |
unsigned reg_count; |
| 162 |
|
| 163 |
struct ir2_instr instr[0x300]; |
| 164 |
unsigned instr_count; |
| 165 |
|
| 166 |
struct ir2_sched_instr instr_sched[0x180]; |
| 167 |
unsigned instr_sched_count; |
| 168 |
}; |
| 169 |
|
| 170 |
void assemble(struct ir2_context *ctx, bool binning); |
| 171 |
|
| 172 |
void ir2_nir_compile(struct ir2_context *ctx, bool binning); |
| 173 |
bool ir2_nir_lower_scalar(nir_shader *shader); |
| 174 |
|
| 175 |
void ra_count_refs(struct ir2_context *ctx); |
| 176 |
void ra_reg(struct ir2_context *ctx, struct ir2_reg *reg, int force_idx, |
| 177 |
bool export, uint8_t export_writemask); |
| 178 |
void ra_src_free(struct ir2_context *ctx, struct ir2_instr *instr); |
| 179 |
void ra_block_free(struct ir2_context *ctx, unsigned block); |
| 180 |
|
| 181 |
void cp_src(struct ir2_context *ctx); |
| 182 |
void cp_export(struct ir2_context *ctx); |
| 183 |
|
| 184 |
/* utils */ |
| 185 |
enum { |
| 186 |
IR2_SWIZZLE_Y = 1 << 0, |
| 187 |
IR2_SWIZZLE_Z = 2 << 0, |
| 188 |
IR2_SWIZZLE_W = 3 << 0, |
| 189 |
|
| 190 |
IR2_SWIZZLE_ZW = 2 << 0 | 2 << 2, |
| 191 |
|
| 192 |
IR2_SWIZZLE_YXW = 1 << 0 | 3 << 2 | 1 << 4, |
| 193 |
|
| 194 |
IR2_SWIZZLE_XXXX = 0 << 0 | 3 << 2 | 2 << 4 | 1 << 6, |
| 195 |
IR2_SWIZZLE_YYYY = 1 << 0 | 0 << 2 | 3 << 4 | 2 << 6, |
| 196 |
IR2_SWIZZLE_ZZZZ = 2 << 0 | 1 << 2 | 0 << 4 | 3 << 6, |
| 197 |
IR2_SWIZZLE_WWWW = 3 << 0 | 2 << 2 | 1 << 4 | 0 << 6, |
| 198 |
IR2_SWIZZLE_WYWW = 3 << 0 | 0 << 2 | 1 << 4 | 0 << 6, |
| 199 |
IR2_SWIZZLE_XYXY = 0 << 0 | 0 << 2 | 2 << 4 | 2 << 6, |
| 200 |
IR2_SWIZZLE_ZZXY = 2 << 0 | 1 << 2 | 2 << 4 | 2 << 6, |
| 201 |
IR2_SWIZZLE_YXZZ = 1 << 0 | 3 << 2 | 0 << 4 | 3 << 6, |
| 202 |
}; |
| 203 |
|
| 204 |
#define compile_error(ctx, args...) \ |
| 205 |
({ \ |
| 206 |
printf(args); \ |
| 207 |
assert(0); \ |
| 208 |
}) |
| 209 |
|
| 210 |
static inline struct ir2_src |
| 211 |
ir2_src(uint16_t num, uint8_t swizzle, enum ir2_src_type type) |
| 212 |
{ |
| 213 |
return (struct ir2_src){.num = num, .swizzle = swizzle, .type = type}; |
| 214 |
} |
| 215 |
|
| 216 |
/* ir2_assemble uses it .. */ |
| 217 |
struct ir2_src ir2_zero(struct ir2_context *ctx); |
| 218 |
|
| 219 |
#define ir2_foreach_instr(it, ctx) \ |
| 220 |
for (struct ir2_instr *it = (ctx)->instr; ({ \ |
| 221 |
while (it != &(ctx)->instr[(ctx)->instr_count] && \ |
| 222 |
it->type == IR2_NONE) \ |
| 223 |
it++; \ |
| 224 |
it != &(ctx)->instr[(ctx)->instr_count]; \ |
| 225 |
}); \ |
| 226 |
