Details | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
5564 | serge | 1 | /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */ |
2 | |||
3 | /* |
||
4 | * Copyright (C) 2013 Rob Clark |
||
5 | * |
||
6 | * Permission is hereby granted, free of charge, to any person obtaining a |
||
7 | * copy of this software and associated documentation files (the "Software"), |
||
8 | * to deal in the Software without restriction, including without limitation |
||
9 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, |
||
10 | * and/or sell copies of the Software, and to permit persons to whom the |
||
11 | * Software is furnished to do so, subject to the following conditions: |
||
12 | * |
||
13 | * The above copyright notice and this permission notice (including the next |
||
14 | * paragraph) shall be included in all copies or substantial portions of the |
||
15 | * Software. |
||
16 | * |
||
17 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
||
18 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
||
19 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
||
20 | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
||
21 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
||
22 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
||
23 | * SOFTWARE. |
||
24 | * |
||
25 | * Authors: |
||
26 | * Rob Clark |
||
27 | */ |
||
28 | |||
29 | #include |
||
30 | |||
31 | #include "pipe/p_state.h" |
||
32 | #include "util/u_string.h" |
||
33 | #include "util/u_memory.h" |
||
34 | #include "util/u_inlines.h" |
||
35 | #include "tgsi/tgsi_lowering.h" |
||
36 | #include "tgsi/tgsi_parse.h" |
||
37 | #include "tgsi/tgsi_ureg.h" |
||
38 | #include "tgsi/tgsi_info.h" |
||
39 | #include "tgsi/tgsi_strings.h" |
||
40 | #include "tgsi/tgsi_dump.h" |
||
41 | #include "tgsi/tgsi_scan.h" |
||
42 | |||
43 | #include "freedreno_util.h" |
||
44 | |||
45 | #include "ir3_compiler.h" |
||
46 | #include "ir3_shader.h" |
||
47 | |||
48 | #include "instr-a3xx.h" |
||
49 | #include "ir3.h" |
||
50 | |||
51 | struct ir3_compile_context { |
||
52 | const struct tgsi_token *tokens; |
||
53 | bool free_tokens; |
||
54 | struct ir3 *ir; |
||
55 | struct ir3_shader_variant *so; |
||
56 | uint16_t integer_s; |
||
57 | |||
58 | struct ir3_block *block; |
||
59 | struct ir3_instruction *current_instr; |
||
60 | |||
61 | /* we need to defer updates to block->outputs[] until the end |
||
62 | * of an instruction (so we don't see new value until *after* |
||
63 | * the src registers are processed) |
||
64 | */ |
||
65 | struct { |
||
66 | struct ir3_instruction *instr, **instrp; |
||
67 | } output_updates[64]; |
||
68 | unsigned num_output_updates; |
||
69 | |||
70 | /* are we in a sequence of "atomic" instructions? |
||
71 | */ |
||
72 | bool atomic; |
||
73 | |||
74 | /* For fragment shaders, from the hw perspective the only |
||
75 | * actual input is r0.xy position register passed to bary.f. |
||
76 | * But TGSI doesn't know that, it still declares things as |
||
77 | * IN[] registers. So we do all the input tracking normally |
||
78 | * and fix things up after compile_instructions() |
||
79 | * |
||
80 | * NOTE that frag_pos is the hardware position (possibly it |
||
81 | * is actually an index or tag or some such.. it is *not* |
||
82 | * values that can be directly used for gl_FragCoord..) |
||
83 | */ |
||
84 | struct ir3_instruction *frag_pos, *frag_face, *frag_coord[4]; |
||
85 | |||
86 | /* For vertex shaders, keep track of the system values sources */ |
||
87 | struct ir3_instruction *vertex_id, *basevertex, *instance_id; |
||
88 | |||
89 | struct tgsi_parse_context parser; |
||
90 | unsigned type; |
||
91 | |||
92 | struct tgsi_shader_info info; |
||
93 | |||
94 | /* hmm, would be nice if tgsi_scan_shader figured this out |
||
95 | * for us: |
||
96 | */ |
||
97 | struct { |
||
98 | unsigned first, last; |
||
99 | struct ir3_instruction *fanin; |
||
100 | } array[MAX_ARRAYS]; |
||
101 | uint32_t array_dirty; |
||
102 | /* offset into array[], per file, of first array info */ |
||
103 | uint8_t array_offsets[TGSI_FILE_COUNT]; |
||
104 | |||
105 | /* for calculating input/output positions/linkages: */ |
||
106 | unsigned next_inloc; |
||
107 | |||
108 | /* a4xx (at least patchlevel 0) cannot seem to flat-interpolate |
||
109 | * so we need to use ldlv.u32 to load the varying directly: |
||
110 | */ |
||
111 | bool flat_bypass; |
||
112 | |||
113 | unsigned num_internal_temps; |
||
114 | struct tgsi_src_register internal_temps[8]; |
||
115 | |||
116 | /* for looking up which system value is which */ |
||
117 | unsigned sysval_semantics[8]; |
||
118 | |||
119 | /* idx/slot for last compiler generated immediate */ |
||
120 | unsigned immediate_idx; |
||
121 | |||
122 | /* stack of branch instructions that mark (potentially nested) |
||
123 | * branch if/else/loop/etc |
||
124 | */ |
||
125 | struct { |
||
126 | struct ir3_instruction *instr, *cond; |
||
127 | bool inv; /* true iff in else leg of branch */ |
||
128 | } branch[16]; |
||
129 | unsigned int branch_count; |
||
130 | |||
131 | /* list of kill instructions: */ |
||
132 | struct ir3_instruction *kill[16]; |
||
133 | unsigned int kill_count; |
||
134 | |||
135 | /* used when dst is same as one of the src, to avoid overwriting a |
||
136 | * src element before the remaining scalar instructions that make |
||
137 | * up the vector operation |
||
138 | */ |
||
139 | struct tgsi_dst_register tmp_dst; |
||
140 | struct tgsi_src_register *tmp_src; |
||
141 | |||
142 | /* just for catching incorrect use of get_dst()/put_dst(): |
||
143 | */ |
||
144 | bool using_tmp_dst; |
||
145 | }; |
||
146 | |||
147 | |||
148 | static void vectorize(struct ir3_compile_context *ctx, |
||
149 | struct ir3_instruction *instr, struct tgsi_dst_register *dst, |
||
150 | int nsrcs, ...); |
||
151 | static void create_mov(struct ir3_compile_context *ctx, |
||
152 | struct tgsi_dst_register *dst, struct tgsi_src_register *src); |
||
153 | static type_t get_ftype(struct ir3_compile_context *ctx); |
||
154 | static type_t get_utype(struct ir3_compile_context *ctx); |
||
155 | |||
156 | static unsigned setup_arrays(struct ir3_compile_context *ctx, unsigned file, unsigned i) |
||
157 | { |
||
158 | /* ArrayID 0 for a given file is the legacy array spanning the entire file: */ |
||
159 | ctx->array[i].first = 0; |
||
160 | ctx->array[i].last = ctx->info.file_max[file]; |
||
161 | ctx->array_offsets[file] = i; |
||
162 | i += ctx->info.array_max[file] + 1; |
||
163 | return i; |
||
164 | } |
||
165 | |||
166 | static unsigned |
||
167 | compile_init(struct ir3_compile_context *ctx, struct ir3_shader_variant *so, |
||
168 | const struct tgsi_token *tokens) |
||
169 | { |
||
170 | unsigned ret, i; |
||
171 | struct tgsi_shader_info *info = &ctx->info; |
||
172 | struct tgsi_lowering_config lconfig = { |
||
173 | .color_two_side = so->key.color_two_side, |
||
174 | .lower_DST = true, |
||
175 | .lower_XPD = true, |
||
176 | .lower_SCS = true, |
||
177 | .lower_LRP = true, |
||
178 | .lower_FRC = true, |
||
179 | .lower_POW = true, |
||
180 | .lower_LIT = true, |
||
181 | .lower_EXP = true, |
||
182 | .lower_LOG = true, |
||
183 | .lower_DP4 = true, |
||
184 | .lower_DP3 = true, |
||
185 | .lower_DPH = true, |
||
186 | .lower_DP2 = true, |
||
187 | .lower_DP2A = true, |
||
188 | }; |
||
189 | |||
190 | switch (so->type) { |
||
191 | case SHADER_FRAGMENT: |
||
192 | case SHADER_COMPUTE: |
||
193 | lconfig.saturate_s = so->key.fsaturate_s; |
||
194 | lconfig.saturate_t = so->key.fsaturate_t; |
||
195 | lconfig.saturate_r = so->key.fsaturate_r; |
||
196 | ctx->integer_s = so->key.finteger_s; |
||
197 | break; |
||
198 | case SHADER_VERTEX: |
||
199 | lconfig.saturate_s = so->key.vsaturate_s; |
||
200 | lconfig.saturate_t = so->key.vsaturate_t; |
||
201 | lconfig.saturate_r = so->key.vsaturate_r; |
||
202 | ctx->integer_s = so->key.vinteger_s; |
||
203 | break; |
||
204 | } |
||
205 | |||
206 | if (!so->shader) { |
||
207 | /* hack for standalone compiler which does not have |
||
208 | * screen/context: |
||
209 | */ |
||
210 | } else if (ir3_shader_gpuid(so->shader) >= 400) { |
||
211 | /* a4xx seems to have *no* sam.p */ |
||
212 | lconfig.lower_TXP = ~0; /* lower all txp */ |
||
213 | /* need special handling for "flat" */ |
||
214 | ctx->flat_bypass = true; |
||
215 | } else { |
||
216 | /* a3xx just needs to avoid sam.p for 3d tex */ |
||
217 | lconfig.lower_TXP = (1 << TGSI_TEXTURE_3D); |
||
218 | /* no special handling for "flat" */ |
||
219 | ctx->flat_bypass = false; |
||
220 | } |
||
221 | |||
222 | ctx->tokens = tgsi_transform_lowering(&lconfig, tokens, &ctx->info); |
||
223 | ctx->free_tokens = !!ctx->tokens; |
||
224 | if (!ctx->tokens) { |
||
225 | /* no lowering */ |
||
226 | ctx->tokens = tokens; |
||
227 | } |
||
228 | ctx->ir = so->ir; |
||
229 | ctx->so = so; |
||
230 | ctx->array_dirty = 0; |
||
231 | ctx->next_inloc = 8; |
||
232 | ctx->num_internal_temps = 0; |
||
233 | ctx->branch_count = 0; |
||
234 | ctx->kill_count = 0; |
||
235 | ctx->block = NULL; |
||
236 | ctx->current_instr = NULL; |
||
237 | ctx->num_output_updates = 0; |
||
238 | ctx->atomic = false; |
||
239 | ctx->frag_pos = NULL; |
||
240 | ctx->frag_face = NULL; |
||
241 | ctx->vertex_id = NULL; |
||
242 | ctx->instance_id = NULL; |
||
243 | ctx->tmp_src = NULL; |
||
244 | ctx->using_tmp_dst = false; |
||
245 | |||
246 | memset(ctx->frag_coord, 0, sizeof(ctx->frag_coord)); |
||
247 | memset(ctx->array, 0, sizeof(ctx->array)); |
||
248 | memset(ctx->array_offsets, 0, sizeof(ctx->array_offsets)); |
||
249 | |||
250 | #define FM(x) (1 << TGSI_FILE_##x) |
||
251 | /* NOTE: if relative addressing is used, we set constlen in |
||
252 | * the compiler (to worst-case value) since we don't know in |
||
253 | * the assembler what the max addr reg value can be: |
||
254 | */ |
||
255 | if (info->indirect_files & FM(CONSTANT)) |
||
256 | so->constlen = MIN2(255, ctx->info.const_file_max[0] + 1); |
||
257 | |||
258 | i = 0; |
||
259 | i += setup_arrays(ctx, TGSI_FILE_INPUT, i); |
||
260 | i += setup_arrays(ctx, TGSI_FILE_TEMPORARY, i); |
||
261 | i += setup_arrays(ctx, TGSI_FILE_OUTPUT, i); |
||
262 | /* any others? we don't track arrays for const..*/ |
||
263 | |||
264 | /* Immediates go after constants: */ |
||
265 | so->first_immediate = so->first_driver_param = |
||
266 | info->const_file_max[0] + 1; |
||
267 | /* 1 unit for the vertex id base */ |
||
268 | if (so->type == SHADER_VERTEX) |
||
269 | so->first_immediate++; |
||
270 | /* 4 (vec4) units for ubo base addresses */ |
||
271 | so->first_immediate += 4; |
||
272 | ctx->immediate_idx = 4 * (ctx->info.file_max[TGSI_FILE_IMMEDIATE] + 1); |
||
273 | |||
274 | ret = tgsi_parse_init(&ctx->parser, ctx->tokens); |
||
275 | if (ret != TGSI_PARSE_OK) |
||
276 | return ret; |
||
277 | |||
278 | ctx->type = ctx->parser.FullHeader.Processor.Processor; |
||
279 | |||
280 | return ret; |
||
281 | } |
||
282 | |||
283 | static void |
||
284 | compile_error(struct ir3_compile_context *ctx, const char *format, ...) |
||
285 | { |
||
286 | va_list ap; |
||
287 | va_start(ap, format); |
||
288 | _debug_vprintf(format, ap); |
||
289 | va_end(ap); |
||
290 | tgsi_dump(ctx->tokens, 0); |
||
291 | debug_assert(0); |
||
292 | } |
||
293 | |||
294 | #define compile_assert(ctx, cond) do { \ |
||
295 | if (!(cond)) compile_error((ctx), "failed assert: "#cond"\n"); \ |
||
296 | } while (0) |
||
297 | |||
298 | static void |
||
299 | compile_free(struct ir3_compile_context *ctx) |
||
300 | { |
||
301 | if (ctx->free_tokens) |
||
302 | free((void *)ctx->tokens); |
||
303 | tgsi_parse_free(&ctx->parser); |
||
304 | } |
||
305 | |||
306 | struct instr_translater { |
||
307 | void (*fxn)(const struct instr_translater *t, |
||
308 | struct ir3_compile_context *ctx, |
||
309 | struct tgsi_full_instruction *inst); |
||
310 | unsigned tgsi_opc; |
||
311 | opc_t opc; |
||
312 | opc_t hopc; /* opc to use for half_precision mode, if different */ |
||
313 | unsigned arg; |
||
314 | }; |
||
315 | |||
316 | static void |
||
317 | instr_finish(struct ir3_compile_context *ctx) |
||
318 | { |
||
319 | unsigned i; |
||
320 | |||
321 | if (ctx->atomic) |
||
322 | return; |
||
323 | |||
324 | for (i = 0; i < ctx->num_output_updates; i++) |
||
325 | *(ctx->output_updates[i].instrp) = ctx->output_updates[i].instr; |
||
326 | |||
327 | ctx->num_output_updates = 0; |
||
328 | |||
329 | while (ctx->array_dirty) { |
||
330 | unsigned aid = ffs(ctx->array_dirty) - 1; |
||
331 | ctx->array[aid].fanin = NULL; |
||
332 | ctx->array_dirty &= ~(1 << aid); |
||
333 | } |
||
334 | } |
||
335 | |||
336 | /* For "atomic" groups of instructions, for example the four scalar |
||
337 | * instructions to perform a vec4 operation. Basically this just |
||
338 | * blocks out handling of output_updates so the next scalar instruction |
||
339 | * still sees the result from before the start of the atomic group. |
||
340 | * |
||
341 | * NOTE: when used properly, this could probably replace get/put_dst() |
||
342 | * stuff. |
||
343 | */ |
||
344 | static void |
||
345 | instr_atomic_start(struct ir3_compile_context *ctx) |
||
346 | { |
||
347 | ctx->atomic = true; |
||
348 | } |
||
349 | |||
350 | static void |
||
351 | instr_atomic_end(struct ir3_compile_context *ctx) |
||
352 | { |
||
353 | ctx->atomic = false; |
||
354 | instr_finish(ctx); |
||
355 | } |
||
356 | |||
357 | static struct ir3_instruction * |
||
358 | instr_create(struct ir3_compile_context *ctx, int category, opc_t opc) |
||
359 | { |
||
360 | instr_finish(ctx); |
||
361 | return (ctx->current_instr = ir3_instr_create(ctx->block, category, opc)); |
||
362 | } |
||
363 | |||
364 | static struct ir3_block * |
||
365 | push_block(struct ir3_compile_context *ctx) |
||
366 | { |
||
367 | struct ir3_block *block; |
||
368 | unsigned ntmp, nin, nout; |
||
369 | |||
370 | #define SCALAR_REGS(file) (4 * (ctx->info.file_max[TGSI_FILE_ ## file] + 1)) |
||
371 | |||
372 | /* hmm, give ourselves room to create 8 extra temporaries (vec4): |
||
373 | */ |
||
374 | ntmp = SCALAR_REGS(TEMPORARY); |
||
375 | ntmp += 8 * 4; |
||
376 | |||
377 | nout = SCALAR_REGS(OUTPUT); |
||
378 | nin = SCALAR_REGS(INPUT) + SCALAR_REGS(SYSTEM_VALUE); |
||
379 | |||
380 | /* for outermost block, 'inputs' are the actual shader INPUT |
||
381 | * register file. Reads from INPUT registers always go back to |
||
382 | * top block. For nested blocks, 'inputs' is used to track any |
||
383 | * TEMPORARY file register from one of the enclosing blocks that |
||
384 | * is ready in this block. |
||
385 | */ |
||
386 | if (!ctx->block) { |
||
387 | /* NOTE: fragment shaders actually have two inputs (r0.xy, the |
||
388 | * position) |
||
389 | */ |
||
390 | if (ctx->type == TGSI_PROCESSOR_FRAGMENT) { |
||
391 | int n = 2; |
||
392 | if (ctx->info.reads_position) |
||
393 | n += 4; |
||
394 | if (ctx->info.uses_frontface) |
||
395 | n += 4; |
||
396 | nin = MAX2(n, nin); |
||
397 | nout += ARRAY_SIZE(ctx->kill); |
||
398 | } |
||
399 | } else { |
||
400 | nin = ntmp; |
||
401 | } |
||
402 | |||
403 | block = ir3_block_create(ctx->ir, ntmp, nin, nout); |
||
404 | |||
405 | if ((ctx->type == TGSI_PROCESSOR_FRAGMENT) && !ctx->block) |
||
406 | block->noutputs -= ARRAY_SIZE(ctx->kill); |
||
407 | |||
408 | block->parent = ctx->block; |
||
409 | ctx->block = block; |
||
410 | |||
411 | return block; |
||
412 | } |
||
413 | |||
414 | static void |
||
415 | pop_block(struct ir3_compile_context *ctx) |
||
416 | { |
||
417 | ctx->block = ctx->block->parent; |
||
418 | compile_assert(ctx, ctx->block); |
||
419 | } |
||
420 | |||
421 | static struct ir3_instruction * |
||
422 | create_output(struct ir3_block *block, struct ir3_instruction *instr, |
||
423 | unsigned n) |
||
424 | { |
||
425 | struct ir3_instruction *out; |
||
426 | |||
427 | out = ir3_instr_create(block, -1, OPC_META_OUTPUT); |
||
428 | out->inout.block = block; |
||
429 | ir3_reg_create(out, n, 0); |
||
430 | if (instr) |
||
431 | ir3_reg_create(out, 0, IR3_REG_SSA)->instr = instr; |
||
432 | |||
433 | return out; |
||
434 | } |
||
435 | |||
436 | static struct ir3_instruction * |
||
437 | create_input(struct ir3_block *block, struct ir3_instruction *instr, |
||
438 | unsigned n) |
||
439 | { |
||
440 | struct ir3_instruction *in; |
||
441 | |||
442 | in = ir3_instr_create(block, -1, OPC_META_INPUT); |
||
443 | in->inout.block = block; |
||
444 | ir3_reg_create(in, n, 0); |
||
445 | if (instr) |
||
446 | ir3_reg_create(in, 0, IR3_REG_SSA)->instr = instr; |
||
447 | |||
448 | return in; |
||
449 | } |
||
450 | |||
451 | static struct ir3_instruction * |
||
452 | block_input(struct ir3_block *block, unsigned n) |
||
453 | { |
||
454 | /* references to INPUT register file always go back up to |
||
455 | * top level: |
||
456 | */ |
||
457 | if (block->parent) |
||
458 | return block_input(block->parent, n); |
||
459 | return block->inputs[n]; |
||
460 | } |
||
461 | |||
462 | /* return temporary in scope, creating if needed meta-input node |
||
463 | * to track block inputs |
||
464 | */ |
||
465 | static struct ir3_instruction * |
||
466 | block_temporary(struct ir3_block *block, unsigned n) |
||
467 | { |
||
468 | /* references to TEMPORARY register file, find the nearest |
||
469 | * enclosing block which has already assigned this temporary, |
||
470 | * creating meta-input instructions along the way to keep |
||
471 | * track of block inputs |
||
472 | */ |
||
473 | if (block->parent && !block->temporaries[n]) { |
||
474 | /* if already have input for this block, reuse: */ |
||
475 | if (!block->inputs[n]) |
||
476 | block->inputs[n] = block_temporary(block->parent, n); |
||
477 | |||
478 | /* and create new input to return: */ |
||
479 | return create_input(block, block->inputs[n], n); |
||
480 | } |
||
