2
* Copyright (C) 2017-2018 Rob Clark <robclark@freedesktop.org>
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:
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
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
24
* Rob Clark <robclark@freedesktop.org>
29
#include "ir3_context.h"
30
#include "ir3_image.h"
33
* Handlers for instructions changed/added in a6xx:
35
* Starting with a6xx, isam and stbi is used for SSBOs as well; stbi and the
36
* atomic instructions (used for both SSBO and image) use a new instruction
37
* encoding compared to a4xx/a5xx.
40
/* src[] = { buffer_index, offset }. No const_index */
42
emit_intrinsic_load_ssbo(struct ir3_context *ctx, nir_intrinsic_instr *intr,
43
struct ir3_instruction **dst)
45
struct ir3_block *b = ctx->block;
46
struct ir3_instruction *offset;
47
struct ir3_instruction *ldib;
49
offset = ir3_get_src(ctx, &intr->src[2])[0];
51
ldib = ir3_LDIB(b, ir3_ssbo_to_ibo(ctx, intr->src[0]), 0, offset, 0);
52
ldib->dsts[0]->wrmask = MASK(intr->num_components);
53
ldib->cat6.iim_val = intr->num_components;
55
ldib->cat6.type = intr->dest.ssa.bit_size == 16 ? TYPE_U16 : TYPE_U32;
56
ldib->barrier_class = IR3_BARRIER_BUFFER_R;
57
ldib->barrier_conflict = IR3_BARRIER_BUFFER_W;
58
ir3_handle_bindless_cat6(ldib, intr->src[0]);
59
ir3_handle_nonuniform(ldib, intr);
61
ir3_split_dest(b, dst, ldib, 0, intr->num_components);
64
/* src[] = { value, block_index, offset }. const_index[] = { write_mask } */
66
emit_intrinsic_store_ssbo(struct ir3_context *ctx, nir_intrinsic_instr *intr)
68
struct ir3_block *b = ctx->block;
69
struct ir3_instruction *stib, *val, *offset;
70
unsigned wrmask = nir_intrinsic_write_mask(intr);
71
unsigned ncomp = ffs(~wrmask) - 1;
73
assert(wrmask == BITFIELD_MASK(intr->num_components));
75
/* src0 is offset, src1 is value:
77
val = ir3_create_collect(b, ir3_get_src(ctx, &intr->src[0]), ncomp);
78
offset = ir3_get_src(ctx, &intr->src[3])[0];
80
stib = ir3_STIB(b, ir3_ssbo_to_ibo(ctx, intr->src[1]), 0, offset, 0, val, 0);
81
stib->cat6.iim_val = ncomp;
83
stib->cat6.type = intr->src[0].ssa->bit_size == 16 ? TYPE_U16 : TYPE_U32;
84
stib->barrier_class = IR3_BARRIER_BUFFER_W;
85
stib->barrier_conflict = IR3_BARRIER_BUFFER_R | IR3_BARRIER_BUFFER_W;
86
ir3_handle_bindless_cat6(stib, intr->src[1]);
87
ir3_handle_nonuniform(stib, intr);
89
array_insert(b, b->keeps, stib);
93
* SSBO atomic intrinsics
95
* All of the SSBO atomic memory operations read a value from memory,
96
* compute a new value using one of the operations below, write the new
97
* value to memory, and return the original value read.
99
* All operations take 3 sources except CompSwap that takes 4. These
102
* 0: The SSBO buffer index.
103
* 1: The offset into the SSBO buffer of the variable that the atomic
104
* operation will operate on.
105
* 2: The data parameter to the atomic function (i.e. the value to add
106
* in ssbo_atomic_add, etc).
107
* 3: For CompSwap only: the second data parameter.
