2 Copyright (C) Intel Corp. 2006. All Rights Reserved.
3 Intel funded Tungsten Graphics to
4 develop this 3D driver.
6 Permission is hereby granted, free of charge, to any person obtaining
7 a copy of this software and associated documentation files (the
8 "Software"), to deal in the Software without restriction, including
9 without limitation the rights to use, copy, modify, merge, publish,
10 distribute, sublicense, and/or sell copies of the Software, and to
11 permit persons to whom the Software is furnished to do so, subject to
12 the following conditions:
14 The above copyright notice and this permission notice (including the
15 next paragraph) shall be included in all copies or substantial
16 portions of the Software.
18 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
19 EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
21 IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
22 LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
23 OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
24 WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **********************************************************************/
29 * Keith Whitwell <keithw@vmware.com>
39 #include "brw_eu_defines.h"
41 #include "brw_disasm_info.h"
47 #define BRW_EU_MAX_INSN_STACK 5
49 struct brw_insn_state
{
50 /* One of BRW_EXECUTE_* */
53 /* Group in units of channels */
56 /* Compression control on gen4-5 */
59 /* One of BRW_MASK_* */
60 unsigned mask_control
:1;
64 /* One of BRW_ALIGN_* */
65 unsigned access_mode
:1;
67 /* One of BRW_PREDICATE_* */
68 enum brw_predicate predicate
:4;
72 /* Flag subreg. Bottom bit is subreg, top bit is reg */
73 unsigned flag_subreg
:2;
75 bool acc_wr_control
:1;
79 /* A helper for accessing the last instruction emitted. This makes it easy
80 * to set various bits on an instruction without having to create temporary
81 * variable and assign the emitted instruction to those.
83 #define brw_last_inst (&p->store[p->nr_insn - 1])
89 unsigned int next_insn_offset
;
93 /* Allow clients to push/pop instruction state:
95 struct brw_insn_state stack
[BRW_EU_MAX_INSN_STACK
];
96 struct brw_insn_state
*current
;
98 /** Whether or not the user wants automatic exec sizes
100 * If true, codegen will try to automatically infer the exec size of an
101 * instruction from the width of the destination register. If false, it
102 * will take whatever is set by brw_set_default_exec_size verbatim.
104 * This is set to true by default in brw_init_codegen.
106 bool automatic_exec_sizes
;
108 bool single_program_flow
;
109 const struct gen_device_info
*devinfo
;
111 /* Control flow stacks:
112 * - if_stack contains IF and ELSE instructions which must be patched
113 * (and popped) once the matching ENDIF instruction is encountered.
115 * Just store the instruction pointer(an index).
119 int if_stack_array_size
;
122 * loop_stack contains the instruction pointers of the starts of loops which
123 * must be patched (and popped) once the matching WHILE instruction is
128 * pre-gen6, the BREAK and CONT instructions had to tell how many IF/ENDIF
129 * blocks they were popping out of, to fix up the mask stack. This tracks
130 * the IF/ENDIF nesting in each current nested loop level.
132 int *if_depth_in_loop
;
133 int loop_stack_depth
;
134 int loop_stack_array_size
;
137 void brw_pop_insn_state( struct brw_codegen
*p
);
138 void brw_push_insn_state( struct brw_codegen
*p
);
139 unsigned brw_get_default_exec_size(struct brw_codegen
*p
);
140 unsigned brw_get_default_group(struct brw_codegen
*p
);
141 unsigned brw_get_default_access_mode(struct brw_codegen
*p
);
142 void brw_set_default_exec_size(struct brw_codegen
*p
, unsigned value
);
143 void brw_set_default_mask_control( struct brw_codegen