481 | return block->temporaries[n]; |
||
482 | } |
||
483 | |||
484 | static struct ir3_instruction * |
||
485 | create_immed(struct ir3_compile_context *ctx, float val) |
||
486 | { |
||
487 | /* NOTE: *don't* use instr_create() here! |
||
488 | */ |
||
489 | struct ir3_instruction *instr; |
||
490 | instr = ir3_instr_create(ctx->block, 1, 0); |
||
491 | instr->cat1.src_type = get_ftype(ctx); |
||
492 | instr->cat1.dst_type = get_ftype(ctx); |
||
493 | ir3_reg_create(instr, 0, 0); |
||
494 | ir3_reg_create(instr, 0, IR3_REG_IMMED)->fim_val = val; |
||
495 | return instr; |
||
496 | } |
||
497 | |||
498 | static void |
||
499 | ssa_instr_set(struct ir3_compile_context *ctx, unsigned file, unsigned n, |
||
500 | struct ir3_instruction *instr) |
||
501 | { |
||
502 | struct ir3_block *block = ctx->block; |
||
503 | unsigned idx = ctx->num_output_updates; |
||
504 | |||
505 | compile_assert(ctx, idx < ARRAY_SIZE(ctx->output_updates)); |
||
506 | |||
507 | /* NOTE: defer update of temporaries[idx] or output[idx] |
||
508 | * until instr_finish(), so that if the current instruction |
||
509 | * reads the same TEMP/OUT[] it gets the old value: |
||
510 | * |
||
511 | * bleh.. this might be a bit easier to just figure out |
||
512 | * in instr_finish(). But at that point we've already |
||
513 | * lost information about OUTPUT vs TEMPORARY register |
||
514 | * file.. |
||
515 | */ |
||
516 | |||
517 | switch (file) { |
||
518 | case TGSI_FILE_OUTPUT: |
||
519 | compile_assert(ctx, n < block->noutputs); |
||
520 | ctx->output_updates[idx].instrp = &block->outputs[n]; |
||
521 | ctx->output_updates[idx].instr = instr; |
||
522 | ctx->num_output_updates++; |
||
523 | break; |
||
524 | case TGSI_FILE_TEMPORARY: |
||
525 | compile_assert(ctx, n < block->ntemporaries); |
||
526 | ctx->output_updates[idx].instrp = &block->temporaries[n]; |
||
527 | ctx->output_updates[idx].instr = instr; |
||
528 | ctx->num_output_updates++; |
||
529 | break; |
||
530 | case TGSI_FILE_ADDRESS: |
||
531 | compile_assert(ctx, n < 1); |
||
532 | ctx->output_updates[idx].instrp = &block->address; |
||
533 | ctx->output_updates[idx].instr = instr; |
||
534 | ctx->num_output_updates++; |
||
535 | break; |
||
536 | } |
||
537 | } |
||
538 | |||
539 | static struct ir3_instruction * |
||
540 | ssa_instr_get(struct ir3_compile_context *ctx, unsigned file, unsigned n) |
||
541 | { |
||
542 | struct ir3_block *block = ctx->block; |
||
543 | struct ir3_instruction *instr = NULL; |
||
544 | |||
545 | switch (file) { |
||
546 | case TGSI_FILE_INPUT: |
||
547 | instr = block_input(ctx->block, n); |
||
548 | break; |
||
549 | case TGSI_FILE_OUTPUT: |
||
550 | /* really this should just happen in case of 'MOV_SAT OUT[n], ..', |
||
551 | * for the following clamp instructions: |
||
552 | */ |
||
553 | instr = block->outputs[n]; |
||
554 | /* we don't have to worry about read from an OUTPUT that was |
||
555 | * assigned outside of the current block, because the _SAT |
||
556 | * clamp instructions will always be in the same block as |
||
557 | * the original instruction which wrote the OUTPUT |
||
558 | */ |
||
559 | compile_assert(ctx, instr); |
||
560 | break; |
||
561 | case TGSI_FILE_TEMPORARY: |
||
562 | instr = block_temporary(ctx->block, n); |
||
563 | if (!instr) { |
||
564 | /* this can happen when registers (or components of a TGSI |
||
565 | * register) are used as src before they have been assigned |
||
566 | * (undefined contents). To avoid confusing the rest of the |
||
567 | * compiler, and to generally keep things peachy, substitute |
||
568 | * an instruction that sets the src to 0.0. Or to keep |
||
569 | * things undefined, I could plug in a random number? :-P |
||
570 | * |
||
571 | * NOTE: *don't* use instr_create() here! |
||
572 | */ |
||
573 | instr = create_immed(ctx, 0.0); |
||
574 | /* no need to recreate the immed for every access: */ |
||
575 | block->temporaries[n] = instr; |
||
576 | } |
||
577 | break; |
||
578 | case TGSI_FILE_SYSTEM_VALUE: |
||
579 | switch (ctx->sysval_semantics[n >> 2]) { |
||
580 | case TGSI_SEMANTIC_VERTEXID_NOBASE: |
||
581 | instr = ctx->vertex_id; |
||
582 | break; |
||
583 | case TGSI_SEMANTIC_BASEVERTEX: |
||
584 | instr = ctx->basevertex; |
||
585 | break; |
||
586 | case TGSI_SEMANTIC_INSTANCEID: |
||
587 | instr = ctx->instance_id; |
||
588 | break; |
||
589 | } |
||
590 | break; |
||
591 | } |
||
592 | |||
593 | return instr; |
||
594 | } |
||
595 | |||
596 | static int dst_array_id(struct ir3_compile_context *ctx, |
||
597 | const struct tgsi_dst_register *dst) |
||
598 | { |
||
599 | // XXX complete hack to recover tgsi_full_dst_register... |
||
600 | // nothing that isn't wrapped in a tgsi_full_dst_register |
||
601 | // should be indirect |
||
602 | const struct tgsi_full_dst_register *fdst = (const void *)dst; |
||
603 | return fdst->Indirect.ArrayID + ctx->array_offsets[dst->File]; |
||
604 | } |
||
605 | |||
606 | static int src_array_id(struct ir3_compile_context *ctx, |
||
607 | const struct tgsi_src_register *src) |
||
608 | { |
||
609 | // XXX complete hack to recover tgsi_full_src_register... |
||
610 | // nothing that isn't wrapped in a tgsi_full_src_register |
||
611 | // should be indirect |
||
612 | const struct tgsi_full_src_register *fsrc = (const void *)src; |
||
613 | debug_assert(src->File != TGSI_FILE_CONSTANT); |
||
614 | return fsrc->Indirect.ArrayID + ctx->array_offsets[src->File]; |
||
615 | } |
||
616 | |||
617 | static struct ir3_instruction * |
||
618 | array_fanin(struct ir3_compile_context *ctx, unsigned aid, unsigned file) |
||
619 | { |
||
620 | struct ir3_instruction *instr; |
||
621 | |||
622 | if (ctx->array[aid].fanin) { |
||
623 | instr = ctx->array[aid].fanin; |
||
624 | } else { |
||
625 | unsigned first = ctx->array[aid].first; |
||
626 | unsigned last = ctx->array[aid].last; |
||
627 | unsigned i, j; |
||
628 | |||
629 | instr = ir3_instr_create2(ctx->block, -1, OPC_META_FI, |
||
630 | 1 + (4 * (last + 1 - first))); |
||
631 | ir3_reg_create(instr, 0, 0); |
||
632 | for (i = first; i <= last; i++) { |
||
633 | for (j = 0; j < 4; j++) { |
||
634 | unsigned n = regid(i, j); |
||
635 | ir3_reg_create(instr, 0, IR3_REG_SSA)->instr = |
||
636 | ssa_instr_get(ctx, file, n); |
||
637 | } |
||
638 | } |
||
639 | ctx->array[aid].fanin = instr; |
||
640 | ctx->array_dirty |= (1 << aid); |
||
641 | } |
||
642 | |||
643 | return instr; |
||
644 | } |
||
645 | |||
646 | static void |
||
647 | ssa_dst(struct ir3_compile_context *ctx, struct ir3_instruction *instr, |
||
648 | const struct tgsi_dst_register *dst, unsigned chan) |
||
649 | { |
||
650 | if (dst->Indirect) { |
||
651 | struct ir3_register *reg = instr->regs[0]; |
||
652 | unsigned i, aid = dst_array_id(ctx, dst); |
||
653 | unsigned first = ctx->array[aid].first; |
||
654 | unsigned last = ctx->array[aid].last; |
||
655 | unsigned off = dst->Index - first; /* vec4 offset */ |
||
656 | |||
657 | reg->size = 4 * (1 + last - first); |
||
658 | reg->offset = regid(off, chan); |
||
659 | |||
660 | instr->fanin = array_fanin(ctx, aid, dst->File); |
||
661 | |||
662 | /* annotate with the array-id, to help out the register- |
||
663 | * assignment stage. At least for the case of indirect |
||
664 | * writes, we should capture enough dependencies to |
||
665 | * preserve the order of reads/writes of the array, so |
||
666 | * the multiple "names" for the array should end up all |
||
667 | * assigned to the same registers. |
||
668 | */ |
||
669 | instr->fanin->fi.aid = aid; |
||
670 | |||
671 | /* Since we are scalarizing vec4 tgsi instructions/regs, we |
||
672 | * run into a slight complication here. To do the naive thing |
||
673 | * and setup a fanout for each scalar array element would end |
||
674 | * up with the result that the instructions generated for each |
||
675 | * component of the vec4 would end up clobbering each other. |
||
676 | * So we take advantage here of knowing that the array index |
||
677 | * (after the shl.b) will be a multiple of four, and only set |
||
678 | * every fourth scalar component in the array. See also |
||
679 | * fixup_ssa_dst_array() |
||
680 | */ |
||
681 | for (i = first; i <= last; i++) { |
||
682 | struct ir3_instruction *split; |
||
683 | unsigned n = regid(i, chan); |
||
684 | int off = (4 * (i - first)) + chan; |
||
685 | |||
686 | if (is_meta(instr) && (instr->opc == OPC_META_FO)) |
||
687 | off -= instr->fo.off; |
||
688 | |||
689 | split = ir3_instr_create(ctx->block, -1, OPC_META_FO); |
||
690 | split->fo.off = off; |
||
691 | ir3_reg_create(split, 0, 0); |
||
692 | ir3_reg_create(split, 0, IR3_REG_SSA)->instr = instr; |
||
693 | |||
694 | ssa_instr_set(ctx, dst->File, n, split); |
||
695 | } |
||
696 | } else { |
||
697 | /* normal case (not relative addressed GPR) */ |
||
698 | ssa_instr_set(ctx, dst->File, regid(dst->Index, chan), instr); |
||
699 | } |
||
700 | } |
||
701 | |||
702 | static void |
||
703 | ssa_src(struct ir3_compile_context *ctx, struct ir3_register *reg, |
||
704 | const struct tgsi_src_register *src, unsigned chan) |
||
705 | { |
||
706 | struct ir3_instruction *instr; |
||
707 | |||
708 | if (src->Indirect && (src->File != TGSI_FILE_CONSTANT)) { |
||
709 | /* for relative addressing of gpr's (due to register assignment) |
||
710 | * we must generate a fanin instruction to collect all possible |
||
711 | * array elements that the instruction could address together: |
||
712 | */ |
||
713 | unsigned aid = src_array_id(ctx, src); |
||
714 | unsigned first = ctx->array[aid].first; |
||
715 | unsigned last = ctx->array[aid].last; |
||
716 | unsigned off = src->Index - first; /* vec4 offset */ |
||
717 | |||
718 | reg->size = 4 * (1 + last - first); |
||
719 | reg->offset = regid(off, chan); |
||
720 | |||
721 | instr = array_fanin(ctx, aid, src->File); |
||
722 | } else if (src->File == TGSI_FILE_CONSTANT && src->Dimension) { |
||
723 | const struct tgsi_full_src_register *fsrc = (const void *)src; |
||
724 | struct ir3_instruction *temp = NULL; |
||
725 | int ubo_regid = regid(ctx->so->first_driver_param, 0) + |
||
726 | fsrc->Dimension.Index - 1; |
||
727 | int offset = 0; |
||
728 | |||
729 | /* We don't handle indirect UBO array accesses... yet. */ |
||
730 | compile_assert(ctx, !fsrc->Dimension.Indirect); |
||
731 | /* UBOs start at index 1. */ |
||
732 | compile_assert(ctx, fsrc->Dimension.Index > 0); |
||
733 | |||
734 | if (src->Indirect) { |
||
735 | /* In case of an indirect index, it will have been loaded into an |
||
736 | * address register. There will be a sequence of |
||
737 | * |
||
738 | * shl.b x, val, 2 |
||
739 | * mova a0, x |
||
740 | * |
||
741 | * We rely on this sequence to get the original val out and shift |
||
742 | * it by 4, since we're dealing in vec4 units. |
||
743 | */ |
||
744 | compile_assert(ctx, ctx->block->address); |
||
745 | compile_assert(ctx, ctx->block->address->regs[1]->instr->opc == |
||
746 | OPC_SHL_B); |
||
747 | |||
748 | temp = instr = instr_create(ctx, 2, OPC_SHL_B); |
||
749 | ir3_reg_create(instr, 0, 0); |
||
750 | ir3_reg_create(instr, 0, IR3_REG_HALF | IR3_REG_SSA)->instr = |
||
751 | ctx->block->address->regs[1]->instr->regs[1]->instr; |
||
752 | ir3_reg_create(instr, 0, IR3_REG_IMMED)->iim_val = 4; |
||
753 | } else if (src->Index >= 64) { |
||
754 | /* Otherwise it's a plain index (in vec4 units). Move it into a |
||
755 | * register. |
||
756 | */ |
||
757 | temp = instr = instr_create(ctx, 1, 0); |
||
758 | instr->cat1.src_type = get_utype(ctx); |
||
759 | instr->cat1.dst_type = get_utype(ctx); |
||
760 | ir3_reg_create(instr, 0, 0); |
||
761 | ir3_reg_create(instr, 0, IR3_REG_IMMED)->iim_val = src->Index * 16; |
||
762 | } else { |
||
763 | /* The offset is small enough to fit into the ldg instruction |
||
764 | * directly. |
||
765 | */ |
||
766 | offset = src->Index * 16; |
||
767 | } |
||
768 | |||
769 | if (temp) { |
||
770 | /* If there was an offset (most common), add it to the buffer |
||
771 | * address. |
||
772 | */ |
||
773 | instr = instr_create(ctx, 2, OPC_ADD_S); |
||
774 | ir3_reg_create(instr, 0, 0); |
||
775 | ir3_reg_create(instr, 0, IR3_REG_SSA)->instr = temp; |
||
776 | ir3_reg_create(instr, ubo_regid, IR3_REG_CONST); |
||
777 | } else { |
||
778 | /* Otherwise just load the buffer address directly */ |
||
779 | instr = instr_create(ctx, 1, 0); |
||
780 | instr->cat1.src_type = get_utype(ctx); |
||
781 | instr->cat1.dst_type = get_utype(ctx); |
||
782 | ir3_reg_create(instr, 0, 0); |
||
783 | ir3_reg_create(instr, ubo_regid, IR3_REG_CONST); |
||
784 | } |
||
785 | |||
786 | temp = instr; |
||
787 | |||
788 | instr = instr_create(ctx, 6, OPC_LDG); |
||
789 | instr->cat6.type = TYPE_U32; |
||
790 | instr->cat6.offset = offset + chan * 4; |
||
791 | ir3_reg_create(instr, 0, 0); |
||
792 | ir3_reg_create(instr, 0, IR3_REG_SSA)->instr = temp; |
||
793 | ir3_reg_create(instr, 0, IR3_REG_IMMED)->iim_val = 1; |
||
794 | |||
795 | reg->flags &= ~(IR3_REG_RELATIV | IR3_REG_CONST); |
||
796 | } else { |
||
797 | /* normal case (not relative addressed GPR) */ |
||
798 | instr = ssa_instr_get(ctx, src->File, regid(src->Index, chan)); |
||
799 | } |
||
800 | |||
801 | if (instr) { |
||
802 | reg->flags |= IR3_REG_SSA; |
||
803 | reg->instr = instr; |
||
804 | } else if (reg->flags & IR3_REG_SSA) { |
||
805 | /* special hack for trans_samp() which calls ssa_src() directly |
||
806 | * to build up the collect (fanin) for const src.. (so SSA flag |
||
807 | * set but no src instr... it basically gets lucky because we |
||
808 | * default to 0.0 for "undefined" src instructions, which is |
||
809 | * what it wants. We probably need to give it a better way to |
||
810 | * do this, but for now this hack: |
||
811 | */ |
||
812 | reg->instr = create_immed(ctx, 0.0); |
||
813 | } |
||
814 | } |
||
815 | |||
816 | static struct ir3_register * |
||
817 | add_dst_reg_wrmask(struct ir3_compile_context *ctx, |
||
818 | struct ir3_instruction *instr, const struct tgsi_dst_register *dst, |
||
819 | unsigned chan, unsigned wrmask) |
||
820 | { |
||
821 | unsigned flags = 0, num = 0; |
||
822 | struct ir3_register *reg; |
||
823 | |||
824 | switch (dst->File) { |
||
825 | case TGSI_FILE_OUTPUT: |
||
826 | case TGSI_FILE_TEMPORARY: |
||
827 | /* uses SSA */ |
||
828 | break; |
||
829 | case TGSI_FILE_ADDRESS: |
||
830 | flags |= IR3_REG_ADDR; |
||
831 | /* uses SSA */ |
||
832 | break; |
||
833 | default: |
||
834 | compile_error(ctx, "unsupported dst register file: %s\n", |
||
835 | tgsi_file_name(dst->File)); |
||
836 | break; |
||
837 | } |
||
838 | |||
839 | if (dst->Indirect) { |
||
840 | flags |= IR3_REG_RELATIV; |
||
841 | |||
842 | /* shouldn't happen, and we can't cope with it below: */ |
||
843 | compile_assert(ctx, wrmask == 0x1); |
||
844 | |||
845 | compile_assert(ctx, ctx->block->address); |
||
846 | if (instr->address) |
||
847 | compile_assert(ctx, ctx->block->address == instr->address); |
||
848 | |||
849 | instr->address = ctx->block->address; |
||
850 | array_insert(ctx->ir->indirects, instr); |
||
851 | } |
||
852 | |||
853 | reg = ir3_reg_create(instr, regid(num, chan), flags); |
||
854 | reg->wrmask = wrmask; |
||
855 | |||
856 | if (wrmask == 0x1) { |
||
857 | /* normal case */ |
||
858 | ssa_dst(ctx, instr, dst, chan); |
||
859 | } else if ((dst->File == TGSI_FILE_TEMPORARY) || |
||
860 | (dst->File == TGSI_FILE_OUTPUT) || |
||
861 | (dst->File == TGSI_FILE_ADDRESS)) { |
||
862 | struct ir3_instruction *prev = NULL; |
||
863 | unsigned i; |
||
864 | |||
865 | compile_assert(ctx, !dst->Indirect); |
||
866 | |||
867 | /* if instruction writes multiple, we need to create |
||
868 | * some place-holder collect the registers: |
||
869 | */ |
||
870 | for (i = 0; i < 4; i++) { |
||
871 | /* NOTE: slightly ugly that we setup neighbor ptrs |
||
872 | * for FO here, but handle FI in CP pass.. we should |
||
873 | * probably just always setup neighbor ptrs in the |
||
874 | * frontend? |
||
875 | */ |
||
876 | struct ir3_instruction *split = |
||
877 | ir3_instr_create(ctx->block, -1, OPC_META_FO); |
||
878 | split->fo.off = i; |
||
879 | /* unused dst reg: */ |
||
880 | /* NOTE: set SSA flag on dst here, because unused FO's |
||
881 | * which don't get scheduled will end up not in the |
||
882 | * instruction list when RA sets SSA flag on each dst. |
||
883 | * Slight hack. We really should set SSA flag on |
||
884 | * every dst register in the frontend. |
||
885 | */ |
||
886 | ir3_reg_create(split, 0, IR3_REG_SSA); |
||
887 | /* and src reg used to hold original instr */ |
||
888 | ir3_reg_create(split, 0, IR3_REG_SSA)->instr = instr; |
||
889 | if (prev) { |
||
890 | split->cp.left = prev; |
||
891 | split->cp.left_cnt++; |
||
892 | prev->cp.right = split; |
||
893 | prev->cp.right_cnt++; |
||
894 | } |
||
895 | if ((wrmask & (1 << i)) && !ctx->atomic) |
||
896 | ssa_dst(ctx, split, dst, chan+i); |
||
897 | prev = split; |
||
898 | } |
||
899 | } |
||
900 | |||
901 | return reg; |
||
902 | } |
||
903 | |||
904 | static struct ir3_register * |
||
905 | add_dst_reg(struct ir3_compile_context *ctx, struct ir3_instruction *instr, |
||
906 | const struct tgsi_dst_register *dst, unsigned chan) |
||
907 | { |
||
908 | return add_dst_reg_wrmask(ctx, instr, dst, chan, 0x1); |
||
909 | } |
||
910 | |||
911 | static struct ir3_register * |
||
912 | add_src_reg_wrmask(struct ir3_compile_context *ctx, |
||
913 | struct ir3_instruction *instr, const struct tgsi_src_register *src, |
||
914 | unsigned chan, unsigned wrmask) |
||
915 | { |
||
916 | unsigned flags = 0, num = 0; |
||
917 | struct ir3_register *reg; |
||
918 | |||
919 | switch (src->File) { |
||
920 | case TGSI_FILE_IMMEDIATE: |
||
921 | /* TODO if possible, use actual immediate instead of const.. but |
||
922 | * TGSI has vec4 immediates, we can only embed scalar (of limited |
