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 const unsigned width
= MIN2(reg_width
, phys_width
);
94 brw_reg
= brw_vecn_reg(width
, brw_file_from_reg(reg
), reg
->nr
, 0);
95 brw_reg
= stride(brw_reg
, width
* reg
->stride
, width
, reg
->stride
);
97 if (devinfo
->gen
== 7 && !devinfo
->is_haswell
) {
98 /* From the IvyBridge PRM (EU Changes by Processor Generation, page 13):
99 * "Each DF (Double Float) operand uses an element size of 4 rather
100 * than 8 and all regioning parameters are twice what the values
101 * would be based on the true element size: ExecSize, Width,
102 * HorzStride, and VertStride. Each DF operand uses a pair of
103 * channels and all masking and swizzing should be adjusted
106 * From the IvyBridge PRM (Special Requirements for Handling Double
107 * Precision Data Types, page 71):
108 * "In Align1 mode, all regioning parameters like stride, execution
109 * size, and width must use the syntax of a pair of packed
110 * floats. The offsets for these data types must be 64-bit
111 * aligned. The execution size and regioning parameters are in terms
114 * Summarized: when handling DF-typed arguments, ExecSize,
115 * VertStride, and Width must be doubled.
117 * It applies to BayTrail too.
119 if (type_sz(reg
->type
) == 8) {
121 if (brw_reg
.vstride
> 0)
123 assert(brw_reg
.hstride
== BRW_HORIZONTAL_STRIDE_1
);
126 /* When converting from DF->F, we set the destination stride to 2
127 * because each d2f conversion implicitly writes 2 floats, being
128 * the first one the converted value. IVB/BYT actually writes two
129 * F components per SIMD channel, and every other component is
130 * filled with garbage.
132 if (reg
== &inst
->dst
&& get_exec_type_size(inst
) == 8 &&
133 type_sz(inst
->dst
.type
) < 8) {
134 assert(brw_reg
.hstride
> BRW_HORIZONTAL_STRIDE_1
);
140 brw_reg
= retype(brw_reg
, reg
->type
);
141 brw_reg
= byte_offset(brw_reg
, reg
->offset
);
142 brw_reg
.abs
= reg
->abs
;
143 brw_reg
.negate
= reg
->negate
;
148 assert(reg
->offset
== 0);
149 brw_reg
= reg
->as_brw_reg();
152 /* Probably unused. */
153 brw_reg
= brw_null_reg();
157 unreachable("not reached");
160 /* On HSW+, scalar DF sources can be accessed using the normal <0,1,0>
161 * region, but on IVB and BYT DF regions must be programmed in terms of
162 * floats. A <0,2,1> region accomplishes this.
164 if (devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
165 type_sz(reg
->type
) == 8 &&
166 brw_reg
.vstride
== BRW_VERTICAL_STRIDE_0
&&
167 brw_reg
.width
== BRW_WIDTH_1
&&
168 brw_reg
.hstride
== BRW_HORIZONTAL_STRIDE_0
) {
169 brw_reg
.width
= BRW_WIDTH_2
;
170 brw_reg
.hstride
= BRW_HORIZONTAL_STRIDE_1
;
176 fs_generator::fs_generator(const struct brw_compiler
*compiler
, void *log_data
,
179 struct brw_stage_prog_data
*prog_data
,
180 unsigned promoted_constants
,
181 bool runtime_check_aads_emit
,
182 gl_shader_stage stage
)
184 : compiler(compiler
), log_data(log_data
),
185 devinfo(compiler
->devinfo
), key(key
),
186 prog_data(prog_data
),
187 promoted_constants(promoted_constants
),
188 runtime_check_aads_emit(runtime_check_aads_emit
), debug_flag(false),
189 stage(stage
), mem_ctx(mem_ctx
)
191 p
= rzalloc(mem_ctx
, struct brw_codegen
);
192 brw_init_codegen(devinfo
, p
, mem_ctx
);
194 /* In the FS code generator, we are very careful to ensure that we always
195 * set the right execution size so we don't need the EU code to "help" us
196 * by trying to infer it. Sometimes, it infers the wrong thing.
198 p
->automatic_exec_sizes
= false;
201 fs_generator::~fs_generator()
205 class ip_record
: public exec_node
{
207 DECLARE_RALLOC_CXX_OPERATORS(ip_record
)
218 fs_generator::patch_discard_jumps_to_fb_writes()
220 if (devinfo
->gen
< 6 || this->discard_halt_patches
.is_empty())
223 int scale
= brw_jump_scale(p
->devinfo
);
225 /* There is a somewhat strange undocumented requirement of using
226 * HALT, according to the simulator. If some channel has HALTed to
227 * a particular UIP, then by the end of the program, every channel
228 * must have HALTed to that UIP. Furthermore, the tracking is a
229 * stack, so you can't do the final halt of a UIP after starting
230 * halting to a new UIP.
232 * Symptoms of not emitting this instruction on actual hardware
233 * included GPU hangs and sparkly rendering on the piglit discard
236 brw_inst
*last_halt
= gen6_HALT(p
);
237 brw_inst_set_uip(p
->devinfo
, last_halt
, 1 * scale
);
238 brw_inst_set_jip(p
->devinfo
, last_halt
, 1 * scale
);
242 foreach_in_list(ip_record
, patch_ip
, &discard_halt_patches
) {
243 brw_inst
*patch
= &p
->store
[patch_ip
->ip
];
245 assert(brw_inst_opcode(p
->devinfo
, patch
) == BRW_OPCODE_HALT
);
246 /* HALT takes a half-instruction distance from the pre-incremented IP. */
247 brw_inst_set_uip(p
->devinfo
, patch
, (ip
- patch_ip
->ip
) * scale
);
250 this->discard_halt_patches
.make_empty();
255 fs_generator::fire_fb_write(fs_inst
*inst
,
256 struct brw_reg payload
,
257 struct brw_reg implied_header
,
260 uint32_t msg_control
;
262 struct brw_wm_prog_data
*prog_data
= brw_wm_prog_data(this->prog_data
);
264 if (devinfo
->gen
< 6) {
265 brw_push_insn_state(p
);
266 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
267 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
268 brw_set_default_predicate_control(p
, BRW_PREDICATE_NONE
);
269 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
270 brw_MOV(p
, offset(payload
, 1), brw_vec8_grf(1, 0));
271 brw_pop_insn_state(p
);
274 if (inst
->opcode
== FS_OPCODE_REP_FB_WRITE
)
275 msg_control
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE_REPLICATED
;
276 else if (prog_data
->dual_src_blend
) {
278 msg_control
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_DUAL_SOURCE_SUBSPAN01
;
280 msg_control
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_DUAL_SOURCE_SUBSPAN23
;
281 } else if (inst
->exec_size
== 16)
282 msg_control
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE
;
284 msg_control
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_SINGLE_SOURCE_SUBSPAN01
;
286 /* We assume render targets start at 0, because headerless FB write
287 * messages set "Render Target Index" to 0. Using a different binding
288 * table index would make it impossible to use headerless messages.
290 const uint32_t surf_index
= inst
->target
;
292 bool last_render_target
= inst
->eot
||
293 (prog_data
->dual_src_blend
&& dispatch_width
== 16);
305 inst
->header_size
!= 0);
307 brw_mark_surface_used(&prog_data
->base
, surf_index
);
311 fs_generator::generate_fb_write(fs_inst
*inst
, struct brw_reg payload
)
313 struct brw_wm_prog_data
*prog_data
= brw_wm_prog_data(this->prog_data
);
314 const brw_wm_prog_key
* const key
= (brw_wm_prog_key
* const) this->key
;
315 struct brw_reg implied_header
;
317 if (devinfo
->gen
< 8 && !devinfo
->is_haswell
) {
318 brw_set_default_predicate_control(p
, BRW_PREDICATE_NONE
);
321 if (inst
->base_mrf
>= 0)
322 payload
= brw_message_reg(inst
->base_mrf
);
324 /* Header is 2 regs, g0 and g1 are the contents. g0 will be implied
327 if (inst
->header_size
!= 0) {
328 brw_push_insn_state(p
);
329 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
330 brw_set_default_exec_size(p
, BRW_EXECUTE_1
);
331 brw_set_default_predicate_control(p
, BRW_PREDICATE_NONE
);
332 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
333 brw_set_default_flag_reg(p
, 0, 0);
335 /* On HSW, the GPU will use the predicate on SENDC, unless the header is
338 if (prog_data
->uses_kill
) {
339 struct brw_reg pixel_mask
;
341 if (devinfo
->gen
>= 6)
342 pixel_mask
= retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW
);
344 pixel_mask
= retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW
);
346 brw_MOV(p
, pixel_mask
, brw_flag_reg(0, 1));
349 if (devinfo
->gen
>= 6) {
350 brw_push_insn_state(p
);
351 brw_set_default_exec_size(p
, BRW_EXECUTE_16
);
352 brw_set_default_compression_control(p
, BRW_COMPRESSION_COMPRESSED
);
354 retype(payload
, BRW_REGISTER_TYPE_UD
),
355 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
356 brw_pop_insn_state(p
);
358 if (inst
->target
> 0 && key
->replicate_alpha
) {
359 /* Set "Source0 Alpha Present to RenderTarget" bit in message
363 vec1(retype(payload
, BRW_REGISTER_TYPE_UD
)),
364 vec1(retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
)),
365 brw_imm_ud(0x1 << 11));
368 if (inst
->target
> 0) {
369 /* Set the render target index for choosing BLEND_STATE. */
370 brw_MOV(p
, retype(vec1(suboffset(payload
, 2)),
371 BRW_REGISTER_TYPE_UD
),
372 brw_imm_ud(inst
->target
));
375 /* Set computes stencil to render target */
376 if (prog_data
->computed_stencil
) {
378 vec1(retype(payload
, BRW_REGISTER_TYPE_UD
)),
379 vec1(retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
)),
380 brw_imm_ud(0x1 << 14));
383 implied_header
= brw_null_reg();
385 implied_header
= retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW
);
388 brw_pop_insn_state(p
);
390 implied_header
= brw_null_reg();
393 if (!runtime_check_aads_emit
) {
394 fire_fb_write(inst
, payload
, implied_header
, inst
->mlen
);
396 /* This can only happen in gen < 6 */
397 assert(devinfo
->gen
< 6);
399 struct brw_reg v1_null_ud
= vec1(retype(brw_null_reg(), BRW_REGISTER_TYPE_UD
));
401 /* Check runtime bit to detect if we have to send AA data or not */
402 brw_push_insn_state(p
);
403 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
404 brw_set_default_exec_size(p
, BRW_EXECUTE_1
);
407 retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_UD
),
409 brw_inst_set_cond_modifier(p
->devinfo
, brw_last_inst
, BRW_CONDITIONAL_NZ
);
411 int jmp
= brw_JMPI(p
, brw_imm_ud(0), BRW_PREDICATE_NORMAL
) - p
->store
;
412 brw_pop_insn_state(p
);
414 /* Don't send AA data */
415 fire_fb_write(inst
, offset(payload
, 1), implied_header
, inst
->mlen
-1);
417 brw_land_fwd_jump(p
, jmp
);
418 fire_fb_write(inst
, payload
, implied_header
, inst
->mlen
);
423 fs_generator::generate_fb_read(fs_inst
*inst
, struct brw_reg dst
,
424 struct brw_reg payload
)
426 assert(inst
->size_written
% REG_SIZE
== 0);
427 struct brw_wm_prog_data
*prog_data
= brw_wm_prog_data(this->prog_data
);
428 /* We assume that render targets start at binding table index 0. */
429 const unsigned surf_index
= inst
->target
;
431 gen9_fb_READ(p
, dst
, payload
, surf_index
,
432 inst
->header_size
, inst
->size_written
/ REG_SIZE
,
433 prog_data
->persample_dispatch
);
435 brw_mark_surface_used(&prog_data
->base
, surf_index
);
439 fs_generator::generate_mov_indirect(fs_inst
*inst
,
442 struct brw_reg indirect_byte_offset
)
444 assert(indirect_byte_offset
.type
== BRW_REGISTER_TYPE_UD
);
445 assert(indirect_byte_offset
.file
== BRW_GENERAL_REGISTER_FILE
);
446 assert(!reg
.abs
&& !reg
.negate
);
447 assert(reg
.type
== dst
.type
);
449 unsigned imm_byte_offset
= reg
.nr
* REG_SIZE
+ reg
.subnr
;
451 if (indirect_byte_offset
.file
== BRW_IMMEDIATE_VALUE
) {
452 imm_byte_offset
+= indirect_byte_offset
.ud
;
454 reg
.nr
= imm_byte_offset
/ REG_SIZE
;
455 reg
.subnr
= imm_byte_offset
% REG_SIZE
;
456 brw_MOV(p
, dst
, reg
);
458 /* Prior to Broadwell, there are only 8 address registers. */
459 assert(inst
->exec_size
<= 8 || devinfo
->gen
>= 8);
461 /* We use VxH indirect addressing, clobbering a0.0 through a0.7. */
462 struct brw_reg addr
= vec8(brw_address_reg(0));
464 /* The destination stride of an instruction (in bytes) must be greater
465 * than or equal to the size of the rest of the instruction. Since the
466 * address register is of type UW, we can't use a D-type instruction.
