projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
pan/mdg: Add pack_unorm_4x8 via 8-bit
[mesa.git] / src / panfrost / midgard / midgard_compile.c
1 /*
2 * Copyright (C) 2018-2019 Alyssa Rosenzweig <alyssa@rosenzweig.io>
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 */
23
24 #include <sys/types.h>
25 #include <sys/stat.h>
26 #include <sys/mman.h>
27 #include <fcntl.h>
28 #include <stdint.h>
29 #include <stdlib.h>
30 #include <stdio.h>
31 #include <err.h>
32
33 #include "main/mtypes.h"
34 #include "compiler/glsl/glsl_to_nir.h"
35 #include "compiler/nir_types.h"
36 #include "compiler/nir/nir_builder.h"
37 #include "util/half_float.h"
38 #include "util/u_math.h"
39 #include "util/u_debug.h"
40 #include "util/u_dynarray.h"
41 #include "util/list.h"
42 #include "main/mtypes.h"
43
44 #include "midgard.h"
45 #include "midgard_nir.h"
46 #include "midgard_compile.h"
47 #include "midgard_ops.h"
48 #include "helpers.h"
49 #include "compiler.h"
50 #include "midgard_quirks.h"
51
52 #include "disassemble.h"
53
54 static const struct debug_named_value debug_options[] = {
55 {"msgs", MIDGARD_DBG_MSGS, "Print debug messages"},
56 {"shaders", MIDGARD_DBG_SHADERS, "Dump shaders in NIR and MIR"},
57 {"shaderdb", MIDGARD_DBG_SHADERDB, "Prints shader-db statistics"},
58 DEBUG_NAMED_VALUE_END
59 };
60
61 DEBUG_GET_ONCE_FLAGS_OPTION(midgard_debug, "MIDGARD_MESA_DEBUG", debug_options, 0)
62
63 unsigned SHADER_DB_COUNT = 0;
64
65 int midgard_debug = 0;
66
67 #define DBG(fmt, ...) \
68 do { if (midgard_debug & MIDGARD_DBG_MSGS) \
69 fprintf(stderr, "%s:%d: "fmt, \
70 __FUNCTION__, __LINE__, ##__VA_ARGS__); } while (0)
71 static midgard_block *
72 create_empty_block(compiler_context *ctx)
73 {
74 midgard_block *blk = rzalloc(ctx, midgard_block);
75
76 blk->base.predecessors = _mesa_set_create(blk,
77 _mesa_hash_pointer,
78 _mesa_key_pointer_equal);
79
80 blk->base.name = ctx->block_source_count++;
81
82 return blk;
83 }
84
85 static void
86 schedule_barrier(compiler_context *ctx)
87 {
88 midgard_block *temp = ctx->after_block;
89 ctx->after_block = create_empty_block(ctx);
90 ctx->block_count++;
91 list_addtail(&ctx->after_block->base.link, &ctx->blocks);
92 list_inithead(&ctx->after_block->base.instructions);
93 pan_block_add_successor(&ctx->current_block->base, &ctx->after_block->base);
94 ctx->current_block = ctx->after_block;
95 ctx->after_block = temp;
96 }
97
98 /* Helpers to generate midgard_instruction's using macro magic, since every
99 * driver seems to do it that way */
100
101 #define EMIT(op, ...) emit_mir_instruction(ctx, v_##op(__VA_ARGS__));
102
103 #define M_LOAD_STORE(name, store, T) \
104 static midgard_instruction m_##name(unsigned ssa, unsigned address) { \
105 midgard_instruction i = { \
106 .type = TAG_LOAD_STORE_4, \
107 .mask = 0xF, \
108 .dest = ~0, \
109 .src = { ~0, ~0, ~0, ~0 }, \
110 .swizzle = SWIZZLE_IDENTITY_4, \
111 .load_store = { \
112 .op = midgard_op_##name, \
113 .address = address \
114 } \
115 }; \
116 \
117 if (store) { \
118 i.src[0] = ssa; \
119 i.src_types[0] = T; \
120 i.dest_type = T; \
121 } else { \
122 i.dest = ssa; \
123 i.dest_type = T; \
124 } \
125 return i; \
126 }
127
128 #define M_LOAD(name, T) M_LOAD_STORE(name, false, T)
129 #define M_STORE(name, T) M_LOAD_STORE(name, true, T)
130
131 M_LOAD(ld_attr_32, nir_type_uint32);
132 M_LOAD(ld_vary_32, nir_type_uint32);
133 M_LOAD(ld_ubo_int4, nir_type_uint32);
134 M_LOAD(ld_int4, nir_type_uint32);
135 M_STORE(st_int4, nir_type_uint32);
136 M_LOAD(ld_color_buffer_32u, nir_type_uint32);
137 M_STORE(st_vary_32, nir_type_uint32);
138 M_LOAD(ld_cubemap_coords, nir_type_uint32);
139 M_LOAD(ld_compute_id, nir_type_uint32);
140
141 static midgard_instruction
142 v_branch(bool conditional, bool invert)
143 {
144 midgard_instruction ins = {
145 .type = TAG_ALU_4,
146 .unit = ALU_ENAB_BRANCH,
147 .compact_branch = true,
148 .branch = {
149 .conditional = conditional,
150 .invert_conditional = invert
151 },
152 .dest = ~0,
153 .src = { ~0, ~0, ~0, ~0 },
154 };
155
156 return ins;
157 }
158
159 static midgard_branch_extended
160 midgard_create_branch_extended( midgard_condition cond,
161 midgard_jmp_writeout_op op,
162 unsigned dest_tag,
163 signed quadword_offset)
164 {
165 /* The condition code is actually a LUT describing a function to
166 * combine multiple condition codes. However, we only support a single
167 * condition code at the moment, so we just duplicate over a bunch of
168 * times. */
169
170 uint16_t duplicated_cond =
171 (cond << 14) |
172 (cond << 12) |
173 (cond << 10) |
174 (cond << 8) |
175 (cond << 6) |
176 (cond << 4) |
177 (cond << 2) |
178 (cond << 0);
179
180 midgard_branch_extended branch = {
181 .op = op,
182 .dest_tag = dest_tag,
183 .offset = quadword_offset,
184 .cond = duplicated_cond
185 };
186
187 return branch;
188 }
189
190 static void
191 attach_constants(compiler_context *ctx, midgard_instruction *ins, void *constants, int name)
192 {
193 ins->has_constants = true;
194 memcpy(&ins->constants, constants, 16);
195 }
196
197 static int
198 glsl_type_size(const struct glsl_type *type, bool bindless)
199 {
200 return glsl_count_attribute_slots(type, false);
201 }
202
203 /* Lower fdot2 to a vector multiplication followed by channel addition */
204 static void
205 midgard_nir_lower_fdot2_body(nir_builder *b, nir_alu_instr *alu)
206 {
207 if (alu->op != nir_op_fdot2)
208 return;
209
210 b->cursor = nir_before_instr(&alu->instr);
211
212 nir_ssa_def *src0 = nir_ssa_for_alu_src(b, alu, 0);
213 nir_ssa_def *src1 = nir_ssa_for_alu_src(b, alu, 1);
214
215 nir_ssa_def *product = nir_fmul(b, src0, src1);
216
217 nir_ssa_def *sum = nir_fadd(b,
218 nir_channel(b, product, 0),
219 nir_channel(b, product, 1));
220
221 /* Replace the fdot2 with this sum */
222 nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(sum));
223 }
224
225 static bool
226 midgard_nir_lower_fdot2(nir_shader *shader)
227 {
228 bool progress = false;
229
230 nir_foreach_function(function, shader) {
231 if (!function->impl) continue;
232
233 nir_builder _b;
234 nir_builder *b = &_b;
235 nir_builder_init(b, function->impl);
236
237 nir_foreach_block(block, function->impl) {
238 nir_foreach_instr_safe(instr, block) {
239 if (instr->type != nir_instr_type_alu) continue;
240
241 nir_alu_instr *alu = nir_instr_as_alu(instr);
242 midgard_nir_lower_fdot2_body(b, alu);
243
244 progress |= true;
245 }
246 }
247
248 nir_metadata_preserve(function->impl, nir_metadata_block_index | nir_metadata_dominance);
249
250 }
251
252 return progress;
253 }
254
255 /* Midgard can't write depth and stencil separately. It has to happen in a
256 * single store operation containing both. Let's add a panfrost specific
257 * intrinsic and turn all depth/stencil stores into a packed depth+stencil
258 * one.
259 */
260 static bool
261 midgard_nir_lower_zs_store(nir_shader *nir)
262 {
263 if (nir->info.stage != MESA_SHADER_FRAGMENT)
264 return false;
265
266 nir_variable *z_var = NULL, *s_var = NULL;
267
268 nir_foreach_variable(var, &nir->outputs) {
269 if (var->data.location == FRAG_RESULT_DEPTH)
270 z_var = var;
271 else if (var->data.location == FRAG_RESULT_STENCIL)
272 s_var = var;
273 }
274
275 if (!z_var && !s_var)
276 return false;
277
278 bool progress = false;
279
280 nir_foreach_function(function, nir) {
281 if (!function->impl) continue;
282
283 nir_intrinsic_instr *z_store = NULL, *s_store = NULL, *last_store = NULL;
284
285 nir_foreach_block(block, function->impl) {
286 nir_foreach_instr_safe(instr, block) {
287 if (instr->type != nir_instr_type_intrinsic)
288 continue;
289
290 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
291 if (intr->intrinsic != nir_intrinsic_store_output)
292 continue;
293
294 if (z_var && nir_intrinsic_base(intr) == z_var->data.driver_location) {
295 assert(!z_store);
296 z_store = intr;
297 last_store = intr;
298 }
299
300 if (s_var && nir_intrinsic_base(intr) == s_var->data.driver_location) {
301 assert(!s_store);
302 s_store = intr;
303 last_store = intr;
304 }
305 }
306 }
307
308 if (!z_store && !s_store) continue;
309
310 nir_builder b;
311 nir_builder_init(&b, function->impl);
312
313 b.cursor = nir_before_instr(&last_store->instr);
314
315 nir_ssa_def *zs_store_src;
316
317 if (z_store && s_store) {
318 nir_ssa_def *srcs[2] = {
319 nir_ssa_for_src(&b, z_store->src[0], 1),
320 nir_ssa_for_src(&b, s_store->src[0], 1),
321 };
322
323 zs_store_src = nir_vec(&b, srcs, 2);
324 } else {
325 zs_store_src = nir_ssa_for_src(&b, last_store->src[0], 1);
326 }
327
328 nir_intrinsic_instr *zs_store;
329
330 zs_store = nir_intrinsic_instr_create(b.shader,
331 nir_intrinsic_store_zs_output_pan);
332 zs_store->src[0] = nir_src_for_ssa(zs_store_src);
333 zs_store->num_components = z_store && s_store ? 2 : 1;
334 nir_intrinsic_set_component(zs_store, z_store ? 0 : 1);
335
336 /* Replace the Z and S store by a ZS store */
337 nir_builder_instr_insert(&b, &zs_store->instr);
338
339 if (z_store)
340 nir_instr_remove(&z_store->instr);
341
342 if (s_store)
343 nir_instr_remove(&s_store->instr);
344
345 nir_metadata_preserve(function->impl, nir_metadata_block_index | nir_metadata_dominance);
346 progress = true;
347 }
348
349 return progress;
350 }
351
352 /* Flushes undefined values to zero */
353
354 static void
355 optimise_nir(nir_shader *nir, unsigned quirks)
356 {
357 bool progress;
358 unsigned lower_flrp =
359 (nir->options->lower_flrp16 ? 16 : 0) |
360 (nir->options->lower_flrp32 ? 32 : 0) |
361 (nir->options->lower_flrp64 ? 64 : 0);
362
363 NIR_PASS(progress, nir, nir_lower_regs_to_ssa);
364 NIR_PASS(progress, nir, nir_lower_idiv, nir_lower_idiv_fast);
365
366 nir_lower_tex_options lower_tex_options = {
367 .lower_txs_lod = true,
368 .lower_txp = ~0,
369 .lower_tex_without_implicit_lod =
370 (quirks & MIDGARD_EXPLICIT_LOD),
371
372 /* TODO: we have native gradient.. */
373 .lower_txd = true,
374 };
375
376 NIR_PASS(progress, nir, nir_lower_tex, &lower_tex_options);
377
378 /* Must lower fdot2 after tex is lowered */
379 NIR_PASS(progress, nir, midgard_nir_lower_fdot2);
380
381 /* T720 is broken. */
382
383 if (quirks & MIDGARD_BROKEN_LOD)
384 NIR_PASS_V(nir, midgard_nir_lod_errata);
385
386 NIR_PASS(progress, nir, midgard_nir_lower_algebraic_early);
387
388 do {
389 progress = false;
390
391 NIR_PASS(progress, nir, nir_lower_var_copies);
392 NIR_PASS(progress, nir, nir_lower_vars_to_ssa);
393
394 NIR_PASS(progress, nir, nir_copy_prop);
395 NIR_PASS(progress, nir, nir_opt_remove_phis);
396 NIR_PASS(progress, nir, nir_opt_dce);
397 NIR_PASS(progress, nir, nir_opt_dead_cf);
398 NIR_PASS(progress, nir, nir_opt_cse);
399 NIR_PASS(progress, nir, nir_opt_peephole_select, 64, false, true);
400 NIR_PASS(progress, nir, nir_opt_algebraic);
401 NIR_PASS(progress, nir, nir_opt_constant_folding);
402
403 if (lower_flrp != 0) {
404 bool lower_flrp_progress = false;
405 NIR_PASS(lower_flrp_progress,
406 nir,
407 nir_lower_flrp,
408 lower_flrp,
409 false /* always_precise */,
410 nir->options->lower_ffma);
411 if (lower_flrp_progress) {
412 NIR_PASS(progress, nir,
413 nir_opt_constant_folding);
414 progress = true;
415 }
416
417 /* Nothing should rematerialize any flrps, so we only
418 * need to do this lowering once.