||
923 | * size, depending on instruction..) |
||
924 | */ |
||
925 | flags |= IR3_REG_CONST; |
||
926 | num = src->Index + ctx->so->first_immediate; |
||
927 | break; |
||
928 | case TGSI_FILE_CONSTANT: |
||
929 | flags |= IR3_REG_CONST; |
||
930 | num = src->Index; |
||
931 | break; |
||
932 | case TGSI_FILE_OUTPUT: |
||
933 | /* NOTE: we should only end up w/ OUTPUT file for things like |
||
934 | * clamp()'ing saturated dst instructions |
||
935 | */ |
||
936 | case TGSI_FILE_INPUT: |
||
937 | case TGSI_FILE_TEMPORARY: |
||
938 | case TGSI_FILE_SYSTEM_VALUE: |
||
939 | /* uses SSA */ |
||
940 | break; |
||
941 | default: |
||
942 | compile_error(ctx, "unsupported src register file: %s\n", |
||
943 | tgsi_file_name(src->File)); |
||
944 | break; |
||
945 | } |
||
946 | |||
947 | /* We seem to have 8 bits (6.2) for dst register always, so I think |
||
948 | * it is safe to assume GPR cannot be >=64 |
||
949 | * |
||
950 | * cat3 instructions only have 8 bits for src2, but cannot take a |
||
951 | * const for src2 |
||
952 | * |
||
953 | * cat5 and cat6 in some cases only has 8 bits, but cannot take a |
||
954 | * const for any src. |
||
955 | * |
||
956 | * Other than that we seem to have 12 bits to encode const src, |
||
957 | * except for cat1 which may only have 11 bits (but that seems like |
||
958 | * a bug) |
||
959 | */ |
||
960 | if (flags & IR3_REG_CONST) |
||
961 | compile_assert(ctx, src->Index < (1 << 9)); |
||
962 | else |
||
963 | compile_assert(ctx, src->Index < (1 << 6)); |
||
964 | |||
965 | /* NOTE: abs/neg modifiers in tgsi only apply to float */ |
||
966 | if (src->Absolute) |
||
967 | flags |= IR3_REG_FABS; |
||
968 | if (src->Negate) |
||
969 | flags |= IR3_REG_FNEG; |
||
970 | |||
971 | if (src->Indirect) { |
||
972 | flags |= IR3_REG_RELATIV; |
||
973 | |||
974 | /* shouldn't happen, and we can't cope with it below: */ |
||
975 | compile_assert(ctx, wrmask == 0x1); |
||
976 | |||
977 | compile_assert(ctx, ctx->block->address); |
||
978 | if (instr->address) |
||
979 | compile_assert(ctx, ctx->block->address == instr->address); |
||
980 | |||
981 | instr->address = ctx->block->address; |
||
982 | array_insert(ctx->ir->indirects, instr); |
||
983 | } |
||
984 | |||
985 | reg = ir3_reg_create(instr, regid(num, chan), flags); |
||
986 | reg->wrmask = wrmask; |
||
987 | |||
988 | if (wrmask == 0x1) { |
||
989 | /* normal case */ |
||
990 | ssa_src(ctx, reg, src, chan); |
||
991 | } else if ((src->File == TGSI_FILE_TEMPORARY) || |
||
992 | (src->File == TGSI_FILE_OUTPUT) || |
||
993 | (src->File == TGSI_FILE_INPUT)) { |
||
994 | struct ir3_instruction *collect; |
||
995 | unsigned i; |
||
996 | |||
997 | compile_assert(ctx, !src->Indirect); |
||
998 | |||
999 | /* if instruction reads multiple, we need to create |
||
1000 | * some place-holder collect the registers: |
||
1001 | */ |
||
1002 | collect = ir3_instr_create(ctx->block, -1, OPC_META_FI); |
||
1003 | ir3_reg_create(collect, 0, 0); /* unused dst reg */ |
||
1004 | |||
1005 | for (i = 0; i < 4; i++) { |
||
1006 | if (wrmask & (1 << i)) { |
||
1007 | /* and src reg used point to the original instr */ |
||
1008 | ssa_src(ctx, ir3_reg_create(collect, 0, IR3_REG_SSA), |
||
1009 | src, chan + i); |
||
1010 | } else if (wrmask & ~((i << i) - 1)) { |
||
1011 | /* if any remaining components, then dummy |
||
1012 | * placeholder src reg to fill in the blanks: |
||
1013 | */ |
||
1014 | ir3_reg_create(collect, 0, 0); |
||
1015 | } |
||
1016 | } |
||
1017 | |||
1018 | reg->flags |= IR3_REG_SSA; |
||
1019 | reg->instr = collect; |
||
1020 | } |
||
1021 | |||
1022 | return reg; |
||
1023 | } |
||
1024 | |||
1025 | static struct ir3_register * |
||
1026 | add_src_reg(struct ir3_compile_context *ctx, struct ir3_instruction *instr, |
||
1027 | const struct tgsi_src_register *src, unsigned chan) |
||
1028 | { |
||
1029 | return add_src_reg_wrmask(ctx, instr, src, chan, 0x1); |
||
1030 | } |
||
1031 | |||
1032 | static void |
||
1033 | src_from_dst(struct tgsi_src_register *src, struct tgsi_dst_register *dst) |
||
1034 | { |
||
1035 | src->File = dst->File; |
||
1036 | src->Indirect = dst->Indirect; |
||
1037 | src->Dimension = dst->Dimension; |
||
1038 | src->Index = dst->Index; |
||
1039 | src->Absolute = 0; |
||
1040 | src->Negate = 0; |
||
1041 | src->SwizzleX = TGSI_SWIZZLE_X; |
||
1042 | src->SwizzleY = TGSI_SWIZZLE_Y; |
||
1043 | src->SwizzleZ = TGSI_SWIZZLE_Z; |
||
1044 | src->SwizzleW = TGSI_SWIZZLE_W; |
||
1045 | } |
||
1046 | |||
1047 | /* Get internal-temp src/dst to use for a sequence of instructions |
||
1048 | * generated by a single TGSI op. |
||
1049 | */ |
||
1050 | static struct tgsi_src_register * |
||
1051 | get_internal_temp(struct ir3_compile_context *ctx, |
||
1052 | struct tgsi_dst_register *tmp_dst) |
||
1053 | { |
||
1054 | struct tgsi_src_register *tmp_src; |
||
1055 | int n; |
||
1056 | |||
1057 | tmp_dst->File = TGSI_FILE_TEMPORARY; |
||
1058 | tmp_dst->WriteMask = TGSI_WRITEMASK_XYZW; |
||
1059 | tmp_dst->Indirect = 0; |
||
1060 | tmp_dst->Dimension = 0; |
||
1061 | |||
1062 | /* assign next temporary: */ |
||
1063 | n = ctx->num_internal_temps++; |
||
1064 | compile_assert(ctx, n < ARRAY_SIZE(ctx->internal_temps)); |
||
1065 | tmp_src = &ctx->internal_temps[n]; |
||
1066 | |||
1067 | tmp_dst->Index = ctx->info.file_max[TGSI_FILE_TEMPORARY] + n + 1; |
||
1068 | |||
1069 | src_from_dst(tmp_src, tmp_dst); |
||
1070 | |||
1071 | return tmp_src; |
||
1072 | } |
||
1073 | |||
1074 | static inline bool |
||
1075 | is_const(struct tgsi_src_register *src) |
||
1076 | { |
||
1077 | return (src->File == TGSI_FILE_CONSTANT) || |
||
1078 | (src->File == TGSI_FILE_IMMEDIATE); |
||
1079 | } |
||
1080 | |||
1081 | static inline bool |
||
1082 | is_relative(struct tgsi_src_register *src) |
||
1083 | { |
||
1084 | return src->Indirect; |
||
1085 | } |
||
1086 | |||
1087 | static inline bool |
||
1088 | is_rel_or_const(struct tgsi_src_register *src) |
||
1089 | { |
||
1090 | return is_relative(src) || is_const(src); |
||
1091 | } |
||
1092 | |||
1093 | static type_t |
||
1094 | get_ftype(struct ir3_compile_context *ctx) |
||
1095 | { |
||
1096 | return TYPE_F32; |
||
1097 | } |
||
1098 | |||
1099 | static type_t |
||
1100 | get_utype(struct ir3_compile_context *ctx) |
||
1101 | { |
||
1102 | return TYPE_U32; |
||
1103 | } |
||
1104 | |||
1105 | static type_t |
||
1106 | get_stype(struct ir3_compile_context *ctx) |
||
1107 | { |
||
1108 | return TYPE_S32; |
||
1109 | } |
||
1110 | |||
1111 | static unsigned |
||
1112 | src_swiz(struct tgsi_src_register *src, int chan) |
||
1113 | { |
||
1114 | switch (chan) { |
||
1115 | case 0: return src->SwizzleX; |
||
1116 | case 1: return src->SwizzleY; |
||
1117 | case 2: return src->SwizzleZ; |
||
1118 | case 3: return src->SwizzleW; |
||
1119 | } |
||
1120 | assert(0); |
||
1121 | return 0; |
||
1122 | } |
||
1123 | |||
1124 | /* for instructions that cannot take a const register as src, if needed |
||
1125 | * generate a move to temporary gpr: |
||
1126 | */ |
||
1127 | static struct tgsi_src_register * |
||
1128 | get_unconst(struct ir3_compile_context *ctx, struct tgsi_src_register *src) |
||
1129 | { |
||
1130 | struct tgsi_dst_register tmp_dst; |
||
1131 | struct tgsi_src_register *tmp_src; |
||
1132 | |||
1133 | compile_assert(ctx, is_rel_or_const(src)); |
||
1134 | |||
1135 | tmp_src = get_internal_temp(ctx, &tmp_dst); |
||
1136 | |||
1137 | create_mov(ctx, &tmp_dst, src); |
||
1138 | |||
1139 | return tmp_src; |
||
1140 | } |
||
1141 | |||
1142 | static void |
||
1143 | get_immediate(struct ir3_compile_context *ctx, |
||
1144 | struct tgsi_src_register *reg, uint32_t val) |
||
1145 | { |
||
1146 | unsigned neg, swiz, idx, i; |
||
1147 | /* actually maps 1:1 currently.. not sure if that is safe to rely on: */ |
||
1148 | static const unsigned swiz2tgsi[] = { |
||
1149 | TGSI_SWIZZLE_X, TGSI_SWIZZLE_Y, TGSI_SWIZZLE_Z, TGSI_SWIZZLE_W, |
||
1150 | }; |
||
1151 | |||
1152 | for (i = 0; i < ctx->immediate_idx; i++) { |
||
1153 | swiz = i % 4; |
||
1154 | idx = i / 4; |
||
1155 | |||
1156 | if (ctx->so->immediates[idx].val[swiz] == val) { |
||
1157 | neg = 0; |
||
1158 | break; |
||
1159 | } |
||
1160 | |||
1161 | if (ctx->so->immediates[idx].val[swiz] == -val) { |
||
1162 | neg = 1; |
||
1163 | break; |
||
1164 | } |
||
1165 | } |
||
1166 | |||
1167 | if (i == ctx->immediate_idx) { |
||
1168 | /* need to generate a new immediate: */ |
||
1169 | swiz = i % 4; |
||
1170 | idx = i / 4; |
||
1171 | neg = 0; |
||
1172 | ctx->so->immediates[idx].val[swiz] = val; |
||
1173 | ctx->so->immediates_count = idx + 1; |
||
1174 | ctx->immediate_idx++; |
||
1175 | } |
||
1176 | |||
1177 | reg->File = TGSI_FILE_IMMEDIATE; |
||
1178 | reg->Indirect = 0; |
||
1179 | reg->Dimension = 0; |
||
1180 | reg->Index = idx; |
||
1181 | reg->Absolute = 0; |
||
1182 | reg->Negate = neg; |
||
1183 | reg->SwizzleX = swiz2tgsi[swiz]; |
||
1184 | reg->SwizzleY = swiz2tgsi[swiz]; |
||
1185 | reg->SwizzleZ = swiz2tgsi[swiz]; |
||
1186 | reg->SwizzleW = swiz2tgsi[swiz]; |
||
1187 | } |
||
1188 | |||
1189 | static void |
||
1190 | create_mov(struct ir3_compile_context *ctx, struct tgsi_dst_register *dst, |
||
1191 | struct tgsi_src_register *src) |
||
1192 | { |
||
1193 | type_t type_mov = get_ftype(ctx); |
||
1194 | unsigned i; |
||
1195 | |||
1196 | for (i = 0; i < 4; i++) { |
||
1197 | /* move to destination: */ |
||
1198 | if (dst->WriteMask & (1 << i)) { |
||
1199 | struct ir3_instruction *instr; |
||
1200 | |||
1201 | if (src->Absolute || src->Negate) { |
||
1202 | /* can't have abs or neg on a mov instr, so use |
||
1203 | * absneg.f instead to handle these cases: |
||
1204 | */ |
||
1205 | instr = instr_create(ctx, 2, OPC_ABSNEG_F); |
||
1206 | } else { |
||
1207 | instr = instr_create(ctx, 1, 0); |
||
1208 | instr->cat1.src_type = type_mov; |
||
1209 | instr->cat1.dst_type = type_mov; |
||
1210 | } |
||
1211 | |||
1212 | add_dst_reg(ctx, instr, dst, i); |
||
1213 | add_src_reg(ctx, instr, src, src_swiz(src, i)); |
||
1214 | } |
||
1215 | } |
||
1216 | } |
||
1217 | |||
1218 | static void |
||
1219 | create_clamp(struct ir3_compile_context *ctx, |
||
1220 | struct tgsi_dst_register *dst, struct tgsi_src_register *val, |
||
1221 | struct tgsi_src_register *minval, struct tgsi_src_register *maxval) |
||
1222 | { |
||
1223 | struct ir3_instruction *instr; |
||
1224 | |||
1225 | instr = instr_create(ctx, 2, OPC_MAX_F); |
||
1226 | vectorize(ctx, instr, dst, 2, val, 0, minval, 0); |
||
1227 | |||
1228 | instr = instr_create(ctx, 2, OPC_MIN_F); |
||
1229 | vectorize(ctx, instr, dst, 2, val, 0, maxval, 0); |
||
1230 | } |
||
1231 | |||
1232 | static void |
||
1233 | create_clamp_imm(struct ir3_compile_context *ctx, |
||
1234 | struct tgsi_dst_register *dst, |
||
1235 | uint32_t minval, uint32_t maxval) |
||
1236 | { |
||
1237 | struct tgsi_src_register minconst, maxconst; |
||
1238 | struct tgsi_src_register src; |
||
1239 | |||
1240 | src_from_dst(&src, dst); |
||
1241 | |||
1242 | get_immediate(ctx, &minconst, minval); |
||
1243 | get_immediate(ctx, &maxconst, maxval); |
||
1244 | |||
1245 | create_clamp(ctx, dst, &src, &minconst, &maxconst); |
||
1246 | } |
||
1247 | |||
1248 | static struct tgsi_dst_register * |
||
1249 | get_dst(struct ir3_compile_context *ctx, struct tgsi_full_instruction *inst) |
||
1250 | { |
||
1251 | struct tgsi_dst_register *dst = &inst->Dst[0].Register; |
||
1252 | unsigned i; |
||
1253 | |||
1254 | compile_assert(ctx, !ctx->using_tmp_dst); |
||
1255 | ctx->using_tmp_dst = true; |
||
1256 | |||
1257 | for (i = 0; i < inst->Instruction.NumSrcRegs; i++) { |
||
1258 | struct tgsi_src_register *src = &inst->Src[i].Register; |
||
1259 | if ((src->File == dst->File) && (src->Index == dst->Index)) { |
||
1260 | if ((dst->WriteMask == TGSI_WRITEMASK_XYZW) && |
||
1261 | (src->SwizzleX == TGSI_SWIZZLE_X) && |
||
1262 | (src->SwizzleY == TGSI_SWIZZLE_Y) && |
||
1263 | (src->SwizzleZ == TGSI_SWIZZLE_Z) && |
||
1264 | (src->SwizzleW == TGSI_SWIZZLE_W)) |
||
1265 | continue; |
||
1266 | ctx->tmp_src = get_internal_temp(ctx, &ctx->tmp_dst); |
||
1267 | ctx->tmp_dst.WriteMask = dst->WriteMask; |
||
1268 | dst = &ctx->tmp_dst; |
||
1269 | break; |
||
1270 | } |
||
1271 | } |
||
1272 | return dst; |
||
1273 | } |
||
1274 | |||
1275 | static void |
||
1276 | put_dst(struct ir3_compile_context *ctx, struct tgsi_full_instruction *inst, |
||
1277 | struct tgsi_dst_register *dst) |
||
1278 | { |
||
1279 | compile_assert(ctx, ctx->using_tmp_dst); |
||
1280 | ctx->using_tmp_dst = false; |
||
1281 | |||
1282 | /* if necessary, add mov back into original dst: */ |
||
1283 | if (dst != &inst->Dst[0].Register) { |
||
1284 | create_mov(ctx, &inst->Dst[0].Register, ctx->tmp_src); |
||
1285 | } |
||
1286 | } |
||
1287 | |||
1288 | /* helper to generate the necessary repeat and/or additional instructions |
||
1289 | * to turn a scalar instruction into a vector operation: |
||
1290 | */ |
||
1291 | static void |
||
1292 | vectorize(struct ir3_compile_context *ctx, struct ir3_instruction *instr, |
||
1293 | struct tgsi_dst_register *dst, int nsrcs, ...) |
||
1294 | { |
||
1295 | va_list ap; |
||
1296 | int i, j, n = 0; |
||
1297 | |||
1298 | instr_atomic_start(ctx); |
||
1299 | |||
1300 | for (i = 0; i < 4; i++) { |
||
1301 | if (dst->WriteMask & (1 << i)) { |
||
1302 | struct ir3_instruction *cur; |
||
1303 | |||
1304 | if (n++ == 0) { |
||
1305 | cur = instr; |
||
1306 | } else { |
||
1307 | cur = instr_create(ctx, instr->category, instr->opc); |
||
1308 | memcpy(cur->info, instr->info, sizeof(cur->info)); |
||
1309 | } |
||
1310 | |||
1311 | add_dst_reg(ctx, cur, dst, i); |
||
1312 | |||
1313 | va_start(ap, nsrcs); |
||
1314 | for (j = 0; j < nsrcs; j++) { |
||
1315 | struct tgsi_src_register *src = |
||
1316 | va_arg(ap, struct tgsi_src_register *); |
||
1317 | unsigned flags = va_arg(ap, unsigned); |
||
1318 | struct ir3_register *reg; |
||
1319 | if (flags & IR3_REG_IMMED) { |
||
1320 | reg = ir3_reg_create(cur, 0, IR3_REG_IMMED); |
||
1321 | /* this is an ugly cast.. should have put flags first! */ |
||
1322 | reg->iim_val = *(int *)&src; |
||
1323 | } else { |
||
1324 | reg = add_src_reg(ctx, cur, src, src_swiz(src, i)); |
||
1325 | } |
||
1326 | reg->flags |= flags & ~(IR3_REG_FNEG | IR3_REG_SNEG); |
||
1327 | if (flags & IR3_REG_FNEG) |
||
1328 | reg->flags ^= IR3_REG_FNEG; |
||
1329 | if (flags & IR3_REG_SNEG) |
||
1330 | reg->flags ^= IR3_REG_SNEG; |
||
1331 | } |
||
1332 | va_end(ap); |
||
1333 | } |
||
1334 | } |
||
1335 | |||
1336 | instr_atomic_end(ctx); |
||
1337 | } |
||
1338 | |||
1339 | /* |
||
1340 | * Handlers for TGSI instructions which do not have a 1:1 mapping to |
||
1341 | * native instructions: |
||
1342 | */ |
||
1343 | |||
1344 | static void |
||
1345 | trans_clamp(const struct instr_translater *t, |
||
1346 | struct ir3_compile_context *ctx, |
||
1347 | struct tgsi_full_instruction *inst) |
||
1348 | { |
||
1349 | struct tgsi_dst_register *dst = get_dst(ctx, inst); |
||
1350 | struct tgsi_src_register *src0 = &inst->Src[0].Register; |
||
1351 | struct tgsi_src_register *src1 = &inst->Src[1].Register; |
||
1352 | struct tgsi_src_register *src2 = &inst->Src[2].Register; |
||
1353 | |||
1354 | create_clamp(ctx, dst, src0, src1, src2); |
||
1355 | |||
1356 | put_dst(ctx, inst, dst); |
||
1357 | } |
||
1358 | |||
1359 | /* ARL(x) = x, but mova from hrN.x to a0.. */ |
||
1360 | static void |
||
1361 | trans_arl(const struct instr_translater *t, |
||
1362 | struct ir3_compile_context *ctx, |
||
1363 | struct tgsi_full_instruction *inst) |
||
1364 | { |
||
1365 | struct ir3_instruction *instr; |
||
1366 | struct tgsi_dst_register tmp_dst; |
||
1367 | struct tgsi_src_register *tmp_src; |
||
1368 | struct tgsi_dst_register *dst = &inst->Dst[0].Register; |
||
1369 | struct tgsi_src_register *src = &inst->Src[0].Register; |
||
1370 | unsigned chan = src->SwizzleX; |
||
1371 | |||
1372 | compile_assert(ctx, dst->File == TGSI_FILE_ADDRESS); |
||
1373 | |||
1374 | /* NOTE: we allocate a temporary from a flat register |
||
1375 | * namespace (ignoring half vs full). It turns out |
||
1376 | * not to really matter since registers get reassigned |
||
1377 | * later in ir3_ra which (hopefully!) can deal a bit |
||
1378 | * better with mixed half and full precision. |
||
1379 | */ |
||
1380 | tmp_src = get_internal_temp(ctx, &tmp_dst); |
||
1381 | |||
1382 | /* cov.{u,f}{32,16}s16 Rtmp, Rsrc */ |
||
1383 | instr = instr_create(ctx, 1, 0); |
||
1384 | instr->cat1.src_type = (t->tgsi_opc == TGSI_OPCODE_ARL) ? |
||
1385 | get_ftype(ctx) : get_utype(ctx); |
||
1386 | instr->cat1.dst_type = TYPE_S16; |
||
1387 | add_dst_reg(ctx, instr, &tmp_dst, chan)->flags |= IR3_REG_HALF; |
||
1388 | add_src_reg(ctx, instr, src, chan); |
||
1389 | |||
1390 | /* shl.b Rtmp, Rtmp, 2 */ |
||
1391 | instr = instr_create(ctx, 2, OPC_SHL_B); |
||
1392 | add_dst_reg(ctx, instr, &tmp_dst, chan)->flags |= IR3_REG_HALF; |
||
1393 | add_src_reg(ctx, instr, tmp_src, chan)->flags |= IR3_REG_HALF; |
||
1394 | ir3_reg_create(instr, 0, IR3_REG_IMMED)->iim_val = 2; |
||
1395 | |||
1396 | /* mova a0, Rtmp */ |
||
1397 | instr = instr_create(ctx, 1, 0); |
||
1398 | instr->cat1.src_type = TYPE_S16; |
||
1399 | instr->cat1.dst_type = TYPE_S16; |
||
1400 | add_dst_reg(ctx, instr, dst, 0)->flags |= IR3_REG_HALF; |
||
1401 | add_src_reg(ctx, instr, tmp_src, chan)->flags |= IR3_REG_HALF; |
||
1402 | } |
||
1403 | |||
1404 | /* |
||
1405 | * texture fetch/sample instructions: |
||
1406 | */ |
||
1407 | |||
1408 | struct tex_info { |
||
1409 | int8_t order[4]; |
||
1410 | int8_t args; |
||
1411 | unsigned src_wrmask, flags; |
||
1412 | }; |
||
1413 | |||
1414 | struct target_info { |
||
1415 | uint8_t dims; |
||
1416 | uint8_t cube; |
||
1417 | uint8_t array; |
||
1418 | uint8_t shadow; |
||
