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.
33 #include "brw_program.h"
35 static enum brw_reg_file
36 brw_file_from_reg(fs_reg
*reg
)
40 return BRW_ARCHITECTURE_REGISTER_FILE
;
43 return BRW_GENERAL_REGISTER_FILE
;
45 return BRW_MESSAGE_REGISTER_FILE
;
47 return BRW_IMMEDIATE_VALUE
;
51 unreachable("not reached");
53 return BRW_ARCHITECTURE_REGISTER_FILE
;
57 brw_reg_from_fs_reg(fs_inst
*inst
, fs_reg
*reg
, unsigned gen
)
59 struct brw_reg brw_reg
;
63 assert((reg
->nr
& ~BRW_MRF_COMPR4
) < BRW_MAX_MRF(gen
));
66 if (reg
->stride
== 0) {
67 brw_reg
= brw_vec1_reg(brw_file_from_reg(reg
), reg
->nr
, 0);
68 } else if (inst
->exec_size
< 8) {
69 brw_reg
= brw_vec8_reg(brw_file_from_reg(reg
), reg
->nr
, 0);
70 brw_reg
= stride(brw_reg
, inst
->exec_size
* reg
->stride
,
71 inst
->exec_size
, reg
->stride
);
73 /* From the Haswell PRM:
75 * VertStride must be used to cross GRF register boundaries. This
76 * rule implies that elements within a 'Width' cannot cross GRF
79 * So, for registers with width > 8, we have to use a width of 8
80 * and trust the compression state to sort out the exec size.
82 brw_reg
= brw_vec8_reg(brw_file_from_reg(reg
), reg
->nr
, 0);
83 brw_reg
= stride(brw_reg
, 8 * reg
->stride
, 8, reg
->stride
);
86 brw_reg
= retype(brw_reg
, reg
->type
);
87 brw_reg
= byte_offset(brw_reg
, reg
->subreg_offset
);
88 brw_reg
.abs
= reg
->abs
;
89 brw_reg
.negate
= reg
->negate
;
94 brw_reg
= reg
->as_brw_reg();
97 /* Probably unused. */
98 brw_reg
= brw_null_reg();
102 unreachable("not reached");
108 fs_generator::fs_generator(const struct brw_compiler
*compiler
, void *log_data
,
111 struct brw_stage_prog_data
*prog_data
,
112 unsigned promoted_constants
,
113 bool runtime_check_aads_emit
,
114 gl_shader_stage stage
)
116 : compiler(compiler
), log_data(log_data
),
117 devinfo(compiler
->devinfo
), key(key
),
118 prog_data(prog_data
),
119 promoted_constants(promoted_constants
),
120 runtime_check_aads_emit(runtime_check_aads_emit
), debug_flag(false),
121 stage(stage
), mem_ctx(mem_ctx
)
123 p
= rzalloc(mem_ctx
, struct brw_codegen
);
124 brw_init_codegen(devinfo
, p
, mem_ctx
);
127 fs_generator::~fs_generator()
131 class ip_record
: public exec_node
{
133 DECLARE_RALLOC_CXX_OPERATORS(ip_record
)
144 fs_generator::patch_discard_jumps_to_fb_writes()
146 if (devinfo
->gen
< 6 || this->discard_halt_patches
.is_empty())
149 int scale
= brw_jump_scale(p
->devinfo
);
151 /* There is a somewhat strange undocumented requirement of using
152 * HALT, according to the simulator. If some channel has HALTed to
153 * a particular UIP, then by the end of the program, every channel
154 * must have HALTed to that UIP. Furthermore, the tracking is a
155 * stack, so you can't do the final halt of a UIP after starting
156 * halting to a new UIP.
158 * Symptoms of not emitting this instruction on actual hardware
159 * included GPU hangs and sparkly rendering on the piglit discard
162 brw_inst
*last_halt
= gen6_HALT(p
);
163 brw_inst_set_uip(p
->devinfo
, last_halt
, 1 * scale
);
164 brw_inst_set_jip(p
->devinfo
, last_halt
, 1 * scale
);
168 foreach_in_list(ip_record
, patch_ip
, &discard_halt_patches
) {
169 brw_inst
*patch
= &p
->store
[patch_ip
->ip
];
171 assert(brw_inst_opcode(p
->devinfo
, patch
) == BRW_OPCODE_HALT
);
172 /* HALT takes a half-instruction distance from the pre-incremented IP. */
173 brw_inst_set_uip(p
->devinfo
, patch
, (ip
- patch_ip
->ip
) * scale
);
176 this->discard_halt_patches
.make_empty();
181 fs_generator::fire_fb_write(fs_inst
*inst
,
182 struct brw_reg payload
,
183 struct brw_reg implied_header
,
186 uint32_t msg_control
;
188 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
190 if (devinfo
->gen
< 6) {
191 brw_push_insn_state(p
);
192 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
193 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
194 brw_set_default_predicate_control(p
, BRW_PREDICATE_NONE
);
195 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
196 brw_MOV(p
, offset(payload
, 1), brw_vec8_grf(1, 0));
197 brw_pop_insn_state(p
);
200 if (inst
->opcode
== FS_OPCODE_REP_FB_WRITE
)
201 msg_control
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE_REPLICATED
;
202 else if (prog_data
->dual_src_blend
) {
203 if (!inst
->force_sechalf
)
204 msg_control
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_DUAL_SOURCE_SUBSPAN01
;
206 msg_control
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_DUAL_SOURCE_SUBSPAN23
;
207 } else if (inst
->exec_size
== 16)
208 msg_control
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE
;
210 msg_control
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_SINGLE_SOURCE_SUBSPAN01
;
212 uint32_t surf_index
=
213 prog_data
->binding_table
.render_target_start
+ inst
->target
;
215 bool last_render_target
= inst
->eot
||
216 (prog_data
->dual_src_blend
&& dispatch_width
== 16);
229 inst
->header_size
!= 0);
231 brw_mark_surface_used(&prog_data
->base
, surf_index
);
235 fs_generator::generate_fb_write(fs_inst
*inst
, struct brw_reg payload
)
237 brw_wm_prog_data
*prog_data
= (brw_wm_prog_data
*) this->prog_data
;
238 const brw_wm_prog_key
* const key
= (brw_wm_prog_key
* const) this->key
;
239 struct brw_reg implied_header
;
241 if (devinfo
->gen
< 8 && !devinfo
->is_haswell
) {
242 brw_set_default_predicate_control(p
, BRW_PREDICATE_NONE
);
245 if (inst
->base_mrf
>= 0)
246 payload
= brw_message_reg(inst
->base_mrf
);
248 /* Header is 2 regs, g0 and g1 are the contents. g0 will be implied
251 if (inst
->header_size
!= 0) {
252 brw_push_insn_state(p
);
253 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
254 brw_set_default_predicate_control(p
, BRW_PREDICATE_NONE
);
255 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
256 brw_set_default_flag_reg(p
, 0, 0);
258 /* On HSW, the GPU will use the predicate on SENDC, unless the header is
261 if (prog_data
->uses_kill
) {
262 struct brw_reg pixel_mask
;
264 if (devinfo
->gen
>= 6)
265 pixel_mask
= retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW
);
267 pixel_mask
= retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW
);
269 brw_MOV(p
, pixel_mask
, brw_flag_reg(0, 1));
272 if (devinfo
->gen
>= 6) {
273 brw_push_insn_state(p
);
274 brw_set_default_exec_size(p
, BRW_EXECUTE_16
);
275 brw_set_default_compression_control(p
, BRW_COMPRESSION_COMPRESSED
);
277 retype(payload
, BRW_REGISTER_TYPE_UD
),
278 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
279 brw_pop_insn_state(p
);
281 if (inst
->target
> 0 && key
->replicate_alpha
) {
282 /* Set "Source0 Alpha Present to RenderTarget" bit in message
286 vec1(retype(payload
, BRW_REGISTER_TYPE_UD
)),
287 vec1(retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
)),
288 brw_imm_ud(0x1 << 11));
291 if (inst
->target
> 0) {
292 /* Set the render target index for choosing BLEND_STATE. */
293 brw_MOV(p
, retype(vec1(suboffset(payload
, 2)),
294 BRW_REGISTER_TYPE_UD
),
295 brw_imm_ud(inst
->target
));
298 /* Set computes stencil to render target */
299 if (prog_data
->computed_stencil
) {
301 vec1(retype(payload
, BRW_REGISTER_TYPE_UD
)),
302 vec1(retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
)),
303 brw_imm_ud(0x1 << 14));
306 implied_header
= brw_null_reg();
308 implied_header
= retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW
);
311 brw_pop_insn_state(p
);
313 implied_header
= brw_null_reg();
316 if (!runtime_check_aads_emit
) {
317 fire_fb_write(inst
, payload
, implied_header
, inst
->mlen
);
319 /* This can only happen in gen < 6 */
320 assert(devinfo
->gen
< 6);
322 struct brw_reg v1_null_ud
= vec1(retype(brw_null_reg(), BRW_REGISTER_TYPE_UD
));
324 /* Check runtime bit to detect if we have to send AA data or not */
325 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
328 retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_UD
),
330 brw_inst_set_cond_modifier(p
->devinfo
, brw_last_inst
, BRW_CONDITIONAL_NZ
);
332 int jmp
= brw_JMPI(p
, brw_imm_ud(0), BRW_PREDICATE_NORMAL
) - p
->store
;
333 brw_inst_set_exec_size(p
->devinfo
, brw_last_inst
, BRW_EXECUTE_1
);
335 /* Don't send AA data */
336 fire_fb_write(inst
, offset(payload
, 1), implied_header
, inst
->mlen
-1);
338 brw_land_fwd_jump(p
, jmp
);
339 fire_fb_write(inst
, payload
, implied_header
, inst
->mlen
);
344 fs_generator::generate_mov_indirect(fs_inst
*inst
,
347 struct brw_reg indirect_byte_offset
)
349 assert(indirect_byte_offset
.type
== BRW_REGISTER_TYPE_UD
);
350 assert(indirect_byte_offset
.file
== BRW_GENERAL_REGISTER_FILE
);
352 unsigned imm_byte_offset
= reg
.nr
* REG_SIZE
+ reg
.subnr
;
354 if (indirect_byte_offset
.file
== BRW_IMMEDIATE_VALUE
) {
355 imm_byte_offset
+= indirect_byte_offset
.ud
;
357 reg
.nr
= imm_byte_offset
/ REG_SIZE
;
358 reg
.subnr
= imm_byte_offset
% REG_SIZE
;
359 brw_MOV(p
, dst
, reg
);
361 /* Prior to Broadwell, there are only 8 address registers. */
362 assert(inst
->exec_size
== 8 || devinfo
->gen
>= 8);
364 /* We use VxH indirect addressing, clobbering a0.0 through a0.7. */
365 struct brw_reg addr
= vec8(brw_address_reg(0));
367 /* The destination stride of an instruction (in bytes) must be greater
368 * than or equal to the size of the rest of the instruction. Since the
369 * address register is of type UW, we can't use a D-type instruction.
