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>
38 #include "brw_eu_defines.h"
40 #include "intel_asm_annotation.h"
46 #define BRW_EU_MAX_INSN_STACK 5
48 /* A helper for accessing the last instruction emitted. This makes it easy
49 * to set various bits on an instruction without having to create temporary
50 * variable and assign the emitted instruction to those.
52 #define brw_last_inst (&p->store[p->nr_insn - 1])
58 unsigned int next_insn_offset
;
62 /* Allow clients to push/pop instruction state:
64 brw_inst stack
[BRW_EU_MAX_INSN_STACK
];
65 bool compressed_stack
[BRW_EU_MAX_INSN_STACK
];
68 bool single_program_flow
;
69 const struct gen_device_info
*devinfo
;
71 /* Control flow stacks:
72 * - if_stack contains IF and ELSE instructions which must be patched
73 * (and popped) once the matching ENDIF instruction is encountered.
75 * Just store the instruction pointer(an index).
79 int if_stack_array_size
;
82 * loop_stack contains the instruction pointers of the starts of loops which
83 * must be patched (and popped) once the matching WHILE instruction is
88 * pre-gen6, the BREAK and CONT instructions had to tell how many IF/ENDIF
89 * blocks they were popping out of, to fix up the mask stack. This tracks
90 * the IF/ENDIF nesting in each current nested loop level.
92 int *if_depth_in_loop
;
94 int loop_stack_array_size
;
97 void brw_pop_insn_state( struct brw_codegen
*p
);
98 void brw_push_insn_state( struct brw_codegen
*p
);
99 void brw_set_default_exec_size(struct brw_codegen
*p
, unsigned value
);
100 void brw_set_default_mask_control( struct brw_codegen
*p
, unsigned value
);
101 void brw_set_default_saturate( struct brw_codegen
*p
, bool enable
);
102 void brw_set_default_access_mode( struct brw_codegen
*p
, unsigned access_mode
);
103 void brw_inst_set_compression(const struct gen_device_info
*devinfo
,
104 brw_inst
*inst
, bool on
);
105 void brw_set_default_compression(struct brw_codegen
*p
, bool on
);
106 void brw_inst_set_group(const struct gen_device_info
*devinfo
,
107 brw_inst
*inst
, unsigned group
);
108 void brw_set_default_group(struct brw_codegen
*p
, unsigned group
);
109 void brw_set_default_compression_control(struct brw_codegen
*p
, enum brw_compression c
);
110 void brw_set_default_predicate_control( struct brw_codegen
*p
, unsigned pc
);
111 void brw_set_default_predicate_inverse(struct brw_codegen
*p
, bool predicate_inverse
);
112 void brw_set_default_flag_reg(struct brw_codegen
*p
, int reg
, int subreg
);
113 void brw_set_default_acc_write_control(struct brw_codegen
*p
, unsigned value
);
115 void brw_init_codegen(const struct gen_device_info
*, struct brw_codegen
*p
,
117 int brw_disassemble_inst(FILE *file
, const struct gen_device_info
*devinfo
,
118 struct brw_inst
*inst
, bool is_compacted
);
119 void brw_disassemble(const struct gen_device_info
*devinfo
, void *assembly
,
120 int start
, int end
, FILE *out
);
121 const unsigned *brw_get_program( struct brw_codegen
*p
, unsigned *sz
);
123 brw_inst
*brw_next_insn(struct brw_codegen
*p
, unsigned opcode
);
124 void brw_set_dest(struct brw_codegen
*p
, brw_inst
*insn
, struct brw_reg dest
);
125 void brw_set_src0(struct brw_codegen
*p
, brw_inst
*insn
, struct brw_reg reg
);
127 void gen6_resolve_implied_move(struct brw_codegen
*p
,
129 unsigned msg_reg_nr
);
131 /* Helpers for regular instructions:
134 brw_inst *brw_##OP(struct brw_codegen *p, \
135 struct brw_reg dest, \
136 struct brw_reg src0);
139 brw_inst *brw_##OP(struct brw_codegen *p, \
140 struct brw_reg dest, \
141 struct brw_reg src0, \
142 struct brw_reg src1);
145 brw_inst *brw_##OP(struct brw_codegen *p, \
146 struct brw_reg dest, \
