2 * Copyright © 2010 Intel Corporation
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
24 /** @file brw_fs_generator.cpp
26 * This file supports generating code from the FS LIR to the actual
27 * native instructions.
34 static enum brw_reg_file
35 brw_file_from_reg(fs_reg
*reg
)
39 return BRW_ARCHITECTURE_REGISTER_FILE
;
42 return BRW_GENERAL_REGISTER_FILE
;
44 return BRW_MESSAGE_REGISTER_FILE
;
46 return BRW_IMMEDIATE_VALUE
;
50 unreachable("not reached");
52 return BRW_ARCHITECTURE_REGISTER_FILE
;
56 brw_reg_from_fs_reg(const struct gen_device_info
*devinfo
, fs_inst
*inst
,
57 fs_reg
*reg
, bool compressed
)
59 struct brw_reg brw_reg
;
63 assert((reg
->nr
& ~BRW_MRF_COMPR4
) < BRW_MAX_MRF(devinfo
->gen
));
66 if (reg
->stride
== 0) {
67 brw_reg
= brw_vec1_reg(brw_file_from_reg(reg
), reg
->nr
, 0);
69 /* From the Haswell PRM:
71 * "VertStride must be used to cross GRF register boundaries. This
72 * rule implies that elements within a 'Width' cannot cross GRF
75 * The maximum width value that could satisfy this restriction is:
77 const unsigned reg_width
= REG_SIZE
/ (reg
->stride
* type_sz(reg
->type
));
79 /* Because the hardware can only split source regions at a whole
80 * multiple of width during decompression (i.e. vertically), clamp
81 * the value obtained above to the physical execution size of a
82 * single decompressed chunk of the instruction:
84 const unsigned phys_width
= compressed
? inst
->exec_size
/ 2 :
87 /* XXX - The equation above is strictly speaking not correct on
88 * hardware that supports unbalanced GRF writes -- On Gen9+
89 * each decompressed chunk of the instruction may have a
90 * different execution size when the number of components
91 * written to each destination GRF is not the same.
93 if (reg
->stride
> 4) {
94 assert(reg
!= &inst
->dst
);
95 assert(reg
->stride
* type_sz(reg
->type
) <= REG_SIZE
);
96 brw_reg
= brw_vecn_reg(1, brw_file_from_reg(reg
), reg
->nr
, 0);
97 brw_reg
= stride(brw_reg
, reg
->stride
, 1, 0);
99 const unsigned width
= MIN2(reg_width
, phys_width
);
100 brw_reg
= brw_vecn_reg(width
, brw_file_from_reg(reg
), reg
->nr
, 0);
101 brw_reg
= stride(brw_reg
, width
* reg
->stride
, width
, reg
->stride
);
104 if (devinfo
->gen
== 7 && !devinfo
->is_haswell
) {
105 /* From the IvyBridge PRM (EU Changes by Processor Generation, page 13):
106 * "Each DF (Double Float) operand uses an element size of 4 rather
107 * than 8 and all regioning parameters are twice what the values
108 * would be based on the true element size: ExecSize, Width,
109 * HorzStride, and VertStride. Each DF operand uses a pair of
110 * channels and all masking and swizzing should be adjusted
113 * From the IvyBridge PRM (Special Requirements for Handling Double
114 * Precision Data Types, page 71):
115 * "In Align1 mode, all regioning parameters like stride, execution
116 * size, and width must use the syntax of a pair of packed
117 * floats. The offsets for these data types must be 64-bit
118 * aligned. The execution size and regioning parameters are in terms
121 * Summarized: when handling DF-typed arguments, ExecSize,
122 * VertStride, and Width must be doubled.
124 * It applies to BayTrail too.
126 if (type_sz(reg
->type
) == 8) {
128 if (brw_reg
.vstride
> 0)
130 assert(brw_reg
.hstride
== BRW_HORIZONTAL_STRIDE_1
);
133 /* When converting from DF->F, we set the destination stride to 2
134 * because each d2f conversion implicitly writes 2 floats, being
135 * the first one the converted value. IVB/BYT actually writes two
136 * F components per SIMD channel, and every other component is
137 * filled with garbage.
139 if (reg
== &inst
->dst
&& get_exec_type_size(inst
) == 8 &&
140 type_sz(inst
->dst
.type
) < 8) {
141 assert(brw_reg
.hstride
> BRW_HORIZONTAL_STRIDE_1
);
147 brw_reg
= retype(brw_reg
, reg
->type
);
148 brw_reg
= byte_offset(brw_reg
, reg
->offset
);
149 brw_reg
.abs
= reg
->abs
;
150 brw_reg
.negate
= reg
->negate
;
155 assert(reg
->offset
== 0);
156 brw_reg
= reg
->as_brw_reg();
159 /* Probably unused. */
160 brw_reg
= brw_null_reg();
164 unreachable("not reached");
167 /* On HSW+, scalar DF sources can be accessed using the normal <0,1,0>
168 * region, but on IVB and BYT DF regions must be programmed in terms of
169 * floats. A <0,2,1> region accomplishes this.
171 if (devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
172 type_sz(reg
->type
) == 8 &&
173 brw_reg
.vstride
== BRW_VERTICAL_STRIDE_0
&&
174 brw_reg
.width
== BRW_WIDTH_1
&&
175 brw_reg
.hstride
== BRW_HORIZONTAL_STRIDE_0
) {
176 brw_reg
.width
= BRW_WIDTH_2
;
177 brw_reg
.hstride
= BRW_HORIZONTAL_STRIDE_1
;
183 fs_generator::fs_generator(const struct brw_compiler
*compiler
, void *log_data
,
185 struct brw_stage_prog_data
*prog_data
,
186 unsigned promoted_constants
,
187 bool runtime_check_aads_emit
,
188 gl_shader_stage stage
)
190 : compiler(compiler
), log_data(log_data
),
191 devinfo(compiler
->devinfo
),
192 prog_data(prog_data
),
193 promoted_constants(promoted_constants
),
194 runtime_check_aads_emit(runtime_check_aads_emit
), debug_flag(false),
195 stage(stage
), mem_ctx(mem_ctx
)
197 p
= rzalloc(mem_ctx
, struct brw_codegen
);
198 brw_init_codegen(devinfo
, p
, mem_ctx
);
200 /* In the FS code generator, we are very careful to ensure that we always
201 * set the right execution size so we don't need the EU code to "help" us
202 * by trying to infer it. Sometimes, it infers the wrong thing.
204 p
->automatic_exec_sizes
= false;
207 fs_generator::~fs_generator()
211 class ip_record
: public exec_node
{
213 DECLARE_RALLOC_CXX_OPERATORS(ip_record
)
224 fs_generator::patch_discard_jumps_to_fb_writes()
226 if (devinfo
->gen
< 6 || this->discard_halt_patches
.is_empty())
229 int scale
= brw_jump_scale(p
->devinfo
);
231 /* There is a somewhat strange undocumented requirement of using
232 * HALT, according to the simulator. If some channel has HALTed to
233 * a particular UIP, then by the end of the program, every channel
234 * must have HALTed to that UIP. Furthermore, the tracking is a
235 * stack, so you can't do the final halt of a UIP after starting
236 * halting to a new UIP.
238 * Symptoms of not emitting this instruction on actual hardware
239 * included GPU hangs and sparkly rendering on the piglit discard
242 brw_inst
*last_halt
= gen6_HALT(p
);
243 brw_inst_set_uip(p
->devinfo
, last_halt
, 1 * scale
);
244 brw_inst_set_jip(p
->devinfo
, last_halt
, 1 * scale
);
248 foreach_in_list(ip_record
, patch_ip
, &discard_halt_patches
) {
249 brw_inst
*patch
= &p
->store
[patch_ip
->ip
];
251 assert(brw_inst_opcode(p
->devinfo
, patch
) == BRW_OPCODE_HALT
);
252 /* HALT takes a half-instruction distance from the pre-incremented IP. */
253 brw_inst_set_uip(p
->devinfo
, patch
, (ip
- patch_ip
->ip
) * scale
);
256 this->discard_halt_patches
.make_empty();
261 fs_generator::generate_send(fs_inst
*inst
,
264 struct brw_reg ex_desc
,
265 struct brw_reg payload
,
266 struct brw_reg payload2
)
268 const bool dst_is_null
= dst
.file
== BRW_ARCHITECTURE_REGISTER_FILE
&&
269 dst
.nr
== BRW_ARF_NULL
;
270 const unsigned rlen
= dst_is_null
? 0 : inst
->size_written
/ REG_SIZE
;
272 uint32_t desc_imm
= inst
->desc
|
273 brw_message_desc(devinfo
, inst
->mlen
, rlen
, inst
->header_size
);
275 uint32_t ex_desc_imm
= brw_message_ex_desc(devinfo
, inst
->ex_mlen
);
277 if (ex_desc
.file
!= BRW_IMMEDIATE_VALUE
|| ex_desc
.ud
|| ex_desc_imm
) {
278 /* If we have any sort of extended descriptor, then we need SENDS. This
279 * also covers the dual-payload case because ex_mlen goes in ex_desc.
281 brw_send_indirect_split_message(p
, inst
->sfid
, dst
, payload
, payload2
,
282 desc
, desc_imm
, ex_desc
, ex_desc_imm
,
285 brw_inst_set_opcode(p
->devinfo
, brw_last_inst
, BRW_OPCODE_SENDSC
);
287 brw_send_indirect_message(p
, inst
->sfid
, dst
, payload
, desc
, desc_imm
,
290 brw_inst_set_opcode(p
->devinfo
, brw_last_inst
, BRW_OPCODE_SENDC
);
295 fs_generator::fire_fb_write(fs_inst
*inst
,
296 struct brw_reg payload
,
297 struct brw_reg implied_header
,
300 uint32_t msg_control
;
302 struct brw_wm_prog_data
*prog_data
= brw_wm_prog_data(this->prog_data
);
304 if (devinfo
->gen
< 6) {
305 brw_push_insn_state(p
);
306 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
307 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
308 brw_set_default_predicate_control(p
, BRW_PREDICATE_NONE
);
309 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
310 brw_MOV(p
, offset(retype(payload
, BRW_REGISTER_TYPE_UD
), 1),
311 offset(retype(implied_header
, BRW_REGISTER_TYPE_UD
), 1));
312 brw_pop_insn_state(p
);
315 if (inst
->opcode
== FS_OPCODE_REP_FB_WRITE
) {
316 assert(inst
->group
== 0 && inst
->exec_size
== 16);
317 msg_control
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE_REPLICATED
;
319 } else if (prog_data
->dual_src_blend
) {
320 assert(inst
->exec_size
== 8);
322 if (inst
->group
% 16 == 0)
323 msg_control
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_DUAL_SOURCE_SUBSPAN01
;
324 else if (inst
->group
% 16 == 8)
325 msg_control
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_DUAL_SOURCE_SUBSPAN23
;
327 unreachable("Invalid dual-source FB write instruction group");
330 assert(inst
->group
== 0 || (inst
->group
== 16 && inst
->exec_size
== 16));
332 if (inst
->exec_size
== 16)
333 msg_control
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE
;
334 else if (inst
->exec_size
== 8)
335 msg_control
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_SINGLE_SOURCE_SUBSPAN01
;
337 unreachable("Invalid FB write execution size");
340 /* We assume render targets start at 0, because headerless FB write
341 * messages set "Render Target Index" to 0. Using a different binding
342 * table index would make it impossible to use headerless messages.