370 * In order to get around this, re retype to UW and use a stride.
372 indirect_byte_offset
=
373 retype(spread(indirect_byte_offset
, 2), BRW_REGISTER_TYPE_UW
);
375 struct brw_reg ind_src
;
376 if (devinfo
->gen
< 8) {
377 /* From the Haswell PRM section "Register Region Restrictions":
379 * "The lower bits of the AddressImmediate must not overflow to
380 * change the register address. The lower 5 bits of Address
381 * Immediate when added to lower 5 bits of address register gives
382 * the sub-register offset. The upper bits of Address Immediate
383 * when added to upper bits of address register gives the register
384 * address. Any overflow from sub-register offset is dropped."
386 * This restriction is only listed in the Haswell PRM but emperical
387 * testing indicates that it applies on all older generations and is
388 * lifted on Broadwell.
390 * Since the indirect may cause us to cross a register boundary, this
391 * makes the base offset almost useless. We could try and do
392 * something clever where we use a actual base offset if
393 * base_offset % 32 == 0 but that would mean we were generating
394 * different code depending on the base offset. Instead, for the
395 * sake of consistency, we'll just do the add ourselves.
397 brw_ADD(p
, addr
, indirect_byte_offset
, brw_imm_uw(imm_byte_offset
));
398 ind_src
= brw_VxH_indirect(0, 0);
400 brw_MOV(p
, addr
, indirect_byte_offset
);
401 ind_src
= brw_VxH_indirect(0, imm_byte_offset
);
404 brw_inst
*mov
= brw_MOV(p
, dst
, retype(ind_src
, dst
.type
));
406 if (devinfo
->gen
== 6 && dst
.file
== BRW_MESSAGE_REGISTER_FILE
&&
407 !inst
->get_next()->is_tail_sentinel() &&
408 ((fs_inst
*)inst
->get_next())->mlen
> 0) {
409 /* From the Sandybridge PRM:
411 * "[Errata: DevSNB(SNB)] If MRF register is updated by any
412 * instruction that “indexed/indirect” source AND is followed by a
413 * send, the instruction requires a “Switch”. This is to avoid
414 * race condition where send may dispatch before MRF is updated."
416 brw_inst_set_thread_control(devinfo
, mov
, BRW_THREAD_SWITCH
);
422 fs_generator::generate_urb_read(fs_inst
*inst
,
424 struct brw_reg header
)
426 assert(header
.file
== BRW_GENERAL_REGISTER_FILE
);
427 assert(header
.type
== BRW_REGISTER_TYPE_UD
);
429 brw_inst
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
430 brw_set_dest(p
, send
, dst
);
431 brw_set_src0(p
, send
, header
);
432 brw_set_src1(p
, send
, brw_imm_ud(0u));
434 brw_inst_set_sfid(p
->devinfo
, send
, BRW_SFID_URB
);
435 brw_inst_set_urb_opcode(p
->devinfo
, send
, GEN8_URB_OPCODE_SIMD8_READ
);
437 if (inst
->opcode
== SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
)
438 brw_inst_set_urb_per_slot_offset(p
->devinfo
, send
, true);
440 brw_inst_set_mlen(p
->devinfo
, send
, inst
->mlen
);
441 brw_inst_set_rlen(p
->devinfo
, send
, inst
->regs_written
);
442 brw_inst_set_header_present(p
->devinfo
, send
, true);
443 brw_inst_set_urb_global_offset(p
->devinfo
, send
, inst
->offset
);
447 fs_generator::generate_urb_write(fs_inst
*inst
, struct brw_reg payload
)
451 insn
= brw_next_insn(p
, BRW_OPCODE_SEND
);
453 brw_set_dest(p
, insn
, brw_null_reg());
454 brw_set_src0(p
, insn
, payload
);
455 brw_set_src1(p
, insn
, brw_imm_d(0));
457 brw_inst_set_sfid(p
->devinfo
, insn
, BRW_SFID_URB
);
458 brw_inst_set_urb_opcode(p
->devinfo
, insn
, GEN8_URB_OPCODE_SIMD8_WRITE
);
460 if (inst
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
||
461 inst
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
)
462 brw_inst_set_urb_per_slot_offset(p
->devinfo
, insn
, true);
464 if (inst
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
||
465 inst
->opcode
== SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
)
466 brw_inst_set_urb_channel_mask_present(p
->devinfo
, insn
, true);
468 brw_inst_set_mlen(p
->devinfo
, insn
, inst
->mlen
);
469 brw_inst_set_rlen(p
->devinfo
, insn
, 0);
470 brw_inst_set_eot(p
->devinfo
, insn
, inst
->eot
);
471 brw_inst_set_header_present(p
->devinfo
, insn
, true);
472 brw_inst_set_urb_global_offset(p
->devinfo
, insn
, inst
->offset
);
476 fs_generator::generate_cs_terminate(fs_inst
*inst
, struct brw_reg payload
)
478 struct brw_inst
*insn
;
480 insn
= brw_next_insn(p
, BRW_OPCODE_SEND
);
482 brw_set_dest(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_UW
));
483 brw_set_src0(p
, insn
, payload
);
484 brw_set_src1(p
, insn
, brw_imm_d(0));
486 /* Terminate a compute shader by sending a message to the thread spawner.
488 brw_inst_set_sfid(devinfo
, insn
, BRW_SFID_THREAD_SPAWNER
);
489 brw_inst_set_mlen(devinfo
, insn
, 1);
490 brw_inst_set_rlen(devinfo
, insn
, 0);
491 brw_inst_set_eot(devinfo
, insn
, inst
->eot
);
492 brw_inst_set_header_present(devinfo
, insn
, false);
494 brw_inst_set_ts_opcode(devinfo
, insn
, 0); /* Dereference resource */
495 brw_inst_set_ts_request_type(devinfo
, insn
, 0); /* Root thread */
497 /* Note that even though the thread has a URB resource associated with it,
498 * we set the "do not dereference URB" bit, because the URB resource is
499 * managed by the fixed-function unit, so it will free it automatically.
501 brw_inst_set_ts_resource_select(devinfo
, insn
, 1); /* Do not dereference URB */
503 brw_inst_set_mask_control(devinfo
, insn
, BRW_MASK_DISABLE
);
507 fs_generator::generate_stencil_ref_packing(fs_inst
*inst
,
511 assert(dispatch_width
== 8);
512 assert(devinfo
->gen
>= 9);
514 /* Stencil value updates are provided in 8 slots of 1 byte per slot.
515 * Presumably, in order to save memory bandwidth, the stencil reference
516 * values written from the FS need to be packed into 2 dwords (this makes
517 * sense because the stencil values are limited to 1 byte each and a SIMD8
518 * send, so stencil slots 0-3 in dw0, and 4-7 in dw1.)
520 * The spec is confusing here because in the payload definition of MDP_RTW_S8
521 * (Message Data Payload for Render Target Writes with Stencil 8b) the
522 * stencil value seems to be dw4.0-dw4.7. However, if you look at the type of
523 * dw4 it is type MDPR_STENCIL (Message Data Payload Register) which is the
524 * packed values specified above and diagrammed below:
527 * --------------------------------
531 * --------------------------------
532 * DW1 | STC | STC | STC | STC |
533 * | slot7 | slot6 | slot5 | slot4|
534 * --------------------------------
535 * DW0 | STC | STC | STC | STC |
536 * | slot3 | slot2 | slot1 | slot0|
537 * --------------------------------
540 src
.vstride
= BRW_VERTICAL_STRIDE_4
;
541 src
.width
= BRW_WIDTH_1
;
542 src
.hstride
= BRW_HORIZONTAL_STRIDE_0
;
543 assert(src
.type
== BRW_REGISTER_TYPE_UB
);
544 brw_MOV(p
, retype(dst
, BRW_REGISTER_TYPE_UB
), src
);
548 fs_generator::generate_barrier(fs_inst
*inst
, struct brw_reg src
)
555 fs_generator::generate_blorp_fb_write(fs_inst
*inst
)
558 16 /* dispatch_width */,
559 brw_message_reg(inst
->base_mrf
),
560 brw_reg_from_fs_reg(inst
, &inst
->src
[0], devinfo
->gen
),
561 BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE
,
567 inst
->header_size
!= 0);
571 fs_generator::generate_linterp(fs_inst
*inst
,
572 struct brw_reg dst
, struct brw_reg
*src
)
576 * -----------------------------------
577 * | src1+0 | src1+1 | src1+2 | src1+3 |
578 * |-----------------------------------|
579 * |(x0, x1)|(y0, y1)|(x2, x3)|(y2, y3)|
580 * -----------------------------------
582 * but for the LINE/MAC pair, the LINE reads Xs and the MAC reads Ys:
584 * -----------------------------------
585 * | src1+0 | src1+1 | src1+2 | src1+3 |
586 * |-----------------------------------|
587 * |(x0, x1)|(y0, y1)| | | in SIMD8
588 * |-----------------------------------|
589 * |(x0, x1)|(x2, x3)|(y0, y1)|(y2, y3)| in SIMD16
590 * -----------------------------------
592 * See also: emit_interpolation_setup_gen4().
