1 /* Copyright © 2011 Intel Corporation
3 * Permission is hereby granted, free of charge, to any person obtaining a
4 * copy of this software and associated documentation files (the "Software"),
5 * to deal in the Software without restriction, including without limitation
6 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
7 * and/or sell copies of the Software, and to permit persons to whom the
8 * Software is furnished to do so, subject to the following conditions:
10 * The above copyright notice and this permission notice (including the next
11 * paragraph) shall be included in all copies or substantial portions of the
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
19 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
29 #include "main/macros.h"
30 #include "program/prog_print.h"
31 #include "program/prog_parameter.h"
37 vec4_instruction::get_dst(void)
39 struct brw_reg brw_reg
;
43 brw_reg
= brw_vec8_grf(dst
.reg
+ dst
.reg_offset
, 0);
44 brw_reg
= retype(brw_reg
, dst
.type
);
45 brw_reg
.dw1
.bits
.writemask
= dst
.writemask
;
49 brw_reg
= brw_message_reg(dst
.reg
+ dst
.reg_offset
);
50 brw_reg
= retype(brw_reg
, dst
.type
);
51 brw_reg
.dw1
.bits
.writemask
= dst
.writemask
;
55 assert(dst
.type
== dst
.fixed_hw_reg
.type
);
56 brw_reg
= dst
.fixed_hw_reg
;
60 brw_reg
= brw_null_reg();
64 unreachable("not reached");
70 vec4_instruction::get_src(const struct brw_vue_prog_data
*prog_data
, int i
)
72 struct brw_reg brw_reg
;
74 switch (src
[i
].file
) {
76 brw_reg
= brw_vec8_grf(src
[i
].reg
+ src
[i
].reg_offset
, 0);
77 brw_reg
= retype(brw_reg
, src
[i
].type
);
78 brw_reg
.dw1
.bits
.swizzle
= src
[i
].swizzle
;
80 brw_reg
= brw_abs(brw_reg
);
82 brw_reg
= negate(brw_reg
);
86 switch (src
[i
].type
) {
87 case BRW_REGISTER_TYPE_F
:
88 brw_reg
= brw_imm_f(src
[i
].fixed_hw_reg
.dw1
.f
);
90 case BRW_REGISTER_TYPE_D
:
91 brw_reg
= brw_imm_d(src
[i
].fixed_hw_reg
.dw1
.d
);
93 case BRW_REGISTER_TYPE_UD
:
94 brw_reg
= brw_imm_ud(src
[i
].fixed_hw_reg
.dw1
.ud
);
96 case BRW_REGISTER_TYPE_VF
:
97 brw_reg
= brw_imm_vf(src
[i
].fixed_hw_reg
.dw1
.ud
);
100 unreachable("not reached");
105 brw_reg
= stride(brw_vec4_grf(prog_data
->base
.dispatch_grf_start_reg
+
106 (src
[i
].reg
+ src
[i
].reg_offset
) / 2,
107 ((src
[i
].reg
+ src
[i
].reg_offset
) % 2) * 4),
109 brw_reg
= retype(brw_reg
, src
[i
].type
);
110 brw_reg
.dw1
.bits
.swizzle
= src
[i
].swizzle
;
112 brw_reg
= brw_abs(brw_reg
);
114 brw_reg
= negate(brw_reg
);
116 /* This should have been moved to pull constants. */
117 assert(!src
[i
].reladdr
);
121 assert(src
[i
].type
== src
[i
].fixed_hw_reg
.type
);
122 brw_reg
= src
[i
].fixed_hw_reg
;
126 /* Probably unused. */
127 brw_reg
= brw_null_reg();
131 unreachable("not reached");
137 vec4_generator::vec4_generator(struct brw_context
*brw
,
138 struct gl_shader_program
*shader_prog
,
139 struct gl_program
*prog
,
140 struct brw_vue_prog_data
*prog_data
,
143 const char *stage_name
,
144 const char *stage_abbrev
)
145 : brw(brw
), shader_prog(shader_prog
), prog(prog
), prog_data(prog_data
),
146 mem_ctx(mem_ctx
), stage_name(stage_name
), stage_abbrev(stage_abbrev
),
147 debug_flag(debug_flag
)
149 p
= rzalloc(mem_ctx
, struct brw_compile
);
150 brw_init_compile(brw
, p
, mem_ctx
);
153 vec4_generator::~vec4_generator()
158 vec4_generator::generate_math1_gen4(vec4_instruction
*inst
,
164 brw_math_function(inst
->opcode
),
167 BRW_MATH_PRECISION_FULL
);
171 check_gen6_math_src_arg(struct brw_reg src
)
173 /* Source swizzles are ignored. */
176 assert(src
.dw1
.bits
.swizzle
== BRW_SWIZZLE_XYZW
);
180 vec4_generator::generate_math_gen6(vec4_instruction
*inst
,
185 /* Can't do writemask because math can't be align16. */
186 assert(dst
.dw1
.bits
.writemask
== WRITEMASK_XYZW
);
187 /* Source swizzles are ignored. */
188 check_gen6_math_src_arg(src0
);
189 if (src1
.file
== BRW_GENERAL_REGISTER_FILE
)
190 check_gen6_math_src_arg(src1
);
192 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
193 gen6_math(p
, dst
, brw_math_function(inst
->opcode
), src0
, src1
);
194 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
198 vec4_generator::generate_math2_gen4(vec4_instruction
*inst
,
203 /* From the Ironlake PRM, Volume 4, Part 1, Section 6.1.13
206 * "Operand0[7]. For the INT DIV functions, this operand is the
209 * "Operand1[7]. For the INT DIV functions, this operand is the
212 bool is_int_div
= inst
->opcode
!= SHADER_OPCODE_POW
;
213 struct brw_reg
&op0
= is_int_div
? src1
: src0
;
214 struct brw_reg
&op1
= is_int_div
? src0
: src1
;
216 brw_push_insn_state(p
);
217 brw_set_default_saturate(p
, false);
218 brw_set_default_predicate_control(p
, BRW_PREDICATE_NONE
);
219 brw_MOV(p
, retype(brw_message_reg(inst
->base_mrf
+ 1), op1
.type
), op1
);
220 brw_pop_insn_state(p
);
224 brw_math_function(inst
->opcode
),
227 BRW_MATH_PRECISION_FULL
);
231 vec4_generator::generate_tex(vec4_instruction
*inst
,
234 struct brw_reg sampler_index
)
239 switch (inst
->opcode
) {
240 case SHADER_OPCODE_TEX
:
241 case SHADER_OPCODE_TXL
:
242 if (inst
->shadow_compare
) {
243 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE
;
245 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LOD
;
248 case SHADER_OPCODE_TXD
:
249 if (inst
->shadow_compare
) {
250 /* Gen7.5+. Otherwise, lowered by brw_lower_texture_gradients(). */
251 assert(brw
->gen
>= 8 || brw
->is_haswell
);
252 msg_type
= HSW_SAMPLER_MESSAGE_SAMPLE_DERIV_COMPARE
;
254 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS
;
257 case SHADER_OPCODE_TXF
:
258 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
260 case SHADER_OPCODE_TXF_CMS
:
262 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DMS
;
264 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
266 case SHADER_OPCODE_TXF_MCS
:
267 assert(brw
->gen
>= 7);
268 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_LD_MCS
;
270 case SHADER_OPCODE_TXS
:
271 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO
;
273 case SHADER_OPCODE_TG4
:
274 if (inst
->shadow_compare
) {
275 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_C
;
277 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4
;
280 case SHADER_OPCODE_TG4_OFFSET
:
281 if (inst
->shadow_compare
) {
282 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO_C
;
284 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO
;
288 unreachable("should not get here: invalid vec4 texture opcode");
291 switch (inst
->opcode
) {
292 case SHADER_OPCODE_TEX
:
293 case SHADER_OPCODE_TXL
:
294 if (inst
->shadow_compare
) {
295 msg_type
= BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_LOD_COMPARE
;
296 assert(inst
->mlen
== 3);
298 msg_type
= BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_LOD
;
299 assert(inst
->mlen
== 2);
302 case SHADER_OPCODE_TXD
:
303 /* There is no sample_d_c message; comparisons are done manually. */
304 msg_type
= BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_GRADIENTS
;
305 assert(inst
->mlen
== 4);
307 case SHADER_OPCODE_TXF
:
308 msg_type
= BRW_SAMPLER_MESSAGE_SIMD4X2_LD
;
309 assert(inst
->mlen
== 2);
311 case SHADER_OPCODE_TXS
:
312 msg_type
= BRW_SAMPLER_MESSAGE_SIMD4X2_RESINFO
;
313 assert(inst
->mlen
== 2);
316 unreachable("should not get here: invalid vec4 texture opcode");
320 assert(msg_type
!= -1);
322 assert(sampler_index
.type
== BRW_REGISTER_TYPE_UD
);
324 /* Load the message header if present. If there's a texture offset, we need
325 * to set it up explicitly and load the offset bitfield. Otherwise, we can
326 * use an implied move from g0 to the first message register.
