2 Copyright (C) Intel Corp. 2006. All Rights Reserved.
3 Intel funded Tungsten Graphics (http://www.tungstengraphics.com) 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 <keith@tungstengraphics.com>
33 #include "brw_context.h"
34 #include "brw_defines.h"
37 #include "glsl/ralloc.h"
39 /***********************************************************************
40 * Internal helper for constructing instructions
43 static void guess_execution_size(struct brw_compile
*p
,
44 struct brw_instruction
*insn
,
47 if (reg
.width
== BRW_WIDTH_8
&& p
->compressed
)
48 insn
->header
.execution_size
= BRW_EXECUTE_16
;
50 insn
->header
.execution_size
= reg
.width
; /* note - definitions are compatible */
55 * Prior to Sandybridge, the SEND instruction accepted non-MRF source
56 * registers, implicitly moving the operand to a message register.
58 * On Sandybridge, this is no longer the case. This function performs the
59 * explicit move; it should be called before emitting a SEND instruction.
62 gen6_resolve_implied_move(struct brw_compile
*p
,
66 struct intel_context
*intel
= &p
->brw
->intel
;
70 if (src
->file
== BRW_MESSAGE_REGISTER_FILE
)
73 if (src
->file
!= BRW_ARCHITECTURE_REGISTER_FILE
|| src
->nr
!= BRW_ARF_NULL
) {
74 brw_push_insn_state(p
);
75 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
76 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
77 brw_MOV(p
, retype(brw_message_reg(msg_reg_nr
), BRW_REGISTER_TYPE_UD
),
78 retype(*src
, BRW_REGISTER_TYPE_UD
));
79 brw_pop_insn_state(p
);
81 *src
= brw_message_reg(msg_reg_nr
);
85 gen7_convert_mrf_to_grf(struct brw_compile
*p
, struct brw_reg
*reg
)
87 struct intel_context
*intel
= &p
->brw
->intel
;
88 if (intel
->gen
== 7 && reg
->file
== BRW_MESSAGE_REGISTER_FILE
) {
89 reg
->file
= BRW_GENERAL_REGISTER_FILE
;
96 brw_set_dest(struct brw_compile
*p
, struct brw_instruction
*insn
,
99 if (dest
.file
!= BRW_ARCHITECTURE_REGISTER_FILE
&&
100 dest
.file
!= BRW_MESSAGE_REGISTER_FILE
)
101 assert(dest
.nr
< 128);
103 gen7_convert_mrf_to_grf(p
, &dest
);
105 insn
->bits1
.da1
.dest_reg_file
= dest
.file
;
106 insn
->bits1
.da1
.dest_reg_type
= dest
.type
;
107 insn
->bits1
.da1
.dest_address_mode
= dest
.address_mode
;
109 if (dest
.address_mode
== BRW_ADDRESS_DIRECT
) {
110 insn
->bits1
.da1
.dest_reg_nr
= dest
.nr
;
112 if (insn
->header
.access_mode
== BRW_ALIGN_1
) {
113 insn
->bits1
.da1
.dest_subreg_nr
= dest
.subnr
;
114 if (dest
.hstride
== BRW_HORIZONTAL_STRIDE_0
)
115 dest
.hstride
= BRW_HORIZONTAL_STRIDE_1
;
116 insn
->bits1
.da1
.dest_horiz_stride
= dest
.hstride
;
119 insn
->bits1
.da16
.dest_subreg_nr
= dest
.subnr
/ 16;
120 insn
->bits1
.da16
.dest_writemask
= dest
.dw1
.bits
.writemask
;
121 /* even ignored in da16, still need to set as '01' */
122 insn
->bits1
.da16
.dest_horiz_stride
= 1;
126 insn
->bits1
.ia1
.dest_subreg_nr
= dest
.subnr
;
128 /* These are different sizes in align1 vs align16:
130 if (insn
->header
.access_mode
== BRW_ALIGN_1
) {
131 insn
->bits1
.ia1
.dest_indirect_offset
= dest
.dw1
.bits
.indirect_offset
;
132 if (dest
.hstride
== BRW_HORIZONTAL_STRIDE_0
)
133 dest
.hstride
= BRW_HORIZONTAL_STRIDE_1
;
134 insn
->bits1
.ia1
.dest_horiz_stride
= dest
.hstride
;
137 insn
->bits1
.ia16
.dest_indirect_offset
= dest
.dw1
.bits
.indirect_offset
;
138 /* even ignored in da16, still need to set as '01' */
139 insn
->bits1
.ia16
.dest_horiz_stride
= 1;
143 /* NEW: Set the execution size based on dest.width and
144 * insn->compression_control:
146 guess_execution_size(p
, insn
, dest
);
149 extern int reg_type_size
[];
152 validate_reg(struct brw_instruction
*insn
, struct brw_reg reg
)
154 int hstride_for_reg
[] = {0, 1, 2, 4};
155 int vstride_for_reg
[] = {0, 1, 2, 4, 8, 16, 32, 64, 128, 256};
156 int width_for_reg
[] = {1, 2, 4, 8, 16};
157 int execsize_for_reg
[] = {1, 2, 4, 8, 16};
158 int width
, hstride
, vstride
, execsize
;
160 if (reg
.file
== BRW_IMMEDIATE_VALUE
) {
161 /* 3.3.6: Region Parameters. Restriction: Immediate vectors
162 * mean the destination has to be 128-bit aligned and the
163 * destination horiz stride has to be a word.