467 * In order to get around this, re retype to UW and use a stride.
469 indirect_byte_offset
=
470 retype(spread(indirect_byte_offset
, 2), BRW_REGISTER_TYPE_UW
);
472 /* There are a number of reasons why we don't use the base offset here.
473 * One reason is that the field is only 9 bits which means we can only
474 * use it to access the first 16 GRFs. Also, from the Haswell PRM
475 * section "Register Region Restrictions":
477 * "The lower bits of the AddressImmediate must not overflow to
478 * change the register address. The lower 5 bits of Address
479 * Immediate when added to lower 5 bits of address register gives
480 * the sub-register offset. The upper bits of Address Immediate
481 * when added to upper bits of address register gives the register
482 * address. Any overflow from sub-register offset is dropped."
484 * Since the indirect may cause us to cross a register boundary, this
485 * makes the base offset almost useless. We could try and do something
486 * clever where we use a actual base offset if base_offset % 32 == 0 but
487 * that would mean we were generating different code depending on the
488 * base offset. Instead, for the sake of consistency, we'll just do the
489 * add ourselves. This restriction is only listed in the Haswell PRM
490 * but empirical testing indicates that it applies on all older
491 * generations and is lifted on Broadwell.
493 * In the end, while base_offset is nice to look at in the generated
494 * code, using it saves us 0 instructions and would require quite a bit
495 * of case-by-case work. It's just not worth it.
497 brw_ADD(p
, addr
, indirect_byte_offset
, brw_imm_uw(imm_byte_offset
));
499 if (type_sz(reg
.type
) > 4 &&
500 ((devinfo
->gen
== 7 && !devinfo
->is_haswell
) ||
501 devinfo
->is_cherryview
|| gen_device_info_is_9lp(devinfo
))) {
502 /* IVB has an issue (which we found empirically) where it reads two
503 * address register components per channel for indirectly addressed
506 * From the Cherryview PRM Vol 7. "Register Region Restrictions":
508 * "When source or destination datatype is 64b or operation is
509 * integer DWord multiply, indirect addressing must not be used."
511 * To work around both of these, we do two integer MOVs insead of one
512 * 64-bit MOV. Because no double value should ever cross a register
513 * boundary, it's safe to use the immediate offset in the indirect
514 * here to handle adding 4 bytes to the offset and avoid the extra
515 * ADD to the register file.
517 brw_MOV(p
, subscript(dst
, BRW_REGISTER_TYPE_D
, 0),
518 retype(brw_VxH_indirect(0, 0), BRW_REGISTER_TYPE_D
));
519 brw_MOV(p
, subscript(dst
, BRW_REGISTER_TYPE_D
, 1),
520 retype(brw_VxH_indirect(0, 4), BRW_REGISTER_TYPE_D
));
522 struct brw_reg ind_src
= brw_VxH_indirect(0, 0);
524 brw_inst
*mov
= brw_MOV(p
, dst
, retype(ind_src
, reg
.type
));
526 if (devinfo
->gen
== 6 && dst
.file
== BRW_MESSAGE_REGISTER_FILE
&&
527 !inst
->get_next()->is_tail_sentinel() &&
528 ((fs_inst
*)inst
->get_next())->mlen
> 0) {
529 /* From the Sandybridge PRM:
531 * "[Errata: DevSNB(SNB)] If MRF register is updated by any
532 * instruction that “indexed/indirect” source AND is followed
533 * by a send, the instruction requires a “Switch”. This is to
534 * avoid race condition where send may dispatch before MRF is
537 brw_inst_set_thread_control(devinfo
, mov
, BRW_THREAD_SWITCH
);
544 fs_generator::generate_shuffle(fs_inst
*inst
,
549 /* Ivy bridge has some strange behavior that makes this a real pain to
550 * implement for 64-bit values so we just don't bother.
552 assert(devinfo
->gen
>= 8 || devinfo
->is_haswell
|| type_sz(src
.type
) <= 4);
554 /* Because we're using the address register, we're limited to 8-wide
555 * execution on gen7. On gen8, we're limited to 16-wide by the address
556 * register file and 8-wide for 64-bit types. We could try and make this
557 * instruction splittable higher up in the compiler but that gets weird
558 * because it reads all of the channels regardless of execution size. It's
559 * easier just to split it here.
561 const unsigned lower_width
=
562 (devinfo
->gen
<= 7 || type_sz(src
.type
) > 4) ?
563 8 : MIN2(16, inst
->exec_size
);
565 brw_set_default_exec_size(p
, cvt(lower_width
) - 1);
566 for (unsigned group
= 0; group
< inst
->exec_size
; group
+= lower_width
) {
567 brw_set_default_group(p
, group
);
569 if ((src
.vstride
== 0 && src
.hstride
== 0) ||
570 idx
.file
== BRW_IMMEDIATE_VALUE
) {
571 /* Trivial, the source is already uniform or the index is a constant.
572 * We will typically not get here if the optimizer is doing its job,
573 * but asserting would be mean.
575 const unsigned i
= idx
.file
== BRW_IMMEDIATE_VALUE
? idx
.ud
: 0;
576 brw_MOV(p
, suboffset(dst
, group
), stride(suboffset(src
, i
), 0, 1, 0));
578 /* We use VxH indirect addressing, clobbering a0.0 through a0.7. */
579 struct brw_reg addr
= vec8(brw_address_reg(0));
581 struct brw_reg group_idx
= suboffset(idx
, group
);
583 if (lower_width
== 8 && group_idx
.width
== BRW_WIDTH_16
) {
584 /* Things get grumpy if the register is too wide. */
589 assert(type_sz(group_idx
.type
) <= 4);
590 if (type_sz(group_idx
.type
) == 4) {
591 /* The destination stride of an instruction (in bytes) must be
592 * greater than or equal to the size of the rest of the
593 * instruction. Since the address register is of type UW, we
594 * can't use a D-type instruction. In order to get around this,
595 * re retype to UW and use a stride.
597 group_idx
= retype(spread(group_idx
, 2), BRW_REGISTER_TYPE_W
);
600 /* Take into account the component size and horizontal stride. */
601 assert(src
.vstride
== src
.hstride
+ src
.width
);
602 brw_SHL(p
, addr
, group_idx
,
603 brw_imm_uw(_mesa_logbase2(type_sz(src
.type
)) +
606 /* Add on the register start offset */
607 brw_ADD(p
, addr
, addr
, brw_imm_uw(src
.nr
* REG_SIZE
+ src
.subnr
));
609 if (type_sz(src
.type
) > 4 &&
610 ((devinfo
->gen
== 7 && !devinfo
->is_haswell
) ||
611 devinfo
->is_cherryview
|| gen_device_info_is_9lp(devinfo
))) {
612 /* IVB has an issue (which we found empirically) where it reads
613 * two address register components per channel for indirectly
614 * addressed 64-bit sources.
616 * From the Cherryview PRM Vol 7. "Register Region Restrictions":
618 * "When source or destination datatype is 64b or operation is
619 * integer DWord multiply, indirect addressing must not be
622 * To work around both of these, we do two integer MOVs insead of
623 * one 64-bit MOV. Because no double value should ever cross a
624 * register boundary, it's safe to use the immediate offset in the
625 * indirect here to handle adding 4 bytes to the offset and avoid
626 * the extra ADD to the register file.
628 struct brw_reg gdst
= suboffset(dst
, group
);
629 struct brw_reg dst_d
= retype(spread(gdst
, 2),
630 BRW_REGISTER_TYPE_D
);
632 retype(brw_VxH_indirect(0, 0), BRW_REGISTER_TYPE_D
));
633 brw_MOV(p
, byte_offset(dst_d
, 4),
634 retype(brw_VxH_indirect(0, 4), BRW_REGISTER_TYPE_D
));
636 brw_MOV(p
, suboffset(dst
, group
),
637 retype(brw_VxH_indirect(0, 0), src
.type
));
644 fs_generator::generate_urb_read(fs_inst
*inst
,
646 struct brw_reg header
)
648 assert(inst
->size_written
% REG_SIZE
== 0);
649 assert(header
.file
== BRW_GENERAL_REGISTER_FILE
);
650 assert(header
.type
== BRW_REGISTER_TYPE_UD
);
652 brw_inst
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
653 brw_set_dest(p
, send
, retype(dst
, BRW_REGISTER_TYPE_UD
));
654 brw_set_src0(p
, send
, header
);
655 brw_set_src1(p
, send
, brw_imm_ud(0u));
657 brw_inst_set_sfid(p
->devinfo
, send
, BRW_SFID_URB
);
658 brw_inst_set_urb_opcode(p
->devinfo
, send
, GEN8_URB_OPCODE_SIMD8_READ
);
660 if (inst
->opcode
== SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
)
661 brw_inst_set_urb_per_slot_offset(p
->devinfo
, send
, true);
663 brw_inst_set_mlen(p
->devinfo
, send
, inst
->mlen
);
664 brw_inst_set_rlen(p
->devinfo
, send
, inst
->size_written
/ REG_SIZE
);
665 brw_inst_set_header_present(p
->devinfo
, send
, true);
666 brw_inst_set_urb_global_offset(p
->devinfo
, send
, inst
->offset
);
670 fs_generator::generate_urb_write(fs_inst
*inst
, struct brw_reg payload
)
674 /* WaClearTDRRegBeforeEOTForNonPS.