419 */
420 lower_flrp = 0;
421 }
422
423 NIR_PASS(progress, nir, nir_opt_undef);
424 NIR_PASS(progress, nir, nir_undef_to_zero);
425
426 NIR_PASS(progress, nir, nir_opt_loop_unroll,
427 nir_var_shader_in |
428 nir_var_shader_out |
429 nir_var_function_temp);
430
431 NIR_PASS(progress, nir, nir_opt_vectorize);
432 } while (progress);
433
434 /* Must be run at the end to prevent creation of fsin/fcos ops */
435 NIR_PASS(progress, nir, midgard_nir_scale_trig);
436
437 do {
438 progress = false;
439
440 NIR_PASS(progress, nir, nir_opt_dce);
441 NIR_PASS(progress, nir, nir_opt_algebraic);
442 NIR_PASS(progress, nir, nir_opt_constant_folding);
443 NIR_PASS(progress, nir, nir_copy_prop);
444 } while (progress);
445
446 NIR_PASS(progress, nir, nir_opt_algebraic_late);
447 NIR_PASS(progress, nir, nir_opt_algebraic_distribute_src_mods);
448
449 /* We implement booleans as 32-bit 0/~0 */
450 NIR_PASS(progress, nir, nir_lower_bool_to_int32);
451
452 /* Now that booleans are lowered, we can run out late opts */
453 NIR_PASS(progress, nir, midgard_nir_lower_algebraic_late);
454 NIR_PASS(progress, nir, midgard_nir_cancel_inot);
455
456 NIR_PASS(progress, nir, nir_copy_prop);
457 NIR_PASS(progress, nir, nir_opt_dce);
458
459 /* Take us out of SSA */
460 NIR_PASS(progress, nir, nir_lower_locals_to_regs);
461 NIR_PASS(progress, nir, nir_convert_from_ssa, true);
462
463 /* We are a vector architecture; write combine where possible */
464 NIR_PASS(progress, nir, nir_move_vec_src_uses_to_dest);
465 NIR_PASS(progress, nir, nir_lower_vec_to_movs);
466
467 NIR_PASS(progress, nir, nir_opt_dce);
468 }
469
470 /* Do not actually emit a load; instead, cache the constant for inlining */
471
472 static void
473 emit_load_const(compiler_context *ctx, nir_load_const_instr *instr)
474 {
475 nir_ssa_def def = instr->def;
476
477 midgard_constants *consts = rzalloc(NULL, midgard_constants);
478
479 assert(instr->def.num_components * instr->def.bit_size <= sizeof(*consts) * 8);
480
481 #define RAW_CONST_COPY(bits) \
482 nir_const_value_to_array(consts->u##bits, instr->value, \
483 instr->def.num_components, u##bits)
484
485 switch (instr->def.bit_size) {
486 case 64:
487 RAW_CONST_COPY(64);
488 break;
489 case 32:
490 RAW_CONST_COPY(32);
491 break;
492 case 16:
493 RAW_CONST_COPY(16);
494 break;
495 case 8:
496 RAW_CONST_COPY(8);
497 break;
498 default:
499 unreachable("Invalid bit_size for load_const instruction\n");
500 }
501
502 /* Shifted for SSA, +1 for off-by-one */
503 _mesa_hash_table_u64_insert(ctx->ssa_constants, (def.index << 1) + 1, consts);
504 }
505
506 /* Normally constants are embedded implicitly, but for I/O and such we have to
507 * explicitly emit a move with the constant source */
508
509 static void
510 emit_explicit_constant(compiler_context *ctx, unsigned node, unsigned to)
511 {
512 void *constant_value = _mesa_hash_table_u64_search(ctx->ssa_constants, node + 1);
513
514 if (constant_value) {
515 midgard_instruction ins = v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT), to);
516 attach_constants(ctx, &ins, constant_value, node + 1);
517 emit_mir_instruction(ctx, ins);
518 }
519 }
520
521 static bool
522 nir_is_non_scalar_swizzle(nir_alu_src *src, unsigned nr_components)
523 {
524 unsigned comp = src->swizzle[0];
525
526 for (unsigned c = 1; c < nr_components; ++c) {
527 if (src->swizzle[c] != comp)
528 return true;
529 }
530
531 return false;
532 }
533
534 #define ALU_CASE(nir, _op) \
535 case nir_op_##nir: \
536 op = midgard_alu_op_##_op; \
537 assert(src_bitsize == dst_bitsize); \
538 break;
539
540 #define ALU_CHECK_CMP(sext) \
541 assert(src_bitsize == 16 || src_bitsize == 32); \
542 assert(dst_bitsize == 16 || dst_bitsize == 32); \
543
544 #define ALU_CASE_BCAST(nir, _op, count) \
545 case nir_op_##nir: \
546 op = midgard_alu_op_##_op; \
547 broadcast_swizzle = count; \
548 ALU_CHECK_CMP(true); \
549 break;
550
551 #define ALU_CASE_CMP(nir, _op, sext) \
552 case nir_op_##nir: \
553 op = midgard_alu_op_##_op; \
554 ALU_CHECK_CMP(sext); \
555 break;
556
557 /* Analyze the sizes of the dest and inputs to determine reg mode. */
558
559 static midgard_reg_mode
560 reg_mode_for_nir(nir_alu_instr *instr)
561 {
562 unsigned src_bitsize = nir_src_bit_size(instr->src[0].src);
563 unsigned dst_bitsize = nir_dest_bit_size(instr->dest.dest);
564 unsigned max_bitsize = MAX2(src_bitsize, dst_bitsize);
565
566 /* We don't have fp16 LUTs, so we'll want to emit code like:
567 *
568 * vlut.fsinr hr0, hr0
569 *
570 * where both input and output are 16-bit but the operation is carried
571 * out in 32-bit
572 */
573
574 switch (instr->op) {
575 case nir_op_fsqrt:
576 case nir_op_frcp:
577 case nir_op_frsq:
578 case nir_op_fsin:
579 case nir_op_fcos:
580 case nir_op_fexp2:
581 case nir_op_flog2:
582 max_bitsize = MAX2(max_bitsize, 32);
583 default:
584 break;
585 }
586
587 switch (max_bitsize) {
588 case 8:
589 return midgard_reg_mode_8;
590 case 16:
591 return midgard_reg_mode_16;
592 case 32:
593 return midgard_reg_mode_32;
594 case 64:
595 return midgard_reg_mode_64;
596 default:
597 unreachable("Invalid bit size");
598 }
599 }
600
601 /* Compare mir_lower_invert */
602 static bool
603 nir_accepts_inot(nir_op op, unsigned src)
604 {
605 switch (op) {
606 case nir_op_ior:
607 case nir_op_iand: /* TODO: b2f16 */
608 case nir_op_ixor:
609 return true;
610 case nir_op_b32csel:
611 /* Only the condition */
612 return (src == 0);
613 default:
614 return false;
615 }
616 }
617
618 static bool
619 mir_accept_dest_mod(compiler_context *ctx, nir_dest **dest, nir_op op)
620 {
621 if (pan_has_dest_mod(dest, op)) {
622 assert((*dest)->is_ssa);
623 BITSET_SET(ctx->already_emitted, (*dest)->ssa.index);
624 return true;
625 }
626
627 return false;
628 }
629
630 static void
631 mir_copy_src(midgard_instruction *ins, nir_alu_instr *instr, unsigned i, unsigned to, bool *abs, bool *neg, bool *not, bool is_int, unsigned bcast_count)
632 {
633 nir_alu_src src = instr->src[i];
634
635 if (!is_int) {
636 if (pan_has_source_mod(&src, nir_op_fneg))
637 *neg = !(*neg);
638
639 if (pan_has_source_mod(&src, nir_op_fabs))
640 *abs = true;
641 }
642
643 if (nir_accepts_inot(instr->op, i) && pan_has_source_mod(&src, nir_op_inot))
644 *not = true;
645
646 unsigned bits = nir_src_bit_size(src.src);
647
648 ins->src[to] = nir_src_index(NULL, &src.src);
649 ins->src_types[to] = nir_op_infos[instr->op].input_types[i] | bits;
650
651 for (unsigned c = 0; c < NIR_MAX_VEC_COMPONENTS; ++c) {
652 ins->swizzle[to][c] = src.swizzle[
653 (!bcast_count || c < bcast_count) ? c :
654 (bcast_count - 1)];
655 }
656 }
657
658 /* Midgard features both fcsel and icsel, depending on whether you want int or
659 * float modifiers. NIR's csel is typeless, so we want a heuristic to guess if
660 * we should emit an int or float csel depending on what modifiers could be
661 * placed. In the absense of modifiers, this is probably arbitrary. */
662
663 static bool
664 mir_is_bcsel_float(nir_alu_instr *instr)
665 {
666 nir_op intmods[] = {
667 nir_op_i2i8, nir_op_i2i16,
668 nir_op_i2i32, nir_op_i2i64
669 };
670
671 nir_op floatmods[] = {
672 nir_op_fabs, nir_op_fneg,
673 nir_op_f2f16, nir_op_f2f32,
674 nir_op_f2f64
675 };
676
677 nir_op floatdestmods[] = {
678 nir_op_fsat, nir_op_fsat_signed, nir_op_fclamp_pos,
679 nir_op_f2f16, nir_op_f2f32
680 };
681
682 signed score = 0;
683
684 for (unsigned i = 1; i < 3; ++i) {
685 nir_alu_src s = instr->src[i];
686 for (unsigned q = 0; q < ARRAY_SIZE(intmods); ++q) {
687 if (pan_has_source_mod(&s, intmods[q]))
688 score--;
689 }
690 }
691
692 for (unsigned i = 1; i < 3; ++i) {
693 nir_alu_src s = instr->src[i];
694 for (unsigned q = 0; q < ARRAY_SIZE(floatmods); ++q) {
695 if (pan_has_source_mod(&s, floatmods[q]))
696 score++;
697 }
698 }
699
700 for (unsigned q = 0; q < ARRAY_SIZE(floatdestmods); ++q) {
701 nir_dest *dest = &instr->dest.dest;
702 if (pan_has_dest_mod(&dest, floatdestmods[q]))
703 score++;
704 }
705
706 return (score > 0);
707 }
708
709 static void
710 emit_alu(compiler_context *ctx, nir_alu_instr *instr)
711 {
712 nir_dest *dest = &instr->dest.dest;
713
714 if (dest->is_ssa && BITSET_TEST(ctx->already_emitted, dest->ssa.index))
715 return;
716
717 /* Derivatives end up emitted on the texture pipe, not the ALUs. This
718 * is handled elsewhere */
719
720 if (instr->op == nir_op_fddx || instr->op == nir_op_fddy) {
721 midgard_emit_derivatives(ctx, instr);
722 return;
723 }
724
725 bool is_ssa = dest->is_ssa;
726
727 unsigned nr_components = nir_dest_num_components(*dest);
728 unsigned nr_inputs = nir_op_infos[instr->op].num_inputs;
729 unsigned op = 0;
730
731 /* Number of components valid to check for the instruction (the rest
732 * will be forced to the last), or 0 to use as-is. Relevant as
733 * ball-type instructions have a channel count in NIR but are all vec4
734 * in Midgard */
735
736 unsigned broadcast_swizzle = 0;
737
738 /* What register mode should we operate in? */
739 midgard_reg_mode reg_mode =
740 reg_mode_for_nir(instr);
741
742 /* Should we swap arguments? */
743 bool flip_src12 = false;
744
745 unsigned src_bitsize = nir_src_bit_size(instr->src[0].src);
746 unsigned dst_bitsize = nir_dest_bit_size(*dest);
747
