xref: /aosp_15_r20/external/mesa3d/src/panfrost/compiler/valhall/va_pack.c (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /*
2  * Copyright (C) 2021 Collabora, Ltd.
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 "bi_builder.h"
25 #include "va_compiler.h"
26 #include "valhall.h"
27 #include "valhall_enums.h"
28 
29 /* This file contains the final passes of the compiler. Running after
30  * scheduling and RA, the IR is now finalized, so we need to emit it to actual
31  * bits on the wire (as well as fixup branches)
32  */
33 
34 /*
35  * Unreachable for encoding failures, when hitting an invalid instruction.
36  * Prints the (first) failing instruction to aid debugging.
37  */
38 NORETURN static void PRINTFLIKE(2, 3)
invalid_instruction(const bi_instr * I,const char * cause,...)39    invalid_instruction(const bi_instr *I, const char *cause, ...)
40 {
41    fputs("\nInvalid ", stderr);
42 
43    va_list ap;
44    va_start(ap, cause);
45    vfprintf(stderr, cause, ap);
46    va_end(ap);
47 
48    fputs(":\n\t", stderr);
49    bi_print_instr(I, stderr);
50    fprintf(stderr, "\n");
51 
52    unreachable("Invalid instruction");
53 }
54 
55 /*
56  * Like assert, but prints the instruction if the assertion fails to aid
57  * debugging invalid inputs to the packing module.
58  */
59 #define pack_assert(I, cond)                                                   \
60    if (!(cond))                                                                \
61       invalid_instruction(I, "invariant " #cond);
62 
63 /*
64  * Validate that two adjacent 32-bit sources form an aligned 64-bit register
65  * pair. This is a compiler invariant, required on Valhall but not on Bifrost.
66  */
67 static void
va_validate_register_pair(const bi_instr * I,unsigned s)68 va_validate_register_pair(const bi_instr *I, unsigned s)
69 {
70    ASSERTED bi_index lo = I->src[s], hi = I->src[s + 1];
71 
72    pack_assert(I, lo.type == hi.type);
73 
74    if (lo.type == BI_INDEX_REGISTER) {
75       pack_assert(I, hi.value & 1);
76       pack_assert(I, hi.value == lo.value + 1);
77    } else if (lo.type == BI_INDEX_FAU && lo.value & BIR_FAU_IMMEDIATE) {
78       /* Small constants are zero extended, so the top word encode zero */
79       pack_assert(I, hi.value == (BIR_FAU_IMMEDIATE | 0));
80    } else {
81       pack_assert(I, hi.offset & 1);
82       pack_assert(I, hi.offset == lo.offset + 1);
83    }
84 }
85 
86 static unsigned
va_pack_reg(const bi_instr * I,bi_index idx)87 va_pack_reg(const bi_instr *I, bi_index idx)
88 {
89    pack_assert(I, idx.type == BI_INDEX_REGISTER);
90    pack_assert(I, idx.value < 64);
91 
92    return idx.value;
93 }
94 
95 static unsigned
va_pack_fau_special(const bi_instr * I,enum bir_fau fau)96 va_pack_fau_special(const bi_instr *I, enum bir_fau fau)
97 {
98    switch (fau) {
99    case BIR_FAU_ATEST_PARAM:
100       return VA_FAU_SPECIAL_PAGE_0_ATEST_DATUM;
101    case BIR_FAU_TLS_PTR:
102       return VA_FAU_SPECIAL_PAGE_1_THREAD_LOCAL_POINTER;
103    case BIR_FAU_WLS_PTR:
104       return VA_FAU_SPECIAL_PAGE_1_WORKGROUP_LOCAL_POINTER;
105    case BIR_FAU_LANE_ID:
106       return VA_FAU_SPECIAL_PAGE_3_LANE_ID;
107    case BIR_FAU_PROGRAM_COUNTER:
108       return VA_FAU_SPECIAL_PAGE_3_PROGRAM_COUNTER;
109    case BIR_FAU_SAMPLE_POS_ARRAY:
110       return VA_FAU_SPECIAL_PAGE_0_SAMPLE;
111 
112    case BIR_FAU_BLEND_0 ...(BIR_FAU_BLEND_0 + 7):
113       return VA_FAU_SPECIAL_PAGE_0_BLEND_DESCRIPTOR_0 + (fau - BIR_FAU_BLEND_0);
114 
115    default:
116       invalid_instruction(I, "FAU");
117    }
118 }
119 
120 /*
121  * Encode a 64-bit FAU source. The offset is ignored, so this function can be
122  * used to encode a 32-bit FAU source by or'ing in the appropriate offset.