165 if (reg
.type
== BRW_REGISTER_TYPE_V
) {
166 assert(hstride_for_reg
[insn
->bits1
.da1
.dest_horiz_stride
] *
167 reg_type_size
[insn
->bits1
.da1
.dest_reg_type
] == 2);
173 if (reg
.file
== BRW_ARCHITECTURE_REGISTER_FILE
&&
174 reg
.file
== BRW_ARF_NULL
)
177 assert(reg
.hstride
>= 0 && reg
.hstride
< Elements(hstride_for_reg
));
178 hstride
= hstride_for_reg
[reg
.hstride
];
180 if (reg
.vstride
== 0xf) {
183 assert(reg
.vstride
>= 0 && reg
.vstride
< Elements(vstride_for_reg
));
184 vstride
= vstride_for_reg
[reg
.vstride
];
187 assert(reg
.width
>= 0 && reg
.width
< Elements(width_for_reg
));
188 width
= width_for_reg
[reg
.width
];
190 assert(insn
->header
.execution_size
>= 0 &&
191 insn
->header
.execution_size
< Elements(execsize_for_reg
));
192 execsize
= execsize_for_reg
[insn
->header
.execution_size
];
194 /* Restrictions from 3.3.10: Register Region Restrictions. */
196 assert(execsize
>= width
);
199 if (execsize
== width
&& hstride
!= 0) {
200 assert(vstride
== -1 || vstride
== width
* hstride
);
204 if (execsize
== width
&& hstride
== 0) {
205 /* no restriction on vstride. */
210 assert(hstride
== 0);
214 if (execsize
== 1 && width
== 1) {
215 assert(hstride
== 0);
216 assert(vstride
== 0);
220 if (vstride
== 0 && hstride
== 0) {
224 /* 10. Check destination issues. */
228 brw_set_src0(struct brw_compile
*p
, struct brw_instruction
*insn
,
231 if (reg
.type
!= BRW_ARCHITECTURE_REGISTER_FILE
)
232 assert(reg
.nr
< 128);
234 gen7_convert_mrf_to_grf(p
, ®
);
236 validate_reg(insn
, reg
);
238 insn
->bits1
.da1
.src0_reg_file
= reg
.file
;
239 insn
->bits1
.da1
.src0_reg_type
= reg
.type
;
240 insn
->bits2
.da1
.src0_abs
= reg
.abs
;
241 insn
->bits2
.da1
.src0_negate
= reg
.negate
;
242 insn
->bits2
.da1
.src0_address_mode
= reg
.address_mode
;
244 if (reg
.file
== BRW_IMMEDIATE_VALUE
) {
245 insn
->bits3
.ud
= reg
.dw1
.ud
;
247 /* Required to set some fields in src1 as well:
249 insn
->bits1
.da1
.src1_reg_file
= 0; /* arf */
250 insn
->bits1
.da1
.src1_reg_type
= reg
.type
;
254 if (reg
.address_mode
== BRW_ADDRESS_DIRECT
) {
255 if (insn
->header
.access_mode
== BRW_ALIGN_1
) {
256 insn
->bits2
.da1
.src0_subreg_nr
= reg
.subnr
;
257 insn
->bits2
.da1
.src0_reg_nr
= reg
.nr
;
260 insn
->bits2
.da16
.src0_subreg_nr
= reg
.subnr
/ 16;
261 insn
->bits2
.da16
.src0_reg_nr
= reg
.nr
;
265 insn
->bits2
.ia1
.src0_subreg_nr
= reg
.subnr
;
267 if (insn
->header
.access_mode
== BRW_ALIGN_1
) {
268 insn
->bits2
.ia1
.src0_indirect_offset
= reg
.dw1
.bits
.indirect_offset
;
271 insn
->bits2
.ia16
.src0_subreg_nr
= reg
.dw1
.bits
.indirect_offset
;
275 if (insn
->header
.access_mode
== BRW_ALIGN_1
) {
276 if (reg
.width
== BRW_WIDTH_1
&&
277 insn
->header
.execution_size
== BRW_EXECUTE_1
) {
278 insn
->bits2
.da1
.src0_horiz_stride
= BRW_HORIZONTAL_STRIDE_0
;
279 insn
->bits2
.da1
.src0_width
= BRW_WIDTH_1
;
280 insn
->bits2
.da1
.src0_vert_stride
= BRW_VERTICAL_STRIDE_0
;
283 insn
->bits2
.da1
.src0_horiz_stride
= reg
.hstride
;
284 insn
->bits2
.da1
.src0_width
= reg
.width
;
285 insn
->bits2
.da1
.src0_vert_stride
= reg
.vstride
;
289 insn
->bits2
.da16
.src0_swz_x