594 struct brw_reg delta_x
= src
[0];
595 struct brw_reg delta_y
= offset(src
[0], dispatch_width
/ 8);
596 struct brw_reg interp
= src
[1];
598 if (devinfo
->has_pln
&&
599 (devinfo
->gen
>= 7 || (delta_x
.nr
& 1) == 0)) {
600 brw_PLN(p
, dst
, interp
, delta_x
);
602 brw_LINE(p
, brw_null_reg(), interp
, delta_x
);
603 brw_MAC(p
, dst
, suboffset(interp
, 1), delta_y
);
608 fs_generator::generate_math_gen6(fs_inst
*inst
,
613 int op
= brw_math_function(inst
->opcode
);
614 bool binop
= src1
.file
!= BRW_ARCHITECTURE_REGISTER_FILE
;
616 if (dispatch_width
== 8) {
617 gen6_math(p
, dst
, op
, src0
, src1
);
618 } else if (dispatch_width
== 16) {
619 brw_push_insn_state(p
);
620 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
621 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
622 gen6_math(p
, firsthalf(dst
), op
, firsthalf(src0
), firsthalf(src1
));
623 brw_set_default_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
624 gen6_math(p
, sechalf(dst
), op
, sechalf(src0
),
625 binop
? sechalf(src1
) : brw_null_reg());
626 brw_pop_insn_state(p
);
631 fs_generator::generate_math_gen4(fs_inst
*inst
,
635 int op
= brw_math_function(inst
->opcode
);
637 assert(inst
->mlen
>= 1);
639 if (dispatch_width
== 8) {
643 BRW_MATH_PRECISION_FULL
);
644 } else if (dispatch_width
== 16) {
645 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
646 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
647 gen4_math(p
, firsthalf(dst
),
649 inst
->base_mrf
, firsthalf(src
),
650 BRW_MATH_PRECISION_FULL
);
651 brw_set_default_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
652 gen4_math(p
, sechalf(dst
),
654 inst
->base_mrf
+ 1, sechalf(src
),
655 BRW_MATH_PRECISION_FULL
);
657 brw_set_default_compression_control(p
, BRW_COMPRESSION_COMPRESSED
);
662 fs_generator::generate_math_g45(fs_inst
*inst
,
666 if (inst
->opcode
== SHADER_OPCODE_POW
||
667 inst
->opcode
== SHADER_OPCODE_INT_QUOTIENT
||
668 inst
->opcode
== SHADER_OPCODE_INT_REMAINDER
) {
669 generate_math_gen4(inst
, dst
, src
);
673 int op
= brw_math_function(inst
->opcode
);
675 assert(inst
->mlen
>= 1);
680 BRW_MATH_PRECISION_FULL
);
684 fs_generator::generate_get_buffer_size(fs_inst
*inst
,
687 struct brw_reg surf_index
)
689 assert(devinfo
->gen
>= 7);
690 assert(surf_index
.file
== BRW_IMMEDIATE_VALUE
);
695 switch (inst
->exec_size
) {
697 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD8
;
700 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
703 unreachable("Invalid width for texture instruction");
706 if (simd_mode
== BRW_SAMPLER_SIMD_MODE_SIMD16
) {
712 retype(dst
, BRW_REGISTER_TYPE_UW
),
717 GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO
,
718 rlen
, /* response length */
720 inst
->header_size
> 0,
722 BRW_SAMPLER_RETURN_FORMAT_SINT32
);
724 brw_mark_surface_used(prog_data
, surf_index
.ud
);
728 fs_generator::generate_tex(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
,
729 struct brw_reg surface_index
,
730 struct brw_reg sampler_index
)
735 uint32_t return_format
;
736 bool is_combined_send
= inst
->eot
;
739 case BRW_REGISTER_TYPE_D
:
740 return_format
= BRW_SAMPLER_RETURN_FORMAT_SINT32
;
742 case BRW_REGISTER_TYPE_UD
:
743 return_format
= BRW_SAMPLER_RETURN_FORMAT_UINT32
;
746 return_format
= BRW_SAMPLER_RETURN_FORMAT_FLOAT32
;
750 /* Stomp the resinfo output type to UINT32. On gens 4-5, the output type
751 * is set as part of the message descriptor. On gen4, the PRM seems to
752 * allow UINT32 and FLOAT32 (i965 PRM, Vol. 4 Section 4.8.1.1), but on
753 * later gens UINT32 is required. Once you hit Sandy Bridge, the bit is
754 * gone from the message descriptor entirely and you just get UINT32 all
755 * the time regasrdless. Since we can really only do non-UINT32 on gen4,
756 * just stomp it to UINT32 all the time.
758 if (inst
->opcode
== SHADER_OPCODE_TXS
)
759 return_format
= BRW_SAMPLER_RETURN_FORMAT_UINT32
;
761 switch (inst
->exec_size
) {
763 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD8
;
766 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
769 unreachable("Invalid width for texture instruction");
772 if (devinfo
->gen
>= 5) {
773 switch (inst
->opcode
) {
774 case SHADER_OPCODE_TEX
:
775 if (inst
->shadow_compare
) {
776 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_COMPARE
;
778 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE
;
782 if (inst
->shadow_compare
) {
783 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS_COMPARE
;
785 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS
;
788 case SHADER_OPCODE_TXL
:
789 if (inst
->shadow_compare
) {
790 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE
;
792 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LOD
;
795 case SHADER_OPCODE_TXS
:
796 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO
;
798 case SHADER_OPCODE_TXD
:
799 if (inst
->shadow_compare
) {
800 /* Gen7.5+. Otherwise, lowered by brw_lower_texture_gradients(). */
801 assert(devinfo
->gen
>= 8 || devinfo
->is_haswell
);
802 msg_type
= HSW_SAMPLER_MESSAGE_SAMPLE_DERIV_COMPARE
;
804 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS
;
807 case SHADER_OPCODE_TXF
:
808 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
810 case SHADER_OPCODE_TXF_CMS_W
:
811 assert(devinfo
->gen
>= 9);
812 msg_type
= GEN9_SAMPLER_MESSAGE_SAMPLE_LD2DMS_W
;
814 case SHADER_OPCODE_TXF_CMS
:
815 if (devinfo
->gen
>= 7)
816 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DMS
;
818 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
820 case SHADER_OPCODE_TXF_UMS
:
821 assert(devinfo
->gen
>= 7);
822 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DSS
;
824 case SHADER_OPCODE_TXF_MCS
:
825 assert(devinfo
->gen
>= 7);
826 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_LD_MCS
;
828 case SHADER_OPCODE_LOD
:
829 msg_type
= GEN5_SAMPLER_MESSAGE_LOD
;
831 case SHADER_OPCODE_TG4
:
832 if (inst
->shadow_compare
) {
833 assert(devinfo
->gen
>= 7);
834 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_C
;
836 assert(devinfo
->gen
>= 6);
837 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4
;
840 case SHADER_OPCODE_TG4_OFFSET
:
841 assert(devinfo
->gen
>= 7);
842 if (inst
->shadow_compare
) {
843 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO_C
;
845 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO
;
848 case SHADER_OPCODE_SAMPLEINFO
:
849 msg_type
= GEN6_SAMPLER_MESSAGE_SAMPLE_SAMPLEINFO
;
852 unreachable("not reached");
855 switch (inst
->opcode
) {
856 case SHADER_OPCODE_TEX
:
857 /* Note that G45 and older determines shadow compare and dispatch width
858 * from message length for most messages.
860 if (inst
->exec_size
== 8) {
861 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE
;
862 if (inst
->shadow_compare
) {
863 assert(inst
->mlen
== 6);
865 assert(inst
->mlen
<= 4);
868 if (inst
->shadow_compare
) {
869 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_COMPARE
;
870 assert(inst
->mlen
== 9);
872 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE
;
873 assert(inst
->mlen
<= 7 && inst
->mlen
% 2 == 1);
878 if (inst
->shadow_compare
) {
879 assert(inst
->exec_size
== 8);
880 assert(inst
->mlen
== 6);
881 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_BIAS_COMPARE
;
883 assert(inst
->mlen
== 9);
884 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_BIAS
;
885 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
888 case SHADER_OPCODE_TXL
:
889 if (inst
->shadow_compare
) {
890 assert(inst
->exec_size
== 8);
891 assert(inst
->mlen
== 6);
892 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_LOD_COMPARE
;
894 assert(inst
->mlen
== 9);
895 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_LOD
;
896 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
899 case SHADER_OPCODE_TXD
:
900 /* There is no sample_d_c message; comparisons are done manually */
901 assert(inst
->exec_size
== 8);
902 assert(inst
->mlen
== 7 || inst
->mlen
== 10);
903 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_GRADIENTS
;
905 case SHADER_OPCODE_TXF
:
906 assert(inst
->mlen
<= 9 && inst
->mlen
% 2 == 1);
907 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_LD
;
908 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
910 case SHADER_OPCODE_TXS
:
911 assert(inst
->mlen
== 3);
912 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_RESINFO
;
913 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
916 unreachable("not reached");
919 assert(msg_type
!= -1);
921 if (simd_mode
== BRW_SAMPLER_SIMD_MODE_SIMD16
) {
926 if (is_combined_send
) {
927 assert(devinfo
->gen
>= 9 || devinfo
->is_cherryview
);
931 assert(devinfo
->gen
< 7 || inst
->header_size
== 0 ||
932 src
.file
== BRW_GENERAL_REGISTER_FILE
);
934 assert(sampler_index
.type
== BRW_REGISTER_TYPE_UD
);
936 /* Load the message header if present. If there's a texture offset,
937 * we need to set it up explicitly and load the offset bitfield.
938 * Otherwise, we can use an implied move from g0 to the first message reg.
940 if (inst
->header_size
!= 0) {
941 if (devinfo
->gen
< 6 && !inst
->offset
) {
942 /* Set up an implied move from g0 to the MRF. */
943 src
= retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW
);
945 struct brw_reg header_reg
;
947 if (devinfo
->gen
>= 7) {
950 assert(inst
->base_mrf
!= -1);
951 header_reg
= brw_message_reg(inst
->base_mrf
);
954 brw_push_insn_state(p
);
955 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
956 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
957 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
958 /* Explicitly set up the message header by copying g0 to the MRF. */
959 brw_MOV(p
, header_reg
, brw_vec8_grf(0, 0));
962 /* Set the offset bits in DWord 2. */
963 brw_MOV(p
, get_element_ud(header_reg
, 2),
964 brw_imm_ud(inst
->offset
));
965 } else if (stage
!= MESA_SHADER_VERTEX
&&
966 stage
!= MESA_SHADER_FRAGMENT
) {
967 /* The vertex and fragment stages have g0.2 set to 0, so
968 * header0.2 is 0 when g0 is copied. Other stages may not, so we
969 * must set it to 0 to avoid setting undesirable bits in the
972 brw_MOV(p
, get_element_ud(header_reg
, 2), brw_imm_ud(0));
975 brw_adjust_sampler_state_pointer(p
, header_reg
, sampler_index
);
976 brw_pop_insn_state(p
);
980 uint32_t base_binding_table_index
= (inst
->opcode
== SHADER_OPCODE_TG4
||
981 inst
->opcode
== SHADER_OPCODE_TG4_OFFSET
)
982 ? prog_data
->binding_table
.gather_texture_start
983 : prog_data
->binding_table
.texture_start
;
985 if (surface_index
.file
== BRW_IMMEDIATE_VALUE
&&
986 sampler_index
.file
== BRW_IMMEDIATE_VALUE
) {
987 uint32_t surface
= surface_index
.ud
;
988 uint32_t sampler
= sampler_index
.ud
;
991 retype(dst
, BRW_REGISTER_TYPE_UW
),
994 surface
+ base_binding_table_index
,
999 inst
->header_size
!= 0,
1003 brw_mark_surface_used(prog_data
, surface
+ base_binding_table_index
);
1005 /* Non-const sampler index */
1007 struct brw_reg addr
= vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD
));
1008 struct brw_reg surface_reg
= vec1(retype(surface_index
, BRW_REGISTER_TYPE_UD
));
1009 struct brw_reg sampler_reg
= vec1(retype(sampler_index
, BRW_REGISTER_TYPE_UD
));
1011 brw_push_insn_state(p
);
1012 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1013 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
1015 if (memcmp(&surface_reg
, &sampler_reg
, sizeof(surface_reg
)) == 0) {
1016 brw_MUL(p
, addr
, sampler_reg
, brw_imm_uw(0x101));
1018 brw_SHL(p
, addr
, sampler_reg
, brw_imm_ud(8));
1019 brw_OR(p
, addr
, addr
, surface_reg
);
1021 if (base_binding_table_index
)
1022 brw_ADD(p
, addr
, addr
, brw_imm_ud(base_binding_table_index
));
1023 brw_AND(p
, addr
, addr
, brw_imm_ud(0xfff));
1025 brw_pop_insn_state(p
);
1027 /* dst = send(offset, a0.0 | <descriptor>) */
1028 brw_inst
*insn
= brw_send_indirect_message(
1029 p
, BRW_SFID_SAMPLER
, dst
, src
, addr
);
1030 brw_set_sampler_message(p
, insn
,
1035 inst
->mlen
/* mlen */,
1036 inst
->header_size
!= 0 /* header */,
1040 /* visitor knows more than we do about the surface limit required,
1041 * so has already done marking.