123  */
124 static unsigned
va_pack_fau_64(const bi_instr * I,bi_index idx)125 va_pack_fau_64(const bi_instr *I, bi_index idx)
126 {
127    pack_assert(I, idx.type == BI_INDEX_FAU);
128 
129    unsigned val = (idx.value & BITFIELD_MASK(5));
130 
131    if (idx.value & BIR_FAU_IMMEDIATE)
132       return (0x3 << 6) | (val << 1);
133    else if (idx.value & BIR_FAU_UNIFORM)
134       return (0x2 << 6) | (val << 1);
135    else
136       return (0x7 << 5) | (va_pack_fau_special(I, idx.value) << 1);
137 }
138 
139 static unsigned
va_pack_src(const bi_instr * I,unsigned s)140 va_pack_src(const bi_instr *I, unsigned s)
141 {
142    bi_index idx = I->src[s];
143 
144    if (idx.type == BI_INDEX_REGISTER) {
145       unsigned value = va_pack_reg(I, idx);
146       if (idx.discard)
147          value |= (1 << 6);
148       return value;
149    } else if (idx.type == BI_INDEX_FAU) {
150       pack_assert(I, idx.offset <= 1);
151       return va_pack_fau_64(I, idx) | idx.offset;
152    }
153 
154    invalid_instruction(I, "type of source %u", s);
155 }
156 
157 static unsigned
va_pack_wrmask(const bi_instr * I)158 va_pack_wrmask(const bi_instr *I)
159 {
160    switch (I->dest[0].swizzle) {
161    case BI_SWIZZLE_H00:
162       return 0x1;
163    case BI_SWIZZLE_H11:
164       return 0x2;
165    case BI_SWIZZLE_H01:
166       return 0x3;
167    default:
168       invalid_instruction(I, "write mask");
169    }
170 }
171 
172 static enum va_atomic_operation
va_pack_atom_opc(const bi_instr * I)173 va_pack_atom_opc(const bi_instr *I)
174 {
175    switch (I->atom_opc) {
176    case BI_ATOM_OPC_AADD:
177       return VA_ATOMIC_OPERATION_AADD;
178    case BI_ATOM_OPC_ASMIN:
179       return VA_ATOMIC_OPERATION_ASMIN;
180    case BI_ATOM_OPC_ASMAX:
181       return VA_ATOMIC_OPERATION_ASMAX;
182    case BI_ATOM_OPC_AUMIN:
183       return VA_ATOMIC_OPERATION_AUMIN;
184    case BI_ATOM_OPC_AUMAX:
185       return VA_ATOMIC_OPERATION_AUMAX;
186    case BI_ATOM_OPC_AAND:
187       return VA_ATOMIC_OPERATION_AAND;
188    case BI_ATOM_OPC_AOR:
189       return VA_ATOMIC_OPERATION_AOR;
190    case BI_ATOM_OPC_AXOR:
191       return VA_ATOMIC_OPERATION_AXOR;
192    case BI_ATOM_OPC_ACMPXCHG:
193    case BI_ATOM_OPC_AXCHG:
194       return VA_ATOMIC_OPERATION_AXCHG;
195    default:
196       invalid_instruction(I, "atomic opcode");
197    }
198 }
199 
200 static enum va_atomic_operation_with_1
va_pack_atom_opc_1(const bi_instr * I)201 va_pack_atom_opc_1(const bi_instr *I)
202 {
203    switch (I->atom_opc) {
204    case BI_ATOM_OPC_AINC:
205       return VA_ATOMIC_OPERATION_WITH_1_AINC;
206    case BI_ATOM_OPC_ADEC:
207       return VA_ATOMIC_OPERATION_WITH_1_ADEC;
208    case BI_ATOM_OPC_AUMAX1:
209       return VA_ATOMIC_OPERATION_WITH_1_AUMAX1;
210    case BI_ATOM_OPC_ASMAX1:
211       return VA_ATOMIC_OPERATION_WITH_1_ASMAX1;
212    case BI_ATOM_OPC_AOR1:
213       return VA_ATOMIC_OPERATION_WITH_1_AOR1;
214    default:
215       invalid_instruction(I, "atomic opcode with implicit 1");
216    }
217 }
218 
219 static unsigned
va_pack_dest(const bi_instr * I)220 va_pack_dest(const bi_instr *I)
221 {
222    assert(I->nr_dests);
223    return va_pack_reg(I, I->dest[0]) | (va_pack_wrmask(I) << 6);
224 }
225 
226 static enum va_widen
va_pack_widen_f32(const bi_instr * I,enum bi_swizzle swz)227 va_pack_widen_f32(const bi_instr *I, enum bi_swizzle swz)
228 {
229    switch (swz) {
230    case BI_SWIZZLE_H01:
231       return VA_WIDEN_NONE;
232    case BI_SWIZZLE_H00:
233       return VA_WIDEN_H0;
234    case BI_SWIZZLE_H11:
235       return VA_WIDEN_H1;
236    default:
237       invalid_instruction(I, "widen");
238    }
239 }
240 
241 static enum va_swizzles_16_bit
va_pack_swizzle_f16(const bi_instr * I,enum bi_swizzle swz)242 va_pack_swizzle_f16(const bi_instr *I, enum bi_swizzle swz)
243 {
244    switch (swz) {
245    case BI_SWIZZLE_H00:
246       return VA_SWIZZLES_16_BIT_H00;
247    case BI_SWIZZLE_H10:
248       return VA_SWIZZLES_16_BIT_H10;
249    case BI_SWIZZLE_H01:
250       return VA_SWIZZLES_16_BIT_H01;
251    case BI_SWIZZLE_H11:
252       return VA_SWIZZLES_16_BIT_H11;
253    default:
254       invalid_instruction(I, "16-bit swizzle");
255    }
256 }
257 
258 static unsigned
va_pack_widen(const bi_instr * I,enum bi_swizzle swz,enum va_size size)259 va_pack_widen(const bi_instr *I, enum bi_swizzle swz, enum va_size size)
260 {
261    if (size == VA_SIZE_8) {
262       switch (swz) {
263       case BI_SWIZZLE_H01:
264          return VA_SWIZZLES_8_BIT_B0123;
265       case BI_SWIZZLE_H00:
266          return VA_SWIZZLES_8_BIT_B0101;
267       case BI_SWIZZLE_H11:
268          return VA_SWIZZLES_8_BIT_B2323;
269       case BI_SWIZZLE_B0000:
270          return VA_SWIZZLES_8_BIT_B0000;
271       case BI_SWIZZLE_B1111:
272          return VA_SWIZZLES_8_BIT_B1111;
273       case BI_SWIZZLE_B2222:
274          return VA_SWIZZLES_8_BIT_B2222;
275       case BI_SWIZZLE_B3333:
276          return VA_SWIZZLES_8_BIT_B3333;
277       default:
278          invalid_instruction(I, "8-bit widen");
279       }
280    } else if (size == VA_SIZE_16) {
281       switch (swz) {
282       case BI_SWIZZLE_H00:
283          return VA_SWIZZLES_16_BIT_H00;
284       case BI_SWIZZLE_H10:
285          return VA_SWIZZLES_16_BIT_H10;
286       case BI_SWIZZLE_H01:
287          return VA_SWIZZLES_16_BIT_H01;
288       case BI_SWIZZLE_H11:
289          return VA_SWIZZLES_16_BIT_H11;
290       case BI_SWIZZLE_B0000:
291          return VA_SWIZZLES_16_BIT_B00;
292       case BI_SWIZZLE_B1111:
293          return VA_SWIZZLES_16_BIT_B11;
294       case BI_SWIZZLE_B2222:
295          return VA_SWIZZLES_16_BIT_B22;
296       case BI_SWIZZLE_B3333:
297          return VA_SWIZZLES_16_BIT_B33;
298       default:
299          invalid_instruction(I, "16-bit widen");
300       }
301    } else if (size == VA_SIZE_32) {
302       switch (swz) {
303       case BI_SWIZZLE_H01:
304          return VA_SWIZZLES_32_BIT_NONE;
305       case BI_SWIZZLE_H00:
306          return VA_SWIZZLES_32_BIT_H0;
307       case BI_SWIZZLE_H11:
308          return VA_SWIZZLES_32_BIT_H1;
309       case BI_SWIZZLE_B0000:
310          return VA_SWIZZLES_32_BIT_B0;
311       case BI_SWIZZLE_B1111:
312          return VA_SWIZZLES_32_BIT_B1;
313       case BI_SWIZZLE_B2222:
314          return VA_SWIZZLES_32_BIT_B2;
315       case BI_SWIZZLE_B3333:
316          return VA_SWIZZLES_32_BIT_B3;
317       default:
318          invalid_instruction(I, "32-bit widen");
319       }
320    } else {
321       invalid_instruction(I, "type size for widen");
322    }
323 }
324 
325 static enum va_half_swizzles_8_bit
va_pack_halfswizzle(const bi_instr * I,enum bi_swizzle swz)326 va_pack_halfswizzle(const bi_instr *I, enum bi_swizzle swz)
327 {
328    switch (swz) {
329    case BI_SWIZZLE_B0000:
330       return VA_HALF_SWIZZLES_8_BIT_B00;
331    case BI_SWIZZLE_B1111:
332       return VA_HALF_SWIZZLES_8_BIT_B11;
333    case BI_SWIZZLE_B2222:
334       return VA_HALF_SWIZZLES_8_BIT_B22;
335    case BI_SWIZZLE_B3333:
336       return VA_HALF_SWIZZLES_8_BIT_B33;
337    case BI_SWIZZLE_B0011:
338       return VA_HALF_SWIZZLES_8_BIT_B01;
339    case BI_SWIZZLE_B2233:
340       return VA_HALF_SWIZZLES_8_BIT_B23;
341    case BI_SWIZZLE_B0022:
342       return VA_HALF_SWIZZLES_8_BIT_B02;
343    default:
344       invalid_instruction(I, "v2u8 swizzle");
345    }
346 }
347 
348 static enum va_lanes_8_bit
va_pack_shift_lanes(const bi_instr * I,enum bi_swizzle swz)349 va_pack_shift_lanes(const bi_instr *I, enum bi_swizzle swz)
350 {
351    switch (swz) {
352    case BI_SWIZZLE_H01:
353       return VA_LANES_8_BIT_B02;
354    case BI_SWIZZLE_B0000:
355       return VA_LANES_8_BIT_B00;
356    case BI_SWIZZLE_B1111:
357       return VA_LANES_8_BIT_B11;
358    case BI_SWIZZLE_B2222:
359       return VA_LANES_8_BIT_B22;
360    case BI_SWIZZLE_B3333:
361       return VA_LANES_8_BIT_B33;
362    default:
363       invalid_instruction(I, "lane shift");
364    }
365 }
366 
367 static enum va_combine
va_pack_combine(const bi_instr * I,enum bi_swizzle swz)368 va_pack_combine(const bi_instr *I, enum bi_swizzle swz)
369 {
370    switch (swz) {
371    case BI_SWIZZLE_H01:
372       return VA_COMBINE_NONE;
373    case BI_SWIZZLE_H00:
374       return VA_COMBINE_H0;
375    case BI_SWIZZLE_H11:
376       return VA_COMBINE_H1;
377    default:
378       invalid_instruction(I, "branch lane");
379    }
380 }
381 
382 static enum va_source_format
va_pack_source_format(const bi_instr * I)383 va_pack_source_format(const bi_instr *I)
384 {
385    switch (I->source_format) {
386    case BI_SOURCE_FORMAT_FLAT32:
387       return VA_SOURCE_FORMAT_SRC_FLAT32;
388    case BI_SOURCE_FORMAT_FLAT16:
389       return VA_SOURCE_FORMAT_SRC_FLAT16;
390    case BI_SOURCE_FORMAT_F32:
391       return VA_SOURCE_FORMAT_SRC_F32;
392    case BI_SOURCE_FORMAT_F16:
393       return VA_SOURCE_FORMAT_SRC_F16;
394    }
395 
396    invalid_instruction(I, "source format");
397 }
398 
399 static uint64_t
va_pack_rhadd(const bi_instr * I)400 va_pack_rhadd(const bi_instr *I)
401 {
