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_urb_read(fs_inst
*inst
,
546 struct brw_reg header
)
548 assert(inst
->size_written
% REG_SIZE
== 0);
549 assert(header
.file
== BRW_GENERAL_REGISTER_FILE
);
550 assert(header
.type
== BRW_REGISTER_TYPE_UD
);
552 brw_inst
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
553 brw_set_dest(p
, send
, retype(dst
, BRW_REGISTER_TYPE_UD
));
554 brw_set_src0(p
, send
, header
);
555 brw_set_src1(p
, send
, brw_imm_ud(0u));
557 brw_inst_set_sfid(p
->devinfo
, send
, BRW_SFID_URB
);
558 brw_inst_set_urb_opcode(p
->devinfo
, send
, GEN8_URB_OPCODE_SIMD8_READ
);
560 if (inst
->opcode
== SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
)
561 brw_inst_set_urb_per_slot_offset(p
->devinfo
, send
, true);
563 brw_inst_set_mlen(p
->devinfo
, send
, inst
->mlen
);
564 brw_inst_set_rlen(p
->devinfo
, send
, inst
->size_written
/ REG_SIZE
);
565 brw_inst_set_header_present(p
->devinfo
, send
, true);
566 brw_inst_set_urb_global_offset(p
->devinfo
, send
, inst
->offset
);
570 fs_generator::generate_urb_write(fs_inst
*inst
, struct brw_reg payload
)
574 /* WaClearTDRRegBeforeEOTForNonPS.
576 * WA: Clear tdr register before send EOT in all non-PS shader kernels
578 * mov(8) tdr0:ud 0x0:ud {NoMask}"
580 if (inst
->eot
&& p
->devinfo
->gen
== 10) {
581 brw_push_insn_state(p
);
582 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
583 brw_MOV(p
, brw_tdr_reg(), brw_imm_uw(0));
584 brw_pop_insn_state(p
);
587 insn
= brw_next_insn(p
, BRW_OPCODE_SEND
);
589 brw_set_dest(p
, insn
, brw_null_reg());
590 brw_set_src0(p
, insn
, payload
);
591 brw_set_src1(p
, insn
, brw_imm_d(0));
593 brw_inst_set_sfid(p
->devinfo
, insn
, BRW_SFID_URB
);
594 brw_inst_set_urb_opcode(p
->devinfo
, insn
, GEN8_URB_OPCODE_SIMD8_WRITE
);
596 if (inst
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
||
597 inst
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
)
598 brw_inst_set_urb_per_slot_offset(p
->devinfo
, insn
, true);
600 if (inst
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
||
601 inst
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
)
602 brw_inst_set_urb_channel_mask_present(p
->devinfo
, insn
, true);
604 brw_inst_set_mlen(p
->devinfo
, insn
, inst
->mlen
);
605 brw_inst_set_rlen(p
->devinfo
, insn
, 0);
606 brw_inst_set_eot(p
->devinfo
, insn
, inst
->eot
);
607 brw_inst_set_header_present(p
->devinfo
, insn
, true);
608 brw_inst_set_urb_global_offset(p
->devinfo
, insn
, inst
->offset
);
612 fs_generator::generate_cs_terminate(fs_inst
*inst
, struct brw_reg payload
)
614 struct brw_inst
*insn
;
616 insn
= brw_next_insn(p
, BRW_OPCODE_SEND
);
618 brw_set_dest(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_UW
));
619 brw_set_src0(p
, insn
, retype(payload
, BRW_REGISTER_TYPE_UW
));
620 brw_set_src1(p
, insn
, brw_imm_d(0));
622 /* Terminate a compute shader by sending a message to the thread spawner.
624 brw_inst_set_sfid(devinfo
, insn
, BRW_SFID_THREAD_SPAWNER
);
625 brw_inst_set_mlen(devinfo
, insn
, 1);
626 brw_inst_set_rlen(devinfo
, insn
, 0);
627 brw_inst_set_eot(devinfo
, insn
, inst
->eot
);
628 brw_inst_set_header_present(devinfo
, insn
, false);
630 brw_inst_set_ts_opcode(devinfo
, insn
, 0); /* Dereference resource */
631 brw_inst_set_ts_request_type(devinfo
, insn
, 0); /* Root thread */
633 /* Note that even though the thread has a URB resource associated with it,
634 * we set the "do not dereference URB" bit, because the URB resource is
635 * managed by the fixed-function unit, so it will free it automatically.
637 brw_inst_set_ts_resource_select(devinfo
, insn
, 1); /* Do not dereference URB */
639 brw_inst_set_mask_control(devinfo
, insn
, BRW_MASK_DISABLE
);
643 fs_generator::generate_barrier(fs_inst
*inst
, struct brw_reg src
)
650 fs_generator::generate_linterp(fs_inst
*inst
,
651 struct brw_reg dst
, struct brw_reg
*src
)
655 * -----------------------------------
656 * | src1+0 | src1+1 | src1+2 | src1+3 |
657 * |-----------------------------------|
658 * |(x0, x1)|(y0, y1)|(x2, x3)|(y2, y3)|
659 * -----------------------------------
661 * but for the LINE/MAC pair, the LINE reads Xs and the MAC reads Ys:
663 * -----------------------------------
664 * | src1+0 | src1+1 | src1+2 | src1+3 |
665 * |-----------------------------------|
666 * |(x0, x1)|(y0, y1)| | | in SIMD8
667 * |-----------------------------------|
668 * |(x0, x1)|(x2, x3)|(y0, y1)|(y2, y3)| in SIMD16
669 * -----------------------------------
671 * See also: emit_interpolation_setup_gen4().
673 struct brw_reg delta_x
= src
[0];
674 struct brw_reg delta_y
= offset(src
[0], inst
->exec_size
/ 8);
675 struct brw_reg interp
= src
[1];
678 if (devinfo
->gen
>= 11) {
679 struct brw_reg acc
= retype(brw_acc_reg(8), BRW_REGISTER_TYPE_NF
);
680 struct brw_reg dwP
= suboffset(interp
, 0);
681 struct brw_reg dwQ
= suboffset(interp
, 1);
682 struct brw_reg dwR
= suboffset(interp
, 3);
684 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
686 if (inst
->exec_size
== 8) {
687 i
[0] = brw_MAD(p
, acc
, dwR
, offset(delta_x
, 0), dwP
);
688 i
[1] = brw_MAD(p
, offset(dst
, 0), acc
, offset(delta_y
, 0), dwQ
);
690 brw_inst_set_cond_modifier(p
->devinfo
, i
[1], inst
->conditional_mod
);
692 /* brw_set_default_saturate() is called before emitting instructions,
693 * so the saturate bit is set in each instruction, so we need to unset
694 * it on the first instruction of each pair.
696 brw_inst_set_saturate(p
->devinfo
, i
[0], false);
698 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
699 i
[0] = brw_MAD(p
, acc
, dwR
, offset(delta_x
, 0), dwP
);
700 i
[1] = brw_MAD(p
, offset(dst
, 0), acc
, offset(delta_x
, 1), dwQ
);
702 brw_set_default_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
703 i
[2] = brw_MAD(p
, acc
, dwR
, offset(delta_y
, 0), dwP
);
704 i
[3] = brw_MAD(p
, offset(dst
, 1), acc
, offset(delta_y
, 1), dwQ
);
706 brw_set_default_compression_control(p
, BRW_COMPRESSION_COMPRESSED
);
708 brw_inst_set_cond_modifier(p
->devinfo
, i
[1], inst
->conditional_mod
);
709 brw_inst_set_cond_modifier(p
->devinfo
, i
[3], inst
->conditional_mod
);
711 /* brw_set_default_saturate() is called before emitting instructions,
712 * so the saturate bit is set in each instruction, so we need to unset
713 * it on the first instruction of each pair.
715 brw_inst_set_saturate(p
->devinfo
, i
[0], false);
716 brw_inst_set_saturate(p
->devinfo
, i
[2], false);
720 } else if (devinfo
->has_pln
&&
721 (devinfo
->gen
>= 7 || (delta_x
.nr
& 1) == 0)) {
722 brw_PLN(p
, dst
, interp
, delta_x
);
726 i
[0] = brw_LINE(p
, brw_null_reg(), interp
, delta_x
);
727 i
[1] = brw_MAC(p
, dst
, suboffset(interp
, 1), delta_y
);
729 brw_inst_set_cond_modifier(p
->devinfo
, i
[1], inst
->conditional_mod
);
731 /* brw_set_default_saturate() is called before emitting instructions, so
732 * the saturate bit is set in each instruction, so we need to unset it on
733 * the first instruction.