= BRW_GET_SWZ(reg
.dw1
.bits
.swizzle
, BRW_CHANNEL_X
);
290 insn
->bits2
.da16
.src0_swz_y
= BRW_GET_SWZ(reg
.dw1
.bits
.swizzle
, BRW_CHANNEL_Y
);
291 insn
->bits2
.da16
.src0_swz_z
= BRW_GET_SWZ(reg
.dw1
.bits
.swizzle
, BRW_CHANNEL_Z
);
292 insn
->bits2
.da16
.src0_swz_w
= BRW_GET_SWZ(reg
.dw1
.bits
.swizzle
, BRW_CHANNEL_W
);
294 /* This is an oddity of the fact we're using the same
295 * descriptions for registers in align_16 as align_1:
297 if (reg
.vstride
== BRW_VERTICAL_STRIDE_8
)
298 insn
->bits2
.da16
.src0_vert_stride
= BRW_VERTICAL_STRIDE_4
;
300 insn
->bits2
.da16
.src0_vert_stride
= reg
.vstride
;
306 void brw_set_src1(struct brw_compile
*p
,
307 struct brw_instruction
*insn
,
310 assert(reg
.file
!= BRW_MESSAGE_REGISTER_FILE
);
312 assert(reg
.nr
< 128);
314 gen7_convert_mrf_to_grf(p
, ®
);
316 validate_reg(insn
, reg
);
318 insn
->bits1
.da1
.src1_reg_file
= reg
.file
;
319 insn
->bits1
.da1
.src1_reg_type
= reg
.type
;
320 insn
->bits3
.da1
.src1_abs
= reg
.abs
;
321 insn
->bits3
.da1
.src1_negate
= reg
.negate
;
323 /* Only src1 can be immediate in two-argument instructions.
325 assert(insn
->bits1
.da1
.src0_reg_file
!= BRW_IMMEDIATE_VALUE
);
327 if (reg
.file
== BRW_IMMEDIATE_VALUE
) {
328 insn
->bits3
.ud
= reg
.dw1
.ud
;
331 /* This is a hardware restriction, which may or may not be lifted
334 assert (reg
.address_mode
== BRW_ADDRESS_DIRECT
);
335 /* assert (reg.file == BRW_GENERAL_REGISTER_FILE); */
337 if (insn
->header
.access_mode
== BRW_ALIGN_1
) {
338 insn
->bits3
.da1
.src1_subreg_nr
= reg
.subnr
;
339 insn
->bits3
.da1
.src1_reg_nr
= reg
.nr
;
342 insn
->bits3
.da16
.src1_subreg_nr
= reg
.subnr
/ 16;
343 insn
->bits3
.da16
.src1_reg_nr
= reg
.nr
;
346 if (insn
->header
.access_mode
== BRW_ALIGN_1
) {
347 if (reg
.width
== BRW_WIDTH_1
&&
348 insn
->header
.execution_size
== BRW_EXECUTE_1
) {
349 insn
->bits3
.da1
.src1_horiz_stride
= BRW_HORIZONTAL_STRIDE_0
;
350 insn
->bits3
.da1
.src1_width
= BRW_WIDTH_1
;
351 insn
->bits3
.da1
.src1_vert_stride
= BRW_VERTICAL_STRIDE_0
;
354 insn
->bits3
.da1
.src1_horiz_stride
= reg
.hstride
;
355 insn
->bits3
.da1
.src1_width
= reg
.width
;
356 insn
->bits3
.da1
.src1_vert_stride
= reg
.vstride
;
360 insn
->bits3
.da16
.src1_swz_x
= BRW_GET_SWZ(reg
.dw1
.bits
.swizzle
, BRW_CHANNEL_X
);
361 insn
->bits3
.da16
.src1_swz_y
= BRW_GET_SWZ(reg
.dw1
.bits
.swizzle
, BRW_CHANNEL_Y
);
362 insn
->bits3
.da16
.src1_swz_z
= BRW_GET_SWZ(reg
.dw1
.bits
.swizzle
, BRW_CHANNEL_Z
);
363 insn
->bits3
.da16
.src1_swz_w
= BRW_GET_SWZ(reg
.dw1
.bits
.swizzle
, BRW_CHANNEL_W
);
365 /* This is an oddity of the fact we're using the same
366 * descriptions for registers in align_16 as align_1:
368 if (reg
.vstride
== BRW_VERTICAL_STRIDE_8
)
369 insn
->bits3
.da16
.src1_vert_stride
= BRW_VERTICAL_STRIDE_4
;
371 insn
->bits3
.da16
.src1_vert_stride
= reg
.vstride
;
377 * Set the Message Descriptor and Extended Message Descriptor fields
380 * \note This zeroes out the Function Control bits, so it must be called
381 * \b before filling out any message-specific data. Callers can
382 * choose not to fill in irrelevant bits; they will be zero.