676 * WA: Clear tdr register before send EOT in all non-PS shader kernels
678 * mov(8) tdr0:ud 0x0:ud {NoMask}"
680 if (inst
->eot
&& p
->devinfo
->gen
== 10) {
681 brw_push_insn_state(p
);
682 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
683 brw_MOV(p
, brw_tdr_reg(), brw_imm_uw(0));
684 brw_pop_insn_state(p
);
687 insn
= brw_next_insn(p
, BRW_OPCODE_SEND
);
689 brw_set_dest(p
, insn
, brw_null_reg());
690 brw_set_src0(p
, insn
, payload
);
691 brw_set_src1(p
, insn
, brw_imm_d(0));
693 brw_inst_set_sfid(p
->devinfo
, insn
, BRW_SFID_URB
);
694 brw_inst_set_urb_opcode(p
->devinfo
, insn
, GEN8_URB_OPCODE_SIMD8_WRITE
);
696 if (inst
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
||
697 inst
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
)
698 brw_inst_set_urb_per_slot_offset(p
->devinfo
, insn
, true);
700 if (inst
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
||
701 inst
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
)
702 brw_inst_set_urb_channel_mask_present(p
->devinfo
, insn
, true);
704 brw_inst_set_mlen(p
->devinfo
, insn
, inst
->mlen
);
705 brw_inst_set_rlen(p
->devinfo
, insn
, 0);
706 brw_inst_set_eot(p
->devinfo
, insn
, inst
->eot
);
707 brw_inst_set_header_present(p
->devinfo
, insn
, true);
708 brw_inst_set_urb_global_offset(p
->devinfo
, insn
, inst
->offset
);
712 fs_generator::generate_cs_terminate(fs_inst
*inst
, struct brw_reg payload
)
714 struct brw_inst
*insn
;
716 insn
= brw_next_insn(p
, BRW_OPCODE_SEND
);
718 brw_set_dest(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_UW
));
719 brw_set_src0(p
, insn
, retype(payload
, BRW_REGISTER_TYPE_UW
));
720 brw_set_src1(p
, insn
, brw_imm_d(0));
722 /* Terminate a compute shader by sending a message to the thread spawner.
724 brw_inst_set_sfid(devinfo
, insn
, BRW_SFID_THREAD_SPAWNER
);
725 brw_inst_set_mlen(devinfo
, insn
, 1);
726 brw_inst_set_rlen(devinfo
, insn
, 0);
727 brw_inst_set_eot(devinfo
, insn
, inst
->eot
);
728 brw_inst_set_header_present(devinfo
, insn
, false);
730 brw_inst_set_ts_opcode(devinfo
, insn
, 0); /* Dereference resource */
731 brw_inst_set_ts_request_type(devinfo
, insn
, 0); /* Root thread */
733 /* Note that even though the thread has a URB resource associated with it,
734 * we set the "do not dereference URB" bit, because the URB resource is
735 * managed by the fixed-function unit, so it will free it automatically.
737 brw_inst_set_ts_resource_select(devinfo
, insn
, 1); /* Do not dereference URB */
739 brw_inst_set_mask_control(devinfo
, insn
, BRW_MASK_DISABLE
);
743 fs_generator::generate_barrier(fs_inst
*, struct brw_reg src
)
750 fs_generator::generate_linterp(fs_inst
*inst
,
751 struct brw_reg dst
, struct brw_reg
*src
)
755 * -----------------------------------
756 * | src1+0 | src1+1 | src1+2 | src1+3 |
757 * |-----------------------------------|
758 * |(x0, x1)|(y0, y1)|(x2, x3)|(y2, y3)|
759 * -----------------------------------
761 * but for the LINE/MAC pair, the LINE reads Xs and the MAC reads Ys:
763 * -----------------------------------
764 * | src1+0 | src1+1 | src1+2 | src1+3 |
765 * |-----------------------------------|
766 * |(x0, x1)|(y0, y1)| | | in SIMD8
767 * |-----------------------------------|
768 * |(x0, x1)|(x2, x3)|(y0, y1)|(y2, y3)| in SIMD16
769 * -----------------------------------
771 * See also: emit_interpolation_setup_gen4().
773 struct brw_reg delta_x
= src
[0];
774 struct brw_reg delta_y
= offset(src
[0], inst
->exec_size
/ 8);
775 struct brw_reg interp
= src
[1];
778 if (devinfo
->gen
>= 11) {
779 struct brw_reg acc
= retype(brw_acc_reg(8), BRW_REGISTER_TYPE_NF
);
780 struct brw_reg dwP
= suboffset(interp
, 0);
781 struct brw_reg dwQ
= suboffset(interp
, 1);
782 struct brw_reg dwR
= suboffset(interp
, 3);
784 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
786 if (inst
->exec_size
== 8) {
787 i
[0] = brw_MAD(p
, acc
, dwR
, offset(delta_x
, 0), dwP
);
788 i
[1] = brw_MAD(p
, offset(dst
, 0), acc
, offset(delta_y
, 0), dwQ
);
790 brw_inst_set_cond_modifier(p
->devinfo
, i
[1], inst
->conditional_mod
);
792 /* brw_set_default_saturate() is called before emitting instructions,
793 * so the saturate bit is set in each instruction, so we need to unset
794 * it on the first instruction of each pair.
796 brw_inst_set_saturate(p
->devinfo
, i
[0], false);
798 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
799 i
[0] = brw_MAD(p
, acc
, dwR
, offset(delta_x
, 0), dwP
);
800 i
[1] = brw_MAD(p
, offset(dst
, 0), acc
, offset(delta_x
, 1), dwQ
);
802 brw_set_default_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
803 i
[2] = brw_MAD(p
, acc
, dwR
, offset(delta_y
, 0), dwP
);
804 i
[3] = brw_MAD(p
, offset(dst
, 1), acc
, offset(delta_y
, 1), dwQ
);
806 brw_set_default_compression_control(p
, BRW_COMPRESSION_COMPRESSED
);
808 brw_inst_set_cond_modifier(p
->devinfo
, i
[1], inst
->conditional_mod
);
809 brw_inst_set_cond_modifier(p
->devinfo
, i
[3], inst
->conditional_mod
);
811 /* brw_set_default_saturate() is called before emitting instructions,
812 * so the saturate bit is set in each instruction, so we need to unset
813 * it on the first instruction of each pair.
815 brw_inst_set_saturate(p
->devinfo
, i
[0], false);
816 brw_inst_set_saturate(p
->devinfo
, i
[2], false);
820 } else if (devinfo
->has_pln
) {
821 /* From the Sandy Bridge PRM Vol. 4, Pt. 2, Section 8.3.53, "Plane":
823 * "[DevSNB]:<src1> must be even register aligned.
825 * This restriction is lifted on Ivy Bridge.
827 assert(devinfo
->gen
>= 7 || (delta_x
.nr
& 1) == 0);
828 brw_PLN(p
, dst
, interp
, delta_x
);
832 i
[0] = brw_LINE(p
, brw_null_reg(), interp
, delta_x
);
833 i
[1] = brw_MAC(p
, dst
, suboffset(interp
, 1), delta_y
);
835 brw_inst_set_cond_modifier(p
->devinfo
, i
[1], inst
->conditional_mod
);
837 /* brw_set_default_saturate() is called before emitting instructions, so
838 * the saturate bit is set in each instruction, so we need to unset it on
839 * the first instruction.
841 brw_inst_set_saturate(p
->devinfo
, i
[0], false);
848 fs_generator::generate_get_buffer_size(fs_inst
*inst
,
851 struct brw_reg surf_index
)
853 assert(devinfo
->gen
>= 7);
854 assert(surf_index
.file
== BRW_IMMEDIATE_VALUE
);
859 switch (inst
->exec_size
) {
861 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD8
;
864 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
867 unreachable("Invalid width for texture instruction");
870 if (simd_mode
== BRW_SAMPLER_SIMD_MODE_SIMD16
) {
876 retype(dst
, BRW_REGISTER_TYPE_UW
),
881 GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO
,
882 rlen
, /* response length */
884 inst
->header_size
> 0,
886 BRW_SAMPLER_RETURN_FORMAT_SINT32
);
888 brw_mark_surface_used(prog_data
, surf_index
.ud
);
892 fs_generator::generate_tex(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
,
893 struct brw_reg surface_index
,
894 struct brw_reg sampler_index
)
896 assert(inst
->size_written
% REG_SIZE
== 0);
899 uint32_t return_format
;
900 bool is_combined_send
= inst
->eot
;
903 case BRW_REGISTER_TYPE_D
:
904 return_format
= BRW_SAMPLER_RETURN_FORMAT_SINT32
;
906 case BRW_REGISTER_TYPE_UD
:
907 return_format
= BRW_SAMPLER_RETURN_FORMAT_UINT32
;
910 return_format
= BRW_SAMPLER_RETURN_FORMAT_FLOAT32
;
914 /* Stomp the resinfo output type to UINT32. On gens 4-5, the output type
915 * is set as part of the message descriptor. On gen4, the PRM seems to
916 * allow UINT32 and FLOAT32 (i965 PRM, Vol. 4 Section 4.8.1.1), but on
917 * later gens UINT32 is required. Once you hit Sandy Bridge, the bit is
918 * gone from the message descriptor entirely and you just get UINT32 all
919 * the time regasrdless. Since we can really only do non-UINT32 on gen4,
920 * just stomp it to UINT32 all the time.
922 if (inst
->opcode
== SHADER_OPCODE_TXS
)
923 return_format
= BRW_SAMPLER_RETURN_FORMAT_UINT32
;
925 switch (inst
->exec_size
) {
927 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD8
;
930 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
933 unreachable("Invalid width for texture instruction");
936 if (devinfo
->gen
>= 5) {
937 switch (inst
->opcode
) {
938 case SHADER_OPCODE_TEX
:
939 if (inst
->shadow_compare
) {
940 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_COMPARE
;
942 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE
;
946 if (inst
->shadow_compare
) {
947 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS_COMPARE
;
949 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS
;
952 case SHADER_OPCODE_TXL
:
953 if (inst
->shadow_compare
) {
954 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE
;
956 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LOD
;
959 case SHADER_OPCODE_TXL_LZ
:
960 assert(devinfo
->gen
>= 9);
961 if (inst
->shadow_compare
) {
962 msg_type
= GEN9_SAMPLER_MESSAGE_SAMPLE_C_LZ
;
964 msg_type
= GEN9_SAMPLER_MESSAGE_SAMPLE_LZ
;
967 case SHADER_OPCODE_TXS
:
968 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO
;
970 case SHADER_OPCODE_TXD
:
971 if (inst
->shadow_compare
) {
972 /* Gen7.5+. Otherwise, lowered in NIR */
973 assert(devinfo
->gen
>= 8 || devinfo
->is_haswell
);
974 msg_type
= HSW_SAMPLER_MESSAGE_SAMPLE_DERIV_COMPARE
;
976 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS
;
979 case SHADER_OPCODE_TXF
:
980 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
982 case SHADER_OPCODE_TXF_LZ
:
983 assert(devinfo
->gen
>= 9);
984 msg_type
= GEN9_SAMPLER_MESSAGE_SAMPLE_LD_LZ
;
986 case SHADER_OPCODE_TXF_CMS_W
:
987 assert(devinfo
->gen
>= 9);
988 msg_type
= GEN9_SAMPLER_MESSAGE_SAMPLE_LD2DMS_W
;
990 case SHADER_OPCODE_TXF_CMS
:
991 if (devinfo
->gen
>= 7)
992 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DMS
;
994 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
996 case SHADER_OPCODE_TXF_UMS
:
997 assert(devinfo
->gen
>= 7);
998 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DSS
;
1000 case SHADER_OPCODE_TXF_MCS
:
1001 assert(devinfo
->gen
>= 7);
1002 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_LD_MCS
;
1004 case SHADER_OPCODE_LOD
:
1005 msg_type
= GEN5_SAMPLER_MESSAGE_LOD
;
1007 case SHADER_OPCODE_TG4
:
1008 if (inst
->shadow_compare
) {
1009 assert(devinfo
->gen
>= 7);
1010 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_C
;
1012 assert(devinfo
->gen
>= 6);
1013 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4
;
1016 case SHADER_OPCODE_TG4_OFFSET
:
1017 assert(devinfo
->gen
>= 7);
1018 if (inst
->shadow_compare
) {
1019 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO_C
;
1021 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO
;
1024 case SHADER_OPCODE_SAMPLEINFO
:
1025 msg_type
= GEN6_SAMPLER_MESSAGE_SAMPLE_SAMPLEINFO
;
1028 unreachable("not reached");
1031 switch (inst
->opcode
) {
1032 case SHADER_OPCODE_TEX
:
1033 /* Note that G45 and older determines shadow compare and dispatch width
1034 * from message length for most messages.