344 const uint32_t surf_index
= inst
->target
;
346 brw_inst
*insn
= brw_fb_WRITE(p
,
348 retype(implied_header
, BRW_REGISTER_TYPE_UW
),
355 inst
->header_size
!= 0);
357 if (devinfo
->gen
>= 6)
358 brw_inst_set_rt_slot_group(devinfo
, insn
, inst
->group
/ 16);
362 fs_generator::generate_fb_write(fs_inst
*inst
, struct brw_reg payload
)
364 if (devinfo
->gen
< 8 && !devinfo
->is_haswell
) {
365 brw_set_default_predicate_control(p
, BRW_PREDICATE_NONE
);
368 const struct brw_reg implied_header
=
369 devinfo
->gen
< 6 ? payload
: brw_null_reg();
371 if (inst
->base_mrf
>= 0)
372 payload
= brw_message_reg(inst
->base_mrf
);
374 if (!runtime_check_aads_emit
) {
375 fire_fb_write(inst
, payload
, implied_header
, inst
->mlen
);
377 /* This can only happen in gen < 6 */
378 assert(devinfo
->gen
< 6);
380 struct brw_reg v1_null_ud
= vec1(retype(brw_null_reg(), BRW_REGISTER_TYPE_UD
));
382 /* Check runtime bit to detect if we have to send AA data or not */
383 brw_push_insn_state(p
);
384 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
385 brw_set_default_exec_size(p
, BRW_EXECUTE_1
);
388 retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_UD
),
390 brw_inst_set_cond_modifier(p
->devinfo
, brw_last_inst
, BRW_CONDITIONAL_NZ
);
392 int jmp
= brw_JMPI(p
, brw_imm_ud(0), BRW_PREDICATE_NORMAL
) - p
->store
;
393 brw_pop_insn_state(p
);
395 /* Don't send AA data */
396 fire_fb_write(inst
, offset(payload
, 1), implied_header
, inst
->mlen
-1);
398 brw_land_fwd_jump(p
, jmp
);
399 fire_fb_write(inst
, payload
, implied_header
, inst
->mlen
);
404 fs_generator::generate_fb_read(fs_inst
*inst
, struct brw_reg dst
,
405 struct brw_reg payload
)
407 assert(inst
->size_written
% REG_SIZE
== 0);
408 struct brw_wm_prog_data
*prog_data
= brw_wm_prog_data(this->prog_data
);
409 /* We assume that render targets start at binding table index 0. */
410 const unsigned surf_index
= inst
->target
;
412 gen9_fb_READ(p
, dst
, payload
, surf_index
,
413 inst
->header_size
, inst
->size_written
/ REG_SIZE
,
414 prog_data
->persample_dispatch
);
418 fs_generator::generate_mov_indirect(fs_inst
*inst
,
421 struct brw_reg indirect_byte_offset
)
423 assert(indirect_byte_offset
.type
== BRW_REGISTER_TYPE_UD
);
424 assert(indirect_byte_offset
.file
== BRW_GENERAL_REGISTER_FILE
);
425 assert(!reg
.abs
&& !reg
.negate
);
426 assert(reg
.type
== dst
.type
);
428 unsigned imm_byte_offset
= reg
.nr
* REG_SIZE
+ reg
.subnr
;
430 if (indirect_byte_offset
.file
== BRW_IMMEDIATE_VALUE
) {
431 imm_byte_offset
+= indirect_byte_offset
.ud
;
433 reg
.nr
= imm_byte_offset
/ REG_SIZE
;
434 reg
.subnr
= imm_byte_offset
% REG_SIZE
;
435 brw_MOV(p
, dst
, reg
);
437 /* Prior to Broadwell, there are only 8 address registers. */
438 assert(inst
->exec_size
<= 8 || devinfo
->gen
>= 8);
440 /* We use VxH indirect addressing, clobbering a0.0 through a0.7. */
441 struct brw_reg addr
= vec8(brw_address_reg(0));
443 /* The destination stride of an instruction (in bytes) must be greater
444 * than or equal to the size of the rest of the instruction. Since the
445 * address register is of type UW, we can't use a D-type instruction.
446 * In order to get around this, re retype to UW and use a stride.
448 indirect_byte_offset
=
449 retype(spread(indirect_byte_offset
, 2), BRW_REGISTER_TYPE_UW
);
451 /* There are a number of reasons why we don't use the base offset here.
452 * One reason is that the field is only 9 bits which means we can only
453 * use it to access the first 16 GRFs. Also, from the Haswell PRM
454 * section "Register Region Restrictions":
456 * "The lower bits of the AddressImmediate must not overflow to
457 * change the register address. The lower 5 bits of Address
458 * Immediate when added to lower 5 bits of address register gives
459 * the sub-register offset. The upper bits of Address Immediate
460 * when added to upper bits of address register gives the register
461 * address. Any overflow from sub-register offset is dropped."
463 * Since the indirect may cause us to cross a register boundary, this
464 * makes the base offset almost useless. We could try and do something
465 * clever where we use a actual base offset if base_offset % 32 == 0 but
466 * that would mean we were generating different code depending on the
467 * base offset. Instead, for the sake of consistency, we'll just do the
468 * add ourselves. This restriction is only listed in the Haswell PRM
469 * but empirical testing indicates that it applies on all older
470 * generations and is lifted on Broadwell.
472 * In the end, while base_offset is nice to look at in the generated
473 * code, using it saves us 0 instructions and would require quite a bit
474 * of case-by-case work. It's just not worth it.
476 brw_ADD(p
, addr
, indirect_byte_offset
, brw_imm_uw(imm_byte_offset
));
478 if (type_sz(reg
.type
) > 4 &&
479 ((devinfo
->gen
== 7 && !devinfo
->is_haswell
) ||
480 devinfo
->is_cherryview
|| gen_device_info_is_9lp(devinfo
) ||
481 !devinfo
->has_64bit_types
)) {
482 /* IVB has an issue (which we found empirically) where it reads two
483 * address register components per channel for indirectly addressed
486 * From the Cherryview PRM Vol 7. "Register Region Restrictions":
488 * "When source or destination datatype is 64b or operation is
489 * integer DWord multiply, indirect addressing must not be used."
491 * To work around both of these, we do two integer MOVs insead of one
492 * 64-bit MOV. Because no double value should ever cross a register
493 * boundary, it's safe to use the immediate offset in the indirect
494 * here to handle adding 4 bytes to the offset and avoid the extra
495 * ADD to the register file.
497 brw_MOV(p
, subscript(dst
, BRW_REGISTER_TYPE_D
, 0),
498 retype(brw_VxH_indirect(0, 0), BRW_REGISTER_TYPE_D
));
499 brw_MOV(p
, subscript(dst
, BRW_REGISTER_TYPE_D
, 1),
500 retype(brw_VxH_indirect(0, 4), BRW_REGISTER_TYPE_D
));
502 struct brw_reg ind_src
= brw_VxH_indirect(0, 0);
504 brw_inst
*mov
= brw_MOV(p
, dst
, retype(ind_src
, reg
.type
));
506 if (devinfo
->gen
== 6 && dst
.file
== BRW_MESSAGE_REGISTER_FILE
&&
507 !inst
->get_next()->is_tail_sentinel() &&
508 ((fs_inst
*)inst
->get_next())->mlen
> 0) {
509 /* From the Sandybridge PRM:
511 * "[Errata: DevSNB(SNB)] If MRF register is updated by any
512 * instruction that “indexed/indirect” source AND is followed
513 * by a send, the instruction requires a “Switch”. This is to
514 * avoid race condition where send may dispatch before MRF is
517 brw_inst_set_thread_control(devinfo
, mov
, BRW_THREAD_SWITCH
);
524 fs_generator::generate_shuffle(fs_inst
*inst
,
529 /* Ivy bridge has some strange behavior that makes this a real pain to
530 * implement for 64-bit values so we just don't bother.
532 assert(devinfo
->gen
>= 8 || devinfo
->is_haswell
|| type_sz(src
.type
) <= 4);
534 /* Because we're using the address register, we're limited to 8-wide
535 * execution on gen7. On gen8, we're limited to 16-wide by the address
536 * register file and 8-wide for 64-bit types. We could try and make this
537 * instruction splittable higher up in the compiler but that gets weird
538 * because it reads all of the channels regardless of execution size. It's
539 * easier just to split it here.
541 const unsigned lower_width
=
542 (devinfo
->gen
<= 7 || type_sz(src
.type
) > 4) ?
543 8 : MIN2(16, inst
->exec_size
);
545 brw_set_default_exec_size(p
, cvt(lower_width
) - 1);
546 for (unsigned group
= 0; group
< inst
->exec_size
; group
+= lower_width
) {
547 brw_set_default_group(p
, group
);
549 if ((src
.vstride
== 0 && src
.hstride
== 0) ||
550 idx
.file
== BRW_IMMEDIATE_VALUE
) {
551 /* Trivial, the source is already uniform or the index is a constant.
552 * We will typically not get here if the optimizer is doing its job,
553 * but asserting would be mean.
555 const unsigned i
= idx
.file
== BRW_IMMEDIATE_VALUE
? idx
.ud
: 0;
556 brw_MOV(p
, suboffset(dst
, group
), stride(suboffset(src
, i
), 0, 1, 0));
558 /* We use VxH indirect addressing, clobbering a0.0 through a0.7. */
559 struct brw_reg addr
= vec8(brw_address_reg(0));
561 struct brw_reg group_idx
= suboffset(idx
, group
);
563 if (lower_width
== 8 && group_idx
.width
== BRW_WIDTH_16
) {
564 /* Things get grumpy if the register is too wide. */
569 assert(type_sz(group_idx
.type
) <= 4);
570 if (type_sz(group_idx
.type
) == 4) {
571 /* The destination stride of an instruction (in bytes) must be
572 * greater than or equal to the size of the rest of the
573 * instruction. Since the address register is of type UW, we
574 * can't use a D-type instruction. In order to get around this,
575 * re retype to UW and use a stride.
577 group_idx
= retype(spread(group_idx
, 2), BRW_REGISTER_TYPE_W
);
580 /* Take into account the component size and horizontal stride. */
581 assert(src
.vstride
== src
.hstride
+ src
.width
);
582 brw_SHL(p
, addr
, group_idx
,
583 brw_imm_uw(_mesa_logbase2(type_sz(src
.type
)) +
586 /* Add on the register start offset */
587 brw_ADD(p
, addr
, addr
, brw_imm_uw(src
.nr
* REG_SIZE
+ src
.subnr
));
589 if (type_sz(src
.type
) > 4 &&
590 ((devinfo
->gen
== 7 && !devinfo
->is_haswell
) ||
591 devinfo
->is_cherryview
|| gen_device_info_is_9lp(devinfo
))) {
592 /* IVB has an issue (which we found empirically) where it reads
593 * two address register components per channel for indirectly
594 * addressed 64-bit sources.
596 * From the Cherryview PRM Vol 7. "Register Region Restrictions":
598 * "When source or destination datatype is 64b or operation is
599 * integer DWord multiply, indirect addressing must not be
602 * To work around both of these, we do two integer MOVs insead of
603 * one 64-bit MOV. Because no double value should ever cross a
604 * register boundary, it's safe to use the immediate offset in the
605 * indirect here to handle adding 4 bytes to the offset and avoid
606 * the extra ADD to the register file.
