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:
248 void brw_set_dp_read_message(struct brw_codegen
*p
,
250 unsigned binding_table_index
,
251 unsigned msg_control
,
253 unsigned target_cache
,
256 unsigned response_length
);
258 void brw_set_dp_write_message(struct brw_codegen
*p
,
260 unsigned binding_table_index
,
261 unsigned msg_control
,
263 unsigned target_cache
,
266 unsigned last_render_target
,
267 unsigned response_length
,
268 unsigned end_of_thread
,
269 unsigned send_commit_msg
);
272 * Construct a message descriptor immediate with the specified common
273 * descriptor controls.
275 static inline uint32_t
276 brw_message_desc(const struct gen_device_info
*devinfo
,
278 unsigned response_length
,
281 if (devinfo
->gen
>= 5) {
282 return (SET_BITS(msg_length
, 28, 25) |
283 SET_BITS(response_length
, 24, 20) |
284 SET_BITS(header_present
, 19, 19));
286 return (SET_BITS(msg_length
, 23, 20) |
287 SET_BITS(response_length
, 19, 16));
292 * Construct a message descriptor immediate with the specified sampler
295 static inline uint32_t
296 brw_sampler_desc(const struct gen_device_info
*devinfo
,
297 unsigned binding_table_index
,
301 unsigned return_format
)
303 const unsigned desc
= (SET_BITS(binding_table_index
, 7, 0) |
304 SET_BITS(sampler
, 11, 8));
305 if (devinfo
->gen
>= 7)
306 return (desc
| SET_BITS(msg_type
, 16, 12) |
307 SET_BITS(simd_mode
, 18, 17));
308 else if (devinfo
->gen
>= 5)
309 return (desc
| SET_BITS(msg_type
, 15, 12) |
310 SET_BITS(simd_mode
, 17, 16));
311 else if (devinfo
->is_g4x
)
312 return desc
| SET_BITS(msg_type
, 15, 12);
314 return (desc
| SET_BITS(return_format
, 13, 12) |
315 SET_BITS(msg_type
, 15, 14));
318 void brw_urb_WRITE(struct brw_codegen
*p
,
322 enum brw_urb_write_flags flags
,
324 unsigned response_length
,
329 * Send message to shared unit \p sfid with a possibly indirect descriptor \p
330 * desc. If \p desc is not an immediate it will be transparently loaded to an
331 * address register using an OR instruction. The returned instruction can be
332 * passed as argument to the usual brw_set_*_message() functions in order to
333 * specify any additional descriptor bits -- If \p desc is an immediate this
334 * will be the SEND instruction itself, otherwise it will be the OR
338 brw_send_indirect_message(struct brw_codegen
*p
,
341 struct brw_reg payload
,
345 void brw_ff_sync(struct brw_codegen
*p
,
350 unsigned response_length
,
353 void brw_svb_write(struct brw_codegen
*p
,
357 unsigned binding_table_index
,
358 bool send_commit_msg
);
360 brw_inst
*brw_fb_WRITE(struct brw_codegen
*p
,
361 struct brw_reg payload
,
362 struct brw_reg implied_header
,
363 unsigned msg_control
,
364 unsigned binding_table_index
,
366 unsigned response_length
,
368 bool last_render_target
,
369 bool header_present
);
371 brw_inst
*gen9_fb_READ(struct brw_codegen
*p
,
373 struct brw_reg payload
,
374 unsigned binding_table_index
,
376 unsigned response_length
,
379 void brw_SAMPLE(struct brw_codegen
*p
,
383 unsigned binding_table_index
,
386 unsigned response_length
,
388 unsigned header_present
,
390 unsigned return_format
);
392 void brw_adjust_sampler_state_pointer(struct brw_codegen
*p
,
393 struct brw_reg header
,
394 struct brw_reg sampler_index
);
396 void gen4_math(struct brw_codegen
*p
,
401 unsigned precision
);
403 void gen6_math(struct brw_codegen
*p
,
407 struct brw_reg src1
);
409 void brw_oword_block_read(struct brw_codegen
*p
,
413 uint32_t bind_table_index
);
415 unsigned brw_scratch_surface_idx(const struct brw_codegen
*p
);
417 void brw_oword_block_read_scratch(struct brw_codegen
*p
,
423 void brw_oword_block_write_scratch(struct brw_codegen
*p
,
428 void gen7_block_read_scratch(struct brw_codegen
*p
,
433 void brw_shader_time_add(struct brw_codegen
*p
,
434 struct brw_reg payload
,
435 uint32_t surf_index
);
438 * Return the generation-specific jump distance scaling factor.