1036 if (inst
->exec_size
== 8) {
1037 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE
;
1038 if (inst
->shadow_compare
) {
1039 assert(inst
->mlen
== 6);
1041 assert(inst
->mlen
<= 4);
1044 if (inst
->shadow_compare
) {
1045 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_COMPARE
;
1046 assert(inst
->mlen
== 9);
1048 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE
;
1049 assert(inst
->mlen
<= 7 && inst
->mlen
% 2 == 1);
1054 if (inst
->shadow_compare
) {
1055 assert(inst
->exec_size
== 8);
1056 assert(inst
->mlen
== 6);
1057 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_BIAS_COMPARE
;
1059 assert(inst
->mlen
== 9);
1060 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_BIAS
;
1061 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1064 case SHADER_OPCODE_TXL
:
1065 if (inst
->shadow_compare
) {
1066 assert(inst
->exec_size
== 8);
1067 assert(inst
->mlen
== 6);
1068 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_LOD_COMPARE
;
1070 assert(inst
->mlen
== 9);
1071 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_LOD
;
1072 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1075 case SHADER_OPCODE_TXD
:
1076 /* There is no sample_d_c message; comparisons are done manually */
1077 assert(inst
->exec_size
== 8);
1078 assert(inst
->mlen
== 7 || inst
->mlen
== 10);
1079 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_GRADIENTS
;
1081 case SHADER_OPCODE_TXF
:
1082 assert(inst
->mlen
<= 9 && inst
->mlen
% 2 == 1);
1083 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_LD
;
1084 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1086 case SHADER_OPCODE_TXS
:
1087 assert(inst
->mlen
== 3);
1088 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_RESINFO
;
1089 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1092 unreachable("not reached");
1095 assert(msg_type
!= -1);
1097 if (simd_mode
== BRW_SAMPLER_SIMD_MODE_SIMD16
) {
1101 assert(devinfo
->gen
< 7 || inst
->header_size
== 0 ||
1102 src
.file
== BRW_GENERAL_REGISTER_FILE
);
1104 assert(sampler_index
.type
== BRW_REGISTER_TYPE_UD
);
1106 /* Load the message header if present. If there's a texture offset,
1107 * we need to set it up explicitly and load the offset bitfield.
1108 * Otherwise, we can use an implied move from g0 to the first message reg.
1110 if (inst
->header_size
!= 0 && devinfo
->gen
< 7) {
1111 if (devinfo
->gen
< 6 && !inst
->offset
) {
1112 /* Set up an implied move from g0 to the MRF. */
1113 src
= retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW
);
1115 assert(inst
->base_mrf
!= -1);
1116 struct brw_reg header_reg
= brw_message_reg(inst
->base_mrf
);
1118 brw_push_insn_state(p
);
1119 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
1120 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1121 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
1122 /* Explicitly set up the message header by copying g0 to the MRF. */
1123 brw_MOV(p
, header_reg
, brw_vec8_grf(0, 0));
1125 brw_set_default_exec_size(p
, BRW_EXECUTE_1
);
1127 /* Set the offset bits in DWord 2. */
1128 brw_MOV(p
, get_element_ud(header_reg
, 2),
1129 brw_imm_ud(inst
->offset
));
1132 brw_pop_insn_state(p
);
1136 uint32_t base_binding_table_index
= (inst
->opcode
== SHADER_OPCODE_TG4
||
1137 inst
->opcode
== SHADER_OPCODE_TG4_OFFSET
)
1138 ? prog_data
->binding_table
.gather_texture_start
1139 : prog_data
->binding_table
.texture_start
;
1141 if (surface_index
.file
== BRW_IMMEDIATE_VALUE
&&
1142 sampler_index
.file
== BRW_IMMEDIATE_VALUE
) {
1143 uint32_t surface
= surface_index
.ud
;
1144 uint32_t sampler
= sampler_index
.ud
;
1147 retype(dst
, BRW_REGISTER_TYPE_UW
),
1150 surface
+ base_binding_table_index
,
1153 inst
->size_written
/ REG_SIZE
,
1155 inst
->header_size
!= 0,
1159 brw_mark_surface_used(prog_data
, surface
+ base_binding_table_index
);
1161 /* Non-const sampler index */
1163 struct brw_reg addr
= vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD
));
1164 struct brw_reg surface_reg
= vec1(retype(surface_index
, BRW_REGISTER_TYPE_UD
));
1165 struct brw_reg sampler_reg
= vec1(retype(sampler_index
, BRW_REGISTER_TYPE_UD
));
1167 brw_push_insn_state(p
);
1168 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1169 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
1170 brw_set_default_exec_size(p
, BRW_EXECUTE_1
);
1172 if (brw_regs_equal(&surface_reg
, &sampler_reg
)) {
1173 brw_MUL(p
, addr
, sampler_reg
, brw_imm_uw(0x101));
1175 if (sampler_reg
.file
== BRW_IMMEDIATE_VALUE
) {
1176 brw_OR(p
, addr
, surface_reg
, brw_imm_ud(sampler_reg
.ud
<< 8));
1178 brw_SHL(p
, addr
, sampler_reg
, brw_imm_ud(8));
1179 brw_OR(p
, addr
, addr
, surface_reg
);
1182 if (base_binding_table_index
)
1183 brw_ADD(p
, addr
, addr
, brw_imm_ud(base_binding_table_index
));
1184 brw_AND(p
, addr
, addr
, brw_imm_ud(0xfff));
1186 brw_pop_insn_state(p
);
1188 /* dst = send(offset, a0.0 | <descriptor>) */
1189 brw_inst
*insn
= brw_send_indirect_message(
1190 p
, BRW_SFID_SAMPLER
, dst
, src
, addr
);
1191 brw_set_sampler_message(p
, insn
,
1195 inst
->size_written
/ REG_SIZE
,
1196 inst
->mlen
/* mlen */,
1197 inst
->header_size
!= 0 /* header */,
1201 /* visitor knows more than we do about the surface limit required,
1202 * so has already done marking.
1206 if (is_combined_send
) {
1207 brw_inst_set_eot(p
->devinfo
, brw_last_inst
, true);
1208 brw_inst_set_opcode(p
->devinfo
, brw_last_inst
, BRW_OPCODE_SENDC
);
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
= src
;
1258 struct brw_reg src1
= src
;
1260 src0
.subnr
= sizeof(float);
1261 src0
.vstride
= vstride
;
1263 src0
.hstride
= BRW_HORIZONTAL_STRIDE_0
;
1264 src1
.vstride
= vstride
;
1266 src1
.hstride
= BRW_HORIZONTAL_STRIDE_0
;
1268 brw_ADD(p
, dst
, src0
, negate(src1
));
1271 /* The negate_value boolean is used to negate the derivative computation for
1272 * FBOs, since they place the origin at the upper left instead of the lower
1276 fs_generator::generate_ddy(const fs_inst
*inst
,
1277 struct brw_reg dst
, struct brw_reg src
)
1279 if (inst
->opcode
== FS_OPCODE_DDY_FINE
) {
1280 /* produce accurate derivatives */
1281 if (devinfo
->gen
>= 11) {
1282 src
= stride(src
, 0, 2, 1);
1283 struct brw_reg src_0
= byte_offset(src
, 0 * sizeof(float));
1284 struct brw_reg src_2
= byte_offset(src
, 2 * sizeof(float));
1285 struct brw_reg src_4
= byte_offset(src
, 4 * sizeof(float));
1286 struct brw_reg src_6
= byte_offset(src
, 6 * sizeof(float));
1287 struct brw_reg src_8
= byte_offset(src
, 8 * sizeof(float));
1288 struct brw_reg src_10
= byte_offset(src
, 10 * sizeof(float));
1289 struct brw_reg src_12
= byte_offset(src
, 12 * sizeof(float));
1290 struct brw_reg src_14
= byte_offset(src
, 14 * sizeof(float));
1292 struct brw_reg dst_0
= byte_offset(dst
, 0 * sizeof(float));
1293 struct brw_reg dst_4
= byte_offset(dst
, 4 * sizeof(float));
1294 struct brw_reg dst_8
= byte_offset(dst
, 8 * sizeof(float));
1295 struct brw_reg dst_12
= byte_offset(dst
, 12 * sizeof(float));
1297 brw_push_insn_state(p
);
1298 brw_set_default_exec_size(p
, BRW_EXECUTE_4
);
1300 brw_ADD(p
, dst_0
, negate(src_0
), src_2
);
1301 brw_ADD(p
, dst_4
, negate(src_4
), src_6
);
1303 if (inst
->exec_size
== 16) {
1304 brw_ADD(p
, dst_8
, negate(src_8
), src_10
);
1305 brw_ADD(p
, dst_12
, negate(src_12
), src_14
);
1308 brw_pop_insn_state(p
);
1310 struct brw_reg src0
= stride(src
, 4, 4, 1);
1311 struct brw_reg src1
= stride(src
, 4, 4, 1);
1312 src0
.swizzle
= BRW_SWIZZLE_XYXY
;
1313 src1
.swizzle
= BRW_SWIZZLE_ZWZW
;
1315 brw_push_insn_state(p
);
1316 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1317 brw_ADD(p
, dst
, negate(src0
), src1
);
1318 brw_pop_insn_state(p
);
1321 /* replicate the derivative at the top-left pixel to other pixels */
1322 struct brw_reg src0
= stride(src
, 4, 4, 0);
1323 struct brw_reg src1
= stride(src
, 4, 4, 0);
1324 src0
.subnr
= 0 * sizeof(float);
1325 src1
.subnr
= 2 * sizeof(float);
1327 brw_ADD(p
, dst
, negate(src0
), src1
);
1332 fs_generator::generate_discard_jump(fs_inst
*)
1334 assert(devinfo
->gen
>= 6);
1336 /* This HALT will be patched up at FB write time to point UIP at the end of
1337 * the program, and at brw_uip_jip() JIP will be set to the end of the
1338 * current block (or the program).