402    switch (I->round) {
403    case BI_ROUND_RTN:
404       return 0; /* hadd */
405    case BI_ROUND_RTP:
406       return BITFIELD_BIT(30); /* rhadd */
407    default:
408       unreachable("Invalid round for HADD");
409    }
410 }
411 
412 static uint64_t
va_pack_alu(const bi_instr * I)413 va_pack_alu(const bi_instr *I)
414 {
415    struct va_opcode_info info = valhall_opcodes[I->op];
416    uint64_t hex = 0;
417 
418    switch (I->op) {
419    /* Add FREXP flags */
420    case BI_OPCODE_FREXPE_F32:
421    case BI_OPCODE_FREXPE_V2F16:
422    case BI_OPCODE_FREXPM_F32:
423    case BI_OPCODE_FREXPM_V2F16:
424       if (I->sqrt)
425          hex |= 1ull << 24;
426       if (I->log)
427          hex |= 1ull << 25;
428       break;
429 
430    /* Add mux type */
431    case BI_OPCODE_MUX_I32:
432    case BI_OPCODE_MUX_V2I16:
433    case BI_OPCODE_MUX_V4I8:
434       hex |= (uint64_t)I->mux << 32;
435       break;
436 
437    /* Add .eq flag */
438    case BI_OPCODE_BRANCHZ_I16:
439    case BI_OPCODE_BRANCHZI:
440       pack_assert(I, I->cmpf == BI_CMPF_EQ || I->cmpf == BI_CMPF_NE);
441 
442       if (I->cmpf == BI_CMPF_EQ)
443          hex |= (1ull << 36);
444 
445       if (I->op == BI_OPCODE_BRANCHZI)
446          hex |= (0x1ull << 40); /* Absolute */
447       else
448          hex |= ((uint64_t)I->branch_offset & BITFIELD_MASK(27)) << 8;
449 
450       break;
451 
452    /* Add arithmetic flag */
453    case BI_OPCODE_RSHIFT_AND_I32:
454    case BI_OPCODE_RSHIFT_AND_V2I16:
455    case BI_OPCODE_RSHIFT_AND_V4I8:
456    case BI_OPCODE_RSHIFT_OR_I32:
457    case BI_OPCODE_RSHIFT_OR_V2I16:
458    case BI_OPCODE_RSHIFT_OR_V4I8:
459    case BI_OPCODE_RSHIFT_XOR_I32:
460    case BI_OPCODE_RSHIFT_XOR_V2I16:
461    case BI_OPCODE_RSHIFT_XOR_V4I8:
462       hex |= (uint64_t)I->arithmetic << 34;
463       break;
464 
465    case BI_OPCODE_LEA_BUF_IMM:
466       /* Buffer table index */
467       hex |= 0xD << 8;
468       break;
469 
470    case BI_OPCODE_LEA_ATTR_IMM:
471       hex |= ((uint64_t)I->table) << 16;
472       hex |= ((uint64_t)I->attribute_index) << 20;
473       break;
474 
475    case BI_OPCODE_IADD_IMM_I32:
476    case BI_OPCODE_IADD_IMM_V2I16:
477    case BI_OPCODE_IADD_IMM_V4I8:
478    case BI_OPCODE_FADD_IMM_F32:
479    case BI_OPCODE_FADD_IMM_V2F16:
480       hex |= ((uint64_t)I->index) << 8;
481       break;
482 
483    case BI_OPCODE_CLPER_I32:
484       hex |= ((uint64_t)I->inactive_result) << 22;
485       hex |= ((uint64_t)I->lane_op) << 32;
486       hex |= ((uint64_t)I->subgroup) << 36;
487       break;
488 
489    case BI_OPCODE_LD_VAR:
490    case BI_OPCODE_LD_VAR_FLAT:
491    case BI_OPCODE_LD_VAR_IMM:
492    case BI_OPCODE_LD_VAR_FLAT_IMM:
493    case BI_OPCODE_LD_VAR_BUF_F16:
494    case BI_OPCODE_LD_VAR_BUF_F32:
495    case BI_OPCODE_LD_VAR_BUF_IMM_F16:
496    case BI_OPCODE_LD_VAR_BUF_IMM_F32:
497    case BI_OPCODE_LD_VAR_SPECIAL:
498       if (I->op == BI_OPCODE_LD_VAR_SPECIAL)
499          hex |= ((uint64_t)I->varying_name) << 12; /* instead of index */
500       else if (I->op == BI_OPCODE_LD_VAR_BUF_IMM_F16 ||
501                I->op == BI_OPCODE_LD_VAR_BUF_IMM_F32) {
502          hex |= ((uint64_t)I->index) << 16;
503       } else if (I->op == BI_OPCODE_LD_VAR_IMM ||
504                  I->op == BI_OPCODE_LD_VAR_FLAT_IMM) {
505          hex |= ((uint64_t)I->table) << 8;
506          hex |= ((uint64_t)I->index) << 12;
507       }
508 
509       hex |= ((uint64_t)va_pack_source_format(I)) << 24;
510       hex |= ((uint64_t)I->update) << 36;
511       hex |= ((uint64_t)I->sample) << 38;
512       break;
513 
514    case BI_OPCODE_LD_ATTR_IMM:
515       hex |= ((uint64_t)I->table) << 16;
516       hex |= ((uint64_t)I->attribute_index) << 20;
517       break;
518 
519    case BI_OPCODE_LD_TEX_IMM:
520    case BI_OPCODE_LEA_TEX_IMM:
521       hex |= ((uint64_t)I->table) << 16;
522       hex |= ((uint64_t)I->texture_index) << 20;
523       break;
524 
525    case BI_OPCODE_ZS_EMIT:
526       if (I->stencil)
527          hex |= (1 << 24);
528       if (I->z)
529          hex |= (1 << 25);
530       break;
531 
532    default:
533       break;
534    }
535 
536    /* FMA_RSCALE.f32 special modes treated as extra opcodes */
537    if (I->op == BI_OPCODE_FMA_RSCALE_F32) {
538       pack_assert(I, I->special < 4);
539       hex |= ((uint64_t)I->special) << 48;
540    }
541 
542    /* Add the normal destination or a placeholder.  Staging destinations are
543     * added elsewhere, as they require special handling for control fields.