1045 if (is_combined_send
) {
1046 brw_inst_set_eot(p
->devinfo
, brw_last_inst
, true);
1047 brw_inst_set_opcode(p
->devinfo
, brw_last_inst
, BRW_OPCODE_SENDC
);
1052 /* For OPCODE_DDX and OPCODE_DDY, per channel of output we've got input
1055 * arg0: ss0.tl ss0.tr ss0.bl ss0.br ss1.tl ss1.tr ss1.bl ss1.br
1057 * Ideally, we want to produce:
1060 * dst: (ss0.tr - ss0.tl) (ss0.tl - ss0.bl)
1061 * (ss0.tr - ss0.tl) (ss0.tr - ss0.br)
1062 * (ss0.br - ss0.bl) (ss0.tl - ss0.bl)
1063 * (ss0.br - ss0.bl) (ss0.tr - ss0.br)
1064 * (ss1.tr - ss1.tl) (ss1.tl - ss1.bl)
1065 * (ss1.tr - ss1.tl) (ss1.tr - ss1.br)
1066 * (ss1.br - ss1.bl) (ss1.tl - ss1.bl)
1067 * (ss1.br - ss1.bl) (ss1.tr - ss1.br)
1069 * and add another set of two more subspans if in 16-pixel dispatch mode.
1071 * For DDX, it ends up being easy: width = 2, horiz=0 gets us the same result
1072 * for each pair, and vertstride = 2 jumps us 2 elements after processing a
1073 * pair. But the ideal approximation may impose a huge performance cost on
1074 * sample_d. On at least Haswell, sample_d instruction does some
1075 * optimizations if the same LOD is used for all pixels in the subspan.
1077 * For DDY, we need to use ALIGN16 mode since it's capable of doing the
1078 * appropriate swizzling.
1081 fs_generator::generate_ddx(enum opcode opcode
,
1082 struct brw_reg dst
, struct brw_reg src
)
1084 unsigned vstride
, width
;
1086 if (opcode
== FS_OPCODE_DDX_FINE
) {
1087 /* produce accurate derivatives */
1088 vstride
= BRW_VERTICAL_STRIDE_2
;
1089 width
= BRW_WIDTH_2
;
1091 /* replicate the derivative at the top-left pixel to other pixels */
1092 vstride
= BRW_VERTICAL_STRIDE_4
;
1093 width
= BRW_WIDTH_4
;
1096 struct brw_reg src0
= brw_reg(src
.file
, src
.nr
, 1,
1097 src
.negate
, src
.abs
,
1098 BRW_REGISTER_TYPE_F
,
1101 BRW_HORIZONTAL_STRIDE_0
,
1102 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
1103 struct brw_reg src1
= brw_reg(src
.file
, src
.nr
, 0,
1104 src
.negate
, src
.abs
,
1105 BRW_REGISTER_TYPE_F
,
1108 BRW_HORIZONTAL_STRIDE_0
,
1109 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
1110 brw_ADD(p
, dst
, src0
, negate(src1
));
1113 /* The negate_value boolean is used to negate the derivative computation for
1114 * FBOs, since they place the origin at the upper left instead of the lower
1118 fs_generator::generate_ddy(enum opcode opcode
,
1119 struct brw_reg dst
, struct brw_reg src
,
1122 if (opcode
== FS_OPCODE_DDY_FINE
) {
1123 /* From the Ivy Bridge PRM, volume 4 part 3, section 3.3.9 (Register
1124 * Region Restrictions):
1126 * In Align16 access mode, SIMD16 is not allowed for DW operations
1127 * and SIMD8 is not allowed for DF operations.
1129 * In this context, "DW operations" means "operations acting on 32-bit
1130 * values", so it includes operations on floats.
1132 * Gen4 has a similar restriction. From the i965 PRM, section 11.5.3
1133 * (Instruction Compression -> Rules and Restrictions):
1135 * A compressed instruction must be in Align1 access mode. Align16
1136 * mode instructions cannot be compressed.
1138 * Similar text exists in the g45 PRM.
1140 * On these platforms, if we're building a SIMD16 shader, we need to
1141 * manually unroll to a pair of SIMD8 instructions.
1143 bool unroll_to_simd8
=
1144 (dispatch_width
== 16 &&
1145 (devinfo
->gen
== 4 || (devinfo
->gen
== 7 && !devinfo
->is_haswell
)));
1147 /* produce accurate derivatives */
1148 struct brw_reg src0
= brw_reg(src
.file
, src
.nr
, 0,
1149 src
.negate
, src
.abs
,
1150 BRW_REGISTER_TYPE_F
,
1151 BRW_VERTICAL_STRIDE_4
,
1153 BRW_HORIZONTAL_STRIDE_1
,
1154 BRW_SWIZZLE_XYXY
, WRITEMASK_XYZW
);
1155 struct brw_reg src1
= brw_reg(src
.file
, src
.nr
, 0,
1156 src
.negate
, src
.abs
,
1157 BRW_REGISTER_TYPE_F
,
1158 BRW_VERTICAL_STRIDE_4
,
1160 BRW_HORIZONTAL_STRIDE_1
,
1161 BRW_SWIZZLE_ZWZW
, WRITEMASK_XYZW
);
1162 brw_push_insn_state(p
);
1163 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1164 if (unroll_to_simd8
) {
1165 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
1166 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
1168 brw_ADD(p
, firsthalf(dst
), firsthalf(src1
), negate(firsthalf(src0
)));
1169 brw_set_default_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
1170 brw_ADD(p
, sechalf(dst
), sechalf(src1
), negate(sechalf(src0
)));
1172 brw_ADD(p
, firsthalf(dst
), firsthalf(src0
), negate(firsthalf(src1
)));
1173 brw_set_default_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
1174 brw_ADD(p
, sechalf(dst
), sechalf(src0
), negate(sechalf(src1
)));
1178 brw_ADD(p
, dst
, src1
, negate(src0
));
1180 brw_ADD(p
, dst
, src0
, negate(src1
));
1182 brw_pop_insn_state(p
);
1184 /* replicate the derivative at the top-left pixel to other pixels */
1185 struct brw_reg src0
= brw_reg(src
.file
, src
.nr
, 0,
1186 src
.negate
, src
.abs
,
1187 BRW_REGISTER_TYPE_F
,
1188 BRW_VERTICAL_STRIDE_4
,
1190 BRW_HORIZONTAL_STRIDE_0
,
1191 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
1192 struct brw_reg src1
= brw_reg(src
.file
, src
.nr
, 2,
1193 src
.negate
, src
.abs
,
1194 BRW_REGISTER_TYPE_F
,
1195 BRW_VERTICAL_STRIDE_4
,
1197 BRW_HORIZONTAL_STRIDE_0
,
1198 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
1200 brw_ADD(p
, dst
, src1
, negate(src0
));
1202 brw_ADD(p
, dst
, src0
, negate(src1
));
1207 fs_generator::generate_discard_jump(fs_inst
*inst
)
1209 assert(devinfo
->gen
>= 6);
1211 /* This HALT will be patched up at FB write time to point UIP at the end of
1212 * the program, and at brw_uip_jip() JIP will be set to the end of the
1213 * current block (or the program).
1215 this->discard_halt_patches
.push_tail(new(mem_ctx
) ip_record(p
->nr_insn
));
1217 brw_push_insn_state(p
);
1218 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1220 brw_pop_insn_state(p
);
1224 fs_generator::generate_scratch_write(fs_inst
*inst
, struct brw_reg src
)
1226 assert(inst
->mlen
!= 0);
1229 brw_uvec_mrf(inst
->exec_size
, (inst
->base_mrf
+ 1), 0),
1230 retype(src
, BRW_REGISTER_TYPE_UD
));
1231 brw_oword_block_write_scratch(p
, brw_message_reg(inst
->base_mrf
),
1232 inst
->exec_size
/ 8, inst
->offset
);
1236 fs_generator::generate_scratch_read(fs_inst
*inst
, struct brw_reg dst
)
1238 assert(inst
->mlen
!= 0);
1240 brw_oword_block_read_scratch(p
, dst
, brw_message_reg(inst
->base_mrf
),
1241 inst
->exec_size
/ 8, inst
->offset
);
1245 fs_generator::generate_scratch_read_gen7(fs_inst
*inst
, struct brw_reg dst
)
1247 gen7_block_read_scratch(p
, dst
, inst
->exec_size
/ 8, inst
->offset
);
1251 fs_generator::generate_uniform_pull_constant_load(fs_inst
*inst
,
1253 struct brw_reg index
,
1254 struct brw_reg offset
)
1256 assert(inst
->mlen
!= 0);
1258 assert(index
.file
== BRW_IMMEDIATE_VALUE
&&
1259 index
.type
== BRW_REGISTER_TYPE_UD
);
1260 uint32_t surf_index
= index
.ud
;
1262 assert(offset
.file
== BRW_IMMEDIATE_VALUE
&&
1263 offset
.type
== BRW_REGISTER_TYPE_UD
);
1264 uint32_t read_offset
= offset
.ud
;
1266 brw_oword_block_read(p
, dst
, brw_message_reg(inst
->base_mrf
),
1267 read_offset
, surf_index
);
1271 fs_generator::generate_uniform_pull_constant_load_gen7(fs_inst
*inst
,
1273 struct brw_reg index
,
1274 struct brw_reg offset
)
1276 assert(index
.type
== BRW_REGISTER_TYPE_UD
);
1278 assert(offset
.file
== BRW_GENERAL_REGISTER_FILE
);
1279 /* Reference just the dword we need, to avoid angering validate_reg(). */
1280 offset
= brw_vec1_grf(offset
.nr
, 0);
1282 /* We use the SIMD4x2 mode because we want to end up with 4 components in
1283 * the destination loaded consecutively from the same offset (which appears
1284 * in the first component, and the rest are ignored).