735 brw_inst_set_saturate(p
->devinfo
, i
[0], false);
742 fs_generator::generate_get_buffer_size(fs_inst
*inst
,
745 struct brw_reg surf_index
)
747 assert(devinfo
->gen
>= 7);
748 assert(surf_index
.file
== BRW_IMMEDIATE_VALUE
);
753 switch (inst
->exec_size
) {
755 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD8
;
758 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
761 unreachable("Invalid width for texture instruction");
764 if (simd_mode
== BRW_SAMPLER_SIMD_MODE_SIMD16
) {
770 retype(dst
, BRW_REGISTER_TYPE_UW
),
775 GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO
,
776 rlen
, /* response length */
778 inst
->header_size
> 0,
780 BRW_SAMPLER_RETURN_FORMAT_SINT32
);
782 brw_mark_surface_used(prog_data
, surf_index
.ud
);
786 fs_generator::generate_tex(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
,
787 struct brw_reg surface_index
,
788 struct brw_reg sampler_index
)
790 assert(inst
->size_written
% REG_SIZE
== 0);
793 uint32_t return_format
;
794 bool is_combined_send
= inst
->eot
;
797 case BRW_REGISTER_TYPE_D
:
798 return_format
= BRW_SAMPLER_RETURN_FORMAT_SINT32
;
800 case BRW_REGISTER_TYPE_UD
:
801 return_format
= BRW_SAMPLER_RETURN_FORMAT_UINT32
;
804 return_format
= BRW_SAMPLER_RETURN_FORMAT_FLOAT32
;
808 /* Stomp the resinfo output type to UINT32. On gens 4-5, the output type
809 * is set as part of the message descriptor. On gen4, the PRM seems to
810 * allow UINT32 and FLOAT32 (i965 PRM, Vol. 4 Section 4.8.1.1), but on
811 * later gens UINT32 is required. Once you hit Sandy Bridge, the bit is
812 * gone from the message descriptor entirely and you just get UINT32 all
813 * the time regasrdless. Since we can really only do non-UINT32 on gen4,
814 * just stomp it to UINT32 all the time.
816 if (inst
->opcode
== SHADER_OPCODE_TXS
)
817 return_format
= BRW_SAMPLER_RETURN_FORMAT_UINT32
;
819 switch (inst
->exec_size
) {
821 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD8
;
824 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
827 unreachable("Invalid width for texture instruction");
830 if (devinfo
->gen
>= 5) {
831 switch (inst
->opcode
) {
832 case SHADER_OPCODE_TEX
:
833 if (inst
->shadow_compare
) {
834 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_COMPARE
;
836 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE
;
840 if (inst
->shadow_compare
) {
841 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS_COMPARE
;
843 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS
;
846 case SHADER_OPCODE_TXL
:
847 if (inst
->shadow_compare
) {
848 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE
;
850 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LOD
;
853 case SHADER_OPCODE_TXL_LZ
:
854 assert(devinfo
->gen
>= 9);
855 if (inst
->shadow_compare
) {
856 msg_type
= GEN9_SAMPLER_MESSAGE_SAMPLE_C_LZ
;
858 msg_type
= GEN9_SAMPLER_MESSAGE_SAMPLE_LZ
;
861 case SHADER_OPCODE_TXS
:
862 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO
;
864 case SHADER_OPCODE_TXD
:
865 if (inst
->shadow_compare
) {
866 /* Gen7.5+. Otherwise, lowered in NIR */
867 assert(devinfo
->gen
>= 8 || devinfo
->is_haswell
);
868 msg_type
= HSW_SAMPLER_MESSAGE_SAMPLE_DERIV_COMPARE
;
870 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS
;
873 case SHADER_OPCODE_TXF
:
874 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
876 case SHADER_OPCODE_TXF_LZ
:
877 assert(devinfo
->gen
>= 9);
878 msg_type
= GEN9_SAMPLER_MESSAGE_SAMPLE_LD_LZ
;
880 case SHADER_OPCODE_TXF_CMS_W
:
881 assert(devinfo
->gen
>= 9);
882 msg_type
= GEN9_SAMPLER_MESSAGE_SAMPLE_LD2DMS_W
;
884 case SHADER_OPCODE_TXF_CMS
:
885 if (devinfo
->gen
>= 7)
886 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DMS
;
888 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
890 case SHADER_OPCODE_TXF_UMS
:
891 assert(devinfo
->gen
>= 7);
892 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DSS
;
894 case SHADER_OPCODE_TXF_MCS
:
895 assert(devinfo
->gen
>= 7);
896 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_LD_MCS
;
898 case SHADER_OPCODE_LOD
:
899 msg_type
= GEN5_SAMPLER_MESSAGE_LOD
;
901 case SHADER_OPCODE_TG4
:
902 if (inst
->shadow_compare
) {
903 assert(devinfo
->gen
>= 7);
904 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_C
;
906 assert(devinfo
->gen
>= 6);
907 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4
;
910 case SHADER_OPCODE_TG4_OFFSET
:
911 assert(devinfo
->gen
>= 7);
912 if (inst
->shadow_compare
) {
913 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO_C
;
915 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO
;
918 case SHADER_OPCODE_SAMPLEINFO
:
919 msg_type
= GEN6_SAMPLER_MESSAGE_SAMPLE_SAMPLEINFO
;
922 unreachable("not reached");
925 switch (inst
->opcode
) {
926 case SHADER_OPCODE_TEX
:
927 /* Note that G45 and older determines shadow compare and dispatch width
928 * from message length for most messages.
930 if (inst
->exec_size
== 8) {
931 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE
;
932 if (inst
->shadow_compare
) {
933 assert(inst
->mlen
== 6);
935 assert(inst
->mlen
<= 4);
938 if (inst
->shadow_compare
) {
939 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_COMPARE
;
940 assert(inst
->mlen
== 9);
942 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE
;
943 assert(inst
->mlen
<= 7 && inst
->mlen
% 2 == 1);
948 if (inst
->shadow_compare
) {
949 assert(inst
->exec_size
== 8);
950 assert(inst
->mlen
== 6);
951 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_BIAS_COMPARE
;
953 assert(inst
->mlen
== 9);
954 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_BIAS
;
955 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
958 case SHADER_OPCODE_TXL
:
959 if (inst
->shadow_compare
) {
960 assert(inst
->exec_size
== 8);
961 assert(inst
->mlen
== 6);
962 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_LOD_COMPARE
;
964 assert(inst
->mlen
== 9);
965 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_LOD
;
966 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
969 case SHADER_OPCODE_TXD
:
970 /* There is no sample_d_c message; comparisons are done manually */
971 assert(inst
->exec_size
== 8);
972 assert(inst
->mlen
== 7 || inst
->mlen
== 10);
973 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_GRADIENTS
;
975 case SHADER_OPCODE_TXF
:
976 assert(inst
->mlen
<= 9 && inst
->mlen
% 2 == 1);
977 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_LD
;
978 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
980 case SHADER_OPCODE_TXS
:
981 assert(inst
->mlen
== 3);
982 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_RESINFO
;
983 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
986 unreachable("not reached");
989 assert(msg_type
!= -1);
991 if (simd_mode
== BRW_SAMPLER_SIMD_MODE_SIMD16
) {
995 assert(devinfo
->gen
< 7 || inst
->header_size
== 0 ||
996 src
.file
== BRW_GENERAL_REGISTER_FILE
);
998 assert(sampler_index
.type
== BRW_REGISTER_TYPE_UD
);
1000 /* Load the message header if present. If there's a texture offset,
1001 * we need to set it up explicitly and load the offset bitfield.
1002 * Otherwise, we can use an implied move from g0 to the first message reg.