*p
, unsigned value
);
144 void brw_set_default_saturate( struct brw_codegen
*p
, bool enable
);
145 void brw_set_default_access_mode( struct brw_codegen
*p
, unsigned access_mode
);
146 void brw_inst_set_compression(const struct gen_device_info
*devinfo
,
147 brw_inst
*inst
, bool on
);
148 void brw_set_default_compression(struct brw_codegen
*p
, bool on
);
149 void brw_inst_set_group(const struct gen_device_info
*devinfo
,
150 brw_inst
*inst
, unsigned group
);
151 void brw_set_default_group(struct brw_codegen
*p
, unsigned group
);
152 void brw_set_default_compression_control(struct brw_codegen
*p
, enum brw_compression c
);
153 void brw_set_default_predicate_control( struct brw_codegen
*p
, unsigned pc
);
154 void brw_set_default_predicate_inverse(struct brw_codegen
*p
, bool predicate_inverse
);
155 void brw_set_default_flag_reg(struct brw_codegen
*p
, int reg
, int subreg
);
156 void brw_set_default_acc_write_control(struct brw_codegen
*p
, unsigned value
);
158 void brw_init_codegen(const struct gen_device_info
*, struct brw_codegen
*p
,
160 int brw_disassemble_inst(FILE *file
, const struct gen_device_info
*devinfo
,
161 const struct brw_inst
*inst
, bool is_compacted
);
162 void brw_disassemble(const struct gen_device_info
*devinfo
,
163 const void *assembly
, int start
, int end
, FILE *out
);
164 const unsigned *brw_get_program( struct brw_codegen
*p
, unsigned *sz
);
166 brw_inst
*brw_next_insn(struct brw_codegen
*p
, unsigned opcode
);
167 void brw_set_dest(struct brw_codegen
*p
, brw_inst
*insn
, struct brw_reg dest
);
168 void brw_set_src0(struct brw_codegen
*p
, brw_inst
*insn
, struct brw_reg reg
);
170 void gen6_resolve_implied_move(struct brw_codegen
*p
,
172 unsigned msg_reg_nr
);
174 /* Helpers for regular instructions:
177 brw_inst *brw_##OP(struct brw_codegen *p, \
178 struct brw_reg dest, \
179 struct brw_reg src0);
182 brw_inst *brw_##OP(struct brw_codegen *p, \
183 struct brw_reg dest, \
184 struct brw_reg src0, \
185 struct brw_reg src1);
188 brw_inst *brw_##OP(struct brw_codegen *p, \
189 struct brw_reg dest, \
190 struct brw_reg src0, \
191 struct brw_reg src1, \
192 struct brw_reg src2);
195 void brw_##OP(struct brw_codegen *p, struct brw_reg dest, struct brw_reg src0);
246 /* Helpers for SEND instruction:
250 * Construct a message descriptor immediate with the specified common
251 * descriptor controls.
253 static inline uint32_t
254 brw_message_desc(const struct gen_device_info
*devinfo
,
256 unsigned response_length
,
259 if (devinfo
->gen
>= 5) {
260 return (SET_BITS(msg_length
, 28, 25) |
261 SET_BITS(response_length
, 24, 20) |
262 SET_BITS(header_present
, 19, 19));
264 return (SET_BITS(msg_length
, 23, 20) |
265 SET_BITS(response_length
, 19, 16));
269 static inline unsigned
270 brw_message_desc_mlen(const struct gen_device_info
*devinfo
, uint32_t desc
)
272 if (devinfo
->gen
>= 5)
273 return GET_BITS(desc
, 28, 25);
275 return GET_BITS(desc
, 23, 20);
278 static inline unsigned
279 brw_message_desc_rlen(const struct gen_device_info
*devinfo
, uint32_t desc
)
281 if (devinfo
->gen
>= 5)
282 return GET_BITS(desc
, 24, 20);
284 return GET_BITS(desc
, 19, 16);
288 brw_message_desc_header_present(const struct gen_device_info
*devinfo
,
291 assert(devinfo
->gen
>= 5);
292 return GET_BITS(desc
, 19, 19);
295 static inline unsigned
296 brw_message_ex_desc(const struct gen_device_info
*devinfo
,
297 unsigned ex_msg_length
)
299 return SET_BITS(ex_msg_length
, 9, 6);
302 static inline unsigned
303 brw_message_ex_desc_ex_mlen(const struct gen_device_info
*devinfo
,
306 return GET_BITS(ex_desc
, 9, 6);
310 * Construct a message descriptor immediate with the specified sampler
313 static inline uint32_t
314 brw_sampler_desc(const struct gen_device_info