1286 dst
.width
= BRW_WIDTH_4
;
1288 struct brw_reg src
= offset
;
1289 bool header_present
= false;
1291 if (devinfo
->gen
>= 9) {
1292 /* Skylake requires a message header in order to use SIMD4x2 mode. */
1293 src
= retype(brw_vec4_grf(offset
.nr
, 0), BRW_REGISTER_TYPE_UD
);
1294 header_present
= true;
1296 brw_push_insn_state(p
);
1297 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1298 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
1299 brw_MOV(p
, vec8(src
), retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
1300 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
1302 brw_MOV(p
, get_element_ud(src
, 2),
1303 brw_imm_ud(GEN9_SAMPLER_SIMD_MODE_EXTENSION_SIMD4X2
));
1304 brw_pop_insn_state(p
);
1307 if (index
.file
== BRW_IMMEDIATE_VALUE
) {
1309 uint32_t surf_index
= index
.ud
;
1311 brw_push_insn_state(p
);
1312 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
1313 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1314 brw_inst
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
1315 brw_inst_set_exec_size(devinfo
, send
, BRW_EXECUTE_4
);
1316 brw_pop_insn_state(p
);
1318 brw_set_dest(p
, send
, dst
);
1319 brw_set_src0(p
, send
, src
);
1320 brw_set_sampler_message(p
, send
,
1322 0, /* LD message ignores sampler unit */
1323 GEN5_SAMPLER_MESSAGE_SAMPLE_LD
,
1327 BRW_SAMPLER_SIMD_MODE_SIMD4X2
,
1331 struct brw_reg addr
= vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD
));
1333 brw_push_insn_state(p
);
1334 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1335 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
1337 /* a0.0 = surf_index & 0xff */
1338 brw_inst
*insn_and
= brw_next_insn(p
, BRW_OPCODE_AND
);
1339 brw_inst_set_exec_size(p
->devinfo
, insn_and
, BRW_EXECUTE_1
);
1340 brw_set_dest(p
, insn_and
, addr
);
1341 brw_set_src0(p
, insn_and
, vec1(retype(index
, BRW_REGISTER_TYPE_UD
)));
1342 brw_set_src1(p
, insn_and
, brw_imm_ud(0x0ff));
1344 /* dst = send(payload, a0.0 | <descriptor>) */
1345 brw_inst
*insn
= brw_send_indirect_message(
1346 p
, BRW_SFID_SAMPLER
, dst
, src
, addr
);
1347 brw_set_sampler_message(p
, insn
,
1349 0, /* LD message ignores sampler unit */
1350 GEN5_SAMPLER_MESSAGE_SAMPLE_LD
,
1354 BRW_SAMPLER_SIMD_MODE_SIMD4X2
,
1357 brw_pop_insn_state(p
);
1362 fs_generator::generate_varying_pull_constant_load(fs_inst
*inst
,
1364 struct brw_reg index
,
1365 struct brw_reg offset
)
1367 assert(devinfo
->gen
< 7); /* Should use the gen7 variant. */
1368 assert(inst
->header_size
!= 0);
1371 assert(index
.file
== BRW_IMMEDIATE_VALUE
&&
1372 index
.type
== BRW_REGISTER_TYPE_UD
);
1373 uint32_t surf_index
= index
.ud
;
1375 uint32_t simd_mode
, rlen
, msg_type
;
1376 if (dispatch_width
== 16) {
1377 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1380 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD8
;
1384 if (devinfo
->gen
>= 5)
1385 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
1387 /* We always use the SIMD16 message so that we only have to load U, and
1390 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_LD
;
1391 assert(inst
->mlen
== 3);
1392 assert(inst
->regs_written
== 8);
1394 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1397 struct brw_reg offset_mrf
= retype(brw_message_reg(inst
->base_mrf
+ 1),
1398 BRW_REGISTER_TYPE_D
);
1399 brw_MOV(p
, offset_mrf
, offset
);
1401 struct brw_reg header
= brw_vec8_grf(0, 0);
1402 gen6_resolve_implied_move(p
, &header
, inst
->base_mrf
);
1404 brw_inst
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
1405 brw_inst_set_qtr_control(p
->devinfo
, send
, BRW_COMPRESSION_NONE
);
1406 brw_set_dest(p
, send
, retype(dst
, BRW_REGISTER_TYPE_UW
));
1407 brw_set_src0(p
, send
, header
);
1408 if (devinfo
->gen
< 6)
1409 brw_inst_set_base_mrf(p
->devinfo
, send
, inst
->base_mrf
);
1411 /* Our surface is set up as floats, regardless of what actual data is
1414 uint32_t return_format
= BRW_SAMPLER_RETURN_FORMAT_FLOAT32
;
1415 brw_set_sampler_message(p
, send
,
1417 0, /* sampler (unused) */
1421 inst
->header_size
!= 0,
1427 fs_generator::generate_varying_pull_constant_load_gen7(fs_inst
*inst
,
1429 struct brw_reg index
,
1430 struct brw_reg offset
)
1432 assert(devinfo
->gen
>= 7);
1433 /* Varying-offset pull constant loads are treated as a normal expression on
1434 * gen7, so the fact that it's a send message is hidden at the IR level.
1436 assert(inst
->header_size
== 0);
1437 assert(!inst
->mlen
);
1438 assert(index
.type
== BRW_REGISTER_TYPE_UD
);
1440 uint32_t simd_mode
, rlen
, mlen
;
1441 if (dispatch_width
== 16) {
1444 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
1448 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD8
;
1451 if (index
.file
== BRW_IMMEDIATE_VALUE
) {
1453 uint32_t surf_index
= index
.ud
;
1455 brw_inst
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
1456 brw_set_dest(p
, send
, retype(dst
, BRW_REGISTER_TYPE_UW
));
1457 brw_set_src0(p
, send
, offset
);
1458 brw_set_sampler_message(p
, send
,
1460 0, /* LD message ignores sampler unit */
1461 GEN5_SAMPLER_MESSAGE_SAMPLE_LD
,
1464 false, /* no header */
1470 struct brw_reg addr
= vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD
));
1472 brw_push_insn_state(p
);
1473 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1474 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
1476 /* a0.0 = surf_index & 0xff */
1477 brw_inst
*insn_and
= brw_next_insn(p
, BRW_OPCODE_AND
);
1478 brw_inst_set_exec_size(p
->devinfo
, insn_and
, BRW_EXECUTE_1
);
1479 brw_set_dest(p
, insn_and
, addr
);
1480 brw_set_src0(p
, insn_and
, vec1(retype(index
, BRW_REGISTER_TYPE_UD
)));
1481 brw_set_src1(p
, insn_and
, brw_imm_ud(0x0ff));
1483 brw_pop_insn_state(p
);
1485 /* dst = send(offset, a0.0 | <descriptor>) */
1486 brw_inst
*insn
= brw_send_indirect_message(
1487 p
, BRW_SFID_SAMPLER
, retype(dst
, BRW_REGISTER_TYPE_UW
),
1489 brw_set_sampler_message(p
, insn
,
1492 GEN5_SAMPLER_MESSAGE_SAMPLE_LD
,
1502 * Cause the current pixel/sample mask (from R1.7 bits 15:0) to be transferred
1503 * into the flags register (f0.0).
1505 * Used only on Gen6 and above.
1508 fs_generator::generate_mov_dispatch_to_flags(fs_inst
*inst
)
1510 struct brw_reg flags
= brw_flag_reg(0, inst
->flag_subreg
);
1511 struct brw_reg dispatch_mask
;
1513 if (devinfo
->gen
>= 6)
1514 dispatch_mask
= retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW
);
1516 dispatch_mask
= retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW
);
1518 brw_push_insn_state(p
);
1519 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1520 brw_MOV(p
, flags
, dispatch_mask
);
1521 brw_pop_insn_state(p
);
1525 fs_generator::generate_pixel_interpolator_query(fs_inst
*inst
,
1528 struct brw_reg msg_data
,
1531 assert(msg_data
.type
== BRW_REGISTER_TYPE_UD
);
1533 brw_pixel_interpolator_query(p
,
1534 retype(dst
, BRW_REGISTER_TYPE_UW
),
1536 inst
->pi_noperspective
,
1540 inst
->regs_written
);
1545 * Sets the first word of a vgrf for gen7+ simd4x2 uniform pull constant
1546 * sampler LD messages.
1548 * We don't want to bake it into the send message's code generation because
1549 * that means we don't get a chance to schedule the instructions.
1552 fs_generator::generate_set_simd4x2_offset(fs_inst
*inst
,
1554 struct brw_reg value
)
1556 assert(value
.file
== BRW_IMMEDIATE_VALUE
);
1558 brw_push_insn_state(p
);
1559 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
1560 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
1561 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1562 brw_MOV(p
, retype(brw_vec1_reg(dst
.file
, dst
.nr
, 0), value
.type
), value
);
1563 brw_pop_insn_state(p
);
1566 /* Sets vstride=1, width=4, hstride=0 of register src1 during
1567 * the ADD instruction.
1570 fs_generator::generate_set_sample_id(fs_inst
*inst
,
1572 struct brw_reg src0
,
1573 struct brw_reg src1
)
1575 assert(dst
.type
== BRW_REGISTER_TYPE_D
||
1576 dst
.type
== BRW_REGISTER_TYPE_UD
);
1577 assert(src0
.type
== BRW_REGISTER_TYPE_D
||
1578 src0
.type
== BRW_REGISTER_TYPE_UD
);
1580 struct brw_reg reg
= stride(src1
, 1, 4, 0);
1581 if (devinfo
->gen
>= 8 || dispatch_width
== 8) {
1582 brw_ADD(p
, dst
, src0
, reg
);
1583 } else if (dispatch_width
== 16) {
1584 brw_push_insn_state(p
);
1585 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
1586 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
1587 brw_ADD(p
, firsthalf(dst
), firsthalf(src0
), reg
);
1588 brw_set_default_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
1589 brw_ADD(p
, sechalf(dst
), sechalf(src0
), suboffset(reg
, 2));
1590 brw_pop_insn_state(p
);
1595 fs_generator::generate_pack_half_2x16_split(fs_inst
*inst
,
1600 assert(devinfo
->gen
>= 7);
1601 assert(dst
.type
== BRW_REGISTER_TYPE_UD
);
1602 assert(x
.type
== BRW_REGISTER_TYPE_F
);
1603 assert(y
.type
== BRW_REGISTER_TYPE_F
);
1605 /* From the Ivybridge PRM, Vol4, Part3, Section 6.27 f32to16:
1607 * Because this instruction does not have a 16-bit floating-point type,
1608 * the destination data type must be Word (W).