748 switch (instr->op) {
749 ALU_CASE(fadd, fadd);
750 ALU_CASE(fmul, fmul);
751 ALU_CASE(fmin, fmin);
752 ALU_CASE(fmax, fmax);
753 ALU_CASE(imin, imin);
754 ALU_CASE(imax, imax);
755 ALU_CASE(umin, umin);
756 ALU_CASE(umax, umax);
757 ALU_CASE(ffloor, ffloor);
758 ALU_CASE(fround_even, froundeven);
759 ALU_CASE(ftrunc, ftrunc);
760 ALU_CASE(fceil, fceil);
761 ALU_CASE(fdot3, fdot3);
762 ALU_CASE(fdot4, fdot4);
763 ALU_CASE(iadd, iadd);
764 ALU_CASE(isub, isub);
765 ALU_CASE(imul, imul);
766
767 /* Zero shoved as second-arg */
768 ALU_CASE(iabs, iabsdiff);
769
770 ALU_CASE(mov, imov);
771
772 ALU_CASE_CMP(feq32, feq, false);
773 ALU_CASE_CMP(fne32, fne, false);
774 ALU_CASE_CMP(flt32, flt, false);
775 ALU_CASE_CMP(ieq32, ieq, true);
776 ALU_CASE_CMP(ine32, ine, true);
777 ALU_CASE_CMP(ilt32, ilt, true);
778 ALU_CASE_CMP(ult32, ult, false);
779
780 /* We don't have a native b2f32 instruction. Instead, like many
781 * GPUs, we exploit booleans as 0/~0 for false/true, and
782 * correspondingly AND
783 * by 1.0 to do the type conversion. For the moment, prime us
784 * to emit:
785 *
786 * iand [whatever], #0
787 *
788 * At the end of emit_alu (as MIR), we'll fix-up the constant
789 */
790
791 ALU_CASE_CMP(b2f32, iand, true);
792 ALU_CASE_CMP(b2f16, iand, true);
793 ALU_CASE_CMP(b2i32, iand, true);
794
795 /* Likewise, we don't have a dedicated f2b32 instruction, but
796 * we can do a "not equal to 0.0" test. */
797
798 ALU_CASE_CMP(f2b32, fne, false);
799 ALU_CASE_CMP(i2b32, ine, true);
800
801 ALU_CASE(frcp, frcp);
802 ALU_CASE(frsq, frsqrt);
803 ALU_CASE(fsqrt, fsqrt);
804 ALU_CASE(fexp2, fexp2);
805 ALU_CASE(flog2, flog2);
806
807 ALU_CASE(f2i64, f2i_rtz);
808 ALU_CASE(f2u64, f2u_rtz);
809 ALU_CASE(i2f64, i2f_rtz);
810 ALU_CASE(u2f64, u2f_rtz);
811
812 ALU_CASE(f2i32, f2i_rtz);
813 ALU_CASE(f2u32, f2u_rtz);
814 ALU_CASE(i2f32, i2f_rtz);
815 ALU_CASE(u2f32, u2f_rtz);
816
817 ALU_CASE(f2i16, f2i_rtz);
818 ALU_CASE(f2u16, f2u_rtz);
819 ALU_CASE(i2f16, i2f_rtz);
820 ALU_CASE(u2f16, u2f_rtz);
821
822 ALU_CASE(fsin, fsin);
823 ALU_CASE(fcos, fcos);
824
825 /* We'll get 0 in the second arg, so:
826 * ~a = ~(a | 0) = nor(a, 0) */
827 ALU_CASE(inot, inor);
828 ALU_CASE(iand, iand);
829 ALU_CASE(ior, ior);
830 ALU_CASE(ixor, ixor);
831 ALU_CASE(ishl, ishl);
832 ALU_CASE(ishr, iasr);
833 ALU_CASE(ushr, ilsr);
834
835 ALU_CASE_BCAST(b32all_fequal2, fball_eq, 2);
836 ALU_CASE_BCAST(b32all_fequal3, fball_eq, 3);
837 ALU_CASE_CMP(b32all_fequal4, fball_eq, true);
838
839 ALU_CASE_BCAST(b32any_fnequal2, fbany_neq, 2);
840 ALU_CASE_BCAST(b32any_fnequal3, fbany_neq, 3);
841 ALU_CASE_CMP(b32any_fnequal4, fbany_neq, true);
842
843 ALU_CASE_BCAST(b32all_iequal2, iball_eq, 2);
844 ALU_CASE_BCAST(b32all_iequal3, iball_eq, 3);
845 ALU_CASE_CMP(b32all_iequal4, iball_eq, true);
846
847 ALU_CASE_BCAST(b32any_inequal2, ibany_neq, 2);
848 ALU_CASE_BCAST(b32any_inequal3, ibany_neq, 3);
849 ALU_CASE_CMP(b32any_inequal4, ibany_neq, true);
850
851 /* Source mods will be shoved in later */
852 ALU_CASE(fabs, fmov);
853 ALU_CASE(fneg, fmov);
854 ALU_CASE(fsat, fmov);
855 ALU_CASE(fsat_signed, fmov);
856 ALU_CASE(fclamp_pos, fmov);
857
858 /* For size conversion, we use a move. Ideally though we would squash
859 * these ops together; maybe that has to happen after in NIR as part of
860 * propagation...? An earlier algebraic pass ensured we step down by
861 * only / exactly one size. If stepping down, we use a dest override to
862 * reduce the size; if stepping up, we use a larger-sized move with a
863 * half source and a sign/zero-extension modifier */
864
865 case nir_op_i2i8:
866 case nir_op_i2i16:
867 case nir_op_i2i32:
868 case nir_op_i2i64:
869 case nir_op_u2u8:
870 case nir_op_u2u16:
871 case nir_op_u2u32:
872 case nir_op_u2u64:
873 case nir_op_f2f16:
874 case nir_op_f2f32:
875 case nir_op_f2f64: {
876 if (instr->op == nir_op_f2f16 || instr->op == nir_op_f2f32 ||
877 instr->op == nir_op_f2f64)
878 op = midgard_alu_op_fmov;
879 else
880 op = midgard_alu_op_imov;
881
882 break;
883 }
884
885 /* For greater-or-equal, we lower to less-or-equal and flip the
886 * arguments */
887
888 case nir_op_fge:
889 case nir_op_fge32:
890 case nir_op_ige32:
891 case nir_op_uge32: {
892 op =
893 instr->op == nir_op_fge ? midgard_alu_op_fle :
894 instr->op == nir_op_fge32 ? midgard_alu_op_fle :
895 instr->op == nir_op_ige32 ? midgard_alu_op_ile :
896 instr->op == nir_op_uge32 ? midgard_alu_op_ule :
897 0;
898
899 flip_src12 = true;
900 ALU_CHECK_CMP(false);
901 break;
902 }
903
904 case nir_op_b32csel: {
905 bool mixed = nir_is_non_scalar_swizzle(&instr->src[0], nr_components);
906 bool is_float = mir_is_bcsel_float(instr);
907 op = is_float ?
908 (mixed ? midgard_alu_op_fcsel_v : midgard_alu_op_fcsel) :
909 (mixed ? midgard_alu_op_icsel_v : midgard_alu_op_icsel);
910
911 break;
912 }
913
914 case nir_op_unpack_32_2x16:
915 case nir_op_unpack_32_4x8:
916 case nir_op_pack_32_2x16:
917 case nir_op_pack_32_4x8: {
918 op = midgard_alu_op_imov;
919 break;
920 }
921
922 default:
923 DBG("Unhandled ALU op %s\n", nir_op_infos[instr->op].name);
924 assert(0);
925 return;
926 }
927
928 /* Promote imov to fmov if it might help inline a constant */
929 if (op == midgard_alu_op_imov && nir_src_is_const(instr->src[0].src)
930 && nir_src_bit_size(instr->src[0].src) == 32
931 && nir_is_same_comp_swizzle(instr->src[0].swizzle,
932 nir_src_num_components(instr->src[0].src))) {
933 op = midgard_alu_op_fmov;
934 }
935
936 /* Midgard can perform certain modifiers on output of an ALU op */
937
938 unsigned outmod = 0;
939 bool is_int = midgard_is_integer_op(op);
940
941 if (midgard_is_integer_out_op(op)) {
942 outmod = midgard_outmod_int_wrap;
943 } else if (instr->op == nir_op_fsat) {
944 outmod = midgard_outmod_sat;
945 } else if (instr->op == nir_op_fsat_signed) {
946 outmod = midgard_outmod_sat_signed;
947 } else if (instr->op == nir_op_fclamp_pos) {
948 outmod = midgard_outmod_pos;
949 }
950
951 /* Fetch unit, quirks, etc information */
952 unsigned opcode_props = alu_opcode_props[op].props;
953 bool quirk_flipped_r24 = opcode_props & QUIRK_FLIPPED_R24;
954
955 /* Look for floating point mods. We have the mods fsat, fsat_signed,
956 * and fpos. We also have the relations (note 3 * 2 = 6 cases):
957 *
958 * fsat_signed(fpos(x)) = fsat(x)
959 * fsat_signed(fsat(x)) = fsat(x)
960 * fpos(fsat_signed(x)) = fsat(x)
961 * fpos(fsat(x)) = fsat(x)
962 * fsat(fsat_signed(x)) = fsat(x)
963 * fsat(fpos(x)) = fsat(x)
964 *
965 * So by cases any composition of output modifiers is equivalent to
966 * fsat alone.
967 */
968
969 if (!is_int && !(opcode_props & OP_TYPE_CONVERT)) {
970 bool fpos = mir_accept_dest_mod(ctx, &dest, nir_op_fclamp_pos);
971 bool fsat = mir_accept_dest_mod(ctx, &dest, nir_op_fsat);
972 bool ssat = mir_accept_dest_mod(ctx, &dest, nir_op_fsat_signed);
973 bool prior = (outmod != midgard_outmod_none);
974 int count = (int) prior + (int) fpos + (int) ssat + (int) fsat;
975
976 outmod = ((count > 1) || fsat) ? midgard_outmod_sat :
977 fpos ? midgard_outmod_pos :
978 ssat ? midgard_outmod_sat_signed :
979 outmod;
980 }
981
982 midgard_instruction ins = {
983 .type = TAG_ALU_4,
984 .dest = nir_dest_index(dest),
985 .dest_type = nir_op_infos[instr->op].output_type
986 | nir_dest_bit_size(*dest),
987 };
988
989 for (unsigned i = nr_inputs; i < ARRAY_SIZE(ins.src); ++i)
990 ins.src[i] = ~0;
991
992 if (quirk_flipped_r24) {
993 ins.src[0] = ~0;
994 mir_copy_src(&ins, instr, 0, 1, &ins.src_abs[1], &ins.src_neg[1], &ins.src_invert[1], is_int, broadcast_swizzle);
995 } else {
996 for (unsigned i = 0; i < nr_inputs; ++i) {
997 unsigned to = i;
998
999 if (instr->op == nir_op_b32csel) {
1000 /* The condition is the first argument; move
1001 * the other arguments up one to be a binary
1002 * instruction for Midgard with the condition
1003 * last */
1004
1005 if (i == 0)
1006 to = 2;
1007 else if (flip_src12)
1008 to = 2 - i;
1009 else
1010 to = i - 1;
1011 } else if (flip_src12) {
1012 to = 1 - to;
1013 }
1014
1015 mir_copy_src(&ins, instr, i, to, &ins.src_abs[to], &ins.src_neg[to], &ins.src_invert[to], is_int, broadcast_swizzle);
1016
1017 /* (!c) ? a : b = c ? b : a */
1018 if (instr->op == nir_op_b32csel && ins.src_invert[2]) {
1019 ins.src_invert[2] = false;
1020 flip_src12 ^= true;
1021 }
1022 }
1023 }
1024
1025 if (instr->op == nir_op_fneg || instr->op == nir_op_fabs) {
1026 /* Lowered to move */
1027 if (instr->op == nir_op_fneg)
1028 ins.src_neg[1] ^= true;
1029
1030 if (instr->op == nir_op_fabs)
1031 ins.src_abs[1] = true;
1032 }
1033
1034 ins.mask = mask_of(nr_components);
1035
1036 midgard_vector_alu alu = {
1037 .op = op,
1038 .reg_mode = reg_mode,
1039 .outmod = outmod,
1040 };
1041
1042 /* Apply writemask if non-SSA, keeping in mind that we can't write to
1043 * components that don't exist. Note modifier => SSA => !reg => no
1044 * writemask, so we don't have to worry about writemasks here.*/
1045
1046 if (!is_ssa)
1047 ins.mask &= instr->dest.write_mask;
1048
1049 ins.alu = alu;
1050
1051 /* Late fixup for emulated instructions */
1052
1053 if (instr->op == nir_op_b2f32 || instr->op == nir_op_b2i32) {
1054 /* Presently, our second argument is an inline #0 constant.
