2
* Copyright © 2016 Broadcom
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
20
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25
#include "nir_builder.h"
26
#include "nir_deref.h"
28
/** @file nir_lower_io_to_scalar.c
30
* Replaces nir_load_input/nir_store_output operations with num_components !=
31
* 1 with individual per-channel operations.
35
set_io_semantics(nir_intrinsic_instr *scalar_intr,
36
nir_intrinsic_instr *vec_intr, unsigned component)
38
nir_io_semantics sem = nir_intrinsic_io_semantics(vec_intr);
39
sem.gs_streams = (sem.gs_streams >> (component * 2)) & 0x3;
40
nir_intrinsic_set_io_semantics(scalar_intr, sem);
44
lower_load_input_to_scalar(nir_builder *b, nir_intrinsic_instr *intr)
46
b->cursor = nir_before_instr(&intr->instr);
48
assert(intr->dest.is_ssa);
50
nir_ssa_def *loads[NIR_MAX_VEC_COMPONENTS];
52
for (unsigned i = 0; i < intr->num_components; i++) {
53
nir_intrinsic_instr *chan_intr =
54
nir_intrinsic_instr_create(b->shader, intr->intrinsic);
55
nir_ssa_dest_init(&chan_intr->instr, &chan_intr->dest,
56
1, intr->dest.ssa.bit_size, NULL);
57
chan_intr->num_components = 1;
59
nir_intrinsic_set_base(chan_intr, nir_intrinsic_base(intr));
60
nir_intrinsic_set_component(chan_intr, nir_intrinsic_component(intr) + i);
61
nir_intrinsic_set_dest_type(chan_intr, nir_intrinsic_dest_type(intr));
62
set_io_semantics(chan_intr, intr, i);
64
nir_src_copy(&chan_intr->src[0], &intr->src[0]);
66
nir_builder_instr_insert(b, &chan_intr->instr);
68
loads[i] = &chan_intr->dest.ssa;
71
nir_ssa_def_rewrite_uses(&intr->dest.ssa,
72
nir_vec(b, loads, intr->num_components));
73
nir_instr_remove(&intr->instr);
77
lower_load_to_scalar(nir_builder *b, nir_intrinsic_instr *intr)
79
b->cursor = nir_before_instr(&intr->instr);
81
assert(intr->dest.is_ssa);
83
nir_ssa_def *loads[NIR_MAX_VEC_COMPONENTS];
84
unsigned offset_idx = intr->intrinsic == nir_intrinsic_load_shared ? 0 : 1;
85
nir_ssa_def *base_offset = intr->src[offset_idx].ssa;
87
for (unsigned i = 0; i < intr->num_components; i++) {
88
nir_intrinsic_instr *chan_intr =
89
nir_intrinsic_instr_create(b->shader, intr->intrinsic);
90
nir_ssa_dest_init(&chan_intr->instr, &chan_intr->dest,
91
1, intr->dest.ssa.bit_size, NULL);
92
chan_intr->num_components = 1;
94
nir_intrinsic_set_align_offset(chan_intr,
95
(nir_intrinsic_align_offset(intr) +
96
i * (intr->dest.ssa.bit_size / 8)) % nir_intrinsic_align_mul(intr));
97
nir_intrinsic_set_align_mul(chan_intr, nir_intrinsic_align_mul(intr));
98
if (nir_intrinsic_has_access(intr))
99
nir_intrinsic_set_access(chan_intr, nir_intrinsic_access(intr));
100
if (nir_intrinsic_has_range(intr))
101
nir_intrinsic_set_range(chan_intr, nir_intrinsic_range(intr));
102
if (nir_intrinsic_has_range_base(intr))
103
nir_intrinsic_set_range_base(chan_intr, nir_intrinsic_range_base(intr));
104
if (nir_intrinsic_has_base(intr))
105
nir_intrinsic_set_base(chan_intr, nir_intrinsic_base(intr));
106
for (unsigned j = 0; j < nir_intrinsic_infos[intr->intrinsic].num_srcs - 1; j++)
107
nir_src_copy(&chan_intr->src[j], &intr->src[j]);
109
/* increment offset per component */
110
nir_ssa_def *offset = nir_iadd_imm(b, base_offset, i * (intr->dest.ssa.bit_size / 8));