1610 * The destination must be DWord-aligned and specify a horizontal stride
1611 * (HorzStride) of 2. The 16-bit result is stored in the lower word of
1612 * each destination channel and the upper word is not modified.
1614 struct brw_reg dst_w
= spread(retype(dst
, BRW_REGISTER_TYPE_W
), 2);
1616 /* Give each 32-bit channel of dst the form below, where "." means
1620 brw_F32TO16(p
, dst_w
, y
);
1625 brw_SHL(p
, dst
, dst
, brw_imm_ud(16u));
1627 /* And, finally the form of packHalf2x16's output:
1630 brw_F32TO16(p
, dst_w
, x
);
1634 fs_generator::generate_unpack_half_2x16_split(fs_inst
*inst
,
1638 assert(devinfo
->gen
>= 7);
1639 assert(dst
.type
== BRW_REGISTER_TYPE_F
);
1640 assert(src
.type
== BRW_REGISTER_TYPE_UD
);
1642 /* From the Ivybridge PRM, Vol4, Part3, Section 6.26 f16to32:
1644 * Because this instruction does not have a 16-bit floating-point type,
1645 * the source data type must be Word (W). The destination type must be
1648 struct brw_reg src_w
= spread(retype(src
, BRW_REGISTER_TYPE_W
), 2);
1650 /* Each channel of src has the form of unpackHalf2x16's input: 0xhhhhllll.
1651 * For the Y case, we wish to access only the upper word; therefore
1652 * a 16-bit subregister offset is needed.
1654 assert(inst
->opcode
== FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X
||
1655 inst
->opcode
== FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y
);
1656 if (inst
->opcode
== FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y
)
1659 brw_F16TO32(p
, dst
, src_w
);
1663 fs_generator::generate_shader_time_add(fs_inst
*inst
,
1664 struct brw_reg payload
,
1665 struct brw_reg offset
,
1666 struct brw_reg value
)
1668 assert(devinfo
->gen
>= 7);
1669 brw_push_insn_state(p
);
1670 brw_set_default_mask_control(p
, true);
1672 assert(payload
.file
== BRW_GENERAL_REGISTER_FILE
);
1673 struct brw_reg payload_offset
= retype(brw_vec1_grf(payload
.nr
, 0),
1675 struct brw_reg payload_value
= retype(brw_vec1_grf(payload
.nr
+ 1, 0),
1678 assert(offset
.file
== BRW_IMMEDIATE_VALUE
);
1679 if (value
.file
== BRW_GENERAL_REGISTER_FILE
) {
1680 value
.width
= BRW_WIDTH_1
;
1681 value
.hstride
= BRW_HORIZONTAL_STRIDE_0
;
1682 value
.vstride
= BRW_VERTICAL_STRIDE_0
;
1684 assert(value
.file
== BRW_IMMEDIATE_VALUE
);
1687 /* Trying to deal with setup of the params from the IR is crazy in the FS8
1688 * case, and we don't really care about squeezing every bit of performance
1689 * out of this path, so we just emit the MOVs from here.
1691 brw_MOV(p
, payload_offset
, offset
);
1692 brw_MOV(p
, payload_value
, value
);
1693 brw_shader_time_add(p
, payload
,
1694 prog_data
->binding_table
.shader_time_start
);
1695 brw_pop_insn_state(p
);
1697 brw_mark_surface_used(prog_data
,
1698 prog_data
->binding_table
.shader_time_start
);
1702 fs_generator::enable_debug(const char *shader_name
)
1705 this->shader_name
= shader_name
;
1709 fs_generator::generate_code(const cfg_t
*cfg
, int dispatch_width
)
1711 /* align to 64 byte boundary. */
1712 while (p
->next_insn_offset
% 64)
1715 this->dispatch_width
= dispatch_width
;
1716 if (dispatch_width
== 16)
1717 brw_set_default_compression_control(p
, BRW_COMPRESSION_COMPRESSED
);
1719 int start_offset
= p
->next_insn_offset
;
1720 int spill_count
= 0, fill_count
= 0;
1723 struct annotation_info annotation
;
1724 memset(&annotation
, 0, sizeof(annotation
));
1726 foreach_block_and_inst (block
, fs_inst
, inst
, cfg
) {
1727 struct brw_reg src
[3], dst
;
1728 unsigned int last_insn_offset
= p
->next_insn_offset
;
1729 bool multiple_instructions_emitted
= false;
1731 if (unlikely(debug_flag
))
1732 annotate(p
->devinfo
, &annotation
, cfg
, inst
, p
->next_insn_offset
);
1734 for (unsigned int i
= 0; i
< inst
->sources
; i
++) {
1735 src
[i
] = brw_reg_from_fs_reg(inst
, &inst
->src
[i
], devinfo
->gen
);
1737 /* The accumulator result appears to get used for the
1738 * conditional modifier generation. When negating a UD
1739 * value, there is a 33rd bit generated for the sign in the
1740 * accumulator value, so now you can't check, for example,
1741 * equality with a 32-bit value. See piglit fs-op-neg-uvec4.
1743 assert(!inst
->conditional_mod
||
1744 inst
->src
[i
].type
!= BRW_REGISTER_TYPE_UD
||
1745 !inst
->src
[i
].negate
);
1747 dst
= brw_reg_from_fs_reg(inst
, &inst
->dst
, devinfo
->gen
);
1749 brw_set_default_predicate_control(p
, inst
->predicate
);
1750 brw_set_default_predicate_inverse(p
, inst
->predicate_inverse
);
1751 brw_set_default_flag_reg(p
, 0, inst
->flag_subreg
);
1752 brw_set_default_saturate(p
, inst
->saturate
);
1753 brw_set_default_mask_control(p
, inst
->force_writemask_all
);
1754 brw_set_default_acc_write_control(p
, inst
->writes_accumulator
);
1755 brw_set_default_exec_size(p
, cvt(inst
->exec_size
) - 1);
1757 assert(inst
->base_mrf
+ inst
->mlen
<= BRW_MAX_MRF(devinfo
->gen
));
1758 assert(inst
->mlen
<= BRW_MAX_MSG_LENGTH
);
1760 switch (inst
->exec_size
) {
1764 assert(inst
->force_writemask_all
);
1765 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
1768 if (inst
->force_sechalf
) {
1769 brw_set_default_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
1771 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
1776 /* If the instruction writes to more than one register, it needs to
1777 * be a "compressed" instruction on Gen <= 5.
1779 if (inst
->dst
.component_size(inst
->exec_size
) > REG_SIZE
)
1780 brw_set_default_compression_control(p
, BRW_COMPRESSION_COMPRESSED
);
1782 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
1785 unreachable("Invalid instruction width");
1788 switch (inst
->opcode
) {
1789 case BRW_OPCODE_MOV
:
1790 brw_MOV(p
, dst
, src
[0]);
1792 case BRW_OPCODE_ADD
:
1793 brw_ADD(p
, dst
, src
[0], src
[1]);
1795 case BRW_OPCODE_MUL
:
1796 brw_MUL(p
, dst
, src
[0], src
[1]);
1798 case BRW_OPCODE_AVG
:
1799 brw_AVG(p
, dst
, src
[0], src
[1]);
1801 case BRW_OPCODE_MACH
:
1802 brw_MACH(p
, dst
, src
[0], src
[1]);
1805 case BRW_OPCODE_LINE
:
1806 brw_LINE(p
, dst
, src
[0], src
[1]);
1809 case BRW_OPCODE_MAD
:
1810 assert(devinfo
->gen
>= 6);
1811 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1812 if (dispatch_width
== 16 && !devinfo
->supports_simd16_3src
) {
1813 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
1814 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
1815 brw_inst
*f
= brw_MAD(p
, firsthalf(dst
), firsthalf(src
[0]), firsthalf(src
[1]), firsthalf(src
[2]));
1816 brw_set_default_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
1817 brw_inst
*s
= brw_MAD(p
, sechalf(dst
), sechalf(src
[0]), sechalf(src
[1]), sechalf(src
[2]));
1818 brw_set_default_compression_control(p
, BRW_COMPRESSION_COMPRESSED
);
1820 if (inst
->conditional_mod
) {
1821 brw_inst_set_cond_modifier(p
->devinfo
, f
, inst
->conditional_mod
);
1822 brw_inst_set_cond_modifier(p
->devinfo
, s
, inst
->conditional_mod
);
1823 multiple_instructions_emitted
= true;
1826 brw_MAD(p
, dst
, src
[0], src
[1], src
[2]);
1828 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
1831 case BRW_OPCODE_LRP
:
1832 assert(devinfo
->gen
>= 6);
1833 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1834 if (dispatch_width
== 16 && !devinfo
->supports_simd16_3src
) {
1835 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
1836 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
1837 brw_inst
*f
= brw_LRP(p
, firsthalf(dst
), firsthalf(src
[0]), firsthalf(src
[1]), firsthalf(src
[2]));
1838 brw_set_default_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
1839 brw_inst
*s
= brw_LRP(p
, sechalf(dst
), sechalf(src
[0]), sechalf(src
[1]), sechalf(src
[2]));
1840 brw_set_default_compression_control(p
, BRW_COMPRESSION_COMPRESSED
);
1842 if (inst
->conditional_mod
) {
1843 brw_inst_set_cond_modifier(p
->devinfo
, f
, inst
->conditional_mod
);
1844 brw_inst_set_cond_modifier(p
->devinfo
, s
, inst
->conditional_mod
);
1845 multiple_instructions_emitted
= true;
1848 brw_LRP(p
, dst
, src
[0], src
[1], src
[2]);
1850 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
1853 case BRW_OPCODE_FRC
:
1854 brw_FRC(p
, dst
, src
[0]);
1856 case BRW_OPCODE_RNDD
:
1857 brw_RNDD(p
, dst
, src
[0]);
1859 case BRW_OPCODE_RNDE
:
1860 brw_RNDE(p
, dst
, src
[0]);
1862 case BRW_OPCODE_RNDZ
:
1863 brw_RNDZ(p
, dst
, src
[0]);
1866 case BRW_OPCODE_AND
:
1867 brw_AND(p
, dst
, src
[0], src
[1]);
1870 brw_OR(p
, dst
, src
[0], src
[1]);
1872 case BRW_OPCODE_XOR
:
1873 brw_XOR(p
, dst
, src
[0], src
[1]);
1875 case BRW_OPCODE_NOT
:
1876 brw_NOT(p
, dst
, src
[0]);
1878 case BRW_OPCODE_ASR
:
1879 brw_ASR(p
, dst
, src
[0], src
[1]);
1881 case BRW_OPCODE_SHR
:
1882 brw_SHR(p
, dst
, src
[0], src
[1]);
1884 case BRW_OPCODE_SHL
:
1885 brw_SHL(p
, dst
, src
[0], src
[1]);
1887 case BRW_OPCODE_F32TO16
:
1888 assert(devinfo
->gen
>= 7);
1889 brw_F32TO16(p
, dst
, src
[0]);
1891 case BRW_OPCODE_F16TO32
:
1892 assert(devinfo
->gen
>= 7);
1893 brw_F16TO32(p
, dst
, src
[0]);
1895 case BRW_OPCODE_CMP
:
1896 /* The Ivybridge/BayTrail WaCMPInstFlagDepClearedEarly workaround says
1897 * that when the destination is a GRF that the dependency-clear bit on
1898 * the flag register is cleared early.