109
static struct ir3_instruction *
110
emit_intrinsic_atomic_ssbo(struct ir3_context *ctx, nir_intrinsic_instr *intr)
112
struct ir3_block *b = ctx->block;
113
struct ir3_instruction *atomic, *ibo, *src0, *src1, *data, *dummy;
114
type_t type = TYPE_U32;
116
ibo = ir3_ssbo_to_ibo(ctx, intr->src[0]);
118
data = ir3_get_src(ctx, &intr->src[2])[0];
120
/* So this gets a bit creative:
122
* src0 - vecN offset/coords
123
* src1.x - is actually destination register
124
* src1.y - is 'data' except for cmpxchg where src2.y is 'compare'
125
* src1.z - is 'data' for cmpxchg
127
* The combining src and dest kinda doesn't work out so well with how
128
* scheduling and RA work. So we create a dummy src2 which is tied to the
129
* destination in RA (i.e. must be allocated to the same vec2/vec3
130
* register) and then immediately extract the first component.
132
* Note that nir already multiplies the offset by four
134
dummy = create_immed(b, 0);
136
if (intr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap_ir3) {
137
src0 = ir3_get_src(ctx, &intr->src[4])[0];
138
struct ir3_instruction *compare = ir3_get_src(ctx, &intr->src[3])[0];
139
src1 = ir3_collect(b, dummy, compare, data);
141
src0 = ir3_get_src(ctx, &intr->src[3])[0];
142
src1 = ir3_collect(b, dummy, data);
145
switch (intr->intrinsic) {
146
case nir_intrinsic_ssbo_atomic_add_ir3:
147
atomic = ir3_ATOMIC_B_ADD(b, ibo, 0, src0, 0, src1, 0);
149
case nir_intrinsic_ssbo_atomic_imin_ir3:
150
atomic = ir3_ATOMIC_B_MIN(b, ibo, 0, src0, 0, src1, 0);
153
case nir_intrinsic_ssbo_atomic_umin_ir3:
154
atomic = ir3_ATOMIC_B_MIN(b, ibo, 0, src0, 0, src1, 0);
156
case nir_intrinsic_ssbo_atomic_imax_ir3:
157
atomic = ir3_ATOMIC_B_MAX(b, ibo, 0, src0, 0, src1, 0);
160
case nir_intrinsic_ssbo_atomic_umax_ir3:
161
atomic = ir3_ATOMIC_B_MAX(b, ibo, 0, src0, 0, src1, 0);
163
case nir_intrinsic_ssbo_atomic_and_ir3:
164
atomic = ir3_ATOMIC_B_AND(b, ibo, 0, src0, 0, src1, 0);
166
case nir_intrinsic_ssbo_atomic_or_ir3:
167
atomic = ir3_ATOMIC_B_OR(b, ibo, 0, src0, 0, src1, 0);
169
case nir_intrinsic_ssbo_atomic_xor_ir3:
170
atomic = ir3_ATOMIC_B_XOR(b, ibo, 0, src0, 0, src1, 0);
172
case nir_intrinsic_ssbo_atomic_exchange_ir3:
173
atomic = ir3_ATOMIC_B_XCHG(b, ibo, 0, src0, 0, src1, 0);
175
case nir_intrinsic_ssbo_atomic_comp_swap_ir3:
176
atomic = ir3_ATOMIC_B_CMPXCHG(b, ibo, 0, src0, 0, src1, 0);
182
atomic->cat6.iim_val = 1;
184
atomic->cat6.type = type;
185
atomic->barrier_class = IR3_BARRIER_BUFFER_W;
186
atomic->barrier_conflict = IR3_BARRIER_BUFFER_R | IR3_BARRIER_BUFFER_W;
187
ir3_handle_bindless_cat6(atomic, intr->src[0]);
189
/* even if nothing consume the result, we can't DCE the instruction: */
190
array_insert(b, b->keeps, atomic);
192
atomic->dsts[0]->wrmask = src1->dsts[0]->wrmask;
193
ir3_reg_tie(atomic->dsts[0], atomic->srcs[2]);
194
struct ir3_instruction *split;
195
ir3_split_dest(b, &split, atomic, 0, 1);
199
/* src[] = { deref, coord, sample_index }. const_index[] = {} */
201
emit_intrinsic_load_image(struct ir3_context *ctx, nir_intrinsic_instr *intr,
