1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
|
/*
* Copyright (C) 2018-2019 Alyssa Rosenzweig <alyssa@rosenzweig.io>
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "compiler.h"
#include "midgard_ops.h"
/* Midgard IR only knows vector ALU types, but we sometimes need to actually
* use scalar ALU instructions, for functional or performance reasons. To do
* this, we just demote vector ALU payloads to scalar. */
static int
component_from_mask(unsigned mask)
{
for (int c = 0; c < 8; ++c) {
if (mask & (1 << c))
return c;
}
assert(0);
return 0;
}
static unsigned
vector_to_scalar_source(unsigned u, bool is_int, bool is_full,
unsigned component)
{
midgard_vector_alu_src v;
memcpy(&v, &u, sizeof(v));
/* TODO: Integers */
midgard_scalar_alu_src s = { 0 };
if (is_full) {
/* For a 32-bit op, just check the source half flag */
s.full = !v.half;
} else if (!v.half) {
/* For a 16-bit op that's not subdivided, never full */
s.full = false;
} else {
/* We can't do 8-bit scalar, abort! */
assert(0);
}
/* Component indexing takes size into account */
if (s.full)
s.component = component << 1;
else
s.component = component;
if (is_int) {
/* TODO */
} else {
s.abs = v.mod & MIDGARD_FLOAT_MOD_ABS;
s.negate = v.mod & MIDGARD_FLOAT_MOD_NEG;
}
unsigned o;
memcpy(&o, &s, sizeof(s));
return o & ((1 << 6) - 1);
}
static midgard_scalar_alu
vector_to_scalar_alu(midgard_vector_alu v, midgard_instruction *ins)
{
bool is_int = midgard_is_integer_op(v.op);
bool is_full = v.reg_mode == midgard_reg_mode_32;
bool is_inline_constant = ins->has_inline_constant;
unsigned comp = component_from_mask(ins->mask);
/* The output component is from the mask */
midgard_scalar_alu s = {
.op = v.op,
.src1 = vector_to_scalar_source(v.src1, is_int, is_full, ins->swizzle[0][comp]),
.src2 = !is_inline_constant ? vector_to_scalar_source(v.src2, is_int, is_full, ins->swizzle[1][comp]) : 0,
.unknown = 0,
.outmod = v.outmod,
.output_full = is_full,
.output_component = comp
};
/* Full components are physically spaced out */
if (is_full) {
assert(s.output_component < 4);
s.output_component <<= 1;
}
/* Inline constant is passed along rather than trying to extract it
* from v */
if (ins->has_inline_constant) {
uint16_t imm = 0;
int lower_11 = ins->inline_constant & ((1 << 12) - 1);
imm |= (lower_11 >> 9) & 3;
imm |= (lower_11 >> 6) & 4;
imm |= (lower_11 >> 2) & 0x38;
imm |= (lower_11 & 63) << 6;
s.src2 = imm;
}
return s;
}
/* 64-bit swizzles are super easy since there are 2 components of 2 components
* in an 8-bit field ... lots of duplication to go around!
*
* Swizzles of 32-bit vectors accessed from 64-bit instructions are a little
* funny -- pack them *as if* they were native 64-bit, using rep_* flags to
* flag upper. For instance, xy would become 64-bit XY but that's just xyzw
* native. Likewise, zz would become 64-bit XX with rep* so it would be xyxy
* with rep. Pretty nifty, huh? */
static unsigned
mir_pack_swizzle_64(unsigned *swizzle, unsigned max_component)
{
unsigned packed = 0;
for (unsigned i = 0; i < 2; ++i) {
assert(swizzle[i] <= max_component);
unsigned a = swizzle[i] & 1 ?