1340 this->discard_halt_patches
.push_tail(new(mem_ctx
) ip_record(p
->nr_insn
));
1345 fs_generator::generate_scratch_write(fs_inst
*inst
, struct brw_reg src
)
1347 /* The 32-wide messages only respect the first 16-wide half of the channel
1348 * enable signals which are replicated identically for the second group of
1349 * 16 channels, so we cannot use them unless the write is marked
1350 * force_writemask_all.
1352 const unsigned lower_size
= inst
->force_writemask_all
? inst
->exec_size
:
1353 MIN2(16, inst
->exec_size
);
1354 const unsigned block_size
= 4 * lower_size
/ REG_SIZE
;
1355 assert(inst
->mlen
!= 0);
1357 brw_push_insn_state(p
);
1358 brw_set_default_exec_size(p
, cvt(lower_size
) - 1);
1359 brw_set_default_compression(p
, lower_size
> 8);
1361 for (unsigned i
= 0; i
< inst
->exec_size
/ lower_size
; i
++) {
1362 brw_set_default_group(p
, inst
->group
+ lower_size
* i
);
1364 brw_MOV(p
, brw_uvec_mrf(lower_size
, inst
->base_mrf
+ 1, 0),
1365 retype(offset(src
, block_size
* i
), BRW_REGISTER_TYPE_UD
));
1367 brw_oword_block_write_scratch(p
, brw_message_reg(inst
->base_mrf
),
1369 inst
->offset
+ block_size
* REG_SIZE
* i
);
1372 brw_pop_insn_state(p
);
1376 fs_generator::generate_scratch_read(fs_inst
*inst
, struct brw_reg dst
)
1378 assert(inst
->exec_size
<= 16 || inst
->force_writemask_all
);
1379 assert(inst
->mlen
!= 0);
1381 brw_oword_block_read_scratch(p
, dst
, brw_message_reg(inst
->base_mrf
),
1382 inst
->exec_size
/ 8, inst
->offset
);
1386 fs_generator::generate_scratch_read_gen7(fs_inst
*inst
, struct brw_reg dst
)
1388 assert(inst
->exec_size
<= 16 || inst
->force_writemask_all
);
1390 gen7_block_read_scratch(p
, dst
, inst
->exec_size
/ 8, inst
->offset
);
1394 fs_generator::generate_uniform_pull_constant_load(fs_inst
*inst
,
1396 struct brw_reg index
,
1397 struct brw_reg offset
)
1399 assert(type_sz(dst
.type
) == 4);
1400 assert(inst
->mlen
!= 0);
1402 assert(index
.file
== BRW_IMMEDIATE_VALUE
&&
1403 index
.type
== BRW_REGISTER_TYPE_UD
);
1404 uint32_t surf_index
= index
.ud
;
1406 assert(offset
.file
== BRW_IMMEDIATE_VALUE
&&
1407 offset
.type
== BRW_REGISTER_TYPE_UD
);
1408 uint32_t read_offset
= offset
.ud
;
1410 brw_oword_block_read(p
, dst
, brw_message_reg(inst
->base_mrf
),
1411 read_offset
, surf_index
);
1415 fs_generator::generate_uniform_pull_constant_load_gen7(fs_inst
*inst
,
1417 struct brw_reg index
,
1418 struct brw_reg payload
)
1420 assert(index
.type
== BRW_REGISTER_TYPE_UD
);
1421 assert(payload
.file
== BRW_GENERAL_REGISTER_FILE
);
1422 assert(type_sz(dst
.type
) == 4);
1424 if (index
.file
== BRW_IMMEDIATE_VALUE
) {
1425 const uint32_t surf_index
= index
.ud
;
1427 brw_push_insn_state(p
);
1428 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1429 brw_inst
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
1430 brw_pop_insn_state(p
);
1432 brw_set_dest(p
, send
, retype(dst
, BRW_REGISTER_TYPE_UD
));
1433 brw_set_src0(p
, send
, retype(payload
, BRW_REGISTER_TYPE_UD
));
1434 brw_set_dp_read_message(p
, send
, surf_index
,
1435 BRW_DATAPORT_OWORD_BLOCK_DWORDS(inst
->exec_size
),
1436 GEN7_DATAPORT_DC_OWORD_BLOCK_READ
,
1437 GEN6_SFID_DATAPORT_CONSTANT_CACHE
,
1440 DIV_ROUND_UP(inst
->size_written
, REG_SIZE
));
1443 struct brw_reg addr
= vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD
));
1445 brw_push_insn_state(p
);
1446 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1448 /* a0.0 = surf_index & 0xff */
1449 brw_inst
*insn_and
= brw_next_insn(p
, BRW_OPCODE_AND
);
1450 brw_inst_set_exec_size(p
->devinfo
, insn_and
, BRW_EXECUTE_1
);
1451 brw_set_dest(p
, insn_and
, addr
);
1452 brw_set_src0(p
, insn_and
, vec1(retype(index
, BRW_REGISTER_TYPE_UD
)));
1453 brw_set_src1(p
, insn_and
, brw_imm_ud(0x0ff));
1455 /* dst = send(payload, a0.0 | <descriptor>) */
1456 brw_inst
*insn
= brw_send_indirect_message(
1457 p
, GEN6_SFID_DATAPORT_CONSTANT_CACHE
,
1458 retype(dst
, BRW_REGISTER_TYPE_UD
),
1459 retype(payload
, BRW_REGISTER_TYPE_UD
), addr
);
1460 brw_set_dp_read_message(p
, insn
, 0 /* surface */,
1461 BRW_DATAPORT_OWORD_BLOCK_DWORDS(inst
->exec_size
),
1462 GEN7_DATAPORT_DC_OWORD_BLOCK_READ
,
1463 GEN6_SFID_DATAPORT_CONSTANT_CACHE
,
1466 DIV_ROUND_UP(inst
->size_written
, REG_SIZE
));
1468 brw_pop_insn_state(p
);
1473 fs_generator::generate_varying_pull_constant_load_gen4(fs_inst
*inst
,
1475 struct brw_reg index
)
1477 assert(devinfo
->gen
< 7); /* Should use the gen7 variant. */
1478 assert(inst
->header_size
!= 0);
1481 assert(index
.file
== BRW_IMMEDIATE_VALUE
&&
1482 index
.type
== BRW_REGISTER_TYPE_UD
);
1483 uint32_t surf_index
= index
.ud
;
1485 uint32_t simd_mode
, rlen
, msg_type
;
1486 if (inst
->exec_size
== 16) {
1487 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1490 assert(inst
->exec_size
== 8);
1491 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD8
;
1495 if (devinfo
->gen
>= 5)
1496 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
1498 /* We always use the SIMD16 message so that we only have to load U, and
1501 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_LD
;
1502 assert(inst
->mlen
== 3);
1503 assert(inst
->size_written
== 8 * REG_SIZE
);
1505 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1508 struct brw_reg header
= brw_vec8_grf(0, 0);
1509 gen6_resolve_implied_move(p
, &header
, inst
->base_mrf
);
1511 brw_inst
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
1512 brw_inst_set_compression(devinfo
, send
, false);
1513 brw_set_dest(p
, send
, retype(dst
, BRW_REGISTER_TYPE_UW
));
1514 brw_set_src0(p
, send
, header
);
1515 if (devinfo
->gen
< 6)
1516 brw_inst_set_base_mrf(p
->devinfo
, send
, inst
->base_mrf
);
1518 /* Our surface is set up as floats, regardless of what actual data is
1521 uint32_t return_format
= BRW_SAMPLER_RETURN_FORMAT_FLOAT32
;
1522 brw_set_sampler_message(p
, send
,
1524 0, /* sampler (unused) */
1528 inst
->header_size
!= 0,
1534 fs_generator::generate_varying_pull_constant_load_gen7(fs_inst
*inst
,
1536 struct brw_reg index
,
1537 struct brw_reg offset
)
1539 assert(devinfo
->gen
>= 7);
1540 /* Varying-offset pull constant loads are treated as a normal expression on
1541 * gen7, so the fact that it's a send message is hidden at the IR level.
1543 assert(inst
->header_size
== 0);
1544 assert(!inst
->mlen
);
1545 assert(index
.type
== BRW_REGISTER_TYPE_UD
);
1547 uint32_t simd_mode
, rlen
, mlen
;
1548 if (inst
->exec_size
== 16) {
1551 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1553 assert(inst
->exec_size
== 8);
1556 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD8
;
1559 if (index
.file
== BRW_IMMEDIATE_VALUE
) {
1561 uint32_t surf_index
= index
.ud
;
1563 brw_inst
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
1564 brw_set_dest(p
, send
, retype(dst
, BRW_REGISTER_TYPE_UW
));
1565 brw_set_src0(p
, send
, offset
);
1566 brw_set_sampler_message(p
, send
,
1568 0, /* LD message ignores sampler unit */
1569 GEN5_SAMPLER_MESSAGE_SAMPLE_LD
,
1572 false, /* no header */
1578 struct brw_reg addr
= vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD
));
1580 brw_push_insn_state(p
);
1581 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1583 /* a0.0 = surf_index & 0xff */
1584 brw_inst
*insn_and
= brw_next_insn(p
, BRW_OPCODE_AND
);
1585 brw_inst_set_exec_size(p
->devinfo
, insn_and
, BRW_EXECUTE_1
);
1586 brw_set_dest(p
, insn_and
, addr
);
1587 brw_set_src0(p
, insn_and
, vec1(retype(index
, BRW_REGISTER_TYPE_UD
)));
1588 brw_set_src1(p
, insn_and
, brw_imm_ud(0x0ff));
1590 brw_pop_insn_state(p
);
1592 /* dst = send(offset, a0.0 | <descriptor>) */
1593 brw_inst
*insn
= brw_send_indirect_message(
1594 p
, BRW_SFID_SAMPLER
, retype(dst
, BRW_REGISTER_TYPE_UW
),
1596 brw_set_sampler_message(p
, insn
,
1599 GEN5_SAMPLER_MESSAGE_SAMPLE_LD
,
1609 * Cause the current pixel/sample mask (from R1.7 bits 15:0) to be transferred
1610 * into the flags register (f0.0).
1612 * Used only on Gen6 and above.
1615 fs_generator::generate_mov_dispatch_to_flags(fs_inst
*inst
)
1617 struct brw_reg flags
= brw_flag_subreg(inst
->flag_subreg
);
1618 struct brw_reg dispatch_mask
;
1620 if (devinfo
->gen
>= 6)
1621 dispatch_mask
= retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW
);
1623 dispatch_mask
= retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW
);
1625 brw_push_insn_state(p
);
1626 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1627 brw_set_default_exec_size(p
, BRW_EXECUTE_1
);
1628 brw_MOV(p
, flags
, dispatch_mask
);
1629 brw_pop_insn_state(p
);
1633 fs_generator::generate_pixel_interpolator_query(fs_inst
*inst
,
1636 struct brw_reg msg_data
,
1639 assert(inst
->size_written
% REG_SIZE
== 0);
1640 assert(msg_data
.type
== BRW_REGISTER_TYPE_UD
);
1642 brw_pixel_interpolator_query(p
,
1643 retype(dst
, BRW_REGISTER_TYPE_UW
),
1645 inst
->pi_noperspective
,
1649 inst
->size_written
/ REG_SIZE
);
1652 /* Sets vstride=1, width=4, hstride=0 of register src1 during
1653 * the ADD instruction.