202
struct ir3_instruction **dst)
204
struct ir3_block *b = ctx->block;
205
struct ir3_instruction *ldib;
206
struct ir3_instruction *const *coords = ir3_get_src(ctx, &intr->src[1]);
207
unsigned ncoords = ir3_get_image_coords(intr, NULL);
209
ldib = ir3_LDIB(b, ir3_image_to_ibo(ctx, intr->src[0]), 0,
210
ir3_create_collect(b, coords, ncoords), 0);
211
ldib->dsts[0]->wrmask = MASK(intr->num_components);
212
ldib->cat6.iim_val = intr->num_components;
213
ldib->cat6.d = ncoords;
214
ldib->cat6.type = ir3_get_type_for_image_intrinsic(intr);
215
ldib->cat6.typed = true;
216
ldib->barrier_class = IR3_BARRIER_IMAGE_R;
217
ldib->barrier_conflict = IR3_BARRIER_IMAGE_W;
218
ir3_handle_bindless_cat6(ldib, intr->src[0]);
219
ir3_handle_nonuniform(ldib, intr);
221
ir3_split_dest(b, dst, ldib, 0, intr->num_components);
224
/* src[] = { deref, coord, sample_index, value }. const_index[] = {} */
226
emit_intrinsic_store_image(struct ir3_context *ctx, nir_intrinsic_instr *intr)
228
struct ir3_block *b = ctx->block;
229
struct ir3_instruction *stib;
230
struct ir3_instruction *const *value = ir3_get_src(ctx, &intr->src[3]);
231
struct ir3_instruction *const *coords = ir3_get_src(ctx, &intr->src[1]);
232
unsigned ncoords = ir3_get_image_coords(intr, NULL);
233
enum pipe_format format = nir_intrinsic_format(intr);
234
unsigned ncomp = ir3_get_num_components_for_image_format(format);
236
/* src0 is offset, src1 is value:
238
stib = ir3_STIB(b, ir3_image_to_ibo(ctx, intr->src[0]), 0,
239
ir3_create_collect(b, coords, ncoords), 0,
240
ir3_create_collect(b, value, ncomp), 0);
241
stib->cat6.iim_val = ncomp;
242
stib->cat6.d = ncoords;
243
stib->cat6.type = ir3_get_type_for_image_intrinsic(intr);
244
stib->cat6.typed = true;
245
stib->barrier_class = IR3_BARRIER_IMAGE_W;
246
stib->barrier_conflict = IR3_BARRIER_IMAGE_R | IR3_BARRIER_IMAGE_W;
247
ir3_handle_bindless_cat6(stib, intr->src[0]);
248
ir3_handle_nonuniform(stib, intr);
250
array_insert(b, b->keeps, stib);
253
/* src[] = { deref, coord, sample_index, value, compare }. const_index[] = {} */
254
static struct ir3_instruction *
255
emit_intrinsic_atomic_image(struct ir3_context *ctx, nir_intrinsic_instr *intr)
257
struct ir3_block *b = ctx->block;
258
struct ir3_instruction *atomic, *ibo, *src0, *src1, *dummy;
259
struct ir3_instruction *const *coords = ir3_get_src(ctx, &intr->src[1]);
260
struct ir3_instruction *value = ir3_get_src(ctx, &intr->src[3])[0];
261
unsigned ncoords = ir3_get_image_coords(intr, NULL);
263
ibo = ir3_image_to_ibo(ctx, intr->src[0]);
265
/* So this gets a bit creative:
267
* src0 - vecN offset/coords
268
* src1.x - is actually destination register
269
* src1.y - is 'value' except for cmpxchg where src2.y is 'compare'
270
* src1.z - is 'value' for cmpxchg
272
* The combining src and dest kinda doesn't work out so well with how
273
* scheduling and RA work. So we create a dummy src2 which is tied to the
274
* destination in RA (i.e. must be allocated to the same vec2/vec3
275
* register) and then immediately extract the first component.