1419 | }; |
||
1420 | |||
1421 | static const struct target_info tex_targets[] = { |
||
1422 | [TGSI_TEXTURE_1D] = { 1, 0, 0, 0 }, |
||
1423 | [TGSI_TEXTURE_2D] = { 2, 0, 0, 0 }, |
||
1424 | [TGSI_TEXTURE_3D] = { 3, 0, 0, 0 }, |
||
1425 | [TGSI_TEXTURE_CUBE] = { 3, 1, 0, 0 }, |
||
1426 | [TGSI_TEXTURE_RECT] = { 2, 0, 0, 0 }, |
||
1427 | [TGSI_TEXTURE_SHADOW1D] = { 1, 0, 0, 1 }, |
||
1428 | [TGSI_TEXTURE_SHADOW2D] = { 2, 0, 0, 1 }, |
||
1429 | [TGSI_TEXTURE_SHADOWRECT] = { 2, 0, 0, 1 }, |
||
1430 | [TGSI_TEXTURE_1D_ARRAY] = { 1, 0, 1, 0 }, |
||
1431 | [TGSI_TEXTURE_2D_ARRAY] = { 2, 0, 1, 0 }, |
||
1432 | [TGSI_TEXTURE_SHADOW1D_ARRAY] = { 1, 0, 1, 1 }, |
||
1433 | [TGSI_TEXTURE_SHADOW2D_ARRAY] = { 2, 0, 1, 1 }, |
||
1434 | [TGSI_TEXTURE_SHADOWCUBE] = { 3, 1, 0, 1 }, |
||
1435 | [TGSI_TEXTURE_2D_MSAA] = { 2, 0, 0, 0 }, |
||
1436 | [TGSI_TEXTURE_2D_ARRAY_MSAA] = { 2, 0, 1, 0 }, |
||
1437 | [TGSI_TEXTURE_CUBE_ARRAY] = { 3, 1, 1, 0 }, |
||
1438 | [TGSI_TEXTURE_SHADOWCUBE_ARRAY] = { 3, 1, 1, 1 }, |
||
1439 | }; |
||
1440 | |||
1441 | static void |
||
1442 | fill_tex_info(struct ir3_compile_context *ctx, |
||
1443 | struct tgsi_full_instruction *inst, |
||
1444 | struct tex_info *info) |
||
1445 | { |
||
1446 | const struct target_info *tgt = &tex_targets[inst->Texture.Texture]; |
||
1447 | |||
1448 | if (tgt->dims == 3) |
||
1449 | info->flags |= IR3_INSTR_3D; |
||
1450 | if (tgt->array) |
||
1451 | info->flags |= IR3_INSTR_A; |
||
1452 | if (tgt->shadow) |
||
1453 | info->flags |= IR3_INSTR_S; |
||
1454 | |||
1455 | switch (inst->Instruction.Opcode) { |
||
1456 | case TGSI_OPCODE_TXB: |
||
1457 | case TGSI_OPCODE_TXB2: |
||
1458 | case TGSI_OPCODE_TXL: |
||
1459 | case TGSI_OPCODE_TXF: |
||
1460 | info->args = 2; |
||
1461 | break; |
||
1462 | case TGSI_OPCODE_TXP: |
||
1463 | info->flags |= IR3_INSTR_P; |
||
1464 | /* fallthrough */ |
||
1465 | case TGSI_OPCODE_TEX: |
||
1466 | case TGSI_OPCODE_TXD: |
||
1467 | info->args = 1; |
||
1468 | break; |
||
1469 | } |
||
1470 | |||
1471 | /* |
||
1472 | * lay out the first argument in the proper order: |
||
1473 | * - actual coordinates first |
||
1474 | * - shadow reference |
||
1475 | * - array index |
||
1476 | * - projection w |
||
1477 | * |
||
1478 | * bias/lod go into the second arg |
||
1479 | */ |
||
1480 | int arg, pos = 0; |
||
1481 | for (arg = 0; arg < tgt->dims; arg++) |
||
1482 | info->order[arg] = pos++; |
||
1483 | if (tgt->dims == 1) |
||
1484 | info->order[pos++] = -1; |
||
1485 | if (tgt->shadow) |
||
1486 | info->order[pos++] = MAX2(arg + tgt->array, 2); |
||
1487 | if (tgt->array) |
||
1488 | info->order[pos++] = arg++; |
||
1489 | if (info->flags & IR3_INSTR_P) |
||
1490 | info->order[pos++] = 3; |
||
1491 | |||
1492 | info->src_wrmask = (1 << pos) - 1; |
||
1493 | |||
1494 | for (; pos < 4; pos++) |
||
1495 | info->order[pos] = -1; |
||
1496 | |||
1497 | assert(pos <= 4); |
||
1498 | } |
||
1499 | |||
1500 | static bool check_swiz(struct tgsi_src_register *src, const int8_t order[4]) |
||
1501 | { |
||
1502 | unsigned i; |
||
1503 | for (i = 1; (i < 4) && order[i] >= 0; i++) |
||
1504 | if (src_swiz(src, i) != (src_swiz(src, 0) + order[i])) |
||
1505 | return false; |
||
1506 | return true; |
||
1507 | } |
||
1508 | |||
1509 | static bool is_1d(unsigned tex) |
||
1510 | { |
||
1511 | return tex_targets[tex].dims == 1; |
||
1512 | } |
||
1513 | |||
1514 | static struct tgsi_src_register * |
||
1515 | get_tex_coord(struct ir3_compile_context *ctx, |
||
1516 | struct tgsi_full_instruction *inst, |
||
1517 | const struct tex_info *tinf) |
||
1518 | { |
||
1519 | struct tgsi_src_register *coord = &inst->Src[0].Register; |
||
1520 | struct ir3_instruction *instr; |
||
1521 | unsigned tex = inst->Texture.Texture; |
||
1522 | struct tgsi_dst_register tmp_dst; |
||
1523 | struct tgsi_src_register *tmp_src; |
||
1524 | type_t type_mov = get_ftype(ctx); |
||
1525 | unsigned j; |
||
1526 | |||
1527 | /* need to move things around: */ |
||
1528 | tmp_src = get_internal_temp(ctx, &tmp_dst); |
||
1529 | |||
1530 | for (j = 0; j < 4; j++) { |
||
1531 | if (tinf->order[j] < 0) |
||
1532 | continue; |
||
1533 | instr = instr_create(ctx, 1, 0); /* mov */ |
||
1534 | instr->cat1.src_type = type_mov; |
||
1535 | instr->cat1.dst_type = type_mov; |
||
1536 | add_dst_reg(ctx, instr, &tmp_dst, j); |
||
1537 | add_src_reg(ctx, instr, coord, |
||
1538 | src_swiz(coord, tinf->order[j])); |
||
1539 | } |
||
1540 | |||
1541 | /* fix up .y coord: */ |
||
1542 | if (is_1d(tex)) { |
||
1543 | struct ir3_register *imm; |
||
1544 | instr = instr_create(ctx, 1, 0); /* mov */ |
||
1545 | instr->cat1.src_type = type_mov; |
||
1546 | instr->cat1.dst_type = type_mov; |
||
1547 | add_dst_reg(ctx, instr, &tmp_dst, 1); /* .y */ |
||
1548 | imm = ir3_reg_create(instr, 0, IR3_REG_IMMED); |
||
1549 | if (inst->Instruction.Opcode == TGSI_OPCODE_TXF) |
||
1550 | imm->iim_val = 0; |
||
1551 | else |
||
1552 | imm->fim_val = 0.5; |
||
1553 | } |
||
1554 | |||
1555 | return tmp_src; |
||
1556 | } |
||
1557 | |||
1558 | static void |
||
1559 | trans_samp(const struct instr_translater *t, |
||
1560 | struct ir3_compile_context *ctx, |
||
1561 | struct tgsi_full_instruction *inst) |
||
1562 | { |
||
1563 | struct ir3_instruction *instr, *collect; |
||
1564 | struct ir3_register *reg; |
||
1565 | struct tgsi_dst_register *dst = &inst->Dst[0].Register; |
||
1566 | struct tgsi_src_register *orig, *coord, *samp, *offset, *dpdx, *dpdy; |
||
1567 | struct tgsi_src_register zero; |
||
1568 | const struct target_info *tgt = &tex_targets[inst->Texture.Texture]; |
||
1569 | struct tex_info tinf; |
||
1570 | int i; |
||
1571 | |||
1572 | memset(&tinf, 0, sizeof(tinf)); |
||
1573 | fill_tex_info(ctx, inst, &tinf); |
||
1574 | coord = get_tex_coord(ctx, inst, &tinf); |
||
1575 | get_immediate(ctx, &zero, 0); |
||
1576 | |||
1577 | switch (inst->Instruction.Opcode) { |
||
1578 | case TGSI_OPCODE_TXB2: |
||
1579 | orig = &inst->Src[1].Register; |
||
1580 | samp = &inst->Src[2].Register; |
||
1581 | break; |
||
1582 | case TGSI_OPCODE_TXD: |
||
1583 | orig = &inst->Src[0].Register; |
||
1584 | dpdx = &inst->Src[1].Register; |
||
1585 | dpdy = &inst->Src[2].Register; |
||
1586 | samp = &inst->Src[3].Register; |
||
1587 | if (is_rel_or_const(dpdx)) |
||
1588 | dpdx = get_unconst(ctx, dpdx); |
||
1589 | if (is_rel_or_const(dpdy)) |
||
1590 | dpdy = get_unconst(ctx, dpdy); |
||
1591 | break; |
||
1592 | default: |
||
1593 | orig = &inst->Src[0].Register; |
||
1594 | samp = &inst->Src[1].Register; |
||
1595 | break; |
||
1596 | } |
||
1597 | if (tinf.args > 1 && is_rel_or_const(orig)) |
||
1598 | orig = get_unconst(ctx, orig); |
||
1599 | |||
1600 | /* scale up integer coords for TXF based on the LOD */ |
||
1601 | if (inst->Instruction.Opcode == TGSI_OPCODE_TXF) { |
||
1602 | struct tgsi_dst_register tmp_dst; |
||
1603 | struct tgsi_src_register *tmp_src; |
||
1604 | type_t type_mov = get_utype(ctx); |
||
1605 | |||
1606 | tmp_src = get_internal_temp(ctx, &tmp_dst); |
||
1607 | for (i = 0; i < tgt->dims; i++) { |
||
1608 | instr = instr_create(ctx, 2, OPC_SHL_B); |
||
1609 | add_dst_reg(ctx, instr, &tmp_dst, i); |
||
1610 | add_src_reg(ctx, instr, coord, src_swiz(coord, i)); |
||
1611 | add_src_reg(ctx, instr, orig, orig->SwizzleW); |
||
1612 | } |
||
1613 | if (tgt->dims < 2) { |
||
1614 | instr = instr_create(ctx, 1, 0); |
||
1615 | instr->cat1.src_type = type_mov; |
||
1616 | instr->cat1.dst_type = type_mov; |
||
1617 | add_dst_reg(ctx, instr, &tmp_dst, i); |
||
1618 | add_src_reg(ctx, instr, &zero, zero.SwizzleX); |
||
1619 | i++; |
||
1620 | } |
||
1621 | if (tgt->array) { |
||
1622 | instr = instr_create(ctx, 1, 0); |
||
1623 | instr->cat1.src_type = type_mov; |
||
1624 | instr->cat1.dst_type = type_mov; |
||
1625 | add_dst_reg(ctx, instr, &tmp_dst, i); |
||
1626 | add_src_reg(ctx, instr, coord, src_swiz(coord, i)); |
||
1627 | } |
||
1628 | coord = tmp_src; |
||
1629 | } |
||
1630 | |||
1631 | if (inst->Texture.NumOffsets) { |
||
1632 | struct tgsi_texture_offset *tex_offset = &inst->TexOffsets[0]; |
||
1633 | struct tgsi_src_register offset_src = {0}; |
||
1634 | |||
1635 | offset_src.File = tex_offset->File; |
||
1636 | offset_src.Index = tex_offset->Index; |
||
1637 | offset_src.SwizzleX = tex_offset->SwizzleX; |
||
1638 | offset_src.SwizzleY = tex_offset->SwizzleY; |
||
1639 | offset_src.SwizzleZ = tex_offset->SwizzleZ; |
||
1640 | offset = get_unconst(ctx, &offset_src); |
||
1641 | tinf.flags |= IR3_INSTR_O; |
||
1642 | } |
||
1643 | |||
1644 | instr = instr_create(ctx, 5, t->opc); |
||
1645 | if (ctx->integer_s & (1 << samp->Index)) |
||
1646 | instr->cat5.type = get_utype(ctx); |
||
1647 | else |
||
1648 | instr->cat5.type = get_ftype(ctx); |
||
1649 | instr->cat5.samp = samp->Index; |
||
1650 | instr->cat5.tex = samp->Index; |
||
1651 | instr->flags |= tinf.flags; |
||
1652 | |||
1653 | add_dst_reg_wrmask(ctx, instr, dst, 0, dst->WriteMask); |
||
1654 | |||
1655 | reg = ir3_reg_create(instr, 0, IR3_REG_SSA); |
||
1656 | |||
1657 | collect = ir3_instr_create2(ctx->block, -1, OPC_META_FI, 12); |
||
1658 | ir3_reg_create(collect, 0, 0); |
||
1659 | for (i = 0; i < 4; i++) { |
||
1660 | if (tinf.src_wrmask & (1 << i)) |
||
1661 | ssa_src(ctx, ir3_reg_create(collect, 0, IR3_REG_SSA), |
||
1662 | coord, src_swiz(coord, i)); |
||
1663 | else if (tinf.src_wrmask & ~((1 << i) - 1)) |
||
1664 | ir3_reg_create(collect, 0, 0); |
||
1665 | } |
||
1666 | |||
1667 | /* Attach derivatives onto the end of the fan-in. Derivatives start after |
||
1668 | * the 4th argument, so make sure that fi is padded up to 4 first. |
||
1669 | */ |
||
1670 | if (inst->Instruction.Opcode == TGSI_OPCODE_TXD) { |
||
1671 | while (collect->regs_count < 5) |
||
1672 | ssa_src(ctx, ir3_reg_create(collect, 0, IR3_REG_SSA), |
||
1673 | &zero, zero.SwizzleX); |
||
1674 | for (i = 0; i < tgt->dims; i++) |
||
1675 | ssa_src(ctx, ir3_reg_create(collect, 0, IR3_REG_SSA), dpdx, i); |
||
1676 | if (tgt->dims < 2) |
||
1677 | ssa_src(ctx, ir3_reg_create(collect, 0, IR3_REG_SSA), |
||
1678 | &zero, zero.SwizzleX); |
||
1679 | for (i = 0; i < tgt->dims; i++) |
||
1680 | ssa_src(ctx, ir3_reg_create(collect, 0, IR3_REG_SSA), dpdy, i); |
||
1681 | if (tgt->dims < 2) |
||
1682 | ssa_src(ctx, ir3_reg_create(collect, 0, IR3_REG_SSA), |
||
1683 | &zero, zero.SwizzleX); |
||
1684 | tinf.src_wrmask |= ((1 << (2 * MAX2(tgt->dims, 2))) - 1) << 4; |
||
1685 | } |
||
1686 | |||
1687 | reg->instr = collect; |
||
1688 | reg->wrmask = tinf.src_wrmask; |
||
1689 | |||
1690 | /* The second argument contains the offsets, followed by the lod/bias |
||
1691 | * argument. This is constructed more manually due to the dynamic nature. |
||
1692 | */ |
||
1693 | if (inst->Texture.NumOffsets == 0 && tinf.args == 1) |
||
1694 | return; |
||
1695 | |||
1696 | reg = ir3_reg_create(instr, 0, IR3_REG_SSA); |
||
1697 | |||
1698 | collect = ir3_instr_create2(ctx->block, -1, OPC_META_FI, 5); |
||
1699 | ir3_reg_create(collect, 0, 0); |
||
1700 | |||
1701 | if (inst->Texture.NumOffsets) { |
||
1702 | for (i = 0; i < tgt->dims; i++) |
||
1703 | ssa_src(ctx, ir3_reg_create(collect, 0, IR3_REG_SSA), |
||
1704 | offset, i); |
||
1705 | if (tgt->dims < 2) |
||
1706 | ssa_src(ctx, ir3_reg_create(collect, 0, IR3_REG_SSA), |
||
1707 | &zero, zero.SwizzleX); |
||
1708 | } |
||
1709 | if (inst->Instruction.Opcode == TGSI_OPCODE_TXB2) |
||
1710 | ssa_src(ctx, ir3_reg_create(collect, 0, IR3_REG_SSA), |
||
1711 | orig, orig->SwizzleX); |
||
1712 | else if (tinf.args > 1) |
||
1713 | ssa_src(ctx, ir3_reg_create(collect, 0, IR3_REG_SSA), |
||
1714 | orig, orig->SwizzleW); |
||
1715 | |||
1716 | reg->instr = collect; |
||
1717 | reg->wrmask = (1 << (collect->regs_count - 1)) - 1; |
||
1718 | } |
||
1719 | |||
1720 | static void |
||
1721 | trans_txq(const struct instr_translater *t, |
||
1722 | struct ir3_compile_context *ctx, |
||
1723 | struct tgsi_full_instruction *inst) |
||
1724 | { |
||
1725 | struct ir3_instruction *instr; |
||
1726 | struct tgsi_dst_register *dst = &inst->Dst[0].Register; |
||
1727 | struct tgsi_src_register *level = &inst->Src[0].Register; |
||
1728 | struct tgsi_src_register *samp = &inst->Src[1].Register; |
||
1729 | const struct target_info *tgt = &tex_targets[inst->Texture.Texture]; |
||
1730 | struct tex_info tinf; |
||
1731 | |||
1732 | memset(&tinf, 0, sizeof(tinf)); |
||
1733 | fill_tex_info(ctx, inst, &tinf); |
||
1734 | if (is_rel_or_const(level)) |
||
1735 | level = get_unconst(ctx, level); |
||
1736 | |||
1737 | instr = instr_create(ctx, 5, OPC_GETSIZE); |
||
1738 | instr->cat5.type = get_utype(ctx); |
||
1739 | instr->cat5.samp = samp->Index; |
||
1740 | instr->cat5.tex = samp->Index; |
||
1741 | instr->flags |= tinf.flags; |
||
1742 | |||
1743 | if (tgt->array && (dst->WriteMask & (1 << tgt->dims))) { |
||
1744 | /* Array size actually ends up in .w rather than .z. This doesn't |
||
1745 | * matter for miplevel 0, but for higher mips the value in z is |
||
1746 | * minified whereas w stays. Also, the value in TEX_CONST_3_DEPTH is |
||
1747 | * returned, which means that we have to add 1 to it for arrays. |
||
1748 | */ |
||
1749 | struct tgsi_dst_register tmp_dst; |
||
1750 | struct tgsi_src_register *tmp_src; |
||
1751 | type_t type_mov = get_utype(ctx); |
||
1752 | |||
1753 | tmp_src = get_internal_temp(ctx, &tmp_dst); |
||
1754 | add_dst_reg_wrmask(ctx, instr, &tmp_dst, 0, |
||
1755 | dst->WriteMask | TGSI_WRITEMASK_W); |
||
1756 | add_src_reg_wrmask(ctx, instr, level, level->SwizzleX, 0x1); |
||
1757 | |||
1758 | if (dst->WriteMask & TGSI_WRITEMASK_X) { |
||
1759 | instr = instr_create(ctx, 1, 0); |
||
1760 | instr->cat1.src_type = type_mov; |
||
1761 | instr->cat1.dst_type = type_mov; |
||
1762 | add_dst_reg(ctx, instr, dst, 0); |
||
1763 | add_src_reg(ctx, instr, tmp_src, src_swiz(tmp_src, 0)); |
||
1764 | } |
||
1765 | |||
1766 | if (tgt->dims == 2) { |
||
1767 | if (dst->WriteMask & TGSI_WRITEMASK_Y) { |
||
1768 | instr = instr_create(ctx, 1, 0); |
||
1769 | instr->cat1.src_type = type_mov; |
||
1770 | instr->cat1.dst_type = type_mov; |
||
1771 | add_dst_reg(ctx, instr, dst, 1); |
||
1772 | add_src_reg(ctx, instr, tmp_src, src_swiz(tmp_src, 1)); |
||
1773 | } |
||
1774 | } |
||
1775 | |||
1776 | instr = instr_create(ctx, 2, OPC_ADD_U); |
||
1777 | add_dst_reg(ctx, instr, dst, tgt->dims); |
||
1778 | add_src_reg(ctx, instr, tmp_src, src_swiz(tmp_src, 3)); |
||
1779 | ir3_reg_create(instr, 0, IR3_REG_IMMED)->iim_val = 1; |
||
1780 | } else { |
||
1781 | add_dst_reg_wrmask(ctx, instr, dst, 0, dst->WriteMask); |
||
1782 | add_src_reg_wrmask(ctx, instr, level, level->SwizzleX, 0x1); |
||
1783 | } |
||
1784 | |||
1785 | if (dst->WriteMask & TGSI_WRITEMASK_W) { |
||
1786 | /* The # of levels comes from getinfo.z. We need to add 1 to it, since |
||
1787 | * the value in TEX_CONST_0 is zero-based. |
||
1788 | */ |
||
1789 | struct tgsi_dst_register tmp_dst; |
||
1790 | struct tgsi_src_register *tmp_src; |
||
1791 | |||
1792 | tmp_src = get_internal_temp(ctx, &tmp_dst); |
||
1793 | instr = instr_create(ctx, 5, OPC_GETINFO); |
||
1794 | instr->cat5.type = get_utype(ctx); |
||
1795 | instr->cat5.samp = samp->Index; |
||
1796 | instr->cat5.tex = samp->Index; |
||
1797 | add_dst_reg_wrmask(ctx, instr, &tmp_dst, 0, TGSI_WRITEMASK_Z); |
||
1798 | |||
1799 | instr = instr_create(ctx, 2, OPC_ADD_U); |
||
1800 | add_dst_reg(ctx, instr, dst, 3); |
||
1801 | add_src_reg(ctx, instr, tmp_src, src_swiz(tmp_src, 2)); |
||
1802 | ir3_reg_create(instr, 0, IR3_REG_IMMED)->iim_val = 1; |
||
1803 | } |
||
1804 | } |
||
1805 | |||
1806 | /* DDX/DDY */ |
||
1807 | static void |
||
1808 | trans_deriv(const struct instr_translater *t, |
||
1809 | struct ir3_compile_context *ctx, |
||
1810 | struct tgsi_full_instruction *inst) |
||
1811 | { |
||
1812 | struct ir3_instruction *instr; |
||
1813 | struct tgsi_dst_register *dst = &inst->Dst[0].Register; |
||
1814 | struct tgsi_src_register *src = &inst->Src[0].Register; |
||
1815 | static const int8_t order[4] = {0, 1, 2, 3}; |
||
1816 | |||
1817 | if (!check_swiz(src, order)) { |
||
1818 | struct tgsi_dst_register tmp_dst; |
||
1819 | struct tgsi_src_register *tmp_src; |
||
1820 | |||
1821 | tmp_src = get_internal_temp(ctx, &tmp_dst); |
||
1822 | create_mov(ctx, &tmp_dst, src); |
||
1823 | |||
1824 | src = tmp_src; |
||
1825 | } |
||
1826 | |||
1827 | /* This might be a workaround for hw bug? Blob compiler always |
||
1828 | * seems to work two components at a time for dsy/dsx. It does |
||
1829 | * actually seem to work in some cases (or at least some piglit |
||
1830 | * tests) for four components at a time. But seems more reliable |
||
1831 | * to split this into two instructions like the blob compiler |
||
1832 | * does: |
||
1833 | */ |
||
1834 | |||
1835 | instr = instr_create(ctx, 5, t->opc); |
||
1836 | instr->cat5.type = get_ftype(ctx); |
||
1837 | add_dst_reg_wrmask(ctx, instr, dst, 0, dst->WriteMask & 0x3); |
||
1838 | add_src_reg_wrmask(ctx, instr, src, 0, dst->WriteMask & 0x3); |
||
1839 | |||
1840 | instr = instr_create(ctx, 5, t->opc); |
||
1841 | instr->cat5.type = get_ftype(ctx); |
||
1842 | add_dst_reg_wrmask(ctx, instr, dst, 2, (dst->WriteMask >> 2) & 0x3); |
||
1843 | add_src_reg_wrmask(ctx, instr, src, 2, (dst->WriteMask >> 2) & 0x3); |
||
1844 | } |
||
1845 | |||
1846 | /* |
||
1847 | * SEQ(a,b) = (a == b) ? 1.0 : 0.0 |
||
1848 | * cmps.f.eq tmp0, a, b |
||
1849 | * cov.u16f16 dst, tmp0 |
||
1850 | * |
||
1851 | * SNE(a,b) = (a != b) ? 1.0 : 0.0 |
||
1852 | * cmps.f.ne tmp0, a, b |
||
1853 | * cov.u16f16 dst, tmp0 |
||
1854 | * |
||
1855 | * SGE(a,b) = (a >= b) ? 1.0 : 0.0 |
||
1856 | * cmps.f.ge tmp0, a, b |
||
1857 | * cov.u16f16 dst, tmp0 |
||
1858 | * |
||
1859 | * SLE(a,b) = (a <= b) ? 1.0 : 0.0 |
||
1860 | * cmps.f.le tmp0, a, b |
||
1861 | * cov.u16f16 dst, tmp0 |