440 * Given the number of instructions to jump, we need to scale by
441 * some number to obtain the actual jump distance to program in an
444 static inline unsigned
445 brw_jump_scale(const struct gen_device_info
*devinfo
)
447 /* Broadwell measures jump targets in bytes. */
448 if (devinfo
->gen
>= 8)
451 /* Ironlake and later measure jump targets in 64-bit data chunks (in order
452 * (to support compaction), so each 128-bit instruction requires 2 chunks.
454 if (devinfo
->gen
>= 5)
457 /* Gen4 simply uses the number of 128-bit instructions. */
461 void brw_barrier(struct brw_codegen
*p
, struct brw_reg src
);
463 /* If/else/endif. Works by manipulating the execution flags on each
466 brw_inst
*brw_IF(struct brw_codegen
*p
, unsigned execute_size
);
467 brw_inst
*gen6_IF(struct brw_codegen
*p
, enum brw_conditional_mod conditional
,
468 struct brw_reg src0
, struct brw_reg src1
);
470 void brw_ELSE(struct brw_codegen
*p
);
471 void brw_ENDIF(struct brw_codegen
*p
);
475 brw_inst
*brw_DO(struct brw_codegen
*p
, unsigned execute_size
);
477 brw_inst
*brw_WHILE(struct brw_codegen
*p
);
479 brw_inst
*brw_BREAK(struct brw_codegen
*p
);
480 brw_inst
*brw_CONT(struct brw_codegen
*p
);
481 brw_inst
*gen6_HALT(struct brw_codegen
*p
);
485 void brw_land_fwd_jump(struct brw_codegen
*p
, int jmp_insn_idx
);
487 brw_inst
*brw_JMPI(struct brw_codegen
*p
, struct brw_reg index
,
488 unsigned predicate_control
);
490 void brw_NOP(struct brw_codegen
*p
);
492 void brw_WAIT(struct brw_codegen
*p
);
494 /* Special case: there is never a destination, execution size will be
497 void brw_CMP(struct brw_codegen
*p
,
499 unsigned conditional
,
501 struct brw_reg src1
);
504 brw_untyped_atomic(struct brw_codegen
*p
,
506 struct brw_reg payload
,
507 struct brw_reg surface
,
510 bool response_expected
,
511 bool header_present
);
514 brw_untyped_surface_read(struct brw_codegen
*p
,
516 struct brw_reg payload
,
517 struct brw_reg surface
,
519 unsigned num_channels
);
522 brw_untyped_surface_write(struct brw_codegen
*p
,
523 struct brw_reg payload
,
524 struct brw_reg surface
,
526 unsigned num_channels
,
527 bool header_present
);
530 brw_typed_atomic(struct brw_codegen
*p
,
532 struct brw_reg payload
,
533 struct brw_reg surface
,
536 bool response_expected
,
537 bool header_present
);
540 brw_typed_surface_read(struct brw_codegen
*p
,
542 struct brw_reg payload
,
543 struct brw_reg surface
,
545 unsigned num_channels
,
546 bool header_present
);
549 brw_typed_surface_write(struct brw_codegen
*p
,
550 struct brw_reg payload
,
551 struct brw_reg surface
,
553 unsigned num_channels
,
554 bool header_present
);
557 brw_byte_scattered_read(struct brw_codegen
*p
,
559 struct brw_reg payload
,
560 struct brw_reg surface
,
565 brw_byte_scattered_write(struct brw_codegen
*p
,
566 struct brw_reg payload
,
567 struct brw_reg surface
,
570 bool header_present
);
573 brw_memory_fence(struct brw_codegen
*p
,
575 enum opcode send_op
);
578 brw_pixel_interpolator_query(struct brw_codegen
*p
,
585 unsigned response_length
);
588 brw_find_live_channel(struct brw_codegen
*p
,
590 struct brw_reg mask