1055 * Switch over to an embedded 1.0 constant (that can't fit
1056 * inline, since we're 32-bit, not 16-bit like the inline
1057 * constants) */
1058
1059 ins.has_inline_constant = false;
1060 ins.src[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT);
1061 ins.src_types[1] = nir_type_float32;
1062 ins.has_constants = true;
1063
1064 if (instr->op == nir_op_b2f32)
1065 ins.constants.f32[0] = 1.0f;
1066 else
1067 ins.constants.i32[0] = 1;
1068
1069 for (unsigned c = 0; c < 16; ++c)
1070 ins.swizzle[1][c] = 0;
1071 } else if (instr->op == nir_op_b2f16) {
1072 ins.src[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT);
1073 ins.src_types[1] = nir_type_float16;
1074 ins.has_constants = true;
1075 ins.constants.i16[0] = _mesa_float_to_half(1.0);
1076
1077 for (unsigned c = 0; c < 16; ++c)
1078 ins.swizzle[1][c] = 0;
1079 } else if (nr_inputs == 1 && !quirk_flipped_r24) {
1080 /* Lots of instructions need a 0 plonked in */
1081 ins.has_inline_constant = false;
1082 ins.src[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT);
1083 ins.src_types[1] = nir_type_uint32;
1084 ins.has_constants = true;
1085 ins.constants.u32[0] = 0;
1086
1087 for (unsigned c = 0; c < 16; ++c)
1088 ins.swizzle[1][c] = 0;
1089 } else if (instr->op == nir_op_pack_32_2x16) {
1090 ins.dest_type = nir_type_uint16;
1091 ins.mask = mask_of(nr_components * 2);
1092 } else if (instr->op == nir_op_pack_32_4x8) {
1093 ins.dest_type = nir_type_uint8;
1094 ins.mask = mask_of(nr_components * 4);
1095 } else if (instr->op == nir_op_unpack_32_2x16) {
1096 ins.dest_type = nir_type_uint32;
1097 ins.mask = mask_of(nr_components >> 1);
1098 } else if (instr->op == nir_op_unpack_32_4x8) {
1099 ins.dest_type = nir_type_uint32;
1100 ins.mask = mask_of(nr_components >> 2);
1101 }
1102
1103 /* Arrange for creation of iandnot/iornot */
1104 if (ins.src_invert[0] && !ins.src_invert[1]) {
1105 mir_flip(&ins);
1106 ins.src_invert[0] = false;
1107 ins.src_invert[1] = true;
1108 }
1109
1110 if ((opcode_props & UNITS_ALL) == UNIT_VLUT) {
1111 /* To avoid duplicating the lookup tables (probably), true LUT
1112 * instructions can only operate as if they were scalars. Lower
1113 * them here by changing the component. */
1114
1115 unsigned orig_mask = ins.mask;
1116
1117 unsigned swizzle_back[MIR_VEC_COMPONENTS];
1118 memcpy(&swizzle_back, ins.swizzle[0], sizeof(swizzle_back));
1119
1120 for (int i = 0; i < nr_components; ++i) {
1121 /* Mask the associated component, dropping the
1122 * instruction if needed */
1123
1124 ins.mask = 1 << i;
1125 ins.mask &= orig_mask;
1126
1127 if (!ins.mask)
1128 continue;
1129
1130 for (unsigned j = 0; j < MIR_VEC_COMPONENTS; ++j)
1131 ins.swizzle[0][j] = swizzle_back[i]; /* Pull from the correct component */
1132
1133 emit_mir_instruction(ctx, ins);
1134 }
1135 } else {
1136 emit_mir_instruction(ctx, ins);
1137 }
1138 }
1139
1140 #undef ALU_CASE
1141
1142 static void
1143 mir_set_intr_mask(nir_instr *instr, midgard_instruction *ins, bool is_read)
1144 {
1145 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
1146 unsigned nir_mask = 0;
1147 unsigned dsize = 0;
1148
1149 if (is_read) {
1150 nir_mask = mask_of(nir_intrinsic_dest_components(intr));
1151 dsize = nir_dest_bit_size(intr->dest);
1152 } else {
1153 nir_mask = nir_intrinsic_write_mask(intr);
1154 dsize = 32;
1155 }
1156
1157 /* Once we have the NIR mask, we need to normalize to work in 32-bit space */
1158 unsigned bytemask = pan_to_bytemask(dsize, nir_mask);
1159 mir_set_bytemask(ins, bytemask);
1160 ins->dest_type = nir_type_uint | dsize;
1161 }
1162
1163 /* Uniforms and UBOs use a shared code path, as uniforms are just (slightly
1164 * optimized) versions of UBO #0 */
1165
1166 static midgard_instruction *
1167 emit_ubo_read(
1168 compiler_context *ctx,
1169 nir_instr *instr,
1170 unsigned dest,
1171 unsigned offset,
1172 nir_src *indirect_offset,
1173 unsigned indirect_shift,
1174 unsigned index)
1175 {
1176 /* TODO: half-floats */
1177
1178 midgard_instruction ins = m_ld_ubo_int4(dest, 0);
1179 ins.constants.u32[0] = offset;
1180
1181 if (instr->type == nir_instr_type_intrinsic)
1182 mir_set_intr_mask(instr, &ins, true);
1183
1184 if (indirect_offset) {
1185 ins.src[2] = nir_src_index(ctx, indirect_offset);
1186 ins.src_types[2] = nir_type_uint32;
1187 ins.load_store.arg_2 = (indirect_shift << 5);
1188 } else {
1189 ins.load_store.arg_2 = 0x1E;
1190 }
1191
1192 ins.load_store.arg_1 = index;
1193
1194 return emit_mir_instruction(ctx, ins);
1195 }
1196
1197 /* Globals are like UBOs if you squint. And shared memory is like globals if
1198 * you squint even harder */
1199
1200 static void
1201 emit_global(
1202 compiler_context *ctx,
1203 nir_instr *instr,
1204 bool is_read,
1205 unsigned srcdest,
1206 nir_src *offset,
1207 bool is_shared)
1208 {
1209 /* TODO: types */
1210
1211 midgard_instruction ins;
1212
1213 if (is_read)
1214 ins = m_ld_int4(srcdest, 0);
1215 else
1216 ins = m_st_int4(srcdest, 0);
1217
1218 mir_set_offset(ctx, &ins, offset, is_shared);
1219 mir_set_intr_mask(instr, &ins, is_read);
1220
1221 emit_mir_instruction(ctx, ins);
1222 }
1223
1224 static void
1225 emit_varying_read(
1226 compiler_context *ctx,
1227 unsigned dest, unsigned offset,
1228 unsigned nr_comp, unsigned component,
1229 nir_src *indirect_offset, nir_alu_type type, bool flat)
1230 {
1231 /* XXX: Half-floats? */
1232 /* TODO: swizzle, mask */
1233
1234 midgard_instruction ins = m_ld_vary_32(dest, offset);
1235 ins.mask = mask_of(nr_comp);
1236
1237 for (unsigned i = 0; i < ARRAY_SIZE(ins.swizzle[0]); ++i)
1238 ins.swizzle[0][i] = MIN2(i + component, COMPONENT_W);
1239
1240 midgard_varying_parameter p = {
1241 .is_varying = 1,
1242 .interpolation = midgard_interp_default,
1243 .flat = flat,
1244 };
1245
1246 unsigned u;
1247 memcpy(&u, &p, sizeof(p));
1248 ins.load_store.varying_parameters = u;
1249
1250 if (indirect_offset) {
1251 ins.src[2] = nir_src_index(ctx, indirect_offset);
1252 ins.src_types[2] = nir_type_uint32;
1253 } else
1254 ins.load_store.arg_2 = 0x1E;
1255
1256 ins.load_store.arg_1 = 0x9E;
1257
1258 /* Use the type appropriate load */
1259 switch (type) {
1260 case nir_type_uint:
1261 case nir_type_bool:
1262 ins.load_store.op = midgard_op_ld_vary_32u;
1263 break;
1264 case nir_type_int:
1265 ins.load_store.op = midgard_op_ld_vary_32i;
1266 break;
1267 case nir_type_float:
1268 ins.load_store.op = midgard_op_ld_vary_32;
1269 break;
1270 default:
1271 unreachable("Attempted to load unknown type");
1272 break;
1273 }
1274
1275 emit_mir_instruction(ctx, ins);
1276 }
1277
1278 static void
1279 emit_attr_read(
1280 compiler_context *ctx,
1281 unsigned dest, unsigned offset,
1282 unsigned nr_comp, nir_alu_type t)
1283 {
1284 midgard_instruction ins = m_ld_attr_32(dest, offset);
1285 ins.load_store.arg_1 = 0x1E;
1286 ins.load_store.arg_2 = 0x1E;
1287 ins.mask = mask_of(nr_comp);
1288
1289 /* Use the type appropriate load */
1290 switch (t) {
1291 case nir_type_uint:
1292 case nir_type_bool:
1293 ins.load_store.op = midgard_op_ld_attr_32u;
1294 break;
1295 case nir_type_int:
1296 ins.load_store.op = midgard_op_ld_attr_32i;
1297 break;
1298 case nir_type_float:
1299 ins.load_store.op = midgard_op_ld_attr_32;
1300 break;
1301 default:
1302 unreachable("Attempted to load unknown type");
1303 break;
1304 }
1305
1306 emit_mir_instruction(ctx, ins);
1307 }
1308
1309 static void
1310 emit_sysval_read(compiler_context *ctx, nir_instr *instr,
1311 unsigned nr_components, unsigned offset)
1312 {
1313 nir_dest nir_dest;
1314
1315 /* Figure out which uniform this is */
1316 int sysval = panfrost_sysval_for_instr(instr, &nir_dest);
1317 void *val = _mesa_hash_table_u64_search(ctx->sysvals.sysval_to_id, sysval);
1318
1319 unsigned dest = nir_dest_index(&nir_dest);
1320
1321 /* Sysvals are prefix uniforms */
1322 unsigned uniform = ((uintptr_t) val) - 1;
1323
1324 /* Emit the read itself -- this is never indirect */
1325 midgard_instruction *ins =
1326 emit_ubo_read(ctx, instr, dest, (uniform * 16) + offset, NULL, 0, 0);
1327
1328 ins->mask = mask_of(nr_components);
1329 }
1330
1331 static unsigned
1332 compute_builtin_arg(nir_op op)
1333 {
1334 switch (op) {
1335 case nir_intrinsic_load_work_group_id:
1336 return 0x14;
1337 case nir_intrinsic_load_local_invocation_id:
1338 return 0x10;
1339 default:
1340 unreachable("Invalid compute paramater loaded");
1341 }
1342 }
1343
1344 static void
1345 emit_fragment_store(compiler_context *ctx, unsigned src, enum midgard_rt_id rt)
1346 {
1347 assert(rt < ARRAY_SIZE(ctx->writeout_branch));
1348
1349 midgard_instruction *br = ctx->writeout_branch[rt];
1350
1351 assert(!br);
1352
1353 emit_explicit_constant(ctx, src, src);
1354
1355 struct midgard_instruction ins =
1356 v_branch(false, false);
1357
1358 ins.writeout = true;
1359
1360 /* Add dependencies */
1361 ins.src[0] = src;
1362 ins.src_types[0] = nir_type_uint32;
1363 ins.constants.u32[0] = rt == MIDGARD_ZS_RT ?
1364 0xFF : (rt - MIDGARD_COLOR_RT0) * 0x100;
1365
1366 /* Emit the branch */
1367 br = emit_mir_instruction(ctx, ins);
1368 schedule_barrier(ctx);
1369 ctx->writeout_branch[rt] = br;
1370
1371 /* Push our current location = current block count - 1 = where we'll
1372 * jump to. Maybe a bit too clever for my own good */
1373
1374 br->branch.target_block = ctx->block_count - 1;
1375 }
1376
1377 static void
1378 emit_compute_builtin(compiler_context *ctx, nir_intrinsic_instr *instr)
1379 {
1380 unsigned reg = nir_dest_index(&instr->dest);
1381 midgard_instruction ins = m_ld_compute_id(reg, 0);
1382 ins.mask = mask_of(3);
1383 ins.swizzle[0][3] = COMPONENT_X; /* xyzx */
1384 ins.load_store.arg_1 = compute_builtin_arg(instr->intrinsic);
1385 emit_mir_instruction(ctx, ins);
1386 }
1387
1388 static unsigned
1389 vertex_builtin_arg(nir_op op)
1390 {
1391 switch (op) {
1392 case nir_intrinsic_load_vertex_id:
1393 return PAN_VERTEX_ID;
1394 case nir_intrinsic_load_instance_id:
1395 return PAN_INSTANCE_ID;
1396 default:
1397 unreachable("Invalid vertex builtin");
1398 }
1399 }
1400
1401 static void
1402 emit_vertex_builtin(compiler_context *ctx, nir_intrinsic_instr *instr)
1403 {
1404 unsigned reg = nir_dest_index(&instr->dest);
1405 emit_attr_read(ctx, reg, vertex_builtin_arg(instr->intrinsic), 1, nir_type_int);
1406 }
1407
1408 static void
1409 emit_control_barrier(compiler_context *ctx)
1410 {
1411 midgard_instruction ins = {
1412 .type = TAG_TEXTURE_4,
1413 .dest = ~0,
1414 .src = { ~0, ~0, ~0, ~0 },
1415 .texture = {
1416 .op = TEXTURE_OP_BARRIER,
1417
1418 /* TODO: optimize */
1419 .barrier_buffer = 1,
1420 .barrier_shared = 1
1421 }
1422 };
1423
1424 emit_mir_instruction(ctx, ins);
1425 }
1426
1427 static const nir_variable *
1428 search_var(struct exec_list *vars, unsigned driver_loc)
1429 {
1430 nir_foreach_variable(var, vars) {
1431 if (var->data.driver_location == driver_loc)
1432 return var;
1433 }
1434
1435 return NULL;
1436 }
1437
1438 static unsigned
1439 mir_get_branch_cond(nir_src *src, bool *invert)
1440 {
1441 /* Wrap it. No swizzle since it's a scalar */
1442
1443 nir_alu_src alu = {
1444 .src = *src
1445 };
1446
1447 *invert = pan_has_source_mod(&alu, nir_op_inot);
1448 return nir_src_index(NULL, &alu.src);
1449 }
1450
1451 static void
1452 emit_intrinsic(compiler_context *ctx, nir_intrinsic_instr *instr)
1453 {
1454 unsigned offset = 0, reg;
1455
1456 switch (instr->intrinsic) {
1457 case nir_intrinsic_discard_if:
1458 case nir_intrinsic_discard: {
1459 bool conditional = instr->intrinsic == nir_intrinsic_discard_if;
1460 struct midgard_instruction discard = v_branch(conditional, false);
1461 discard.branch.target_type = TARGET_DISCARD;
1462
1463 if (conditional) {
1464 discard.src[0] = mir_get_branch_cond(&instr->src[0],
1465 &discard.branch.invert_conditional);
1466 discard.src_types[0] = nir_type_uint32;
1467 }
1468
1469 emit_mir_instruction(ctx, discard);
1470 schedule_barrier(ctx);
1471
1472 break;
1473 }
1474
1475 case nir_intrinsic_load_uniform:
1476 case nir_intrinsic_load_ubo:
1477 case nir_intrinsic_load_global:
1478 case nir_intrinsic_load_shared:
1479 case nir_intrinsic_load_input:
1480 case nir_intrinsic_load_interpolated_input: {
1481 bool is_uniform = instr->intrinsic == nir_intrinsic_load_uniform;
1482 bool is_ubo = instr->intrinsic == nir_intrinsic_load_ubo;
1483 bool is_global = instr->intrinsic == nir_intrinsic_load_global;
1484 bool is_shared = instr->intrinsic == nir_intrinsic_load_shared;
1485 bool is_flat = instr->intrinsic == nir_intrinsic_load_input;
1486 bool is_interp = instr->intrinsic == nir_intrinsic_load_interpolated_input;
1487
1488 /* Get the base type of the intrinsic */
1489 /* TODO: Infer type? Does it matter? */
1490 nir_alu_type t =
1491 (is_ubo || is_global || is_shared) ? nir_type_uint :
1492 (is_interp) ? nir_type_float :
1493 nir_intrinsic_type(instr);
1494
1495 t = nir_alu_type_get_base_type(t);
1496
1497 if (!(is_ubo || is_global)) {
1498 offset = nir_intrinsic_base(instr);
1499 }
1500
1501 unsigned nr_comp = nir_intrinsic_dest_components(instr);
1502
1503 nir_src *src_offset = nir_get_io_offset_src(instr);
1504
1505 bool direct = nir_src_is_const(*src_offset);
1506 nir_src *indirect_offset = direct ? NULL : src_offset;
1507
1508 if (direct)
1509 offset += nir_src_as_uint(*src_offset);
1510
1511 /* We may need to apply a fractional offset */
1512 int component = (is_flat || is_interp) ?