*devinfo
,
315 unsigned binding_table_index
,
319 unsigned return_format
)
321 const unsigned desc
= (SET_BITS(binding_table_index
, 7, 0) |
322 SET_BITS(sampler
, 11, 8));
323 if (devinfo
->gen
>= 7)
324 return (desc
| SET_BITS(msg_type
, 16, 12) |
325 SET_BITS(simd_mode
, 18, 17));
326 else if (devinfo
->gen
>= 5)
327 return (desc
| SET_BITS(msg_type
, 15, 12) |
328 SET_BITS(simd_mode
, 17, 16));
329 else if (devinfo
->is_g4x
)
330 return desc
| SET_BITS(msg_type
, 15, 12);
332 return (desc
| SET_BITS(return_format
, 13, 12) |
333 SET_BITS(msg_type
, 15, 14));
336 static inline unsigned
337 brw_sampler_desc_binding_table_index(const struct gen_device_info
*devinfo
,
340 return GET_BITS(desc
, 7, 0);
343 static inline unsigned
344 brw_sampler_desc_sampler(const struct gen_device_info
*devinfo
, uint32_t desc
)
346 return GET_BITS(desc
, 11, 8);
349 static inline unsigned
350 brw_sampler_desc_msg_type(const struct gen_device_info
*devinfo
, uint32_t desc
)
352 if (devinfo
->gen
>= 7)
353 return GET_BITS(desc
, 16, 12);
354 else if (devinfo
->gen
>= 5 || devinfo
->is_g4x
)
355 return GET_BITS(desc
, 15, 12);
357 return GET_BITS(desc
, 15, 14);
360 static inline unsigned
361 brw_sampler_desc_simd_mode(const struct gen_device_info
*devinfo
, uint32_t desc
)
363 assert(devinfo
->gen
>= 5);
364 if (devinfo
->gen
>= 7)
365 return GET_BITS(desc
, 18, 17);
367 return GET_BITS(desc
, 17, 16);
370 static inline unsigned
371 brw_sampler_desc_return_format(const struct gen_device_info
*devinfo
,
374 assert(devinfo
->gen
== 4 && !devinfo
->is_g4x
);
375 return GET_BITS(desc
, 13, 12);
379 * Construct a message descriptor for the dataport
381 static inline uint32_t
382 brw_dp_desc(const struct gen_device_info
*devinfo
,
383 unsigned binding_table_index
,
385 unsigned msg_control
)
387 /* Prior to gen6, things are too inconsistent; use the dp_read/write_desc
390 assert(devinfo
->gen
>= 6);
391 const unsigned desc
= SET_BITS(binding_table_index
, 7, 0);
392 if (devinfo
->gen
>= 8) {
393 return (desc
| SET_BITS(msg_control
, 13, 8) |
394 SET_BITS(msg_type
, 18, 14));
395 } else if (devinfo
->gen
>= 7) {
396 return (desc
| SET_BITS(msg_control
, 13, 8) |
397 SET_BITS(msg_type
, 17, 14));
399 return (desc
| SET_BITS(msg_control
, 12, 8) |
400 SET_BITS(msg_type
, 16, 13));
404 static inline unsigned
405 brw_dp_desc_binding_table_index(const struct gen_device_info
*devinfo
,
408 return GET_BITS(desc
, 7, 0);
411 static inline unsigned
412 brw_dp_desc_msg_type(const struct gen_device_info
*devinfo
, uint32_t desc
)
414 assert(devinfo
->gen
>= 6);
415 if (devinfo
->gen
>= 8)
416 return GET_BITS(desc
, 18, 14);
417 else if (devinfo
->gen
>= 7)
418 return GET_BITS(desc
, 17, 14);
420 return GET_BITS(desc
, 16, 13);
423 static inline unsigned
424 brw_dp_desc_msg_control(const struct gen_device_info
*devinfo
, uint32_t desc
)
426 assert(devinfo
->gen
>= 6);
427 if (devinfo
->gen
>= 7)
428 return GET_BITS(desc
, 13, 8);
430 return GET_BITS(desc
, 12, 8);
434 * Construct a message descriptor immediate with the specified dataport read
437 static inline uint32_t
438 brw_dp_read_desc(const struct gen_device_info
*devinfo
,
439 unsigned binding_table_index
,
440 unsigned msg_control
,
442 unsigned target_cache
)
444 if (devinfo
->gen
>= 6)
445 return brw_dp_desc(devinfo
, binding_table_index
, msg_type
, msg_control
);
446 else if (devinfo
->gen
>= 5 || devinfo
->is_g4x
)
447 return (SET_BITS(binding_table_index
, 7, 0) |
448 SET_BITS(msg_control
, 10, 8) |
449 SET_BITS(msg_type
, 13, 11) |
450 SET_BITS(target_cache
, 15, 14));
452 return (SET_BITS(binding_table_index
, 7, 0) |
453 SET_BITS(msg_control
, 11, 8) |
454 SET_BITS(msg_type
, 13, 12) |
455 SET_BITS(target_cache