608 struct brw_reg gdst
= suboffset(dst
, group
);
609 struct brw_reg dst_d
= retype(spread(gdst
, 2),
610 BRW_REGISTER_TYPE_D
);
612 retype(brw_VxH_indirect(0, 0), BRW_REGISTER_TYPE_D
));
613 brw_MOV(p
, byte_offset(dst_d
, 4),
614 retype(brw_VxH_indirect(0, 4), BRW_REGISTER_TYPE_D
));
616 brw_MOV(p
, suboffset(dst
, group
),
617 retype(brw_VxH_indirect(0, 0), src
.type
));
624 fs_generator::generate_quad_swizzle(const fs_inst
*inst
,
625 struct brw_reg dst
, struct brw_reg src
,
628 /* Requires a quad. */
629 assert(inst
->exec_size
>= 4);
631 if (src
.file
== BRW_IMMEDIATE_VALUE
||
632 has_scalar_region(src
)) {
633 /* The value is uniform across all channels */
634 brw_MOV(p
, dst
, src
);
636 } else if (devinfo
->gen
< 11 && type_sz(src
.type
) == 4) {
637 /* This only works on 8-wide 32-bit values */
638 assert(inst
->exec_size
== 8);
639 assert(src
.hstride
== BRW_HORIZONTAL_STRIDE_1
);
640 assert(src
.vstride
== src
.width
+ 1);
641 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
642 struct brw_reg swiz_src
= stride(src
, 4, 4, 1);
643 swiz_src
.swizzle
= swiz
;
644 brw_MOV(p
, dst
, swiz_src
);
647 assert(src
.hstride
== BRW_HORIZONTAL_STRIDE_1
);
648 assert(src
.vstride
== src
.width
+ 1);
649 const struct brw_reg src_0
= suboffset(src
, BRW_GET_SWZ(swiz
, 0));
652 case BRW_SWIZZLE_XXXX
:
653 case BRW_SWIZZLE_YYYY
:
654 case BRW_SWIZZLE_ZZZZ
:
655 case BRW_SWIZZLE_WWWW
:
656 brw_MOV(p
, dst
, stride(src_0
, 4, 4, 0));
659 case BRW_SWIZZLE_XXZZ
:
660 case BRW_SWIZZLE_YYWW
:
661 brw_MOV(p
, dst
, stride(src_0
, 2, 2, 0));
664 case BRW_SWIZZLE_XYXY
:
665 case BRW_SWIZZLE_ZWZW
:
666 assert(inst
->exec_size
== 4);
667 brw_MOV(p
, dst
, stride(src_0
, 0, 2, 1));
671 assert(inst
->force_writemask_all
);
672 brw_set_default_exec_size(p
, cvt(inst
->exec_size
/ 4) - 1);
674 for (unsigned c
= 0; c
< 4; c
++) {
675 brw_inst
*insn
= brw_MOV(
676 p
, stride(suboffset(dst
, c
),
677 4 * inst
->dst
.stride
, 1, 4 * inst
->dst
.stride
),
678 stride(suboffset(src
, BRW_GET_SWZ(swiz
, c
)), 4, 1, 0));
680 brw_inst_set_no_dd_clear(devinfo
, insn
, c
< 3);
681 brw_inst_set_no_dd_check(devinfo
, insn
, c
> 0);
690 fs_generator::generate_urb_read(fs_inst
*inst
,
692 struct brw_reg header
)
694 assert(inst
->size_written
% REG_SIZE
== 0);
695 assert(header
.file
== BRW_GENERAL_REGISTER_FILE
);
696 assert(header
.type
== BRW_REGISTER_TYPE_UD
);
698 brw_inst
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
699 brw_set_dest(p
, send
, retype(dst
, BRW_REGISTER_TYPE_UD
));
700 brw_set_src0(p
, send
, header
);
701 brw_set_src1(p
, send
, brw_imm_ud(0u));
703 brw_inst_set_sfid(p
->devinfo
, send
, BRW_SFID_URB
);
704 brw_inst_set_urb_opcode(p
->devinfo
, send
, GEN8_URB_OPCODE_SIMD8_READ
);
706 if (inst
->opcode
== SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
)
707 brw_inst_set_urb_per_slot_offset(p
->devinfo
, send
, true);
709 brw_inst_set_mlen(p
->devinfo
, send
, inst
->mlen
);
710 brw_inst_set_rlen(p
->devinfo
, send
, inst
->size_written
/ REG_SIZE
);
711 brw_inst_set_header_present(p
->devinfo
, send
, true);
712 brw_inst_set_urb_global_offset(p
->devinfo
, send
, inst
->offset
);
716 fs_generator::generate_urb_write(fs_inst
*inst
, struct brw_reg payload
)
720 /* WaClearTDRRegBeforeEOTForNonPS.
722 * WA: Clear tdr register before send EOT in all non-PS shader kernels
724 * mov(8) tdr0:ud 0x0:ud {NoMask}"
726 if (inst
->eot
&& p
->devinfo
->gen
== 10) {
727 brw_push_insn_state(p
);
728 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
729 brw_MOV(p
, brw_tdr_reg(), brw_imm_uw(0));
730 brw_pop_insn_state(p
);
733 insn
= brw_next_insn(p
, BRW_OPCODE_SEND
);
735 brw_set_dest(p
, insn
, brw_null_reg());
736 brw_set_src0(p
, insn
, payload
);
737 brw_set_src1(p
, insn
, brw_imm_ud(0u));
739 brw_inst_set_sfid(p
->devinfo
, insn
, BRW_SFID_URB
);
740 brw_inst_set_urb_opcode(p
->devinfo
, insn
, GEN8_URB_OPCODE_SIMD8_WRITE
);
742 if (inst
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
||
743 inst
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
)
744 brw_inst_set_urb_per_slot_offset(p
->devinfo
, insn
, true);
746 if (inst
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
||
747 inst
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
)
748 brw_inst_set_urb_channel_mask_present(p
->devinfo
, insn
, true);
750 brw_inst_set_mlen(p
->devinfo
, insn
, inst
->mlen
);
751 brw_inst_set_rlen(p
->devinfo
, insn
, 0);
752 brw_inst_set_eot(p
->devinfo
, insn
, inst
->eot
);
753 brw_inst_set_header_present(p
->devinfo
, insn
, true);
754 brw_inst_set_urb_global_offset(p
->devinfo
, insn
, inst
->offset
);
758 fs_generator::generate_cs_terminate(fs_inst
*inst
, struct brw_reg payload
)
760 struct brw_inst
*insn
;
762 insn
= brw_next_insn(p
, BRW_OPCODE_SEND
);
764 brw_set_dest(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_UW
));
765 brw_set_src0(p
, insn
, retype(payload
, BRW_REGISTER_TYPE_UW
));
766 brw_set_src1(p
, insn
, brw_imm_ud(0u));
768 /* Terminate a compute shader by sending a message to the thread spawner.
770 brw_inst_set_sfid(devinfo
, insn
, BRW_SFID_THREAD_SPAWNER
);
771 brw_inst_set_mlen(devinfo
, insn
, 1);
772 brw_inst_set_rlen(devinfo
, insn
, 0);
773 brw_inst_set_eot(devinfo
, insn
, inst
->eot
);
774 brw_inst_set_header_present(devinfo
, insn
, false);
776 brw_inst_set_ts_opcode(devinfo
, insn
, 0); /* Dereference resource */
777 brw_inst_set_ts_request_type(devinfo
, insn
, 0); /* Root thread */
779 /* Note that even though the thread has a URB resource associated with it,
780 * we set the "do not dereference URB" bit, because the URB resource is
781 * managed by the fixed-function unit, so it will free it automatically.
783 brw_inst_set_ts_resource_select(devinfo
, insn
, 1); /* Do not dereference URB */
785 brw_inst_set_mask_control(devinfo
, insn
, BRW_MASK_DISABLE
);
789 fs_generator::generate_barrier(fs_inst
*, struct brw_reg src
)
796 fs_generator::generate_linterp(fs_inst
*inst
,
797 struct brw_reg dst
, struct brw_reg
*src
)
801 * -----------------------------------
802 * | src1+0 | src1+1 | src1+2 | src1+3 |
803 * |-----------------------------------|
804 * |(x0, x1)|(y0, y1)|(x2, x3)|(y2, y3)|
805 * -----------------------------------
807 * but for the LINE/MAC pair, the LINE reads Xs and the MAC reads Ys:
809 * -----------------------------------
810 * | src1+0 | src1+1 | src1+2 | src1+3 |
811 * |-----------------------------------|
812 * |(x0, x1)|(y0, y1)| | | in SIMD8
813 * |-----------------------------------|
814 * |(x0, x1)|(x2, x3)|(y0, y1)|(y2, y3)| in SIMD16
815 * -----------------------------------
817 * See also: emit_interpolation_setup_gen4().
819 struct brw_reg delta_x
= src
[0];
820 struct brw_reg delta_y
= offset(src
[0], inst
->exec_size
/ 8);
821 struct brw_reg interp
= src
[1];
824 if (devinfo
->gen
>= 11) {
825 struct brw_reg acc
= retype(brw_acc_reg(8), BRW_REGISTER_TYPE_NF
);
826 struct brw_reg dwP
= suboffset(interp
, 0);
827 struct brw_reg dwQ
= suboffset(interp
, 1);
828 struct brw_reg dwR
= suboffset(interp
, 3);
830 brw_push_insn_state(p
);
831 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
833 if (inst
->exec_size
== 8) {
834 i
[0] = brw_MAD(p
, acc
, dwR
, offset(delta_x
, 0), dwP
);
835 i
[1] = brw_MAD(p
, offset(dst
, 0), acc
, offset(delta_y
, 0), dwQ
);
837 brw_inst_set_cond_modifier(p
->devinfo
, i
[1], inst
->conditional_mod
);
839 /* brw_set_default_saturate() is called before emitting instructions,
840 * so the saturate bit is set in each instruction, so we need to unset
841 * it on the first instruction of each pair.
843 brw_inst_set_saturate(p
->devinfo
, i
[0], false);
845 brw_set_default_group(p
, inst
->group
);
846 i
[0] = brw_MAD(p
, acc
, dwR
, offset(delta_x
, 0), dwP
);
847 i
[1] = brw_MAD(p
, offset(dst
, 0), acc
, offset(delta_x
, 1), dwQ
);
849 brw_set_default_group(p
, inst
->group
+ 8);
850 i
[2] = brw_MAD(p
, acc
, dwR
, offset(delta_y
, 0), dwP
);
851 i
[3] = brw_MAD(p
, offset(dst
, 1), acc
, offset(delta_y
, 1), dwQ
);
853 brw_inst_set_cond_modifier(p
->devinfo
, i
[1], inst
->conditional_mod
);
854 brw_inst_set_cond_modifier(p
->devinfo
, i
[3], inst
->conditional_mod
);
856 /* brw_set_default_saturate() is called before emitting instructions,
857 * so the saturate bit is set in each instruction, so we need to unset
858 * it on the first instruction of each pair.
860 brw_inst_set_saturate(p
->devinfo
, i
[0], false);
861 brw_inst_set_saturate(p
->devinfo
, i
[2], false);
864 brw_pop_insn_state(p
);
867 } else if (devinfo
->has_pln
) {
868 if (devinfo
->gen
<= 6 && (delta_x
.nr
& 1) != 0) {
869 /* From the Sandy Bridge PRM Vol. 4, Pt. 2, Section 8.3.53, "Plane":
871 * "[DevSNB]:<src1> must be even register aligned.