it++) |
| 227 |
|
| 228 |
#define ir2_foreach_live_reg(it, ctx) \ |
| 229 |
for (struct ir2_reg **__ptr = (ctx)->live_regs, *it; ({ \ |
| 230 |
while (__ptr != &(ctx)->live_regs[64] && *__ptr == NULL) \ |
| 231 |
__ptr++; \ |
| 232 |
__ptr != &(ctx)->live_regs[64] ? (it = *__ptr) : NULL; \ |
| 233 |
}); \ |
| 234 |
it++) |
| 235 |
|
| 236 |
#define ir2_foreach_avail(it) \ |
| 237 |
for (struct ir2_instr **__instrp = avail, *it; \ |
| 238 |
it = *__instrp, __instrp != &avail[avail_count]; __instrp++) |
| 239 |
|
| 240 |
#define ir2_foreach_src(it, instr) \ |
| 241 |
for (struct ir2_src *it = instr->src; it != &instr->src[instr->src_count]; \ |
| 242 |
it++) |
| 243 |
|
| 244 |
/* mask for register allocation |
| 245 |
* 64 registers with 4 components each = 256 bits |
| 246 |
*/ |
| 247 |
/* typedef struct { |
| 248 |
uint64_t data[4]; |
| 249 |
} regmask_t; */ |
| 250 |
|
| 251 |
static inline bool |
| 252 |
mask_isset(uint32_t *mask, unsigned num) |
| 253 |
{ |
| 254 |
return !!(mask[num / 32] & 1 << num % 32); |
| 255 |
} |
| 256 |
|
| 257 |
static inline void |
| 258 |
mask_set(uint32_t *mask, unsigned num) |
| 259 |
{ |
| 260 |
mask[num / 32] |= 1 << num % 32; |
| 261 |
} |
| 262 |
|
| 263 |
static inline void |
| 264 |
mask_unset(uint32_t *mask, unsigned num) |
| 265 |
{ |
| 266 |
mask[num / 32] &= ~(1 << num % 32); |
| 267 |
} |
| 268 |
|
| 269 |
static inline unsigned |
| 270 |
mask_reg(uint32_t *mask, unsigned num) |
| 271 |
{ |
| 272 |
return mask[num / 8] >> num % 8 * 4 & 0xf; |
| 273 |
} |
| 274 |
|
| 275 |
static inline bool |
| 276 |
is_export(struct ir2_instr *instr) |
| 277 |
{ |
| 278 |
return instr->type == IR2_ALU && instr->alu.export >= 0; |
| 279 |
} |
| 280 |
|
| 281 |
static inline instr_alloc_type_t |
| 282 |
export_buf(unsigned num) |
| 283 |
{ |
| 284 |
return num < 32 ? SQ_PARAMETER_PIXEL : num >= 62 ? SQ_POSITION : SQ_MEMORY; |
| 285 |
} |
| 286 |
|
| 287 |
/* component c for channel i */ |
| 288 |
static inline unsigned |
| 289 |
swiz_set(unsigned c, unsigned i) |
| 290 |
{ |
| 291 |
return ((c - i) & 3) << i * 2; |
| 292 |
} |
| 293 |
|
| 294 |
/* get swizzle in channel i */ |
| 295 |
static inline unsigned |
| 296 |
swiz_get(unsigned swiz, unsigned i) |
| 297 |
{ |
| 298 |
return ((swiz >> i * 2) + i) & 3; |
| 299 |
} |
| 300 |
|
| 301 |
static inline unsigned |
| 302 |
swiz_merge(unsigned swiz0, unsigned swiz1) |
| 303 |
{ |
| 304 |
unsigned swiz = 0; |
| 305 |
for (int i = 0; i < 4; i++) |
| 306 |