, 15, 14));
458 static inline unsigned
459 brw_dp_read_desc_msg_type(const struct gen_device_info
*devinfo
, uint32_t desc
)
461 if (devinfo
->gen
>= 6)
462 return brw_dp_desc_msg_type(devinfo
, desc
);
463 else if (devinfo
->gen
>= 5 || devinfo
->is_g4x
)
464 return GET_BITS(desc
, 13, 11);
466 return GET_BITS(desc
, 13, 12);
469 static inline unsigned
470 brw_dp_read_desc_msg_control(const struct gen_device_info
*devinfo
,
473 if (devinfo
->gen
>= 6)
474 return brw_dp_desc_msg_control(devinfo
, desc
);
475 else if (devinfo
->gen
>= 5 || devinfo
->is_g4x
)
476 return GET_BITS(desc
, 10, 8);
478 return GET_BITS(desc
, 11, 8);
482 * Construct a message descriptor immediate with the specified dataport write
485 static inline uint32_t
486 brw_dp_write_desc(const struct gen_device_info
*devinfo
,
487 unsigned binding_table_index
,
488 unsigned msg_control
,
490 unsigned last_render_target
,
491 unsigned send_commit_msg
)
493 assert(devinfo
->gen
<= 6 || !send_commit_msg
);
494 if (devinfo
->gen
>= 6)
495 return brw_dp_desc(devinfo
, binding_table_index
, msg_type
, msg_control
) |
496 SET_BITS(last_render_target
, 12, 12) |
497 SET_BITS(send_commit_msg
, 17, 17);
499 return (SET_BITS(binding_table_index
, 7, 0) |
500 SET_BITS(msg_control
, 11, 8) |
501 SET_BITS(last_render_target
, 11, 11) |
502 SET_BITS(msg_type
, 14, 12) |
503 SET_BITS(send_commit_msg
, 15, 15));
506 static inline unsigned
507 brw_dp_write_desc_msg_type(const struct gen_device_info
*devinfo
,
510 if (devinfo
->gen
>= 6)
511 return brw_dp_desc_msg_type(devinfo
, desc
);
513 return GET_BITS(desc
, 14, 12);
516 static inline unsigned
517 brw_dp_write_desc_msg_control(const struct gen_device_info
*devinfo
,
520 if (devinfo
->gen
>= 6)
521 return brw_dp_desc_msg_control(devinfo
, desc
);
523 return GET_BITS(desc
, 11, 8);
527 brw_dp_write_desc_last_render_target(const struct gen_device_info
*devinfo
,
530 if (devinfo
->gen
>= 6)
531 return GET_BITS(desc
, 12, 12);
533 return GET_BITS(desc
, 11, 11);
537 brw_dp_write_desc_write_commit(const struct gen_device_info
*devinfo
,
540 assert(devinfo
->gen
<= 6);
541 if (devinfo
->gen
>= 6)
542 return GET_BITS(desc
, 17, 17);
544 return GET_BITS(desc
, 15, 15);
548 * Construct a message descriptor immediate with the specified dataport
549 * surface function controls.
551 static inline uint32_t
552 brw_dp_surface_desc(const struct gen_device_info
*devinfo
,
554 unsigned msg_control
)
556 assert(devinfo
->gen
>= 7);
557 /* We'll OR in the binding table index later */
558 return brw_dp_desc(devinfo
, 0, msg_type
, msg_control
);
561 static inline uint32_t
562 brw_dp_untyped_atomic_desc(const struct gen_device_info
*devinfo
,
563 unsigned exec_size
, /**< 0 for SIMD4x2 */
565 bool response_expected
)
567 assert(exec_size
<= 8 || exec_size
== 16);
570 if (devinfo
->gen
>= 8 || devinfo
->is_haswell
) {
572 msg_type
= HSW_DATAPORT_DC_PORT1_UNTYPED_ATOMIC_OP
;
574 msg_type
= HSW_DATAPORT_DC_PORT1_UNTYPED_ATOMIC_OP_SIMD4X2
;
577 msg_type
= GEN7_DATAPORT_DC_UNTYPED_ATOMIC_OP
;
580 const unsigned msg_control
=
581 SET_BITS(atomic_op
, 3, 0) |
582 SET_BITS(0 < exec_size
&& exec_size
<= 8, 4, 4) |
583 SET_BITS(response_expected
, 5, 5);
585 return brw_dp_surface_desc(devinfo
, msg_type
, msg_control
);
588 static inline uint32_t
589 brw_dp_untyped_atomic_float_desc(const struct gen_device_info
*devinfo
,
592 bool response_expected
)
594 assert(exec_size
<= 8 || exec_size
== 16);
595 assert(devinfo
->gen
>= 9);
597 assert(exec_size
> 0);
598 const unsigned msg_type
= GEN9_DATAPORT_DC_PORT1_UNTYPED_ATOMIC_FLOAT_OP
;
600 const unsigned msg_control
=
601 SET_BITS(atomic_op
, 1, 0) |
602 SET_BITS(exec_size
<= 8, 4, 4) |
603 SET_BITS(response_expected
, 5, 5);
605 return brw_dp_surface_desc(devinfo