1900 * Suggested workarounds are to disable coissuing CMP instructions
1901 * or to split CMP(16) instructions into two CMP(8) instructions.
1903 * We choose to split into CMP(8) instructions since disabling
1904 * coissuing would affect CMP instructions not otherwise affected by
1907 if (dispatch_width
== 16 && devinfo
->gen
== 7 && !devinfo
->is_haswell
) {
1908 if (dst
.file
== BRW_GENERAL_REGISTER_FILE
) {
1909 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
1910 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
1911 brw_CMP(p
, firsthalf(dst
), inst
->conditional_mod
,
1912 firsthalf(src
[0]), firsthalf(src
[1]));
1913 brw_set_default_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
1914 brw_CMP(p
, sechalf(dst
), inst
->conditional_mod
,
1915 sechalf(src
[0]), sechalf(src
[1]));
1916 brw_set_default_compression_control(p
, BRW_COMPRESSION_COMPRESSED
);
1918 multiple_instructions_emitted
= true;
1919 } else if (dst
.file
== BRW_ARCHITECTURE_REGISTER_FILE
) {
1920 /* For unknown reasons, the aforementioned workaround is not
1921 * sufficient. Overriding the type when the destination is the
1922 * null register is necessary but not sufficient by itself.
1924 assert(dst
.nr
== BRW_ARF_NULL
);
1925 dst
.type
= BRW_REGISTER_TYPE_D
;
1926 brw_CMP(p
, dst
, inst
->conditional_mod
, src
[0], src
[1]);
1928 unreachable("not reached");
1931 brw_CMP(p
, dst
, inst
->conditional_mod
, src
[0], src
[1]);
1934 case BRW_OPCODE_SEL
:
1935 brw_SEL(p
, dst
, src
[0], src
[1]);
1937 case BRW_OPCODE_BFREV
:
1938 assert(devinfo
->gen
>= 7);
1939 /* BFREV only supports UD type for src and dst. */
1940 brw_BFREV(p
, retype(dst
, BRW_REGISTER_TYPE_UD
),
1941 retype(src
[0], BRW_REGISTER_TYPE_UD
));
1943 case BRW_OPCODE_FBH
:
1944 assert(devinfo
->gen
>= 7);
1945 /* FBH only supports UD type for dst. */
1946 brw_FBH(p
, retype(dst
, BRW_REGISTER_TYPE_UD
), src
[0]);
1948 case BRW_OPCODE_FBL
:
1949 assert(devinfo
->gen
>= 7);
1950 /* FBL only supports UD type for dst. */
1951 brw_FBL(p
, retype(dst
, BRW_REGISTER_TYPE_UD
), src
[0]);
1953 case BRW_OPCODE_CBIT
:
1954 assert(devinfo
->gen
>= 7);
1955 /* CBIT only supports UD type for dst. */
1956 brw_CBIT(p
, retype(dst
, BRW_REGISTER_TYPE_UD
), src
[0]);
1958 case BRW_OPCODE_ADDC
:
1959 assert(devinfo
->gen
>= 7);
1960 brw_ADDC(p
, dst
, src
[0], src
[1]);
1962 case BRW_OPCODE_SUBB
:
1963 assert(devinfo
->gen
>= 7);
1964 brw_SUBB(p
, dst
, src
[0], src
[1]);
1966 case BRW_OPCODE_MAC
:
1967 brw_MAC(p
, dst
, src
[0], src
[1]);
1970 case BRW_OPCODE_BFE
:
1971 assert(devinfo
->gen
>= 7);
1972 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1973 if (dispatch_width
== 16 && !devinfo
->supports_simd16_3src
) {
1974 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
1975 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
1976 brw_BFE(p
, firsthalf(dst
), firsthalf(src
[0]), firsthalf(src
[1]), firsthalf(src
[2]));
1977 brw_set_default_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
1978 brw_BFE(p
, sechalf(dst
), sechalf(src
[0]), sechalf(src
[1]), sechalf(src
[2]));
1979 brw_set_default_compression_control(p
, BRW_COMPRESSION_COMPRESSED
);
1981 brw_BFE(p
, dst
, src
[0], src
[1], src
[2]);
1983 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
1986 case BRW_OPCODE_BFI1
:
1987 assert(devinfo
->gen
>= 7);
1988 /* The Haswell WaForceSIMD8ForBFIInstruction workaround says that we
1991 * "Force BFI instructions to be executed always in SIMD8."
1993 if (dispatch_width
== 16 && devinfo
->is_haswell
) {
1994 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
1995 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
1996 brw_BFI1(p
, firsthalf(dst
), firsthalf(src
[0]), firsthalf(src
[1]));
1997 brw_set_default_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
1998 brw_BFI1(p
, sechalf(dst
), sechalf(src
[0]), sechalf(src
[1]));
1999 brw_set_default_compression_control(p
, BRW_COMPRESSION_COMPRESSED
);
2001 brw_BFI1(p
, dst
, src
[0], src
[1]);
2004 case BRW_OPCODE_BFI2
:
2005 assert(devinfo
->gen
>= 7);
2006 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
2007 /* The Haswell WaForceSIMD8ForBFIInstruction workaround says that we
2010 * "Force BFI instructions to be executed always in SIMD8."
2012 * Otherwise we would be able to emit compressed instructions like we
2013 * do for the other three-source instructions.
2015 if (dispatch_width
== 16 &&
2016 (devinfo
->is_haswell
|| !devinfo
->supports_simd16_3src
)) {
2017 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
2018 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
2019 brw_BFI2(p
, firsthalf(dst
), firsthalf(src
[0]), firsthalf(src
[1]), firsthalf(src
[2]));
2020 brw_set_default_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
2021 brw_BFI2(p
, sechalf(dst
), sechalf(src
[0]), sechalf(src
[1]), sechalf(src
[2]));
2022 brw_set_default_compression_control(p
, BRW_COMPRESSION_COMPRESSED
);
2024 brw_BFI2(p
, dst
, src
[0], src
[1], src
[2]);
2026 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
2030 if (inst
->src
[0].file
!= BAD_FILE
) {
2031 /* The instruction has an embedded compare (only allowed on gen6) */
2032 assert(devinfo
->gen
== 6);
2033 gen6_IF(p
, inst
->conditional_mod
, src
[0], src
[1]);
2035 brw_IF(p
, dispatch_width
== 16 ? BRW_EXECUTE_16
: BRW_EXECUTE_8
);
2039 case BRW_OPCODE_ELSE
:
2042 case BRW_OPCODE_ENDIF
:
2047 brw_DO(p
, dispatch_width
== 16 ? BRW_EXECUTE_16
: BRW_EXECUTE_8
);
2050 case BRW_OPCODE_BREAK
:
2052 brw_set_default_predicate_control(p
, BRW_PREDICATE_NONE
);
2054 case BRW_OPCODE_CONTINUE
:
2056 brw_set_default_predicate_control(p
, BRW_PREDICATE_NONE
);
2059 case BRW_OPCODE_WHILE
:
2064 case SHADER_OPCODE_RCP
:
2065 case SHADER_OPCODE_RSQ
:
2066 case SHADER_OPCODE_SQRT
:
2067 case SHADER_OPCODE_EXP2
:
2068 case SHADER_OPCODE_LOG2
:
2069 case SHADER_OPCODE_SIN
:
2070 case SHADER_OPCODE_COS
:
2071 assert(devinfo
->gen
< 6 || inst
->mlen
== 0);
2072 assert(inst
->conditional_mod
== BRW_CONDITIONAL_NONE
);
2073 if (devinfo
->gen
>= 7) {
2074 gen6_math(p
, dst
, brw_math_function(inst
->opcode
), src
[0],
2076 } else if (devinfo
->gen
== 6) {
2077 generate_math_gen6(inst
, dst
, src
[0], brw_null_reg());
2078 } else if (devinfo
->gen
== 5 || devinfo
->is_g4x
) {
2079 generate_math_g45(inst
, dst
, src
[0]);
2081 generate_math_gen4(inst
, dst
, src
[0]);
2084 case SHADER_OPCODE_INT_QUOTIENT
:
2085 case SHADER_OPCODE_INT_REMAINDER
:
2086 case SHADER_OPCODE_POW
:
2087 assert(devinfo
->gen
< 6 || inst
->mlen
== 0);
2088 assert(inst
->conditional_mod
== BRW_CONDITIONAL_NONE
);
2089 if (devinfo
->gen
>= 7 && inst
->opcode
== SHADER_OPCODE_POW
) {
2090 gen6_math(p
, dst
, brw_math_function(inst
->opcode
), src
[0], src
[1]);
2091 } else if (devinfo
->gen
>= 6) {
2092 generate_math_gen6(inst
, dst
, src
[0], src
[1]);
2094 generate_math_gen4(inst
, dst
, src
[0]);
2097 case FS_OPCODE_CINTERP
:
2098 brw_MOV(p
, dst
, src
[0]);
2100 case FS_OPCODE_LINTERP
:
2101 generate_linterp(inst
, dst
, src
);
2103 case FS_OPCODE_PIXEL_X
:
2104 assert(src
[0].type
== BRW_REGISTER_TYPE_UW
);
2105 src
[0].subnr
= 0 * type_sz(src
[0].type
);
2106 brw_MOV(p
, dst
, stride(src
[0], 8, 4, 1));
2108 case FS_OPCODE_PIXEL_Y
:
2109 assert(src
[0].type
== BRW_REGISTER_TYPE_UW
);
2110 src
[0].subnr
= 4 * type_sz(src
[0].type
);
2111 brw_MOV(p
, dst
, stride(src
[0], 8, 4, 1));
2113 case FS_OPCODE_GET_BUFFER_SIZE
:
2114 generate_get_buffer_size(inst
, dst
, src
[0], src
[1]);
2116 case SHADER_OPCODE_TEX
:
2118 case SHADER_OPCODE_TXD
:
2119 case SHADER_OPCODE_TXF
:
2120 case SHADER_OPCODE_TXF_CMS
:
2121 case SHADER_OPCODE_TXF_CMS_W
:
2122 case SHADER_OPCODE_TXF_UMS
:
2123 case SHADER_OPCODE_TXF_MCS
:
2124 case SHADER_OPCODE_TXL
:
2125 case SHADER_OPCODE_TXS
:
2126 case SHADER_OPCODE_LOD
:
2127 case SHADER_OPCODE_TG4
:
2128 case SHADER_OPCODE_TG4_OFFSET
:
2129 case SHADER_OPCODE_SAMPLEINFO
:
2130 generate_tex(inst
, dst
, src
[0], src
[1], src
[2]);
2132 case FS_OPCODE_DDX_COARSE
:
2133 case FS_OPCODE_DDX_FINE
:
2134 generate_ddx(inst
->opcode
, dst
, src