328 if (inst
->header_present
) {
329 if (brw
->gen
< 6 && !inst
->offset
) {
330 /* Set up an implied move from g0 to the MRF. */
331 src
= brw_vec8_grf(0, 0);
333 struct brw_reg header
=
334 retype(brw_message_reg(inst
->base_mrf
), BRW_REGISTER_TYPE_UD
);
337 /* Explicitly set up the message header by copying g0 to the MRF. */
338 brw_push_insn_state(p
);
339 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
340 brw_MOV(p
, header
, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
342 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
345 /* Set the texel offset bits in DWord 2. */
349 /* SKL+ overloads BRW_SAMPLER_SIMD_MODE_SIMD4X2 to also do SIMD8D,
350 * based on bit 22 in the header.
352 dw2
|= GEN9_SAMPLER_SIMD_MODE_EXTENSION_SIMD4X2
;
355 brw_MOV(p
, get_element_ud(header
, 2), brw_imm_ud(dw2
));
357 brw_adjust_sampler_state_pointer(p
, header
, sampler_index
);
358 brw_pop_insn_state(p
);
362 uint32_t return_format
;
365 case BRW_REGISTER_TYPE_D
:
366 return_format
= BRW_SAMPLER_RETURN_FORMAT_SINT32
;
368 case BRW_REGISTER_TYPE_UD
:
369 return_format
= BRW_SAMPLER_RETURN_FORMAT_UINT32
;
372 return_format
= BRW_SAMPLER_RETURN_FORMAT_FLOAT32
;
376 uint32_t base_binding_table_index
= (inst
->opcode
== SHADER_OPCODE_TG4
||
377 inst
->opcode
== SHADER_OPCODE_TG4_OFFSET
)
378 ? prog_data
->base
.binding_table
.gather_texture_start
379 : prog_data
->base
.binding_table
.texture_start
;
381 if (sampler_index
.file
== BRW_IMMEDIATE_VALUE
) {
382 uint32_t sampler
= sampler_index
.dw1
.ud
;
388 sampler
+ base_binding_table_index
,
391 1, /* response length */
393 inst
->header_present
,
394 BRW_SAMPLER_SIMD_MODE_SIMD4X2
,
397 brw_mark_surface_used(&prog_data
->base
, sampler
+ base_binding_table_index
);
399 /* Non-constant sampler index. */
400 /* Note: this clobbers `dst` as a temporary before emitting the send */
402 struct brw_reg addr
= vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD
));
403 struct brw_reg temp
= vec1(retype(dst
, BRW_REGISTER_TYPE_UD
));
405 struct brw_reg sampler_reg
= vec1(retype(sampler_index
, BRW_REGISTER_TYPE_UD
));
407 brw_push_insn_state(p
);
408 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
409 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
411 /* Some care required: `sampler` and `temp` may alias:
412 * addr = sampler & 0xff
413 * temp = (sampler << 8) & 0xf00
416 brw_ADD(p
, addr
, sampler_reg
, brw_imm_ud(base_binding_table_index
));
417 brw_SHL(p
, temp
, sampler_reg
, brw_imm_ud(8u));
418 brw_AND(p
, temp
, temp
, brw_imm_ud(0x0f00));
419 brw_AND(p
, addr
, addr
, brw_imm_ud(0x0ff));
420 brw_OR(p
, addr
, addr
, temp
);
422 /* dst = send(offset, a0.0 | <descriptor>) */
423 brw_inst
*insn
= brw_send_indirect_message(
424 p
, BRW_SFID_SAMPLER
, dst
, src
, addr
);
425 brw_set_sampler_message(p
, insn
,
430 inst
->mlen
/* mlen */,
431 inst
->header_present
/* header */,
432 BRW_SAMPLER_SIMD_MODE_SIMD4X2
,
435 brw_pop_insn_state(p
);
437 /* visitor knows more than we do about the surface limit required,
438 * so has already done marking.
444 vec4_generator::generate_vs_urb_write(vec4_instruction
*inst
)
447 brw_null_reg(), /* dest */
448 inst
->base_mrf
, /* starting mrf reg nr */
449 brw_vec8_grf(0, 0), /* src */
450 inst
->urb_write_flags
,
452 0, /* response len */
453 inst
->offset
, /* urb destination offset */
454 BRW_URB_SWIZZLE_INTERLEAVE
);
458 vec4_generator::generate_gs_urb_write(vec4_instruction
*inst
)
460 struct brw_reg src
= brw_message_reg(inst
->base_mrf
);
462 brw_null_reg(), /* dest */
463 inst
->base_mrf
, /* starting mrf reg nr */
465 inst
->urb_write_flags
,
467 0, /* response len */
468 inst
->offset
, /* urb destination offset */
469 BRW_URB_SWIZZLE_INTERLEAVE
);
473 vec4_generator::generate_gs_urb_write_allocate(vec4_instruction
*inst
)
475 struct brw_reg src
= brw_message_reg(inst
->base_mrf
);
477 /* We pass the temporary passed in src0 as the writeback register */
479 inst
->get_src(this->prog_data
, 0), /* dest */
480 inst
->base_mrf
, /* starting mrf reg nr */
482 BRW_URB_WRITE_ALLOCATE_COMPLETE
,
484 1, /* response len */
485 inst
->offset
, /* urb destination offset */
486 BRW_URB_SWIZZLE_INTERLEAVE
);
488 /* Now put allocated urb handle in dst.0 */
489 brw_push_insn_state(p
);
490 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
491 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
492 brw_MOV(p
, get_element_ud(inst
->get_dst(), 0),
493 get_element_ud(inst
->get_src(this->prog_data
, 0), 0));
494 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
495 brw_pop_insn_state(p
);
499 vec4_generator::generate_gs_thread_end(vec4_instruction
*inst
)
501 struct brw_reg src
= brw_message_reg(inst
->base_mrf
);
503 brw_null_reg(), /* dest */
504 inst
->base_mrf
, /* starting mrf reg nr */
506 BRW_URB_WRITE_EOT
| inst
->urb_write_flags
,
507 brw
->gen
>= 8 ? 2 : 1,/* message len */
508 0, /* response len */
509 0, /* urb destination offset */
510 BRW_URB_SWIZZLE_INTERLEAVE
);
514 vec4_generator::generate_gs_set_write_offset(struct brw_reg dst
,
518 /* From p22 of volume 4 part 2 of the Ivy Bridge PRM (2.4.3.1 Message
521 * Slot 0 Offset. This field, after adding to the Global Offset field
522 * in the message descriptor, specifies the offset (in 256-bit units)
523 * from the start of the URB entry, as referenced by URB Handle 0, at
524 * which the data will be accessed.