, msg_type
, msg_control
);
608 static inline unsigned
609 brw_mdc_cmask(unsigned num_channels
)
611 /* See also MDC_CMASK in the SKL PRM Vol 2d. */
612 return 0xf & (0xf << num_channels
);
615 static inline uint32_t
616 brw_dp_untyped_surface_rw_desc(const struct gen_device_info
*devinfo
,
617 unsigned exec_size
, /**< 0 for SIMD4x2 */
618 unsigned num_channels
,
621 assert(exec_size
<= 8 || exec_size
== 16);
625 if (devinfo
->gen
>= 8 || devinfo
->is_haswell
) {
626 msg_type
= HSW_DATAPORT_DC_PORT1_UNTYPED_SURFACE_WRITE
;
628 msg_type
= GEN7_DATAPORT_DC_UNTYPED_SURFACE_WRITE
;
632 if (devinfo
->gen
>= 8 || devinfo
->is_haswell
) {
633 msg_type
= HSW_DATAPORT_DC_PORT1_UNTYPED_SURFACE_READ
;
635 msg_type
= GEN7_DATAPORT_DC_UNTYPED_SURFACE_READ
;
639 /* SIMD4x2 is only valid for read messages on IVB; use SIMD8 instead */
640 if (write
&& devinfo
->gen
== 7 && !devinfo
->is_haswell
&& exec_size
== 0)
643 /* See also MDC_SM3 in the SKL PRM Vol 2d. */
644 const unsigned simd_mode
= exec_size
== 0 ? 0 : /* SIMD4x2 */
645 exec_size
<= 8 ? 2 : 1;
647 const unsigned msg_control
=
648 SET_BITS(brw_mdc_cmask(num_channels
), 3, 0) |
649 SET_BITS(simd_mode
, 5, 4);
651 return brw_dp_surface_desc(devinfo
, msg_type
, msg_control
);
654 static inline unsigned
655 brw_mdc_ds(unsigned bit_size
)
659 return GEN7_BYTE_SCATTERED_DATA_ELEMENT_BYTE
;
661 return GEN7_BYTE_SCATTERED_DATA_ELEMENT_WORD
;
663 return GEN7_BYTE_SCATTERED_DATA_ELEMENT_DWORD
;
665 unreachable("Unsupported bit_size for byte scattered messages");
669 static inline uint32_t
670 brw_dp_byte_scattered_rw_desc(const struct gen_device_info
*devinfo
,
675 assert(exec_size
<= 8 || exec_size
== 16);
677 assert(devinfo
->gen
> 7 || devinfo
->is_haswell
);
678 const unsigned msg_type
=
679 write
? HSW_DATAPORT_DC_PORT0_BYTE_SCATTERED_WRITE
:
680 HSW_DATAPORT_DC_PORT0_BYTE_SCATTERED_READ
;
682 assert(exec_size
> 0);
683 const unsigned msg_control
=
684 SET_BITS(exec_size
== 16, 0, 0) |
685 SET_BITS(brw_mdc_ds(bit_size
), 3, 2);
687 return brw_dp_surface_desc(devinfo
, msg_type
, msg_control
);
690 static inline uint32_t
691 brw_dp_a64_untyped_surface_rw_desc(const struct gen_device_info
*devinfo
,
692 unsigned exec_size
, /**< 0 for SIMD4x2 */
693 unsigned num_channels
,
696 assert(exec_size
<= 8 || exec_size
== 16);
697 assert(devinfo
->gen
>= 8);
700 write
? GEN8_DATAPORT_DC_PORT1_A64_UNTYPED_SURFACE_WRITE
:
701 GEN8_DATAPORT_DC_PORT1_A64_UNTYPED_SURFACE_READ
;
703 /* See also MDC_SM3 in the SKL PRM Vol 2d. */
704 const unsigned simd_mode
= exec_size
== 0 ? 0 : /* SIMD4x2 */
705 exec_size
<= 8 ? 2 : 1;
707 const unsigned msg_control
=
708 SET_BITS(brw_mdc_cmask(num_channels
), 3, 0) |
709 SET_BITS(simd_mode
, 5, 4);
711 return brw_dp_desc(devinfo
, BRW_BTI_STATELESS
, msg_type
, msg_control
);
715 * Calculate the data size (see MDC_A64_DS in the "Structures" volume of the
718 static inline uint32_t
719 brw_mdc_a64_ds(unsigned elems
)
727 unreachable("Unsupported elmeent count for A64 scattered message");
731 static inline uint32_t
732 brw_dp_a64_byte_scattered_rw_desc(const struct gen_device_info
*devinfo
,
733 unsigned exec_size
, /**< 0 for SIMD4x2 */
737 assert(exec_size
<= 8 || exec_size
== 16);
738 assert(devinfo
->gen
>= 8);
741 write
? GEN8_DATAPORT_DC_PORT1_A64_SCATTERED_WRITE
:
742 GEN9_DATAPORT_DC_PORT1_A64_SCATTERED_READ
;
744 const unsigned msg_control
=
745 SET_BITS(GEN8_A64_SCATTERED_SUBTYPE_BYTE
, 1, 0) |
746 SET_BITS(brw_mdc_a64_ds(bit_size
/ 8), 3, 2) |
747 SET_BITS(exec_size
== 16, 4, 4);
749 return brw_dp_desc(devinfo
, BRW_BTI_STATELESS
, msg_type
, msg_control
);
752 static inline uint32_t
753 brw_dp_a64_untyped_atomic_desc(const struct gen_device_info
*devinfo
,
754 unsigned exec_size
, /**< 0 for SIMD4x2 */