[0]);
2136 case FS_OPCODE_DDY_COARSE
:
2137 case FS_OPCODE_DDY_FINE
:
2138 assert(src
[1].file
== BRW_IMMEDIATE_VALUE
);
2139 generate_ddy(inst
->opcode
, dst
, src
[0], src
[1].ud
);
2142 case SHADER_OPCODE_GEN4_SCRATCH_WRITE
:
2143 generate_scratch_write(inst
, src
[0]);
2147 case SHADER_OPCODE_GEN4_SCRATCH_READ
:
2148 generate_scratch_read(inst
, dst
);
2152 case SHADER_OPCODE_GEN7_SCRATCH_READ
:
2153 generate_scratch_read_gen7(inst
, dst
);
2157 case SHADER_OPCODE_MOV_INDIRECT
:
2158 generate_mov_indirect(inst
, dst
, src
[0], src
[1]);
2161 case SHADER_OPCODE_URB_READ_SIMD8
:
2162 case SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
:
2163 generate_urb_read(inst
, dst
, src
[0]);
2166 case SHADER_OPCODE_URB_WRITE_SIMD8
:
2167 case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
2168 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
:
2169 case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
2170 generate_urb_write(inst
, src
[0]);
2173 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
2174 generate_uniform_pull_constant_load(inst
, dst
, src
[0], src
[1]);
2177 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
:
2178 generate_uniform_pull_constant_load_gen7(inst
, dst
, src
[0], src
[1]);
2181 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD
:
2182 generate_varying_pull_constant_load(inst
, dst
, src
[0], src
[1]);
2185 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7
:
2186 generate_varying_pull_constant_load_gen7(inst
, dst
, src
[0], src
[1]);
2189 case FS_OPCODE_REP_FB_WRITE
:
2190 case FS_OPCODE_FB_WRITE
:
2191 generate_fb_write(inst
, src
[0]);
2194 case FS_OPCODE_BLORP_FB_WRITE
:
2195 generate_blorp_fb_write(inst
);
2198 case FS_OPCODE_MOV_DISPATCH_TO_FLAGS
:
2199 generate_mov_dispatch_to_flags(inst
);
2202 case FS_OPCODE_DISCARD_JUMP
:
2203 generate_discard_jump(inst
);
2206 case SHADER_OPCODE_SHADER_TIME_ADD
:
2207 generate_shader_time_add(inst
, src
[0], src
[1], src
[2]);
2210 case SHADER_OPCODE_UNTYPED_ATOMIC
:
2211 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2212 brw_untyped_atomic(p
, dst
, src
[0], src
[1], src
[2].ud
,
2213 inst
->mlen
, !inst
->dst
.is_null());
2216 case SHADER_OPCODE_UNTYPED_SURFACE_READ
:
2217 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2218 brw_untyped_surface_read(p
, dst
, src
[0], src
[1],
2219 inst
->mlen
, src
[2].ud
);
2222 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE
:
2223 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2224 brw_untyped_surface_write(p
, src
[0], src
[1],
2225 inst
->mlen
, src
[2].ud
);
2228 case SHADER_OPCODE_TYPED_ATOMIC
:
2229 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2230 brw_typed_atomic(p
, dst
, src
[0], src
[1],
2231 src
[2].ud
, inst
->mlen
, !inst
->dst
.is_null());
2234 case SHADER_OPCODE_TYPED_SURFACE_READ
:
2235 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2236 brw_typed_surface_read(p
, dst
, src
[0], src
[1],
2237 inst
->mlen
, src
[2].ud
);
2240 case SHADER_OPCODE_TYPED_SURFACE_WRITE
:
2241 assert(src
[2].file
== BRW_IMMEDIATE_VALUE
);
2242 brw_typed_surface_write(p
, src
[0], src
[1], inst
->mlen
, src
[2].ud
);
2245 case SHADER_OPCODE_MEMORY_FENCE
:
2246 brw_memory_fence(p
, dst
);
2249 case FS_OPCODE_SET_SIMD4X2_OFFSET
:
2250 generate_set_simd4x2_offset(inst
, dst
, src
[0]);
2253 case SHADER_OPCODE_FIND_LIVE_CHANNEL
:
2254 brw_find_live_channel(p
, dst
);
2257 case SHADER_OPCODE_BROADCAST
:
2258 brw_broadcast(p
, dst
, src
[0], src
[1]);
2261 case SHADER_OPCODE_EXTRACT_BYTE
: {
2262 assert(src
[0].type
== BRW_REGISTER_TYPE_D
||
2263 src
[0].type
== BRW_REGISTER_TYPE_UD
);
2265 enum brw_reg_type type
=
2266 src
[0].type
== BRW_REGISTER_TYPE_D
? BRW_REGISTER_TYPE_B
2267 : BRW_REGISTER_TYPE_UB
;
2268 brw_MOV(p
, dst
, spread(suboffset(retype(src
[0], type
), src
[1].ud
), 4));
2272 case SHADER_OPCODE_EXTRACT_WORD
: {
2273 assert(src
[0].type
== BRW_REGISTER_TYPE_D
||
2274 src
[0].type
== BRW_REGISTER_TYPE_UD
);
2276 enum brw_reg_type type
=
2277 src
[0].type
== BRW_REGISTER_TYPE_D
? BRW_REGISTER_TYPE_W
2278 : BRW_REGISTER_TYPE_UW
;
2279 brw_MOV(p
, dst
, spread(suboffset(retype(src
[0], type
), src
[1].ud
), 2));
2283 case FS_OPCODE_SET_SAMPLE_ID
:
2284 generate_set_sample_id(inst
, dst
, src
[0], src
[1]);
2287 case FS_OPCODE_PACK_HALF_2x16_SPLIT
:
2288 generate_pack_half_2x16_split(inst
, dst
, src
[0], src
[1]);
2291 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X
:
2292 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y
:
2293 generate_unpack_half_2x16_split(inst
, dst
, src
[0]);
2296 case FS_OPCODE_PLACEHOLDER_HALT
:
2297 /* This is the place where the final HALT needs to be inserted if
2298 * we've emitted any discards. If not, this will emit no code.
2300 if (!patch_discard_jumps_to_fb_writes()) {
2301 if (unlikely(debug_flag
)) {
2302 annotation
.ann_count
--;
2307 case FS_OPCODE_INTERPOLATE_AT_CENTROID
:
2308 generate_pixel_interpolator_query(inst
, dst
, src
[0], src
[1],
2309 GEN7_PIXEL_INTERPOLATOR_LOC_CENTROID
);
2312 case FS_OPCODE_INTERPOLATE_AT_SAMPLE
:
2313 generate_pixel_interpolator_query(inst
, dst
, src
[0], src
[1],
2314 GEN7_PIXEL_INTERPOLATOR_LOC_SAMPLE
);
2317 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
:
2318 generate_pixel_interpolator_query(inst
, dst
, src
[0], src
[1],
2319 GEN7_PIXEL_INTERPOLATOR_LOC_SHARED_OFFSET
);
2322 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
:
2323 generate_pixel_interpolator_query(inst
, dst
, src
[0], src
[1],
2324 GEN7_PIXEL_INTERPOLATOR_LOC_PER_SLOT_OFFSET
);
2327 case CS_OPCODE_CS_TERMINATE
:
2328 generate_cs_terminate(inst
, src
[0]);
2331 case SHADER_OPCODE_BARRIER
:
2332 generate_barrier(inst
, src
[0]);
2335 case FS_OPCODE_PACK_STENCIL_REF
:
2336 generate_stencil_ref_packing(inst
, dst
, src
[0]);
2340 unreachable("Unsupported opcode");
2342 case SHADER_OPCODE_LOAD_PAYLOAD
:
2343 unreachable("Should be lowered by lower_load_payload()");
2346 if (multiple_instructions_emitted
)
2349 if (inst
->no_dd_clear
|| inst
->no_dd_check
|| inst
->conditional_mod
) {
2350 assert(p
->next_insn_offset
== last_insn_offset
+ 16 ||
2351 !"conditional_mod, no_dd_check, or no_dd_clear set for IR "
2352 "emitting more than 1 instruction");
2354 brw_inst
*last
= &p
->store
[last_insn_offset
/ 16];
2356 if (inst
->conditional_mod
)
2357 brw_inst_set_cond_modifier(p
->devinfo
, last
, inst
->conditional_mod
);
2358 brw_inst_set_no_dd_clear(p
->devinfo
, last
, inst
->no_dd_clear
);
2359 brw_inst_set_no_dd_check(p
->devinfo
, last
, inst
->no_dd_check
);
2364 annotation_finalize(&annotation
, p
->next_insn_offset
);
2367 bool validated
= brw_validate_instructions(p
, start_offset
, &annotation
);
2369 if (unlikely(debug_flag
))
2370 brw_validate_instructions(p
, start_offset
, &annotation
);
2373 int before_size
= p
->next_insn_offset
- start_offset
;
2374 brw_compact_instructions(p
, start_offset
, annotation
.ann_count
,
2376 int after_size
= p
->next_insn_offset
- start_offset
;
2378 if (unlikely(debug_flag
)) {
2379 fprintf(stderr
, "Native code for %s\n"
2380 "SIMD%d shader: %d instructions. %d loops. %u cycles. %d:%d spills:fills. Promoted %u constants. Compacted %d to %d"
2381 " bytes (%.0f%%)\n",
2382 shader_name
, dispatch_width
, before_size
/ 16, loop_count
, cfg
->cycle_count
,
2383 spill_count
, fill_count
, promoted_constants
, before_size
, after_size
,
2384 100.0f
* (before_size
- after_size
) / before_size
);
2386 dump_assembly(p
->store
, annotation
.ann_count
, annotation
.ann
,
2388 ralloc_free(annotation
.mem_ctx
);
2392 compiler
->shader_debug_log(log_data
,
2393 "%s SIMD%d shader: %d inst, %d loops, %u cycles, "
2394 "%d:%d spills:fills, Promoted %u constants, "
2395 "compacted %d to %d bytes.",
2396 _mesa_shader_stage_to_abbrev(stage
),
2397 dispatch_width
, before_size
/ 16,
2398 loop_count
, cfg
->cycle_count
, spill_count
,
2399 fill_count
, promoted_constants
, before_size
,
2402 return start_offset
;
2406 fs_generator::get_assembly(unsigned int *assembly_size
)
2408 return brw_get_program(p
, assembly_size
);