1513 nir_intrinsic_component(instr) : 0;
1514 reg = nir_dest_index(&instr->dest);
1515
1516 if (is_uniform && !ctx->is_blend) {
1517 emit_ubo_read(ctx, &instr->instr, reg, (ctx->sysvals.sysval_count + offset) * 16, indirect_offset, 4, 0);
1518 } else if (is_ubo) {
1519 nir_src index = instr->src[0];
1520
1521 /* TODO: Is indirect block number possible? */
1522 assert(nir_src_is_const(index));
1523
1524 uint32_t uindex = nir_src_as_uint(index) + 1;
1525 emit_ubo_read(ctx, &instr->instr, reg, offset, indirect_offset, 0, uindex);
1526 } else if (is_global || is_shared) {
1527 emit_global(ctx, &instr->instr, true, reg, src_offset, is_shared);
1528 } else if (ctx->stage == MESA_SHADER_FRAGMENT && !ctx->is_blend) {
1529 emit_varying_read(ctx, reg, offset, nr_comp, component, indirect_offset, t, is_flat);
1530 } else if (ctx->is_blend) {
1531 /* For blend shaders, load the input color, which is
1532 * preloaded to r0 */
1533
1534 midgard_instruction move = v_mov(SSA_FIXED_REGISTER(0), reg);
1535 emit_mir_instruction(ctx, move);
1536 schedule_barrier(ctx);
1537 } else if (ctx->stage == MESA_SHADER_VERTEX) {
1538 emit_attr_read(ctx, reg, offset, nr_comp, t);
1539 } else {
1540 DBG("Unknown load\n");
1541 assert(0);
1542 }
1543
1544 break;
1545 }
1546
1547 /* Artefact of load_interpolated_input. TODO: other barycentric modes */
1548 case nir_intrinsic_load_barycentric_pixel:
1549 case nir_intrinsic_load_barycentric_centroid:
1550 break;
1551
1552 /* Reads 128-bit value raw off the tilebuffer during blending, tasty */
1553
1554 case nir_intrinsic_load_raw_output_pan:
1555 case nir_intrinsic_load_output_u8_as_fp16_pan:
1556 reg = nir_dest_index(&instr->dest);
1557 assert(ctx->is_blend);
1558
1559 /* T720 and below use different blend opcodes with slightly
1560 * different semantics than T760 and up */
1561
1562 midgard_instruction ld = m_ld_color_buffer_32u(reg, 0);
1563 bool old_blend = ctx->quirks & MIDGARD_OLD_BLEND;
1564
1565 if (instr->intrinsic == nir_intrinsic_load_output_u8_as_fp16_pan) {
1566 ld.load_store.op = old_blend ?
1567 midgard_op_ld_color_buffer_u8_as_fp16_old :
1568 midgard_op_ld_color_buffer_u8_as_fp16;
1569
1570 if (old_blend) {
1571 ld.load_store.address = 1;
1572 ld.load_store.arg_2 = 0x1E;
1573 }
1574
1575 for (unsigned c = 4; c < 16; ++c)
1576 ld.swizzle[0][c] = 0;
1577
1578 ld.dest_type = nir_type_float16;
1579 }
1580
1581 emit_mir_instruction(ctx, ld);
1582 break;
1583
1584 case nir_intrinsic_load_blend_const_color_rgba: {
1585 assert(ctx->is_blend);
1586 reg = nir_dest_index(&instr->dest);
1587
1588 /* Blend constants are embedded directly in the shader and
1589 * patched in, so we use some magic routing */
1590
1591 midgard_instruction ins = v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT), reg);
1592 ins.has_constants = true;
1593 ins.has_blend_constant = true;
1594 emit_mir_instruction(ctx, ins);
1595 break;
1596 }
1597
1598 case nir_intrinsic_store_zs_output_pan: {
1599 assert(ctx->stage == MESA_SHADER_FRAGMENT);
1600 emit_fragment_store(ctx, nir_src_index(ctx, &instr->src[0]),
1601 MIDGARD_ZS_RT);
1602
1603 midgard_instruction *br = ctx->writeout_branch[MIDGARD_ZS_RT];
1604
1605 if (!nir_intrinsic_component(instr))
1606 br->writeout_depth = true;
1607 if (nir_intrinsic_component(instr) ||
1608 instr->num_components)
1609 br->writeout_stencil = true;
1610 assert(br->writeout_depth | br->writeout_stencil);
1611 break;
1612 }
1613
1614 case nir_intrinsic_store_output:
1615 assert(nir_src_is_const(instr->src[1]) && "no indirect outputs");
1616
1617 offset = nir_intrinsic_base(instr) + nir_src_as_uint(instr->src[1]);
1618
1619 reg = nir_src_index(ctx, &instr->src[0]);
1620
1621 if (ctx->stage == MESA_SHADER_FRAGMENT) {
1622 const nir_variable *var;
1623 enum midgard_rt_id rt;
1624
1625 var = search_var(&ctx->nir->outputs,
1626 nir_intrinsic_base(instr));
1627 assert(var);
1628 if (var->data.location == FRAG_RESULT_COLOR)
1629 rt = MIDGARD_COLOR_RT0;
1630 else if (var->data.location >= FRAG_RESULT_DATA0)
1631 rt = MIDGARD_COLOR_RT0 + var->data.location -
1632 FRAG_RESULT_DATA0;
1633 else
1634 assert(0);
1635
1636 emit_fragment_store(ctx, reg, rt);
1637 } else if (ctx->stage == MESA_SHADER_VERTEX) {
1638 /* We should have been vectorized, though we don't
1639 * currently check that st_vary is emitted only once
1640 * per slot (this is relevant, since there's not a mask
1641 * parameter available on the store [set to 0 by the
1642 * blob]). We do respect the component by adjusting the
1643 * swizzle. If this is a constant source, we'll need to
1644 * emit that explicitly. */
1645
1646 emit_explicit_constant(ctx, reg, reg);
1647
1648 unsigned dst_component = nir_intrinsic_component(instr);
1649 unsigned nr_comp = nir_src_num_components(instr->src[0]);
1650
1651 midgard_instruction st = m_st_vary_32(reg, offset);
1652 st.load_store.arg_1 = 0x9E;
1653 st.load_store.arg_2 = 0x1E;
1654
1655 switch (nir_alu_type_get_base_type(nir_intrinsic_type(instr))) {
1656 case nir_type_uint:
1657 case nir_type_bool:
1658 st.load_store.op = midgard_op_st_vary_32u;
1659 break;
1660 case nir_type_int:
1661 st.load_store.op = midgard_op_st_vary_32i;
1662 break;
1663 case nir_type_float:
1664 st.load_store.op = midgard_op_st_vary_32;
1665 break;
1666 default:
1667 unreachable("Attempted to store unknown type");
1668 break;
1669 }
1670
1671 /* nir_intrinsic_component(store_intr) encodes the
1672 * destination component start. Source component offset
1673 * adjustment is taken care of in
1674 * install_registers_instr(), when offset_swizzle() is
1675 * called.
1676 */
1677 unsigned src_component = COMPONENT_X;
1678
1679 assert(nr_comp > 0);
1680 for (unsigned i = 0; i < ARRAY_SIZE(st.swizzle); ++i) {
1681 st.swizzle[0][i] = src_component;
1682 if (i >= dst_component && i < dst_component + nr_comp - 1)
1683 src_component++;
1684 }
1685
1686 emit_mir_instruction(ctx, st);
1687 } else {
1688 DBG("Unknown store\n");
1689 assert(0);
1690 }
1691
1692 break;
1693
1694 /* Special case of store_output for lowered blend shaders */
1695 case nir_intrinsic_store_raw_output_pan:
1696 assert (ctx->stage == MESA_SHADER_FRAGMENT);
1697 reg = nir_src_index(ctx, &instr->src[0]);
1698
1699 if (ctx->quirks & MIDGARD_OLD_BLEND) {
1700 /* Suppose reg = qr0.xyzw. That means 4 8-bit ---> 1 32-bit. So
1701 * reg = r0.x. We want to splatter. So we can do a 32-bit move
1702 * of:
1703 *
1704 * imov r0.xyzw, r0.xxxx
1705 */
1706
1707 unsigned expanded = make_compiler_temp(ctx);
1708
1709 midgard_instruction splatter = v_mov(reg, expanded);
1710
1711 for (unsigned c = 0; c < 16; ++c)
1712 splatter.swizzle[1][c] = 0;
1713
1714 emit_mir_instruction(ctx, splatter);
1715 emit_fragment_store(ctx, expanded, ctx->blend_rt);
1716 } else
1717 emit_fragment_store(ctx, reg, ctx->blend_rt);
1718
1719 break;
1720
1721 case nir_intrinsic_store_global:
1722 case nir_intrinsic_store_shared:
1723 reg = nir_src_index(ctx, &instr->src[0]);
1724 emit_explicit_constant(ctx, reg, reg);
1725
1726 emit_global(ctx, &instr->instr, false, reg, &instr->src[1], instr->intrinsic == nir_intrinsic_store_shared);
1727 break;
1728
1729 case nir_intrinsic_load_ssbo_address:
1730 emit_sysval_read(ctx, &instr->instr, 1, 0);
1731 break;
1732
1733 case nir_intrinsic_get_buffer_size:
1734 emit_sysval_read(ctx, &instr->instr, 1, 8);
1735 break;
1736
1737 case nir_intrinsic_load_viewport_scale:
1738 case nir_intrinsic_load_viewport_offset:
1739 case nir_intrinsic_load_num_work_groups:
1740 case nir_intrinsic_load_sampler_lod_parameters_pan:
1741 emit_sysval_read(ctx, &instr->instr, 3, 0);
1742 break;
1743
1744 case nir_intrinsic_load_work_group_id:
1745 case nir_intrinsic_load_local_invocation_id:
1746 emit_compute_builtin(ctx, instr);
1747 break;
1748
1749 case nir_intrinsic_load_vertex_id:
1750 case nir_intrinsic_load_instance_id:
1751 emit_vertex_builtin(ctx, instr);
1752 break;
1753
1754 case nir_intrinsic_memory_barrier_buffer:
1755 case nir_intrinsic_memory_barrier_shared:
1756 break;
1757
1758 case nir_intrinsic_control_barrier:
1759 schedule_barrier(ctx);
1760 emit_control_barrier(ctx);
1761 schedule_barrier(ctx);
1762 break;
1763
1764 default:
1765 fprintf(stderr, "Unhandled intrinsic %s\n", nir_intrinsic_infos[instr->intrinsic].name);
1766 assert(0);
1767 break;
1768 }
1769 }
1770
1771 static unsigned
1772 midgard_tex_format(enum glsl_sampler_dim dim)
1773 {
1774 switch (dim) {
1775 case GLSL_SAMPLER_DIM_1D:
1776 case GLSL_SAMPLER_DIM_BUF:
1777 return MALI_TEX_1D;
1778
1779 case GLSL_SAMPLER_DIM_2D:
1780 case GLSL_SAMPLER_DIM_EXTERNAL:
1781 case GLSL_SAMPLER_DIM_RECT:
1782 return MALI_TEX_2D;
1783
1784 case GLSL_SAMPLER_DIM_3D:
1785 return MALI_TEX_3D;
1786
1787 case GLSL_SAMPLER_DIM_CUBE:
1788 return MALI_TEX_CUBE;
1789
1790 default:
1791 DBG("Unknown sampler dim type\n");
1792 assert(0);
1793 return 0;
1794 }
1795 }
1796
1797 /* Tries to attach an explicit LOD or bias as a constant. Returns whether this
1798 * was successful */
1799
1800 static bool
1801 pan_attach_constant_bias(
1802 compiler_context *ctx,
1803 nir_src lod,
1804 midgard_texture_word *word)
1805 {
1806 /* To attach as constant, it has to *be* constant */
1807
1808 if (!nir_src_is_const(lod))
1809 return false;
1810
1811 float f = nir_src_as_float(lod);
1812
1813 /* Break into fixed-point */
1814 signed lod_int = f;
1815 float lod_frac = f - lod_int;
1816
1817 /* Carry over negative fractions */
1818 if (lod_frac < 0.0) {
1819 lod_int--;
1820 lod_frac += 1.0;
1821 }
1822
1823 /* Encode */
1824 word->bias = float_to_ubyte(lod_frac);
1825 word->bias_int = lod_int;
1826
1827 return true;
1828 }
1829
1830 static void
1831 emit_texop_native(compiler_context *ctx, nir_tex_instr *instr,
1832 unsigned midgard_texop)
1833 {
1834 /* TODO */
1835 //assert (!instr->sampler);
1836
1837 int texture_index = instr->texture_index;
1838 int sampler_index = texture_index;
1839
1840 nir_alu_type dest_base = nir_alu_type_get_base_type(instr->dest_type);
1841 nir_alu_type dest_type = dest_base | nir_dest_bit_size(instr->dest);
1842
1843 midgard_instruction ins = {
1844 .type = TAG_TEXTURE_4,
1845 .mask = 0xF,
1846 .dest = nir_dest_index(&instr->dest),
1847 .src = { ~0, ~0, ~0, ~0 },
1848 .dest_type = dest_type,
1849 .swizzle = SWIZZLE_IDENTITY_4,
1850 .texture = {
1851 .op = midgard_texop,
1852 .format = midgard_tex_format(instr->sampler_dim),
1853 .texture_handle = texture_index,
1854 .sampler_handle = sampler_index,
1855 .shadow = instr->is_shadow,
1856 }
1857 };
1858
1859 if (instr->is_shadow && !instr->is_new_style_shadow)