385 brw_set_message_descriptor(struct brw_compile
*p
,
386 struct brw_instruction
*inst
,
387 enum brw_message_target sfid
,
389 unsigned response_length
,
393 struct intel_context
*intel
= &p
->brw
->intel
;
395 brw_set_src1(p
, inst
, brw_imm_d(0));
397 if (intel
->gen
>= 5) {
398 inst
->bits3
.generic_gen5
.header_present
= header_present
;
399 inst
->bits3
.generic_gen5
.response_length
= response_length
;
400 inst
->bits3
.generic_gen5
.msg_length
= msg_length
;
401 inst
->bits3
.generic_gen5
.end_of_thread
= end_of_thread
;
403 if (intel
->gen
>= 6) {
404 /* On Gen6+ Message target/SFID goes in bits 27:24 of the header */
405 inst
->header
.destreg__conditionalmod
= sfid
;
407 /* Set Extended Message Descriptor (ex_desc) */
408 inst
->bits2
.send_gen5
.sfid
= sfid
;
409 inst
->bits2
.send_gen5
.end_of_thread
= end_of_thread
;
412 inst
->bits3
.generic
.response_length
= response_length
;
413 inst
->bits3
.generic
.msg_length
= msg_length
;
414 inst
->bits3
.generic
.msg_target
= sfid
;
415 inst
->bits3
.generic
.end_of_thread
= end_of_thread
;
419 static void brw_set_math_message( struct brw_compile
*p
,
420 struct brw_instruction
*insn
,
427 struct brw_context
*brw
= p
->brw
;
428 struct intel_context
*intel
= &brw
->intel
;
430 unsigned response_length
;
432 /* Infer message length from the function */
434 case BRW_MATH_FUNCTION_POW
:
435 case BRW_MATH_FUNCTION_INT_DIV_QUOTIENT
:
436 case BRW_MATH_FUNCTION_INT_DIV_REMAINDER
:
437 case BRW_MATH_FUNCTION_INT_DIV_QUOTIENT_AND_REMAINDER
:
445 /* Infer response length from the function */
447 case BRW_MATH_FUNCTION_SINCOS
:
448 case BRW_MATH_FUNCTION_INT_DIV_QUOTIENT_AND_REMAINDER
:
456 brw_set_message_descriptor(p
, insn
, BRW_SFID_MATH
,
457 msg_length
, response_length
, false, false);
458 if (intel
->gen
== 5) {
459 insn
->bits3
.math_gen5
.function
= function
;
460 insn
->bits3
.math_gen5
.int_type
= integer_type
;
461 insn
->bits3
.math_gen5
.precision
= low_precision
;
462 insn
->bits3
.math_gen5
.saturate
= saturate
;
463 insn
->bits3
.math_gen5
.data_type
= dataType
;
464 insn
->bits3
.math_gen5
.snapshot
= 0;
466 insn
->bits3
.math
.function
= function
;
467 insn
->bits3
.math
.int_type
= integer_type
;
468 insn
->bits3
.math
.precision
= low_precision
;
469 insn
->bits3
.math
.saturate
= saturate
;
470 insn
->bits3
.math
.data_type
= dataType
;
475 static void brw_set_ff_sync_message(struct brw_compile
*p
,
476 struct brw_instruction
*insn
,
478 GLuint response_length
,
481 brw_set_message_descriptor(p
, insn
, BRW_SFID_URB
,
482 1, response_length
, true, end_of_thread
);
483 insn
->bits3
.urb_gen5
.opcode
= 1; /* FF_SYNC */
484 insn
->bits3
.urb_gen5
.offset
= 0; /* Not used by FF_SYNC */
485 insn
->bits3
.urb_gen5
.swizzle_control
= 0; /* Not used by FF_SYNC */
486 insn
->bits3
.urb_gen5
.allocate
= allocate
;
487 insn
->bits3
.urb_gen5
.used
= 0; /* Not used by FF_SYNC */
488 insn
->bits3
.urb_gen5
.complete
= 0; /* Not used by FF_SYNC */
491 static void brw_set_urb_message( struct brw_compile
*p
,
492 struct brw_instruction
*insn
,
496 GLuint response_length
,
500 GLuint swizzle_control
)
502 struct brw_context
*brw
= p
->brw
;
503 struct intel_context
*intel
= &brw
->intel
;
505 brw_set_message_descriptor(p
, insn
, BRW_SFID_URB
,
506 msg_length
, response_length
, true, end_of_thread
);
507 if (intel
->gen
== 7) {
508 insn
->bits3
.urb_gen7
.opcode
= 0; /* URB_WRITE_HWORD */
509 insn
->bits3
.urb_gen7
.offset
= offset
;
510 assert(swizzle_control
!= BRW_URB_SWIZZLE_TRANSPOSE
);
511 insn
->bits3
.urb_gen7
.swizzle_control
= swizzle_control
;
512 /* per_slot_offset = 0 makes it ignore offsets in message header */
513 insn
->bits3
.urb_gen7
.per_slot_offset
= 0;
514 insn