526 * Similar text describes DWORD M0.4, which is slot 1 offset.
528 * Therefore, we want to multiply DWORDs 0 and 4 of src0 (the x components
529 * of the register for geometry shader invocations 0 and 1) by the
530 * immediate value in src1, and store the result in DWORDs 3 and 4 of dst.
532 * We can do this with the following EU instruction:
534 * mul(2) dst.3<1>UD src0<8;2,4>UD src1<...>UW { Align1 WE_all }
536 brw_push_insn_state(p
);
537 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
538 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
539 assert(brw
->gen
>= 7 &&
540 src1
.file
== BRW_IMMEDIATE_VALUE
&&
541 src1
.type
== BRW_REGISTER_TYPE_UD
&&
542 src1
.dw1
.ud
<= USHRT_MAX
);
543 brw_MUL(p
, suboffset(stride(dst
, 2, 2, 1), 3), stride(src0
, 8, 2, 4),
544 retype(src1
, BRW_REGISTER_TYPE_UW
));
545 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
546 brw_pop_insn_state(p
);
550 vec4_generator::generate_gs_set_vertex_count(struct brw_reg dst
,
553 brw_push_insn_state(p
);
554 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
557 /* Move the vertex count into the second MRF for the EOT write. */
558 brw_MOV(p
, retype(brw_message_reg(dst
.nr
+ 1), BRW_REGISTER_TYPE_UD
),
561 /* If we think of the src and dst registers as composed of 8 DWORDs each,
562 * we want to pick up the contents of DWORDs 0 and 4 from src, truncate
563 * them to WORDs, and then pack them into DWORD 2 of dst.
565 * It's easier to get the EU to do this if we think of the src and dst
566 * registers as composed of 16 WORDS each; then, we want to pick up the
567 * contents of WORDs 0 and 8 from src, and pack them into WORDs 4 and 5
570 * We can do that by the following EU instruction:
572 * mov (2) dst.4<1>:uw src<8;1,0>:uw { Align1, Q1, NoMask }
574 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
576 suboffset(stride(retype(dst
, BRW_REGISTER_TYPE_UW
), 2, 2, 1), 4),
577 stride(retype(src
, BRW_REGISTER_TYPE_UW
), 8, 1, 0));
578 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
580 brw_pop_insn_state(p
);
584 vec4_generator::generate_gs_svb_write(vec4_instruction
*inst
,
589 int binding
= inst
->sol_binding
;
590 bool final_write
= inst
->sol_final_write
;
592 brw_push_insn_state(p
);
593 /* Copy Vertex data into M0.x */
594 brw_MOV(p
, stride(dst
, 4, 4, 1),
595 stride(retype(src0
, BRW_REGISTER_TYPE_UD
), 4, 4, 1));
599 final_write
? src1
: brw_null_reg(), /* dest == src1 */
601 dst
, /* src0 == previous dst */
602 SURF_INDEX_GEN6_SOL_BINDING(binding
), /* binding_table_index */
603 final_write
); /* send_commit_msg */
605 /* Finally, wait for the write commit to occur so that we can proceed to
606 * other things safely.
608 * From the Sandybridge PRM, Volume 4, Part 1, Section 3.3:
610 * The write commit does not modify the destination register, but
611 * merely clears the dependency associated with the destination
612 * register. Thus, a simple “mov” instruction using the register as a
613 * source is sufficient to wait for the write commit to occur.
616 brw_MOV(p
, src1
, src1
);
618 brw_pop_insn_state(p
);
622 vec4_generator::generate_gs_svb_set_destination_index(vec4_instruction
*inst
,
627 int vertex
= inst
->sol_vertex
;
628 brw_push_insn_state(p
);
629 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
630 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
631 brw_MOV(p
, get_element_ud(dst
, 5), get_element_ud(src
, vertex
));
632 brw_pop_insn_state(p
);
636 vec4_generator::generate_gs_set_dword_2(struct brw_reg dst
, struct brw_reg src
)
638 brw_push_insn_state(p
);
639 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
640 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
641 brw_MOV(p
, suboffset(vec1(dst
), 2), suboffset(vec1(src
), 0));
642 brw_pop_insn_state(p
);
646 vec4_generator::generate_gs_prepare_channel_masks(struct brw_reg dst
)
648 /* We want to left shift just DWORD 4 (the x component belonging to the
649 * second geometry shader invocation) by 4 bits. So generate the
652 * shl(1) dst.4<1>UD dst.4<0,1,0>UD 4UD { align1 WE_all }
654 dst
= suboffset(vec1(dst
), 4);
655 brw_push_insn_state(p
);
656 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
657 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
658 brw_SHL(p
, dst
, dst
, brw_imm_ud(4));
659 brw_pop_insn_state(p
);
663 vec4_generator::generate_gs_set_channel_masks(struct brw_reg dst
,
666 /* From p21 of volume 4 part 2 of the Ivy Bridge PRM (2.4.3.1 Message
669 * 15 Vertex 1 DATA [3] / Vertex 0 DATA[7] Channel Mask
671 * When Swizzle Control = URB_INTERLEAVED this bit controls Vertex 1
672 * DATA[3], when Swizzle Control = URB_NOSWIZZLE this bit controls
673 * Vertex 0 DATA[7]. This bit is ANDed with the corresponding
674 * channel enable to determine the final channel enable. For the
675 * URB_READ_OWORD & URB_READ_HWORD messages, when final channel
676 * enable is 1 it indicates that Vertex 1 DATA [3] will be included
677 * in the writeback message. For the URB_WRITE_OWORD &
678 * URB_WRITE_HWORD messages, when final channel enable is 1 it
679 * indicates that Vertex 1 DATA [3] will be written to the surface.