1860 for (int i = 0; i < 4; ++i)
1861 ins.swizzle[0][i] = COMPONENT_X;
1862
1863 /* We may need a temporary for the coordinate */
1864
1865 bool needs_temp_coord =
1866 (midgard_texop == TEXTURE_OP_TEXEL_FETCH) ||
1867 (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) ||
1868 (instr->is_shadow);
1869
1870 unsigned coords = needs_temp_coord ? make_compiler_temp_reg(ctx) : 0;
1871
1872 for (unsigned i = 0; i < instr->num_srcs; ++i) {
1873 int index = nir_src_index(ctx, &instr->src[i].src);
1874 unsigned nr_components = nir_src_num_components(instr->src[i].src);
1875 unsigned sz = nir_src_bit_size(instr->src[i].src);
1876 nir_alu_type T = nir_tex_instr_src_type(instr, i) | sz;
1877
1878 switch (instr->src[i].src_type) {
1879 case nir_tex_src_coord: {
1880 emit_explicit_constant(ctx, index, index);
1881
1882 unsigned coord_mask = mask_of(instr->coord_components);
1883
1884 bool flip_zw = (instr->sampler_dim == GLSL_SAMPLER_DIM_2D) && (coord_mask & (1 << COMPONENT_Z));
1885
1886 if (flip_zw)
1887 coord_mask ^= ((1 << COMPONENT_Z) | (1 << COMPONENT_W));
1888
1889 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
1890 /* texelFetch is undefined on samplerCube */
1891 assert(midgard_texop != TEXTURE_OP_TEXEL_FETCH);
1892
1893 /* For cubemaps, we use a special ld/st op to
1894 * select the face and copy the xy into the
1895 * texture register */
1896
1897 midgard_instruction ld = m_ld_cubemap_coords(coords, 0);
1898 ld.src[1] = index;
1899 ld.src_types[1] = T;
1900 ld.mask = 0x3; /* xy */
1901 ld.load_store.arg_1 = 0x20;
1902 ld.swizzle[1][3] = COMPONENT_X;
1903 emit_mir_instruction(ctx, ld);
1904
1905 /* xyzw -> xyxx */
1906 ins.swizzle[1][2] = instr->is_shadow ? COMPONENT_Z : COMPONENT_X;
1907 ins.swizzle[1][3] = COMPONENT_X;
1908 } else if (needs_temp_coord) {
1909 /* mov coord_temp, coords */
1910 midgard_instruction mov = v_mov(index, coords);
1911 mov.mask = coord_mask;
1912
1913 if (flip_zw)
1914 mov.swizzle[1][COMPONENT_W] = COMPONENT_Z;
1915
1916 emit_mir_instruction(ctx, mov);
1917 } else {
1918 coords = index;
1919 }
1920
1921 ins.src[1] = coords;
1922 ins.src_types[1] = T;
1923
1924 /* Texelfetch coordinates uses all four elements
1925 * (xyz/index) regardless of texture dimensionality,
1926 * which means it's necessary to zero the unused
1927 * components to keep everything happy */
1928
1929 if (midgard_texop == TEXTURE_OP_TEXEL_FETCH) {
1930 /* mov index.zw, #0, or generalized */
1931 midgard_instruction mov =
1932 v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT), coords);
1933 mov.has_constants = true;
1934 mov.mask = coord_mask ^ 0xF;
1935 emit_mir_instruction(ctx, mov);
1936 }
1937
1938 if (instr->sampler_dim == GLSL_SAMPLER_DIM_2D) {
1939 /* Array component in w but NIR wants it in z,
1940 * but if we have a temp coord we already fixed
1941 * that up */
1942
1943 if (nr_components == 3) {
1944 ins.swizzle[1][2] = COMPONENT_Z;
1945 ins.swizzle[1][3] = needs_temp_coord ? COMPONENT_W : COMPONENT_Z;
1946 } else if (nr_components == 2) {
1947 ins.swizzle[1][2] =
1948 instr->is_shadow ? COMPONENT_Z : COMPONENT_X;
1949 ins.swizzle[1][3] = COMPONENT_X;
1950 } else
1951 unreachable("Invalid texture 2D components");
1952 }
1953
1954 if (midgard_texop == TEXTURE_OP_TEXEL_FETCH) {
1955 /* We zeroed */
1956 ins.swizzle[1][2] = COMPONENT_Z;
1957 ins.swizzle[1][3] = COMPONENT_W;
1958 }
1959
1960 break;
1961 }
1962
1963 case nir_tex_src_bias:
1964 case nir_tex_src_lod: {
1965 /* Try as a constant if we can */
1966
1967 bool is_txf = midgard_texop == TEXTURE_OP_TEXEL_FETCH;
1968 if (!is_txf && pan_attach_constant_bias(ctx, instr->src[i].src, &ins.texture))
1969 break;
1970
1971 ins.texture.lod_register = true;
1972 ins.src[2] = index;
1973 ins.src_types[2] = T;
1974
1975 for (unsigned c = 0; c < MIR_VEC_COMPONENTS; ++c)
1976 ins.swizzle[2][c] = COMPONENT_X;
1977
1978 emit_explicit_constant(ctx, index, index);
1979
1980 break;
1981 };
1982
1983 case nir_tex_src_offset: {
1984 ins.texture.offset_register = true;
1985 ins.src[3] = index;
1986 ins.src_types[3] = T;
1987
1988 for (unsigned c = 0; c < MIR_VEC_COMPONENTS; ++c)
1989 ins.swizzle[3][c] = (c > COMPONENT_Z) ? 0 : c;
1990
1991 emit_explicit_constant(ctx, index, index);
1992 break;
1993 };
1994
1995 case nir_tex_src_comparator: {
1996 unsigned comp = COMPONENT_Z;
1997
1998 /* mov coord_temp.foo, coords */
1999 midgard_instruction mov = v_mov(index, coords);
2000 mov.mask = 1 << comp;
2001
2002 for (unsigned i = 0; i < MIR_VEC_COMPONENTS; ++i)
2003 mov.swizzle[1][i] = COMPONENT_X;
2004
2005 emit_mir_instruction(ctx, mov);
2006 break;
2007 }
2008
2009 default: {
2010 fprintf(stderr, "Unknown texture source type: %d\n", instr->src[i].src_type);
2011 assert(0);
2012 }
2013 }
2014 }
2015
2016 emit_mir_instruction(ctx, ins);
2017 }
2018
2019 static void
2020 emit_tex(compiler_context *ctx, nir_tex_instr *instr)
2021 {
2022 switch (instr->op) {
2023 case nir_texop_tex:
2024 case nir_texop_txb:
2025 emit_texop_native(ctx, instr, TEXTURE_OP_NORMAL);
2026 break;
2027 case nir_texop_txl:
2028 emit_texop_native(ctx, instr, TEXTURE_OP_LOD);
2029 break;
2030 case nir_texop_txf:
2031 emit_texop_native(ctx, instr, TEXTURE_OP_TEXEL_FETCH);
2032 break;
2033 case nir_texop_txs:
2034 emit_sysval_read(ctx, &instr->instr, 4, 0);
2035 break;
2036 default: {
2037 fprintf(stderr, "Unhandled texture op: %d\n", instr->op);
2038 assert(0);
2039 }
2040 }
2041 }
2042
2043 static void
2044 emit_jump(compiler_context *ctx, nir_jump_instr *instr)
2045 {
2046 switch (instr->type) {
2047 case nir_jump_break: {
2048 /* Emit a branch out of the loop */
2049 struct midgard_instruction br = v_branch(false, false);
2050 br.branch.target_type = TARGET_BREAK;
2051 br.branch.target_break = ctx->current_loop_depth;
2052 emit_mir_instruction(ctx, br);
2053 break;
2054 }
2055
2056 default:
2057 DBG("Unknown jump type %d\n", instr->type);
2058 break;
2059 }
2060 }
2061
2062 static void
2063 emit_instr(compiler_context *ctx, struct nir_instr *instr)
2064 {
2065 switch (instr->type) {
2066 case nir_instr_type_load_const:
2067 emit_load_const(ctx, nir_instr_as_load_const(instr));
2068 break;
2069
2070 case nir_instr_type_intrinsic:
2071 emit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
2072 break;
2073
2074 case nir_instr_type_alu:
2075 emit_alu(ctx, nir_instr_as_alu(instr));
2076 break;
2077
2078 case nir_instr_type_tex:
2079 emit_tex(ctx, nir_instr_as_tex(instr));
2080 break;
2081
2082 case nir_instr_type_jump:
2083 emit_jump(ctx, nir_instr_as_jump(instr));
2084 break;
2085
2086 case nir_instr_type_ssa_undef:
2087 /* Spurious */
2088 break;
2089
2090 default:
2091 DBG("Unhandled instruction type\n");
2092 break;
2093 }
2094 }
2095
2096
2097 /* ALU instructions can inline or embed constants, which decreases register
2098 * pressure and saves space. */
2099
2100 #define CONDITIONAL_ATTACH(idx) { \
2101 void *entry = _mesa_hash_table_u64_search(ctx->ssa_constants, alu->src[idx] + 1); \
2102 \
2103 if (entry) { \
2104 attach_constants(ctx, alu, entry, alu->src[idx] + 1); \
2105 alu->src[idx] = SSA_FIXED_REGISTER(REGISTER_CONSTANT); \
2106 } \
2107 }
2108
2109 static void
2110 inline_alu_constants(compiler_context *ctx, midgard_block *block)
2111 {
2112 mir_foreach_instr_in_block(block, alu) {
2113 /* Other instructions cannot inline constants */
2114 if (alu->type != TAG_ALU_4) continue;
2115 if (alu->compact_branch) continue;
2116
2117 /* If there is already a constant here, we can do nothing */
2118 if (alu->has_constants) continue;
2119
2120 CONDITIONAL_ATTACH(0);
2121
2122 if (!alu->has_constants) {
2123 CONDITIONAL_ATTACH(1)
2124 } else if (!alu->inline_constant) {
2125 /* Corner case: _two_ vec4 constants, for instance with a
2126 * csel. For this case, we can only use a constant
2127 * register for one, we'll have to emit a move for the
2128 * other. */
2129
2130 void *entry = _mesa_hash_table_u64_search(ctx->ssa_constants, alu->src[1] + 1);
2131 unsigned scratch = make_compiler_temp(ctx);
2132
2133 if (entry) {
2134 midgard_instruction ins = v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT), scratch);
2135 attach_constants(ctx, &ins, entry, alu->src[1] + 1);
2136
2137 /* Set the source */
2138 alu->src[1] = scratch;
2139
2140 /* Inject us -before- the last instruction which set r31 */
2141 mir_insert_instruction_before(ctx, mir_prev_op(alu), ins);
2142 }
2143 }
2144 }
2145 }
2146
2147 /* Midgard supports two types of constants, embedded constants (128-bit) and
2148 * inline constants (16-bit). Sometimes, especially with scalar ops, embedded
2149 * constants can be demoted to inline constants, for space savings and
2150 * sometimes a performance boost */
2151
2152 static void
2153 embedded_to_inline_constant(compiler_context *ctx, midgard_block *block)
2154 {
2155 mir_foreach_instr_in_block(block, ins) {
2156 if (!ins->has_constants) continue;
2157 if (ins->has_inline_constant) continue;
2158
2159 /* Blend constants must not be inlined by definition */
2160 if (ins->has_blend_constant) continue;
2161
2162 /* We can inline 32-bit (sometimes) or 16-bit (usually) */
2163 bool is_16 = ins->alu.reg_mode == midgard_reg_mode_16;
2164 bool is_32 = ins->alu.reg_mode == midgard_reg_mode_32;
2165
2166 if (!(is_16 || is_32))
2167 continue;
2168
2169 /* src1 cannot be an inline constant due to encoding
2170 * restrictions. So, if possible we try to flip the arguments
2171 * in that case */
2172
2173 int op = ins->alu.op;
2174
2175 if (ins->src[0] == SSA_FIXED_REGISTER(REGISTER_CONSTANT) &&
2176 alu_opcode_props[op].props & OP_COMMUTES) {
2177 mir_flip(ins);
2178 }
2179
2180 if (ins->src[1] == SSA_FIXED_REGISTER(REGISTER_CONSTANT)) {
2181 /* Component is from the swizzle. Take a nonzero component */
2182 assert(ins->mask);
2183 unsigned first_comp = ffs(ins->mask) - 1;
2184 unsigned component = ins->swizzle[1][first_comp];
2185
2186 /* Scale constant appropriately, if we can legally */
2187 uint16_t scaled_constant = 0;
2188
2189 if (is_16) {
2190 scaled_constant = ins->constants.u16[component];
2191 } else if (midgard_is_integer_op(op)) {
2192 scaled_constant = ins->constants.u32[component];
2193
2194 /* Constant overflow after resize */
2195 if (scaled_constant != ins->constants.u32[component])
2196 continue;
2197 } else {
2198 float original = ins->constants.f32[component];
2199 scaled_constant = _mesa_float_to_half(original);
2200
2201 /* Check for loss of precision. If this is
2202 * mediump, we don't care, but for a highp
2203 * shader, we need to pay attention. NIR
2204 * doesn't yet tell us which mode we're in!