111
chan_intr->src[offset_idx] = nir_src_for_ssa(offset);
113
nir_builder_instr_insert(b, &chan_intr->instr);
115
loads[i] = &chan_intr->dest.ssa;
118
nir_ssa_def_rewrite_uses(&intr->dest.ssa,
119
nir_vec(b, loads, intr->num_components));
120
nir_instr_remove(&intr->instr);
124
lower_store_output_to_scalar(nir_builder *b, nir_intrinsic_instr *intr)
126
b->cursor = nir_before_instr(&intr->instr);
128
nir_ssa_def *value = nir_ssa_for_src(b, intr->src[0], intr->num_components);
130
for (unsigned i = 0; i < intr->num_components; i++) {
131
if (!(nir_intrinsic_write_mask(intr) & (1 << i)))
134
nir_intrinsic_instr *chan_intr =
135
nir_intrinsic_instr_create(b->shader, intr->intrinsic);
136
chan_intr->num_components = 1;
138
nir_intrinsic_set_base(chan_intr, nir_intrinsic_base(intr));
139
nir_intrinsic_set_write_mask(chan_intr, 0x1);
140
nir_intrinsic_set_component(chan_intr, nir_intrinsic_component(intr) + i);
141
nir_intrinsic_set_src_type(chan_intr, nir_intrinsic_src_type(intr));
142
set_io_semantics(chan_intr, intr, i);
144
/* Scalarize transform feedback info. */
145
unsigned component = nir_intrinsic_component(chan_intr);
147
for (unsigned c = 0; c <= component; c++) {
148
nir_io_xfb xfb = c < 2 ? nir_intrinsic_io_xfb(intr) :
149
nir_intrinsic_io_xfb2(intr);
151
if (component < c + xfb.out[c % 2].num_components) {
152
nir_io_xfb scalar_xfb;
154
memset(&scalar_xfb, 0, sizeof(scalar_xfb));
155
scalar_xfb.out[component % 2].num_components = 1;
156
scalar_xfb.out[component % 2].buffer = xfb.out[c % 2].buffer;
157
scalar_xfb.out[component % 2].offset = xfb.out[c % 2].offset +
160
nir_intrinsic_set_io_xfb(chan_intr, scalar_xfb);
162
nir_intrinsic_set_io_xfb2(chan_intr, scalar_xfb);
168
chan_intr->src[0] = nir_src_for_ssa(nir_channel(b, value, i));
170
nir_src_copy(&chan_intr->src[1], &intr->src[1]);
172
nir_builder_instr_insert(b, &chan_intr->instr);
175
nir_instr_remove(&intr->instr);
179
lower_store_to_scalar(nir_builder *b, nir_intrinsic_instr *intr)
181
b->cursor = nir_before_instr(&intr->instr);
183
nir_ssa_def *value = nir_ssa_for_src(b, intr->src[0], intr->num_components);
184
unsigned offset_idx = intr->intrinsic == nir_intrinsic_store_shared ? 1 : 2;
185
nir_ssa_def *base_offset = intr->src[offset_idx].ssa;
187
/* iterate wrmask instead of num_components to handle split components */
188
u_foreach_bit(i, nir_intrinsic_write_mask(intr)) {
189
nir_intrinsic_instr *chan_intr =
190
nir_intrinsic_instr_create(b->shader, intr->intrinsic);
191
chan_intr->num_components = 1;
193
nir_intrinsic_set_write_mask(chan_intr, 0x1);
194
nir_intrinsic_set_align_offset(chan_intr,
195
(nir_intrinsic_align_offset(intr) +
196
i * (value->bit_size / 8)) % nir_intrinsic_align_mul(intr));
197
nir_intrinsic_set_align_mul(chan_intr, nir_intrinsic_align_mul(intr));
198
if (nir_intrinsic_has_access(intr))
199
nir_intrinsic_set_access(chan_intr, nir_intrinsic_access(intr));
200
if (nir_intrinsic_has_base(intr))
201
nir_intrinsic_set_base(chan_intr, nir_intrinsic_base(intr));
204
chan_intr->src[0] = nir_src_for_ssa(nir_channel(b, value, i));
205
for (unsigned j = 1; j < nir_intrinsic_infos[intr->intrinsic].num_srcs - 1; j++)