);
593 brw_broadcast(struct brw_codegen
*p
,
599 brw_rounding_mode(struct brw_codegen
*p
,
600 enum brw_rnd_mode mode
);
602 /***********************************************************************
606 void brw_copy_indirect_to_indirect(struct brw_codegen
*p
,
607 struct brw_indirect dst_ptr
,
608 struct brw_indirect src_ptr
,
611 void brw_copy_from_indirect(struct brw_codegen
*p
,
613 struct brw_indirect ptr
,
616 void brw_copy4(struct brw_codegen
*p
,
621 void brw_copy8(struct brw_codegen
*p
,
626 void brw_math_invert( struct brw_codegen
*p
,
630 void brw_set_src1(struct brw_codegen
*p
, brw_inst
*insn
, struct brw_reg reg
);
632 void brw_set_desc_ex(struct brw_codegen
*p
, brw_inst
*insn
,
633 unsigned desc
, unsigned ex_desc
);
636 brw_set_desc(struct brw_codegen
*p
, brw_inst
*insn
, unsigned desc
)
638 brw_set_desc_ex(p
, insn
, desc
, 0);
641 void brw_set_uip_jip(struct brw_codegen
*p
, int start_offset
);
643 enum brw_conditional_mod
brw_negate_cmod(uint32_t cmod
);
644 enum brw_conditional_mod
brw_swap_cmod(uint32_t cmod
);
646 /* brw_eu_compact.c */
647 void brw_init_compaction_tables(const struct gen_device_info
*devinfo
);
648 void brw_compact_instructions(struct brw_codegen
*p
, int start_offset
,
649 struct disasm_info
*disasm
);
650 void brw_uncompact_instruction(const struct gen_device_info
*devinfo
,
651 brw_inst
*dst
, brw_compact_inst
*src
);
652 bool brw_try_compact_instruction(const struct gen_device_info
*devinfo
,
653 brw_compact_inst
*dst
, const brw_inst
*src
);
655 void brw_debug_compact_uncompact(const struct gen_device_info
*devinfo
,
656 brw_inst
*orig
, brw_inst
*uncompacted
);
658 /* brw_eu_validate.c */
659 bool brw_validate_instructions(const struct gen_device_info
*devinfo
,
660 const void *assembly
, int start_offset
, int end_offset
,
661 struct disasm_info
*disasm
);
664 next_offset(const struct gen_device_info
*devinfo
, void *store
, int offset
)
666 brw_inst
*insn
= (brw_inst
*)((char *)store
+ offset
);
668 if (brw_inst_cmpt_control(devinfo
, insn
))
675 /* The union is an implementation detail used by brw_opcode_desc() to handle
676 * opcodes that have been reused for different instructions across hardware
679 * The gens field acts as a tag. If it is non-zero, name points to a string
680 * containing the instruction mnemonic. If it is zero, the table field is
681 * valid and either points to a secondary opcode_desc table with 'size'
682 * elements or is NULL and no such instruction exists for the opcode.
690 const struct opcode_desc
*table
;
698 const struct opcode_desc
*
699 brw_opcode_desc(const struct gen_device_info
*devinfo
, enum opcode opcode
);
702 is_3src(const struct gen_device_info
*devinfo
, enum opcode opcode
)
704 const struct opcode_desc
*desc
= brw_opcode_desc(devinfo
, opcode
);
705 return desc
&& desc
->nsrc
== 3;
708 /** Maximum SEND message length */
709 #define BRW_MAX_MSG_LENGTH 15
711 /** First MRF register used by pull loads */
712 #define FIRST_SPILL_MRF(gen) ((gen) == 6 ? 21 : 13)
714 /** First MRF register used by spills */
715 #define FIRST_PULL_LOAD_MRF(gen) ((gen) == 6 ? 16 : 13)