2205 * Practically this prevents most constants
2206 * from being inlined, sadly. */
2207
2208 float fp32 = _mesa_half_to_float(scaled_constant);
2209
2210 if (fp32 != original)
2211 continue;
2212 }
2213
2214 /* Should've been const folded */
2215 if (ins->src_abs[1] || ins->src_neg[1])
2216 continue;
2217
2218 /* Make sure that the constant is not itself a vector
2219 * by checking if all accessed values are the same. */
2220
2221 const midgard_constants *cons = &ins->constants;
2222 uint32_t value = is_16 ? cons->u16[component] : cons->u32[component];
2223
2224 bool is_vector = false;
2225 unsigned mask = effective_writemask(&ins->alu, ins->mask);
2226
2227 for (unsigned c = 0; c < MIR_VEC_COMPONENTS; ++c) {
2228 /* We only care if this component is actually used */
2229 if (!(mask & (1 << c)))
2230 continue;
2231
2232 uint32_t test = is_16 ?
2233 cons->u16[ins->swizzle[1][c]] :
2234 cons->u32[ins->swizzle[1][c]];
2235
2236 if (test != value) {
2237 is_vector = true;
2238 break;
2239 }
2240 }
2241
2242 if (is_vector)
2243 continue;
2244
2245 /* Get rid of the embedded constant */
2246 ins->has_constants = false;
2247 ins->src[1] = ~0;
2248 ins->has_inline_constant = true;
2249 ins->inline_constant = scaled_constant;
2250 }
2251 }
2252 }
2253
2254 /* Dead code elimination for branches at the end of a block - only one branch
2255 * per block is legal semantically */
2256
2257 static void
2258 midgard_cull_dead_branch(compiler_context *ctx, midgard_block *block)
2259 {
2260 bool branched = false;
2261
2262 mir_foreach_instr_in_block_safe(block, ins) {
2263 if (!midgard_is_branch_unit(ins->unit)) continue;
2264
2265 if (branched)
2266 mir_remove_instruction(ins);
2267
2268 branched = true;
2269 }
2270 }
2271
2272 static unsigned
2273 emit_fragment_epilogue(compiler_context *ctx, unsigned rt)
2274 {
2275 /* Loop to ourselves */
2276 midgard_instruction *br = ctx->writeout_branch[rt];
2277 struct midgard_instruction ins = v_branch(false, false);
2278 ins.writeout = true;
2279 ins.writeout_depth = br->writeout_depth;
2280 ins.writeout_stencil = br->writeout_stencil;
2281 ins.branch.target_block = ctx->block_count - 1;
2282 ins.constants.u32[0] = br->constants.u32[0];
2283 emit_mir_instruction(ctx, ins);
2284
2285 ctx->current_block->epilogue = true;
2286 schedule_barrier(ctx);
2287 return ins.branch.target_block;
2288 }
2289
2290 static midgard_block *
2291 emit_block(compiler_context *ctx, nir_block *block)
2292 {
2293 midgard_block *this_block = ctx->after_block;
2294 ctx->after_block = NULL;
2295
2296 if (!this_block)
2297 this_block = create_empty_block(ctx);
2298
2299 list_addtail(&this_block->base.link, &ctx->blocks);
2300
2301 this_block->scheduled = false;
2302 ++ctx->block_count;
2303
2304 /* Set up current block */
2305 list_inithead(&this_block->base.instructions);
2306 ctx->current_block = this_block;
2307
2308 nir_foreach_instr(instr, block) {
2309 emit_instr(ctx, instr);
2310 ++ctx->instruction_count;
2311 }
2312
2313 return this_block;
2314 }
2315
2316 static midgard_block *emit_cf_list(struct compiler_context *ctx, struct exec_list *list);
2317
2318 static void
2319 emit_if(struct compiler_context *ctx, nir_if *nif)
2320 {
2321 midgard_block *before_block = ctx->current_block;
2322
2323 /* Speculatively emit the branch, but we can't fill it in until later */
2324 bool inv = false;
2325 EMIT(branch, true, true);
2326 midgard_instruction *then_branch = mir_last_in_block(ctx->current_block);
2327 then_branch->src[0] = mir_get_branch_cond(&nif->condition, &inv);
2328 then_branch->src_types[0] = nir_type_uint32;
2329 then_branch->branch.invert_conditional = !inv;
2330
2331 /* Emit the two subblocks. */
2332 midgard_block *then_block = emit_cf_list(ctx, &nif->then_list);
2333 midgard_block *end_then_block = ctx->current_block;
2334
2335 /* Emit a jump from the end of the then block to the end of the else */
2336 EMIT(branch, false, false);
2337 midgard_instruction *then_exit = mir_last_in_block(ctx->current_block);
2338
2339 /* Emit second block, and check if it's empty */
2340
2341 int else_idx = ctx->block_count;
2342 int count_in = ctx->instruction_count;
2343 midgard_block *else_block = emit_cf_list(ctx, &nif->else_list);
2344 midgard_block *end_else_block = ctx->current_block;
2345 int after_else_idx = ctx->block_count;
2346
2347 /* Now that we have the subblocks emitted, fix up the branches */
2348
2349 assert(then_block);
2350 assert(else_block);
2351
2352 if (ctx->instruction_count == count_in) {
2353 /* The else block is empty, so don't emit an exit jump */
2354 mir_remove_instruction(then_exit);
2355 then_branch->branch.target_block = after_else_idx;
2356 } else {
2357 then_branch->branch.target_block = else_idx;
2358 then_exit->branch.target_block = after_else_idx;
2359 }
2360
2361 /* Wire up the successors */
2362
2363 ctx->after_block = create_empty_block(ctx);
2364
2365 pan_block_add_successor(&before_block->base, &then_block->base);
2366 pan_block_add_successor(&before_block->base, &else_block->base);
2367
2368 pan_block_add_successor(&end_then_block->base, &ctx->after_block->base);
2369 pan_block_add_successor(&end_else_block->base, &ctx->after_block->base);
2370 }
2371
2372 static void
2373 emit_loop(struct compiler_context *ctx, nir_loop *nloop)
2374 {
2375 /* Remember where we are */
2376 midgard_block *start_block = ctx->current_block;
2377
2378 /* Allocate a loop number, growing the current inner loop depth */
2379 int loop_idx = ++ctx->current_loop_depth;
2380
2381 /* Get index from before the body so we can loop back later */
2382 int start_idx = ctx->block_count;
2383
2384 /* Emit the body itself */
2385 midgard_block *loop_block = emit_cf_list(ctx, &nloop->body);
2386
2387 /* Branch back to loop back */
2388 struct midgard_instruction br_back = v_branch(false, false);
2389 br_back.branch.target_block = start_idx;
2390 emit_mir_instruction(ctx, br_back);
2391
2392 /* Mark down that branch in the graph. */
2393 pan_block_add_successor(&start_block->base, &loop_block->base);
2394 pan_block_add_successor(&ctx->current_block->base, &loop_block->base);
2395
2396 /* Find the index of the block about to follow us (note: we don't add
2397 * one; blocks are 0-indexed so we get a fencepost problem) */
2398 int break_block_idx = ctx->block_count;
2399
2400 /* Fix up the break statements we emitted to point to the right place,
2401 * now that we can allocate a block number for them */
2402 ctx->after_block = create_empty_block(ctx);
2403
2404 mir_foreach_block_from(ctx, start_block, _block) {
2405 mir_foreach_instr_in_block(((midgard_block *) _block), ins) {
2406 if (ins->type != TAG_ALU_4) continue;
2407 if (!ins->compact_branch) continue;
2408
2409 /* We found a branch -- check the type to see if we need to do anything */
2410 if (ins->branch.target_type != TARGET_BREAK) continue;
2411
2412 /* It's a break! Check if it's our break */
2413 if (ins->branch.target_break != loop_idx) continue;
2414
2415 /* Okay, cool, we're breaking out of this loop.