757 bool response_expected
)
759 assert(exec_size
== 8);
760 assert(devinfo
->gen
>= 8);
761 assert(bit_size
== 32 || bit_size
== 64);
763 const unsigned msg_type
= GEN8_DATAPORT_DC_PORT1_A64_UNTYPED_ATOMIC_OP
;
765 const unsigned msg_control
=
766 SET_BITS(atomic_op
, 3, 0) |
767 SET_BITS(bit_size
== 64, 4, 4) |
768 SET_BITS(response_expected
, 5, 5);
770 return brw_dp_desc(devinfo
, BRW_BTI_STATELESS
, msg_type
, msg_control
);
773 static inline uint32_t
774 brw_dp_a64_untyped_atomic_float_desc(const struct gen_device_info
*devinfo
,
777 bool response_expected
)
779 assert(exec_size
== 8);
780 assert(devinfo
->gen
>= 9);
782 assert(exec_size
> 0);
783 const unsigned msg_type
= GEN9_DATAPORT_DC_PORT1_A64_UNTYPED_ATOMIC_FLOAT_OP
;
785 const unsigned msg_control
=
786 SET_BITS(atomic_op
, 1, 0) |
787 SET_BITS(response_expected
, 5, 5);
789 return brw_dp_desc(devinfo
, BRW_BTI_STATELESS
, msg_type
, msg_control
);
792 static inline uint32_t
793 brw_dp_typed_atomic_desc(const struct gen_device_info
*devinfo
,
797 bool response_expected
)
799 assert(exec_size
> 0 || exec_group
== 0);
800 assert(exec_group
% 8 == 0);
803 if (devinfo
->gen
>= 8 || devinfo
->is_haswell
) {
804 if (exec_size
== 0) {
805 msg_type
= HSW_DATAPORT_DC_PORT1_TYPED_ATOMIC_OP_SIMD4X2
;
807 msg_type
= HSW_DATAPORT_DC_PORT1_TYPED_ATOMIC_OP
;
810 /* SIMD4x2 typed surface R/W messages only exist on HSW+ */
811 assert(exec_size
> 0);
812 msg_type
= GEN7_DATAPORT_RC_TYPED_ATOMIC_OP
;
815 const bool high_sample_mask
= (exec_group
/ 8) % 2 == 1;
817 const unsigned msg_control
=
818 SET_BITS(atomic_op
, 3, 0) |
819 SET_BITS(high_sample_mask
, 4, 4) |
820 SET_BITS(response_expected
, 5, 5);
822 return brw_dp_surface_desc(devinfo
, msg_type
, msg_control
);
825 static inline uint32_t
826 brw_dp_typed_surface_rw_desc(const struct gen_device_info
*devinfo
,
829 unsigned num_channels
,
832 assert(exec_size
> 0 || exec_group
== 0);
833 assert(exec_group
% 8 == 0);
835 /* Typed surface reads and writes don't support SIMD16 */
836 assert(exec_size
<= 8);
840 if (devinfo
->gen
>= 8 || devinfo
->is_haswell
) {
841 msg_type
= HSW_DATAPORT_DC_PORT1_TYPED_SURFACE_WRITE
;
843 msg_type
= GEN7_DATAPORT_RC_TYPED_SURFACE_WRITE
;
846 if (devinfo
->gen
>= 8 || devinfo
->is_haswell
) {
847 msg_type
= HSW_DATAPORT_DC_PORT1_TYPED_SURFACE_READ
;
849 msg_type
= GEN7_DATAPORT_RC_TYPED_SURFACE_READ
;
853 /* See also MDC_SG3 in the SKL PRM Vol 2d. */
854 unsigned msg_control
;
855 if (devinfo
->gen
>= 8 || devinfo
->is_haswell
) {
856 /* See also MDC_SG3 in the SKL PRM Vol 2d. */
857 const unsigned slot_group
= exec_size
== 0 ? 0 : /* SIMD4x2 */
858 1 + ((exec_group
/ 8) % 2);
861 SET_BITS(brw_mdc_cmask(num_channels
), 3, 0) |
862 SET_BITS(slot_group
, 5, 4);
864 /* SIMD4x2 typed surface R/W messages only exist on HSW+ */
865 assert(exec_size
> 0);
866 const unsigned slot_group
= ((exec_group
/ 8) % 2);
869 SET_BITS(brw_mdc_cmask(num_channels
), 3, 0) |
870 SET_BITS(slot_group
, 5, 5);
873 return brw_dp_surface_desc(devinfo
, msg_type
, msg_control
);
877 * Construct a message descriptor immediate with the specified pixel
878 * interpolator function controls.
880 static inline uint32_t
881 brw_pixel_interp_desc(UNUSED
const struct gen_device_info
*devinfo
,
887 return (SET_BITS(slot_group
, 11, 11) |
888 SET_BITS(msg_type
, 13, 12) |
889 SET_BITS(!!noperspective
, 14, 14) |
890 SET_BITS(simd_mode
, 16, 16));
893 void brw_urb_WRITE(struct brw_codegen
*p
,
897 enum brw_urb_write_flags flags
,
899 unsigned response_length
,
904 * Send message to shared unit \p sfid with a possibly indirect descriptor \p
905 * desc. If \p desc is not an immediate it will be transparently loaded to an
906 * address register using an OR instruction.