681 * 0: Vertex 1 DATA [3] / Vertex 0 DATA[7] channel not included
682 * 1: Vertex DATA [3] / Vertex 0 DATA[7] channel included
684 * 14 Vertex 1 DATA [2] Channel Mask
685 * 13 Vertex 1 DATA [1] Channel Mask
686 * 12 Vertex 1 DATA [0] Channel Mask
687 * 11 Vertex 0 DATA [3] Channel Mask
688 * 10 Vertex 0 DATA [2] Channel Mask
689 * 9 Vertex 0 DATA [1] Channel Mask
690 * 8 Vertex 0 DATA [0] Channel Mask
692 * (This is from a section of the PRM that is agnostic to the particular
693 * type of shader being executed, so "Vertex 0" and "Vertex 1" refer to
694 * geometry shader invocations 0 and 1, respectively). Since we have the
695 * enable flags for geometry shader invocation 0 in bits 3:0 of DWORD 0,
696 * and the enable flags for geometry shader invocation 1 in bits 7:0 of
697 * DWORD 4, we just need to OR them together and store the result in bits
700 * It's easier to get the EU to do this if we think of the src and dst
701 * registers as composed of 32 bytes each; then, we want to pick up the
702 * contents of bytes 0 and 16 from src, OR them together, and store them in
705 * We can do that by the following EU instruction:
707 * or(1) dst.21<1>UB src<0,1,0>UB src.16<0,1,0>UB { align1 WE_all }
709 * Note: this relies on the source register having zeros in (a) bits 7:4 of
710 * DWORD 0 and (b) bits 3:0 of DWORD 4. We can rely on (b) because the
711 * source register was prepared by GS_OPCODE_PREPARE_CHANNEL_MASKS (which
712 * shifts DWORD 4 left by 4 bits), and we can rely on (a) because prior to
713 * the execution of GS_OPCODE_PREPARE_CHANNEL_MASKS, DWORDs 0 and 4 need to
714 * contain valid channel mask values (which are in the range 0x0-0xf).
716 dst
= retype(dst
, BRW_REGISTER_TYPE_UB
);
717 src
= retype(src
, BRW_REGISTER_TYPE_UB
);
718 brw_push_insn_state(p
);
719 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
720 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
721 brw_OR(p
, suboffset(vec1(dst
), 21), vec1(src
), suboffset(vec1(src
), 16));
722 brw_pop_insn_state(p
);
726 vec4_generator::generate_gs_get_instance_id(struct brw_reg dst
)
728 /* We want to right shift R0.0 & R0.1 by GEN7_GS_PAYLOAD_INSTANCE_ID_SHIFT
729 * and store into dst.0 & dst.4. So generate the instruction:
731 * shr(8) dst<1> R0<1,4,0> GEN7_GS_PAYLOAD_INSTANCE_ID_SHIFT { align1 WE_normal 1Q }
733 brw_push_insn_state(p
);
734 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
735 dst
= retype(dst
, BRW_REGISTER_TYPE_UD
);
736 struct brw_reg
r0(retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
737 brw_SHR(p
, dst
, stride(r0
, 1, 4, 0),
738 brw_imm_ud(GEN7_GS_PAYLOAD_INSTANCE_ID_SHIFT
));
739 brw_pop_insn_state(p
);
743 vec4_generator::generate_gs_ff_sync_set_primitives(struct brw_reg dst
,
748 brw_push_insn_state(p
);
749 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
750 /* Save src0 data in 16:31 bits of dst.0 */
751 brw_AND(p
, suboffset(vec1(dst
), 0), suboffset(vec1(src0
), 0),
752 brw_imm_ud(0xffffu
));
753 brw_SHL(p
, suboffset(vec1(dst
), 0), suboffset(vec1(dst
), 0), brw_imm_ud(16));
754 /* Save src1 data in 0:15 bits of dst.0 */
755 brw_AND(p
, suboffset(vec1(src2
), 0), suboffset(vec1(src1
), 0),
756 brw_imm_ud(0xffffu
));
757 brw_OR(p
, suboffset(vec1(dst
), 0),
758 suboffset(vec1(dst
), 0),
759 suboffset(vec1(src2
), 0));
760 brw_pop_insn_state(p
);
764 vec4_generator::generate_gs_ff_sync(vec4_instruction
*inst
,
769 /* This opcode uses an implied MRF register for:
770 * - the header of the ff_sync message. And as such it is expected to be
771 * initialized to r0 before calling here.
772 * - the destination where we will write the allocated URB handle.
774 struct brw_reg header
=
775 retype(brw_message_reg(inst
->base_mrf
), BRW_REGISTER_TYPE_UD
);
777 /* Overwrite dword 0 of the header (SO vertices to write) and
778 * dword 1 (number of primitives written).
780 brw_push_insn_state(p
);
781 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
782 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
783 brw_MOV(p
, get_element_ud(header
, 0), get_element_ud(src1
, 0));
784 brw_MOV(p
, get_element_ud(header
, 1), get_element_ud(src0
, 0));
785 brw_pop_insn_state(p
);
787 /* Allocate URB handle in dst */
793 1, /* response length */
796 /* Now put allocated urb handle in header.0 */
797 brw_push_insn_state(p
);
798 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
799 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
800 brw_MOV(p
, get_element_ud(header
, 0), get_element_ud(dst
, 0));
802 /* src1 is not an immediate when we use transform feedback */
803 if (src1
.file
!= BRW_IMMEDIATE_VALUE
)
804 brw_MOV(p
, brw_vec4_grf(src1
.nr
, 0), brw_vec4_grf(dst
.nr
, 1));
806 brw_pop_insn_state(p
);
810 vec4_generator::generate_gs_set_primitive_id(struct brw_reg dst
)
812 /* In gen6, PrimitiveID is delivered in R0.1 of the payload */
813 struct brw_reg src
= brw_vec8_grf(0, 0);
814 brw_push_insn_state(p
);
815 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
816 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
817 brw_MOV(p
, get_element_ud(dst
, 0), get_element_ud(src
, 1));
818 brw_pop_insn_state(p
);
822 vec4_generator::generate_oword_dual_block_offsets(struct brw_reg m1
,
823 struct brw_reg index
)
825 int second_vertex_offset
;
828 second_vertex_offset
= 1;
830 second_vertex_offset
= 16;
832 m1
= retype(m1
, BRW_REGISTER_TYPE_D
);
834 /* Set up M1 (message payload). Only the block offsets in M1.0 and
835 * M1.4 are used, and the rest are ignored.