1656 fs_generator::generate_set_sample_id(fs_inst
*inst
,
1658 struct brw_reg src0
,
1659 struct brw_reg src1
)
1661 assert(dst
.type
== BRW_REGISTER_TYPE_D
||
1662 dst
.type
== BRW_REGISTER_TYPE_UD
);
1663 assert(src0
.type
== BRW_REGISTER_TYPE_D
||
1664 src0
.type
== BRW_REGISTER_TYPE_UD
);
1666 struct brw_reg reg
= stride(src1
, 1, 4, 0);
1667 if (devinfo
->gen
>= 8 || inst
->exec_size
== 8) {
1668 brw_ADD(p
, dst
, src0
, reg
);
1669 } else if (inst
->exec_size
== 16) {
1670 brw_push_insn_state(p
);
1671 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
1672 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
1673 brw_ADD(p
, firsthalf(dst
), firsthalf(src0
), reg
);
1674 brw_set_default_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
1675 brw_ADD(p
, sechalf(dst
), sechalf(src0
), suboffset(reg
, 2));
1676 brw_pop_insn_state(p
);
1681 fs_generator::generate_pack_half_2x16_split(fs_inst
*,
1686 assert(devinfo
->gen
>= 7);
1687 assert(dst
.type
== BRW_REGISTER_TYPE_UD
);
1688 assert(x
.type
== BRW_REGISTER_TYPE_F
);
1689 assert(y
.type
== BRW_REGISTER_TYPE_F
);
1691 /* From the Ivybridge PRM, Vol4, Part3, Section 6.27 f32to16:
1693 * Because this instruction does not have a 16-bit floating-point type,
1694 * the destination data type must be Word (W).
1696 * The destination must be DWord-aligned and specify a horizontal stride
1697 * (HorzStride) of 2. The 16-bit result is stored in the lower word of
1698 * each destination channel and the upper word is not modified.
1700 struct brw_reg dst_w
= spread(retype(dst
, BRW_REGISTER_TYPE_W
), 2);
1702 /* Give each 32-bit channel of dst the form below, where "." means
1706 brw_F32TO16(p
, dst_w
, y
);
1711 brw_SHL(p
, dst
, dst
, brw_imm_ud(16u));
1713 /* And, finally the form of packHalf2x16's output:
1716 brw_F32TO16(p
, dst_w
, x
);
1720 fs_generator::generate_unpack_half_2x16_split(fs_inst
*inst
,
1724 assert(devinfo
->gen
>= 7);
1725 assert(dst
.type
== BRW_REGISTER_TYPE_F
);
1726 assert(src
.type
== BRW_REGISTER_TYPE_UD
);
1728 /* From the Ivybridge PRM, Vol4, Part3, Section 6.26 f16to32:
1730 * Because this instruction does not have a 16-bit floating-point type,
1731 * the source data type must be Word (W). The destination type must be
1734 struct brw_reg src_w
= spread(retype(src
, BRW_REGISTER_TYPE_W
), 2);
1736 /* Each channel of src has the form of unpackHalf2x16's input: 0xhhhhllll.
1737 * For the Y case, we wish to access only the upper word; therefore
1738 * a 16-bit subregister offset is needed.
1740 assert(inst
->opcode
== FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X
||
1741 inst
->opcode
== FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y
);
1742 if (inst
->opcode
== FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y
)
1745 brw_F16TO32(p
, dst
, src_w
);
1749 fs_generator::generate_shader_time_add(fs_inst
*,
1750 struct brw_reg payload
,
1751 struct brw_reg offset
,
1752 struct brw_reg value
)
1754 assert(devinfo
->gen
>= 7);
1755 brw_push_insn_state(p
);
1756 brw_set_default_mask_control(p
, true);
1758 assert(payload
.file
== BRW_GENERAL_REGISTER_FILE
);
1759 struct brw_reg payload_offset
= retype(brw_vec1_grf(payload
.nr
, 0),
1761 struct brw_reg payload_value
= retype(brw_vec1_grf(payload
.nr
+ 1, 0),
1764 assert(offset
.file
== BRW_IMMEDIATE_VALUE
);
1765 if (value
.file
== BRW_GENERAL_REGISTER_FILE
) {
1766 value
.width
= BRW_WIDTH_1
;
1767 value
.hstride
= BRW_HORIZONTAL_STRIDE_0
;
1768 value
.vstride
= BRW_VERTICAL_STRIDE_0
;
1770 assert(value
.file
== BRW_IMMEDIATE_VALUE
);
1773 /* Trying to deal with setup of the params from the IR is crazy in the FS8
1774 * case, and we don't really care about squeezing every bit of performance
1775 * out of this path, so we just emit the MOVs from here.
1777 brw_MOV(p
, payload_offset
, offset
);
1778 brw_MOV(p
, payload_value
, value
);
1779 brw_shader_time_add(p
, payload
,
1780 prog_data
->binding_table
.shader_time_start
);
1781 brw_pop_insn_state(p
);
1783 brw_mark_surface_used(prog_data
,
1784 prog_data
->binding_table
.shader_time_start
);
1788 fs_generator::enable_debug(const char *shader_name
)
1791 this->shader_name
= shader_name
;
1795 fs_generator::generate_code(const cfg_t
*cfg
, int dispatch_width
)
1797 /* align to 64 byte boundary. */
1798 while (p
->next_insn_offset
% 64)
1801 this->dispatch_width
= dispatch_width
;
1803 int start_offset
= p
->next_insn_offset
;
1804 int spill_count
= 0, fill_count
= 0;
1807 struct disasm_info
*disasm_info
= disasm_initialize(devinfo
, cfg
);
1809 foreach_block_and_inst (block
, fs_inst
, inst
, cfg
) {
1810 struct brw_reg src
[3], dst
;
1811 unsigned int last_insn_offset
= p
->next_insn_offset
;
1812 bool multiple_instructions_emitted
= false;
1814 /* From the Broadwell PRM, Volume 7, "3D-Media-GPGPU", in the
1815 * "Register Region Restrictions" section: for BDW, SKL:
1817 * "A POW/FDIV operation must not be followed by an instruction
1818 * that requires two destination registers."
1820 * The documentation is often lacking annotations for Atom parts,
1821 * and empirically this affects CHV as well.
1823 if (devinfo
->gen
>= 8 &&
1824 devinfo
->gen
<= 9 &&
1826 brw_inst_opcode(devinfo
, brw_last_inst
) == BRW_OPCODE_MATH
&&
1827 brw_inst_math_function(devinfo
, brw_last_inst
) == BRW_MATH_FUNCTION_POW
&&
1828 inst
->dst
.component_size(inst
->exec_size
) > REG_SIZE
) {
1830 last_insn_offset
= p
->next_insn_offset
;
1833 if (unlikely(debug_flag
))
1834 disasm_annotate(disasm_info
, inst
, p
->next_insn_offset
);
1836 /* If the instruction writes to more than one register, it needs to be
1837 * explicitly marked as compressed on Gen <= 5. On Gen >= 6 the
1838 * hardware figures out by itself what the right compression mode is,
1839 * but we still need to know whether the instruction is compressed to
1840 * set up the source register regions appropriately.
1842 * XXX - This is wrong for instructions that write a single register but
1843 * read more than one which should strictly speaking be treated as
1844 * compressed. For instructions that don't write any registers it
1845 * relies on the destination being a null register of the correct
1846 * type and regioning so the instruction is considered compressed
1847 * or not accordingly.
1849 const bool compressed
=
1850 inst
->dst
.component_size(inst
->exec_size
) > REG_SIZE
;
1851 brw_set_default_compression(p
, compressed
);
1852 brw_set_default_group(p
, inst
->group
);
1854 for (unsigned int i
= 0; i
< inst
->sources
; i
++) {
1855 src
[i
] = brw_reg_from_fs_reg(devinfo
, inst
,
1856 &inst
->src
[i
], compressed
);
1857 /* The accumulator result appears to get used for the
1858 * conditional modifier generation. When negating a UD
1859 * value, there is a 33rd bit generated for the sign in the
1860 * accumulator value, so now you can't check, for example,
1861 * equality with a 32-bit value. See piglit fs-op-neg-uvec4.
1863 assert(!inst
->conditional_mod
||
1864 inst
->src
[i
].type
!= BRW_REGISTER_TYPE_UD
||
1865 !inst
->src
[i
].negate
);
1867 dst
= brw_reg_from_fs_reg(devinfo
, inst
,
1868 &inst
->dst
, compressed
);
1870 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
1871 brw_set_default_predicate_control(p
, inst
->predicate
);
1872 brw_set_default_predicate_inverse(p
, inst
->predicate_inverse
);
1873 brw_set_default_flag_reg(p
, inst
->flag_subreg
/ 2, inst
->flag_subreg
% 2);
1874 brw_set_default_saturate(p
, inst
->saturate
);
1875 brw_set_default_mask_control(p
, inst
->force_writemask_all
);
1876 brw_set_default_acc_write_control(p
, inst
->writes_accumulator
);
1878 unsigned exec_size
= inst
->exec_size
;
1879 if (devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
1880 (get_exec_type_size(inst
) == 8 || type_sz(inst
->dst
.type
) == 8)) {
1884 brw_set_default_exec_size(p
, cvt(exec_size
) - 1);
1886 assert(inst
->force_writemask_all
|| inst
->exec_size
>= 4);
1887 assert(inst
->force_writemask_all
|| inst
->group
% inst
->exec_size
== 0);
1888 assert(inst
->base_mrf
+ inst
->mlen
<= BRW_MAX_MRF(devinfo
->gen
));
1889 assert(inst
->mlen
<= BRW_MAX_MSG_LENGTH
);
1891 switch (inst
->opcode
) {
1892 case BRW_OPCODE_MOV
:
1893 brw_MOV(p
, dst
, src
[0]);
1895 case BRW_OPCODE_ADD
:
1896 brw_ADD(p
, dst
, src
[0], src
[1]);
1898 case BRW_OPCODE_MUL
:
1899 brw_MUL(p
, dst
, src
[0], src
[1]);
1901 case BRW_OPCODE_AVG
:
1902 brw_AVG(p
, dst
, src
[0], src
[1]);
1904 case BRW_OPCODE_MACH
:
1905 brw_MACH(p
, dst
, src
[0], src
[1]);
1908 case BRW_OPCODE_LINE
:
1909 brw_LINE(p
, dst
, src
[0], src
[1]);
1912 case BRW_OPCODE_MAD
:
1913 assert(devinfo
->gen
>= 6);
1914 if (devinfo
->gen
< 10)
1915 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1916 brw_MAD(p
, dst
, src
[0], src
[1], src
[2]);
1919 case BRW_OPCODE_LRP
:
1920 assert(devinfo
->gen
>= 6 && devinfo
->gen
<= 10);
1921 if (devinfo
->gen
< 10)
1922 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1923 brw_LRP(p
, dst
, src
[0], src
[1], src
[2]);
1926 case BRW_OPCODE_FRC
:
1927 brw_FRC(p
, dst
, src
[0]);
1929 case BRW_OPCODE_RNDD
:
1930 brw_RNDD(p
, dst
, src
[0]);
1932 case BRW_OPCODE_RNDE
:
1933 brw_RNDE(p
, dst
, src
[0]);
1935 case BRW_OPCODE_RNDZ
:
1936 brw_RNDZ(p
, dst
, src
[0]);
1939 case BRW_OPCODE_AND
:
1940 brw_AND(p
, dst
, src
[0], src
[1]);
1943 brw_OR(p
, dst
, src
[0], src
[1]);
1945 case BRW_OPCODE_XOR
:
1946 brw_XOR(p
, dst
, src
[0], src
[1]);
1948 case BRW_OPCODE_NOT
:
1949 brw_NOT(p
, dst
, src
[0]);
1951 case BRW_OPCODE_ASR
:
1952 brw_ASR(p
, dst
, src
[0], src
[1]);
1954 case BRW_OPCODE_SHR
:
1955 brw_SHR(p
, dst
, src
[0], src
[1]);
1957 case BRW_OPCODE_SHL
:
1958 brw_SHL(p
, dst
, src
[0], src
[1]);
1960 case BRW_OPCODE_F32TO16
:
1961 assert(devinfo
->gen
>= 7);
1962 brw_F32TO16(p
, dst
, src
[0]);
1964 case BRW_OPCODE_F16TO32
:
1965 assert(devinfo
->gen
>= 7);
1966 brw_F16TO32(p
, dst
, src
[0]);
1968 case BRW_OPCODE_CMP
:
1969 if (inst
->exec_size
>= 16 && devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
1970 dst
.file
== BRW_ARCHITECTURE_REGISTER_FILE
) {
1971 /* For unknown reasons the WaCMPInstFlagDepClearedEarly workaround
1972 * implemented in the compiler is not sufficient. Overriding the
1973 * type when the destination is the null register is necessary but
1974 * not sufficient by itself.