277
dummy = create_immed(b, 0);
278
src0 = ir3_create_collect(b, coords, ncoords);
280
if (intr->intrinsic == nir_intrinsic_image_atomic_comp_swap ||
281
intr->intrinsic == nir_intrinsic_bindless_image_atomic_comp_swap) {
282
struct ir3_instruction *compare = ir3_get_src(ctx, &intr->src[4])[0];
283
src1 = ir3_collect(b, dummy, compare, value);
285
src1 = ir3_collect(b, dummy, value);
288
switch (intr->intrinsic) {
289
case nir_intrinsic_image_atomic_add:
290
case nir_intrinsic_bindless_image_atomic_add:
291
atomic = ir3_ATOMIC_B_ADD(b, ibo, 0, src0, 0, src1, 0);
293
case nir_intrinsic_image_atomic_imin:
294
case nir_intrinsic_image_atomic_umin:
295
case nir_intrinsic_bindless_image_atomic_imin:
296
case nir_intrinsic_bindless_image_atomic_umin:
297
atomic = ir3_ATOMIC_B_MIN(b, ibo, 0, src0, 0, src1, 0);
299
case nir_intrinsic_image_atomic_imax:
300
case nir_intrinsic_image_atomic_umax:
301
case nir_intrinsic_bindless_image_atomic_imax:
302
case nir_intrinsic_bindless_image_atomic_umax:
303
atomic = ir3_ATOMIC_B_MAX(b, ibo, 0, src0, 0, src1, 0);
305
case nir_intrinsic_image_atomic_and:
306
case nir_intrinsic_bindless_image_atomic_and:
307
atomic = ir3_ATOMIC_B_AND(b, ibo, 0, src0, 0, src1, 0);
309
case nir_intrinsic_image_atomic_or:
310
case nir_intrinsic_bindless_image_atomic_or:
311
atomic = ir3_ATOMIC_B_OR(b, ibo, 0, src0, 0, src1, 0);
313
case nir_intrinsic_image_atomic_xor:
314
case nir_intrinsic_bindless_image_atomic_xor:
315
atomic = ir3_ATOMIC_B_XOR(b, ibo, 0, src0, 0, src1, 0);
317
case nir_intrinsic_image_atomic_exchange:
318
case nir_intrinsic_bindless_image_atomic_exchange:
319
atomic = ir3_ATOMIC_B_XCHG(b, ibo, 0, src0, 0, src1, 0);
321
case nir_intrinsic_image_atomic_comp_swap:
322
case nir_intrinsic_bindless_image_atomic_comp_swap:
323
atomic = ir3_ATOMIC_B_CMPXCHG(b, ibo, 0, src0, 0, src1, 0);
329
atomic->cat6.iim_val = 1;
330
atomic->cat6.d = ncoords;
331
atomic->cat6.type = ir3_get_type_for_image_intrinsic(intr);
332
atomic->cat6.typed = true;
333
atomic->barrier_class = IR3_BARRIER_IMAGE_W;
334
atomic->barrier_conflict = IR3_BARRIER_IMAGE_R | IR3_BARRIER_IMAGE_W;
335
ir3_handle_bindless_cat6(atomic, intr->src[0]);
337
/* even if nothing consume the result, we can't DCE the instruction: */
338
array_insert(b, b->keeps, atomic);
340
atomic->dsts[0]->wrmask = src1->dsts[0]->wrmask;
341
ir3_reg_tie(atomic->dsts[0], atomic->srcs[2]);
342
struct ir3_instruction *split;
343
ir3_split_dest(b, &split, atomic, 0, 1);
348
emit_intrinsic_image_size(struct ir3_context *ctx, nir_intrinsic_instr *intr,
349
struct ir3_instruction **dst)
351
struct ir3_block *b = ctx->block;
352
struct ir3_instruction *ibo = ir3_image_to_ibo(ctx, intr->src[0]);
353
struct ir3_instruction *resinfo = ir3_RESINFO(b, ibo, 0);