||
1862 | * |
||
1863 | * SGT(a,b) = (a > b) ? 1.0 : 0.0 |
||
1864 | * cmps.f.gt tmp0, a, b |
||
1865 | * cov.u16f16 dst, tmp0 |
||
1866 | * |
||
1867 | * SLT(a,b) = (a < b) ? 1.0 : 0.0 |
||
1868 | * cmps.f.lt tmp0, a, b |
||
1869 | * cov.u16f16 dst, tmp0 |
||
1870 | * |
||
1871 | * CMP(a,b,c) = (a < 0.0) ? b : c |
||
1872 | * cmps.f.lt tmp0, a, {0.0} |
||
1873 | * sel.b16 dst, b, tmp0, c |
||
1874 | */ |
||
1875 | static void |
||
1876 | trans_cmp(const struct instr_translater *t, |
||
1877 | struct ir3_compile_context *ctx, |
||
1878 | struct tgsi_full_instruction *inst) |
||
1879 | { |
||
1880 | struct ir3_instruction *instr; |
||
1881 | struct tgsi_dst_register tmp_dst; |
||
1882 | struct tgsi_src_register *tmp_src; |
||
1883 | struct tgsi_src_register constval0; |
||
1884 | /* final instruction for CMP() uses orig src1 and src2: */ |
||
1885 | struct tgsi_dst_register *dst = get_dst(ctx, inst); |
||
1886 | struct tgsi_src_register *a0, *a1, *a2; |
||
1887 | unsigned condition; |
||
1888 | |||
1889 | tmp_src = get_internal_temp(ctx, &tmp_dst); |
||
1890 | |||
1891 | a0 = &inst->Src[0].Register; /* a */ |
||
1892 | a1 = &inst->Src[1].Register; /* b */ |
||
1893 | |||
1894 | switch (t->tgsi_opc) { |
||
1895 | case TGSI_OPCODE_SEQ: |
||
1896 | case TGSI_OPCODE_FSEQ: |
||
1897 | condition = IR3_COND_EQ; |
||
1898 | break; |
||
1899 | case TGSI_OPCODE_SNE: |
||
1900 | case TGSI_OPCODE_FSNE: |
||
1901 | condition = IR3_COND_NE; |
||
1902 | break; |
||
1903 | case TGSI_OPCODE_SGE: |
||
1904 | case TGSI_OPCODE_FSGE: |
||
1905 | condition = IR3_COND_GE; |
||
1906 | break; |
||
1907 | case TGSI_OPCODE_SLT: |
||
1908 | case TGSI_OPCODE_FSLT: |
||
1909 | condition = IR3_COND_LT; |
||
1910 | break; |
||
1911 | case TGSI_OPCODE_SLE: |
||
1912 | condition = IR3_COND_LE; |
||
1913 | break; |
||
1914 | case TGSI_OPCODE_SGT: |
||
1915 | condition = IR3_COND_GT; |
||
1916 | break; |
||
1917 | case TGSI_OPCODE_CMP: |
||
1918 | get_immediate(ctx, &constval0, fui(0.0)); |
||
1919 | a0 = &inst->Src[0].Register; /* a */ |
||
1920 | a1 = &constval0; /* {0.0} */ |
||
1921 | condition = IR3_COND_LT; |
||
1922 | break; |
||
1923 | default: |
||
1924 | compile_assert(ctx, 0); |
||
1925 | return; |
||
1926 | } |
||
1927 | |||
1928 | if (is_const(a0) && is_const(a1)) |
||
1929 | a0 = get_unconst(ctx, a0); |
||
1930 | |||
1931 | /* cmps.f. |
||
1932 | instr = instr_create(ctx, 2, OPC_CMPS_F); |
||
1933 | instr->cat2.condition = condition; |
||
1934 | vectorize(ctx, instr, &tmp_dst, 2, a0, 0, a1, 0); |
||
1935 | |||
1936 | switch (t->tgsi_opc) { |
||
1937 | case TGSI_OPCODE_SEQ: |
||
1938 | case TGSI_OPCODE_SGE: |
||
1939 | case TGSI_OPCODE_SLE: |
||
1940 | case TGSI_OPCODE_SNE: |
||
1941 | case TGSI_OPCODE_SGT: |
||
1942 | case TGSI_OPCODE_SLT: |
||
1943 | /* cov.u16f16 dst, tmp0 */ |
||
1944 | instr = instr_create(ctx, 1, 0); |
||
1945 | instr->cat1.src_type = get_utype(ctx); |
||
1946 | instr->cat1.dst_type = get_ftype(ctx); |
||
1947 | vectorize(ctx, instr, dst, 1, tmp_src, 0); |
||
1948 | break; |
||
1949 | case TGSI_OPCODE_FSEQ: |
||
1950 | case TGSI_OPCODE_FSGE: |
||
1951 | case TGSI_OPCODE_FSNE: |
||
1952 | case TGSI_OPCODE_FSLT: |
||
1953 | /* absneg.s dst, (neg)tmp0 */ |
||
1954 | instr = instr_create(ctx, 2, OPC_ABSNEG_S); |
||
1955 | vectorize(ctx, instr, dst, 1, tmp_src, IR3_REG_SNEG); |
||
1956 | break; |
||
1957 | case TGSI_OPCODE_CMP: |
||
1958 | a1 = &inst->Src[1].Register; |
||
1959 | a2 = &inst->Src[2].Register; |
||
1960 | /* sel.{b32,b16} dst, src2, tmp, src1 */ |
||
1961 | instr = instr_create(ctx, 3, OPC_SEL_B32); |
||
1962 | vectorize(ctx, instr, dst, 3, a1, 0, tmp_src, 0, a2, 0); |
||
1963 | |||
1964 | break; |
||
1965 | } |
||
1966 | |||
1967 | put_dst(ctx, inst, dst); |
||
1968 | } |
||
1969 | |||
1970 | /* |
||
1971 | * USNE(a,b) = (a != b) ? ~0 : 0 |
||
1972 | * cmps.u32.ne dst, a, b |
||
1973 | * |
||
1974 | * USEQ(a,b) = (a == b) ? ~0 : 0 |
||
1975 | * cmps.u32.eq dst, a, b |
||
1976 | * |
||
1977 | * ISGE(a,b) = (a > b) ? ~0 : 0 |
||
1978 | * cmps.s32.ge dst, a, b |
||
1979 | * |
||
1980 | * USGE(a,b) = (a > b) ? ~0 : 0 |
||
1981 | * cmps.u32.ge dst, a, b |
||
1982 | * |
||
1983 | * ISLT(a,b) = (a < b) ? ~0 : 0 |
||
1984 | * cmps.s32.lt dst, a, b |
||
1985 | * |
||
1986 | * USLT(a,b) = (a < b) ? ~0 : 0 |
||
1987 | * cmps.u32.lt dst, a, b |
||
1988 | * |
||
1989 | */ |
||
1990 | static void |
||
1991 | trans_icmp(const struct instr_translater *t, |
||
1992 | struct ir3_compile_context *ctx, |
||
1993 | struct tgsi_full_instruction *inst) |
||
1994 | { |
||
1995 | struct ir3_instruction *instr; |
||
1996 | struct tgsi_dst_register *dst = get_dst(ctx, inst); |
||
1997 | struct tgsi_dst_register tmp_dst; |
||
1998 | struct tgsi_src_register *tmp_src; |
||
1999 | struct tgsi_src_register *a0, *a1; |
||
2000 | unsigned condition; |
||
2001 | |||
2002 | a0 = &inst->Src[0].Register; /* a */ |
||
2003 | a1 = &inst->Src[1].Register; /* b */ |
||
2004 | |||
2005 | switch (t->tgsi_opc) { |
||
2006 | case TGSI_OPCODE_USNE: |
||
2007 | condition = IR3_COND_NE; |
||
2008 | break; |
||
2009 | case TGSI_OPCODE_USEQ: |
||
2010 | condition = IR3_COND_EQ; |
||
2011 | break; |
||
2012 | case TGSI_OPCODE_ISGE: |
||
2013 | case TGSI_OPCODE_USGE: |
||
2014 | condition = IR3_COND_GE; |
||
2015 | break; |
||
2016 | case TGSI_OPCODE_ISLT: |
||
2017 | case TGSI_OPCODE_USLT: |
||
2018 | condition = IR3_COND_LT; |
||
2019 | break; |
||
2020 | |||
2021 | default: |
||
2022 | compile_assert(ctx, 0); |
||
2023 | return; |
||
2024 | } |
||
2025 | |||
2026 | if (is_const(a0) && is_const(a1)) |
||
2027 | a0 = get_unconst(ctx, a0); |
||
2028 | |||
2029 | tmp_src = get_internal_temp(ctx, &tmp_dst); |
||
2030 | /* cmps.{u32,s32}. |
||
2031 | instr = instr_create(ctx, 2, t->opc); |
||
2032 | instr->cat2.condition = condition; |
||
2033 | vectorize(ctx, instr, &tmp_dst, 2, a0, 0, a1, 0); |
||
2034 | |||
2035 | /* absneg.s dst, (neg)tmp */ |
||
2036 | instr = instr_create(ctx, 2, OPC_ABSNEG_S); |
||
2037 | vectorize(ctx, instr, dst, 1, tmp_src, IR3_REG_SNEG); |
||
2038 | |||
2039 | put_dst(ctx, inst, dst); |
||
2040 | } |
||
2041 | |||
2042 | /* |
||
2043 | * UCMP(a,b,c) = a ? b : c |
||
2044 | * sel.b16 dst, b, a, c |
||
2045 | */ |
||
2046 | static void |
||
2047 | trans_ucmp(const struct instr_translater *t, |
||
2048 | struct ir3_compile_context *ctx, |
||
2049 | struct tgsi_full_instruction *inst) |
||
2050 | { |
||
2051 | struct ir3_instruction *instr; |
||
2052 | struct tgsi_dst_register *dst = get_dst(ctx, inst); |
||
2053 | struct tgsi_src_register *a0, *a1, *a2; |
||
2054 | |||
2055 | a0 = &inst->Src[0].Register; /* a */ |
||
2056 | a1 = &inst->Src[1].Register; /* b */ |
||
2057 | a2 = &inst->Src[2].Register; /* c */ |
||
2058 | |||
2059 | if (is_rel_or_const(a0)) |
||
2060 | a0 = get_unconst(ctx, a0); |
||
2061 | |||
2062 | /* sel.{b32,b16} dst, b, a, c */ |
||
2063 | instr = instr_create(ctx, 3, OPC_SEL_B32); |
||
2064 | vectorize(ctx, instr, dst, 3, a1, 0, a0, 0, a2, 0); |
||
2065 | put_dst(ctx, inst, dst); |
||
2066 | } |
||
2067 | |||
2068 | /* |
||
2069 | * ISSG(a) = a < 0 ? -1 : a > 0 ? 1 : 0 |
||
2070 | * cmps.s.lt tmp_neg, a, 0 # 1 if a is negative |
||
2071 | * cmps.s.gt tmp_pos, a, 0 # 1 if a is positive |
||
2072 | * sub.u dst, tmp_pos, tmp_neg |
||
2073 | */ |
||
2074 | static void |
||
2075 | trans_issg(const struct instr_translater *t, |
||
2076 | struct ir3_compile_context *ctx, |
||
2077 | struct tgsi_full_instruction *inst) |
||
2078 | { |
||
2079 | struct ir3_instruction *instr; |
||
2080 | struct tgsi_dst_register *dst = get_dst(ctx, inst); |
||
2081 | struct tgsi_src_register *a = &inst->Src[0].Register; |
||
2082 | struct tgsi_dst_register neg_dst, pos_dst; |
||
2083 | struct tgsi_src_register *neg_src, *pos_src; |
||
2084 | |||
2085 | neg_src = get_internal_temp(ctx, &neg_dst); |
||
2086 | pos_src = get_internal_temp(ctx, &pos_dst); |
||
2087 | |||
2088 | /* cmps.s.lt neg, a, 0 */ |
||
2089 | instr = instr_create(ctx, 2, OPC_CMPS_S); |
||
2090 | instr->cat2.condition = IR3_COND_LT; |
||
2091 | vectorize(ctx, instr, &neg_dst, 2, a, 0, 0, IR3_REG_IMMED); |
||
2092 | |||
2093 | /* cmps.s.gt pos, a, 0 */ |
||
2094 | instr = instr_create(ctx, 2, OPC_CMPS_S); |
||
2095 | instr->cat2.condition = IR3_COND_GT; |
||
2096 | vectorize(ctx, instr, &pos_dst, 2, a, 0, 0, IR3_REG_IMMED); |
||
2097 | |||
2098 | /* sub.u dst, pos, neg */ |
||
2099 | instr = instr_create(ctx, 2, OPC_SUB_U); |
||
2100 | vectorize(ctx, instr, dst, 2, pos_src, 0, neg_src, 0); |
||
2101 | |||
2102 | put_dst(ctx, inst, dst); |
||
2103 | } |
||
2104 | |||
2105 | |||
2106 | |||
2107 | /* |
||
2108 | * Conditional / Flow control |
||
2109 | */ |
||
2110 | |||
2111 | static void |
||
2112 | push_branch(struct ir3_compile_context *ctx, bool inv, |
||
2113 | struct ir3_instruction *instr, struct ir3_instruction *cond) |
||
2114 | { |
||
2115 | unsigned int idx = ctx->branch_count++; |
||
2116 | compile_assert(ctx, idx < ARRAY_SIZE(ctx->branch)); |
||
2117 | ctx->branch[idx].instr = instr; |
||
2118 | ctx->branch[idx].inv = inv; |
||
2119 | /* else side of branch has same condition: */ |
||
2120 | if (!inv) |
||
2121 | ctx->branch[idx].cond = cond; |
||
2122 | } |
||
2123 | |||
2124 | static struct ir3_instruction * |
||
2125 | pop_branch(struct ir3_compile_context *ctx) |
||
2126 | { |
||
2127 | unsigned int idx = --ctx->branch_count; |
||
2128 | return ctx->branch[idx].instr; |
||
2129 | } |
||
2130 | |||
2131 | static void |
||
2132 | trans_if(const struct instr_translater *t, |
||
2133 | struct ir3_compile_context *ctx, |
||
2134 | struct tgsi_full_instruction *inst) |
||
2135 | { |
||
2136 | struct ir3_instruction *instr, *cond; |
||
2137 | struct tgsi_src_register *src = &inst->Src[0].Register; |
||
2138 | struct tgsi_dst_register tmp_dst; |
||
2139 | struct tgsi_src_register *tmp_src; |
||
2140 | struct tgsi_src_register constval; |
||
2141 | |||
2142 | get_immediate(ctx, &constval, fui(0.0)); |
||
2143 | tmp_src = get_internal_temp(ctx, &tmp_dst); |
||
2144 | |||
2145 | if (is_const(src)) |
||
2146 | src = get_unconst(ctx, src); |
||
2147 | |||
2148 | /* cmps.{f,u}.ne tmp0, b, {0.0} */ |
||
2149 | instr = instr_create(ctx, 2, t->opc); |
||
2150 | add_dst_reg(ctx, instr, &tmp_dst, 0); |
||
2151 | add_src_reg(ctx, instr, src, src->SwizzleX); |
||
2152 | add_src_reg(ctx, instr, &constval, constval.SwizzleX); |
||
2153 | instr->cat2.condition = IR3_COND_NE; |
||
2154 | |||
2155 | compile_assert(ctx, instr->regs[1]->flags & IR3_REG_SSA); /* because get_unconst() */ |
||
2156 | cond = instr->regs[1]->instr; |
||
2157 | |||
2158 | /* meta:flow tmp0 */ |
||
2159 | instr = instr_create(ctx, -1, OPC_META_FLOW); |
||
2160 | ir3_reg_create(instr, 0, 0); /* dummy dst */ |
||
2161 | add_src_reg(ctx, instr, tmp_src, TGSI_SWIZZLE_X); |
||
2162 | |||
2163 | push_branch(ctx, false, instr, cond); |
||
2164 | instr->flow.if_block = push_block(ctx); |
||
2165 | } |
||
2166 | |||
2167 | static void |
||
2168 | trans_else(const struct instr_translater *t, |
||
2169 | struct ir3_compile_context *ctx, |
||
2170 | struct tgsi_full_instruction *inst) |
||
2171 | { |
||
2172 | struct ir3_instruction *instr; |
||
2173 | |||
2174 | pop_block(ctx); |
||
2175 | |||
2176 | instr = pop_branch(ctx); |
||
2177 | |||
2178 | compile_assert(ctx, (instr->category == -1) && |
||
2179 | (instr->opc == OPC_META_FLOW)); |
||
2180 | |||
2181 | push_branch(ctx, true, instr, NULL); |
||
2182 | instr->flow.else_block = push_block(ctx); |
||
2183 | } |
||
2184 | |||
2185 | static struct ir3_instruction * |
||
2186 | find_temporary(struct ir3_block *block, unsigned n) |
||
2187 | { |
||
2188 | if (block->parent && !block->temporaries[n]) |
||
2189 | return find_temporary(block->parent, n); |
||
2190 | return block->temporaries[n]; |
||
2191 | } |
||
2192 | |||
2193 | static struct ir3_instruction * |
||
2194 | find_output(struct ir3_block *block, unsigned n) |
||
2195 | { |
||
2196 | if (block->parent && !block->outputs[n]) |
||
2197 | return find_output(block->parent, n); |
||
2198 | return block->outputs[n]; |
||
2199 | } |
||
2200 | |||
2201 | static struct ir3_instruction * |
||
2202 | create_phi(struct ir3_compile_context *ctx, struct ir3_instruction *cond, |
||
2203 | struct ir3_instruction *a, struct ir3_instruction *b) |
||
2204 | { |
||
2205 | struct ir3_instruction *phi; |
||
2206 | |||
2207 | compile_assert(ctx, cond); |
||
2208 | |||
2209 | /* Either side of the condition could be null.. which |
||
2210 | * indicates a variable written on only one side of the |
||
2211 | * branch. Normally this should only be variables not |
||
2212 | * used outside of that side of the branch. So we could |
||
2213 | * just 'return a ? a : b;' in that case. But for better |
||
2214 | * defined undefined behavior we just stick in imm{0.0}. |
||
2215 | * In the common case of a value only used within the |
||
2216 | * one side of the branch, the PHI instruction will not |
||
2217 | * get scheduled |
||
2218 | */ |
||
2219 | if (!a) |
||
2220 | a = create_immed(ctx, 0.0); |
||
2221 | if (!b) |
||
2222 | b = create_immed(ctx, 0.0); |
||
2223 | |||
2224 | phi = instr_create(ctx, -1, OPC_META_PHI); |
||
2225 | ir3_reg_create(phi, 0, 0); /* dummy dst */ |
||
2226 | ir3_reg_create(phi, 0, IR3_REG_SSA)->instr = cond; |
||
2227 | ir3_reg_create(phi, 0, IR3_REG_SSA)->instr = a; |
||
2228 | ir3_reg_create(phi, 0, IR3_REG_SSA)->instr = b; |
||
2229 | |||
2230 | return phi; |
||
2231 | } |
||
2232 | |||
2233 | static void |
||
2234 | trans_endif(const struct instr_translater *t, |
||
2235 | struct ir3_compile_context *ctx, |
||
2236 | struct tgsi_full_instruction *inst) |
||
2237 | { |
||
2238 | struct ir3_instruction *instr; |
||
2239 | struct ir3_block *ifb, *elseb; |
||
2240 | struct ir3_instruction **ifout, **elseout; |
||
2241 | unsigned i, ifnout = 0, elsenout = 0; |
||
2242 | |||
2243 | pop_block(ctx); |
||
2244 | |||
2245 | instr = pop_branch(ctx); |
||
2246 | |||
2247 | compile_assert(ctx, (instr->category == -1) && |
||
2248 | (instr->opc == OPC_META_FLOW)); |
||
2249 | |||
2250 | ifb = instr->flow.if_block; |
||
2251 | elseb = instr->flow.else_block; |
||
2252 | /* if there is no else block, the parent block is used for the |
||
2253 | * branch-not-taken src of the PHI instructions: |
||
2254 | */ |
||
2255 | if (!elseb) |
||
2256 | elseb = ifb->parent; |
||
2257 | |||
2258 | /* worst case sizes: */ |
||
2259 | ifnout = ifb->ntemporaries + ifb->noutputs; |
||
2260 | elsenout = elseb->ntemporaries + elseb->noutputs; |
||
2261 | |||
2262 | ifout = ir3_alloc(ctx->ir, sizeof(ifb->outputs[0]) * ifnout); |
||
2263 | if (elseb != ifb->parent) |
||
2264 | elseout = ir3_alloc(ctx->ir, sizeof(ifb->outputs[0]) * elsenout); |
||
2265 | |||
2266 | ifnout = 0; |
||
2267 | elsenout = 0; |
||
2268 | |||
2269 | /* generate PHI instructions for any temporaries written: */ |
||
2270 | for (i = 0; i < ifb->ntemporaries; i++) { |
||
2271 | struct ir3_instruction *a = ifb->temporaries[i]; |
||
2272 | struct ir3_instruction *b = elseb->temporaries[i]; |
||
2273 | |||
2274 | /* if temporary written in if-block, or if else block |
||
2275 | * is present and temporary written in else-block: |
||
2276 | */ |
||
2277 | if (a || ((elseb != ifb->parent) && b)) { |
||
2278 | struct ir3_instruction *phi; |
||
2279 | |||
2280 | /* if only written on one side, find the closest |
||
2281 | * enclosing update on other side: |
||
2282 | */ |
||
2283 | if (!a) |
||
2284 | a = find_temporary(ifb, i); |
||
2285 | if (!b) |
||
2286 | b = find_temporary(elseb, i); |
||
2287 | |||
2288 | ifout[ifnout] = a; |
||
2289 | a = create_output(ifb, a, ifnout++); |
||
2290 | |||
2291 | if (elseb != ifb->parent) { |
||
2292 | elseout[elsenout] = b; |
||
2293 | b = create_output(elseb, b, elsenout++); |
||
2294 | } |
||
2295 | |||
2296 | phi = create_phi(ctx, instr, a, b); |
||
2297 | ctx->block->temporaries[i] = phi; |
||
2298 | } |
||
2299 | } |
||
2300 | |||
2301 | compile_assert(ctx, ifb->noutputs == elseb->noutputs); |
||
2302 | |||
2303 | /* .. and any outputs written: */ |
||
2304 | for (i = 0; i < ifb->noutputs; i++) { |
||
2305 | struct ir3_instruction *a = ifb->outputs[i]; |
||
2306 | struct ir3_instruction *b = elseb->outputs[i]; |
||
2307 | |||
2308 | /* if output written in if-block, or if else block |
||
2309 | * is present and output written in else-block: |
||
2310 | */ |
||
2311 | if (a || ((elseb != ifb->parent) && b)) { |
||
2312 | struct ir3_instruction *phi; |
||
2313 | |||
2314 | /* if only written on one side, find the closest |
||
2315 | * enclosing update on other side: |
||
2316 | */ |
||
2317 | if (!a) |
||
2318 | a = find_output(ifb, i); |
||
2319 | if (!b) |
||
2320 | b = find_output(elseb, i); |
||
2321 | |||
2322 | ifout[ifnout] = a; |
||
2323 | a = create_output(ifb, a, ifnout++); |
||
2324 | |||
2325 | if (elseb != ifb->parent) { |
||
2326 | elseout[elsenout] = b; |
||
2327 | b = create_output(elseb, b, elsenout++); |
||
2328 | } |
||
2329 | |||
2330 | phi = create_phi(ctx, instr, a, b); |
||
2331 | ctx->block->outputs[i] = phi; |
||
2332 | } |
||
2333 | } |
||
2334 | |||
2335 | ifb->noutputs = ifnout; |
||
2336 | ifb->outputs = ifout; |
||
2337 | |||
2338 | if (elseb != ifb->parent) { |
||
2339 | elseb->noutputs = elsenout; |
||
2340 | elseb->outputs = elseout; |
||
2341 | } |
||
2342 | |||
2343 | // TODO maybe we want to compact block->inputs? |
||
2344 | } |
||
2345 | |||
2346 | /* |
||
2347 | * Kill |
||
2348 | */ |
||
2349 | |||
2350 | static void |
||
2351 | trans_kill(const struct instr_translater *t, |
||
2352 | struct ir3_compile_context *ctx, |
||
2353 | struct tgsi_full_instruction *inst) |
||
2354 | { |
||
2355 | struct ir3_instruction *instr, *immed, *cond = NULL; |
||
2356 | bool inv = false; |
||
2357 | |||
2358 | /* unconditional kill, use enclosing if condition: */ |