->bits3
.urb_gen7
.complete
= complete
;
515 } else if (intel
->gen
>= 5) {
516 insn
->bits3
.urb_gen5
.opcode
= 0; /* URB_WRITE */
517 insn
->bits3
.urb_gen5
.offset
= offset
;
518 insn
->bits3
.urb_gen5
.swizzle_control
= swizzle_control
;
519 insn
->bits3
.urb_gen5
.allocate
= allocate
;
520 insn
->bits3
.urb_gen5
.used
= used
; /* ? */
521 insn
->bits3
.urb_gen5
.complete
= complete
;
523 insn
->bits3
.urb
.opcode
= 0; /* ? */
524 insn
->bits3
.urb
.offset
= offset
;
525 insn
->bits3
.urb
.swizzle_control
= swizzle_control
;
526 insn
->bits3
.urb
.allocate
= allocate
;
527 insn
->bits3
.urb
.used
= used
; /* ? */
528 insn
->bits3
.urb
.complete
= complete
;
533 brw_set_dp_write_message(struct brw_compile
*p
,
534 struct brw_instruction
*insn
,
535 GLuint binding_table_index
,
540 GLuint last_render_target
,
541 GLuint response_length
,
542 GLuint end_of_thread
,
543 GLuint send_commit_msg
)
545 struct brw_context
*brw
= p
->brw
;
546 struct intel_context
*intel
= &brw
->intel
;
549 if (intel
->gen
>= 7) {
550 /* Use the Render Cache for RT writes; otherwise use the Data Cache */
551 if (msg_type
== GEN6_DATAPORT_WRITE_MESSAGE_RENDER_TARGET_WRITE
)
552 sfid
= GEN6_SFID_DATAPORT_RENDER_CACHE
;
554 sfid
= GEN7_SFID_DATAPORT_DATA_CACHE
;
555 } else if (intel
->gen
== 6) {
556 /* Use the render cache for all write messages. */
557 sfid
= GEN6_SFID_DATAPORT_RENDER_CACHE
;
559 sfid
= BRW_SFID_DATAPORT_WRITE
;
562 brw_set_message_descriptor(p
, insn
, sfid
, msg_length
, response_length
,
563 header_present
, end_of_thread
);
565 if (intel
->gen
>= 7) {
566 insn
->bits3
.gen7_dp
.binding_table_index
= binding_table_index
;
567 insn
->bits3
.gen7_dp
.msg_control
= msg_control
;
568 insn
->bits3
.gen7_dp
.last_render_target
= last_render_target
;
569 insn
->bits3
.gen7_dp
.msg_type
= msg_type
;
570 } else if (intel
->gen
== 6) {
571 insn
->bits3
.gen6_dp
.binding_table_index
= binding_table_index
;
572 insn
->bits3
.gen6_dp
.msg_control
= msg_control
;
573 insn
->bits3
.gen6_dp
.last_render_target
= last_render_target
;
574 insn
->bits3
.gen6_dp
.msg_type
= msg_type
;
575 insn
->bits3
.gen6_dp
.send_commit_msg
= send_commit_msg
;
576 } else if (intel
->gen
== 5) {
577 insn
->bits3
.dp_write_gen5
.binding_table_index
= binding_table_index
;
578 insn
->bits3
.dp_write_gen5
.msg_control
= msg_control
;
579 insn
->bits3
.dp_write_gen5
.last_render_target
= last_render_target
;
580 insn
->bits3
.dp_write_gen5
.msg_type
= msg_type
;
581 insn
->bits3
.dp_write_gen5
.send_commit_msg
= send_commit_msg
;
583 insn
->bits3
.dp_write
.binding_table_index
= binding_table_index
;
584 insn
->bits3
.dp_write
.msg_control
= msg_control
;
585 insn
->bits3
.dp_write
.last_render_target
= last_render_target
;
586 insn
->bits3
.dp_write
.msg_type
= msg_type
;
587 insn
->bits3
.dp_write
.send_commit_msg
= send_commit_msg
;
592 brw_set_dp_read_message(struct brw_compile
*p
,
593 struct brw_instruction
*insn
,
594 GLuint binding_table_index
,
599 GLuint response_length
)
601 struct brw_context
*brw
= p
->brw
;
602 struct intel_context
*intel
= &brw
->intel
;
605 if (intel
->gen
>= 7) {
606 sfid
= GEN7_SFID_DATAPORT_DATA_CACHE
;
607 } else if (intel
->gen
== 6) {
608 if (target_cache
== BRW_DATAPORT_READ_TARGET_RENDER_CACHE
)
609 sfid
= GEN6_SFID_DATAPORT_RENDER_CACHE
;
611 sfid
= GEN6_SFID_DATAPORT_SAMPLER_CACHE
;
613 sfid
= BRW_SFID_DATAPORT_READ
;
616 brw_set_message_descriptor(p
, insn
, sfid
, msg_length
, response_length
,
619 if (intel
->gen
>= 7) {
620 insn
->bits3
.gen7_dp
.binding_table_index
= binding_table_index
;
621 insn
->bits3
.gen7_dp
.msg_control
= msg_control
;
622 insn
->bits3
.gen7_dp
.last_render_target
= 0;
623 insn
->bits3
.gen7_dp
.msg_type
= msg_type