1004 if (inst
->header_size
!= 0) {
1005 if (devinfo
->gen
< 6 && !inst
->offset
) {
1006 /* Set up an implied move from g0 to the MRF. */
1007 src
= retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW
);
1009 struct brw_reg header_reg
;
1011 if (devinfo
->gen
>= 7) {
1014 assert(inst
->base_mrf
!= -1);
1015 header_reg
= brw_message_reg(inst
->base_mrf
);
1018 brw_push_insn_state(p
);
1019 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
1020 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1021 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
1022 /* Explicitly set up the message header by copying g0 to the MRF. */
1023 brw_MOV(p
, header_reg
, brw_vec8_grf(0, 0));
1025 brw_set_default_exec_size(p
, BRW_EXECUTE_1
);
1027 /* Set the offset bits in DWord 2. */
1028 brw_MOV(p
, get_element_ud(header_reg
, 2),
1029 brw_imm_ud(inst
->offset
));
1030 } else if (stage
!= MESA_SHADER_VERTEX
&&
1031 stage
!= MESA_SHADER_FRAGMENT
) {
1032 /* The vertex and fragment stages have g0.2 set to 0, so
1033 * header0.2 is 0 when g0 is copied. Other stages may not, so we
1034 * must set it to 0 to avoid setting undesirable bits in the
1037 brw_MOV(p
, get_element_ud(header_reg
, 2), brw_imm_ud(0));
1040 brw_adjust_sampler_state_pointer(p
, header_reg
, sampler_index
);
1041 brw_pop_insn_state(p
);
1045 uint32_t base_binding_table_index
= (inst
->opcode
== SHADER_OPCODE_TG4
||
1046 inst
->opcode
== SHADER_OPCODE_TG4_OFFSET
)
1047 ? prog_data
->binding_table
.gather_texture_start
1048 : prog_data
->binding_table
.texture_start
;
1050 if (surface_index
.file
== BRW_IMMEDIATE_VALUE
&&
1051 sampler_index
.file
== BRW_IMMEDIATE_VALUE
) {
1052 uint32_t surface
= surface_index
.ud
;
1053 uint32_t sampler
= sampler_index
.ud
;
1056 retype(dst
, BRW_REGISTER_TYPE_UW
),
1059 surface
+ base_binding_table_index
,
1062 inst
->size_written
/ REG_SIZE
,
1064 inst
->header_size
!= 0,
1068 brw_mark_surface_used(prog_data
, surface
+ base_binding_table_index
);
1070 /* Non-const sampler index */
1072 struct brw_reg addr
= vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD
));
1073 struct brw_reg surface_reg
= vec1(retype(surface_index
, BRW_REGISTER_TYPE_UD
));
1074 struct brw_reg sampler_reg
= vec1(retype(sampler_index
, BRW_REGISTER_TYPE_UD
));
1076 brw_push_insn_state(p
);
1077 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1078 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
1079 brw_set_default_exec_size(p
, BRW_EXECUTE_1
);
1081 if (brw_regs_equal(&surface_reg
, &sampler_reg
)) {
1082 brw_MUL(p
, addr
, sampler_reg
, brw_imm_uw(0x101));
1084 if (sampler_reg
.file
== BRW_IMMEDIATE_VALUE
) {
1085 brw_OR(p
, addr
, surface_reg
, brw_imm_ud(sampler_reg
.ud
<< 8));
1087 brw_SHL(p
, addr
, sampler_reg
, brw_imm_ud(8));
1088 brw_OR(p
, addr
, addr
, surface_reg
);
1091 if (base_binding_table_index
)
1092 brw_ADD(p
, addr
, addr
, brw_imm_ud(base_binding_table_index
));
1093 brw_AND(p
, addr
, addr
, brw_imm_ud(0xfff));
1095 brw_pop_insn_state(p
);
1097 /* dst = send(offset, a0.0 | <descriptor>) */
1098 brw_inst
*insn
= brw_send_indirect_message(
1099 p
, BRW_SFID_SAMPLER
, dst
, src
, addr
);
1100 brw_set_sampler_message(p
, insn
,
1104 inst
->size_written
/ REG_SIZE
,
1105 inst
->mlen
/* mlen */,
1106 inst
->header_size
!= 0 /* header */,
1110 /* visitor knows more than we do about the surface limit required,
1111 * so has already done marking.
1115 if (is_combined_send
) {
1116 brw_inst_set_eot(p
->devinfo
, brw_last_inst
, true);
1117 brw_inst_set_opcode(p
->devinfo
, brw_last_inst
, BRW_OPCODE_SENDC
);
1122 /* For OPCODE_DDX and OPCODE_DDY, per channel of output we've got input
1125 * arg0: ss0.tl ss0.tr ss0.bl ss0.br ss1.tl ss1.tr ss1.bl ss1.br
1127 * Ideally, we want to produce:
1130 * dst: (ss0.tr - ss0.tl) (ss0.tl - ss0.bl)
1131 * (ss0.tr - ss0.tl) (ss0.tr - ss0.br)
1132 * (ss0.br - ss0.bl) (ss0.tl - ss0.bl)
1133 * (ss0.br - ss0.bl) (ss0.tr - ss0.br)
1134 * (ss1.tr - ss1.tl) (ss1.tl - ss1.bl)
1135 * (ss1.tr - ss1.tl) (ss1.tr - ss1.br)
1136 * (ss1.br - ss1.bl) (ss1.tl - ss1.bl)
1137 * (ss1.br - ss1.bl) (ss1.tr - ss1.br)
1139 * and add another set of two more subspans if in 16-pixel dispatch mode.
1141 * For DDX, it ends up being easy: width = 2, horiz=0 gets us the same result
1142 * for each pair, and vertstride = 2 jumps us 2 elements after processing a
1143 * pair. But the ideal approximation may impose a huge performance cost on
1144 * sample_d. On at least Haswell, sample_d instruction does some
1145 * optimizations if the same LOD is used for all pixels in the subspan.
1147 * For DDY, we need to use ALIGN16 mode since it's capable of doing the
1148 * appropriate swizzling.
1151 fs_generator::generate_ddx(const fs_inst
*inst
,
1152 struct brw_reg dst
, struct brw_reg src
)
1154 unsigned vstride
, width
;
1156 if (inst
->opcode
== FS_OPCODE_DDX_FINE
) {
1157 /* produce accurate derivatives */
1158 vstride
= BRW_VERTICAL_STRIDE_2
;
1159 width
= BRW_WIDTH_2
;
1161 /* replicate the derivative at the top-left pixel to other pixels */
1162 vstride
= BRW_VERTICAL_STRIDE_4
;
1163 width
= BRW_WIDTH_4
;
1166 struct brw_reg src0
= src
;
1167 struct brw_reg src1
= src
;
1169 src0
.subnr
= sizeof(float);
1170 src0
.vstride
= vstride
;
1172 src0
.hstride
= BRW_HORIZONTAL_STRIDE_0
;
1173 src1
.vstride
= vstride
;
1175 src1
.hstride
= BRW_HORIZONTAL_STRIDE_0
;
1177 brw_ADD(p
, dst
, src0
, negate(src1
));
1180 /* The negate_value boolean is used to negate the derivative computation for
1181 * FBOs, since they place the origin at the upper left instead of the lower
1185 fs_generator::generate_ddy(const fs_inst
*inst
,
1186 struct brw_reg dst
, struct brw_reg src
)
1188 if (inst
->opcode
== FS_OPCODE_DDY_FINE
) {
1189 /* produce accurate derivatives */
1190 if (devinfo
->gen
>= 11) {
1191 src
= stride(src
, 0, 2, 1);
1192 struct brw_reg src_0
= byte_offset(src
, 0 * sizeof(float));
1193 struct brw_reg src_2
= byte_offset(src
, 2 * sizeof(float));
1194 struct brw_reg src_4
= byte_offset(src
, 4 * sizeof(float));
1195 struct brw_reg src_6
= byte_offset(src
, 6 * sizeof(float));
1196 struct brw_reg src_8
= byte_offset(src
, 8 * sizeof(float));
1197 struct brw_reg src_10
= byte_offset(src
, 10 * sizeof(float));
1198 struct brw_reg src_12
= byte_offset(src
, 12 * sizeof(float));
1199 struct brw_reg src_14
= byte_offset(src
, 14 * sizeof(float));
1201 struct brw_reg dst_0
= byte_offset(dst
, 0 * sizeof(float));
1202 struct brw_reg dst_4
= byte_offset(dst
, 4 * sizeof(float));
1203 struct brw_reg dst_8
= byte_offset(dst
, 8 * sizeof(float));
1204 struct brw_reg dst_12
= byte_offset(dst
, 12 * sizeof(float));
1206 brw_push_insn_state(p
);
1207 brw_set_default_exec_size(p
, BRW_EXECUTE_4
);
1209 brw_ADD(p
, dst_0
, negate(src_0
), src_2
);
1210 brw_ADD(p
, dst_4
, negate(src_4
), src_6
);
1212 if (inst
->exec_size
== 16) {
1213 brw_ADD(p
, dst_8
, negate(src_8
), src_10
);
1214 brw_ADD(p
, dst_12
, negate(src_12
), src_14
);
1217 brw_pop_insn_state(p
);
1219 struct brw_reg src0
= stride(src
, 4, 4, 1);
1220 struct brw_reg src1
= stride(src
, 4, 4, 1);
1221 src0
.swizzle
= BRW_SWIZZLE_XYXY
;
1222 src1
.swizzle
= BRW_SWIZZLE_ZWZW
;
1224 brw_push_insn_state(p
);
1225 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1226 brw_ADD(p
, dst
, negate(src0
), src1
);
1227 brw_pop_insn_state(p
);
1230 /* replicate the derivative at the top-left pixel to other pixels */
1231 struct brw_reg src0
= stride(src
, 4, 4, 0);
1232 struct brw_reg src1
= stride(src
, 4, 4, 0);
1233 src0
.subnr
= 0 * sizeof(float);
1234 src1
.subnr
= 2 * sizeof(float);
1236 brw_ADD(p
, dst
, negate(src0
), src1
);
1241 fs_generator::generate_discard_jump(fs_inst
*inst
)
1243 assert(devinfo
->gen
>= 6);
1245 /* This HALT will be patched up at FB write time to point UIP at the end of
1246 * the program, and at brw_uip_jip() JIP will be set to the end of the
1247 * current block (or the program).