swiz |= swiz_set(swiz_get(swiz0, swiz_get(swiz1, i)), i); |
| 307 |
return swiz; |
| 308 |
} |
| 309 |
|
| 310 |
static inline void |
| 311 |
swiz_merge_p(uint8_t *swiz0, unsigned swiz1) |
| 312 |
{ |
| 313 |
unsigned swiz = 0; |
| 314 |
for (int i = 0; i < 4; i++) |
| 315 |
swiz |= swiz_set(swiz_get(*swiz0, swiz_get(swiz1, i)), i); |
| 316 |
*swiz0 = swiz; |
| 317 |
} |
| 318 |
|
| 319 |
static inline struct ir2_reg * |
| 320 |
get_reg(struct ir2_instr *instr) |
| 321 |
{ |
| 322 |
return instr->is_ssa ? &instr->ssa : instr->reg; |
| 323 |
} |
| 324 |
|
| 325 |
static inline struct ir2_reg * |
| 326 |
get_reg_src(struct ir2_context *ctx, struct ir2_src *src) |
| 327 |
{ |
| 328 |
switch (src->type) { |
| 329 |
case IR2_SRC_INPUT: |
| 330 |
return &ctx->input[src->num]; |
| 331 |
case IR2_SRC_SSA: |
| 332 |
return &ctx->instr[src->num].ssa; |
| 333 |
case IR2_SRC_REG: |
| 334 |
return &ctx->reg[src->num]; |
| 335 |
default: |
| 336 |
return NULL; |
| 337 |
} |
| 338 |
} |
| 339 |
|
| 340 |
/* gets a ncomp value for the dst */ |
| 341 |
static inline unsigned |
| 342 |
dst_ncomp(struct ir2_instr *instr) |
| 343 |
{ |
| 344 |
if (instr->is_ssa) |
| 345 |
return instr->ssa.ncomp; |
| 346 |
|
| 347 |
if (instr->type == IR2_FETCH) |
| 348 |
return instr->reg->ncomp; |
| 349 |
|
| 350 |
assert(instr->type == IR2_ALU); |
| 351 |
|
| 352 |
unsigned ncomp = 0; |
| 353 |
for (int i = 0; i < instr->reg->ncomp; i++) |
| 354 |
ncomp += !!(instr->alu.write_mask & 1 << i); |
| 355 |
return ncomp; |
| 356 |
} |
| 357 |
|
| 358 |
/* gets a ncomp value for the src registers */ |
| 359 |
static inline unsigned |
| 360 |
src_ncomp(struct ir2_instr *instr) |
| 361 |
{ |
| 362 |
if (instr->type == IR2_FETCH) { |
| 363 |
switch (instr->fetch.opc) { |
| 364 |
case VTX_FETCH: |
| 365 |
return 1; |
| 366 |
case TEX_FETCH: |
| 367 |
return instr->fetch.tex.is_cube ? 3 : 2; |
| 368 |
case TEX_SET_TEX_LOD: |
| 369 |
return 1; |
| 370 |
default: |
| 371 |
assert(0); |
| 372 |
} |
| 373 |
} |
| 374 |
|
| 375 |
switch (instr->alu.scalar_opc) { |
| 376 |
case PRED_SETEs ... KILLONEs: |
| 377 |
return 1; |
| 378 |
default: |
| 379 |
break; |
| 380 |
} |
| 381 |
|
| 382 |
switch (instr->alu.vector_opc) { |
| 383 |
case DOT2ADDv: |
| 384 |
return 2; |
| 385 |
case DOT3v: |
| 386 |
return 3; |
| 387 |
case DOT4v: |
| 388 |
case CUBEv: |
| 389 |
case PRED_SETE_PUSHv: |
| 390 |
return 4; |
| 391 |
default: |
| 392 |
return dst_ncomp(instr); |
| 393 |
} |
| 394 |
} |