;
624 } else if (intel
->gen
== 6) {
625 insn
->bits3
.gen6_dp
.binding_table_index
= binding_table_index
;
626 insn
->bits3
.gen6_dp
.msg_control
= msg_control
;
627 insn
->bits3
.gen6_dp
.last_render_target
= 0;
628 insn
->bits3
.gen6_dp
.msg_type
= msg_type
;
629 insn
->bits3
.gen6_dp
.send_commit_msg
= 0;
630 } else if (intel
->gen
== 5) {
631 insn
->bits3
.dp_read_gen5
.binding_table_index
= binding_table_index
;
632 insn
->bits3
.dp_read_gen5
.msg_control
= msg_control
;
633 insn
->bits3
.dp_read_gen5
.msg_type
= msg_type
;
634 insn
->bits3
.dp_read_gen5
.target_cache
= target_cache
;
635 } else if (intel
->is_g4x
) {
636 insn
->bits3
.dp_read_g4x
.binding_table_index
= binding_table_index
; /*0:7*/
637 insn
->bits3
.dp_read_g4x
.msg_control
= msg_control
; /*8:10*/
638 insn
->bits3
.dp_read_g4x
.msg_type
= msg_type
; /*11:13*/
639 insn
->bits3
.dp_read_g4x
.target_cache
= target_cache
; /*14:15*/
641 insn
->bits3
.dp_read
.binding_table_index
= binding_table_index
; /*0:7*/
642 insn
->bits3
.dp_read
.msg_control
= msg_control
; /*8:11*/
643 insn
->bits3
.dp_read
.msg_type
= msg_type
; /*12:13*/
644 insn
->bits3
.dp_read
.target_cache
= target_cache
; /*14:15*/
648 static void brw_set_sampler_message(struct brw_compile
*p
,
649 struct brw_instruction
*insn
,
650 GLuint binding_table_index
,
653 GLuint response_length
,
655 GLuint header_present
,
657 GLuint return_format
)
659 struct brw_context
*brw
= p
->brw
;
660 struct intel_context
*intel
= &brw
->intel
;
662 brw_set_message_descriptor(p
, insn
, BRW_SFID_SAMPLER
, msg_length
,
663 response_length
, header_present
, false);
665 if (intel
->gen
>= 7) {
666 insn
->bits3
.sampler_gen7
.binding_table_index
= binding_table_index
;
667 insn
->bits3
.sampler_gen7
.sampler
= sampler
;
668 insn
->bits3
.sampler_gen7
.msg_type
= msg_type
;
669 insn
->bits3
.sampler_gen7
.simd_mode
= simd_mode
;
670 } else if (intel
->gen
>= 5) {
671 insn
->bits3
.sampler_gen5
.binding_table_index
= binding_table_index
;
672 insn
->bits3
.sampler_gen5
.sampler
= sampler
;
673 insn
->bits3
.sampler_gen5
.msg_type
= msg_type
;
674 insn
->bits3
.sampler_gen5
.simd_mode
= simd_mode
;
675 } else if (intel
->is_g4x
) {
676 insn
->bits3
.sampler_g4x
.binding_table_index
= binding_table_index
;
677 insn
->bits3
.sampler_g4x
.sampler
= sampler
;
678 insn
->bits3
.sampler_g4x
.msg_type
= msg_type
;
680 insn
->bits3
.sampler
.binding_table_index
= binding_table_index
;
681 insn
->bits3
.sampler
.sampler
= sampler
;
682 insn
->bits3
.sampler
.msg_type
= msg_type
;
683 insn
->bits3
.sampler
.return_format
= return_format
;
688 #define next_insn brw_next_insn
689 struct brw_instruction
*
690 brw_next_insn(struct brw_compile
*p
, GLuint opcode
)
692 struct brw_instruction
*insn
;
694 assert(p
->nr_insn
+ 1 < BRW_EU_MAX_INSN
);
696 insn
= &p
->store
[p
->nr_insn
++];
697 memcpy(insn
, p
->current
, sizeof(*insn
));
699 /* Reset this one-shot flag:
702 if (p
->current
->header
.destreg__conditionalmod
) {
703 p
->current
->header
.destreg__conditionalmod
= 0;
704 p
->current
->header
.predicate_control
= BRW_PREDICATE_NORMAL
;
707 insn
->header
.opcode
= opcode
;
711 static struct brw_instruction
*brw_alu1( struct brw_compile
*p
,
716 struct brw_instruction
*insn
= next_insn(p
, opcode
);
717 brw_set_dest(p
, insn
, dest
);
718 brw_set_src0(p
, insn
, src
);
722 static struct brw_instruction
*brw_alu2(struct brw_compile
*p
,
726 struct brw_reg src1
)
728 struct brw_instruction
*insn
= next_insn(p
, opcode
);
729 brw_set_dest(p
, insn
, dest
);
730 brw_set_src0(p
, insn
, src0
);
731 brw_set_src1(p
, insn
, src1
);
736 /***********************************************************************
737 * Convenience routines.