||
2359 | if (ctx->branch_count > 0) { |
||
2360 | unsigned int idx = ctx->branch_count - 1; |
||
2361 | cond = ctx->branch[idx].cond; |
||
2362 | inv = ctx->branch[idx].inv; |
||
2363 | } else { |
||
2364 | cond = create_immed(ctx, 1.0); |
||
2365 | } |
||
2366 | |||
2367 | compile_assert(ctx, cond); |
||
2368 | |||
2369 | immed = create_immed(ctx, 0.0); |
||
2370 | |||
2371 | /* cmps.f.ne p0.x, cond, {0.0} */ |
||
2372 | instr = instr_create(ctx, 2, OPC_CMPS_F); |
||
2373 | instr->cat2.condition = IR3_COND_NE; |
||
2374 | ir3_reg_create(instr, regid(REG_P0, 0), 0); |
||
2375 | ir3_reg_create(instr, 0, IR3_REG_SSA)->instr = cond; |
||
2376 | ir3_reg_create(instr, 0, IR3_REG_SSA)->instr = immed; |
||
2377 | cond = instr; |
||
2378 | |||
2379 | /* kill p0.x */ |
||
2380 | instr = instr_create(ctx, 0, OPC_KILL); |
||
2381 | instr->cat0.inv = inv; |
||
2382 | ir3_reg_create(instr, 0, 0); /* dummy dst */ |
||
2383 | ir3_reg_create(instr, 0, IR3_REG_SSA)->instr = cond; |
||
2384 | |||
2385 | ctx->kill[ctx->kill_count++] = instr; |
||
2386 | |||
2387 | ctx->so->has_kill = true; |
||
2388 | } |
||
2389 | |||
2390 | /* |
||
2391 | * Kill-If |
||
2392 | */ |
||
2393 | |||
2394 | static void |
||
2395 | trans_killif(const struct instr_translater *t, |
||
2396 | struct ir3_compile_context *ctx, |
||
2397 | struct tgsi_full_instruction *inst) |
||
2398 | { |
||
2399 | struct tgsi_src_register *src = &inst->Src[0].Register; |
||
2400 | struct ir3_instruction *instr, *immed, *cond = NULL; |
||
2401 | bool inv = false; |
||
2402 | |||
2403 | immed = create_immed(ctx, 0.0); |
||
2404 | |||
2405 | /* cmps.f.ne p0.x, cond, {0.0} */ |
||
2406 | instr = instr_create(ctx, 2, OPC_CMPS_F); |
||
2407 | instr->cat2.condition = IR3_COND_NE; |
||
2408 | ir3_reg_create(instr, regid(REG_P0, 0), 0); |
||
2409 | ir3_reg_create(instr, 0, IR3_REG_SSA)->instr = immed; |
||
2410 | add_src_reg(ctx, instr, src, src->SwizzleX); |
||
2411 | |||
2412 | cond = instr; |
||
2413 | |||
2414 | /* kill p0.x */ |
||
2415 | instr = instr_create(ctx, 0, OPC_KILL); |
||
2416 | instr->cat0.inv = inv; |
||
2417 | ir3_reg_create(instr, 0, 0); /* dummy dst */ |
||
2418 | ir3_reg_create(instr, 0, IR3_REG_SSA)->instr = cond; |
||
2419 | |||
2420 | ctx->kill[ctx->kill_count++] = instr; |
||
2421 | |||
2422 | ctx->so->has_kill = true; |
||
2423 | |||
2424 | } |
||
2425 | /* |
||
2426 | * I2F / U2F / F2I / F2U |
||
2427 | */ |
||
2428 | |||
2429 | static void |
||
2430 | trans_cov(const struct instr_translater *t, |
||
2431 | struct ir3_compile_context *ctx, |
||
2432 | struct tgsi_full_instruction *inst) |
||
2433 | { |
||
2434 | struct ir3_instruction *instr; |
||
2435 | struct tgsi_dst_register *dst = get_dst(ctx, inst); |
||
2436 | struct tgsi_src_register *src = &inst->Src[0].Register; |
||
2437 | |||
2438 | // cov.f32s32 dst, tmp0 / |
||
2439 | instr = instr_create(ctx, 1, 0); |
||
2440 | switch (t->tgsi_opc) { |
||
2441 | case TGSI_OPCODE_U2F: |
||
2442 | instr->cat1.src_type = TYPE_U32; |
||
2443 | instr->cat1.dst_type = TYPE_F32; |
||
2444 | break; |
||
2445 | case TGSI_OPCODE_I2F: |
||
2446 | instr->cat1.src_type = TYPE_S32; |
||
2447 | instr->cat1.dst_type = TYPE_F32; |
||
2448 | break; |
||
2449 | case TGSI_OPCODE_F2U: |
||
2450 | instr->cat1.src_type = TYPE_F32; |
||
2451 | instr->cat1.dst_type = TYPE_U32; |
||
2452 | break; |
||
2453 | case TGSI_OPCODE_F2I: |
||
2454 | instr->cat1.src_type = TYPE_F32; |
||
2455 | instr->cat1.dst_type = TYPE_S32; |
||
2456 | break; |
||
2457 | |||
2458 | } |
||
2459 | vectorize(ctx, instr, dst, 1, src, 0); |
||
2460 | put_dst(ctx, inst, dst); |
||
2461 | } |
||
2462 | |||
2463 | /* |
||
2464 | * UMUL / UMAD |
||
2465 | * |
||
2466 | * There is no 32-bit multiply instruction, so splitting a and b into high and |
||
2467 | * low components, we get that |
||
2468 | * |
||
2469 | * dst = al * bl + ah * bl << 16 + al * bh << 16 |
||
2470 | * |
||
2471 | * mull.u tmp0, a, b (mul low, i.e. al * bl) |
||
2472 | * madsh.m16 tmp1, a, b, tmp0 (mul-add shift high mix, i.e. ah * bl << 16) |
||
2473 | * madsh.m16 dst, b, a, tmp1 (i.e. al * bh << 16) |
||
2474 | * |
||
2475 | * For UMAD, add in the extra argument after mull.u. |
||
2476 | */ |
||
2477 | static void |
||
2478 | trans_umul(const struct instr_translater *t, |
||
2479 | struct ir3_compile_context *ctx, |
||
2480 | struct tgsi_full_instruction *inst) |
||
2481 | { |
||
2482 | struct ir3_instruction *instr; |
||
2483 | struct tgsi_dst_register *dst = get_dst(ctx, inst); |
||
2484 | struct tgsi_src_register *a = &inst->Src[0].Register; |
||
2485 | struct tgsi_src_register *b = &inst->Src[1].Register; |
||
2486 | |||
2487 | struct tgsi_dst_register tmp0_dst, tmp1_dst; |
||
2488 | struct tgsi_src_register *tmp0_src, *tmp1_src; |
||
2489 | |||
2490 | tmp0_src = get_internal_temp(ctx, &tmp0_dst); |
||
2491 | tmp1_src = get_internal_temp(ctx, &tmp1_dst); |
||
2492 | |||
2493 | if (is_rel_or_const(a)) |
||
2494 | a = get_unconst(ctx, a); |
||
2495 | if (is_rel_or_const(b)) |
||
2496 | b = get_unconst(ctx, b); |
||
2497 | |||
2498 | /* mull.u tmp0, a, b */ |
||
2499 | instr = instr_create(ctx, 2, OPC_MULL_U); |
||
2500 | vectorize(ctx, instr, &tmp0_dst, 2, a, 0, b, 0); |
||
2501 | |||
2502 | if (t->tgsi_opc == TGSI_OPCODE_UMAD) { |
||
2503 | struct tgsi_src_register *c = &inst->Src[2].Register; |
||
2504 | |||
2505 | /* add.u tmp0, tmp0, c */ |
||
2506 | instr = instr_create(ctx, 2, OPC_ADD_U); |
||
2507 | vectorize(ctx, instr, &tmp0_dst, 2, tmp0_src, 0, c, 0); |
||
2508 | } |
||
2509 | |||
2510 | /* madsh.m16 tmp1, a, b, tmp0 */ |
||
2511 | instr = instr_create(ctx, 3, OPC_MADSH_M16); |
||
2512 | vectorize(ctx, instr, &tmp1_dst, 3, a, 0, b, 0, tmp0_src, 0); |
||
2513 | |||
2514 | /* madsh.m16 dst, b, a, tmp1 */ |
||
2515 | instr = instr_create(ctx, 3, OPC_MADSH_M16); |
||
2516 | vectorize(ctx, instr, dst, 3, b, 0, a, 0, tmp1_src, 0); |
||
2517 | put_dst(ctx, inst, dst); |
||
2518 | } |
||
2519 | |||
2520 | /* |
||
2521 | * IDIV / UDIV / MOD / UMOD |
||
2522 | * |
||
2523 | * See NV50LegalizeSSA::handleDIV for the origin of this implementation. For |
||
2524 | * MOD/UMOD, it becomes a - [IU]DIV(a, modulus) * modulus. |
||
2525 | */ |
||
2526 | static void |
||
2527 | trans_idiv(const struct instr_translater *t, |
||
2528 | struct ir3_compile_context *ctx, |
||
2529 | struct tgsi_full_instruction *inst) |
||
2530 | { |
||
2531 | struct ir3_instruction *instr; |
||
2532 | struct tgsi_dst_register *dst = get_dst(ctx, inst), *premod_dst = dst; |
||
2533 | struct tgsi_src_register *a = &inst->Src[0].Register; |
||
2534 | struct tgsi_src_register *b = &inst->Src[1].Register; |
||
2535 | |||
2536 | struct tgsi_dst_register af_dst, bf_dst, q_dst, r_dst, a_dst, b_dst; |
||
2537 | struct tgsi_src_register *af_src, *bf_src, *q_src, *r_src, *a_src, *b_src; |
||
2538 | |||
2539 | struct tgsi_src_register negative_2, thirty_one; |
||
2540 | type_t src_type; |
||
2541 | |||
2542 | if (t->tgsi_opc == TGSI_OPCODE_IDIV || t->tgsi_opc == TGSI_OPCODE_MOD) |
||
2543 | src_type = get_stype(ctx); |
||
2544 | else |
||
2545 | src_type = get_utype(ctx); |
||
2546 | |||
2547 | af_src = get_internal_temp(ctx, &af_dst); |
||
2548 | bf_src = get_internal_temp(ctx, &bf_dst); |
||
2549 | q_src = get_internal_temp(ctx, &q_dst); |
||
2550 | r_src = get_internal_temp(ctx, &r_dst); |
||
2551 | a_src = get_internal_temp(ctx, &a_dst); |
||
2552 | b_src = get_internal_temp(ctx, &b_dst); |
||
2553 | |||
2554 | get_immediate(ctx, &negative_2, -2); |
||
2555 | get_immediate(ctx, &thirty_one, 31); |
||
2556 | |||
2557 | if (t->tgsi_opc == TGSI_OPCODE_MOD || t->tgsi_opc == TGSI_OPCODE_UMOD) |
||
2558 | premod_dst = &q_dst; |
||
2559 | |||
2560 | /* cov.[us]32f32 af, numerator */ |
||
2561 | instr = instr_create(ctx, 1, 0); |
||
2562 | instr->cat1.src_type = src_type; |
||
2563 | instr->cat1.dst_type = get_ftype(ctx); |
||
2564 | vectorize(ctx, instr, &af_dst, 1, a, 0); |
||
2565 | |||
2566 | /* cov.[us]32f32 bf, denominator */ |
||
2567 | instr = instr_create(ctx, 1, 0); |
||
2568 | instr->cat1.src_type = src_type; |
||
2569 | instr->cat1.dst_type = get_ftype(ctx); |
||
2570 | vectorize(ctx, instr, &bf_dst, 1, b, 0); |
||
2571 | |||
2572 | /* Get the absolute values for IDIV */ |
||
2573 | if (type_sint(src_type)) { |
||
2574 | /* absneg.f af, (abs)af */ |
||
2575 | instr = instr_create(ctx, 2, OPC_ABSNEG_F); |
||
2576 | vectorize(ctx, instr, &af_dst, 1, af_src, IR3_REG_FABS); |
||
2577 | |||
2578 | /* absneg.f bf, (abs)bf */ |
||
2579 | instr = instr_create(ctx, 2, OPC_ABSNEG_F); |
||
2580 | vectorize(ctx, instr, &bf_dst, 1, bf_src, IR3_REG_FABS); |
||
2581 | |||
2582 | /* absneg.s a, (abs)numerator */ |
||
2583 | instr = instr_create(ctx, 2, OPC_ABSNEG_S); |
||
2584 | vectorize(ctx, instr, &a_dst, 1, a, IR3_REG_SABS); |
||
2585 | |||
2586 | /* absneg.s b, (abs)denominator */ |
||
2587 | instr = instr_create(ctx, 2, OPC_ABSNEG_S); |
||
2588 | vectorize(ctx, instr, &b_dst, 1, b, IR3_REG_SABS); |
||
2589 | } else { |
||
2590 | /* mov.u32u32 a, numerator */ |
||
2591 | instr = instr_create(ctx, 1, 0); |
||
2592 | instr->cat1.src_type = src_type; |
||
2593 | instr->cat1.dst_type = src_type; |
||
2594 | vectorize(ctx, instr, &a_dst, 1, a, 0); |
||
2595 | |||
2596 | /* mov.u32u32 b, denominator */ |
||
2597 | instr = instr_create(ctx, 1, 0); |
||
2598 | instr->cat1.src_type = src_type; |
||
2599 | instr->cat1.dst_type = src_type; |
||
2600 | vectorize(ctx, instr, &b_dst, 1, b, 0); |
||
2601 | } |
||
2602 | |||
2603 | /* rcp.f bf, bf */ |
||
2604 | instr = instr_create(ctx, 4, OPC_RCP); |
||
2605 | vectorize(ctx, instr, &bf_dst, 1, bf_src, 0); |
||
2606 | |||
2607 | /* That's right, subtract 2 as an integer from the float */ |
||
2608 | /* add.u bf, bf, -2 */ |
||
2609 | instr = instr_create(ctx, 2, OPC_ADD_U); |
||
2610 | vectorize(ctx, instr, &bf_dst, 2, bf_src, 0, &negative_2, 0); |
||
2611 | |||
2612 | /* mul.f q, af, bf */ |
||
2613 | instr = instr_create(ctx, 2, OPC_MUL_F); |
||
2614 | vectorize(ctx, instr, &q_dst, 2, af_src, 0, bf_src, 0); |
||
2615 | |||
2616 | /* cov.f32[us]32 q, q */ |
||
2617 | instr = instr_create(ctx, 1, 0); |
||
2618 | instr->cat1.src_type = get_ftype(ctx); |
||
2619 | instr->cat1.dst_type = src_type; |
||
2620 | vectorize(ctx, instr, &q_dst, 1, q_src, 0); |
||
2621 | |||
2622 | /* integer multiply q by b */ |
||
2623 | /* mull.u r, q, b */ |
||
2624 | instr = instr_create(ctx, 2, OPC_MULL_U); |
||
2625 | vectorize(ctx, instr, &r_dst, 2, q_src, 0, b_src, 0); |
||
2626 | |||
2627 | /* madsh.m16 r, q, b, r */ |
||
2628 | instr = instr_create(ctx, 3, OPC_MADSH_M16); |
||
2629 | vectorize(ctx, instr, &r_dst, 3, q_src, 0, b_src, 0, r_src, 0); |
||
2630 | |||
2631 | /* madsh.m16, r, b, q, r */ |
||
2632 | instr = instr_create(ctx, 3, OPC_MADSH_M16); |
||
2633 | vectorize(ctx, instr, &r_dst, 3, b_src, 0, q_src, 0, r_src, 0); |
||
2634 | |||
2635 | /* sub.u r, a, r */ |
||
2636 | instr = instr_create(ctx, 2, OPC_SUB_U); |
||
2637 | vectorize(ctx, instr, &r_dst, 2, a_src, 0, r_src, 0); |
||
2638 | |||
2639 | /* cov.u32f32, r, r */ |
||
2640 | instr = instr_create(ctx, 1, 0); |
||
2641 | instr->cat1.src_type = get_utype(ctx); |
||
2642 | instr->cat1.dst_type = get_ftype(ctx); |
||
2643 | vectorize(ctx, instr, &r_dst, 1, r_src, 0); |
||
2644 | |||
2645 | /* mul.f r, r, bf */ |
||
2646 | instr = instr_create(ctx, 2, OPC_MUL_F); |
||
2647 | vectorize(ctx, instr, &r_dst, 2, r_src, 0, bf_src, 0); |
||
2648 | |||
2649 | /* cov.f32u32 r, r */ |
||
2650 | instr = instr_create(ctx, 1, 0); |
||
2651 | instr->cat1.src_type = get_ftype(ctx); |
||
2652 | instr->cat1.dst_type = get_utype(ctx); |
||
2653 | vectorize(ctx, instr, &r_dst, 1, r_src, 0); |
||
2654 | |||
2655 | /* add.u q, q, r */ |
||
2656 | instr = instr_create(ctx, 2, OPC_ADD_U); |
||
2657 | vectorize(ctx, instr, &q_dst, 2, q_src, 0, r_src, 0); |
||
2658 | |||
2659 | /* mull.u r, q, b */ |
||
2660 | instr = instr_create(ctx, 2, OPC_MULL_U); |
||
2661 | vectorize(ctx, instr, &r_dst, 2, q_src, 0, b_src, 0); |
||
2662 | |||
2663 | /* madsh.m16 r, q, b, r */ |
||
2664 | instr = instr_create(ctx, 3, OPC_MADSH_M16); |
||
2665 | vectorize(ctx, instr, &r_dst, 3, q_src, 0, b_src, 0, r_src, 0); |
||
2666 | |||
2667 | /* madsh.m16 r, b, q, r */ |
||
2668 | instr = instr_create(ctx, 3, OPC_MADSH_M16); |
||
2669 | vectorize(ctx, instr, &r_dst, 3, b_src, 0, q_src, 0, r_src, 0); |
||
2670 | |||
2671 | /* sub.u r, a, r */ |
||
2672 | instr = instr_create(ctx, 2, OPC_SUB_U); |
||
2673 | vectorize(ctx, instr, &r_dst, 2, a_src, 0, r_src, 0); |
||
2674 | |||
2675 | /* cmps.u.ge r, r, b */ |
||
2676 | instr = instr_create(ctx, 2, OPC_CMPS_U); |
||
2677 | instr->cat2.condition = IR3_COND_GE; |
||
2678 | vectorize(ctx, instr, &r_dst, 2, r_src, 0, b_src, 0); |
||
2679 | |||
2680 | if (type_uint(src_type)) { |
||
2681 | /* add.u dst, q, r */ |
||
2682 | instr = instr_create(ctx, 2, OPC_ADD_U); |
||
2683 | vectorize(ctx, instr, premod_dst, 2, q_src, 0, r_src, 0); |
||
2684 | } else { |
||
2685 | /* add.u q, q, r */ |
||
2686 | instr = instr_create(ctx, 2, OPC_ADD_U); |
||
2687 | vectorize(ctx, instr, &q_dst, 2, q_src, 0, r_src, 0); |
||
2688 | |||
2689 | /* negate result based on the original arguments */ |
||
2690 | if (is_const(a) && is_const(b)) |
||
2691 | a = get_unconst(ctx, a); |
||
2692 | |||
2693 | /* xor.b r, numerator, denominator */ |
||
2694 | instr = instr_create(ctx, 2, OPC_XOR_B); |
||
2695 | vectorize(ctx, instr, &r_dst, 2, a, 0, b, 0); |
||
2696 | |||
2697 | /* shr.b r, r, 31 */ |
||
2698 | instr = instr_create(ctx, 2, OPC_SHR_B); |
||
2699 | vectorize(ctx, instr, &r_dst, 2, r_src, 0, &thirty_one, 0); |
||
2700 | |||
2701 | /* absneg.s b, (neg)q */ |
||
2702 | instr = instr_create(ctx, 2, OPC_ABSNEG_S); |
||
2703 | vectorize(ctx, instr, &b_dst, 1, q_src, IR3_REG_SNEG); |
||
2704 | |||
2705 | /* sel.b dst, b, r, q */ |
||
2706 | instr = instr_create(ctx, 3, OPC_SEL_B32); |
||
2707 | vectorize(ctx, instr, premod_dst, 3, b_src, 0, r_src, 0, q_src, 0); |
||
2708 | } |
||
2709 | |||
2710 | if (t->tgsi_opc == TGSI_OPCODE_MOD || t->tgsi_opc == TGSI_OPCODE_UMOD) { |
||
2711 | /* The division result will have ended up in q. */ |
||
2712 | |||
2713 | if (is_rel_or_const(b)) |
||
2714 | b = get_unconst(ctx, b); |
||
2715 | |||
2716 | /* mull.u r, q, b */ |
||
2717 | instr = instr_create(ctx, 2, OPC_MULL_U); |
||
2718 | vectorize(ctx, instr, &r_dst, 2, q_src, 0, b, 0); |
||
2719 | |||
2720 | /* madsh.m16 r, q, b, r */ |
||
2721 | instr = instr_create(ctx, 3, OPC_MADSH_M16); |
||
2722 | vectorize(ctx, instr, &r_dst, 3, q_src, 0, b, 0, r_src, 0); |
||
2723 | |||
2724 | /* madsh.m16 r, b, q, r */ |
||
2725 | instr = instr_create(ctx, 3, OPC_MADSH_M16); |
||
2726 | vectorize(ctx, instr, &r_dst, 3, b, 0, q_src, 0, r_src, 0); |
||
2727 | |||
2728 | /* sub.u dst, a, r */ |
||
2729 | instr = instr_create(ctx, 2, OPC_SUB_U); |
||
2730 | vectorize(ctx, instr, dst, 2, a, 0, r_src, 0); |
||
2731 | } |
||
2732 | |||
2733 | put_dst(ctx, inst, dst); |
||
2734 | } |
||
2735 | |||
2736 | /* |
||
2737 | * Handlers for TGSI instructions which do have 1:1 mapping to native |
||
2738 | * instructions: |
||
2739 | */ |
||
2740 | |||
2741 | static void |
||
2742 | instr_cat0(const struct instr_translater *t, |
||
2743 | struct ir3_compile_context *ctx, |
||
2744 | struct tgsi_full_instruction *inst) |
||
2745 | { |
||
2746 | instr_create(ctx, 0, t->opc); |
||
2747 | } |
||
2748 | |||
2749 | static void |
||
2750 | instr_cat1(const struct instr_translater *t, |
||
2751 | struct ir3_compile_context *ctx, |
||
2752 | struct tgsi_full_instruction *inst) |
||
2753 | { |
||
2754 | struct tgsi_dst_register *dst = &inst->Dst[0].Register; |
||
2755 | struct tgsi_src_register *src = &inst->Src[0].Register; |
||
2756 | |||
2757 | /* NOTE: atomic start/end, rather than in create_mov() since |
||
2758 | * create_mov() is used already w/in atomic sequences (and |
||
2759 | * we aren't clever enough to deal with the nesting) |
||
2760 | */ |
||
2761 | instr_atomic_start(ctx); |
||
2762 | create_mov(ctx, dst, src); |
||
2763 | instr_atomic_end(ctx); |
||
2764 | } |
||
2765 | |||
2766 | static void |
||
2767 | instr_cat2(const struct instr_translater *t, |
||
2768 | struct ir3_compile_context *ctx, |
||
2769 | struct tgsi_full_instruction *inst) |
||
2770 | { |
||
2771 | struct tgsi_dst_register *dst = get_dst(ctx, inst); |
||
2772 | struct tgsi_src_register *src0 = &inst->Src[0].Register; |
||
2773 | struct tgsi_src_register *src1 = &inst->Src[1].Register; |
||
2774 | struct ir3_instruction *instr; |
||
2775 | unsigned src0_flags = 0, src1_flags = 0; |
||
2776 | |||
2777 | switch (t->tgsi_opc) { |
||
2778 | case TGSI_OPCODE_ABS: |
||
2779 | src0_flags = IR3_REG_FABS; |
||
2780 | break; |
||
2781 | case TGSI_OPCODE_IABS: |
||
2782 | src0_flags = IR3_REG_SABS; |
||
2783 | break; |
||
2784 | case TGSI_OPCODE_INEG: |
||
2785 | src0_flags = IR3_REG_SNEG; |
||
2786 | break; |
||
2787 | case TGSI_OPCODE_SUB: |
||
2788 | src1_flags = IR3_REG_FNEG; |
||
2789 | break; |
||
2790 | } |
||
2791 | |||
2792 | switch (t->opc) { |
||
2793 | case OPC_ABSNEG_F: |
||
2794 | case OPC_ABSNEG_S: |
||
2795 | case OPC_CLZ_B: |
||
2796 | case OPC_CLZ_S: |
||
2797 | case OPC_SIGN_F: |
||
2798 | case OPC_FLOOR_F: |
||
2799 | case OPC_CEIL_F: |
||
2800 | case OPC_RNDNE_F: |
||
2801 | case OPC_RNDAZ_F: |
||
2802 | case OPC_TRUNC_F: |
||
2803 | case OPC_NOT_B: |
||
2804 | case OPC_BFREV_B: |
||
2805 | case OPC_SETRM: |
||
2806 | case OPC_CBITS_B: |
||
2807 | /* these only have one src reg */ |
||
2808 | instr = instr_create(ctx, 2, t->opc); |
||
2809 | vectorize(ctx, instr, dst, 1, src0, src0_flags); |
||
2810 | break; |
||
2811 | default: |
||
2812 | if (is_const(src0) && is_const(src1)) |
||
2813 | src0 = get_unconst(ctx, src0); |
||
2814 | |||