544     */
545    if (info.has_dest && info.nr_staging_dests == 0) {
546       hex |= (uint64_t)va_pack_dest(I) << 40;
547    } else if (info.nr_staging_dests == 0 && info.nr_staging_srcs == 0) {
548       pack_assert(I, I->nr_dests == 0);
549       hex |= 0xC0ull << 40; /* Placeholder */
550    }
551 
552    bool swap12 = va_swap_12(I->op);
553 
554    /* First src is staging if we read, skip it when packing sources */
555    unsigned src_offset = bi_opcode_props[I->op].sr_read ? 1 : 0;
556 
557    for (unsigned i = 0; i < info.nr_srcs; ++i) {
558       unsigned logical_i = (swap12 && i == 1) ? 2 : (swap12 && i == 2) ? 1 : i;
559 
560       struct va_src_info src_info = info.srcs[i];
561       enum va_size size = src_info.size;
562 
563       bi_index src = I->src[logical_i + src_offset];
564       hex |= (uint64_t)va_pack_src(I, logical_i + src_offset) << (8 * i);
565 
566       if (src_info.notted) {
567          if (src.neg)
568             hex |= (1ull << 35);
569       } else if (src_info.absneg) {
570          unsigned neg_offs = 32 + 2 + ((2 - i) * 2);
571          unsigned abs_offs = 33 + 2 + ((2 - i) * 2);
572 
573          if (src.neg)
574             hex |= 1ull << neg_offs;
575          if (src.abs)
576             hex |= 1ull << abs_offs;
577       } else {
578          if (src.neg)
579             invalid_instruction(I, "negate");
580          if (src.abs)
581             invalid_instruction(I, "absolute value");
582       }
583 
584       if (src_info.swizzle) {
585          unsigned offs = 24 + ((2 - i) * 2);
586          unsigned S = src.swizzle;
587          pack_assert(I, size == VA_SIZE_16 || size == VA_SIZE_32);
588 
589          uint64_t v = (size == VA_SIZE_32 ? va_pack_widen_f32(I, S)
590                                           : va_pack_swizzle_f16(I, S));
591          hex |= v << offs;
592       } else if (src_info.widen) {
593          unsigned offs = (i == 1) ? 26 : 36;
594          hex |= (uint64_t)va_pack_widen(I, src.swizzle, src_info.size) << offs;
595       } else if (src_info.lane) {
596          unsigned offs =
597             (I->op == BI_OPCODE_MKVEC_V2I8) ? ((i == 0) ? 38 : 36) : 28;
598 
599          if (src_info.size == VA_SIZE_16) {
600             hex |= (src.swizzle == BI_SWIZZLE_H11 ? 1 : 0) << offs;
601          } else if (I->op == BI_OPCODE_BRANCHZ_I16) {
602             hex |= ((uint64_t)va_pack_combine(I, src.swizzle) << 37);
603          } else {
604             pack_assert(I, src_info.size == VA_SIZE_8);
605             unsigned comp = src.swizzle - BI_SWIZZLE_B0000;
606             pack_assert(I, comp < 4);
607             hex |= (uint64_t)comp << offs;
608          }
609       } else if (src_info.lanes) {
610          pack_assert(I, src_info.size == VA_SIZE_8);
611          pack_assert(I, i == 1);
612          hex |= (uint64_t)va_pack_shift_lanes(I, src.swizzle) << 26;
613       } else if (src_info.combine) {
614          /* Treat as swizzle, subgroup ops not yet supported */
615          pack_assert(I, src_info.size == VA_SIZE_32);
616          pack_assert(I, i == 0);
617          hex |= (uint64_t)va_pack_widen_f32(I, src.swizzle) << 37;
618       } else if (src_info.halfswizzle) {
619          pack_assert(I, src_info.size == VA_SIZE_8);
620          pack_assert(I, i == 0);
621          hex |= (uint64_t)va_pack_halfswizzle(I, src.swizzle) << 36;
622       } else if (src.swizzle != BI_SWIZZLE_H01) {
623          invalid_instruction(I, "swizzle");
624       }
625    }
626 
627    if (info.saturate)
628       hex |= (uint64_t)I->saturate << 30;
629    if (info.rhadd)
630       hex |= va_pack_rhadd(I);
631    if (info.clamp)
632       hex |= (uint64_t)I->clamp << 32;
633    if (info.round_mode)
634       hex |= (uint64_t)I->round << 30;
635    if (info.condition)
636       hex |= (uint64_t)I->cmpf << 32;
637    if (info.result_type)
638       hex |= (uint64_t)I->result_type << 30;
639 
640    return hex;
641 }
642 
643 static uint64_t
va_pack_byte_offset(const bi_instr * I)644 va_pack_byte_offset(const bi_instr *I)
645 {
646    int16_t offset = I->byte_offset;
647    if (offset != I->byte_offset)
648       invalid_instruction(I, "byte offset");
649 
650    uint16_t offset_as_u16 = offset;
651    return ((uint64_t)offset_as_u16) << 8;
652 }
653 
654 static uint64_t
va_pack_byte_offset_8(const bi_instr * I)655 va_pack_byte_offset_8(const bi_instr *I)
656 {
657    uint8_t offset = I->byte_offset;
658    if (offset != I->byte_offset)
659       invalid_instruction(I, "byte offset");
660 
661    return ((uint64_t)offset) << 8;
662 }
663 
664 static uint64_t
va_pack_load(const bi_instr * I,bool buffer_descriptor)665 va_pack_load(const bi_instr *I, bool buffer_descriptor)
666 {
667    const uint8_t load_lane_identity[8] = {