147 struct brw_reg src0, \
148 struct brw_reg src1, \
149 struct brw_reg src2);
152 void brw_##OP(struct brw_codegen *p, struct brw_reg dest, struct brw_reg src0);
202 /* Helpers for SEND instruction:
204 void brw_set_sampler_message(struct brw_codegen
*p
,
206 unsigned binding_table_index
,
209 unsigned response_length
,
211 unsigned header_present
,
213 unsigned return_format
);
215 void brw_set_message_descriptor(struct brw_codegen
*p
,
217 enum brw_message_target sfid
,
219 unsigned response_length
,
223 void brw_set_dp_read_message(struct brw_codegen
*p
,
225 unsigned binding_table_index
,
226 unsigned msg_control
,
228 unsigned target_cache
,
231 unsigned response_length
);
233 void brw_set_dp_write_message(struct brw_codegen
*p
,
235 unsigned binding_table_index
,
236 unsigned msg_control
,
238 unsigned target_cache
,
241 unsigned last_render_target
,
242 unsigned response_length
,
243 unsigned end_of_thread
,
244 unsigned send_commit_msg
);
246 void brw_urb_WRITE(struct brw_codegen
*p
,
250 enum brw_urb_write_flags flags
,
252 unsigned response_length
,
257 * Send message to shared unit \p sfid with a possibly indirect descriptor \p
258 * desc. If \p desc is not an immediate it will be transparently loaded to an
259 * address register using an OR instruction. The returned instruction can be
260 * passed as argument to the usual brw_set_*_message() functions in order to
261 * specify any additional descriptor bits -- If \p desc is an immediate this
262 * will be the SEND instruction itself, otherwise it will be the OR
266 brw_send_indirect_message(struct brw_codegen
*p
,
269 struct brw_reg payload
,
270 struct brw_reg desc
);
272 void brw_ff_sync(struct brw_codegen
*p
,
277 unsigned response_length
,
280 void brw_svb_write(struct brw_codegen
*p
,
284 unsigned binding_table_index
,
285 bool send_commit_msg
);
287 void brw_fb_WRITE(struct brw_codegen
*p
,
288 struct brw_reg payload
,
289 struct brw_reg implied_header
,
290 unsigned msg_control
,
291 unsigned binding_table_index
,
293 unsigned response_length
,
295 bool last_render_target
,
296 bool header_present
);
298 brw_inst
*gen9_fb_READ(struct brw_codegen
*p
,
300 struct brw_reg payload
,
301 unsigned binding_table_index
,
303 unsigned response_length
,
306 void brw_SAMPLE(struct brw_codegen
*p
,
310 unsigned binding_table_index
,
313 unsigned response_length
,
315 unsigned header_present
,
317 unsigned return_format
);
319 void brw_adjust_sampler_state_pointer(struct brw_codegen
*p
,
320 struct brw_reg header
,
321 struct brw_reg sampler_index
);
323 void gen4_math(struct brw_codegen
*p
,
328 unsigned precision
);
330 void gen6_math(struct brw_codegen
*p
,
334 struct brw_reg src1
);
336 void brw_oword_block_read(struct brw_codegen
*p
,
340 uint32_t bind_table_index
);
342 unsigned brw_scratch_surface_idx(const struct brw_codegen
*p
);
344 void brw_oword_block_read_scratch(struct brw_codegen
*p
,
350 void brw_oword_block_write_scratch(struct brw_codegen
*p
,
355 void gen7_block_read_scratch(struct brw_codegen
*p
,
360 void brw_shader_time_add(struct brw_codegen
*p
,
361 struct brw_reg payload
,
362 uint32_t surf_index
);
365 * Return the generation-specific jump distance scaling factor.
367 * Given the number of instructions to jump, we need to scale by
368 * some number to obtain the actual jump distance to program in an
371 static inline unsigned
372 brw_jump_scale(const struct gen_device_info
*devinfo
)
374 /* Broadwell measures jump targets in bytes. */
375 if (devinfo
->gen
>= 8)
378 /* Ironlake and later measure jump targets in 64-bit data chunks (in order
379 * (to support compaction), so each 128-bit instruction requires 2 chunks.