837 struct brw_reg m1_0
= suboffset(vec1(m1
), 0);
838 struct brw_reg m1_4
= suboffset(vec1(m1
), 4);
839 struct brw_reg index_0
= suboffset(vec1(index
), 0);
840 struct brw_reg index_4
= suboffset(vec1(index
), 4);
842 brw_push_insn_state(p
);
843 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
844 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
846 brw_MOV(p
, m1_0
, index_0
);
848 if (index
.file
== BRW_IMMEDIATE_VALUE
) {
849 index_4
.dw1
.ud
+= second_vertex_offset
;
850 brw_MOV(p
, m1_4
, index_4
);
852 brw_ADD(p
, m1_4
, index_4
, brw_imm_d(second_vertex_offset
));
855 brw_pop_insn_state(p
);
859 vec4_generator::generate_unpack_flags(struct brw_reg dst
)
861 brw_push_insn_state(p
);
862 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
863 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
865 struct brw_reg flags
= brw_flag_reg(0, 0);
866 struct brw_reg dst_0
= suboffset(vec1(dst
), 0);
867 struct brw_reg dst_4
= suboffset(vec1(dst
), 4);
869 brw_AND(p
, dst_0
, flags
, brw_imm_ud(0x0f));
870 brw_AND(p
, dst_4
, flags
, brw_imm_ud(0xf0));
871 brw_SHR(p
, dst_4
, dst_4
, brw_imm_ud(4));
873 brw_pop_insn_state(p
);
877 vec4_generator::generate_scratch_read(vec4_instruction
*inst
,
879 struct brw_reg index
)
881 struct brw_reg header
= brw_vec8_grf(0, 0);
883 gen6_resolve_implied_move(p
, &header
, inst
->base_mrf
);
885 generate_oword_dual_block_offsets(brw_message_reg(inst
->base_mrf
+ 1),
891 msg_type
= GEN6_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ
;
892 else if (brw
->gen
== 5 || brw
->is_g4x
)
893 msg_type
= G45_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ
;
895 msg_type
= BRW_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ
;
897 /* Each of the 8 channel enables is considered for whether each
900 brw_inst
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
901 brw_set_dest(p
, send
, dst
);
902 brw_set_src0(p
, send
, header
);
904 brw_inst_set_cond_modifier(brw
, send
, inst
->base_mrf
);
905 brw_set_dp_read_message(p
, send
,
906 255, /* binding table index: stateless access */
907 BRW_DATAPORT_OWORD_DUAL_BLOCK_1OWORD
,
909 BRW_DATAPORT_READ_TARGET_RENDER_CACHE
,
911 true, /* header_present */
916 vec4_generator::generate_scratch_write(vec4_instruction
*inst
,
919 struct brw_reg index
)
921 struct brw_reg header
= brw_vec8_grf(0, 0);
924 /* If the instruction is predicated, we'll predicate the send, not
927 brw_set_default_predicate_control(p
, false);
929 gen6_resolve_implied_move(p
, &header
, inst
->base_mrf
);
931 generate_oword_dual_block_offsets(brw_message_reg(inst
->base_mrf
+ 1),
935 retype(brw_message_reg(inst
->base_mrf
+ 2), BRW_REGISTER_TYPE_D
),
936 retype(src
, BRW_REGISTER_TYPE_D
));
941 msg_type
= GEN7_DATAPORT_DC_OWORD_DUAL_BLOCK_WRITE
;
942 else if (brw
->gen
== 6)
943 msg_type
= GEN6_DATAPORT_WRITE_MESSAGE_OWORD_DUAL_BLOCK_WRITE
;
945 msg_type
= BRW_DATAPORT_WRITE_MESSAGE_OWORD_DUAL_BLOCK_WRITE
;
947 brw_set_default_predicate_control(p
, inst
->predicate
);
949 /* Pre-gen6, we have to specify write commits to ensure ordering
950 * between reads and writes within a thread. Afterwards, that's
951 * guaranteed and write commits only matter for inter-thread
955 write_commit
= false;
957 /* The visitor set up our destination register to be g0. This
958 * means that when the next read comes along, we will end up
959 * reading from g0 and causing a block on the write commit. For
960 * write-after-read, we are relying on the value of the previous
961 * read being used (and thus blocking on completion) before our
962 * write is executed. This means we have to be careful in
963 * instruction scheduling to not violate this assumption.
968 /* Each of the 8 channel enables is considered for whether each
971 brw_inst
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
972 brw_set_dest(p
, send
, dst
);
973 brw_set_src0(p
, send
, header
);
975 brw_inst_set_cond_modifier(brw
, send
, inst
->base_mrf
);
976 brw_set_dp_write_message(p
, send
,
977 255, /* binding table index: stateless access */
978 BRW_DATAPORT_OWORD_DUAL_BLOCK_1OWORD
,
981 true, /* header present */
982 false, /* not a render target write */
983 write_commit
, /* rlen */
989 vec4_generator::generate_pull_constant_load(vec4_instruction
*inst
,
991 struct brw_reg index
,
992 struct brw_reg offset
)
994 assert(index
.file
== BRW_IMMEDIATE_VALUE
&&
995 index
.type
== BRW_REGISTER_TYPE_UD
);
996 uint32_t surf_index
= index
.dw1
.ud
;
998 struct brw_reg header
= brw_vec8_grf(0, 0);
1000 gen6_resolve_implied_move(p
, &header
, inst
->base_mrf
);
1002 brw_MOV(p
, retype(brw_message_reg(inst
->base_mrf
+ 1), BRW_REGISTER_TYPE_D
),
1008 msg_type
= GEN6_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ
;
1009 else if (brw
->gen
== 5 || brw
->is_g4x
)
1010 msg_type
= G45_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ
;
1012 msg_type
= BRW_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ
;
1014 /* Each of the 8 channel enables is considered for whether each
1017 brw_inst
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
1018 brw_set_dest(p
, send
, dst
);
1019 brw_set_src0(p
, send
, header
);
1021 brw_inst_set_cond_modifier(brw
, send
, inst
->base_mrf
);
1022 brw_set_dp_read_message(p
, send
,
1024 BRW_DATAPORT_OWORD_DUAL_BLOCK_1OWORD
,
1026 BRW_DATAPORT_READ_TARGET_DATA_CACHE
,
1028 true, /* header_present */
1031 brw_mark_surface_used(&prog_data
->base
, surf_index
);
1035 vec4_generator::generate_pull_constant_load_gen7(vec4_instruction
*inst
,
1037 struct brw_reg surf_index
,
1038 struct brw_reg offset
)
1040 assert(surf_index
.type
== BRW_REGISTER_TYPE_UD
);
1042 struct brw_reg src
= offset
;
1043 bool header_present
= false;
1046 if (brw
->gen
>= 9) {
1047 /* Skylake requires a message header in order to use SIMD4x2 mode. */
1048 src
= retype(brw_vec4_grf(offset
.nr
- 1, 0), BRW_REGISTER_TYPE_UD
);
1050 header_present
= true;
1052 brw_push_insn_state(p
);
1053 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1054 brw_MOV(p
, src
, retype(brw_vec4_grf(0, 0), BRW_REGISTER_TYPE_UD
));
1055 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
1057 brw_MOV(p
, get_element_ud(src
, 2),
1058 brw_imm_ud(GEN9_SAMPLER_SIMD_MODE_EXTENSION_SIMD4X2
));
1059 brw_pop_insn_state(p
);
1062 if (surf_index
.file
== BRW_IMMEDIATE_VALUE
) {
1064 brw_inst
*insn
= brw_next_insn(p
, BRW_OPCODE_SEND
);
1065 brw_set_dest(p
, insn
, dst
);
1066 brw_set_src0(p
, insn
, src
);
1067 brw_set_sampler_message(p
, insn
,
1069 0, /* LD message ignores sampler unit */
1070 GEN5_SAMPLER_MESSAGE_SAMPLE_LD
,
1074 BRW_SAMPLER_SIMD_MODE_SIMD4X2
,
1077 brw_mark_surface_used(&prog_data
->base
, surf_index
.dw1
.ud
);
1081 struct brw_reg addr
= vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD
));
1083 brw_push_insn_state(p
);
1084 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
1085 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
1087 /* a0.0 = surf_index & 0xff */
1088 brw_inst
*insn_and
= brw_next_insn(p
, BRW_OPCODE_AND
);
1089 brw_inst_set_exec_size(p
->brw
, insn_and
, BRW_EXECUTE_1
);
1090 brw_set_dest(p
, insn_and
, addr
);
1091 brw_set_src0(p
, insn_and
, vec1(retype(surf_index
, BRW_REGISTER_TYPE_UD
)));
1092 brw_set_src1(p
, insn_and
, brw_imm_ud(0x0ff));
1094 /* dst = send(offset, a0.0 | <descriptor>) */
1095 brw_inst
*insn
= brw_send_indirect_message(
1096 p
, BRW_SFID_SAMPLER
, dst
, src
, addr
);
1097 brw_set_sampler_message(p
, insn
,
1100 GEN5_SAMPLER_MESSAGE_SAMPLE_LD
,
1103 header_present
/* header */,
1104 BRW_SAMPLER_SIMD_MODE_SIMD4X2
,
1107 brw_pop_insn_state(p
);
1109 /* visitor knows more than we do about the surface limit required,
1110 * so has already done marking.