740 struct brw_instruction *brw_##OP(struct brw_compile *p, \
741 struct brw_reg dest, \
742 struct brw_reg src0) \
744 return brw_alu1(p, BRW_OPCODE_##OP, dest, src0); \
748 struct brw_instruction *brw_##OP(struct brw_compile *p, \
749 struct brw_reg dest, \
750 struct brw_reg src0, \
751 struct brw_reg src1) \
753 return brw_alu2(p, BRW_OPCODE_##OP, dest, src0, src1); \
756 /* Rounding operations (other than RNDD) require two instructions - the first
757 * stores a rounded value (possibly the wrong way) in the dest register, but
758 * also sets a per-channel "increment bit" in the flag register. A predicated
759 * add of 1.0 fixes dest to contain the desired result.
761 * Sandybridge and later appear to round correctly without an ADD.
764 void brw_##OP(struct brw_compile *p, \
765 struct brw_reg dest, \
766 struct brw_reg src) \
768 struct brw_instruction *rnd, *add; \
769 rnd = next_insn(p, BRW_OPCODE_##OP); \
770 brw_set_dest(p, rnd, dest); \
771 brw_set_src0(p, rnd, src); \
773 if (p->brw->intel.gen < 6) { \
774 /* turn on round-increments */ \
775 rnd->header.destreg__conditionalmod = BRW_CONDITIONAL_R; \
776 add = brw_ADD(p, dest, dest, brw_imm_f(1.0f)); \
777 add->header.predicate_control = BRW_PREDICATE_NORMAL; \
810 struct brw_instruction
*brw_ADD(struct brw_compile
*p
,
816 if (src0
.type
== BRW_REGISTER_TYPE_F
||
817 (src0
.file
== BRW_IMMEDIATE_VALUE
&&
818 src0
.type
== BRW_REGISTER_TYPE_VF
)) {
819 assert(src1
.type
!= BRW_REGISTER_TYPE_UD
);
820 assert(src1
.type
!= BRW_REGISTER_TYPE_D
);
823 if (src1
.type
== BRW_REGISTER_TYPE_F
||
824 (src1
.file
== BRW_IMMEDIATE_VALUE
&&
825 src1
.type
== BRW_REGISTER_TYPE_VF
)) {
826 assert(src0
.type
!= BRW_REGISTER_TYPE_UD
);
827 assert(src0
.type
!= BRW_REGISTER_TYPE_D
);
830 return brw_alu2(p
, BRW_OPCODE_ADD
, dest
, src0
, src1
);
833 struct brw_instruction
*brw_MUL(struct brw_compile
*p
,
839 if (src0
.type
== BRW_REGISTER_TYPE_D
||
840 src0
.type
== BRW_REGISTER_TYPE_UD
||
841 src1
.type
== BRW_REGISTER_TYPE_D
||
842 src1
.type
== BRW_REGISTER_TYPE_UD
) {
843 assert(dest
.type
!= BRW_REGISTER_TYPE_F
);
846 if (src0
.type
== BRW_REGISTER_TYPE_F
||
847 (src0
.file
== BRW_IMMEDIATE_VALUE
&&
848 src0
.type
== BRW_REGISTER_TYPE_VF
)) {
849 assert(src1
.type
!= BRW_REGISTER_TYPE_UD
);
850 assert(src1
.type
!= BRW_REGISTER_TYPE_D
);
853 if (src1
.type
== BRW_REGISTER_TYPE_F
||
854 (src1
.file
== BRW_IMMEDIATE_VALUE
&&
855 src1
.type
== BRW_REGISTER_TYPE_VF
)) {
856 assert(src0
.type
!= BRW_REGISTER_TYPE_UD
);
857 assert(src0
.type
!= BRW_REGISTER_TYPE_D
);
860 assert(src0
.file
!= BRW_ARCHITECTURE_REGISTER_FILE
||
861 src0
.nr
!= BRW_ARF_ACCUMULATOR
);
862 assert(src1
.file
!= BRW_ARCHITECTURE_REGISTER_FILE
||
863 src1
.nr
!= BRW_ARF_ACCUMULATOR
);
865 return brw_alu2(p
, BRW_OPCODE_MUL
, dest
, src0
, src1
);
869 void brw_NOP(struct brw_compile
*p
)
871 struct brw_instruction
*insn
= next_insn(p
, BRW_OPCODE_NOP
);
872 brw_set_dest(p
, insn
, retype(brw_vec4_grf(0,0), BRW_REGISTER_TYPE_UD
));
873 brw_set_src0(p
, insn
, retype(brw_vec4_grf(0,0), BRW_REGISTER_TYPE_UD
));
874 brw_set_src1(p
, insn
, brw_imm_ud(0x0));
881 /***********************************************************************
882 * Comparisons, if/else/endif