1249 this->discard_halt_patches
.push_tail(new(mem_ctx
) ip_record(p
->nr_insn
));
1254 fs_generator::generate_scratch_write(fs_inst
*inst
, struct brw_reg src
)
1256 /* The 32-wide messages only respect the first 16-wide half of the channel
1257 * enable signals which are replicated identically for the second group of
1258 * 16 channels, so we cannot use them unless the write is marked
1259 * force_writemask_all.
1261 const unsigned lower_size
= inst
->force_writemask_all
? inst
->exec_size
:
1262 MIN2(16, inst
->exec_size
);
1263 const unsigned block_size
= 4 * lower_size
/ REG_SIZE
;
1264 assert(inst
->mlen
!= 0);
1266 brw_push_insn_state(p
);
1267 brw_set_default_exec_size(p
, cvt(lower_size
) - 1);
1268 brw_set_default_compression(p
, lower_size
> 8);
1270 for (unsigned i
= 0; i
< inst
->exec_size
/ lower_size
; i
++) {
1271 brw_set_default_group(p
, inst
->group
+ lower_size
* i
);
1273 brw_MOV(p
, brw_uvec_mrf(lower_size
, inst
->base_mrf
+ 1, 0),
1274 retype(offset(src
, block_size
* i
), BRW_REGISTER_TYPE_UD
));
1276 brw_oword_block_write_scratch(p
, brw_message_reg(inst
->base_mrf
),
1278 inst
->offset
+ block_size
* REG_SIZE
* i
);
1281 brw_pop_insn_state(p
);
1285 fs_generator::generate_scratch_read(fs_inst
*inst
, struct brw_reg dst
)
1287 assert(inst
->exec_size
<= 16 || inst
->force_writemask_all
);
1288 assert(inst
->mlen
!= 0);
1290 brw_oword_block_read_scratch(p
, dst
, brw_message_reg(inst
->base_mrf
),
1291 inst
->exec_size
/ 8, inst
->offset
);
1295 fs_generator::generate_scratch_read_gen7(fs_inst
*inst
, struct brw_reg dst
)
1297 assert(inst
->exec_size
<= 16 || inst
->force_writemask_all
);
1299 gen7_block_read_scratch(p
, dst
, inst
->exec_size
/ 8, inst
->offset
);
1303 fs_generator::generate_uniform_pull_constant_load(fs_inst
*inst
,
1305 struct brw_reg index
,
1306 struct brw_reg offset
)
1308 assert(type_sz(dst
.type
) == 4);
1309 assert(inst
->mlen
!= 0);
1311 assert(index
.file
== BRW_IMMEDIATE_VALUE
&&
1312 index
.type
== BRW_REGISTER_TYPE_UD
);
1313 uint32_t surf_index
= index
.ud
;
1315 assert(offset
.file
== BRW_IMMEDIATE_VALUE
&&
1316 offset
.type
== BRW_REGISTER_TYPE_UD
);
1317 uint32_t read_offset
= offset
.ud
;
1319 brw_oword_block_read(p
, dst
, brw_message_reg(inst
->base_mrf
),
1320 read_offset
, surf_index
);
1324 fs_generator::generate_uniform_pull_constant_load_gen7(fs_inst
*inst
,
1326 struct brw_reg index
,
1327 struct brw_reg payload
)
1329 assert(index
.type
== BRW_REGISTER_TYPE_UD
);
1330 assert(payload
.file
== BRW_GENERAL_REGISTER_FILE
);
1331 assert(type_sz(dst
.type
) == 4);
1333 if (index
.file
== BRW_IMMEDIATE_VALUE
) {
1334 const uint32_t surf_index
= index
.ud
;
1336 brw_push_insn_state(p
);
1337 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1338 brw_inst
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
1339 brw_pop_insn_state(p
);
1341 brw_set_dest(p
, send
, retype(dst
, BRW_REGISTER_TYPE_UD
));
1342 brw_set_src0(p
, send
, retype(payload
, BRW_REGISTER_TYPE_UD
));
1343 brw_set_dp_read_message(p
, send
, surf_index
,
1344 BRW_DATAPORT_OWORD_BLOCK_DWORDS(inst
->exec_size
),
1345 GEN7_DATAPORT_DC_OWORD_BLOCK_READ
,
1346 GEN6_SFID_DATAPORT_CONSTANT_CACHE
,
1349 DIV_ROUND_UP(inst
->size_written
, REG_SIZE
));
1352 struct brw_reg addr
= vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD
));
1354 brw_push_insn_state(p
);
1355 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1357 /* a0.0 = surf_index & 0xff */
1358 brw_inst
*insn_and
= brw_next_insn(p
, BRW_OPCODE_AND
);
1359 brw_inst_set_exec_size(p
->devinfo
, insn_and
, BRW_EXECUTE_1
);
1360 brw_set_dest(p
, insn_and
, addr
);
1361 brw_set_src0(p
, insn_and
, vec1(retype(index
, BRW_REGISTER_TYPE_UD
)));
1362 brw_set_src1(p
, insn_and
, brw_imm_ud(0x0ff));
1364 /* dst = send(payload, a0.0 | <descriptor>) */
1365 brw_inst
*insn
= brw_send_indirect_message(
1366 p
, GEN6_SFID_DATAPORT_CONSTANT_CACHE
,
1367 retype(dst
, BRW_REGISTER_TYPE_UD
),
1368 retype(payload
, BRW_REGISTER_TYPE_UD
), addr
);
1369 brw_set_dp_read_message(p
, insn
, 0 /* surface */,
1370 BRW_DATAPORT_OWORD_BLOCK_DWORDS(inst
->exec_size
),
1371 GEN7_DATAPORT_DC_OWORD_BLOCK_READ
,
1372 GEN6_SFID_DATAPORT_CONSTANT_CACHE
,
1375 DIV_ROUND_UP(inst
->size_written
, REG_SIZE
));
1377 brw_pop_insn_state(p
);
1382 fs_generator::generate_varying_pull_constant_load_gen4(fs_inst
*inst
,
1384 struct brw_reg index
)
1386 assert(devinfo
->gen
< 7); /* Should use the gen7 variant. */
1387 assert(inst
->header_size
!= 0);
1390 assert(index
.file
== BRW_IMMEDIATE_VALUE
&&
1391 index
.type
== BRW_REGISTER_TYPE_UD
);
1392 uint32_t surf_index
= index
.ud
;
1394 uint32_t simd_mode
, rlen
, msg_type
;
1395 if (inst
->exec_size
== 16) {
1396 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1399 assert(inst
->exec_size
== 8);
1400 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD8
;
1404 if (devinfo
->gen
>= 5)
1405 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
1407 /* We always use the SIMD16 message so that we only have to load U, and
1410 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_LD
;
1411 assert(inst
->mlen
== 3);
1412 assert(inst
->size_written
== 8 * REG_SIZE
);
1414 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1417 struct brw_reg header
= brw_vec8_grf(0, 0);
1418 gen6_resolve_implied_move(p
, &header
, inst
->base_mrf
);
1420 brw_inst
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
1421 brw_inst_set_compression(devinfo
, send
, false);
1422 brw_set_dest(p
, send
, retype(dst
, BRW_REGISTER_TYPE_UW
));
1423 brw_set_src0(p
, send
, header
);
1424 if (devinfo
->gen
< 6)
1425 brw_inst_set_base_mrf(p
->devinfo
, send
, inst
->base_mrf
);
1427 /* Our surface is set up as floats, regardless of what actual data is
1430 uint32_t return_format
= BRW_SAMPLER_RETURN_FORMAT_FLOAT32
;
1431 brw_set_sampler_message(p
, send
,
1433 0, /* sampler (unused) */
1437 inst
->header_size
!= 0,
1443 fs_generator::generate_varying_pull_constant_load_gen7(fs_inst
*inst
,
1445 struct brw_reg index
,
1446 struct brw_reg offset
)
1448 assert(devinfo
->gen
>= 7);
1449 /* Varying-offset pull constant loads are treated as a normal expression on
1450 * gen7, so the fact that it's a send message is hidden at the IR level.