206
nir_src_copy(&chan_intr->src[j], &intr->src[j]);
208
/* increment offset per component */
209
nir_ssa_def *offset = nir_iadd_imm(b, base_offset, i * (value->bit_size / 8));
210
chan_intr->src[offset_idx] = nir_src_for_ssa(offset);
212
nir_builder_instr_insert(b, &chan_intr->instr);
215
nir_instr_remove(&intr->instr);
219
nir_lower_io_to_scalar_instr(nir_builder *b, nir_instr *instr, void *data)
221
nir_variable_mode mask = *(nir_variable_mode *)data;
223
if (instr->type != nir_instr_type_intrinsic)
226
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
228
if (intr->num_components == 1)
231
if (intr->intrinsic == nir_intrinsic_load_input &&
232
(mask & nir_var_shader_in)) {
233
lower_load_input_to_scalar(b, intr);
237
if ((intr->intrinsic == nir_intrinsic_load_ubo && (mask & nir_var_mem_ubo)) ||
238
(intr->intrinsic == nir_intrinsic_load_ssbo && (mask & nir_var_mem_ssbo)) ||
239
(intr->intrinsic == nir_intrinsic_load_shared && (mask & nir_var_mem_shared))) {
240
lower_load_to_scalar(b, intr);
244
if (intr->intrinsic == nir_intrinsic_store_output &&
245
mask & nir_var_shader_out) {
246
lower_store_output_to_scalar(b, intr);
250
if ((intr->intrinsic == nir_intrinsic_store_ssbo && (mask & nir_var_mem_ssbo)) ||
251
(intr->intrinsic == nir_intrinsic_store_shared && (mask & nir_var_mem_shared))) {
252
lower_store_to_scalar(b, intr);
260
nir_lower_io_to_scalar(nir_shader *shader, nir_variable_mode mask)
262
nir_shader_instructions_pass(shader,
263
nir_lower_io_to_scalar_instr,
264
nir_metadata_block_index |
265
nir_metadata_dominance,
269
static nir_variable **
270
get_channel_variables(struct hash_table *ht, nir_variable *var)
272
nir_variable **chan_vars;
273
struct hash_entry *entry = _mesa_hash_table_search(ht, var);
275
chan_vars = (nir_variable **) calloc(4, sizeof(nir_variable *));
276
_mesa_hash_table_insert(ht, var, chan_vars);
278
chan_vars = (nir_variable **) entry->data;
285
* Note that the src deref that we are cloning is the head of the
286
* chain of deref instructions from the original intrinsic, but
287
* the dst we are cloning to is the tail (because chains of deref
288
* instructions are created back to front)
291
static nir_deref_instr *
292
clone_deref_array(nir_builder *b, nir_deref_instr *dst_tail,
293
const nir_deref_instr *src_head)
295
const nir_deref_instr *parent = nir_deref_instr_parent(src_head);
300
assert(src_head->deref_type == nir_deref_type_array);
302
dst_tail = clone_deref_array(b, dst_tail, parent);
304
return nir_build_deref_array(b, dst_tail,
305
nir_ssa_for_src(b, src_head->arr.index, 1));
309
lower_load_to_scalar_early(nir_builder *b, nir_intrinsic_instr *intr,
310
nir_variable *var, struct hash_table *split_inputs,
311
struct hash_table *split_outputs)
313
b->cursor = nir_before_instr(&intr->instr);
315
assert(intr->dest.is_ssa);
317
nir_ssa_def *loads[NIR_MAX_VEC_COMPONENTS];
319
nir_variable **chan_vars;
320
if (var->data.mode == nir_var_shader_in) {
321
chan_vars = get_channel_variables(split_inputs, var);
323
chan_vars = get_channel_variables(split_outputs, var);
326
for (unsigned i = 0; i < intr->num_components; i++) {
327