354
resinfo->cat6.iim_val = 1;
355
resinfo->cat6.d = intr->num_components;
356
resinfo->cat6.type = TYPE_U32;
357
resinfo->cat6.typed = false;
358
/* resinfo has no writemask and always writes out 3 components: */
359
compile_assert(ctx, intr->num_components <= 3);
360
resinfo->dsts[0]->wrmask = MASK(3);
361
ir3_handle_bindless_cat6(resinfo, intr->src[0]);
362
ir3_handle_nonuniform(resinfo, intr);
364
ir3_split_dest(b, dst, resinfo, 0, intr->num_components);
368
emit_intrinsic_load_global_ir3(struct ir3_context *ctx,
369
nir_intrinsic_instr *intr,
370
struct ir3_instruction **dst)
372
struct ir3_block *b = ctx->block;
373
unsigned dest_components = nir_intrinsic_dest_components(intr);
374
struct ir3_instruction *addr, *offset;
376
addr = ir3_collect(b, ir3_get_src(ctx, &intr->src[0])[0],
377
ir3_get_src(ctx, &intr->src[0])[1]);
379
struct ir3_instruction *load;
381
bool const_offset_in_bounds = nir_src_is_const(intr->src[1]) &&
382
nir_src_as_int(intr->src[1]) < (1 << 13) &&
383
nir_src_as_int(intr->src[1]) > -(1 << 13);
385
if (const_offset_in_bounds) {
386
load = ir3_LDG(b, addr, 0, create_immed(b, nir_src_as_int(intr->src[1])),
387
0, create_immed(b, dest_components), 0);
389
offset = ir3_get_src(ctx, &intr->src[1])[0];
391
ir3_LDG_A(b, addr, 0, offset, 0, create_immed(b, 0), 0,
392
create_immed(b, 0), 0, create_immed(b, dest_components), 0);
395
load->cat6.type = type_uint_size(intr->dest.ssa.bit_size);
396
load->dsts[0]->wrmask = MASK(dest_components);
398
load->barrier_class = IR3_BARRIER_BUFFER_R;
399
load->barrier_conflict = IR3_BARRIER_BUFFER_W;
401
ir3_split_dest(b, dst, load, 0, dest_components);
405
emit_intrinsic_store_global_ir3(struct ir3_context *ctx,
406
nir_intrinsic_instr *intr)
408
struct ir3_block *b = ctx->block;
409
struct ir3_instruction *value, *addr, *offset;
410
unsigned ncomp = nir_intrinsic_src_components(intr, 0);
412
addr = ir3_collect(b, ir3_get_src(ctx, &intr->src[1])[0],
413
ir3_get_src(ctx, &intr->src[1])[1]);
415
value = ir3_create_collect(b, ir3_get_src(ctx, &intr->src[0]), ncomp);
417
struct ir3_instruction *stg;
419
bool const_offset_in_bounds = nir_src_is_const(intr->src[2]) &&
420
nir_src_as_int(intr->src[2]) < (1 << 13) &&
421
nir_src_as_int(intr->src[2]) > -(1 << 13);
423
if (const_offset_in_bounds) {
424
stg = ir3_STG(b, addr, 0,
425
create_immed(b, nir_src_as_int(intr->src[2])), 0,
427
create_immed(b, ncomp), 0);
429
offset = ir3_get_src(ctx, &intr->src[2])[0];
431
ir3_STG_A(b, addr, 0, offset, 0, create_immed(b, 0), 0,
432
create_immed(b, 0), 0, value, 0, create_immed(b, ncomp), 0);
435
stg->cat6.type = type_uint_size(intr->src[0].ssa->bit_size);
436
stg->cat6.iim_val = 1;
438
array_insert(b, b->keeps, stg);
440
stg->barrier_class = IR3_BARRIER_BUFFER_W;