2416 * Rewrite from a break to a goto */
2417
2418 ins->branch.target_type = TARGET_GOTO;
2419 ins->branch.target_block = break_block_idx;
2420
2421 pan_block_add_successor(_block, &ctx->after_block->base);
2422 }
2423 }
2424
2425 /* Now that we've finished emitting the loop, free up the depth again
2426 * so we play nice with recursion amid nested loops */
2427 --ctx->current_loop_depth;
2428
2429 /* Dump loop stats */
2430 ++ctx->loop_count;
2431 }
2432
2433 static midgard_block *
2434 emit_cf_list(struct compiler_context *ctx, struct exec_list *list)
2435 {
2436 midgard_block *start_block = NULL;
2437
2438 foreach_list_typed(nir_cf_node, node, node, list) {
2439 switch (node->type) {
2440 case nir_cf_node_block: {
2441 midgard_block *block = emit_block(ctx, nir_cf_node_as_block(node));
2442
2443 if (!start_block)
2444 start_block = block;
2445
2446 break;
2447 }
2448
2449 case nir_cf_node_if:
2450 emit_if(ctx, nir_cf_node_as_if(node));
2451 break;
2452
2453 case nir_cf_node_loop:
2454 emit_loop(ctx, nir_cf_node_as_loop(node));
2455 break;
2456
2457 case nir_cf_node_function:
2458 assert(0);
2459 break;
2460 }
2461 }
2462
2463 return start_block;
2464 }
2465
2466 /* Due to lookahead, we need to report the first tag executed in the command
2467 * stream and in branch targets. An initial block might be empty, so iterate
2468 * until we find one that 'works' */
2469
2470 static unsigned
2471 midgard_get_first_tag_from_block(compiler_context *ctx, unsigned block_idx)
2472 {
2473 midgard_block *initial_block = mir_get_block(ctx, block_idx);
2474
2475 mir_foreach_block_from(ctx, initial_block, _v) {
2476 midgard_block *v = (midgard_block *) _v;
2477 if (v->quadword_count) {
2478 midgard_bundle *initial_bundle =
2479 util_dynarray_element(&v->bundles, midgard_bundle, 0);
2480
2481 return initial_bundle->tag;
2482 }
2483 }
2484
2485 /* Default to a tag 1 which will break from the shader, in case we jump
2486 * to the exit block (i.e. `return` in a compute shader) */
2487
2488 return 1;
2489 }
2490
2491 /* For each fragment writeout instruction, generate a writeout loop to
2492 * associate with it */
2493
2494 static void
2495 mir_add_writeout_loops(compiler_context *ctx)
2496 {
2497 for (unsigned rt = 0; rt < ARRAY_SIZE(ctx->writeout_branch); ++rt) {
2498 midgard_instruction *br = ctx->writeout_branch[rt];
2499 if (!br) continue;
2500
2501 unsigned popped = br->branch.target_block;
2502 pan_block_add_successor(&(mir_get_block(ctx, popped - 1)->base), &ctx->current_block->base);
2503 br->branch.target_block = emit_fragment_epilogue(ctx, rt);
2504 br->branch.target_type = TARGET_GOTO;
2505
2506 /* If we have more RTs, we'll need to restore back after our
2507 * loop terminates */
2508
2509 if ((rt + 1) < ARRAY_SIZE(ctx->writeout_branch) && ctx->writeout_branch[rt + 1]) {
2510 midgard_instruction uncond = v_branch(false, false);
2511 uncond.branch.target_block = popped;
2512 uncond.branch.target_type = TARGET_GOTO;
2513 emit_mir_instruction(ctx, uncond);
2514 pan_block_add_successor(&ctx->current_block->base, &(mir_get_block(ctx, popped)->base));
2515 schedule_barrier(ctx);
2516 } else {
2517 /* We're last, so we can terminate here */
2518 br->last_writeout = true;
2519 }
2520 }
2521 }
2522
2523 int
2524 midgard_compile_shader_nir(nir_shader *nir, panfrost_program *program, bool is_blend, unsigned blend_rt, unsigned gpu_id, bool shaderdb)
2525 {
2526 struct util_dynarray *compiled = &program->compiled;
2527
2528 midgard_debug = debug_get_option_midgard_debug();
2529
2530 /* TODO: Bound against what? */
2531 compiler_context *ctx = rzalloc(NULL, compiler_context);
2532
2533 ctx->nir = nir;
2534 ctx->stage = nir->info.stage;
2535 ctx->is_blend = is_blend;
2536 ctx->alpha_ref = program->alpha_ref;
2537 ctx->blend_rt = MIDGARD_COLOR_RT0 + blend_rt;
2538 ctx->quirks = midgard_get_quirks(gpu_id);
2539
2540 /* Start off with a safe cutoff, allowing usage of all 16 work
2541 * registers. Later, we'll promote uniform reads to uniform registers
2542 * if we determine it is beneficial to do so */
2543 ctx->uniform_cutoff = 8;
2544
2545 /* Initialize at a global (not block) level hash tables */
2546
2547 ctx->ssa_constants = _mesa_hash_table_u64_create(NULL);
2548 ctx->hash_to_temp = _mesa_hash_table_u64_create(NULL);
2549
2550 /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
2551 * (so we don't accidentally duplicate the epilogue since mesa/st has
2552 * messed with our I/O quite a bit already) */
2553
2554 NIR_PASS_V(nir, nir_lower_vars_to_ssa);
2555
2556 if (ctx->stage == MESA_SHADER_VERTEX) {
2557 NIR_PASS_V(nir, nir_lower_viewport_transform);
2558 NIR_PASS_V(nir, nir_lower_point_size, 1.0, 1024.0);
2559 }
2560
2561 NIR_PASS_V(nir, nir_lower_var_copies);
2562 NIR_PASS_V(nir, nir_lower_vars_to_ssa);
2563 NIR_PASS_V(nir, nir_split_var_copies);
2564 NIR_PASS_V(nir, nir_lower_var_copies);
2565 NIR_PASS_V(nir, nir_lower_global_vars_to_local);
2566 NIR_PASS_V(nir, nir_lower_var_copies);
2567 NIR_PASS_V(nir, nir_lower_vars_to_ssa);
2568
2569 NIR_PASS_V(nir, nir_lower_io, nir_var_all, glsl_type_size, 0);
2570 NIR_PASS_V(nir, nir_lower_ssbo);
2571 NIR_PASS_V(nir, midgard_nir_lower_zs_store);
2572
2573 /* Optimisation passes */
2574
2575 optimise_nir(nir, ctx->quirks);
2576
2577 if (midgard_debug & MIDGARD_DBG_SHADERS) {
2578 nir_print_shader(nir, stdout);
2579 }
2580
2581 /* Assign sysvals and counts, now that we're sure
2582 * (post-optimisation) */
2583
2584 panfrost_nir_assign_sysvals(&ctx->sysvals, nir);
2585 program->sysval_count = ctx->sysvals.sysval_count;
2586 memcpy(program->sysvals, ctx->sysvals.sysvals, sizeof(ctx->sysvals.sysvals[0]) * ctx->sysvals.sysval_count);
2587
2588 nir_foreach_function(func, nir) {
2589 if (!func->impl)
2590 continue;
2591
2592 list_inithead(&ctx->blocks);
2593 ctx->block_count = 0;
2594 ctx->func = func;
2595 ctx->already_emitted = calloc(BITSET_WORDS(func->impl->ssa_alloc), sizeof(BITSET_WORD));
2596
2597 emit_cf_list(ctx, &func->impl->body);
2598 free(ctx->already_emitted);
2599 break; /* TODO: Multi-function shaders */
2600 }
2601
2602 util_dynarray_init(compiled, NULL);
2603
2604 /* Per-block lowering before opts */
2605
2606 mir_foreach_block(ctx, _block) {
2607 midgard_block *block = (midgard_block *) _block;
2608 inline_alu_constants(ctx, block);
2609 embedded_to_inline_constant(ctx, block);
2610 }
2611 /* MIR-level optimizations */
2612
2613 bool progress = false;
2614
2615 do {
2616 progress = false;
2617 progress |= midgard_opt_dead_code_eliminate(ctx);
2618
2619 mir_foreach_block(ctx, _block) {
2620 midgard_block *block = (midgard_block *) _block;
2621 progress |= midgard_opt_copy_prop(ctx, block);
2622 progress |= midgard_opt_combine_projection(ctx, block);
2623 progress |= midgard_opt_varying_projection(ctx, block);
2624 }
2625 } while (progress);
2626
2627 mir_foreach_block(ctx, _block) {
2628 midgard_block *block = (midgard_block *) _block;
2629 midgard_lower_derivatives(ctx, block);
2630 midgard_cull_dead_branch(ctx, block);
2631 }
2632
2633 if (ctx->stage == MESA_SHADER_FRAGMENT)
2634 mir_add_writeout_loops(ctx);
2635
2636 /* Analyze now that the code is known but before scheduling creates
2637 * pipeline registers which are harder to track */
2638 mir_analyze_helper_terminate(ctx);
2639 mir_analyze_helper_requirements(ctx);
2640
2641 /* Schedule! */
2642 midgard_schedule_program(ctx);
2643 mir_ra(ctx);
2644
2645 /* Now that all the bundles are scheduled and we can calculate block
2646 * sizes, emit actual branch instructions rather than placeholders */
2647
2648 int br_block_idx = 0;
2649
2650 mir_foreach_block(ctx, _block) {
2651 midgard_block *block = (midgard_block *) _block;
2652 util_dynarray_foreach(&block->bundles, midgard_bundle, bundle) {
2653 for (int c = 0; c < bundle->instruction_count; ++c) {
2654 midgard_instruction *ins = bundle->instructions[c];
2655
2656 if (!midgard_is_branch_unit(ins->unit)) continue;
2657
2658 /* Parse some basic branch info */
2659 bool is_compact = ins->unit == ALU_ENAB_BR_COMPACT;
2660 bool is_conditional = ins->branch.conditional;
2661 bool is_inverted = ins->branch.invert_conditional;
2662 bool is_discard = ins->branch.target_type == TARGET_DISCARD;
2663 bool is_writeout = ins->writeout;
2664
2665 /* Determine the block we're jumping to */
2666 int target_number = ins->branch.target_block;
2667
2668 /* Report the destination tag */
2669 int dest_tag = is_discard ? 0 : midgard_get_first_tag_from_block(ctx, target_number);
2670
2671 /* Count up the number of quadwords we're
2672 * jumping over = number of quadwords until
2673 * (br_block_idx, target_number) */
2674
2675 int quadword_offset = 0;
2676
2677 if (is_discard) {
2678 /* Ignored */
2679 } else if (target_number > br_block_idx) {
2680 /* Jump forward */
2681
2682 for (int idx = br_block_idx + 1; idx < target_number; ++idx) {
2683 midgard_block *blk = mir_get_block(ctx, idx);
2684 assert(blk);
2685
2686 quadword_offset += blk->quadword_count;
2687 }
2688 } else {
2689 /* Jump backwards */
2690
2691 for (int idx = br_block_idx; idx >= target_number; --idx) {
2692 midgard_block *blk = mir_get_block(ctx, idx);
2693 assert(blk);
2694
2695 quadword_offset -= blk->quadword_count;
2696 }
2697 }
2698
2699 /* Unconditional extended branches (far jumps)
2700 * have issues, so we always use a conditional
2701 * branch, setting the condition to always for
2702 * unconditional. For compact unconditional
2703 * branches, cond isn't used so it doesn't
2704 * matter what we pick. */
2705
2706 midgard_condition cond =
2707 !is_conditional ? midgard_condition_always :
2708 is_inverted ? midgard_condition_false :
2709 midgard_condition_true;
2710
2711 midgard_jmp_writeout_op op =
2712 is_discard ? midgard_jmp_writeout_op_discard :
2713 is_writeout ? midgard_jmp_writeout_op_writeout :
2714 (is_compact && !is_conditional) ? midgard_jmp_writeout_op_branch_uncond :
2715 midgard_jmp_writeout_op_branch_cond;
2716
2717 if (!is_compact) {
2718 midgard_branch_extended branch =
2719 midgard_create_branch_extended(
2720 cond, op,
2721 dest_tag,
2722 quadword_offset);
2723
2724 memcpy(&ins->branch_extended, &branch, sizeof(branch));
2725 } else if (is_conditional || is_discard) {
2726 midgard_branch_cond branch = {
2727 .op = op,
2728 .dest_tag = dest_tag,
2729 .offset = quadword_offset,
2730 .cond = cond
2731 };
2732
2733 assert(branch.offset == quadword_offset);
2734
2735 memcpy(&ins->br_compact, &branch, sizeof(branch));
2736 } else {
2737 assert(op == midgard_jmp_writeout_op_branch_uncond);
2738
2739 midgard_branch_uncond branch = {
2740 .op = op,
2741 .dest_tag = dest_tag,
2742 .offset = quadword_offset,
2743 .unknown = 1
2744 };
2745
2746 assert(branch.offset == quadword_offset);
2747
2748 memcpy(&ins->br_compact, &branch, sizeof(branch));
2749 }
2750 }
2751 }
2752
2753 ++br_block_idx;
2754 }
2755
2756 /* Emit flat binary from the instruction arrays. Iterate each block in
2757 * sequence. Save instruction boundaries such that lookahead tags can
2758 * be assigned easily */
2759
2760 /* Cache _all_ bundles in source order for lookahead across failed branches */
2761
2762 int bundle_count = 0;
2763 mir_foreach_block(ctx, _block) {
2764 midgard_block *block = (midgard_block *) _block;
2765 bundle_count += block->bundles.size / sizeof(midgard_bundle);
2766 }
2767 midgard_bundle **source_order_bundles = malloc(sizeof(midgard_bundle *) * bundle_count);
2768 int bundle_idx = 0;
2769 mir_foreach_block(ctx, _block) {
2770 midgard_block *block = (midgard_block *) _block;
2771 util_dynarray_foreach(&block->bundles, midgard_bundle, bundle) {
2772 source_order_bundles[bundle_idx++] = bundle;
2773 }
2774 }
2775
2776 int current_bundle = 0;
2777
2778 /* Midgard prefetches instruction types, so during emission we
2779 * need to lookahead. Unless this is the last instruction, in
2780 * which we return 1. */
2781
2782 mir_foreach_block(ctx, _block) {
2783 midgard_block *block = (midgard_block *) _block;
2784 mir_foreach_bundle_in_block(block, bundle) {
2785 int lookahead = 1;
2786
2787 if (!bundle->last_writeout && (current_bundle + 1 < bundle_count))
2788 lookahead = source_order_bundles[current_bundle + 1]->tag;
2789
2790 emit_binary_bundle(ctx, bundle, compiled, lookahead);
2791 ++current_bundle;
2792 }
2793
2794 /* TODO: Free deeper */
2795 //util_dynarray_fini(&block->instructions);
2796 }
2797
2798 free(source_order_bundles);
2799
2800 /* Report the very first tag executed */
2801 program->first_tag = midgard_get_first_tag_from_block(ctx, 0);
2802
2803 /* Deal with off-by-one related to the fencepost problem */
2804 program->work_register_count = ctx->work_registers + 1;
2805 program->uniform_cutoff = ctx->uniform_cutoff;
2806
2807 program->blend_patch_offset = ctx->blend_constant_offset;
2808 program->tls_size = ctx->tls_size;
2809
2810 if (midgard_debug & MIDGARD_DBG_SHADERS)
2811 disassemble_midgard(stdout, program->compiled.data, program->compiled.size, gpu_id, ctx->stage);
2812
2813 if (midgard_debug & MIDGARD_DBG_SHADERDB || shaderdb) {
2814 unsigned nr_bundles = 0, nr_ins = 0;
2815
2816 /* Count instructions and bundles */
2817
2818 mir_foreach_block(ctx, _block) {
2819 midgard_block *block = (midgard_block *) _block;
2820 nr_bundles += util_dynarray_num_elements(
2821 &block->bundles, midgard_bundle);
2822
2823 mir_foreach_bundle_in_block(block, bun)
2824 nr_ins += bun->instruction_count;
2825 }
2826
2827 /* Calculate thread count. There are certain cutoffs by
2828 * register count for thread count */
2829
2830 unsigned nr_registers = program->work_register_count;
2831
2832 unsigned nr_threads =
2833 (nr_registers <= 4) ? 4 :
2834 (nr_registers <= 8) ? 2 :
2835 1;
2836
2837 /* Dump stats */
2838
2839 fprintf(stderr, "shader%d - %s shader: "
2840 "%u inst, %u bundles, %u quadwords, "
2841 "%u registers, %u threads, %u loops, "
2842 "%u:%u spills:fills\n",
2843 SHADER_DB_COUNT++,
2844 gl_shader_stage_name(ctx->stage),
2845 nr_ins, nr_bundles, ctx->quadword_count,
2846 nr_registers, nr_threads,
2847 ctx->loop_count,
2848 ctx->spills, ctx->fills);
2849 }
2850
2851 ralloc_free(ctx);
2852
2853 return 0;
2854 }