1452 assert(inst
->header_size
== 0);
1453 assert(!inst
->mlen
);
1454 assert(index
.type
== BRW_REGISTER_TYPE_UD
);
1456 uint32_t simd_mode
, rlen
, mlen
;
1457 if (inst
->exec_size
== 16) {
1460 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1462 assert(inst
->exec_size
== 8);
1465 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD8
;
1468 if (index
.file
== BRW_IMMEDIATE_VALUE
) {
1470 uint32_t surf_index
= index
.ud
;
1472 brw_inst
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
1473 brw_set_dest(p
, send
, retype(dst
, BRW_REGISTER_TYPE_UW
));
1474 brw_set_src0(p
, send
, offset
);
1475 brw_set_sampler_message(p
, send
,
1477 0, /* LD message ignores sampler unit */
1478 GEN5_SAMPLER_MESSAGE_SAMPLE_LD
,
1481 false, /* no header */
1487 struct brw_reg addr
= vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD
));
1489 brw_push_insn_state(p
);
1490 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1492 /* a0.0 = surf_index & 0xff */
1493 brw_inst
*insn_and
= brw_next_insn(p
, BRW_OPCODE_AND
);
1494 brw_inst_set_exec_size(p
->devinfo
, insn_and
, BRW_EXECUTE_1
);
1495 brw_set_dest(p
, insn_and
, addr
);
1496 brw_set_src0(p
, insn_and
, vec1(retype(index
, BRW_REGISTER_TYPE_UD
)));
1497 brw_set_src1(p
, insn_and
, brw_imm_ud(0x0ff));
1499 brw_pop_insn_state(p
);
1501 /* dst = send(offset, a0.0 | <descriptor>) */
1502 brw_inst
*insn
= brw_send_indirect_message(
1503 p
, BRW_SFID_SAMPLER
, retype(dst
, BRW_REGISTER_TYPE_UW
),
1505 brw_set_sampler_message(p
, insn
,
1508 GEN5_SAMPLER_MESSAGE_SAMPLE_LD
,
1518 * Cause the current pixel/sample mask (from R1.7 bits 15:0) to be transferred
1519 * into the flags register (f0.0).
1521 * Used only on Gen6 and above.
1524 fs_generator::generate_mov_dispatch_to_flags(fs_inst
*inst
)
1526 struct brw_reg flags
= brw_flag_reg(0, inst
->flag_subreg
);
1527 struct brw_reg dispatch_mask
;
1529 if (devinfo
->gen
>= 6)
1530 dispatch_mask
= retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW
);
1532 dispatch_mask
= retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW
);
1534 brw_push_insn_state(p
);
1535 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1536 brw_set_default_exec_size(p
, BRW_EXECUTE_1
);
1537 brw_MOV(p
, flags
, dispatch_mask
);
1538 brw_pop_insn_state(p
);
1542 fs_generator::generate_pixel_interpolator_query(fs_inst
*inst
,
1545 struct brw_reg msg_data
,
1548 assert(inst
->size_written
% REG_SIZE
== 0);
1549 assert(msg_data
.type
== BRW_REGISTER_TYPE_UD
);
1551 brw_pixel_interpolator_query(p
,
1552 retype(dst
, BRW_REGISTER_TYPE_UW
),
1554 inst
->pi_noperspective
,
1558 inst
->size_written
/ REG_SIZE
);
1561 /* Sets vstride=1, width=4, hstride=0 of register src1 during
1562 * the ADD instruction.
1565 fs_generator::generate_set_sample_id(fs_inst
*inst
,
1567 struct brw_reg src0
,
1568 struct brw_reg src1
)
1570 assert(dst
.type
== BRW_REGISTER_TYPE_D
||
1571 dst
.type
== BRW_REGISTER_TYPE_UD
);
1572 assert(src0
.type
== BRW_REGISTER_TYPE_D
||
1573 src0
.type
== BRW_REGISTER_TYPE_UD
);
1575 struct brw_reg reg
= stride(src1
, 1, 4, 0);
1576 if (devinfo
->gen
>= 8 || inst
->exec_size
== 8) {
1577 brw_ADD(p
, dst
, src0
, reg
);
1578 } else if (inst
->exec_size
== 16) {
1579 brw_push_insn_state(p
);
1580 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
1581 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
1582 brw_ADD(p
, firsthalf(dst
), firsthalf(src0
), reg
);
1583 brw_set_default_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
1584 brw_ADD(p
, sechalf(dst
), sechalf(src0
), suboffset(reg
, 2));
1585 brw_pop_insn_state(p
);
1590 fs_generator::generate_pack_half_2x16_split(fs_inst
*inst
,
1595 assert(devinfo
->gen
>= 7);
1596 assert(dst
.type
== BRW_REGISTER_TYPE_UD
);
1597 assert(x
.type
== BRW_REGISTER_TYPE_F
);
1598 assert(y
.type
== BRW_REGISTER_TYPE_F
);
1600 /* From the Ivybridge PRM, Vol4, Part3, Section 6.27 f32to16:
1602 * Because this instruction does not have a 16-bit floating-point type,
1603 * the destination data type must be Word (W).
1605 * The destination must be DWord-aligned and specify a horizontal stride
1606 * (HorzStride) of 2. The 16-bit result is stored in the lower word of
1607 * each destination channel and the upper word is not modified.
1609 struct brw_reg dst_w
= spread(retype(dst
, BRW_REGISTER_TYPE_W
), 2);
1611 /* Give each 32-bit channel of dst the form below, where "." means
1615 brw_F32TO16(p
, dst_w
, y
);
1620 brw_SHL(p
, dst
, dst
, brw_imm_ud(16u));
1622 /* And, finally the form of packHalf2x16's output:
1625 brw_F32TO16(p
, dst_w
, x
);
1629 fs_generator::generate_unpack_half_2x16_split(fs_inst
*inst
,
1633 assert(devinfo
->gen
>= 7);
1634 assert(dst
.type
== BRW_REGISTER_TYPE_F
);
1635 assert(src
.type
== BRW_REGISTER_TYPE_UD
);
1637 /* From the Ivybridge PRM, Vol4, Part3, Section 6.26 f16to32:
1639 * Because this instruction does not have a 16-bit floating-point type,
1640 * the source data type must be Word (W). The destination type must be
1643 struct brw_reg src_w
= spread(retype(src
, BRW_REGISTER_TYPE_W
), 2);
1645 /* Each channel of src has the form of unpackHalf2x16's input: 0xhhhhllll.
1646 * For the Y case, we wish to access only the upper word; therefore
1647 * a 16-bit subregister offset is needed.
1649 assert(inst
->opcode
== FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X
||
1650 inst
->opcode
== FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y
);
1651 if (inst
->opcode
== FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y
)
1654 brw_F16TO32(p
, dst
, src_w
);
1658 fs_generator::generate_shader_time_add(fs_inst
*inst
,
1659 struct brw_reg payload
,
1660 struct brw_reg offset
,
1661 struct brw_reg value
)
1663 assert(devinfo
->gen
>= 7);
1664 brw_push_insn_state(p
);
1665 brw_set_default_mask_control(p
, true);
1667 assert(payload
.file
== BRW_GENERAL_REGISTER_FILE
);
1668 struct brw_reg payload_offset
= retype(brw_vec1_grf(payload
.nr
, 0),
1670 struct brw_reg payload_value
= retype(brw_vec1_grf(payload
.nr
+ 1, 0),
1673 assert(offset
.file
== BRW_IMMEDIATE_VALUE
);
1674 if (value
.file
== BRW_GENERAL_REGISTER_FILE
) {
1675 value
.width
= BRW_WIDTH_1
;
1676 value
.hstride
= BRW_HORIZONTAL_STRIDE_0
;
1677 value
.vstride
= BRW_VERTICAL_STRIDE_0
;
1679 assert(value
.file
== BRW_IMMEDIATE_VALUE
);
1682 /* Trying to deal with setup of the params from the IR is crazy in the FS8
1683 * case, and we don't really care about squeezing every bit of performance
1684 * out of this path, so we just emit the MOVs from here.