873 * This restriction is lifted on Ivy Bridge.
875 * This means that we need to split PLN into LINE+MAC on-the-fly.
876 * Unfortunately, the inputs are laid out for PLN and not LINE+MAC so
877 * we have to split into SIMD8 pieces. For gen4 (!has_pln), the
878 * coordinate registers are laid out differently so we leave it as a
879 * SIMD16 instruction.
881 assert(inst
->exec_size
== 8 || inst
->exec_size
== 16);
882 assert(inst
->group
% 16 == 0);
884 brw_push_insn_state(p
);
885 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
887 /* Thanks to two accumulators, we can emit all the LINEs and then all
888 * the MACs. This improves parallelism a bit.
890 for (unsigned g
= 0; g
< inst
->exec_size
/ 8; g
++) {
891 brw_inst
*line
= brw_LINE(p
, brw_null_reg(), interp
,
892 offset(delta_x
, g
* 2));
893 brw_inst_set_group(devinfo
, line
, inst
->group
+ g
* 8);
895 /* LINE writes the accumulator automatically on gen4-5. On Sandy
896 * Bridge and later, we have to explicitly enable it.
898 if (devinfo
->gen
>= 6)
899 brw_inst_set_acc_wr_control(p
->devinfo
, line
, true);
901 /* brw_set_default_saturate() is called before emitting
902 * instructions, so the saturate bit is set in each instruction,
903 * so we need to unset it on the LINE instructions.
905 brw_inst_set_saturate(p
->devinfo
, line
, false);
908 for (unsigned g
= 0; g
< inst
->exec_size
/ 8; g
++) {
909 brw_inst
*mac
= brw_MAC(p
, offset(dst
, g
), suboffset(interp
, 1),
910 offset(delta_x
, g
* 2 + 1));
911 brw_inst_set_group(devinfo
, mac
, inst
->group
+ g
* 8);
912 brw_inst_set_cond_modifier(p
->devinfo
, mac
, inst
->conditional_mod
);
915 brw_pop_insn_state(p
);
919 brw_PLN(p
, dst
, interp
, delta_x
);
924 i
[0] = brw_LINE(p
, brw_null_reg(), interp
, delta_x
);
925 i
[1] = brw_MAC(p
, dst
, suboffset(interp
, 1), delta_y
);
927 brw_inst_set_cond_modifier(p
->devinfo
, i
[1], inst
->conditional_mod
);
929 /* brw_set_default_saturate() is called before emitting instructions, so
930 * the saturate bit is set in each instruction, so we need to unset it on
931 * the first instruction.
933 brw_inst_set_saturate(p
->devinfo
, i
[0], false);
940 fs_generator::generate_get_buffer_size(fs_inst
*inst
,
943 struct brw_reg surf_index
)
945 assert(devinfo
->gen
>= 7);
946 assert(surf_index
.file
== BRW_IMMEDIATE_VALUE
);
951 switch (inst
->exec_size
) {
953 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD8
;
956 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
959 unreachable("Invalid width for texture instruction");
962 if (simd_mode
== BRW_SAMPLER_SIMD_MODE_SIMD16
) {
968 retype(dst
, BRW_REGISTER_TYPE_UW
),
973 GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO
,
974 rlen
, /* response length */
976 inst
->header_size
> 0,
978 BRW_SAMPLER_RETURN_FORMAT_SINT32
);
982 fs_generator::generate_tex(fs_inst
*inst
, struct brw_reg dst
,
983 struct brw_reg surface_index
,
984 struct brw_reg sampler_index
)
986 assert(devinfo
->gen
< 7);
987 assert(inst
->size_written
% REG_SIZE
== 0);
990 uint32_t return_format
;
992 /* Sampler EOT message of less than the dispatch width would kill the
993 * thread prematurely.
995 assert(!inst
->eot
|| inst
->exec_size
== dispatch_width
);
998 case BRW_REGISTER_TYPE_D
:
999 return_format
= BRW_SAMPLER_RETURN_FORMAT_SINT32
;
1001 case BRW_REGISTER_TYPE_UD
:
1002 return_format
= BRW_SAMPLER_RETURN_FORMAT_UINT32
;
1005 return_format
= BRW_SAMPLER_RETURN_FORMAT_FLOAT32
;
1009 /* Stomp the resinfo output type to UINT32. On gens 4-5, the output type
1010 * is set as part of the message descriptor. On gen4, the PRM seems to
1011 * allow UINT32 and FLOAT32 (i965 PRM, Vol. 4 Section 4.8.1.1), but on
1012 * later gens UINT32 is required. Once you hit Sandy Bridge, the bit is
1013 * gone from the message descriptor entirely and you just get UINT32 all
1014 * the time regasrdless. Since we can really only do non-UINT32 on gen4,
1015 * just stomp it to UINT32 all the time.
1017 if (inst
->opcode
== SHADER_OPCODE_TXS
)
1018 return_format
= BRW_SAMPLER_RETURN_FORMAT_UINT32
;
1020 switch (inst
->exec_size
) {
1022 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD8
;
1025 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1028 unreachable("Invalid width for texture instruction");
1031 if (devinfo
->gen
>= 5) {
1032 switch (inst
->opcode
) {
1033 case SHADER_OPCODE_TEX
:
1034 if (inst
->shadow_compare
) {
1035 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_COMPARE
;
1037 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE
;
1041 if (inst
->shadow_compare
) {
1042 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS_COMPARE
;
1044 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS
;
1047 case SHADER_OPCODE_TXL
:
1048 if (inst
->shadow_compare
) {
1049 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE
;
1051 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LOD
;
1054 case SHADER_OPCODE_TXS
:
1055 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO
;
1057 case SHADER_OPCODE_TXD
:
1058 assert(!inst
->shadow_compare
);
1059 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS
;
1061 case SHADER_OPCODE_TXF
:
1062 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
1064 case SHADER_OPCODE_TXF_CMS
:
1065 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
1067 case SHADER_OPCODE_LOD
:
1068 msg_type
= GEN5_SAMPLER_MESSAGE_LOD
;
1070 case SHADER_OPCODE_TG4
:
1071 assert(devinfo
->gen
== 6);
1072 assert(!inst
->shadow_compare
);
1073 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4
;
1075 case SHADER_OPCODE_SAMPLEINFO
:
1076 msg_type
= GEN6_SAMPLER_MESSAGE_SAMPLE_SAMPLEINFO
;
1079 unreachable("not reached");
1082 switch (inst
->opcode
) {
1083 case SHADER_OPCODE_TEX
:
1084 /* Note that G45 and older determines shadow compare and dispatch width
1085 * from message length for most messages.
1087 if (inst
->exec_size
== 8) {
1088 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE
;
1089 if (inst
->shadow_compare
) {
1090 assert(inst
->mlen
== 6);
1092 assert(inst
->mlen
<= 4);
1095 if (inst
->shadow_compare
) {
1096 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_COMPARE
;
1097 assert(inst
->mlen
== 9);
1099 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE
;
1100 assert(inst
->mlen
<= 7 && inst
->mlen
% 2 == 1);
1105 if (inst
->shadow_compare
) {
1106 assert(inst
->exec_size
== 8);
1107 assert(inst
->mlen
== 6);
1108 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_BIAS_COMPARE
;
1110 assert(inst
->mlen
== 9);
1111 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_BIAS
;
1112 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1115 case SHADER_OPCODE_TXL
:
1116 if (inst
->shadow_compare
) {
1117 assert(inst
->exec_size
== 8);
1118 assert(inst
->mlen
== 6);
1119 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_LOD_COMPARE
;
1121 assert(inst
->mlen
== 9);
1122 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_LOD
;
1123 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1126 case SHADER_OPCODE_TXD
:
1127 /* There is no sample_d_c message; comparisons are done manually */
1128 assert(inst
->exec_size
== 8);
1129 assert(inst
->mlen
== 7 || inst
->mlen
== 10);
1130 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_GRADIENTS
;
1132 case SHADER_OPCODE_TXF
:
1133 assert(inst
->mlen
<= 9 && inst
->mlen
% 2 == 1);
1134 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_LD
;
1135 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1137 case SHADER_OPCODE_TXS
:
1138 assert(inst
->mlen
== 3);
1139 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_RESINFO
;
1140 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1143 unreachable("not reached");
1146 assert(msg_type
!= -1);
1148 if (simd_mode
== BRW_SAMPLER_SIMD_MODE_SIMD16
) {
1152 assert(sampler_index
.type
== BRW_REGISTER_TYPE_UD
);
1154 /* Load the message header if present. If there's a texture offset,
1155 * we need to set it up explicitly and load the offset bitfield.
1156 * Otherwise, we can use an implied move from g0 to the first message reg.
1158 struct brw_reg src
= brw_null_reg();
1159 if (inst
->header_size
!= 0) {
1160 if (devinfo
->gen
< 6 && !inst
->offset
) {
1161 /* Set up an implied move from g0 to the MRF. */
1162 src
= retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW
);
1164 assert(inst
->base_mrf
!= -1);
1165 struct brw_reg header_reg
= brw_message_reg(inst
->base_mrf
);
1167 brw_push_insn_state(p
);
1168 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
1169 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1170 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
1171 /* Explicitly set up the message header by copying g0 to the MRF. */
1172 brw_MOV(p
, header_reg
, brw_vec8_grf(0, 0));
1174 brw_set_default_exec_size(p
, BRW_EXECUTE_1
);
1176 /* Set the offset bits in DWord 2. */
1177 brw_MOV(p
, get_element_ud(header_reg
, 2),
1178 brw_imm_ud(inst
->offset
));
1181 brw_pop_insn_state(p
);
1185 uint32_t base_binding_table_index
;
1186 switch (inst
->opcode
) {
1187 case SHADER_OPCODE_TG4
:
1188 base_binding_table_index
= prog_data
->binding_table
.gather_texture_start
;
1191 base_binding_table_index
= prog_data
->binding_table
.texture_start
;
1195 assert(surface_index
.file
== BRW_IMMEDIATE_VALUE
);
1196 assert(sampler_index
.file
== BRW_IMMEDIATE_VALUE
);
1199 retype(dst
, BRW_REGISTER_TYPE_UW
),
1202 surface_index
.ud
+ base_binding_table_index
,
1203 sampler_index
.ud
% 16,
1205 inst
->size_written
/ REG_SIZE
,
1207 inst
->header_size
!= 0,
1213 /* For OPCODE_DDX and OPCODE_DDY, per channel of output we've got input
1216 * arg0: ss0.tl ss0.tr ss0.bl ss0.br ss1.tl ss1.tr ss1.bl ss1.br
1218 * Ideally, we want to produce:
1221 * dst: (ss0.tr - ss0.tl) (ss0.tl - ss0.bl)
1222 * (ss0.tr - ss0.tl) (ss0.tr - ss0.br)
1223 * (ss0.br - ss0.bl) (ss0.tl - ss0.bl)
1224 * (ss0.br - ss0.bl) (ss0.tr - ss0.br)
1225 * (ss1.tr - ss1.tl) (ss1.tl - ss1.bl)
1226 * (ss1.tr - ss1.tl) (ss1.tr - ss1.br)
1227 * (ss1.br - ss1.bl) (ss1.tl - ss1.bl)
1228 * (ss1.br - ss1.bl) (ss1.tr - ss1.br)
1230 * and add another set of two more subspans if in 16-pixel dispatch mode.