909 brw_send_indirect_message(struct brw_codegen
*p
,
912 struct brw_reg payload
,
918 brw_send_indirect_split_message(struct brw_codegen
*p
,
921 struct brw_reg payload0
,
922 struct brw_reg payload1
,
925 struct brw_reg ex_desc
,
926 unsigned ex_desc_imm
,
929 void brw_ff_sync(struct brw_codegen
*p
,
934 unsigned response_length
,
937 void brw_svb_write(struct brw_codegen
*p
,
941 unsigned binding_table_index
,
942 bool send_commit_msg
);
944 brw_inst
*brw_fb_WRITE(struct brw_codegen
*p
,
945 struct brw_reg payload
,
946 struct brw_reg implied_header
,
947 unsigned msg_control
,
948 unsigned binding_table_index
,
950 unsigned response_length
,
952 bool last_render_target
,
953 bool header_present
);
955 brw_inst
*gen9_fb_READ(struct brw_codegen
*p
,
957 struct brw_reg payload
,
958 unsigned binding_table_index
,
960 unsigned response_length
,
963 void brw_SAMPLE(struct brw_codegen
*p
,
967 unsigned binding_table_index
,
970 unsigned response_length
,
972 unsigned header_present
,
974 unsigned return_format
);
976 void brw_adjust_sampler_state_pointer(struct brw_codegen
*p
,
977 struct brw_reg header
,
978 struct brw_reg sampler_index
);
980 void gen4_math(struct brw_codegen
*p
,
985 unsigned precision
);
987 void gen6_math(struct brw_codegen
*p
,
991 struct brw_reg src1
);
993 void brw_oword_block_read(struct brw_codegen
*p
,
997 uint32_t bind_table_index
);
999 unsigned brw_scratch_surface_idx(const struct brw_codegen
*p
);
1001 void brw_oword_block_read_scratch(struct brw_codegen
*p
,
1002 struct brw_reg dest
,
1007 void brw_oword_block_write_scratch(struct brw_codegen
*p
,
1012 void gen7_block_read_scratch(struct brw_codegen
*p
,
1013 struct brw_reg dest
,
1017 void brw_shader_time_add(struct brw_codegen
*p
,
1018 struct brw_reg payload
,
1019 uint32_t surf_index
);
1022 * Return the generation-specific jump distance scaling factor.
1024 * Given the number of instructions to jump, we need to scale by
1025 * some number to obtain the actual jump distance to program in an
1028 static inline unsigned
1029 brw_jump_scale(const struct gen_device_info
*devinfo
)
1031 /* Broadwell measures jump targets in bytes. */
1032 if (devinfo
->gen
>= 8)
1035 /* Ironlake and later measure jump targets in 64-bit data chunks (in order
1036 * (to support compaction), so each 128-bit instruction requires 2 chunks.
1038 if (devinfo
->gen
>= 5)
1041 /* Gen4 simply uses the number of 128-bit instructions. */
1045 void brw_barrier(struct brw_codegen
*p
, struct brw_reg src
);
1047 /* If/else/endif. Works by manipulating the execution flags on each
1050 brw_inst
*brw_IF(struct brw_codegen
*p
, unsigned execute_size
);
1051 brw_inst
*gen6_IF(struct brw_codegen
*p
, enum brw_conditional_mod conditional
,
1052 struct brw_reg src0
, struct brw_reg src1
);
1054 void brw_ELSE(struct brw_codegen
*p
);
1055 void brw_ENDIF(struct brw_codegen
*p
);
1059 brw_inst
*brw_DO(struct brw_codegen
*p
, unsigned execute_size
);
1061 brw_inst
*brw_WHILE(struct brw_codegen
*p
);
1063 brw_inst
*brw_BREAK(struct brw_codegen
*p
);
1064 brw_inst
*brw_CONT(struct brw_codegen
*p
);
1065 brw_inst
*gen6_HALT(struct brw_codegen
*p
);
1069 void brw_land_fwd_jump(struct brw_codegen
*p
, int jmp_insn_idx
);
1071 brw_inst
*brw_JMPI(struct brw_codegen
*p
, struct brw_reg index
,
1072 unsigned predicate_control
);
1074 void brw_NOP(struct brw_codegen
*p
);
1076 void brw_WAIT(struct brw_codegen
*p
);
1078 /* Special case: there is never a destination, execution size will be
1081 void brw_CMP(struct brw_codegen
*p
,
1082 struct brw_reg dest
,
1083 unsigned conditional
,
1084 struct brw_reg src0
,
1085 struct brw_reg src1
);
1088 brw_untyped_atomic(struct brw_codegen
*p
,
1090 struct brw_reg payload
,
1091 struct brw_reg surface
,
1093 unsigned msg_length
,
1094 bool response_expected
,
1095 bool header_present
);
1098 brw_untyped_surface_read(struct brw_codegen
*p
,