1116 vec4_generator::generate_untyped_atomic(vec4_instruction
*inst
,
1118 struct brw_reg atomic_op
,
1119 struct brw_reg surf_index
)
1121 assert(atomic_op
.file
== BRW_IMMEDIATE_VALUE
&&
1122 atomic_op
.type
== BRW_REGISTER_TYPE_UD
&&
1123 surf_index
.file
== BRW_IMMEDIATE_VALUE
&&
1124 surf_index
.type
== BRW_REGISTER_TYPE_UD
);
1126 brw_untyped_atomic(p
, dst
, brw_message_reg(inst
->base_mrf
),
1127 atomic_op
.dw1
.ud
, surf_index
.dw1
.ud
,
1130 brw_mark_surface_used(&prog_data
->base
, surf_index
.dw1
.ud
);
1134 vec4_generator::generate_untyped_surface_read(vec4_instruction
*inst
,
1136 struct brw_reg surf_index
)
1138 assert(surf_index
.file
== BRW_IMMEDIATE_VALUE
&&
1139 surf_index
.type
== BRW_REGISTER_TYPE_UD
);
1141 brw_untyped_surface_read(p
, dst
, brw_message_reg(inst
->base_mrf
),
1145 brw_mark_surface_used(&prog_data
->base
, surf_index
.dw1
.ud
);
1149 vec4_generator::generate_code(const cfg_t
*cfg
)
1151 struct annotation_info annotation
;
1152 memset(&annotation
, 0, sizeof(annotation
));
1155 foreach_block_and_inst (block
, vec4_instruction
, inst
, cfg
) {
1156 struct brw_reg src
[3], dst
;
1158 if (unlikely(debug_flag
))
1159 annotate(brw
, &annotation
, cfg
, inst
, p
->next_insn_offset
);
1161 for (unsigned int i
= 0; i
< 3; i
++) {
1162 src
[i
] = inst
->get_src(this->prog_data
, i
);
1164 dst
= inst
->get_dst();
1166 brw_set_default_predicate_control(p
, inst
->predicate
);
1167 brw_set_default_predicate_inverse(p
, inst
->predicate_inverse
);
1168 brw_set_default_flag_reg(p
, 0, inst
->flag_subreg
);
1169 brw_set_default_saturate(p
, inst
->saturate
);
1170 brw_set_default_mask_control(p
, inst
->force_writemask_all
);
1171 brw_set_default_acc_write_control(p
, inst
->writes_accumulator
);
1173 unsigned pre_emit_nr_insn
= p
->nr_insn
;
1175 if (dst
.width
== BRW_WIDTH_4
) {
1176 /* This happens in attribute fixups for "dual instanced" geometry
1177 * shaders, since they use attributes that are vec4's. Since the exec
1178 * width is only 4, it's essential that the caller set
1179 * force_writemask_all in order to make sure the instruction is executed
1180 * regardless of which channels are enabled.
1182 assert(inst
->force_writemask_all
);
1184 /* Fix up any <8;8,1> or <0;4,1> source registers to <4;4,1> to satisfy
1185 * the following register region restrictions (from Graphics BSpec:
1186 * 3D-Media-GPGPU Engine > EU Overview > Registers and Register Regions
1187 * > Register Region Restrictions)
1189 * 1. ExecSize must be greater than or equal to Width.
1191 * 2. If ExecSize = Width and HorzStride != 0, VertStride must be set
1192 * to Width * HorzStride."
1194 for (int i
= 0; i
< 3; i
++) {
1195 if (src
[i
].file
== BRW_GENERAL_REGISTER_FILE
)
1196 src
[i
] = stride(src
[i
], 4, 4, 1);
1200 switch (inst
->opcode
) {
1201 case VEC4_OPCODE_UNPACK_UNIFORM
:
1202 case BRW_OPCODE_MOV
:
1203 brw_MOV(p
, dst
, src
[0]);
1205 case BRW_OPCODE_ADD
:
1206 brw_ADD(p
, dst
, src
[0], src
[1]);
1208 case BRW_OPCODE_MUL
:
1209 brw_MUL(p
, dst
, src
[0], src
[1]);
1211 case BRW_OPCODE_MACH
:
1212 brw_MACH(p
, dst
, src
[0], src
[1]);
1215 case BRW_OPCODE_MAD
:
1216 assert(brw
->gen
>= 6);
1217 brw_MAD(p
, dst
, src
[0], src
[1], src
[2]);
1220 case BRW_OPCODE_FRC
:
1221 brw_FRC(p
, dst
, src
[0]);
1223 case BRW_OPCODE_RNDD
:
1224 brw_RNDD(p
, dst
, src
[0]);
1226 case BRW_OPCODE_RNDE
:
1227 brw_RNDE(p
, dst
, src
[0]);
1229 case BRW_OPCODE_RNDZ
:
1230 brw_RNDZ(p
, dst
, src
[0]);
1233 case BRW_OPCODE_AND
:
1234 brw_AND(p
, dst
, src
[0], src
[1]);
1237 brw_OR(p
, dst
, src
[0], src
[1]);
1239 case BRW_OPCODE_XOR
:
1240 brw_XOR(p
, dst
, src
[0], src
[1]);
1242 case BRW_OPCODE_NOT
:
1243 brw_NOT(p
, dst
, src
[0]);
1245 case BRW_OPCODE_ASR
:
1246 brw_ASR(p
, dst
, src
[0], src
[1]);
1248 case BRW_OPCODE_SHR
:
1249 brw_SHR(p
, dst
, src
[0], src
[1]);
1251 case BRW_OPCODE_SHL
:
1252 brw_SHL(p
, dst
, src
[0], src
[1]);
1255 case BRW_OPCODE_CMP
:
1256 brw_CMP(p
, dst
, inst
->conditional_mod
, src
[0], src
[1]);
1258 case BRW_OPCODE_SEL
:
1259 brw_SEL(p
, dst
, src
[0], src
[1]);
1262 case BRW_OPCODE_DPH
:
1263 brw_DPH(p
, dst
, src
[0], src
[1]);
1266 case BRW_OPCODE_DP4
:
1267 brw_DP4(p
, dst
, src
[0], src
[1]);
1270 case BRW_OPCODE_DP3
:
1271 brw_DP3(p
, dst
, src
[0], src
[1]);
1274 case BRW_OPCODE_DP2
:
1275 brw_DP2(p
, dst
, src
[0], src
[1]);
1278 case BRW_OPCODE_F32TO16
:
1279 assert(brw
->gen
>= 7);
1280 brw_F32TO16(p
, dst
, src
[0]);
1283 case BRW_OPCODE_F16TO32
:
1284 assert(brw
->gen
>= 7);
1285 brw_F16TO32(p
, dst
, src
[0]);
1288 case BRW_OPCODE_LRP
:
1289 assert(brw
->gen
>= 6);
1290 brw_LRP(p
, dst
, src
[0], src
[1], src
[2]);
1293 case BRW_OPCODE_BFREV
:
1294 assert(brw
->gen
>= 7);
1295 /* BFREV only supports UD type for src and dst. */
1296 brw_BFREV(p
, retype(dst
, BRW_REGISTER_TYPE_UD
),
1297 retype(src
[0], BRW_REGISTER_TYPE_UD
));
1299 case BRW_OPCODE_FBH
:
1300 assert(brw
->gen
>= 7);
1301 /* FBH only supports UD type for dst. */
1302 brw_FBH(p
, retype(dst
, BRW_REGISTER_TYPE_UD
), src
[0]);
1304 case BRW_OPCODE_FBL
:
1305 assert(brw
->gen
>= 7);
1306 /* FBL only supports UD type for dst. */