668       VA_LOAD_LANE_8_BIT_B0,        VA_LOAD_LANE_16_BIT_H0,
669       VA_LOAD_LANE_24_BIT_IDENTITY, VA_LOAD_LANE_32_BIT_W0,
670       VA_LOAD_LANE_48_BIT_IDENTITY, VA_LOAD_LANE_64_BIT_IDENTITY,
671       VA_LOAD_LANE_96_BIT_IDENTITY, VA_LOAD_LANE_128_BIT_IDENTITY,
672    };
673 
674    unsigned memory_size = (valhall_opcodes[I->op].exact >> 27) & 0x7;
675    uint64_t hex = (uint64_t)load_lane_identity[memory_size] << 36;
676 
677    // unsigned
678    hex |= (1ull << 39);
679 
680    if (!buffer_descriptor)
681       hex |= va_pack_byte_offset(I);
682 
683    hex |= (uint64_t)va_pack_src(I, 0) << 0;
684 
685    if (buffer_descriptor)
686       hex |= (uint64_t)va_pack_src(I, 1) << 8;
687 
688    return hex;
689 }
690 
691 static uint64_t
va_pack_memory_access(const bi_instr * I)692 va_pack_memory_access(const bi_instr *I)
693 {
694    switch (I->seg) {
695    case BI_SEG_TL:
696       return VA_MEMORY_ACCESS_FORCE;
697    case BI_SEG_POS:
698       return VA_MEMORY_ACCESS_ISTREAM;
699    case BI_SEG_VARY:
700       return VA_MEMORY_ACCESS_ESTREAM;
701    default:
702       return VA_MEMORY_ACCESS_NONE;
703    }
704 }
705 
706 static uint64_t
va_pack_store(const bi_instr * I)707 va_pack_store(const bi_instr *I)
708 {
709    uint64_t hex = va_pack_memory_access(I) << 24;
710 
711    va_validate_register_pair(I, 1);
712    hex |= (uint64_t)va_pack_src(I, 1) << 0;
713 
714    hex |= va_pack_byte_offset(I);
715 
716    return hex;
717 }
718 
719 static enum va_lod_mode
va_pack_lod_mode(const bi_instr * I)720 va_pack_lod_mode(const bi_instr *I)
721 {
722    switch (I->va_lod_mode) {
723    case BI_VA_LOD_MODE_ZERO_LOD:
724       return VA_LOD_MODE_ZERO;
725    case BI_VA_LOD_MODE_COMPUTED_LOD:
726       return VA_LOD_MODE_COMPUTED;
727    case BI_VA_LOD_MODE_EXPLICIT:
728       return VA_LOD_MODE_EXPLICIT;
729    case BI_VA_LOD_MODE_COMPUTED_BIAS:
730       return VA_LOD_MODE_COMPUTED_BIAS;
731    case BI_VA_LOD_MODE_GRDESC:
732       return VA_LOD_MODE_GRDESC;
733    }
734 
735    invalid_instruction(I, "LOD mode");
736 }
737 
738 static enum va_register_type
va_pack_register_type(const bi_instr * I)739 va_pack_register_type(const bi_instr *I)
740 {
741    switch (I->register_format) {
742    case BI_REGISTER_FORMAT_F16:
743    case BI_REGISTER_FORMAT_F32:
744       return VA_REGISTER_TYPE_F;
745 
746    case BI_REGISTER_FORMAT_U16:
747    case BI_REGISTER_FORMAT_U32:
748       return VA_REGISTER_TYPE_U;
749 
750    case BI_REGISTER_FORMAT_S16:
751    case BI_REGISTER_FORMAT_S32:
752       return VA_REGISTER_TYPE_S;
753 
754    default:
755       invalid_instruction(I, "register type");
756    }
757 }
758 
759 static enum va_register_format
va_pack_register_format(const bi_instr * I)760 va_pack_register_format(const bi_instr *I)
761 {
762    switch (I->register_format) {
763    case BI_REGISTER_FORMAT_AUTO:
764       return VA_REGISTER_FORMAT_AUTO;
765    case BI_REGISTER_FORMAT_F32:
766       return VA_REGISTER_FORMAT_F32;
767    case BI_REGISTER_FORMAT_F16:
768       return VA_REGISTER_FORMAT_F16;
769    case BI_REGISTER_FORMAT_S32:
770       return VA_REGISTER_FORMAT_S32;
771    case BI_REGISTER_FORMAT_S16:
772       return VA_REGISTER_FORMAT_S16;
773    case BI_REGISTER_FORMAT_U32:
774       return VA_REGISTER_FORMAT_U32;
775    case BI_REGISTER_FORMAT_U16:
776       return VA_REGISTER_FORMAT_U16;
777    default:
778       invalid_instruction(I, "register format");
779    }
780 }
781 
782 uint64_t
va_pack_instr(const bi_instr * I)783 va_pack_instr(const bi_instr *I)
784 {
785    struct va_opcode_info info = valhall_opcodes[I->op];
786 
787    uint64_t hex = info.exact | (((uint64_t)I->flow) << 59);
788    hex |= ((uint64_t)va_select_fau_page(I)) << 57;
789 
790    if (info.slot)
791       hex |= ((uint64_t)I->slot << 30);
792 
793    if (info.sr_count) {
794       bool read = bi_opcode_props[I->op].sr_read;
795       bi_index sr = read ? I->src[0] : I->dest[0];
796 
797       unsigned count =
798          read ? bi_count_read_registers(I, 0) : bi_count_write_registers(I, 0);
799 
800       hex |= ((uint64_t)count << 33);
801       hex |= (uint64_t)va_pack_reg(I, sr) << 40;
802       hex |= ((uint64_t)info.sr_control << 46);
803    }
804 
805    if (info.sr_write_count) {
806       hex |= ((uint64_t)bi_count_write_registers(I, 0) - 1) << 36;
807       hex |= ((uint64_t)va_pack_reg(I, I->dest[0])) << 16;
808    }
809 
810    if (info.vecsize)
811       hex |= ((uint64_t)I->vecsize << 28);
812 
813    if (info.register_format)
814       hex |= ((uint64_t)va_pack_register_format(I)) << 24;
815 
816    switch (I->op) {
817    case BI_OPCODE_LOAD_I8:
818    case BI_OPCODE_LOAD_I16:
819    case BI_OPCODE_LOAD_I24:
820    case BI_OPCODE_LOAD_I32:
821    case BI_OPCODE_LOAD_I48:
822    case BI_OPCODE_LOAD_I64:
823    case BI_OPCODE_LOAD_I96:
824    case BI_OPCODE_LOAD_I128:
825       hex |= va_pack_load(I, false);
826       break;
827 
828    case BI_OPCODE_LD_BUFFER_I8:
829    case BI_OPCODE_LD_BUFFER_I16:
830    case BI_OPCODE_LD_BUFFER_I24:
831    case BI_OPCODE_LD_BUFFER_I32:
832    case BI_OPCODE_LD_BUFFER_I48:
833    case BI_OPCODE_LD_BUFFER_I64:
834    case BI_OPCODE_LD_BUFFER_I96:
835    case BI_OPCODE_LD_BUFFER_I128:
836       hex |= va_pack_load(I, true);
837       break;
838 
839    case BI_OPCODE_STORE_I8:
840    case BI_OPCODE_STORE_I16:
841    case BI_OPCODE_STORE_I24:
842    case BI_OPCODE_STORE_I32:
843    case BI_OPCODE_STORE_I48:
844    case BI_OPCODE_STORE_I64:
845    case BI_OPCODE_STORE_I96:
846    case BI_OPCODE_STORE_I128:
847       hex |= va_pack_store(I);
848       break;
849 
850    case BI_OPCODE_ATOM1_RETURN_I32:
851       /* Permit omitting the destination for plain ATOM1 */
852       if (!bi_count_write_registers(I, 0)) {
853          hex |= (0x40ull << 40); // fake read
854       }
855 
856       /* 64-bit source */
857       va_validate_register_pair(I, 0);
858       hex |= (uint64_t)va_pack_src(I, 0) << 0;
859       hex |= va_pack_byte_offset_8(I);
860       hex |= ((uint64_t)va_pack_atom_opc_1(I)) << 22;
861       break;
862 
863    case BI_OPCODE_ATOM_I32:
864    case BI_OPCODE_ATOM_RETURN_I32:
865       /* 64-bit source */
866       va_validate_register_pair(I, 1);
867       hex |= (uint64_t)va_pack_src(I, 1) << 0;
868       hex |= va_pack_byte_offset_8(I);
869       hex |= ((uint64_t)va_pack_atom_opc(I)) << 22;
870 
871       if (I->op == BI_OPCODE_ATOM_RETURN_I32)
872          hex |= (0xc0ull << 40); // flags
873 
874       if (I->atom_opc == BI_ATOM_OPC_ACMPXCHG)
875          hex |= (1 << 26); /* .compare */
876 
877       break;
878 
879    case BI_OPCODE_ST_CVT:
880       /* Staging read */
881       hex |= va_pack_store(I);
882 
883       /* Conversion descriptor */
884       hex |= (uint64_t)va_pack_src(I, 3) << 16;
885       break;
886 
887    case BI_OPCODE_BLEND: {
888       /* Source 0 - Blend descriptor (64-bit) */
889       hex |= ((uint64_t)va_pack_src(I, 2)) << 0;
890       va_validate_register_pair(I, 2);
891 
892       /* Target */
893       if (I->branch_offset & 0x7)
894          invalid_instruction(I, "unaligned branch");
895       hex |= ((I->branch_offset >> 3) << 8);
896 
897       /* Source 2 - coverage mask */
898       hex |= ((uint64_t)va_pack_reg(I, I->src[1])) << 16;
899 
900       /* Vector size */
901       unsigned vecsize = 4;
902       hex |= ((uint64_t)(vecsize - 1) << 28);
903 
904       break;
905    }
906 
907    case BI_OPCODE_TEX_SINGLE:
908    case BI_OPCODE_TEX_FETCH:
909    case BI_OPCODE_TEX_GATHER: {
910       /* Image to read from */
911       hex |= ((uint64_t)va_pack_src(I, 1)) << 0;
912 
913       if (I->op == BI_OPCODE_TEX_FETCH && I->shadow)
914          invalid_instruction(I, "TEX_FETCH does not support .shadow");
915 
916       if (I->wide_indices)
917          hex |= (1ull << 8);
918       if (I->array_enable)
919          hex |= (1ull << 10);
920       if (I->texel_offset)
921          hex |= (1ull << 11);
922       if (I->shadow)
923          hex |= (1ull << 12);
924       if (I->skip)
925          hex |= (1ull << 39);
926       if (!bi_is_regfmt_16(I->register_format))
927          hex |= (1ull << 46);
928 
929       if (I->op == BI_OPCODE_TEX_SINGLE)
930          hex |= ((uint64_t)va_pack_lod_mode(I)) << 13;
931 
932       if (I->op == BI_OPCODE_TEX_GATHER) {
933          if (I->integer_coordinates)
934             hex |= (1 << 13);
935          hex |= ((uint64_t)I->fetch_component) << 14;
936       }
937 
938       hex |= (I->write_mask << 22);
939       hex |= ((uint64_t)va_pack_register_type(I)) << 26;
940       hex |= ((uint64_t)I->dimension) << 28;
941 
942       break;
943    }
944 
945    default:
946       if (!info.exact && I->op != BI_OPCODE_NOP)
947          invalid_instruction(I, "opcode");
948 
949       hex |= va_pack_alu(I);
950       break;
951    }
952 
953    return hex;
954 }
955 
956 static unsigned
va_instructions_in_block(bi_block * block)957 va_instructions_in_block(bi_block *block)
958 {
959    unsigned offset = 0;
960 
961    bi_foreach_instr_in_block(block, _) {
962       offset++;
963    }
964 
965    return offset;
966 }
967 
968 /* Calculate branch_offset from a branch_target for a direct relative branch */
969 
970 static void
va_lower_branch_target(bi_context * ctx,bi_block * start,bi_instr * I)971 va_lower_branch_target(bi_context *ctx, bi_block *start, bi_instr *I)
972 {
973    /* Precondition: unlowered relative branch */
974    bi_block *target = I->branch_target;
975    assert(target != NULL);
976 
977    /* Signed since we might jump backwards */