1686 brw_MOV(p
, payload_offset
, offset
);
1687 brw_MOV(p
, payload_value
, value
);
1688 brw_shader_time_add(p
, payload
,
1689 prog_data
->binding_table
.shader_time_start
);
1690 brw_pop_insn_state(p
);
1692 brw_mark_surface_used(prog_data
,
1693 prog_data
->binding_table
.shader_time_start
);
1697 fs_generator::enable_debug(const char *shader_name
)
1700 this->shader_name
= shader_name
;
1704 fs_generator::generate_code(const cfg_t
*cfg
, int dispatch_width
)
1706 /* align to 64 byte boundary. */
1707 while (p
->next_insn_offset
% 64)
1710 this->dispatch_width
= dispatch_width
;
1712 int start_offset
= p
->next_insn_offset
;
1713 int spill_count
= 0, fill_count
= 0;
1716 struct disasm_info
*disasm_info
= disasm_initialize(devinfo
, cfg
);
1718 foreach_block_and_inst (block
, fs_inst
, inst
, cfg
) {
1719 struct brw_reg src
[3], dst
;
1720 unsigned int last_insn_offset
= p
->next_insn_offset
;
1721 bool multiple_instructions_emitted
= false;
1723 /* From the Broadwell PRM, Volume 7, "3D-Media-GPGPU", in the
1724 * "Register Region Restrictions" section: for BDW, SKL:
1726 * "A POW/FDIV operation must not be followed by an instruction
1727 * that requires two destination registers."
1729 * The documentation is often lacking annotations for Atom parts,
1730 * and empirically this affects CHV as well.
1732 if (devinfo
->gen
>= 8 &&
1733 devinfo
->gen
<= 9 &&
1735 brw_inst_opcode(devinfo
, brw_last_inst
) == BRW_OPCODE_MATH
&&
1736 brw_inst_math_function(devinfo
, brw_last_inst
) == BRW_MATH_FUNCTION_POW
&&
1737 inst
->dst
.component_size(inst
->exec_size
) > REG_SIZE
) {
1739 last_insn_offset
= p
->next_insn_offset
;
1742 if (unlikely(debug_flag
))
1743 disasm_annotate(disasm_info
, inst
, p
->next_insn_offset
);
1745 /* If the instruction writes to more than one register, it needs to be
1746 * explicitly marked as compressed on Gen <= 5. On Gen >= 6 the
1747 * hardware figures out by itself what the right compression mode is,
1748 * but we still need to know whether the instruction is compressed to
1749 * set up the source register regions appropriately.
1751 * XXX - This is wrong for instructions that write a single register but
1752 * read more than one which should strictly speaking be treated as
1753 * compressed. For instructions that don't write any registers it
1754 * relies on the destination being a null register of the correct
1755 * type and regioning so the instruction is considered compressed
1756 * or not accordingly.
1758 const bool compressed
=
1759 inst
->dst
.component_size(inst
->exec_size
) > REG_SIZE
;
1760 brw_set_default_compression(p
, compressed
);
1761 brw_set_default_group(p
, inst
->group
);
1763 for (unsigned int i
= 0; i
< inst
->sources
; i
++) {
1764 src
[i
] = brw_reg_from_fs_reg(devinfo
, inst
,
1765 &inst
->src
[i
], compressed
);
1766 /* The accumulator result appears to get used for the
1767 * conditional modifier generation. When negating a UD
1768 * value, there is a 33rd bit generated for the sign in the
1769 * accumulator value, so now you can't check, for example,
1770 * equality with a 32-bit value. See piglit fs-op-neg-uvec4.
1772 assert(!inst
->conditional_mod
||
1773 inst
->src
[i
].type
!= BRW_REGISTER_TYPE_UD
||
1774 !inst
->src
[i
].negate
);
1776 dst
= brw_reg_from_fs_reg(devinfo
, inst
,
1777 &inst
->dst
, compressed
);
1779 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
1780 brw_set_default_predicate_control(p
, inst
->predicate
);
1781 brw_set_default_predicate_inverse(p
, inst
->predicate_inverse
);
1782 brw_set_default_flag_reg(p
, 0, inst
->flag_subreg
);
1783 brw_set_default_saturate(p
, inst
->saturate
);
1784 brw_set_default_mask_control(p
, inst
->force_writemask_all
);
1785 brw_set_default_acc_write_control(p
, inst
->writes_accumulator
);
1787 unsigned exec_size
= inst
->exec_size
;
1788 if (devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
1789 (get_exec_type_size(inst
) == 8 || type_sz(inst
->dst
.type
) == 8)) {
1793 brw_set_default_exec_size(p
, cvt(exec_size
) - 1);
1795 assert(inst
->force_writemask_all
|| inst
->exec_size
>= 4);
1796 assert(inst
->force_writemask_all
|| inst
->group
% inst
->exec_size
== 0);
1797 assert(inst
->base_mrf
+ inst
->mlen
<= BRW_MAX_MRF(devinfo
->gen
));
1798 assert(inst
->mlen
<= BRW_MAX_MSG_LENGTH
);
1800 switch (inst
->opcode
) {
1801 case BRW_OPCODE_MOV
:
1802 brw_MOV(p
, dst
, src
[0]);
1804 case BRW_OPCODE_ADD
:
1805 brw_ADD(p
, dst
, src
[0], src
[1]);
1807 case BRW_OPCODE_MUL
:
1808 brw_MUL(p
, dst
, src
[0], src
[1]);
1810 case BRW_OPCODE_AVG
:
1811 brw_AVG(p
, dst
, src
[0], src
[1]);
1813 case BRW_OPCODE_MACH
:
1814 brw_MACH(p
, dst
, src
[0], src
[1]);
1817 case BRW_OPCODE_LINE
:
1818 brw_LINE(p
, dst
, src
[0], src
[1]);
1821 case BRW_OPCODE_MAD
:
1822 assert(devinfo
->gen
>= 6);
1823 if (devinfo
->gen
< 10)
1824 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1825 brw_MAD(p
, dst
, src
[0], src
[1], src
[2]);
1828 case BRW_OPCODE_LRP
:
1829 assert(devinfo
->gen
>= 6 && devinfo
->gen
<= 10);
1830 if (devinfo
->gen
< 10)
1831 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1832 brw_LRP(p
, dst
, src
[0], src
[1], src
[2]);
1835 case BRW_OPCODE_FRC
:
1836 brw_FRC(p
, dst
, src
[0]);
1838 case BRW_OPCODE_RNDD
:
1839 brw_RNDD(p
, dst
, src
[0]);
1841 case BRW_OPCODE_RNDE
:
1842 brw_RNDE(p
, dst
, src
[0]);
1844 case BRW_OPCODE_RNDZ
:
1845 brw_RNDZ(p
, dst
, src
[0]);
1848 case BRW_OPCODE_AND
:
1849 brw_AND(p
, dst
, src
[0], src
[1]);
1852 brw_OR(p
, dst
, src
[0], src
[1]);
1854 case BRW_OPCODE_XOR
:
1855 brw_XOR(p
, dst
, src
[0], src
[1]);
1857 case BRW_OPCODE_NOT
:
1858 brw_NOT(p
, dst
, src
[0]);
1860 case BRW_OPCODE_ASR
:
1861 brw_ASR(p
, dst
, src
[0], src
[1]);
1863 case BRW_OPCODE_SHR
:
1864 brw_SHR(p
, dst
, src
[0], src
[1]);
1866 case BRW_OPCODE_SHL
:
1867 brw_SHL(p
, dst
, src
[0], src
[1]);
1869 case BRW_OPCODE_F32TO16
:
1870 assert(devinfo
->gen
>= 7);
1871 brw_F32TO16(p
, dst
, src
[0]);
1873 case BRW_OPCODE_F16TO32
:
1874 assert(devinfo
->gen
>= 7);
1875 brw_F16TO32(p
, dst
, src
[0]);
1877 case BRW_OPCODE_CMP
:
1878 if (inst
->exec_size
>= 16 && devinfo
->gen
== 7 && !devinfo
->is_haswell
&&
1879 dst
.file
== BRW_ARCHITECTURE_REGISTER_FILE
) {
1880 /* For unknown reasons the WaCMPInstFlagDepClearedEarly workaround
1881 * implemented in the compiler is not sufficient. Overriding the
1882 * type when the destination is the null register is necessary but
1883 * not sufficient by itself.