1976 assert(dst
.nr
== BRW_ARF_NULL
);
1977 dst
.type
= BRW_REGISTER_TYPE_D
;
1979 brw_CMP(p
, dst
, inst
->conditional_mod
, src
[0], src
[1]);
1981 case BRW_OPCODE_SEL
:
1982 brw_SEL(p
, dst
, src
[0], src
[1]);
1984 case BRW_OPCODE_CSEL
:
1985 assert(devinfo
->gen
>= 8);
1986 if (devinfo
->gen
< 10)
1987 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1988 brw_CSEL(p
, dst
, src
[0], src
[1], src
[2]);
1990 case BRW_OPCODE_BFREV
:
1991 assert(devinfo
->gen
>= 7);
1992 brw_BFREV(p
, retype(dst
, BRW_REGISTER_TYPE_UD
),
1993 retype(src
[0], BRW_REGISTER_TYPE_UD
));
1995 case BRW_OPCODE_FBH
:
1996 assert(devinfo
->gen
>= 7);
1997 brw_FBH(p
, retype(dst
, src
[0].type
), src
[0]);
1999 case BRW_OPCODE_FBL
:
2000 assert(devinfo
->gen
>= 7);
2001 brw_FBL(p
, retype(dst
, BRW_REGISTER_TYPE_UD
),
2002 retype(src
[0], BRW_REGISTER_TYPE_UD
));
2004 case BRW_OPCODE_LZD
:
2005 brw_LZD(p
, dst
, src
[0]);
2007 case BRW_OPCODE_CBIT
:
2008 assert(devinfo
->gen
>= 7);
2009 brw_CBIT(p
, retype(dst
, BRW_REGISTER_TYPE_UD
),
2010 retype(src
[0], BRW_REGISTER_TYPE_UD
));
2012 case BRW_OPCODE_ADDC
:
2013 assert(devinfo
->gen
>= 7);
2014 brw_ADDC(p
, dst
, src
[0], src
[1]);
2016 case BRW_OPCODE_SUBB
:
2017 assert(devinfo
->gen
>= 7);
2018 brw_SUBB(p
, dst
, src
[0], src
[1]);
2020 case BRW_OPCODE_MAC
:
2021 brw_MAC(p
, dst
, src
[0], src
[1]);
2024 case BRW_OPCODE_BFE
:
2025 assert(devinfo
->gen
>= 7);
2026 if (devinfo
->gen
< 10)
2027 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
2028 brw_BFE(p
, dst
, src
[0], src
[1], src
[2]);
2031 case BRW_OPCODE_BFI1
:
2032 assert(devinfo
->gen
>= 7);
2033 brw_BFI1(p
, dst
, src
[0], src
[1]);
2035 case BRW_OPCODE_BFI2
:
2036 assert(devinfo
->gen
>= 7);
2037 if (devinfo
->gen
< 10)
2038 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
2039 brw_BFI2(p
, dst
, src
[0], src
[1], src
[2]);
2043 if (inst
->src
[0].file
!= BAD_FILE
) {
2044 /* The instruction has an embedded compare (only allowed on gen6) */
2045 assert(devinfo
->gen
== 6);
2046 gen6_IF(p
, inst
->conditional_mod
, src
[0], src
[1]);
2048 brw_IF(p
, brw_inst_exec_size(devinfo
, p
->current
));
2052 case BRW_OPCODE_ELSE
:
2055 case BRW_OPCODE_ENDIF
:
2060 brw_DO(p
, brw_inst_exec_size(devinfo
, p
->current
));
2063 case BRW_OPCODE_BREAK
:
2066 case BRW_OPCODE_CONTINUE
:
2070 case BRW_OPCODE_WHILE
:
2075 case SHADER_OPCODE_RCP
:
2076 case SHADER_OPCODE_RSQ
:
2077 case SHADER_OPCODE_SQRT
:
2078 case SHADER_OPCODE_EXP2
:
2079 case SHADER_OPCODE_LOG2
:
2080 case SHADER_OPCODE_SIN
:
2081 case SHADER_OPCODE_COS
:
2082 assert(inst
->conditional_mod
== BRW_CONDITIONAL_NONE
);
2083 if (devinfo
->gen
>= 6) {
2084 assert(inst
->mlen
== 0);
2085 assert(devinfo
->gen
>= 7 || inst
->exec_size
== 8);
2086 gen6_math(p
, dst
, brw_math_function(inst
->opcode
),
2087 src
[0], brw_null_reg());
2089 assert(inst
->mlen
>= 1);
2090 assert(devinfo
->gen
== 5 || devinfo
->is_g4x
|| inst
->exec_size
== 8);
2092 brw_math_function(inst
->opcode
),
2093 inst
->base_mrf
, src
[0],
2094 BRW_MATH_PRECISION_FULL
);
2097 case SHADER_OPCODE_INT_QUOTIENT
:
2098 case SHADER_OPCODE_INT_REMAINDER
:
2099 case SHADER_OPCODE_POW
:
2100 assert(inst
->conditional_mod
== BRW_CONDITIONAL_NONE
);
2101 if (devinfo
->gen
>= 6) {
2102 assert(inst
->mlen
== 0);
2103 assert((devinfo
->gen
>= 7 && inst
->opcode
== SHADER_OPCODE_POW
) ||
2104 inst
->exec_size
== 8);
2105 gen6_math(p
, dst
, brw_math_function(inst
->opcode
), src
[0], src
[1]);
2107 assert(inst
->mlen
>= 1);
2108 assert(inst
->exec_size
== 8);
2109 gen4_math(p
, dst
, brw_math_function(inst
->opcode
),
2110 inst
->base_mrf
, src
[0],
2111 BRW_MATH_PRECISION_FULL
);
2114 case FS_OPCODE_CINTERP
:
2115 brw_MOV(p
, dst
, src
[0]);
2117 case FS_OPCODE_LINTERP
:
2118 multiple_instructions_emitted
= generate_linterp(inst
, dst
, src
);
2120 case FS_OPCODE_PIXEL_X
:
2121 assert(src
[0].type
== BRW_REGISTER_TYPE_UW
);
2122 src
[0].subnr
= 0 * type_sz(src
[0].type
);
2123 brw_MOV(p
, dst
, stride(src
[0], 8, 4, 1));
2125 case FS_OPCODE_PIXEL_Y
:
2126 assert(src
[0].type
== BRW_REGISTER_TYPE_UW
);
2127 src
[0].subnr
= 4 * type_sz(src
[0].type
);
2128 brw_MOV(p
, dst
, stride(src
[0], 8, 4, 1));
2130 case SHADER_OPCODE_GET_BUFFER_SIZE
:
2131 generate_get_buffer_size(inst
, dst
, src
[0], src
[1]);
2133 case SHADER_OPCODE_TEX
:
2135 case SHADER_OPCODE_TXD
:
2136 case SHADER_OPCODE_TXF
:
2137 case SHADER_OPCODE_TXF_LZ
:
2138 case SHADER_OPCODE_TXF_CMS
:
2139 case SHADER_OPCODE_TXF_CMS_W
:
2140 case SHADER_OPCODE_TXF_UMS
:
2141 case SHADER_OPCODE_TXF_MCS
:
2142 case SHADER_OPCODE_TXL
:
2143 case SHADER_OPCODE_TXL_LZ
:
2144 case SHADER_OPCODE_TXS
:
2145 case SHADER_OPCODE_LOD
:
2146 case SHADER_OPCODE_TG4
:
2147 case SHADER_OPCODE_TG4_OFFSET
:
2148 case SHADER_OPCODE_SAMPLEINFO
:
2149 generate_tex(inst
, dst
, src
[0], src
[1], src
[2]);
2151 case FS_OPCODE_DDX_COARSE
:
2152 case FS_OPCODE_DDX_FINE
:
2153 generate_ddx(inst
, dst
, src
[0]);
2155 case FS_OPCODE_DDY_COARSE
:
2156 case FS_OPCODE_DDY_FINE
:
2157 generate_ddy(inst
, dst
, src
[0]);
2160 case SHADER_OPCODE_GEN4_SCRATCH_WRITE
:
2161 generate_scratch_write(inst
, src
[0]);
2165 case SHADER_OPCODE_GEN4_SCRATCH_READ
:
2166 generate_scratch_read(inst
, dst
);
2170 case SHADER_OPCODE_GEN7_SCRATCH_READ
:
2171 generate_scratch_read_gen7(inst
, dst
);
2175 case SHADER_OPCODE_MOV_INDIRECT
:
2176 generate_mov_indirect(inst
, dst
, src
[0], src
[1]);
2179 case SHADER_OPCODE_URB_READ_SIMD8
:
2180 case SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
:
2181 generate_urb_read(inst
, dst
, src
[0]);
2184 case SHADER_OPCODE_URB_WRITE_SIMD8
:
2185 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
2186 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
2187 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
2188 generate_urb_write(inst
, src
[0]);
2191 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
2192 assert(inst
->force_writemask_all
);
2193 generate_uniform_pull_constant_load(inst
, dst
, src
[0], src
[1]);
2196 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
:
2197 assert(inst
->force_writemask_all
);
2198 generate_uniform_pull_constant_load_gen7(inst
, dst
, src
[0], src
[1]);
2201 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN4
:
2202 generate_varying_pull_constant_load_gen4(inst
, dst
, src
[0]);
2205 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7
:
2206 generate_varying_pull_constant_load_gen7(inst
, dst
, src
[0], src
[1]);
2209 case FS_OPCODE_REP_FB_WRITE
:
2210 case FS_OPCODE_FB_WRITE
:
2211 generate_fb_write(inst
, src
[0]);
2214 case FS_OPCODE_FB_READ
:
2215 generate_fb_read(inst
, dst
, src
[0]);
2218 case FS_OPCODE_MOV_DISPATCH_TO_FLAGS
:
2219 generate_mov_dispatch_to_flags(inst
);
2222 case FS_OPCODE_DISCARD_JUMP
:
2223 generate_discard_jump(inst
);
2226 case SHADER_OPCODE_SHADER_TIME_ADD
:
2227 generate_shader_time_add(inst
, src
[0], src
[1], src
[2]);
2230 case SHADER_OPCODE_UNTYPED_ATOMIC
:
2231 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2232 brw_untyped_atomic(p
, dst
, src
[0], src
[1], src
[2].ud
,
2233 inst
->mlen
, !inst
->dst
.is_null(),
2237 case SHADER_OPCODE_UNTYPED_SURFACE_READ
:
2238 assert(!inst
->header_size
);
2239 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2240 brw_untyped_surface_read(p
, dst
, src
[0], src
[1],
2241 inst
->mlen
, src
[2].ud
);
2244 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE
:
2245 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2246 brw_untyped_surface_write(p
, src
[0], src
[1],
2247 inst
->mlen
, src
[2].ud
,
2251 case SHADER_OPCODE_BYTE_SCATTERED_READ
:
2252 assert(!inst
->header_size
);
2253 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2254 brw_byte_scattered_read(p
, dst
, src
[0], src
[1],
2255 inst
->mlen
, src
[2].ud
);
2258 case SHADER_OPCODE_BYTE_SCATTERED_WRITE