2815 | instr = instr_create(ctx, 2, t->opc); |
||
2816 | vectorize(ctx, instr, dst, 2, src0, src0_flags, |
||
2817 | src1, src1_flags); |
||
2818 | break; |
||
2819 | } |
||
2820 | |||
2821 | put_dst(ctx, inst, dst); |
||
2822 | } |
||
2823 | |||
2824 | static void |
||
2825 | instr_cat3(const struct instr_translater *t, |
||
2826 | struct ir3_compile_context *ctx, |
||
2827 | struct tgsi_full_instruction *inst) |
||
2828 | { |
||
2829 | struct tgsi_dst_register *dst = get_dst(ctx, inst); |
||
2830 | struct tgsi_src_register *src0 = &inst->Src[0].Register; |
||
2831 | struct tgsi_src_register *src1 = &inst->Src[1].Register; |
||
2832 | struct ir3_instruction *instr; |
||
2833 | |||
2834 | /* in particular, can't handle const for src1 for cat3.. |
||
2835 | * for mad, we can swap first two src's if needed: |
||
2836 | */ |
||
2837 | if (is_rel_or_const(src1)) { |
||
2838 | if (is_mad(t->opc) && !is_rel_or_const(src0)) { |
||
2839 | struct tgsi_src_register *tmp; |
||
2840 | tmp = src0; |
||
2841 | src0 = src1; |
||
2842 | src1 = tmp; |
||
2843 | } else { |
||
2844 | src1 = get_unconst(ctx, src1); |
||
2845 | } |
||
2846 | } |
||
2847 | |||
2848 | instr = instr_create(ctx, 3, t->opc); |
||
2849 | vectorize(ctx, instr, dst, 3, src0, 0, src1, 0, |
||
2850 | &inst->Src[2].Register, 0); |
||
2851 | put_dst(ctx, inst, dst); |
||
2852 | } |
||
2853 | |||
2854 | static void |
||
2855 | instr_cat4(const struct instr_translater *t, |
||
2856 | struct ir3_compile_context *ctx, |
||
2857 | struct tgsi_full_instruction *inst) |
||
2858 | { |
||
2859 | struct tgsi_dst_register *dst = get_dst(ctx, inst); |
||
2860 | struct tgsi_src_register *src = &inst->Src[0].Register; |
||
2861 | struct ir3_instruction *instr; |
||
2862 | unsigned i; |
||
2863 | |||
2864 | /* seems like blob compiler avoids const as src.. */ |
||
2865 | if (is_const(src)) |
||
2866 | src = get_unconst(ctx, src); |
||
2867 | |||
2868 | /* we need to replicate into each component: */ |
||
2869 | for (i = 0; i < 4; i++) { |
||
2870 | if (dst->WriteMask & (1 << i)) { |
||
2871 | instr = instr_create(ctx, 4, t->opc); |
||
2872 | add_dst_reg(ctx, instr, dst, i); |
||
2873 | add_src_reg(ctx, instr, src, src->SwizzleX); |
||
2874 | } |
||
2875 | } |
||
2876 | |||
2877 | put_dst(ctx, inst, dst); |
||
2878 | } |
||
2879 | |||
2880 | static const struct instr_translater translaters[TGSI_OPCODE_LAST] = { |
||
2881 | #define INSTR(n, f, ...) \ |
||
2882 | [TGSI_OPCODE_ ## n] = { .fxn = (f), .tgsi_opc = TGSI_OPCODE_ ## n, ##__VA_ARGS__ } |
||
2883 | |||
2884 | INSTR(MOV, instr_cat1), |
||
2885 | INSTR(RCP, instr_cat4, .opc = OPC_RCP), |
||
2886 | INSTR(RSQ, instr_cat4, .opc = OPC_RSQ), |
||
2887 | INSTR(SQRT, instr_cat4, .opc = OPC_SQRT), |
||
2888 | INSTR(MUL, instr_cat2, .opc = OPC_MUL_F), |
||
2889 | INSTR(ADD, instr_cat2, .opc = OPC_ADD_F), |
||
2890 | INSTR(SUB, instr_cat2, .opc = OPC_ADD_F), |
||
2891 | INSTR(MIN, instr_cat2, .opc = OPC_MIN_F), |
||
2892 | INSTR(MAX, instr_cat2, .opc = OPC_MAX_F), |
||
2893 | INSTR(UADD, instr_cat2, .opc = OPC_ADD_U), |
||
2894 | INSTR(IMIN, instr_cat2, .opc = OPC_MIN_S), |
||
2895 | INSTR(UMIN, instr_cat2, .opc = OPC_MIN_U), |
||
2896 | INSTR(IMAX, instr_cat2, .opc = OPC_MAX_S), |
||
2897 | INSTR(UMAX, instr_cat2, .opc = OPC_MAX_U), |
||
2898 | INSTR(AND, instr_cat2, .opc = OPC_AND_B), |
||
2899 | INSTR(OR, instr_cat2, .opc = OPC_OR_B), |
||
2900 | INSTR(NOT, instr_cat2, .opc = OPC_NOT_B), |
||
2901 | INSTR(XOR, instr_cat2, .opc = OPC_XOR_B), |
||
2902 | INSTR(UMUL, trans_umul), |
||
2903 | INSTR(UMAD, trans_umul), |
||
2904 | INSTR(UDIV, trans_idiv), |
||
2905 | INSTR(IDIV, trans_idiv), |
||
2906 | INSTR(MOD, trans_idiv), |
||
2907 | INSTR(UMOD, trans_idiv), |
||
2908 | INSTR(SHL, instr_cat2, .opc = OPC_SHL_B), |
||
2909 | INSTR(USHR, instr_cat2, .opc = OPC_SHR_B), |
||
2910 | INSTR(ISHR, instr_cat2, .opc = OPC_ASHR_B), |
||
2911 | INSTR(IABS, instr_cat2, .opc = OPC_ABSNEG_S), |
||
2912 | INSTR(INEG, instr_cat2, .opc = OPC_ABSNEG_S), |
||
2913 | INSTR(AND, instr_cat2, .opc = OPC_AND_B), |
||
2914 | INSTR(MAD, instr_cat3, .opc = OPC_MAD_F32, .hopc = OPC_MAD_F16), |
||
2915 | INSTR(TRUNC, instr_cat2, .opc = OPC_TRUNC_F), |
||
2916 | INSTR(CLAMP, trans_clamp), |
||
2917 | INSTR(FLR, instr_cat2, .opc = OPC_FLOOR_F), |
||
2918 | INSTR(ROUND, instr_cat2, .opc = OPC_RNDNE_F), |
||
2919 | INSTR(SSG, instr_cat2, .opc = OPC_SIGN_F), |
||
2920 | INSTR(CEIL, instr_cat2, .opc = OPC_CEIL_F), |
||
2921 | INSTR(ARL, trans_arl), |
||
2922 | INSTR(UARL, trans_arl), |
||
2923 | INSTR(EX2, instr_cat4, .opc = OPC_EXP2), |
||
2924 | INSTR(LG2, instr_cat4, .opc = OPC_LOG2), |
||
2925 | INSTR(ABS, instr_cat2, .opc = OPC_ABSNEG_F), |
||
2926 | INSTR(COS, instr_cat4, .opc = OPC_COS), |
||
2927 | INSTR(SIN, instr_cat4, .opc = OPC_SIN), |
||
2928 | INSTR(TEX, trans_samp, .opc = OPC_SAM), |
||
2929 | INSTR(TXP, trans_samp, .opc = OPC_SAM), |
||
2930 | INSTR(TXB, trans_samp, .opc = OPC_SAMB), |
||
2931 | INSTR(TXB2, trans_samp, .opc = OPC_SAMB), |
||
2932 | INSTR(TXL, trans_samp, .opc = OPC_SAML), |
||
2933 | INSTR(TXD, trans_samp, .opc = OPC_SAMGQ), |
||
2934 | INSTR(TXF, trans_samp, .opc = OPC_ISAML), |
||
2935 | INSTR(TXQ, trans_txq), |
||
2936 | INSTR(DDX, trans_deriv, .opc = OPC_DSX), |
||
2937 | INSTR(DDY, trans_deriv, .opc = OPC_DSY), |
||
2938 | INSTR(SGT, trans_cmp), |
||
2939 | INSTR(SLT, trans_cmp), |
||
2940 | INSTR(FSLT, trans_cmp), |
||
2941 | INSTR(SGE, trans_cmp), |
||
2942 | INSTR(FSGE, trans_cmp), |
||
2943 | INSTR(SLE, trans_cmp), |
||
2944 | INSTR(SNE, trans_cmp), |
||
2945 | INSTR(FSNE, trans_cmp), |
||
2946 | INSTR(SEQ, trans_cmp), |
||
2947 | INSTR(FSEQ, trans_cmp), |
||
2948 | INSTR(CMP, trans_cmp), |
||
2949 | INSTR(USNE, trans_icmp, .opc = OPC_CMPS_U), |
||
2950 | INSTR(USEQ, trans_icmp, .opc = OPC_CMPS_U), |
||
2951 | INSTR(ISGE, trans_icmp, .opc = OPC_CMPS_S), |
||
2952 | INSTR(USGE, trans_icmp, .opc = OPC_CMPS_U), |
||
2953 | INSTR(ISLT, trans_icmp, .opc = OPC_CMPS_S), |
||
2954 | INSTR(USLT, trans_icmp, .opc = OPC_CMPS_U), |
||
2955 | INSTR(UCMP, trans_ucmp), |
||
2956 | INSTR(ISSG, trans_issg), |
||
2957 | INSTR(IF, trans_if, .opc = OPC_CMPS_F), |
||
2958 | INSTR(UIF, trans_if, .opc = OPC_CMPS_U), |
||
2959 | INSTR(ELSE, trans_else), |
||
2960 | INSTR(ENDIF, trans_endif), |
||
2961 | INSTR(END, instr_cat0, .opc = OPC_END), |
||
2962 | INSTR(KILL, trans_kill, .opc = OPC_KILL), |
||
2963 | INSTR(KILL_IF, trans_killif, .opc = OPC_KILL), |
||
2964 | INSTR(I2F, trans_cov), |
||
2965 | INSTR(U2F, trans_cov), |
||
2966 | INSTR(F2I, trans_cov), |
||
2967 | INSTR(F2U, trans_cov), |
||
2968 | }; |
||
2969 | |||
2970 | static ir3_semantic |
||
2971 | decl_semantic(const struct tgsi_declaration_semantic *sem) |
||
2972 | { |
||
2973 | return ir3_semantic_name(sem->Name, sem->Index); |
||
2974 | } |
||
2975 | |||
2976 | static struct ir3_instruction * |
||
2977 | decl_in_frag_bary(struct ir3_compile_context *ctx, unsigned regid, |
||
2978 | unsigned j, unsigned inloc, bool use_ldlv) |
||
2979 | { |
||
2980 | struct ir3_instruction *instr; |
||
2981 | struct ir3_register *src; |
||
2982 | |||
2983 | if (use_ldlv) { |
||
2984 | /* ldlv.u32 dst, l[#inloc], 1 */ |
||
2985 | instr = instr_create(ctx, 6, OPC_LDLV); |
||
2986 | instr->cat6.type = TYPE_U32; |
||
2987 | instr->cat6.iim_val = 1; |
||
2988 | ir3_reg_create(instr, regid, 0); /* dummy dst */ |
||
2989 | ir3_reg_create(instr, 0, IR3_REG_IMMED)->iim_val = inloc; |
||
2990 | ir3_reg_create(instr, 0, IR3_REG_IMMED)->iim_val = 1; |
||
2991 | |||
2992 | return instr; |
||
2993 | } |
||
2994 | |||
2995 | /* bary.f dst, #inloc, r0.x */ |
||
2996 | instr = instr_create(ctx, 2, OPC_BARY_F); |
||
2997 | ir3_reg_create(instr, regid, 0); /* dummy dst */ |
||
2998 | ir3_reg_create(instr, 0, IR3_REG_IMMED)->iim_val = inloc; |
||
2999 | src = ir3_reg_create(instr, 0, IR3_REG_SSA); |
||
3000 | src->wrmask = 0x3; |
||
3001 | src->instr = ctx->frag_pos; |
||
3002 | |||
3003 | return instr; |
||
3004 | } |
||
3005 | |||
3006 | /* TGSI_SEMANTIC_POSITION |
||
3007 | * """""""""""""""""""""" |
||
3008 | * |
||
3009 | * For fragment shaders, TGSI_SEMANTIC_POSITION is used to indicate that |
||
3010 | * fragment shader input contains the fragment's window position. The X |
||
3011 | * component starts at zero and always increases from left to right. |
||
3012 | * The Y component starts at zero and always increases but Y=0 may either |
||
3013 | * indicate the top of the window or the bottom depending on the fragment |
||
3014 | * coordinate origin convention (see TGSI_PROPERTY_FS_COORD_ORIGIN). |
||
3015 | * The Z coordinate ranges from 0 to 1 to represent depth from the front |
||
3016 | * to the back of the Z buffer. The W component contains the reciprocol |
||
3017 | * of the interpolated vertex position W component. |
||
3018 | */ |
||
3019 | static struct ir3_instruction * |
||
3020 | decl_in_frag_coord(struct ir3_compile_context *ctx, unsigned regid, |
||
3021 | unsigned j) |
||
3022 | { |
||
3023 | struct ir3_instruction *instr, *src; |
||
3024 | |||
3025 | compile_assert(ctx, !ctx->frag_coord[j]); |
||
3026 | |||
3027 | ctx->frag_coord[j] = create_input(ctx->block, NULL, 0); |
||
3028 | |||
3029 | |||
3030 | switch (j) { |
||
3031 | case 0: /* .x */ |
||
3032 | case 1: /* .y */ |
||
3033 | /* for frag_coord, we get unsigned values.. we need |
||
3034 | * to subtract (integer) 8 and divide by 16 (right- |
||
3035 | * shift by 4) then convert to float: |
||
3036 | */ |
||
3037 | |||
3038 | /* add.s tmp, src, -8 */ |
||
3039 | instr = instr_create(ctx, 2, OPC_ADD_S); |
||
3040 | ir3_reg_create(instr, regid, 0); /* dummy dst */ |
||
3041 | ir3_reg_create(instr, 0, IR3_REG_SSA)->instr = ctx->frag_coord[j]; |
||
3042 | ir3_reg_create(instr, 0, IR3_REG_IMMED)->iim_val = -8; |
||
3043 | src = instr; |
||
3044 | |||
3045 | /* shr.b tmp, tmp, 4 */ |
||
3046 | instr = instr_create(ctx, 2, OPC_SHR_B); |
||
3047 | ir3_reg_create(instr, regid, 0); /* dummy dst */ |
||
3048 | ir3_reg_create(instr, 0, IR3_REG_SSA)->instr = src; |
||
3049 | ir3_reg_create(instr, 0, IR3_REG_IMMED)->iim_val = 4; |
||
3050 | src = instr; |
||
3051 | |||
3052 | /* mov.u32f32 dst, tmp */ |
||
3053 | instr = instr_create(ctx, 1, 0); |
||
3054 | instr->cat1.src_type = TYPE_U32; |
||
3055 | instr->cat1.dst_type = TYPE_F32; |
||
3056 | ir3_reg_create(instr, regid, 0); /* dummy dst */ |
||
3057 | ir3_reg_create(instr, 0, IR3_REG_SSA)->instr = src; |
||
3058 | |||
3059 | break; |
||
3060 | case 2: /* .z */ |
||
3061 | case 3: /* .w */ |
||
3062 | /* seems that we can use these as-is: */ |
||
3063 | instr = ctx->frag_coord[j]; |
||
3064 | break; |
||
3065 | default: |
||
3066 | compile_error(ctx, "invalid channel\n"); |
||
3067 | instr = create_immed(ctx, 0.0); |
||
3068 | break; |
||
3069 | } |
||
3070 | |||
3071 | return instr; |
||
3072 | } |
||
3073 | |||
3074 | /* TGSI_SEMANTIC_FACE |
||
3075 | * """""""""""""""""" |
||
3076 | * |
||
3077 | * This label applies to fragment shader inputs only and indicates that |
||
3078 | * the register contains front/back-face information of the form (F, 0, |
||
3079 | * 0, 1). The first component will be positive when the fragment belongs |
||
3080 | * to a front-facing polygon, and negative when the fragment belongs to a |
||
3081 | * back-facing polygon. |
||
3082 | */ |
||
3083 | static struct ir3_instruction * |
||
3084 | decl_in_frag_face(struct ir3_compile_context *ctx, unsigned regid, |
||
3085 | unsigned j) |
||
3086 | { |
||
3087 | struct ir3_instruction *instr, *src; |
||
3088 | |||
3089 | switch (j) { |
||
3090 | case 0: /* .x */ |
||
3091 | compile_assert(ctx, !ctx->frag_face); |
||
3092 | |||
3093 | ctx->frag_face = create_input(ctx->block, NULL, 0); |
||
3094 | |||
3095 | /* for faceness, we always get -1 or 0 (int).. but TGSI expects |
||
3096 | * positive vs negative float.. and piglit further seems to |
||
3097 | * expect -1.0 or 1.0: |
||
3098 | * |
||
3099 | * mul.s tmp, hr0.x, 2 |
||
3100 | * add.s tmp, tmp, 1 |
||
3101 | * mov.s16f32, dst, tmp |
||
3102 | * |
||
3103 | */ |
||
3104 | |||
3105 | instr = instr_create(ctx, 2, OPC_MUL_S); |
||
3106 | ir3_reg_create(instr, regid, 0); /* dummy dst */ |
||
3107 | ir3_reg_create(instr, 0, IR3_REG_SSA)->instr = ctx->frag_face; |
||
3108 | ir3_reg_create(instr, 0, IR3_REG_IMMED)->iim_val = 2; |
||
3109 | src = instr; |
||
3110 | |||
3111 | instr = instr_create(ctx, 2, OPC_ADD_S); |
||
3112 | ir3_reg_create(instr, regid, 0); /* dummy dst */ |
||
3113 | ir3_reg_create(instr, 0, IR3_REG_SSA)->instr = src; |
||
3114 | ir3_reg_create(instr, 0, IR3_REG_IMMED)->iim_val = 1; |
||
3115 | src = instr; |
||
3116 | |||
3117 | instr = instr_create(ctx, 1, 0); /* mov */ |
||
3118 | instr->cat1.src_type = TYPE_S32; |
||
3119 | instr->cat1.dst_type = TYPE_F32; |
||
3120 | ir3_reg_create(instr, regid, 0); /* dummy dst */ |
||
3121 | ir3_reg_create(instr, 0, IR3_REG_SSA)->instr = src; |
||
3122 | |||
3123 | break; |
||
3124 | case 1: /* .y */ |
||
3125 | case 2: /* .z */ |
||
3126 | instr = create_immed(ctx, 0.0); |
||
3127 | break; |
||
3128 | case 3: /* .w */ |
||
3129 | instr = create_immed(ctx, 1.0); |
||
3130 | break; |
||
3131 | default: |
||
3132 | compile_error(ctx, "invalid channel\n"); |
||
3133 | instr = create_immed(ctx, 0.0); |
||
3134 | break; |
||
3135 | } |
||
3136 | |||
3137 | return instr; |
||
3138 | } |
||
3139 | |||
3140 | static void |
||
3141 | decl_in(struct ir3_compile_context *ctx, struct tgsi_full_declaration *decl) |
||
3142 | { |
||
3143 | struct ir3_shader_variant *so = ctx->so; |
||
3144 | unsigned name = decl->Semantic.Name; |
||
3145 | unsigned i; |
||
3146 | |||
3147 | /* I don't think we should get frag shader input without |
||
3148 | * semantic info? Otherwise how do inputs get linked to |
||
3149 | * vert outputs? |
||
3150 | */ |
||
3151 | compile_assert(ctx, (ctx->type == TGSI_PROCESSOR_VERTEX) || |
||
3152 | decl->Declaration.Semantic); |
||
3153 | |||
3154 | for (i = decl->Range.First; i <= decl->Range.Last; i++) { |
||
3155 | unsigned n = so->inputs_count++; |
||
3156 | unsigned r = regid(i, 0); |
||
3157 | unsigned ncomp, j; |
||
3158 | |||
3159 | /* we'll figure out the actual components used after scheduling */ |
||
3160 | ncomp = 4; |
||
3161 | |||
3162 | DBG("decl in -> r%d", i); |
||
3163 | |||
3164 | compile_assert(ctx, n < ARRAY_SIZE(so->inputs)); |
||
3165 | |||
3166 | so->inputs[n].semantic = decl_semantic(&decl->Semantic); |
||
3167 | so->inputs[n].compmask = (1 << ncomp) - 1; |
||
3168 | so->inputs[n].regid = r; |
||
3169 | so->inputs[n].inloc = ctx->next_inloc; |
||
3170 | so->inputs[n].interpolate = decl->Interp.Interpolate; |
||
3171 | |||
3172 | for (j = 0; j < ncomp; j++) { |
||
3173 | struct ir3_instruction *instr = NULL; |
||
3174 | |||
3175 | if (ctx->type == TGSI_PROCESSOR_FRAGMENT) { |
||
3176 | /* for fragment shaders, POSITION and FACE are handled |
||
3177 | * specially, not using normal varying / bary.f |
||
3178 | */ |
||
3179 | if (name == TGSI_SEMANTIC_POSITION) { |
||
3180 | so->inputs[n].bary = false; |
||
3181 | so->frag_coord = true; |
||
3182 | instr = decl_in_frag_coord(ctx, r + j, j); |
||
3183 | } else if (name == TGSI_SEMANTIC_FACE) { |
||
3184 | so->inputs[n].bary = false; |
||
3185 | so->frag_face = true; |
||
3186 | instr = decl_in_frag_face(ctx, r + j, j); |
||
3187 | } else { |
||
3188 | bool use_ldlv = false; |
||
3189 | |||
3190 | /* if no interpolation given, pick based on |
||
3191 | * semantic: |
||
3192 | */ |
||
3193 | if (!decl->Declaration.Interpolate) { |
||
3194 | switch (decl->Semantic.Name) { |
||
3195 | case TGSI_SEMANTIC_COLOR: |
||
3196 | so->inputs[n].interpolate = |
||
3197 | TGSI_INTERPOLATE_COLOR; |
||
3198 | break; |
||
3199 | default: |
||
3200 | so->inputs[n].interpolate = |
||
3201 | TGSI_INTERPOLATE_LINEAR; |
||
3202 | } |
||
3203 | } |
||
3204 | |||
3205 | if (ctx->flat_bypass) { |
||
3206 | switch (so->inputs[n].interpolate) { |
||
3207 | case TGSI_INTERPOLATE_COLOR: |
||
3208 | if (!ctx->so->key.rasterflat) |
||
3209 | break; |
||
3210 | /* fallthrough */ |
||
3211 | case TGSI_INTERPOLATE_CONSTANT: |
||
3212 | use_ldlv = true; |
||
3213 | break; |
||
3214 | } |
||
3215 | } |
||
3216 | |||
3217 | so->inputs[n].bary = true; |
||
3218 | |||
3219 | instr = decl_in_frag_bary(ctx, r + j, j, |
||
3220 | so->inputs[n].inloc + j - 8, use_ldlv); |
||
3221 | } |
||
3222 | } else { |
||
3223 | instr = create_input(ctx->block, NULL, (i * 4) + j); |
||
3224 | } |
||
3225 | |||
3226 | ctx->block->inputs[(i * 4) + j] = instr; |
||
3227 | } |
||
3228 | |||
3229 | if (so->inputs[n].bary || (ctx->type == TGSI_PROCESSOR_VERTEX)) { |
||
3230 | ctx->next_inloc += ncomp; |
||
3231 | so->total_in += ncomp; |
||
3232 | } |
||
3233 | } |
||
3234 | } |
||
3235 | |||
3236 | static void |
||
3237 | decl_sv(struct ir3_compile_context *ctx, struct tgsi_full_declaration *decl) |
||
3238 | { |
||
3239 | struct ir3_shader_variant *so = ctx->so; |
||
3240 | unsigned r = regid(so->inputs_count, 0); |
||
3241 | unsigned n = so->inputs_count++; |
||
3242 | |||
3243 | DBG("decl sv -> r%d", n); |
||
3244 | |||
3245 | compile_assert(ctx, n < ARRAY_SIZE(so->inputs)); |
||
3246 | compile_assert(ctx, decl->Range.First < ARRAY_SIZE(ctx->sysval_semantics)); |
||
3247 | |||
3248 | ctx->sysval_semantics[decl->Range.First] = decl->Semantic.Name; |
||
3249 | so->inputs[n].semantic = decl_semantic(&decl->Semantic); |
||
3250 | so->inputs[n].compmask = 1; |
||
3251 | so->inputs[n].regid = r; |
||
3252 | so->inputs[n].inloc = ctx->next_inloc; |
||
3253 | so->inputs[n].interpolate = TGSI_INTERPOLATE_CONSTANT; |
||
3254 | |||
3255 | struct ir3_instruction *instr = NULL; |
||
3256 | |||
3257 | switch (decl->Semantic.Name) { |
||
3258 | case TGSI_SEMANTIC_VERTEXID_NOBASE: |
||
3259 | ctx->vertex_id = instr = create_input(ctx->block, NULL, r); |
||