1885 assert(dst
.nr
== BRW_ARF_NULL
);
1886 dst
.type
= BRW_REGISTER_TYPE_D
;
1888 brw_CMP(p
, dst
, inst
->conditional_mod
, src
[0], src
[1]);
1890 case BRW_OPCODE_SEL
:
1891 brw_SEL(p
, dst
, src
[0], src
[1]);
1893 case BRW_OPCODE_BFREV
:
1894 assert(devinfo
->gen
>= 7);
1895 brw_BFREV(p
, retype(dst
, BRW_REGISTER_TYPE_UD
),
1896 retype(src
[0], BRW_REGISTER_TYPE_UD
));
1898 case BRW_OPCODE_FBH
:
1899 assert(devinfo
->gen
>= 7);
1900 brw_FBH(p
, retype(dst
, src
[0].type
), src
[0]);
1902 case BRW_OPCODE_FBL
:
1903 assert(devinfo
->gen
>= 7);
1904 brw_FBL(p
, retype(dst
, BRW_REGISTER_TYPE_UD
),
1905 retype(src
[0], BRW_REGISTER_TYPE_UD
));
1907 case BRW_OPCODE_LZD
:
1908 brw_LZD(p
, dst
, src
[0]);
1910 case BRW_OPCODE_CBIT
:
1911 assert(devinfo
->gen
>= 7);
1912 brw_CBIT(p
, retype(dst
, BRW_REGISTER_TYPE_UD
),
1913 retype(src
[0], BRW_REGISTER_TYPE_UD
));
1915 case BRW_OPCODE_ADDC
:
1916 assert(devinfo
->gen
>= 7);
1917 brw_ADDC(p
, dst
, src
[0], src
[1]);
1919 case BRW_OPCODE_SUBB
:
1920 assert(devinfo
->gen
>= 7);
1921 brw_SUBB(p
, dst
, src
[0], src
[1]);
1923 case BRW_OPCODE_MAC
:
1924 brw_MAC(p
, dst
, src
[0], src
[1]);
1927 case BRW_OPCODE_BFE
:
1928 assert(devinfo
->gen
>= 7);
1929 if (devinfo
->gen
< 10)
1930 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1931 brw_BFE(p
, dst
, src
[0], src
[1], src
[2]);
1934 case BRW_OPCODE_BFI1
:
1935 assert(devinfo
->gen
>= 7);
1936 brw_BFI1(p
, dst
, src
[0], src
[1]);
1938 case BRW_OPCODE_BFI2
:
1939 assert(devinfo
->gen
>= 7);
1940 if (devinfo
->gen
< 10)
1941 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1942 brw_BFI2(p
, dst
, src
[0], src
[1], src
[2]);
1946 if (inst
->src
[0].file
!= BAD_FILE
) {
1947 /* The instruction has an embedded compare (only allowed on gen6) */
1948 assert(devinfo
->gen
== 6);
1949 gen6_IF(p
, inst
->conditional_mod
, src
[0], src
[1]);
1951 brw_IF(p
, brw_inst_exec_size(devinfo
, p
->current
));
1955 case BRW_OPCODE_ELSE
:
1958 case BRW_OPCODE_ENDIF
:
1963 brw_DO(p
, brw_inst_exec_size(devinfo
, p
->current
));
1966 case BRW_OPCODE_BREAK
:
1969 case BRW_OPCODE_CONTINUE
:
1973 case BRW_OPCODE_WHILE
:
1978 case SHADER_OPCODE_RCP
:
1979 case SHADER_OPCODE_RSQ
:
1980 case SHADER_OPCODE_SQRT
:
1981 case SHADER_OPCODE_EXP2
:
1982 case SHADER_OPCODE_LOG2
:
1983 case SHADER_OPCODE_SIN
:
1984 case SHADER_OPCODE_COS
:
1985 assert(inst
->conditional_mod
== BRW_CONDITIONAL_NONE
);
1986 if (devinfo
->gen
>= 6) {
1987 assert(inst
->mlen
== 0);
1988 assert(devinfo
->gen
>= 7 || inst
->exec_size
== 8);
1989 gen6_math(p
, dst
, brw_math_function(inst
->opcode
),
1990 src
[0], brw_null_reg());
1992 assert(inst
->mlen
>= 1);
1993 assert(devinfo
->gen
== 5 || devinfo
->is_g4x
|| inst
->exec_size
== 8);
1995 brw_math_function(inst
->opcode
),
1996 inst
->base_mrf
, src
[0],
1997 BRW_MATH_PRECISION_FULL
);
2000 case SHADER_OPCODE_INT_QUOTIENT
:
2001 case SHADER_OPCODE_INT_REMAINDER
:
2002 case SHADER_OPCODE_POW
:
2003 assert(inst
->conditional_mod
== BRW_CONDITIONAL_NONE
);
2004 if (devinfo
->gen
>= 6) {
2005 assert(inst
->mlen
== 0);
2006 assert((devinfo
->gen
>= 7 && inst
->opcode
== SHADER_OPCODE_POW
) ||
2007 inst
->exec_size
== 8);
2008 gen6_math(p
, dst
, brw_math_function(inst
->opcode
), src
[0], src
[1]);
2010 assert(inst
->mlen
>= 1);
2011 assert(inst
->exec_size
== 8);
2012 gen4_math(p
, dst
, brw_math_function(inst
->opcode
),
2013 inst
->base_mrf
, src
[0],
2014 BRW_MATH_PRECISION_FULL
);
2017 case FS_OPCODE_CINTERP
:
2018 brw_MOV(p
, dst
, src
[0]);
2020 case FS_OPCODE_LINTERP
:
2021 multiple_instructions_emitted
= generate_linterp(inst
, dst
, src
);
2023 case FS_OPCODE_PIXEL_X
:
2024 assert(src
[0].type
== BRW_REGISTER_TYPE_UW
);
2025 src
[0].subnr
= 0 * type_sz(src
[0].type
);
2026 brw_MOV(p
, dst
, stride(src
[0], 8, 4, 1));
2028 case FS_OPCODE_PIXEL_Y
:
2029 assert(src
[0].type
== BRW_REGISTER_TYPE_UW
);
2030 src
[0].subnr
= 4 * type_sz(src
[0].type
);
2031 brw_MOV(p
, dst
, stride(src
[0], 8, 4, 1));
2033 case SHADER_OPCODE_GET_BUFFER_SIZE
:
2034 generate_get_buffer_size(inst
, dst
, src
[0], src
[1]);
2036 case SHADER_OPCODE_TEX
:
2038 case SHADER_OPCODE_TXD
:
2039 case SHADER_OPCODE_TXF
:
2040 case SHADER_OPCODE_TXF_LZ
:
2041 case SHADER_OPCODE_TXF_CMS
:
2042 case SHADER_OPCODE_TXF_CMS_W
:
2043 case SHADER_OPCODE_TXF_UMS
:
2044 case SHADER_OPCODE_TXF_MCS
:
2045 case SHADER_OPCODE_TXL
:
2046 case SHADER_OPCODE_TXL_LZ
:
2047 case SHADER_OPCODE_TXS
:
2048 case SHADER_OPCODE_LOD
:
2049 case SHADER_OPCODE_TG4
:
2050 case SHADER_OPCODE_TG4_OFFSET
:
2051 case SHADER_OPCODE_SAMPLEINFO
:
2052 generate_tex(inst
, dst
, src
[0], src
[1], src
[2]);
2054 case FS_OPCODE_DDX_COARSE
:
2055 case FS_OPCODE_DDX_FINE
:
2056 generate_ddx(inst
, dst
, src
[0]);
2058 case FS_OPCODE_DDY_COARSE
:
2059 case FS_OPCODE_DDY_FINE
:
2060 generate_ddy(inst
, dst
, src
[0]);
2063 case SHADER_OPCODE_GEN4_SCRATCH_WRITE
:
2064 generate_scratch_write(inst
, src
[0]);
2068 case SHADER_OPCODE_GEN4_SCRATCH_READ
:
2069 generate_scratch_read(inst
, dst
);
2073 case SHADER_OPCODE_GEN7_SCRATCH_READ
:
2074 generate_scratch_read_gen7(inst
, dst
);
2078 case SHADER_OPCODE_MOV_INDIRECT
:
2079 generate_mov_indirect(inst
, dst
, src
[0], src
[1]);
2082 case SHADER_OPCODE_URB_READ_SIMD8
:
2083 case SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
:
2084 generate_urb_read(inst
, dst
, src
[0]);
2087 case SHADER_OPCODE_URB_WRITE_SIMD8
:
2088 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
2089 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
2090 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
2091 generate_urb_write(inst
, src
[0]);
2094 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
2095 assert(inst
->force_writemask_all
);
2096 generate_uniform_pull_constant_load(inst
, dst
, src
[0], src
[1]);
2099 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
:
2100 assert(inst
->force_writemask_all
);
2101 generate_uniform_pull_constant_load_gen7(inst
, dst
, src
[0], src
[1]);
2104 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN4
:
2105 generate_varying_pull_constant_load_gen4(inst
, dst
, src
[0]);
2108 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7
:
2109 generate_varying_pull_constant_load_gen7(inst
, dst
, src
[0], src
[1]);
2112 case FS_OPCODE_REP_FB_WRITE
:
2113 case FS_OPCODE_FB_WRITE
:
2114 generate_fb_write(inst
, src
[0]);