:
2259 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2260 brw_byte_scattered_write(p
, src
[0], src
[1],
2261 inst
->mlen
, src
[2].ud
,
2265 case SHADER_OPCODE_TYPED_ATOMIC
:
2266 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2267 brw_typed_atomic(p
, dst
, src
[0], src
[1],
2268 src
[2].ud
, inst
->mlen
, !inst
->dst
.is_null(),
2272 case SHADER_OPCODE_TYPED_SURFACE_READ
:
2273 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2274 brw_typed_surface_read(p
, dst
, src
[0], src
[1],
2275 inst
->mlen
, src
[2].ud
,
2279 case SHADER_OPCODE_TYPED_SURFACE_WRITE
:
2280 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2281 brw_typed_surface_write(p
, src
[0], src
[1], inst
->mlen
, src
[2].ud
,
2285 case SHADER_OPCODE_MEMORY_FENCE
:
2286 brw_memory_fence(p
, dst
);
2289 case SHADER_OPCODE_FIND_LIVE_CHANNEL
: {
2290 const struct brw_reg mask
=
2291 brw_stage_has_packed_dispatch(devinfo
, stage
,
2292 prog_data
) ? brw_imm_ud(~0u) :
2293 stage
== MESA_SHADER_FRAGMENT
? brw_vmask_reg() :
2295 brw_find_live_channel(p
, dst
, mask
);
2299 case SHADER_OPCODE_BROADCAST
:
2300 assert(inst
->force_writemask_all
);
2301 brw_broadcast(p
, dst
, src
[0], src
[1]);
2304 case SHADER_OPCODE_SHUFFLE
:
2305 generate_shuffle(inst
, dst
, src
[0], src
[1]);
2308 case SHADER_OPCODE_SEL_EXEC
:
2309 assert(inst
->force_writemask_all
);
2310 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
2311 brw_MOV(p
, dst
, src
[1]);
2312 brw_set_default_mask_control(p
, BRW_MASK_ENABLE
);
2313 brw_MOV(p
, dst
, src
[0]);
2316 case SHADER_OPCODE_QUAD_SWIZZLE
:
2317 /* This only works on 8-wide 32-bit values */
2318 assert(inst
->exec_size
== 8);
2319 assert(type_sz(src
[0].type
) == 4);
2320 assert(inst
->force_writemask_all
);
2321 assert(src
[1].file
== BRW_IMMEDIATE_VALUE
);
2322 assert(src
[1].type
== BRW_REGISTER_TYPE_UD
);
2324 if (src
[0].file
== BRW_IMMEDIATE_VALUE
||
2325 (src
[0].vstride
== 0 && src
[0].hstride
== 0)) {
2326 /* The value is uniform across all channels */
2327 brw_MOV(p
, dst
, src
[0]);
2329 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
2330 struct brw_reg swiz_src
= stride(src
[0], 4, 4, 1);
2331 swiz_src
.swizzle
= inst
->src
[1].ud
;
2332 brw_MOV(p
, dst
, swiz_src
);
2336 case SHADER_OPCODE_CLUSTER_BROADCAST
: {
2337 assert(src
[0].type
== dst
.type
);
2338 assert(!src
[0].negate
&& !src
[0].abs
);
2339 assert(src
[1].file
== BRW_IMMEDIATE_VALUE
);
2340 assert(src
[1].type
== BRW_REGISTER_TYPE_UD
);
2341 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2342 assert(src
[2].type
== BRW_REGISTER_TYPE_UD
);
2343 const unsigned component
= src
[1].ud
;
2344 const unsigned cluster_size
= src
[2].ud
;
2345 struct brw_reg strided
= stride(suboffset(src
[0], component
),
2346 cluster_size
, cluster_size
, 0);
2347 if (type_sz(src
[0].type
) > 4 &&
2348 (devinfo
->is_cherryview
|| gen_device_info_is_9lp(devinfo
))) {
2349 /* IVB has an issue (which we found empirically) where it reads
2350 * two address register components per channel for indirectly
2351 * addressed 64-bit sources.
2353 * From the Cherryview PRM Vol 7. "Register Region Restrictions":
2355 * "When source or destination datatype is 64b or operation is
2356 * integer DWord multiply, indirect addressing must not be
2359 * To work around both of these, we do two integer MOVs insead of
2360 * one 64-bit MOV. Because no double value should ever cross a
2361 * register boundary, it's safe to use the immediate offset in the
2362 * indirect here to handle adding 4 bytes to the offset and avoid
2363 * the extra ADD to the register file.
2365 brw_MOV(p
, subscript(dst
, BRW_REGISTER_TYPE_D
, 0),
2366 subscript(strided
, BRW_REGISTER_TYPE_D
, 0));
2367 brw_MOV(p
, subscript(dst
, BRW_REGISTER_TYPE_D
, 1),
2368 subscript(strided
, BRW_REGISTER_TYPE_D
, 1));
2370 brw_MOV(p
, dst
, strided
);
2375 case FS_OPCODE_SET_SAMPLE_ID
:
2376 generate_set_sample_id(inst
, dst
, src
[0], src
[1]);
2379 case FS_OPCODE_PACK_HALF_2x16_SPLIT
:
2380 generate_pack_half_2x16_split(inst
, dst
, src
[0], src
[1]);
2383 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X
:
2384 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y
:
2385 generate_unpack_half_2x16_split(inst
, dst
, src
[0]);
2388 case FS_OPCODE_PLACEHOLDER_HALT
:
2389 /* This is the place where the final HALT needs to be inserted if
2390 * we've emitted any discards. If not, this will emit no code.
2392 if (!patch_discard_jumps_to_fb_writes()) {
2393 if (unlikely(debug_flag
)) {
2394 disasm_info
->use_tail
= true;
2399 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
2400 generate_pixel_interpolator_query(inst
, dst
, src
[0], src
[1],
2401 GEN7_PIXEL_INTERPOLATOR_LOC_SAMPLE
);
2404 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
2405 generate_pixel_interpolator_query(inst
, dst
, src
[0], src
[1],
2406 GEN7_PIXEL_INTERPOLATOR_LOC_SHARED_OFFSET
);
2409 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
2410 generate_pixel_interpolator_query(inst
, dst
, src
[0], src
[1],
2411 GEN7_PIXEL_INTERPOLATOR_LOC_PER_SLOT_OFFSET
);
2414 case CS_OPCODE_CS_TERMINATE
:
2415 generate_cs_terminate(inst
, src
[0]);
2418 case SHADER_OPCODE_BARRIER
:
2419 generate_barrier(inst
, src
[0]);
2422 case BRW_OPCODE_DIM
:
2423 assert(devinfo
->is_haswell
);
2424 assert(src
[0].type
== BRW_REGISTER_TYPE_DF
);
2425 assert(dst
.type
== BRW_REGISTER_TYPE_DF
);
2426 brw_DIM(p
, dst
, retype(src
[0], BRW_REGISTER_TYPE_F
));
2429 case SHADER_OPCODE_RND_MODE
:
2430 assert(src
[0].file
== BRW_IMMEDIATE_VALUE
);
2431 brw_rounding_mode(p
, (brw_rnd_mode
) src
[0].d
);
2435 unreachable("Unsupported opcode");
2437 case SHADER_OPCODE_LOAD_PAYLOAD
:
2438 unreachable("Should be lowered by lower_load_payload()");
2441 if (multiple_instructions_emitted
)
2444 if (inst
->no_dd_clear
|| inst
->no_dd_check
|| inst
->conditional_mod
) {
2445 assert(p
->next_insn_offset
== last_insn_offset
+ 16 ||
2446 !"conditional_mod, no_dd_check, or no_dd_clear set for IR "
2447 "emitting more than 1 instruction");
2449 brw_inst
*last
= &p
->store
[last_insn_offset
/ 16];
2451 if (inst
->conditional_mod
)
2452 brw_inst_set_cond_modifier(p
->devinfo
, last
, inst
->conditional_mod
);
2453 brw_inst_set_no_dd_clear(p
->devinfo
, last
, inst
->no_dd_clear
);
2454 brw_inst_set_no_dd_check(p
->devinfo
, last
, inst
->no_dd_check
);
2458 brw_set_uip_jip(p
, start_offset
);
2460 /* end of program sentinel */
2461 disasm_new_inst_group(disasm_info
, p
->next_insn_offset
);
2466 if (unlikely(debug_flag
))
2468 brw_validate_instructions(devinfo
, p
->store
,
2470 p
->next_insn_offset
,
2473 int before_size
= p
->next_insn_offset
- start_offset
;
2474 brw_compact_instructions(p
, start_offset
, disasm_info
);
2475 int after_size
= p
->next_insn_offset
- start_offset
;
2477 if (unlikely(debug_flag
)) {
2478 fprintf(stderr
, "Native code for %s\n"
2479 "SIMD%d shader: %d instructions. %d loops. %u cycles. %d:%d spills:fills. Promoted %u constants. Compacted %d to %d"
2480 " bytes (%.0f%%)\n",
2481 shader_name
, dispatch_width
, before_size
/ 16, loop_count
, cfg
->cycle_count
,
2482 spill_count
, fill_count
, promoted_constants
, before_size
, after_size
,
2483 100.0f
* (before_size
- after_size
) / before_size
);
2485 dump_assembly(p
->store
, disasm_info
);
2487 ralloc_free(disasm_info
);
2490 compiler
->shader_debug_log(log_data
,
2491 "%s SIMD%d shader: %d inst, %d loops, %u cycles, "
2492 "%d:%d spills:fills, Promoted %u constants, "
2493 "compacted %d to %d bytes.",
2494 _mesa_shader_stage_to_abbrev(stage
),
2495 dispatch_width
, before_size
/ 16,
2496 loop_count
, cfg
->cycle_count
, spill_count
,
2497 fill_count
, promoted_constants
, before_size
,
2500 return start_offset
;
2504 fs_generator::get_assembly()
2506 return brw_get_program(p
, &prog_data
->program_size
);