1307 brw_FBL(p
, retype(dst
, BRW_REGISTER_TYPE_UD
), src
[0]);
1309 case BRW_OPCODE_CBIT
:
1310 assert(brw
->gen
>= 7);
1311 /* CBIT only supports UD type for dst. */
1312 brw_CBIT(p
, retype(dst
, BRW_REGISTER_TYPE_UD
), src
[0]);
1314 case BRW_OPCODE_ADDC
:
1315 assert(brw
->gen
>= 7);
1316 brw_ADDC(p
, dst
, src
[0], src
[1]);
1318 case BRW_OPCODE_SUBB
:
1319 assert(brw
->gen
>= 7);
1320 brw_SUBB(p
, dst
, src
[0], src
[1]);
1322 case BRW_OPCODE_MAC
:
1323 brw_MAC(p
, dst
, src
[0], src
[1]);
1326 case BRW_OPCODE_BFE
:
1327 assert(brw
->gen
>= 7);
1328 brw_BFE(p
, dst
, src
[0], src
[1], src
[2]);
1331 case BRW_OPCODE_BFI1
:
1332 assert(brw
->gen
>= 7);
1333 brw_BFI1(p
, dst
, src
[0], src
[1]);
1335 case BRW_OPCODE_BFI2
:
1336 assert(brw
->gen
>= 7);
1337 brw_BFI2(p
, dst
, src
[0], src
[1], src
[2]);
1341 if (inst
->src
[0].file
!= BAD_FILE
) {
1342 /* The instruction has an embedded compare (only allowed on gen6) */
1343 assert(brw
->gen
== 6);
1344 gen6_IF(p
, inst
->conditional_mod
, src
[0], src
[1]);
1346 brw_inst
*if_inst
= brw_IF(p
, BRW_EXECUTE_8
);
1347 brw_inst_set_pred_control(brw
, if_inst
, inst
->predicate
);
1351 case BRW_OPCODE_ELSE
:
1354 case BRW_OPCODE_ENDIF
:
1359 brw_DO(p
, BRW_EXECUTE_8
);
1362 case BRW_OPCODE_BREAK
:
1364 brw_set_default_predicate_control(p
, BRW_PREDICATE_NONE
);
1366 case BRW_OPCODE_CONTINUE
:
1368 brw_set_default_predicate_control(p
, BRW_PREDICATE_NONE
);
1371 case BRW_OPCODE_WHILE
:
1376 case SHADER_OPCODE_RCP
:
1377 case SHADER_OPCODE_RSQ
:
1378 case SHADER_OPCODE_SQRT
:
1379 case SHADER_OPCODE_EXP2
:
1380 case SHADER_OPCODE_LOG2
:
1381 case SHADER_OPCODE_SIN
:
1382 case SHADER_OPCODE_COS
:
1383 assert(inst
->conditional_mod
== BRW_CONDITIONAL_NONE
);
1384 if (brw
->gen
>= 7) {
1385 gen6_math(p
, dst
, brw_math_function(inst
->opcode
), src
[0],
1387 } else if (brw
->gen
== 6) {
1388 generate_math_gen6(inst
, dst
, src
[0], brw_null_reg());
1390 generate_math1_gen4(inst
, dst
, src
[0]);
1394 case SHADER_OPCODE_POW
:
1395 case SHADER_OPCODE_INT_QUOTIENT
:
1396 case SHADER_OPCODE_INT_REMAINDER
:
1397 assert(inst
->conditional_mod
== BRW_CONDITIONAL_NONE
);
1398 if (brw
->gen
>= 7) {
1399 gen6_math(p
, dst
, brw_math_function(inst
->opcode
), src
[0], src
[1]);
1400 } else if (brw
->gen
== 6) {
1401 generate_math_gen6(inst
, dst
, src
[0], src
[1]);
1403 generate_math2_gen4(inst
, dst
, src
[0], src
[1]);
1407 case SHADER_OPCODE_TEX
:
1408 case SHADER_OPCODE_TXD
:
1409 case SHADER_OPCODE_TXF
:
1410 case SHADER_OPCODE_TXF_CMS
:
1411 case SHADER_OPCODE_TXF_MCS
:
1412 case SHADER_OPCODE_TXL
:
1413 case SHADER_OPCODE_TXS
:
1414 case SHADER_OPCODE_TG4
:
1415 case SHADER_OPCODE_TG4_OFFSET
:
1416 generate_tex(inst
, dst
, src
[0], src
[1]);
1419 case VS_OPCODE_URB_WRITE
:
1420 generate_vs_urb_write(inst
);
1423 case SHADER_OPCODE_GEN4_SCRATCH_READ
:
1424 generate_scratch_read(inst
, dst
, src
[0]);
1427 case SHADER_OPCODE_GEN4_SCRATCH_WRITE
:
1428 generate_scratch_write(inst
, dst
, src
[0], src
[1]);
1431 case VS_OPCODE_PULL_CONSTANT_LOAD
:
1432 generate_pull_constant_load(inst
, dst
, src
[0], src
[1]);
1435 case VS_OPCODE_PULL_CONSTANT_LOAD_GEN7
:
1436 generate_pull_constant_load_gen7(inst
, dst
, src
[0], src
[1]);
1439 case GS_OPCODE_URB_WRITE
:
1440 generate_gs_urb_write(inst
);
1443 case GS_OPCODE_URB_WRITE_ALLOCATE
:
1444 generate_gs_urb_write_allocate(inst
);
1447 case GS_OPCODE_SVB_WRITE
:
1448 generate_gs_svb_write(inst
, dst
, src
[0], src
[1]);
1451 case GS_OPCODE_SVB_SET_DST_INDEX
:
1452 generate_gs_svb_set_destination_index(inst
, dst
, src
[0]);
1455 case GS_OPCODE_THREAD_END
:
1456 generate_gs_thread_end(inst
);
1459 case GS_OPCODE_SET_WRITE_OFFSET
:
1460 generate_gs_set_write_offset(dst
, src
[0], src
[1]);
1463 case GS_OPCODE_SET_VERTEX_COUNT
:
1464 generate_gs_set_vertex_count(dst
, src
[0]);
1467 case GS_OPCODE_FF_SYNC
:
1468 generate_gs_ff_sync(inst
, dst
, src
[0], src
[1]);
1471 case GS_OPCODE_FF_SYNC_SET_PRIMITIVES
:
1472 generate_gs_ff_sync_set_primitives(dst
, src
[0], src
[1], src
[2]);
1475 case GS_OPCODE_SET_PRIMITIVE_ID
:
1476 generate_gs_set_primitive_id(dst
);
1479 case GS_OPCODE_SET_DWORD_2
:
1480 generate_gs_set_dword_2(dst
, src
[0]);
1483 case GS_OPCODE_PREPARE_CHANNEL_MASKS
:
1484 generate_gs_prepare_channel_masks(dst
);
1487 case GS_OPCODE_SET_CHANNEL_MASKS
:
1488 generate_gs_set_channel_masks(dst
, src
[0]);
1491 case GS_OPCODE_GET_INSTANCE_ID
:
1492 generate_gs_get_instance_id(dst
);
1495 case SHADER_OPCODE_SHADER_TIME_ADD
:
1496 brw_shader_time_add(p
, src
[0],
1497 prog_data
->base
.binding_table
.shader_time_start
);
1498 brw_mark_surface_used(&prog_data
->base
,
1499 prog_data
->base
.binding_table
.shader_time_start
);
1502 case SHADER_OPCODE_UNTYPED_ATOMIC
:
1503 generate_untyped_atomic(inst
, dst
, src
[0], src
[1]);
1506 case SHADER_OPCODE_UNTYPED_SURFACE_READ
:
1507 generate_untyped_surface_read(inst
, dst
, src
[0]);
1510 case VS_OPCODE_UNPACK_FLAGS_SIMD4X2
:
1511 generate_unpack_flags(dst
);
1514 case VEC4_OPCODE_MOV_BYTES
: {
1515 /* Moves the low byte from each channel, using an Align1 access mode
1516 * and a <4,1,0> source region.