2117 case FS_OPCODE_FB_READ
:
2118 generate_fb_read(inst
, dst
, src
[0]);
2121 case FS_OPCODE_MOV_DISPATCH_TO_FLAGS
:
2122 generate_mov_dispatch_to_flags(inst
);
2125 case FS_OPCODE_DISCARD_JUMP
:
2126 generate_discard_jump(inst
);
2129 case SHADER_OPCODE_SHADER_TIME_ADD
:
2130 generate_shader_time_add(inst
, src
[0], src
[1], src
[2]);
2133 case SHADER_OPCODE_UNTYPED_ATOMIC
:
2134 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2135 brw_untyped_atomic(p
, dst
, src
[0], src
[1], src
[2].ud
,
2136 inst
->mlen
, !inst
->dst
.is_null());
2139 case SHADER_OPCODE_UNTYPED_SURFACE_READ
:
2140 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2141 brw_untyped_surface_read(p
, dst
, src
[0], src
[1],
2142 inst
->mlen
, src
[2].ud
);
2145 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE
:
2146 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2147 brw_untyped_surface_write(p
, src
[0], src
[1],
2148 inst
->mlen
, src
[2].ud
);
2151 case SHADER_OPCODE_BYTE_SCATTERED_READ
:
2152 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2153 brw_byte_scattered_read(p
, dst
, src
[0], src
[1],
2154 inst
->mlen
, src
[2].ud
);
2157 case SHADER_OPCODE_BYTE_SCATTERED_WRITE
:
2158 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2159 brw_byte_scattered_write(p
, src
[0], src
[1],
2160 inst
->mlen
, src
[2].ud
);
2163 case SHADER_OPCODE_TYPED_ATOMIC
:
2164 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2165 brw_typed_atomic(p
, dst
, src
[0], src
[1],
2166 src
[2].ud
, inst
->mlen
, !inst
->dst
.is_null());
2169 case SHADER_OPCODE_TYPED_SURFACE_READ
:
2170 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2171 brw_typed_surface_read(p
, dst
, src
[0], src
[1],
2172 inst
->mlen
, src
[2].ud
);
2175 case SHADER_OPCODE_TYPED_SURFACE_WRITE
:
2176 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2177 brw_typed_surface_write(p
, src
[0], src
[1], inst
->mlen
, src
[2].ud
);
2180 case SHADER_OPCODE_MEMORY_FENCE
:
2181 brw_memory_fence(p
, dst
);
2184 case SHADER_OPCODE_FIND_LIVE_CHANNEL
: {
2185 const struct brw_reg mask
=
2186 brw_stage_has_packed_dispatch(devinfo
, stage
,
2187 prog_data
) ? brw_imm_ud(~0u) :
2188 stage
== MESA_SHADER_FRAGMENT
? brw_vmask_reg() :
2190 brw_find_live_channel(p
, dst
, mask
);
2194 case SHADER_OPCODE_BROADCAST
:
2195 assert(inst
->force_writemask_all
);
2196 brw_broadcast(p
, dst
, src
[0], src
[1]);
2199 case FS_OPCODE_SET_SAMPLE_ID
:
2200 generate_set_sample_id(inst
, dst
, src
[0], src
[1]);
2203 case FS_OPCODE_PACK_HALF_2x16_SPLIT
:
2204 generate_pack_half_2x16_split(inst
, dst
, src
[0], src
[1]);
2207 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X
:
2208 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y
:
2209 generate_unpack_half_2x16_split(inst
, dst
, src
[0]);
2212 case FS_OPCODE_PLACEHOLDER_HALT
:
2213 /* This is the place where the final HALT needs to be inserted if
2214 * we've emitted any discards. If not, this will emit no code.
2216 if (!patch_discard_jumps_to_fb_writes()) {
2217 if (unlikely(debug_flag
)) {
2218 disasm_info
->use_tail
= true;
2223 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
2224 generate_pixel_interpolator_query(inst
, dst
, src
[0], src
[1],
2225 GEN7_PIXEL_INTERPOLATOR_LOC_SAMPLE
);
2228 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
2229 generate_pixel_interpolator_query(inst
, dst
, src
[0], src
[1],
2230 GEN7_PIXEL_INTERPOLATOR_LOC_SHARED_OFFSET
);
2233 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
2234 generate_pixel_interpolator_query(inst
, dst
, src
[0], src
[1],
2235 GEN7_PIXEL_INTERPOLATOR_LOC_PER_SLOT_OFFSET
);
2238 case CS_OPCODE_CS_TERMINATE
:
2239 generate_cs_terminate(inst
, src
[0]);
2242 case SHADER_OPCODE_BARRIER
:
2243 generate_barrier(inst
, src
[0]);
2246 case BRW_OPCODE_DIM
:
2247 assert(devinfo
->is_haswell
);
2248 assert(src
[0].type
== BRW_REGISTER_TYPE_DF
);
2249 assert(dst
.type
== BRW_REGISTER_TYPE_DF
);
2250 brw_DIM(p
, dst
, retype(src
[0], BRW_REGISTER_TYPE_F
));
2253 case SHADER_OPCODE_RND_MODE
:
2254 assert(src
[0].file
== BRW_IMMEDIATE_VALUE
);
2255 brw_rounding_mode(p
, (brw_rnd_mode
) src
[0].d
);
2259 unreachable("Unsupported opcode");
2261 case SHADER_OPCODE_LOAD_PAYLOAD
:
2262 unreachable("Should be lowered by lower_load_payload()");
2265 if (multiple_instructions_emitted
)
2268 if (inst
->no_dd_clear
|| inst
->no_dd_check
|| inst
->conditional_mod
) {
2269 assert(p
->next_insn_offset
== last_insn_offset
+ 16 ||
2270 !"conditional_mod, no_dd_check, or no_dd_clear set for IR "
2271 "emitting more than 1 instruction");
2273 brw_inst
*last
= &p
->store
[last_insn_offset
/ 16];
2275 if (inst
->conditional_mod
)
2276 brw_inst_set_cond_modifier(p
->devinfo
, last
, inst
->conditional_mod
);
2277 brw_inst_set_no_dd_clear(p
->devinfo
, last
, inst
->no_dd_clear
);
2278 brw_inst_set_no_dd_check(p
->devinfo
, last
, inst
->no_dd_check
);
2282 brw_set_uip_jip(p
, start_offset
);
2284 /* end of program sentinel */
2285 disasm_new_inst_group(disasm_info
, p
->next_insn_offset
);
2290 if (unlikely(debug_flag
))
2292 brw_validate_instructions(devinfo
, p
->store
,
2294 p
->next_insn_offset
,
2297 int before_size
= p
->next_insn_offset
- start_offset
;
2298 brw_compact_instructions(p
, start_offset
, disasm_info
);
2299 int after_size
= p
->next_insn_offset
- start_offset
;
2301 if (unlikely(debug_flag
)) {
2302 fprintf(stderr
, "Native code for %s\n"
2303 "SIMD%d shader: %d instructions. %d loops. %u cycles. %d:%d spills:fills. Promoted %u constants. Compacted %d to %d"
2304 " bytes (%.0f%%)\n",
2305 shader_name
, dispatch_width
, before_size
/ 16, loop_count
, cfg
->cycle_count
,
2306 spill_count
, fill_count
, promoted_constants
, before_size
, after_size
,
2307 100.0f
* (before_size
- after_size
) / before_size
);
2309 dump_assembly(p
->store
, disasm_info
);
2311 ralloc_free(disasm_info
);
2314 compiler
->shader_debug_log(log_data
,
2315 "%s SIMD%d shader: %d inst, %d loops, %u cycles, "
2316 "%d:%d spills:fills, Promoted %u constants, "
2317 "compacted %d to %d bytes.",
2318 _mesa_shader_stage_to_abbrev(stage
),
2319 dispatch_width
, before_size
/ 16,
2320 loop_count
, cfg
->cycle_count
, spill_count
,
2321 fill_count
, promoted_constants
, before_size
,
2324 return start_offset
;
2328 fs_generator::get_assembly(unsigned int *assembly_size
)
2330 return brw_get_program(p
, assembly_size
);