1232 * For DDX, it ends up being easy: width = 2, horiz=0 gets us the same result
1233 * for each pair, and vertstride = 2 jumps us 2 elements after processing a
1234 * pair. But the ideal approximation may impose a huge performance cost on
1235 * sample_d. On at least Haswell, sample_d instruction does some
1236 * optimizations if the same LOD is used for all pixels in the subspan.
1238 * For DDY, we need to use ALIGN16 mode since it's capable of doing the
1239 * appropriate swizzling.
1242 fs_generator::generate_ddx(const fs_inst
*inst
,
1243 struct brw_reg dst
, struct brw_reg src
)
1245 unsigned vstride
, width
;
1247 if (inst
->opcode
== FS_OPCODE_DDX_FINE
) {
1248 /* produce accurate derivatives */
1249 vstride
= BRW_VERTICAL_STRIDE_2
;
1250 width
= BRW_WIDTH_2
;
1252 /* replicate the derivative at the top-left pixel to other pixels */
1253 vstride
= BRW_VERTICAL_STRIDE_4
;
1254 width
= BRW_WIDTH_4
;
1257 struct brw_reg src0
= byte_offset(src
, type_sz(src
.type
));;
1258 struct brw_reg src1
= src
;
1260 src0
.vstride
= vstride
;
1262 src0
.hstride
= BRW_HORIZONTAL_STRIDE_0
;
1263 src1
.vstride
= vstride
;
1265 src1
.hstride
= BRW_HORIZONTAL_STRIDE_0
;
1267 brw_ADD(p
, dst
, src0
, negate(src1
));
1270 /* The negate_value boolean is used to negate the derivative computation for
1271 * FBOs, since they place the origin at the upper left instead of the lower
1275 fs_generator::generate_ddy(const fs_inst
*inst
,
1276 struct brw_reg dst
, struct brw_reg src
)
1278 const uint32_t type_size
= type_sz(src
.type
);
1280 if (inst
->opcode
== FS_OPCODE_DDY_FINE
) {
1281 /* produce accurate derivatives.
1283 * From the Broadwell PRM, Volume 7 (3D-Media-GPGPU)
1284 * "Register Region Restrictions", Section "1. Special Restrictions":
1286 * "In Align16 mode, the channel selects and channel enables apply to
1287 * a pair of half-floats, because these parameters are defined for
1288 * DWord elements ONLY. This is applicable when both source and
1289 * destination are half-floats."
1291 * So for half-float operations we use the Gen11+ Align1 path. CHV
1292 * inherits its FP16 hardware from SKL, so it is not affected.
1294 if (devinfo
->gen
>= 11 ||
1295 (devinfo
->is_broadwell
&& src
.type
== BRW_REGISTER_TYPE_HF
)) {
1296 src
= stride(src
, 0, 2, 1);
1297 struct brw_reg src_0
= byte_offset(src
, 0 * type_size
);
1298 struct brw_reg src_2
= byte_offset(src
, 2 * type_size
);
1299 struct brw_reg src_4
= byte_offset(src
, 4 * type_size
);
1300 struct brw_reg src_6
= byte_offset(src
, 6 * type_size
);
1301 struct brw_reg src_8
= byte_offset(src
, 8 * type_size
);
1302 struct brw_reg src_10
= byte_offset(src
, 10 * type_size
);
1303 struct brw_reg src_12
= byte_offset(src
, 12 * type_size
);
1304 struct brw_reg src_14
= byte_offset(src
, 14 * type_size
);
1306 struct brw_reg dst_0
= byte_offset(dst
, 0 * type_size
);
1307 struct brw_reg dst_4
= byte_offset(dst
, 4 * type_size
);
1308 struct brw_reg dst_8
= byte_offset(dst
, 8 * type_size
);
1309 struct brw_reg dst_12
= byte_offset(dst
, 12 * type_size
);
1311 brw_push_insn_state(p
);
1312 brw_set_default_exec_size(p
, BRW_EXECUTE_4
);
1314 brw_ADD(p
, dst_0
, negate(src_0
), src_2
);
1315 brw_ADD(p
, dst_4
, negate(src_4
), src_6
);
1317 if (inst
->exec_size
== 16) {
1318 brw_ADD(p
, dst_8
, negate(src_8
), src_10
);
1319 brw_ADD(p
, dst_12
, negate(src_12
), src_14
);
1322 brw_pop_insn_state(p
);
1324 struct brw_reg src0
= stride(src
, 4, 4, 1);
1325 struct brw_reg src1
= stride(src
, 4, 4, 1);
1326 src0
.swizzle
= BRW_SWIZZLE_XYXY
;
1327 src1
.swizzle
= BRW_SWIZZLE_ZWZW
;
1329 brw_push_insn_state(p
);
1330 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1331 brw_ADD(p
, dst
, negate(src0
), src1
);
1332 brw_pop_insn_state(p
);
1335 /* replicate the derivative at the top-left pixel to other pixels */
1336 struct brw_reg src0
= byte_offset(stride(src
, 4, 4, 0), 0 * type_size
);
1337 struct brw_reg src1
= byte_offset(stride(src
, 4, 4, 0), 2 * type_size
);
1339 brw_ADD(p
, dst
, negate(src0
), src1
);
1344 fs_generator::generate_discard_jump(fs_inst
*)
1346 assert(devinfo
->gen
>= 6);
1348 /* This HALT will be patched up at FB write time to point UIP at the end of
1349 * the program, and at brw_uip_jip() JIP will be set to the end of the
1350 * current block (or the program).
1352 this->discard_halt_patches
.push_tail(new(mem_ctx
) ip_record(p
->nr_insn
));
1357 fs_generator::generate_scratch_write(fs_inst
*inst
, struct brw_reg src
)
1359 /* The 32-wide messages only respect the first 16-wide half of the channel
1360 * enable signals which are replicated identically for the second group of
1361 * 16 channels, so we cannot use them unless the write is marked
1362 * force_writemask_all.
1364 const unsigned lower_size
= inst
->force_writemask_all
? inst
->exec_size
:
1365 MIN2(16, inst
->exec_size
);
1366 const unsigned block_size
= 4 * lower_size
/ REG_SIZE
;
1367 assert(inst
->mlen
!= 0);
1369 brw_push_insn_state(p
);
1370 brw_set_default_exec_size(p
, cvt(lower_size
) - 1);
1371 brw_set_default_compression(p
, lower_size
> 8);
1373 for (unsigned i
= 0; i
< inst
->exec_size
/ lower_size
; i
++) {
1374 brw_set_default_group(p
, inst
->group
+ lower_size
* i
);
1376 brw_MOV(p
, brw_uvec_mrf(lower_size
, inst
->base_mrf
+ 1, 0),
1377 retype(offset(src
, block_size
* i
), BRW_REGISTER_TYPE_UD
));
1379 brw_oword_block_write_scratch(p
, brw_message_reg(inst
->base_mrf
),
1381 inst
->offset
+ block_size
* REG_SIZE
* i
);
1384 brw_pop_insn_state(p
);
1388 fs_generator::generate_scratch_read(fs_inst
*inst
, struct brw_reg dst
)
1390 assert(inst
->exec_size
<= 16 || inst
->force_writemask_all
);
1391 assert(inst
->mlen
!= 0);
1393 brw_oword_block_read_scratch(p
, dst
, brw_message_reg(inst
->base_mrf
),
1394 inst
->exec_size
/ 8, inst
->offset
);
1398 fs_generator::generate_scratch_read_gen7(fs_inst
*inst
, struct brw_reg dst
)
1400 assert(inst
->exec_size
<= 16 || inst
->force_writemask_all
);
1402 gen7_block_read_scratch(p
, dst
, inst
->exec_size
/ 8, inst
->offset
);
1406 fs_generator::generate_uniform_pull_constant_load(fs_inst
*inst
,
1408 struct brw_reg index
,
1409 struct brw_reg offset
)
1411 assert(type_sz(dst
.type
) == 4);
1412 assert(inst
->mlen
!= 0);
1414 assert(index
.file
== BRW_IMMEDIATE_VALUE
&&
1415 index
.type
== BRW_REGISTER_TYPE_UD
);
1416 uint32_t surf_index
= index
.ud
;
1418 assert(offset
.file
== BRW_IMMEDIATE_VALUE
&&
1419 offset
.type
== BRW_REGISTER_TYPE_UD
);
1420 uint32_t read_offset
= offset
.ud
;
1422 brw_oword_block_read(p
, dst
, brw_message_reg(inst
->base_mrf
),
1423 read_offset
, surf_index
);
1427 fs_generator::generate_uniform_pull_constant_load_gen7(fs_inst
*inst
,
1429 struct brw_reg index
,
1430 struct brw_reg payload
)
1432 assert(index
.type
== BRW_REGISTER_TYPE_UD
);
1433 assert(payload
.file
== BRW_GENERAL_REGISTER_FILE
);
1434 assert(type_sz(dst
.type
) == 4);
1436 if (index
.file
== BRW_IMMEDIATE_VALUE
) {
1437 const uint32_t surf_index
= index
.ud
;
1439 brw_push_insn_state(p
);
1440 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1441 brw_inst
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
1442 brw_pop_insn_state(p
);
1444 brw_inst_set_sfid(devinfo
, send
, GEN6_SFID_DATAPORT_CONSTANT_CACHE
);
1445 brw_set_dest(p
, send
, retype(dst
, BRW_REGISTER_TYPE_UD
));
1446 brw_set_src0(p
, send
, retype(payload
, BRW_REGISTER_TYPE_UD
));
1447 brw_set_desc(p
, send
,
1448 brw_message_desc(devinfo
, 1, DIV_ROUND_UP(inst
->size_written
,
1450 brw_dp_read_desc(devinfo
, surf_index
,
1451 BRW_DATAPORT_OWORD_BLOCK_DWORDS(inst
->exec_size
),
1452 GEN7_DATAPORT_DC_OWORD_BLOCK_READ
,
1453 BRW_DATAPORT_READ_TARGET_DATA_CACHE
));
1456 struct brw_reg addr
= vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD
));
1458 brw_push_insn_state(p
);
1459 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1461 /* a0.0 = surf_index & 0xff */
1462 brw_inst
*insn_and
= brw_next_insn(p
, BRW_OPCODE_AND
);
1463 brw_inst_set_exec_size(p
->devinfo
, insn_and
, BRW_EXECUTE_1
);
1464 brw_set_dest(p
, insn_and
, addr
);
1465 brw_set_src0(p
, insn_and
, vec1(retype(index
, BRW_REGISTER_TYPE_UD
)));
1466 brw_set_src1(p
, insn_and
, brw_imm_ud(0x0ff));
1468 /* dst = send(payload, a0.0 | <descriptor>) */
1469 brw_send_indirect_message(
1470 p
, GEN6_SFID_DATAPORT_CONSTANT_CACHE
,
1471 retype(dst
, BRW_REGISTER_TYPE_UD
),
1472 retype(payload
, BRW_REGISTER_TYPE_UD
), addr
,
1473 brw_message_desc(devinfo
, 1,
1474 DIV_ROUND_UP(inst
->size_written
, REG_SIZE
), true) |
1475 brw_dp_read_desc(devinfo
, 0 /* surface */,
1476 BRW_DATAPORT_OWORD_BLOCK_DWORDS(inst
->exec_size
),
1477 GEN7_DATAPORT_DC_OWORD_BLOCK_READ
,
1478 BRW_DATAPORT_READ_TARGET_DATA_CACHE
),
1481 brw_pop_insn_state(p
);
1486 fs_generator::generate_varying_pull_constant_load_gen4(fs_inst
*inst
,
1488 struct brw_reg index
)
1490 assert(devinfo
->gen
< 7); /* Should use the gen7 variant. */
1491 assert(inst
->header_size
!= 0);
1494 assert(index
.file
== BRW_IMMEDIATE_VALUE
&&
1495 index
.type
== BRW_REGISTER_TYPE_UD
);
1496 uint32_t surf_index
= index
.ud
;
1498 uint32_t simd_mode
, rlen
, msg_type
;
1499 if (inst
->exec_size
== 16) {
1500 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1503 assert(inst
->exec_size
== 8);
1504 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD8
;
1508 if (devinfo
->gen
>= 5)
1509 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
1511 /* We always use the SIMD16 message so that we only have to load U, and
1514 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_LD
;
1515 assert(inst
->mlen
== 3);
1516 assert(inst
->size_written
== 8 * REG_SIZE
);
1518 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1521 struct brw_reg header
= brw_vec8_grf(0, 0);
1522 gen6_resolve_implied_move(p
, &header
, inst
->base_mrf
);
1524 brw_inst
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
1525 brw_inst_set_compression(devinfo
, send
, false);
1526 brw_inst_set_sfid(devinfo
, send
, BRW_SFID_SAMPLER
);
1527 brw_set_dest(p
, send
, retype(dst
, BRW_REGISTER_TYPE_UW
));
1528 brw_set_src0(p
, send
, header
);
1529 if (devinfo
->gen
< 6)
1530 brw_inst_set_base_mrf(p
->devinfo
, send
, inst
->base_mrf
);
1532 /* Our surface is set up as floats, regardless of what actual data is
1535 uint32_t return_format
= BRW_SAMPLER_RETURN_FORMAT_FLOAT32
;
1536 brw_set_desc(p
, send
,
1537 brw_message_desc(devinfo
, inst
->mlen
, rlen
, inst
->header_size
) |
1538 brw_sampler_desc(devinfo
, surf_index
,
1539 0, /* sampler (unused) */
1540 msg_type
, simd_mode
, return_format
));
1544 fs_generator::generate_pixel_interpolator_query(fs_inst
*inst
,
1547 struct brw_reg msg_data
,
1550 const bool has_payload
= inst
->src
[0].file
!= BAD_FILE
;
1551 assert(msg_data
.type
== BRW_REGISTER_TYPE_UD
);
1552 assert(inst
->size_written
% REG_SIZE
== 0);
1554 brw_pixel_interpolator_query(p
,
1555 retype(dst
, BRW_REGISTER_TYPE_UW
),
1556 /* If we don't have a payload, what we send doesn't matter */
1557 has_payload
? src
: brw_vec8_grf(0, 0),
1558 inst
->pi_noperspective
,
1561 has_payload
? 2 * inst
->exec_size
/ 8 : 1,
1562 inst
->size_written
/ REG_SIZE
);
1565 /* Sets vstride=1, width=4, hstride=0 of register src1 during
1566 * the ADD instruction.