(COMPONENT_W << 2) | COMPONENT_Z :
(COMPONENT_Y << 2) | COMPONENT_X;
packed |= a << (i * 4);
}
return packed;
}
static void
mir_pack_swizzle_alu(midgard_instruction *ins)
{
midgard_vector_alu_src src[] = {
vector_alu_from_unsigned(ins->alu.src1),
vector_alu_from_unsigned(ins->alu.src2)
};
for (unsigned i = 0; i < 2; ++i) {
unsigned packed = 0;
if (ins->alu.reg_mode == midgard_reg_mode_64) {
midgard_reg_mode mode = mir_srcsize(ins, i);
unsigned components = 16 / mir_bytes_for_mode(mode);
packed = mir_pack_swizzle_64(ins->swizzle[i], components);
if (mode == midgard_reg_mode_32) {
src[i].rep_low |= (ins->swizzle[i][0] >= COMPONENT_Z);
src[i].rep_high |= (ins->swizzle[i][1] >= COMPONENT_Z);
} else if (mode < midgard_reg_mode_32) {
unreachable("Cannot encode 8/16 swizzle in 64-bit");
}
} else {
/* For 32-bit, swizzle packing is stupid-simple. For 16-bit,
* the strategy is to check whether the nibble we're on is
* upper or lower. We need all components to be on the same
* "side"; that much is enforced by the ISA and should have
* been lowered. TODO: 8-bit packing. TODO: vec8 */
unsigned first = ins->mask ? ffs(ins->mask) - 1 : 0;
bool upper = ins->swizzle[i][first] > 3;
if (upper && ins->mask)
assert(mir_srcsize(ins, i) <= midgard_reg_mode_16);
for (unsigned c = 0; c < 4; ++c) {
unsigned v = ins->swizzle[i][c];
bool t_upper = v > 3;
/* Ensure we're doing something sane */
if (ins->mask & (1 << c)) {
assert(t_upper == upper);
assert(v <= 7);
}
/* Use the non upper part */
v &= 0x3;
packed |= v << (2 * c);
}
src[i].rep_high = upper;
}
src[i].swizzle = packed;
}
ins->alu.src1 = vector_alu_srco_unsigned(src[0]);
if (!ins->has_inline_constant)
ins->alu.src2 = vector_alu_srco_unsigned(src[1]);
}
static void
mir_pack_swizzle_ldst(midgard_instruction *ins)
{
/* TODO: non-32-bit, non-vec4 */
for (unsigned c = 0; c < 4; ++c) {
unsigned v = ins->swizzle[0][c];
/* Check vec4 */
assert(v <= 3);
ins->load_store.swizzle |= v << (2 * c);
}
/* TODO: arg_1/2 */
}
static void
mir_pack_swizzle_tex(midgard_instruction *ins)
{
for (unsigned i = 0; i < 2; ++i) {
unsigned packed = 0;
for (unsigned c = 0; c < 4; ++c) {
unsigned v = ins->swizzle[i][c];
/* Check vec4 */
assert(v <= 3);
packed |= v << (2 * c);
}
if (i == 0)
ins->texture.swizzle = packed;
else
ins->texture.in_reg_swizzle = packed;
}
/* TODO: bias component */
}
/* Load store masks are 4-bits. Load/store ops pack for that. vec4 is the
* natural mask width; vec8 is constrained to be in pairs, vec2 is duplicated. TODO: 8-bit?
*/
static void
mir_pack_ldst_mask(midgard_instruction *ins)
{
midgard_reg_mode mode = mir_typesize(ins);
unsigned packed = ins->mask;
if (mode == midgard_reg_mode_64) {
packed = ((ins->mask & 0x2) ? (0x8 | 0x4) : 0) |
((ins->mask & 0x1) ? (0x2 | 0x1) : 0);
} else if (mode == midgard_reg_mode_16) {
packed = 0;
for (unsigned i = 0; i < 4; ++i) {
/* Make sure we're duplicated */
bool u = (ins->mask & (1 << (2*i + 0))) != 0;
bool v = (ins->mask & (1 << (2*i + 1))) != 0;
assert(u == v);
packed |= (u << i);
}
}
ins->load_store.mask = packed;
}
static void
emit_alu_bundle(compiler_context *ctx,
midgard_bundle *bundle,
struct util_dynarray *emission,
unsigned lookahead)
{
/* Emit the control word */
util_dynarray_append(emission, uint32_t, bundle->control | lookahead);
/* Next up, emit register words */
for (unsigned i = 0; i < bundle->instruction_count; ++i) {
midgard_instruction *ins = bundle->instructions[i];
/* Check if this instruction has registers */
if (ins->compact_branch || ins->prepacked_branch) continue;
/* Otherwise, just emit the registers */
uint16_t reg_word = 0;
memcpy(®_word, &ins->registers, sizeof(uint16_t));
util_dynarray_append(emission, uint16_t, reg_word);