441
stg->barrier_conflict = IR3_BARRIER_BUFFER_R | IR3_BARRIER_BUFFER_W;
444
static struct ir3_instruction *
445
emit_intrinsic_atomic_global(struct ir3_context *ctx, nir_intrinsic_instr *intr)
447
struct ir3_block *b = ctx->block;
448
struct ir3_instruction *addr, *atomic, *src1;
449
struct ir3_instruction *value = ir3_get_src(ctx, &intr->src[1])[0];
450
type_t type = TYPE_U32;
452
addr = ir3_collect(b, ir3_get_src(ctx, &intr->src[0])[0],
453
ir3_get_src(ctx, &intr->src[0])[1]);
455
if (intr->intrinsic == nir_intrinsic_global_atomic_comp_swap_ir3) {
456
struct ir3_instruction *compare = ir3_get_src(ctx, &intr->src[2])[0];
457
src1 = ir3_collect(b, compare, value);
462
switch (intr->intrinsic) {
463
case nir_intrinsic_global_atomic_add_ir3:
464
atomic = ir3_ATOMIC_G_ADD(b, addr, 0, src1, 0);
466
case nir_intrinsic_global_atomic_imin_ir3:
467
atomic = ir3_ATOMIC_G_MIN(b, addr, 0, src1, 0);
470
case nir_intrinsic_global_atomic_umin_ir3:
471
atomic = ir3_ATOMIC_G_MIN(b, addr, 0, src1, 0);
473
case nir_intrinsic_global_atomic_imax_ir3:
474
atomic = ir3_ATOMIC_G_MAX(b, addr, 0, src1, 0);
477
case nir_intrinsic_global_atomic_umax_ir3:
478
atomic = ir3_ATOMIC_G_MAX(b, addr, 0, src1, 0);
480
case nir_intrinsic_global_atomic_and_ir3:
481
atomic = ir3_ATOMIC_G_AND(b, addr, 0, src1, 0);
483
case nir_intrinsic_global_atomic_or_ir3:
484
atomic = ir3_ATOMIC_G_OR(b, addr, 0, src1, 0);
486
case nir_intrinsic_global_atomic_xor_ir3:
487
atomic = ir3_ATOMIC_G_XOR(b, addr, 0, src1, 0);
489
case nir_intrinsic_global_atomic_exchange_ir3:
490
atomic = ir3_ATOMIC_G_XCHG(b, addr, 0, src1, 0);
492
case nir_intrinsic_global_atomic_comp_swap_ir3:
493
atomic = ir3_ATOMIC_G_CMPXCHG(b, addr, 0, src1, 0);
496
unreachable("Unknown global atomic op");
499
atomic->cat6.iim_val = 1;
501
atomic->cat6.type = type;
502
atomic->barrier_class = IR3_BARRIER_BUFFER_W;
503
atomic->barrier_conflict = IR3_BARRIER_BUFFER_R | IR3_BARRIER_BUFFER_W;
505
/* even if nothing consume the result, we can't DCE the instruction: */
506
array_insert(b, b->keeps, atomic);
511
const struct ir3_context_funcs ir3_a6xx_funcs = {
512
.emit_intrinsic_load_ssbo = emit_intrinsic_load_ssbo,
513
.emit_intrinsic_store_ssbo = emit_intrinsic_store_ssbo,
514
.emit_intrinsic_atomic_ssbo = emit_intrinsic_atomic_ssbo,
515
.emit_intrinsic_load_image = emit_intrinsic_load_image,
516
.emit_intrinsic_store_image = emit_intrinsic_store_image,
517
.emit_intrinsic_atomic_image = emit_intrinsic_atomic_image,
518
.emit_intrinsic_image_size = emit_intrinsic_image_size,
519
.emit_intrinsic_load_global_ir3 = emit_intrinsic_load_global_ir3,
520
.emit_intrinsic_store_global_ir3 = emit_intrinsic_store_global_ir3,
521
.emit_intrinsic_atomic_global = emit_intrinsic_atomic_global,