1569 fs_generator::generate_set_sample_id(fs_inst
*inst
,
1571 struct brw_reg src0
,
1572 struct brw_reg src1
)
1574 assert(dst
.type
== BRW_REGISTER_TYPE_D
||
1575 dst
.type
== BRW_REGISTER_TYPE_UD
);
1576 assert(src0
.type
== BRW_REGISTER_TYPE_D
||
1577 src0
.type
== BRW_REGISTER_TYPE_UD
);
1579 const struct brw_reg reg
= stride(src1
, 1, 4, 0);
1580 const unsigned lower_size
= MIN2(inst
->exec_size
,
1581 devinfo
->gen
>= 8 ? 16 : 8);
1583 for (unsigned i
= 0; i
< inst
->exec_size
/ lower_size
; i
++) {
1584 brw_inst
*insn
= brw_ADD(p
, offset(dst
, i
* lower_size
/ 8),
1585 offset(src0
, (src0
.vstride
== 0 ? 0 : (1 << (src0
.vstride
- 1)) *
1586 (i
* lower_size
/ (1 << src0
.width
))) *
1587 type_sz(src0
.type
) / REG_SIZE
),
1588 suboffset(reg
, i
* lower_size
/ 4));
1589 brw_inst_set_exec_size(devinfo
, insn
, cvt(lower_size
) - 1);
1590 brw_inst_set_group(devinfo
, insn
, inst
->group
+ lower_size
* i
);
1591 brw_inst_set_compression(devinfo
, insn
, lower_size
> 8);
1596 fs_generator::generate_pack_half_2x16_split(fs_inst
*,
1601 assert(devinfo
->gen
>= 7);
1602 assert(dst
.type
== BRW_REGISTER_TYPE_UD
);
1603 assert(x
.type
== BRW_REGISTER_TYPE_F
);
1604 assert(y
.type
== BRW_REGISTER_TYPE_F
);
1606 /* From the Ivybridge PRM, Vol4, Part3, Section 6.27 f32to16:
1608 * Because this instruction does not have a 16-bit floating-point type,
1609 * the destination data type must be Word (W).
1611 * The destination must be DWord-aligned and specify a horizontal stride
1612 * (HorzStride) of 2. The 16-bit result is stored in the lower word of
1613 * each destination channel and the upper word is not modified.
1615 struct brw_reg dst_w
= spread(retype(dst
, BRW_REGISTER_TYPE_W
), 2);
1617 /* Give each 32-bit channel of dst the form below, where "." means
1621 brw_F32TO16(p
, dst_w
, y
);
1626 brw_SHL(p
, dst
, dst
, brw_imm_ud(16u));
1628 /* And, finally the form of packHalf2x16's output:
1631 brw_F32TO16(p
, dst_w
, x
);
1635 fs_generator::generate_shader_time_add(fs_inst
*,
1636 struct brw_reg payload
,
1637 struct brw_reg offset
,
1638 struct brw_reg value
)
1640 assert(devinfo
->gen
>= 7);
1641 brw_push_insn_state(p
);
1642 brw_set_default_mask_control(p
, true);
1644 assert(payload
.file
== BRW_GENERAL_REGISTER_FILE
);
1645 struct brw_reg payload_offset
= retype(brw_vec1_grf(payload
.nr
, 0),
1647 struct brw_reg payload_value
= retype(brw_vec1_grf(payload
.nr
+ 1, 0),
1650 assert(offset
.file
== BRW_IMMEDIATE_VALUE
);
1651 if (value
.file
== BRW_GENERAL_REGISTER_FILE
) {
1652 value
.width
= BRW_WIDTH_1
;
1653 value
.hstride
= BRW_HORIZONTAL_STRIDE_0
;
1654 value
.vstride
= BRW_VERTICAL_STRIDE_0
;
1656 assert(value
.file
== BRW_IMMEDIATE_VALUE
);
1659 /* Trying to deal with setup of the params from the IR is crazy in the FS8
1660 * case, and we don't really care about squeezing every bit of performance
1661 * out of this path, so we just emit the MOVs from here.
1663 brw_MOV(p
, payload_offset
, offset
);
1664 brw_MOV(p
, payload_value
, value
);
1665 brw_shader_time_add(p
, payload
,
1666 prog_data
->binding_table
.shader_time_start
);
1667 brw_pop_insn_state(p
);
1671 fs_generator::enable_debug(const char *shader_name
)
1674 this->shader_name
= shader_name
;
1678 fs_generator::generate_code(const cfg_t
*cfg
, int dispatch_width
)
1680 /* align to 64 byte boundary. */
1681 while (p
->next_insn_offset
% 64)
1684 this->dispatch_width
= dispatch_width
;
1686 int start_offset
= p
->next_insn_offset
;
1687 int spill_count
= 0, fill_count
= 0;
1690 struct disasm_info
*disasm_info
= disasm_initialize(devinfo
, cfg
);
1692 foreach_block_and_inst (block
, fs_inst
, inst
, cfg
) {
1693 struct brw_reg src
[4], dst
;
1694 unsigned int last_insn_offset
= p
->next_insn_offset
;
1695 bool multiple_instructions_emitted
= false;
1697 /* From the Broadwell PRM, Volume 7, "3D-Media-GPGPU", in the
1698 * "Register Region Restrictions" section: for BDW, SKL:
1700 * "A POW/FDIV operation must not be followed by an instruction
1701 * that requires two destination registers."
1703 * The documentation is often lacking annotations for Atom parts,
1704 * and empirically this affects CHV as well.
1706 if (devinfo
->gen
>= 8 &&
1707 devinfo
->gen
<= 9 &&
1709 brw_inst_opcode(devinfo
, brw_last_inst
) == BRW_OPCODE_MATH
&&
1710 brw_inst_math_function(devinfo
, brw_last_inst
) == BRW_MATH_FUNCTION_POW
&&
1711 inst
->dst
.component_size(inst
->exec_size
) > REG_SIZE
) {
1713 last_insn_offset
= p
->next_insn_offset
;
1716 if (unlikely(debug_flag
))
1717 disasm_annotate(disasm_info
, inst
, p
->next_insn_offset
);
1719 /* If the instruction writes to more than one register, it needs to be
1720 * explicitly marked as compressed on Gen <= 5. On Gen >= 6 the
1721 * hardware figures out by itself what the right compression mode is,
1722 * but we still need to know whether the instruction is compressed to
1723 * set up the source register regions appropriately.
1725 * XXX - This is wrong for instructions that write a single register but
1726 * read more than one which should strictly speaking be treated as
1727 * compressed. For instructions that don't write any registers it
1728 * relies on the destination being a null register of the correct
1729 * type and regioning so the instruction is considered compressed
1730 * or not accordingly.
1732 const bool compressed
=
1733 inst
->dst
.component_size(inst
->exec_size
) > REG_SIZE
;
1734 brw_set_default_compression(p
, compressed
);
1735 brw_set_default_group(p
, inst
->group
);
1737 for (unsigned int i
= 0; i
< inst
->sources
; i
++) {
1738 src
[i
] = brw_reg_from_fs_reg(devinfo
, inst
,
1739 &inst
->src
[i
], compressed
);
1740 /* The accumulator result appears to get used for the
1741 * conditional modifier generation. When negating a UD
1742 * value, there is a 33rd bit generated for the sign in the
1743 * accumulator value, so now you can't check, for example,
1744 * equality with a 32-bit value. See piglit fs-op-neg-uvec4.