381 if (devinfo
->gen
>= 5)
384 /* Gen4 simply uses the number of 128-bit instructions. */
388 void brw_barrier(struct brw_codegen
*p
, struct brw_reg src
);
390 /* If/else/endif. Works by manipulating the execution flags on each
393 brw_inst
*brw_IF(struct brw_codegen
*p
, unsigned execute_size
);
394 brw_inst
*gen6_IF(struct brw_codegen
*p
, enum brw_conditional_mod conditional
,
395 struct brw_reg src0
, struct brw_reg src1
);
397 void brw_ELSE(struct brw_codegen
*p
);
398 void brw_ENDIF(struct brw_codegen
*p
);
402 brw_inst
*brw_DO(struct brw_codegen
*p
, unsigned execute_size
);
404 brw_inst
*brw_WHILE(struct brw_codegen
*p
);
406 brw_inst
*brw_BREAK(struct brw_codegen
*p
);
407 brw_inst
*brw_CONT(struct brw_codegen
*p
);
408 brw_inst
*gen6_HALT(struct brw_codegen
*p
);
412 void brw_land_fwd_jump(struct brw_codegen
*p
, int jmp_insn_idx
);
414 brw_inst
*brw_JMPI(struct brw_codegen
*p
, struct brw_reg index
,
415 unsigned predicate_control
);
417 void brw_NOP(struct brw_codegen
*p
);
419 void brw_WAIT(struct brw_codegen
*p
);
421 /* Special case: there is never a destination, execution size will be
424 void brw_CMP(struct brw_codegen
*p
,
426 unsigned conditional
,
428 struct brw_reg src1
);
431 brw_untyped_atomic(struct brw_codegen
*p
,
433 struct brw_reg payload
,
434 struct brw_reg surface
,
437 bool response_expected
);
440 brw_untyped_surface_read(struct brw_codegen
*p
,
442 struct brw_reg payload
,
443 struct brw_reg surface
,
445 unsigned num_channels
);
448 brw_untyped_surface_write(struct brw_codegen
*p
,
449 struct brw_reg payload
,
450 struct brw_reg surface
,
452 unsigned num_channels
);
455 brw_typed_atomic(struct brw_codegen
*p
,
457 struct brw_reg payload
,
458 struct brw_reg surface
,
461 bool response_expected
);
464 brw_typed_surface_read(struct brw_codegen
*p
,
466 struct brw_reg payload
,
467 struct brw_reg surface
,
469 unsigned num_channels
);
472 brw_typed_surface_write(struct brw_codegen
*p
,
473 struct brw_reg payload
,
474 struct brw_reg surface
,
476 unsigned num_channels
);
479 brw_memory_fence(struct brw_codegen
*p
,
483 brw_pixel_interpolator_query(struct brw_codegen
*p
,
490 unsigned response_length
);
493 brw_find_live_channel(struct brw_codegen
*p
,
495 struct brw_reg mask
);
498 brw_broadcast(struct brw_codegen
*p
,
503 /***********************************************************************
507 void brw_copy_indirect_to_indirect(struct brw_codegen
*p
,
508 struct brw_indirect dst_ptr
,
509 struct brw_indirect src_ptr
,
512 void brw_copy_from_indirect(struct brw_codegen
*p
,
514 struct brw_indirect ptr
,
517 void brw_copy4(struct brw_codegen
*p
,
522 void brw_copy8(struct brw_codegen
*p
,
527 void brw_math_invert( struct brw_codegen
*p
,
531 void brw_set_src1(struct brw_codegen
*p
, brw_inst
*insn
, struct brw_reg reg
);
533 void brw_set_uip_jip(struct brw_codegen
*p
, int start_offset
);
535 enum brw_conditional_mod
brw_negate_cmod(uint32_t cmod
);
536 enum brw_conditional_mod
brw_swap_cmod(uint32_t cmod
);
538 /* brw_eu_compact.c */
539 void brw_init_compaction_tables(const struct gen_device_info
*devinfo
);
540 void brw_compact_instructions(struct brw_codegen
*p
, int start_offset
,
541 int num_annotations
, struct annotation
*annotation
);
542 void brw_uncompact_instruction(const struct gen_device_info
*devinfo
,
543 brw_inst
*dst
, brw_compact_inst
*src
);
544 bool brw_try_compact_instruction(const struct gen_device_info
*devinfo
,
545 brw_compact_inst
*dst
, brw_inst
*src
);
547 void brw_debug_compact_uncompact(const struct gen_device_info
*devinfo
,
548 brw_inst
*orig
, brw_inst
*uncompacted
);
550 /* brw_eu_validate.c */
551 bool brw_validate_instructions(const struct brw_codegen
*p
, int start_offset
,
552 struct annotation_info
*annotation
);
555 next_offset(const struct gen_device_info
*devinfo
, void *store
, int offset
)
557 brw_inst
*insn
= (brw_inst
*)((char *)store
+ offset
);
559 if (brw_inst_cmpt_control(devinfo
, insn
))
566 /* The union is an implementation detail used by brw_opcode_desc() to handle
567 * opcodes that have been reused for different instructions across hardware
570 * The gens field acts as a tag. If it is non-zero, name points to a string
571 * containing the instruction mnemonic. If it is zero, the table field is
572 * valid and either points to a secondary opcode_desc table with 'size'
573 * elements or is NULL and no such instruction exists for the opcode.
581 const struct opcode_desc
*table
;
589 const struct opcode_desc
*
590 brw_opcode_desc(const struct gen_device_info
*devinfo
, enum opcode opcode
);
593 is_3src(const struct gen_device_info
*devinfo
, enum opcode opcode
)
595 const struct opcode_desc
*desc
= brw_opcode_desc(devinfo
, opcode
);
596 return desc
&& desc
->nsrc
== 3;
599 /** Maximum SEND message length */
600 #define BRW_MAX_MSG_LENGTH 15
602 /** First MRF register used by pull loads */
603 #define FIRST_SPILL_MRF(gen) ((gen) == 6 ? 21 : 13)
605 /** First MRF register used by spills */
606 #define FIRST_PULL_LOAD_MRF(gen) ((gen) == 6 ? 16 : 13)