}
/* Now, we emit the body itself */
for (unsigned i = 0; i < bundle->instruction_count; ++i) {
midgard_instruction *ins = bundle->instructions[i];
/* Where is this body */
unsigned size = 0;
void *source = NULL;
/* In case we demote to a scalar */
midgard_scalar_alu scalarized;
if (ins->unit & UNITS_ANY_VECTOR) {
if (ins->alu.reg_mode == midgard_reg_mode_64)
ins->alu.mask = expand_writemask(ins->mask, 2);
else if (ins->alu.reg_mode == midgard_reg_mode_32)
ins->alu.mask = expand_writemask(ins->mask, 4);
else
ins->alu.mask = ins->mask;
mir_pack_swizzle_alu(ins);
size = sizeof(midgard_vector_alu);
source = &ins->alu;
} else if (ins->unit == ALU_ENAB_BR_COMPACT) {
size = sizeof(midgard_branch_cond);
source = &ins->br_compact;
} else if (ins->compact_branch) { /* misnomer */
size = sizeof(midgard_branch_extended);
source = &ins->branch_extended;
} else {
size = sizeof(midgard_scalar_alu);
scalarized = vector_to_scalar_alu(ins->alu, ins);
source = &scalarized;
}
memcpy(util_dynarray_grow_bytes(emission, 1, size), source, size);
}
/* Emit padding (all zero) */
memset(util_dynarray_grow_bytes(emission, 1, bundle->padding), 0, bundle->padding);
/* Tack on constants */
if (bundle->has_embedded_constants) {
util_dynarray_append(emission, float, bundle->constants[0]);
util_dynarray_append(emission, float, bundle->constants[1]);
util_dynarray_append(emission, float, bundle->constants[2]);
util_dynarray_append(emission, float, bundle->constants[3]);
}
}
/* After everything is scheduled, emit whole bundles at a time */
void
emit_binary_bundle(compiler_context *ctx,
midgard_bundle *bundle,
struct util_dynarray *emission,
int next_tag)
{
int lookahead = next_tag << 4;
switch (bundle->tag) {
case TAG_ALU_4:
case TAG_ALU_8:
case TAG_ALU_12:
case TAG_ALU_16:
emit_alu_bundle(ctx, bundle, emission, lookahead);
break;
case TAG_LOAD_STORE_4: {
/* One or two composing instructions */
uint64_t current64, next64 = LDST_NOP;
/* Copy masks */
for (unsigned i = 0; i < bundle->instruction_count; ++i) {
mir_pack_ldst_mask(bundle->instructions[i]);
mir_pack_swizzle_ldst(bundle->instructions[i]);
/* Apply a constant offset */
unsigned offset = bundle->instructions[i]->constants[0];
if (offset) {
bundle->instructions[i]->load_store.varying_parameters |= (offset & 0x7F) << 3;
bundle->instructions[i]->load_store.address |= (offset >> 7);
}
}
memcpy(¤t64, &bundle->instructions[0]->load_store, sizeof(current64));
if (bundle->instruction_count == 2)
memcpy(&next64, &bundle->instructions[1]->load_store, sizeof(next64));
midgard_load_store instruction = {
.type = bundle->tag,
.next_type = next_tag,
.word1 = current64,
.word2 = next64
};
util_dynarray_append(emission, midgard_load_store, instruction);
break;
}
case TAG_TEXTURE_4:
case TAG_TEXTURE_4_VTX: {
/* Texture instructions are easy, since there is no pipelining
* nor VLIW to worry about. We may need to set .cont/.last
* flags. */
midgard_instruction *ins = bundle->instructions[0];
ins->texture.type = bundle->tag;
ins->texture.next_type = next_tag;
ins->texture.mask = ins->mask;
mir_pack_swizzle_tex(ins);
ctx->texture_op_count--;
if (mir_op_computes_derivatives(ctx->stage, ins->texture.op)) {
bool continues = ctx->texture_op_count > 0;
/* Control flow complicates helper invocation
* lifespans, so for now just keep helper threads
* around indefinitely with loops. TODO: Proper
* analysis */
continues |= ctx->loop_count > 0;
ins->texture.cont = continues;
ins->texture.last = !continues;
} else {
ins->texture.cont = ins->texture.last = 1;
}
util_dynarray_append(emission, midgard_texture_word, ins->texture);
break;
}
default:
unreachable("Unknown midgard instruction type\n");
}
}
|