1746 assert(!inst
->conditional_mod
||
1747 inst
->src
[i
].type
!= BRW_REGISTER_TYPE_UD
||
1748 !inst
->src
[i
].negate
);
1750 dst
= brw_reg_from_fs_reg(devinfo
, inst
,
1751 &inst
->dst
, compressed
);
1753 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
1754 brw_set_default_predicate_control(p
, inst
->predicate
);
1755 brw_set_default_predicate_inverse(p
, inst
->predicate_inverse
);
1756 /* On gen7 and above, hardware automatically adds the group onto the
1757 * flag subregister number. On Sandy Bridge and older, we have to do it
1760 const unsigned flag_subreg
= inst
->flag_subreg
+
1761 (devinfo
->gen
>= 7 ? 0 : inst
->group
/ 16);
1762 brw_set_default_flag_reg(p
, flag_subreg
/ 2, flag_subreg
% 2);
1763 brw_set_default_saturate(p
, inst
->saturate
);
1764 brw_set_default_mask_control(p
, inst
->force_writemask_all
);
1765 brw_set_default_acc_write_control(p
, inst
->writes_accumulator
);
1767 unsigned exec_size
= inst
->exec_size
;
1768 if (devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
1769 (get_exec_type_size(inst
) == 8 || type_sz(inst
->dst
.type
) == 8)) {
1773 brw_set_default_exec_size(p
, cvt(exec_size
) - 1);
1775 assert(inst
->force_writemask_all
|| inst
->exec_size
>= 4);
1776 assert(inst
->force_writemask_all
|| inst
->group
% inst
->exec_size
== 0);
1777 assert(inst
->base_mrf
+ inst
->mlen
<= BRW_MAX_MRF(devinfo
->gen
));
1778 assert(inst
->mlen
<= BRW_MAX_MSG_LENGTH
);
1780 switch (inst
->opcode
) {
1781 case BRW_OPCODE_MOV
:
1782 brw_MOV(p
, dst
, src
[0]);
1784 case BRW_OPCODE_ADD
:
1785 brw_ADD(p
, dst
, src
[0], src
[1]);
1787 case BRW_OPCODE_MUL
:
1788 brw_MUL(p
, dst
, src
[0], src
[1]);
1790 case BRW_OPCODE_AVG
:
1791 brw_AVG(p
, dst
, src
[0], src
[1]);
1793 case BRW_OPCODE_MACH
:
1794 brw_MACH(p
, dst
, src
[0], src
[1]);
1797 case BRW_OPCODE_LINE
:
1798 brw_LINE(p
, dst
, src
[0], src
[1]);
1801 case BRW_OPCODE_MAD
:
1802 assert(devinfo
->gen
>= 6);
1803 if (devinfo
->gen
< 10)
1804 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1805 brw_MAD(p
, dst
, src
[0], src
[1], src
[2]);
1808 case BRW_OPCODE_LRP
:
1809 assert(devinfo
->gen
>= 6 && devinfo
->gen
<= 10);
1810 if (devinfo
->gen
< 10)
1811 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1812 brw_LRP(p
, dst
, src
[0], src
[1], src
[2]);
1815 case BRW_OPCODE_FRC
:
1816 brw_FRC(p
, dst
, src
[0]);
1818 case BRW_OPCODE_RNDD
:
1819 brw_RNDD(p
, dst
, src
[0]);
1821 case BRW_OPCODE_RNDE
:
1822 brw_RNDE(p
, dst
, src
[0]);
1824 case BRW_OPCODE_RNDZ
:
1825 brw_RNDZ(p
, dst
, src
[0]);
1828 case BRW_OPCODE_AND
:
1829 brw_AND(p
, dst
, src
[0], src
[1]);
1832 brw_OR(p
, dst
, src
[0], src
[1]);
1834 case BRW_OPCODE_XOR
:
1835 brw_XOR(p
, dst
, src
[0], src
[1]);
1837 case BRW_OPCODE_NOT
:
1838 brw_NOT(p
, dst
, src
[0]);
1840 case BRW_OPCODE_ASR
:
1841 brw_ASR(p
, dst
, src
[0], src
[1]);
1843 case BRW_OPCODE_SHR
:
1844 brw_SHR(p
, dst
, src
[0], src
[1]);
1846 case BRW_OPCODE_SHL
:
1847 brw_SHL(p
, dst
, src
[0], src
[1]);
1849 case BRW_OPCODE_F32TO16
:
1850 assert(devinfo
->gen
>= 7);
1851 brw_F32TO16(p
, dst
, src
[0]);
1853 case BRW_OPCODE_F16TO32
:
1854 assert(devinfo
->gen
>= 7);
1855 brw_F16TO32(p
, dst
, src
[0]);
1857 case BRW_OPCODE_CMP
:
1858 if (inst
->exec_size
>= 16 && devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
1859 dst
.file
== BRW_ARCHITECTURE_REGISTER_FILE
) {
1860 /* For unknown reasons the WaCMPInstFlagDepClearedEarly workaround
1861 * implemented in the compiler is not sufficient. Overriding the
1862 * type when the destination is the null register is necessary but
1863 * not sufficient by itself.
1865 assert(dst
.nr
== BRW_ARF_NULL
);
1866 dst
.type
= BRW_REGISTER_TYPE_D
;
1868 brw_CMP(p
, dst
, inst
->conditional_mod
, src
[0], src
[1]);
1870 case BRW_OPCODE_SEL
:
1871 brw_SEL(p
, dst
, src
[0], src
[1]);
1873 case BRW_OPCODE_CSEL
:
1874 assert(devinfo
->gen
>= 8);
1875 if (devinfo
->gen
< 10)
1876 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1877 brw_CSEL(p
, dst
, src
[0], src
[1], src
[2]);
1879 case BRW_OPCODE_BFREV
:
1880 assert(devinfo
->gen
>= 7);
1881 brw_BFREV(p
, retype(dst
, BRW_REGISTER_TYPE_UD
),
1882 retype(src
[0], BRW_REGISTER_TYPE_UD
));
1884 case BRW_OPCODE_FBH
:
1885 assert(devinfo
->gen
>= 7);
1886 brw_FBH(p
, retype(dst
, src
[0].type
), src
[0]);
1888 case BRW_OPCODE_FBL
:
1889 assert(devinfo
->gen
>= 7);
1890 brw_FBL(p
, retype(dst
, BRW_REGISTER_TYPE_UD
),
1891 retype(src
[0], BRW_REGISTER_TYPE_UD
));
1893 case BRW_OPCODE_LZD
:
1894 brw_LZD(p
, dst
, src
[0]);
1896 case BRW_OPCODE_CBIT
:
1897 assert(devinfo
->gen
>= 7);
1898 brw_CBIT(p
, retype(dst
, BRW_REGISTER_TYPE_UD
),
1899 retype(src
[0], BRW_REGISTER_TYPE_UD
));
1901 case BRW_OPCODE_ADDC
:
1902 assert(devinfo
->gen
>= 7);
1903 brw_ADDC(p
, dst
, src
[0], src
[1]);
1905 case BRW_OPCODE_SUBB
:
1906 assert(devinfo
->gen
>= 7);
1907 brw_SUBB(p
, dst
, src
[0], src
[1]);
1909 case BRW_OPCODE_MAC
:
1910 brw_MAC(p
, dst
, src
[0], src
[1]);
1913 case BRW_OPCODE_BFE
:
1914 assert(devinfo
->gen
>= 7);
1915 if (devinfo
->gen
< 10)
1916 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1917 brw_BFE(p
, dst
, src
[0], src
[1], src
[2]);
1920 case BRW_OPCODE_BFI1
:
1921 assert(devinfo
->gen
>= 7);
1922 brw_BFI1(p
, dst
, src
[0], src
[1]);
1924 case BRW_OPCODE_BFI2
:
1925 assert(devinfo
->gen
>= 7);
1926 if (devinfo
->gen
< 10)
1927 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1928 brw_BFI2(p
, dst
, src
[0], src
[1], src
[2]);
1932 if (inst
->src
[0].file
!= BAD_FILE
) {
1933 /* The instruction has an embedded compare (only allowed on gen6) */
1934 assert(devinfo
->gen
== 6);
1935 gen6_IF(p
, inst
->conditional_mod
, src
[0], src
[1]);
1937 brw_IF(p
, brw_get_default_exec_size(p
));
1941 case BRW_OPCODE_ELSE
:
1944 case BRW_OPCODE_ENDIF
:
1949 brw_DO(p
, brw_get_default_exec_size(p
));
1952 case BRW_OPCODE_BREAK
:
1955 case BRW_OPCODE_CONTINUE
:
1959 case BRW_OPCODE_WHILE
:
1964 case SHADER_OPCODE_RCP
:
1965 case SHADER_OPCODE_RSQ
:
1966 case SHADER_OPCODE_SQRT
:
1967 case SHADER_OPCODE_EXP2
:
1968 case SHADER_OPCODE_LOG2
:
1969 case SHADER_OPCODE_SIN
:
1970 case SHADER_OPCODE_COS
:
1971 assert(inst
->conditional_mod
== BRW_CONDITIONAL_NONE
);
1972 if (devinfo
->gen
>= 6) {
1973 assert(inst
->mlen
== 0);
1974 assert(devinfo
->gen
>= 7 || inst
->exec_size
== 8);
1975 gen6_math(p
, dst
, brw_math_function(inst
->opcode
),
1976 src
[0], brw_null_reg());
1978 assert(inst
->mlen
>= 1);
1979 assert(devinfo
->gen
== 5 || devinfo
->is_g4x
|| inst
->exec_size
== 8);
1981 brw_math_function(inst
->opcode
),
1982 inst
->base_mrf
, src
[0],
1983 BRW_MATH_PRECISION_FULL
);
1986 case SHADER_OPCODE_INT_QUOTIENT
:
1987 case SHADER_OPCODE_INT_REMAINDER
:
1988 case SHADER_OPCODE_POW
:
1989 assert(inst
->conditional_mod
== BRW_CONDITIONAL_NONE
);
1990 if (devinfo
->gen
>= 6) {
1991 assert(inst
->mlen
== 0);
1992 assert((devinfo
->gen
>= 7 && inst
->opcode
== SHADER_OPCODE_POW
) ||
1993 inst
->exec_size
== 8);
1994 gen6_math(p
, dst
, brw_math_function(inst
->opcode
), src
[0], src
[1]);
1996 assert(inst
->mlen
>= 1);
1997 assert(inst
->exec_size
== 8);
1998 gen4_math(p
, dst
, brw_math_function(inst
->opcode
),
1999 inst
->base_mrf
, src
[0],
2000 BRW_MATH_PRECISION_FULL
);
2003 case FS_OPCODE_LINTERP
:
2004 multiple_instructions_emitted
= generate_linterp(inst
, dst
, src
);
2006 case FS_OPCODE_PIXEL_X
:
2007 assert(src
[0].type
== BRW_REGISTER_TYPE_UW
);
2008 src
[0].subnr
= 0 * type_sz(src
[0].type
);
2009 brw_MOV(p
, dst
, stride(src
[0], 8, 4, 1));
2011 case FS_OPCODE_PIXEL_Y
:
2012 assert(src
[0].type
== BRW_REGISTER_TYPE_UW
);
2013 src
[0].subnr
= 4 * type_sz(src
[0].type
);
2014 brw_MOV(p
, dst
, stride(src
[0], 8, 4, 1));
2017 case SHADER_OPCODE_SEND
:
2018 generate_send(inst
, dst
, src
[0], src
[1], src
[2],
2019 inst
->ex_mlen
> 0 ? src
[3] : brw_null_reg());
2022 case SHADER_OPCODE_GET_BUFFER_SIZE
:
2023 generate_get_buffer_size(inst
, dst
, src
[0], src
[1]);
2025 case SHADER_OPCODE_TEX
:
2027 case SHADER_OPCODE_TXD
:
2028 case SHADER_OPCODE_TXF
:
2029 case SHADER_OPCODE_TXF_CMS
:
2030 case SHADER_OPCODE_TXL
:
2031 case SHADER_OPCODE_TXS
:
2032 case SHADER_OPCODE_LOD
:
2033 case SHADER_OPCODE_TG4
:
2034 case SHADER_OPCODE_SAMPLEINFO
:
2035 assert(inst
->src
[0].file
== BAD_FILE
);
2036 generate_tex(inst
, dst
, src
[1], src
[2]);
2039 case FS_OPCODE_DDX_COARSE
:
2040 case FS_OPCODE_DDX_FINE