3260 | break; |
||
3261 | case TGSI_SEMANTIC_BASEVERTEX: |
||
3262 | ctx->basevertex = instr = instr_create(ctx, 1, 0); |
||
3263 | instr->cat1.src_type = get_stype(ctx); |
||
3264 | instr->cat1.dst_type = get_stype(ctx); |
||
3265 | ir3_reg_create(instr, 0, 0); |
||
3266 | ir3_reg_create(instr, regid(so->first_driver_param + 4, 0), |
||
3267 | IR3_REG_CONST); |
||
3268 | break; |
||
3269 | case TGSI_SEMANTIC_INSTANCEID: |
||
3270 | ctx->instance_id = instr = create_input(ctx->block, NULL, r); |
||
3271 | break; |
||
3272 | default: |
||
3273 | compile_error(ctx, "Unknown semantic: %s\n", |
||
3274 | tgsi_semantic_names[decl->Semantic.Name]); |
||
3275 | } |
||
3276 | |||
3277 | ctx->block->inputs[r] = instr; |
||
3278 | ctx->next_inloc++; |
||
3279 | so->total_in++; |
||
3280 | } |
||
3281 | |||
3282 | static void |
||
3283 | decl_out(struct ir3_compile_context *ctx, struct tgsi_full_declaration *decl) |
||
3284 | { |
||
3285 | struct ir3_shader_variant *so = ctx->so; |
||
3286 | unsigned comp = 0; |
||
3287 | unsigned name = decl->Semantic.Name; |
||
3288 | unsigned i; |
||
3289 | |||
3290 | compile_assert(ctx, decl->Declaration.Semantic); |
||
3291 | |||
3292 | DBG("decl out[%d] -> r%d", name, decl->Range.First); |
||
3293 | |||
3294 | if (ctx->type == TGSI_PROCESSOR_VERTEX) { |
||
3295 | switch (name) { |
||
3296 | case TGSI_SEMANTIC_POSITION: |
||
3297 | so->writes_pos = true; |
||
3298 | break; |
||
3299 | case TGSI_SEMANTIC_PSIZE: |
||
3300 | so->writes_psize = true; |
||
3301 | break; |
||
3302 | case TGSI_SEMANTIC_COLOR: |
||
3303 | case TGSI_SEMANTIC_BCOLOR: |
||
3304 | case TGSI_SEMANTIC_GENERIC: |
||
3305 | case TGSI_SEMANTIC_FOG: |
||
3306 | case TGSI_SEMANTIC_TEXCOORD: |
||
3307 | break; |
||
3308 | default: |
||
3309 | compile_error(ctx, "unknown VS semantic name: %s\n", |
||
3310 | tgsi_semantic_names[name]); |
||
3311 | } |
||
3312 | } else { |
||
3313 | switch (name) { |
||
3314 | case TGSI_SEMANTIC_POSITION: |
||
3315 | comp = 2; /* tgsi will write to .z component */ |
||
3316 | so->writes_pos = true; |
||
3317 | break; |
||
3318 | case TGSI_SEMANTIC_COLOR: |
||
3319 | break; |
||
3320 | default: |
||
3321 | compile_error(ctx, "unknown FS semantic name: %s\n", |
||
3322 | tgsi_semantic_names[name]); |
||
3323 | } |
||
3324 | } |
||
3325 | |||
3326 | for (i = decl->Range.First; i <= decl->Range.Last; i++) { |
||
3327 | unsigned n = so->outputs_count++; |
||
3328 | unsigned ncomp, j; |
||
3329 | |||
3330 | ncomp = 4; |
||
3331 | |||
3332 | compile_assert(ctx, n < ARRAY_SIZE(so->outputs)); |
||
3333 | |||
3334 | so->outputs[n].semantic = decl_semantic(&decl->Semantic); |
||
3335 | so->outputs[n].regid = regid(i, comp); |
||
3336 | |||
3337 | /* avoid undefined outputs, stick a dummy mov from imm{0.0}, |
||
3338 | * which if the output is actually assigned will be over- |
||
3339 | * written |
||
3340 | */ |
||
3341 | for (j = 0; j < ncomp; j++) |
||
3342 | ctx->block->outputs[(i * 4) + j] = create_immed(ctx, 0.0); |
||
3343 | } |
||
3344 | } |
||
3345 | |||
3346 | /* from TGSI perspective, we actually have inputs. But most of the "inputs" |
||
3347 | * for a fragment shader are just bary.f instructions. The *actual* inputs |
||
3348 | * from the hw perspective are the frag_pos and optionally frag_coord and |
||
3349 | * frag_face. |
||
3350 | */ |
||
3351 | static void |
||
3352 | fixup_frag_inputs(struct ir3_compile_context *ctx) |
||
3353 | { |
||
3354 | struct ir3_shader_variant *so = ctx->so; |
||
3355 | struct ir3_block *block = ctx->block; |
||
3356 | struct ir3_instruction **inputs; |
||
3357 | struct ir3_instruction *instr; |
||
3358 | int n, regid = 0; |
||
3359 | |||
3360 | block->ninputs = 0; |
||
3361 | |||
3362 | n = 4; /* always have frag_pos */ |
||
3363 | n += COND(so->frag_face, 4); |
||
3364 | n += COND(so->frag_coord, 4); |
||
3365 | |||
3366 | inputs = ir3_alloc(ctx->ir, n * (sizeof(struct ir3_instruction *))); |
||
3367 | |||
3368 | if (so->frag_face) { |
||
3369 | /* this ultimately gets assigned to hr0.x so doesn't conflict |
||
3370 | * with frag_coord/frag_pos.. |
||
3371 | */ |
||
3372 | inputs[block->ninputs++] = ctx->frag_face; |
||
3373 | ctx->frag_face->regs[0]->num = 0; |
||
3374 | |||
3375 | /* remaining channels not used, but let's avoid confusing |
||
3376 | * other parts that expect inputs to come in groups of vec4 |
||
3377 | */ |
||
3378 | inputs[block->ninputs++] = NULL; |
||
3379 | inputs[block->ninputs++] = NULL; |
||
3380 | inputs[block->ninputs++] = NULL; |
||
3381 | } |
||
3382 | |||
3383 | /* since we don't know where to set the regid for frag_coord, |
||
3384 | * we have to use r0.x for it. But we don't want to *always* |
||
3385 | * use r1.x for frag_pos as that could increase the register |
||
3386 | * footprint on simple shaders: |
||
3387 | */ |
||
3388 | if (so->frag_coord) { |
||
3389 | ctx->frag_coord[0]->regs[0]->num = regid++; |
||
3390 | ctx->frag_coord[1]->regs[0]->num = regid++; |
||
3391 | ctx->frag_coord[2]->regs[0]->num = regid++; |
||
3392 | ctx->frag_coord[3]->regs[0]->num = regid++; |
||
3393 | |||
3394 | inputs[block->ninputs++] = ctx->frag_coord[0]; |
||
3395 | inputs[block->ninputs++] = ctx->frag_coord[1]; |
||
3396 | inputs[block->ninputs++] = ctx->frag_coord[2]; |
||
3397 | inputs[block->ninputs++] = ctx->frag_coord[3]; |
||
3398 | } |
||
3399 | |||
3400 | /* we always have frag_pos: */ |
||
3401 | so->pos_regid = regid; |
||
3402 | |||
3403 | /* r0.x */ |
||
3404 | instr = create_input(block, NULL, block->ninputs); |
||
3405 | instr->regs[0]->num = regid++; |
||
3406 | inputs[block->ninputs++] = instr; |
||
3407 | ctx->frag_pos->regs[1]->instr = instr; |
||
3408 | |||
3409 | /* r0.y */ |
||
3410 | instr = create_input(block, NULL, block->ninputs); |
||
3411 | instr->regs[0]->num = regid++; |
||
3412 | inputs[block->ninputs++] = instr; |
||
3413 | ctx->frag_pos->regs[2]->instr = instr; |
||
3414 | |||
3415 | block->inputs = inputs; |
||
3416 | } |
||
3417 | |||
3418 | static void |
||
3419 | compile_instructions(struct ir3_compile_context *ctx) |
||
3420 | { |
||
3421 | push_block(ctx); |
||
3422 | |||
3423 | /* for fragment shader, we have a single input register (usually |
||
3424 | * r0.xy) which is used as the base for bary.f varying fetch instrs: |
||
3425 | */ |
||
3426 | if (ctx->type == TGSI_PROCESSOR_FRAGMENT) { |
||
3427 | struct ir3_instruction *instr; |
||
3428 | instr = ir3_instr_create(ctx->block, -1, OPC_META_FI); |
||
3429 | ir3_reg_create(instr, 0, 0); |
||
3430 | ir3_reg_create(instr, 0, IR3_REG_SSA); /* r0.x */ |
||
3431 | ir3_reg_create(instr, 0, IR3_REG_SSA); /* r0.y */ |
||
3432 | ctx->frag_pos = instr; |
||
3433 | } |
||
3434 | |||
3435 | while (!tgsi_parse_end_of_tokens(&ctx->parser)) { |
||
3436 | tgsi_parse_token(&ctx->parser); |
||
3437 | |||
3438 | switch (ctx->parser.FullToken.Token.Type) { |
||
3439 | case TGSI_TOKEN_TYPE_DECLARATION: { |
||
3440 | struct tgsi_full_declaration *decl = |
||
3441 | &ctx->parser.FullToken.FullDeclaration; |
||
3442 | unsigned file = decl->Declaration.File; |
||
3443 | if (file == TGSI_FILE_OUTPUT) { |
||
3444 | decl_out(ctx, decl); |
||
3445 | } else if (file == TGSI_FILE_INPUT) { |
||
3446 | decl_in(ctx, decl); |
||
3447 | } else if (decl->Declaration.File == TGSI_FILE_SYSTEM_VALUE) { |
||
3448 | decl_sv(ctx, decl); |
||
3449 | } |
||
3450 | |||
3451 | if ((file != TGSI_FILE_CONSTANT) && decl->Declaration.Array) { |
||
3452 | int aid = decl->Array.ArrayID + ctx->array_offsets[file]; |
||
3453 | |||
3454 | compile_assert(ctx, aid < ARRAY_SIZE(ctx->array)); |
||
3455 | |||
3456 | /* legacy ArrayID==0 stuff probably isn't going to work |
||
3457 | * well (and is at least untested).. let's just scream: |
||
3458 | */ |
||
3459 | compile_assert(ctx, aid != 0); |
||
3460 | |||
3461 | ctx->array[aid].first = decl->Range.First; |
||
3462 | ctx->array[aid].last = decl->Range.Last; |
||
3463 | } |
||
3464 | break; |
||
3465 | } |
||
3466 | case TGSI_TOKEN_TYPE_IMMEDIATE: { |
||
3467 | /* TODO: if we know the immediate is small enough, and only |
||
3468 | * used with instructions that can embed an immediate, we |
||
3469 | * can skip this: |
||
3470 | */ |
||
3471 | struct tgsi_full_immediate *imm = |
||
3472 | &ctx->parser.FullToken.FullImmediate; |
||
3473 | unsigned n = ctx->so->immediates_count++; |
||
3474 | compile_assert(ctx, n < ARRAY_SIZE(ctx->so->immediates)); |
||
3475 | memcpy(ctx->so->immediates[n].val, imm->u, 16); |
||
3476 | break; |
||
3477 | } |
||
3478 | case TGSI_TOKEN_TYPE_INSTRUCTION: { |
||
3479 | struct tgsi_full_instruction *inst = |
||
3480 | &ctx->parser.FullToken.FullInstruction; |
||
3481 | unsigned opc = inst->Instruction.Opcode; |
||
3482 | const struct instr_translater *t = &translaters[opc]; |
||
3483 | |||
3484 | if (t->fxn) { |
||
3485 | t->fxn(t, ctx, inst); |
||
3486 | ctx->num_internal_temps = 0; |
||
3487 | |||
3488 | compile_assert(ctx, !ctx->using_tmp_dst); |
||
3489 | } else { |
||
3490 | compile_error(ctx, "unknown TGSI opc: %s\n", |
||
3491 | tgsi_get_opcode_name(opc)); |
||
3492 | } |
||
3493 | |||
3494 | switch (inst->Instruction.Saturate) { |
||
3495 | case TGSI_SAT_ZERO_ONE: |
||
3496 | create_clamp_imm(ctx, &inst->Dst[0].Register, |
||
3497 | fui(0.0), fui(1.0)); |
||
3498 | break; |
||
3499 | case TGSI_SAT_MINUS_PLUS_ONE: |
||
3500 | create_clamp_imm(ctx, &inst->Dst[0].Register, |
||
3501 | fui(-1.0), fui(1.0)); |
||
3502 | break; |
||
3503 | } |
||
3504 | |||
3505 | instr_finish(ctx); |
||
3506 | |||
3507 | break; |
||
3508 | } |
||
3509 | case TGSI_TOKEN_TYPE_PROPERTY: { |
||
3510 | struct tgsi_full_property *prop = |
||
3511 | &ctx->parser.FullToken.FullProperty; |
||
3512 | switch (prop->Property.PropertyName) { |
||
3513 | case TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS: |
||
3514 | ctx->so->color0_mrt = !!prop->u[0].Data; |
||
3515 | break; |
||
3516 | } |
||
3517 | } |
||
3518 | default: |
||
3519 | break; |
||
3520 | } |
||
3521 | } |
||
3522 | } |
||
3523 | |||
3524 | static void |
||
3525 | compile_dump(struct ir3_compile_context *ctx) |
||
3526 | { |
||
3527 | const char *name = (ctx->so->type == SHADER_VERTEX) ? "vert" : "frag"; |
||
3528 | static unsigned n = 0; |
||
3529 | char fname[16]; |
||
3530 | FILE *f; |
||
3531 | snprintf(fname, sizeof(fname), "%s-%04u.dot", name, n++); |
||
3532 | f = fopen(fname, "w"); |
||
3533 | if (!f) |
||
3534 | return; |
||
3535 | ir3_block_depth(ctx->block); |
||
3536 | ir3_dump(ctx->ir, name, ctx->block, f); |
||
3537 | fclose(f); |
||
3538 | } |
||
3539 | |||
3540 | int |
||
3541 | ir3_compile_shader(struct ir3_shader_variant *so, |
||
3542 | const struct tgsi_token *tokens, struct ir3_shader_key key, |
||
3543 | bool cp) |
||
3544 | { |
||
3545 | struct ir3_compile_context ctx; |
||
3546 | struct ir3_block *block; |
||
3547 | struct ir3_instruction **inputs; |
||
3548 | unsigned i, j, actual_in; |
||
3549 | int ret = 0, max_bary; |
||
3550 | |||
3551 | assert(!so->ir); |
||
3552 | |||
3553 | so->ir = ir3_create(); |
||
3554 | |||
3555 | assert(so->ir); |
||
3556 | |||
3557 | if (compile_init(&ctx, so, tokens) != TGSI_PARSE_OK) { |
||
3558 | DBG("INIT failed!"); |
||
3559 | ret = -1; |
||
3560 | goto out; |
||
3561 | } |
||
3562 | |||
3563 | /* for now, until the edge cases are worked out: */ |
||
3564 | if (ctx.info.indirect_files_written & (FM(TEMPORARY) | FM(INPUT) | FM(OUTPUT))) |
||
3565 | cp = false; |
||
3566 | |||
3567 | compile_instructions(&ctx); |
||
3568 | |||
3569 | block = ctx.block; |
||
3570 | so->ir->block = block; |
||
3571 | |||
3572 | /* keep track of the inputs from TGSI perspective.. */ |
||
3573 | inputs = block->inputs; |
||
3574 | |||
3575 | /* but fixup actual inputs for frag shader: */ |
||
3576 | if (ctx.type == TGSI_PROCESSOR_FRAGMENT) |
||
3577 | fixup_frag_inputs(&ctx); |
||
3578 | |||
3579 | /* at this point, for binning pass, throw away unneeded outputs: */ |
||
3580 | if (key.binning_pass) { |
||
3581 | for (i = 0, j = 0; i < so->outputs_count; i++) { |
||
3582 | unsigned name = sem2name(so->outputs[i].semantic); |
||
3583 | unsigned idx = sem2idx(so->outputs[i].semantic); |
||
3584 | |||
3585 | /* throw away everything but first position/psize */ |
||
3586 | if ((idx == 0) && ((name == TGSI_SEMANTIC_POSITION) || |
||
3587 | (name == TGSI_SEMANTIC_PSIZE))) { |
||
3588 | if (i != j) { |
||
3589 | so->outputs[j] = so->outputs[i]; |
||
3590 | block->outputs[(j*4)+0] = block->outputs[(i*4)+0]; |
||
3591 | block->outputs[(j*4)+1] = block->outputs[(i*4)+1]; |
||
3592 | block->outputs[(j*4)+2] = block->outputs[(i*4)+2]; |
||
3593 | block->outputs[(j*4)+3] = block->outputs[(i*4)+3]; |
||
3594 | } |
||
3595 | j++; |
||
3596 | } |
||
3597 | } |
||
3598 | so->outputs_count = j; |
||
3599 | block->noutputs = j * 4; |
||
3600 | } |
||
3601 | |||
3602 | /* if we want half-precision outputs, mark the output registers |
||
3603 | * as half: |
||
3604 | */ |
||
3605 | if (key.half_precision) { |
||
3606 | for (i = 0; i < block->noutputs; i++) { |
||
3607 | if (!block->outputs[i]) |
||
3608 | continue; |
||
3609 | block->outputs[i]->regs[0]->flags |= IR3_REG_HALF; |
||
3610 | } |
||
3611 | } |
||
3612 | |||
3613 | /* at this point, we want the kill's in the outputs array too, |
||
3614 | * so that they get scheduled (since they have no dst).. we've |
||
3615 | * already ensured that the array is big enough in push_block(): |
||
3616 | */ |
||
3617 | if (ctx.type == TGSI_PROCESSOR_FRAGMENT) { |
||
3618 | for (i = 0; i < ctx.kill_count; i++) |
||
3619 | block->outputs[block->noutputs++] = ctx.kill[i]; |
||
3620 | } |
||
3621 | |||
3622 | if (fd_mesa_debug & FD_DBG_OPTDUMP) |
||
3623 | compile_dump(&ctx); |
||
3624 | |||
3625 | ret = ir3_block_flatten(block); |
||
3626 | if (ret < 0) { |
||
3627 | DBG("FLATTEN failed!"); |
||
3628 | goto out; |
||
3629 | } |
||
3630 | if ((ret > 0) && (fd_mesa_debug & FD_DBG_OPTDUMP)) |
||
3631 | compile_dump(&ctx); |
||
3632 | |||
3633 | if (fd_mesa_debug & FD_DBG_OPTMSGS) { |
||
3634 | printf("BEFORE CP:\n"); |
||
3635 | ir3_dump_instr_list(block->head); |
||
3636 | } |
||
3637 | |||
3638 | ir3_block_depth(block); |
||
3639 | |||
3640 | /* First remove all the extra mov's (which we could skip if the |
||
3641 | * front-end was clever enough not to insert them in the first |
||
3642 | * place). Then figure out left/right neighbors, re-inserting |
||
3643 | * extra mov's when needed to avoid conflicts. |
||
3644 | */ |
||
3645 | if (cp && !(fd_mesa_debug & FD_DBG_NOCP)) |
||
3646 | ir3_block_cp(block); |
||
3647 | |||
3648 | if (fd_mesa_debug & FD_DBG_OPTMSGS) { |
||
3649 | printf("BEFORE GROUPING:\n"); |
||
3650 | ir3_dump_instr_list(block->head); |
||
3651 | } |
||
3652 | |||
3653 | /* Group left/right neighbors, inserting mov's where needed to |
||
3654 | * solve conflicts: |
||
3655 | */ |
||
3656 | ir3_block_group(block); |
||
3657 | |||
3658 | if (fd_mesa_debug & FD_DBG_OPTDUMP) |
||
3659 | compile_dump(&ctx); |
||
3660 | |||
3661 | ir3_block_depth(block); |
||
3662 | |||
3663 | if (fd_mesa_debug & FD_DBG_OPTMSGS) { |
||
3664 | printf("AFTER DEPTH:\n"); |
||
3665 | ir3_dump_instr_list(block->head); |
||
3666 | } |
||
3667 | |||
3668 | ret = ir3_block_sched(block); |
||
3669 | if (ret) { |
||
3670 | DBG("SCHED failed!"); |
||
3671 | goto out; |
||
3672 | } |
||
3673 | |||
3674 | if (fd_mesa_debug & FD_DBG_OPTMSGS) { |
||
3675 | printf("AFTER SCHED:\n"); |
||
3676 | ir3_dump_instr_list(block->head); |
||
3677 | } |
||
3678 | |||
3679 | ret = ir3_block_ra(block, so->type, so->frag_coord, so->frag_face); |
||
3680 | if (ret) { |
||
3681 | DBG("RA failed!"); |
||
3682 | goto out; |
||
3683 | } |
||
3684 | |||
3685 | if (fd_mesa_debug & FD_DBG_OPTMSGS) { |
||
3686 | printf("AFTER RA:\n"); |
||
3687 | ir3_dump_instr_list(block->head); |
||
3688 | } |
||
3689 | |||
3690 | ir3_block_legalize(block, &so->has_samp, &max_bary); |
||
3691 | |||
3692 | /* fixup input/outputs: */ |
||
3693 | for (i = 0; i < so->outputs_count; i++) { |
||
3694 | so->outputs[i].regid = block->outputs[i*4]->regs[0]->num; |
||
3695 | /* preserve hack for depth output.. tgsi writes depth to .z, |
||
3696 | * but what we give the hw is the scalar register: |
||
3697 | */ |
||
3698 | if ((ctx.type == TGSI_PROCESSOR_FRAGMENT) && |
||
3699 | (sem2name(so->outputs[i].semantic) == TGSI_SEMANTIC_POSITION)) |
||
3700 | so->outputs[i].regid += 2; |
||
3701 | } |
||
3702 | /* Note that some or all channels of an input may be unused: */ |
||
3703 | actual_in = 0; |
||
3704 | for (i = 0; i < so->inputs_count; i++) { |
||
3705 | unsigned j, regid = ~0, compmask = 0; |
||
3706 | so->inputs[i].ncomp = 0; |
||
3707 | for (j = 0; j < 4; j++) { |
||
3708 | struct ir3_instruction *in = inputs[(i*4) + j]; |
||
3709 | if (in) { |
||
3710 | compmask |= (1 << j); |
||
3711 | regid = in->regs[0]->num - j; |
||
3712 | actual_in++; |
||
3713 | so->inputs[i].ncomp++; |
||
3714 | } |
||
3715 | } |
||
3716 | so->inputs[i].regid = regid; |
||
3717 | so->inputs[i].compmask = compmask; |
||
3718 | } |
||
3719 | |||
3720 | /* fragment shader always gets full vec4's even if it doesn't |
||
3721 | * fetch all components, but vertex shader we need to update |
||
3722 | * with the actual number of components fetch, otherwise thing |
||
3723 | * will hang due to mismaptch between VFD_DECODE's and |
||
3724 | * TOTALATTRTOVS |
||
3725 | */ |
||
3726 | if (so->type == SHADER_VERTEX) |
||
3727 | so->total_in = actual_in; |
||
3728 | else |
||
3729 | so->total_in = align(max_bary + 1, 4); |
||
3730 | |||
3731 | out: |
||
3732 | if (ret) { |
||
3733 | ir3_destroy(so->ir); |
||
3734 | so->ir = NULL; |
||
3735 | } |
||
3736 | compile_free(&ctx); |
||
3737 | |||
3738 | return ret; |
||
3739 | }><>>>>>>>>>>>>=>>>>><>>=>><>>><>><>><>><>>>>>>>>>=>><>><>>>><>><>>>>>>><>><>>><>>>>>>=>>><>>>><>>>><>>>>><>><>>><>>><>>><>>=>><>>=>>>>>><>>><>><> |