1100 struct brw_reg payload
,
1101 struct brw_reg surface
,
1102 unsigned msg_length
,
1103 unsigned num_channels
);
1106 brw_untyped_surface_write(struct brw_codegen
*p
,
1107 struct brw_reg payload
,
1108 struct brw_reg surface
,
1109 unsigned msg_length
,
1110 unsigned num_channels
,
1111 bool header_present
);
1114 brw_memory_fence(struct brw_codegen
*p
,
1116 enum opcode send_op
);
1119 brw_pixel_interpolator_query(struct brw_codegen
*p
,
1120 struct brw_reg dest
,
1124 struct brw_reg data
,
1125 unsigned msg_length
,
1126 unsigned response_length
);
1129 brw_find_live_channel(struct brw_codegen
*p
,
1131 struct brw_reg mask
);
1134 brw_broadcast(struct brw_codegen
*p
,
1137 struct brw_reg idx
);
1140 brw_rounding_mode(struct brw_codegen
*p
,
1141 enum brw_rnd_mode mode
);
1143 /***********************************************************************
1147 void brw_copy_indirect_to_indirect(struct brw_codegen
*p
,
1148 struct brw_indirect dst_ptr
,
1149 struct brw_indirect src_ptr
,
1152 void brw_copy_from_indirect(struct brw_codegen
*p
,
1154 struct brw_indirect ptr
,
1157 void brw_copy4(struct brw_codegen
*p
,
1162 void brw_copy8(struct brw_codegen
*p
,
1167 void brw_math_invert( struct brw_codegen
*p
,
1169 struct brw_reg src
);
1171 void brw_set_src1(struct brw_codegen
*p
, brw_inst
*insn
, struct brw_reg reg
);
1173 void brw_set_desc_ex(struct brw_codegen
*p
, brw_inst
*insn
,
1174 unsigned desc
, unsigned ex_desc
);
1177 brw_set_desc(struct brw_codegen
*p
, brw_inst
*insn
, unsigned desc
)
1179 brw_set_desc_ex(p
, insn
, desc
, 0);
1182 void brw_set_uip_jip(struct brw_codegen
*p
, int start_offset
);
1184 enum brw_conditional_mod
brw_negate_cmod(uint32_t cmod
);
1185 enum brw_conditional_mod
brw_swap_cmod(uint32_t cmod
);
1187 /* brw_eu_compact.c */
1188 void brw_init_compaction_tables(const struct gen_device_info
*devinfo
);
1189 void brw_compact_instructions(struct brw_codegen
*p
, int start_offset
,
1190 struct disasm_info
*disasm
);
1191 void brw_uncompact_instruction(const struct gen_device_info
*devinfo
,
1192 brw_inst
*dst
, brw_compact_inst
*src
);
1193 bool brw_try_compact_instruction(const struct gen_device_info
*devinfo
,
1194 brw_compact_inst
*dst
, const brw_inst
*src
);
1196 void brw_debug_compact_uncompact(const struct gen_device_info
*devinfo
,
1197 brw_inst
*orig
, brw_inst
*uncompacted
);
1199 /* brw_eu_validate.c */
1200 bool brw_validate_instructions(const struct gen_device_info
*devinfo
,
1201 const void *assembly
, int start_offset
, int end_offset
,
1202 struct disasm_info
*disasm
);
1205 next_offset(const struct gen_device_info
*devinfo
, void *store
, int offset
)
1207 brw_inst
*insn
= (brw_inst
*)((char *)store
+ offset
);
1209 if (brw_inst_cmpt_control(devinfo
, insn
))
1215 struct opcode_desc
{
1216 /* The union is an implementation detail used by brw_opcode_desc() to handle
1217 * opcodes that have been reused for different instructions across hardware
1220 * The gens field acts as a tag. If it is non-zero, name points to a string
1221 * containing the instruction mnemonic. If it is zero, the table field is
1222 * valid and either points to a secondary opcode_desc table with 'size'
1223 * elements or is NULL and no such instruction exists for the opcode.
1231 const struct opcode_desc
*table
;
1239 const struct opcode_desc
*
1240 brw_opcode_desc(const struct gen_device_info
*devinfo
, enum opcode opcode
);
1243 is_3src(const struct gen_device_info
*devinfo
, enum opcode opcode
)
1245 const struct opcode_desc
*desc
= brw_opcode_desc(devinfo
, opcode
);
1246 return desc
&& desc
->nsrc
== 3;
1249 /** Maximum SEND message length */
1250 #define BRW_MAX_MSG_LENGTH 15
1252 /** First MRF register used by pull loads */
1253 #define FIRST_SPILL_MRF(gen) ((gen) == 6 ? 21 : 13)
1255 /** First MRF register used by spills */
1256 #define FIRST_PULL_LOAD_MRF(gen) ((gen) == 6 ? 16 : 13)