:
2041 generate_ddx(inst
, dst
, src
[0]);
2043 case FS_OPCODE_DDY_COARSE
:
2044 case FS_OPCODE_DDY_FINE
:
2045 generate_ddy(inst
, dst
, src
[0]);
2048 case SHADER_OPCODE_GEN4_SCRATCH_WRITE
:
2049 generate_scratch_write(inst
, src
[0]);
2053 case SHADER_OPCODE_GEN4_SCRATCH_READ
:
2054 generate_scratch_read(inst
, dst
);
2058 case SHADER_OPCODE_GEN7_SCRATCH_READ
:
2059 generate_scratch_read_gen7(inst
, dst
);
2063 case SHADER_OPCODE_MOV_INDIRECT
:
2064 generate_mov_indirect(inst
, dst
, src
[0], src
[1]);
2067 case SHADER_OPCODE_URB_READ_SIMD8
:
2068 case SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
:
2069 generate_urb_read(inst
, dst
, src
[0]);
2072 case SHADER_OPCODE_URB_WRITE_SIMD8
:
2073 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
2074 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
2075 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
2076 generate_urb_write(inst
, src
[0]);
2079 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
2080 assert(inst
->force_writemask_all
);
2081 generate_uniform_pull_constant_load(inst
, dst
, src
[0], src
[1]);
2084 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
:
2085 assert(inst
->force_writemask_all
);
2086 generate_uniform_pull_constant_load_gen7(inst
, dst
, src
[0], src
[1]);
2089 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN4
:
2090 generate_varying_pull_constant_load_gen4(inst
, dst
, src
[0]);
2093 case FS_OPCODE_REP_FB_WRITE
:
2094 case FS_OPCODE_FB_WRITE
:
2095 generate_fb_write(inst
, src
[0]);
2098 case FS_OPCODE_FB_READ
:
2099 generate_fb_read(inst
, dst
, src
[0]);
2102 case FS_OPCODE_DISCARD_JUMP
:
2103 generate_discard_jump(inst
);
2106 case SHADER_OPCODE_SHADER_TIME_ADD
:
2107 generate_shader_time_add(inst
, src
[0], src
[1], src
[2]);
2110 case SHADER_OPCODE_MEMORY_FENCE
:
2111 brw_memory_fence(p
, dst
, BRW_OPCODE_SEND
);
2114 case SHADER_OPCODE_INTERLOCK
:
2115 assert(devinfo
->gen
>= 9);
2116 /* The interlock is basically a memory fence issued via sendc */
2117 brw_memory_fence(p
, dst
, BRW_OPCODE_SENDC
);
2120 case SHADER_OPCODE_FIND_LIVE_CHANNEL
: {
2121 const struct brw_reg mask
=
2122 brw_stage_has_packed_dispatch(devinfo
, stage
,
2123 prog_data
) ? brw_imm_ud(~0u) :
2124 stage
== MESA_SHADER_FRAGMENT
? brw_vmask_reg() :
2126 brw_find_live_channel(p
, dst
, mask
);
2130 case SHADER_OPCODE_BROADCAST
:
2131 assert(inst
->force_writemask_all
);
2132 brw_broadcast(p
, dst
, src
[0], src
[1]);
2135 case SHADER_OPCODE_SHUFFLE
:
2136 generate_shuffle(inst
, dst
, src
[0], src
[1]);
2139 case SHADER_OPCODE_SEL_EXEC
:
2140 assert(inst
->force_writemask_all
);
2141 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
2142 brw_MOV(p
, dst
, src
[1]);
2143 brw_set_default_mask_control(p
, BRW_MASK_ENABLE
);
2144 brw_MOV(p
, dst
, src
[0]);
2147 case SHADER_OPCODE_QUAD_SWIZZLE
:
2148 assert(src
[1].file
== BRW_IMMEDIATE_VALUE
);
2149 assert(src
[1].type
== BRW_REGISTER_TYPE_UD
);
2150 generate_quad_swizzle(inst
, dst
, src
[0], src
[1].ud
);
2153 case SHADER_OPCODE_CLUSTER_BROADCAST
: {
2154 assert(src
[0].type
== dst
.type
);
2155 assert(!src
[0].negate
&& !src
[0].abs
);
2156 assert(src
[1].file
== BRW_IMMEDIATE_VALUE
);
2157 assert(src
[1].type
== BRW_REGISTER_TYPE_UD
);
2158 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2159 assert(src
[2].type
== BRW_REGISTER_TYPE_UD
);
2160 const unsigned component
= src
[1].ud
;
2161 const unsigned cluster_size
= src
[2].ud
;
2162 struct brw_reg strided
= stride(suboffset(src
[0], component
),
2163 cluster_size
, cluster_size
, 0);
2164 if (type_sz(src
[0].type
) > 4 &&
2165 (devinfo
->is_cherryview
|| gen_device_info_is_9lp(devinfo
))) {
2166 /* IVB has an issue (which we found empirically) where it reads
2167 * two address register components per channel for indirectly
2168 * addressed 64-bit sources.
2170 * From the Cherryview PRM Vol 7. "Register Region Restrictions":
2172 * "When source or destination datatype is 64b or operation is
2173 * integer DWord multiply, indirect addressing must not be
2176 * To work around both of these, we do two integer MOVs insead of
2177 * one 64-bit MOV. Because no double value should ever cross a
2178 * register boundary, it's safe to use the immediate offset in the
2179 * indirect here to handle adding 4 bytes to the offset and avoid
2180 * the extra ADD to the register file.
2182 brw_MOV(p
, subscript(dst
, BRW_REGISTER_TYPE_D
, 0),
2183 subscript(strided
, BRW_REGISTER_TYPE_D
, 0));
2184 brw_MOV(p
, subscript(dst
, BRW_REGISTER_TYPE_D
, 1),
2185 subscript(strided
, BRW_REGISTER_TYPE_D
, 1));
2187 brw_MOV(p
, dst
, strided
);
2192 case FS_OPCODE_SET_SAMPLE_ID
:
2193 generate_set_sample_id(inst
, dst
, src
[0], src
[1]);
2196 case FS_OPCODE_PACK_HALF_2x16_SPLIT
:
2197 generate_pack_half_2x16_split(inst
, dst
, src
[0], src
[1]);
2200 case FS_OPCODE_PLACEHOLDER_HALT
:
2201 /* This is the place where the final HALT needs to be inserted if
2202 * we've emitted any discards. If not, this will emit no code.
2204 if (!patch_discard_jumps_to_fb_writes()) {
2205 if (unlikely(debug_flag
)) {
2206 disasm_info
->use_tail
= true;
2211 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
2212 generate_pixel_interpolator_query(inst
, dst
, src
[0], src
[1],
2213 GEN7_PIXEL_INTERPOLATOR_LOC_SAMPLE
);
2216 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
2217 generate_pixel_interpolator_query(inst
, dst
, src
[0], src
[1],
2218 GEN7_PIXEL_INTERPOLATOR_LOC_SHARED_OFFSET
);
2221 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
2222 generate_pixel_interpolator_query(inst
, dst
, src
[0], src
[1],
2223 GEN7_PIXEL_INTERPOLATOR_LOC_PER_SLOT_OFFSET
);
2226 case CS_OPCODE_CS_TERMINATE
:
2227 generate_cs_terminate(inst
, src
[0]);
2230 case SHADER_OPCODE_BARRIER
:
2231 generate_barrier(inst
, src
[0]);
2234 case BRW_OPCODE_DIM
:
2235 assert(devinfo
->is_haswell
);
2236 assert(src
[0].type
== BRW_REGISTER_TYPE_DF
);
2237 assert(dst
.type
== BRW_REGISTER_TYPE_DF
);
2238 brw_DIM(p
, dst
, retype(src
[0], BRW_REGISTER_TYPE_F
));
2241 case SHADER_OPCODE_RND_MODE
:
2242 assert(src
[0].file
== BRW_IMMEDIATE_VALUE
);
2243 brw_rounding_mode(p
, (brw_rnd_mode
) src
[0].d
);
2247 unreachable("Unsupported opcode");
2249 case SHADER_OPCODE_LOAD_PAYLOAD
:
2250 unreachable("Should be lowered by lower_load_payload()");
2253 if (multiple_instructions_emitted
)
2256 if (inst
->no_dd_clear
|| inst
->no_dd_check
|| inst
->conditional_mod
) {
2257 assert(p
->next_insn_offset
== last_insn_offset
+ 16 ||
2258 !"conditional_mod, no_dd_check, or no_dd_clear set for IR "
2259 "emitting more than 1 instruction");
2261 brw_inst
*last
= &p
->store
[last_insn_offset
/ 16];
2263 if (inst
->conditional_mod
)
2264 brw_inst_set_cond_modifier(p
->devinfo
, last
, inst
->conditional_mod
);
2265 brw_inst_set_no_dd_clear(p
->devinfo
, last
, inst
->no_dd_clear
);
2266 brw_inst_set_no_dd_check(p
->devinfo
, last
, inst
->no_dd_check
);
2270 brw_set_uip_jip(p
, start_offset
);
2272 /* end of program sentinel */
2273 disasm_new_inst_group(disasm_info
, p
->next_insn_offset
);
2278 if (unlikely(debug_flag
))
2280 brw_validate_instructions(devinfo
, p
->store
,
2282 p
->next_insn_offset
,
2285 int before_size
= p
->next_insn_offset
- start_offset
;
2286 brw_compact_instructions(p
, start_offset
, disasm_info
);
2287 int after_size
= p
->next_insn_offset
- start_offset
;
2289 if (unlikely(debug_flag
)) {
2290 fprintf(stderr
, "Native code for %s\n"
2291 "SIMD%d shader: %d instructions. %d loops. %u cycles. %d:%d spills:fills. Promoted %u constants. Compacted %d to %d"
2292 " bytes (%.0f%%)\n",
2293 shader_name
, dispatch_width
, before_size
/ 16, loop_count
, cfg
->cycle_count
,
2294 spill_count
, fill_count
, promoted_constants
, before_size
, after_size
,
2295 100.0f
* (before_size
- after_size
) / before_size
);
2297 dump_assembly(p
->store
, disasm_info
);
2299 ralloc_free(disasm_info
);
2302 compiler
->shader_debug_log(log_data
,
2303 "%s SIMD%d shader: %d inst, %d loops, %u cycles, "
2304 "%d:%d spills:fills, Promoted %u constants, "
2305 "compacted %d to %d bytes.",
2306 _mesa_shader_stage_to_abbrev(stage
),
2307 dispatch_width
, before_size
/ 16,
2308 loop_count
, cfg
->cycle_count
, spill_count
,
2309 fill_count
, promoted_constants
, before_size
,
2312 return start_offset
;
2316 fs_generator::get_assembly()
2318 return brw_get_program(p
, &prog_data
->program_size
);