nir_variable *chan_var = chan_vars[var->data.location_frac + i];
328
if (!chan_vars[var->data.location_frac + i]) {
329
chan_var = nir_variable_clone(var, b->shader);
330
chan_var->data.location_frac = var->data.location_frac + i;
331
chan_var->type = glsl_channel_type(chan_var->type);
332
if (var->data.explicit_offset) {
333
unsigned comp_size = glsl_get_bit_size(chan_var->type) / 8;
334
chan_var->data.offset = var->data.offset + i * comp_size;
337
chan_vars[var->data.location_frac + i] = chan_var;
339
nir_shader_add_variable(b->shader, chan_var);
342
nir_intrinsic_instr *chan_intr =
343
nir_intrinsic_instr_create(b->shader, intr->intrinsic);
344
nir_ssa_dest_init(&chan_intr->instr, &chan_intr->dest,
345
1, intr->dest.ssa.bit_size, NULL);
346
chan_intr->num_components = 1;
348
nir_deref_instr *deref = nir_build_deref_var(b, chan_var);
350
deref = clone_deref_array(b, deref, nir_src_as_deref(intr->src[0]));
352
chan_intr->src[0] = nir_src_for_ssa(&deref->dest.ssa);
354
if (intr->intrinsic == nir_intrinsic_interp_deref_at_offset ||
355
intr->intrinsic == nir_intrinsic_interp_deref_at_sample ||
356
intr->intrinsic == nir_intrinsic_interp_deref_at_vertex)
357
nir_src_copy(&chan_intr->src[1], &intr->src[1]);
359
nir_builder_instr_insert(b, &chan_intr->instr);
361
loads[i] = &chan_intr->dest.ssa;
364
nir_ssa_def_rewrite_uses(&intr->dest.ssa,
365
nir_vec(b, loads, intr->num_components));
367
/* Remove the old load intrinsic */
368
nir_instr_remove(&intr->instr);
372
lower_store_output_to_scalar_early(nir_builder *b, nir_intrinsic_instr *intr,
374
struct hash_table *split_outputs)
376
b->cursor = nir_before_instr(&intr->instr);
378
nir_ssa_def *value = nir_ssa_for_src(b, intr->src[1], intr->num_components);
380
nir_variable **chan_vars = get_channel_variables(split_outputs, var);
381
for (unsigned i = 0; i < intr->num_components; i++) {
382
if (!(nir_intrinsic_write_mask(intr) & (1 << i)))
385
nir_variable *chan_var = chan_vars[var->data.location_frac + i];
386
if (!chan_vars[var->data.location_frac + i]) {
387
chan_var = nir_variable_clone(var, b->shader);
388
chan_var->data.location_frac = var->data.location_frac + i;
389
chan_var->type = glsl_channel_type(chan_var->type);
390
if (var->data.explicit_offset) {
391
unsigned comp_size = glsl_get_bit_size(chan_var->type) / 8;
392
chan_var->data.offset = var->data.offset + i * comp_size;
395
chan_vars[var->data.location_frac + i] = chan_var;
397
nir_shader_add_variable(b->shader, chan_var);
400
nir_intrinsic_instr *chan_intr =
401
nir_intrinsic_instr_create(b->shader, intr->intrinsic);
402
chan_intr->num_components = 1;
404
nir_intrinsic_set_write_mask(chan_intr, 0x1);
406
nir_deref_instr *deref = nir_build_deref_var(b, chan_var);
408
deref = clone_deref_array(b, deref, nir_src_as_deref(intr->src[0]));
410
chan_intr->src[0] = nir_src_for_ssa(&deref->dest.ssa);
411
chan_intr->src[1] = nir_src_for_ssa(nir_channel(b, value, i));
413
nir_builder_instr_insert(b, &chan_intr->instr);
416
/* Remove the old store intrinsic */
417
nir_instr_remove(&intr->instr);
420
struct io_to_scalar_early_state {
421
struct hash_table *split_inputs, *split_outputs;
422
nir_variable_mode mask;
426