1518 assert(src
[0].type
== BRW_REGISTER_TYPE_UB
||
1519 src
[0].type
== BRW_REGISTER_TYPE_B
);
1521 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
1522 src
[0].vstride
= BRW_VERTICAL_STRIDE_4
;
1523 src
[0].width
= BRW_WIDTH_1
;
1524 src
[0].hstride
= BRW_HORIZONTAL_STRIDE_0
;
1525 brw_MOV(p
, dst
, src
[0]);
1526 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1530 case VEC4_OPCODE_PACK_BYTES
: {
1533 * mov(8) dst<16,4,1>:UB src<4,1,0>:UB
1535 * but destinations' only regioning is horizontal stride, so instead we
1536 * have to use two instructions:
1538 * mov(4) dst<1>:UB src<4,1,0>:UB
1539 * mov(4) dst.16<1>:UB src.16<4,1,0>:UB
1541 * where they pack the four bytes from the low and high four DW.
1543 assert(is_power_of_two(dst
.dw1
.bits
.writemask
) &&
1544 dst
.dw1
.bits
.writemask
!= 0);
1545 unsigned offset
= __builtin_ctz(dst
.dw1
.bits
.writemask
);
1547 dst
.type
= BRW_REGISTER_TYPE_UB
;
1549 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
1551 src
[0].type
= BRW_REGISTER_TYPE_UB
;
1552 src
[0].vstride
= BRW_VERTICAL_STRIDE_4
;
1553 src
[0].width
= BRW_WIDTH_1
;
1554 src
[0].hstride
= BRW_HORIZONTAL_STRIDE_0
;
1555 dst
.subnr
= offset
* 4;
1556 struct brw_inst
*insn
= brw_MOV(p
, dst
, src
[0]);
1557 brw_inst_set_exec_size(brw
, insn
, BRW_EXECUTE_4
);
1558 brw_inst_set_no_dd_clear(brw
, insn
, true);
1559 brw_inst_set_no_dd_check(brw
, insn
, inst
->no_dd_check
);
1562 dst
.subnr
= 16 + offset
* 4;
1563 insn
= brw_MOV(p
, dst
, src
[0]);
1564 brw_inst_set_exec_size(brw
, insn
, BRW_EXECUTE_4
);
1565 brw_inst_set_no_dd_clear(brw
, insn
, inst
->no_dd_clear
);
1566 brw_inst_set_no_dd_check(brw
, insn
, true);
1568 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1573 if (inst
->opcode
< (int) ARRAY_SIZE(opcode_descs
)) {
1574 _mesa_problem(&brw
->ctx
, "Unsupported opcode in `%s' in vec4\n",
1575 opcode_descs
[inst
->opcode
].name
);
1577 _mesa_problem(&brw
->ctx
, "Unsupported opcode %d in vec4", inst
->opcode
);
1582 if (inst
->opcode
== VEC4_OPCODE_PACK_BYTES
) {
1583 /* Handled dependency hints in the generator. */
1585 assert(!inst
->conditional_mod
);
1586 } else if (inst
->no_dd_clear
|| inst
->no_dd_check
|| inst
->conditional_mod
) {
1587 assert(p
->nr_insn
== pre_emit_nr_insn
+ 1 ||
1588 !"conditional_mod, no_dd_check, or no_dd_clear set for IR "
1589 "emitting more than 1 instruction");
1591 brw_inst
*last
= &p
->store
[pre_emit_nr_insn
];
1593 if (inst
->conditional_mod
)
1594 brw_inst_set_cond_modifier(brw
, last
, inst
->conditional_mod
);
1595 brw_inst_set_no_dd_clear(brw
, last
, inst
->no_dd_clear
);
1596 brw_inst_set_no_dd_check(brw
, last
, inst
->no_dd_check
);
1601 annotation_finalize(&annotation
, p
->next_insn_offset
);
1603 int before_size
= p
->next_insn_offset
;
1604 brw_compact_instructions(p
, 0, annotation
.ann_count
, annotation
.ann
);
1605 int after_size
= p
->next_insn_offset
;
1607 if (unlikely(debug_flag
)) {
1609 fprintf(stderr
, "Native code for %s %s shader %d:\n",
1610 shader_prog
->Label
? shader_prog
->Label
: "unnamed",
1611 stage_name
, shader_prog
->Name
);
1613 fprintf(stderr
, "Native code for %s program %d:\n", stage_name
,
1616 fprintf(stderr
, "%s vec4 shader: %d instructions. %d loops. Compacted %d to %d"
1617 " bytes (%.0f%%)\n",
1619 before_size
/ 16, loop_count
, before_size
, after_size
,
1620 100.0f
* (before_size
- after_size
) / before_size
);
1622 dump_assembly(p
->store
, annotation
.ann_count
, annotation
.ann
, brw
, prog
);
1623 ralloc_free(annotation
.ann
);
1626 static GLuint msg_id
= 0;
1627 _mesa_gl_debug(&brw
->ctx
, &msg_id
,
1628 MESA_DEBUG_SOURCE_SHADER_COMPILER
,
1629 MESA_DEBUG_TYPE_OTHER
,
1630 MESA_DEBUG_SEVERITY_NOTIFICATION
,
1631 "%s vec4 shader: %d inst, %d loops, "
1632 "compacted %d to %d bytes.\n",
1634 before_size
/ 16, loop_count
,
1635 before_size
, after_size
);
1639 vec4_generator::generate_assembly(const cfg_t
*cfg
,
1640 unsigned *assembly_size
)
1642 brw_set_default_access_mode(p
, BRW_ALIGN_16
);
1645 return brw_get_program(p
, assembly_size
);
1648 } /* namespace brw */