978    signed offset = 0;
979 
980    /* Determine if the target block is strictly greater in source order */
981    bool forwards = target->index > start->index;
982 
983    if (forwards) {
984       /* We have to jump through this block */
985       bi_foreach_instr_in_block_from(start, _, I) {
986          offset++;
987       }
988 
989       /* We then need to jump over every following block until the target */
990       bi_foreach_block_from(ctx, start, blk) {
991          /* End just before the target */
992          if (blk == target)
993             break;
994 
995          /* Count other blocks */
996          if (blk != start)
997             offset += va_instructions_in_block(blk);
998       }
999    } else {
1000       /* Jump through the beginning of this block */
1001       bi_foreach_instr_in_block_from_rev(start, ins, I) {
1002          if (ins != I)
1003             offset--;
1004       }
1005 
1006       /* Jump over preceding blocks up to and including the target to get to
1007        * the beginning of the target */
1008       bi_foreach_block_from_rev(ctx, start, blk) {
1009          if (blk == start)
1010             continue;
1011 
1012          offset -= va_instructions_in_block(blk);
1013 
1014          /* End just after the target */
1015          if (blk == target)
1016             break;
1017       }
1018    }
1019 
1020    /* Offset is relative to the next instruction, so bias */
1021    offset--;
1022 
1023    /* Update the instruction */
1024    I->branch_offset = offset;
1025 }
1026 
1027 /*
1028  * Late lowering to insert blend shader calls after BLEND instructions. Required
1029  * to support blend shaders, so this pass may be omitted if it is known that
1030  * blend shaders are never used.
1031  *
1032  * This lowering runs late because it introduces control flow changes without
1033  * modifying the control flow graph. It hardcodes registers, meaning running
1034  * after RA makes sense. Finally, it hardcodes a manually sized instruction
1035  * sequence, requiring it to run after scheduling.
1036  *
1037  * As it is Valhall specific, running it as a pre-pack lowering is sensible.
1038  */
1039 static void
va_lower_blend(bi_context * ctx)1040 va_lower_blend(bi_context *ctx)
1041 {
1042    /* Program counter for *next* instruction */
1043    bi_index pc = bi_fau(BIR_FAU_PROGRAM_COUNTER, false);
1044 
1045    bi_foreach_instr_global_safe(ctx, I) {
1046       if (I->op != BI_OPCODE_BLEND)
1047          continue;
1048 
1049       bi_builder b = bi_init_builder(ctx, bi_after_instr(I));
1050 
1051       unsigned prolog_length = 2 * 8;
1052 
1053       /* By ABI, r48 is the link register shared with blend shaders */
1054       assert(bi_is_equiv(I->dest[0], bi_register(48)));
1055 
1056       if (I->flow == VA_FLOW_END)
1057          bi_iadd_imm_i32_to(&b, I->dest[0], va_zero_lut(), 0);
1058       else
1059          bi_iadd_imm_i32_to(&b, I->dest[0], pc, prolog_length - 8);
1060 
1061       bi_branchzi(&b, va_zero_lut(), I->src[3], BI_CMPF_EQ);
1062 
1063       /* For fixed function: skip the prologue, or return */
1064       if (I->flow != VA_FLOW_END)
1065          I->branch_offset = prolog_length;
1066    }
1067 }
1068 
1069 void
bi_pack_valhall(bi_context * ctx,struct util_dynarray * emission)1070 bi_pack_valhall(bi_context *ctx, struct util_dynarray *emission)
1071 {
1072    unsigned orig_size = emission->size;
1073 
1074    va_validate(stderr, ctx);
1075 
1076    /* Late lowering */
1077    if (ctx->stage == MESA_SHADER_FRAGMENT && !ctx->inputs->is_blend)
1078       va_lower_blend(ctx);
1079 
1080    bi_foreach_block(ctx, block) {
1081       bi_foreach_instr_in_block(block, I) {
1082          if (I->op == BI_OPCODE_BRANCHZ_I16)
1083             va_lower_branch_target(ctx, block, I);
1084 
1085          uint64_t hex = va_pack_instr(I);
1086          util_dynarray_append(emission, uint64_t, hex);
1087       }
1088    }
1089 
1090    /* Pad with zeroes, but keep empty programs empty so they may be omitted
1091     * altogether. Failing to do this would result in a program containing only
1092     * zeroes, which is invalid and will raise an encoding fault.
1093     *
1094     * Pad an extra 16 byte (one instruction) to separate primary and secondary
1095     * shader disassembles. This is not strictly necessary, but it's a good
1096     * practice. 128 bytes is the optimal program alignment on Trym, so pad
1097     * secondary shaders up to 128 bytes. This may help the instruction cache.
1098     */
1099    if (orig_size != emission->size) {
1100       unsigned aligned = ALIGN_POT(emission->size + 16, 128);
1101       unsigned count = aligned - emission->size;
1102 
1103       memset(util_dynarray_grow(emission, uint8_t, count), 0, count);
1104    }
1105 }
1106