nir_lower_io_to_scalar_early_instr(nir_builder *b, nir_instr *instr, void *data)
428
struct io_to_scalar_early_state *state = data;
430
if (instr->type != nir_instr_type_intrinsic)
433
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
435
if (intr->num_components == 1)
438
if (intr->intrinsic != nir_intrinsic_load_deref &&
439
intr->intrinsic != nir_intrinsic_store_deref &&
440
intr->intrinsic != nir_intrinsic_interp_deref_at_centroid &&
441
intr->intrinsic != nir_intrinsic_interp_deref_at_sample &&
442
intr->intrinsic != nir_intrinsic_interp_deref_at_offset &&
443
intr->intrinsic != nir_intrinsic_interp_deref_at_vertex)
446
nir_deref_instr *deref = nir_src_as_deref(intr->src[0]);
447
if (!nir_deref_mode_is_one_of(deref, state->mask))
450
nir_variable *var = nir_deref_instr_get_variable(deref);
451
nir_variable_mode mode = var->data.mode;
453
/* TODO: add patch support */
457
/* TODO: add doubles support */
458
if (glsl_type_is_64bit(glsl_without_array(var->type)))
461
if (!(b->shader->info.stage == MESA_SHADER_VERTEX &&
462
mode == nir_var_shader_in) &&
463
var->data.location < VARYING_SLOT_VAR0 &&
464
var->data.location >= 0)
467
/* Don't bother splitting if we can't opt away any unused
470
if (var->data.always_active_io)
473
/* Skip types we cannot split */
474
if (glsl_type_is_matrix(glsl_without_array(var->type)) ||
475
glsl_type_is_struct_or_ifc(glsl_without_array(var->type)))
478
switch (intr->intrinsic) {
479
case nir_intrinsic_interp_deref_at_centroid:
480
case nir_intrinsic_interp_deref_at_sample:
481
case nir_intrinsic_interp_deref_at_offset:
482
case nir_intrinsic_interp_deref_at_vertex:
483
case nir_intrinsic_load_deref:
484
if ((state->mask & nir_var_shader_in && mode == nir_var_shader_in) ||
485
(state->mask & nir_var_shader_out && mode == nir_var_shader_out)) {
486
lower_load_to_scalar_early(b, intr, var, state->split_inputs,
487
state->split_outputs);
491
case nir_intrinsic_store_deref:
492
if (state->mask & nir_var_shader_out &&
493
mode == nir_var_shader_out) {
494
lower_store_output_to_scalar_early(b, intr, var, state->split_outputs);
506
* This function is intended to be called earlier than nir_lower_io_to_scalar()
507
* i.e. before nir_lower_io() is called.
510
nir_lower_io_to_scalar_early(nir_shader *shader, nir_variable_mode mask)
512
struct io_to_scalar_early_state state = {
513
.split_inputs = _mesa_pointer_hash_table_create(NULL),
514
.split_outputs = _mesa_pointer_hash_table_create(NULL),
518
bool progress = nir_shader_instructions_pass(shader,
519
nir_lower_io_to_scalar_early_instr,
520
nir_metadata_block_index |
521
nir_metadata_dominance,
524
/* Remove old input from the shaders inputs list */
525
hash_table_foreach(state.split_inputs, entry) {
526
nir_variable *var = (nir_variable *) entry->key;
527
exec_node_remove(&var->node);
532
/* Remove old output from the shaders outputs list */
533
hash_table_foreach(state.split_outputs, entry) {
534
nir_variable *var = (nir_variable *) entry->key;
535
exec_node_remove(&var->node);
540
_mesa_hash_table_destroy(state.split_inputs, NULL);
541
_mesa_hash_table_destroy(state.split_outputs, NULL);
543
nir_remove_dead_derefs(shader);