885 struct brw_instruction
*brw_JMPI(struct brw_compile
*p
,
890 struct brw_instruction
*insn
= brw_alu2(p
, BRW_OPCODE_JMPI
, dest
, src0
, src1
);
892 insn
->header
.execution_size
= 1;
893 insn
->header
.compression_control
= BRW_COMPRESSION_NONE
;
894 insn
->header
.mask_control
= BRW_MASK_DISABLE
;
896 p
->current
->header
.predicate_control
= BRW_PREDICATE_NONE
;
902 push_if_stack(struct brw_compile
*p
, struct brw_instruction
*inst
)
904 p
->if_stack
[p
->if_stack_depth
] = inst
- p
->store
;
907 if (p
->if_stack_array_size
<= p
->if_stack_depth
) {
908 p
->if_stack_array_size
*= 2;
909 p
->if_stack
= reralloc(p
->mem_ctx
, p
->if_stack
, int,
910 p
->if_stack_array_size
);
914 static struct brw_instruction
*
915 pop_if_stack(struct brw_compile
*p
)
918 return &p
->store
[p
->if_stack
[p
->if_stack_depth
]];
922 push_loop_stack(struct brw_compile
*p
, struct brw_instruction
*inst
)
924 if (p
->loop_stack_array_size
< p
->loop_stack_depth
) {
925 p
->loop_stack_array_size
*= 2;
926 p
->loop_stack
= reralloc(p
->mem_ctx
, p
->loop_stack
, int,
927 p
->loop_stack_array_size
);
928 p
->if_depth_in_loop
= reralloc(p
->mem_ctx
, p
->if_depth_in_loop
, int,
929 p
->loop_stack_array_size
);
932 p
->loop_stack
[p
->loop_stack_depth
] = inst
- p
->store
;
933 p
->loop_stack_depth
++;
934 p
->if_depth_in_loop
[p
->loop_stack_depth
] = 0;
937 static struct brw_instruction
*
938 get_inner_do_insn(struct brw_compile
*p
)
940 return &p
->store
[p
->loop_stack
[p
->loop_stack_depth
- 1]];
943 /* EU takes the value from the flag register and pushes it onto some
944 * sort of a stack (presumably merging with any flag value already on
945 * the stack). Within an if block, the flags at the top of the stack
946 * control execution on each channel of the unit, eg. on each of the
947 * 16 pixel values in our wm programs.
949 * When the matching 'else' instruction is reached (presumably by
950 * countdown of the instruction count patched in by our ELSE/ENDIF
951 * functions), the relevent flags are inverted.
953 * When the matching 'endif' instruction is reached, the flags are
954 * popped off. If the stack is now empty, normal execution resumes.
956 struct brw_instruction
*
957 brw_IF(struct brw_compile
*p
, GLuint execute_size
)
959 struct intel_context
*intel
= &p
->brw
->intel
;
960 struct brw_instruction
*insn
;
962 insn
= next_insn(p
, BRW_OPCODE_IF
);
964 /* Override the defaults for this instruction:
966 if (intel
->gen
< 6) {
967 brw_set_dest(p
, insn
, brw_ip_reg());
968 brw_set_src0(p
, insn
, brw_ip_reg());
969 brw_set_src1(p
, insn
, brw_imm_d(0x0));
970 } else if (intel
->gen
== 6) {
971 brw_set_dest(p
, insn
, brw_imm_w(0));
972 insn
->bits1
.branch_gen6
.jump_count
= 0;
973 brw_set_src0(p
, insn
, vec1(retype(brw_null_reg(), BRW_REGISTER_TYPE_D
)));
974 brw_set_src1(p
, insn
, vec1(retype(brw_null_reg(), BRW_REGISTER_TYPE_D
)));
976 brw_set_dest(p
, insn
, vec1(retype(brw_null_reg(), BRW_REGISTER_TYPE_D
)));
977 brw_set_src0(p
, insn
, vec1(retype(brw_null_reg(), BRW_REGISTER_TYPE_D
)));
978 brw_set_src1(p
, insn
, brw_imm_ud(0));
979 insn
->bits3
.break_cont
.jip
= 0;
980 insn
->bits3
.break_cont
.uip
= 0;
983 insn
->header
.execution_size
= execute_size
;
984 insn
->header
.compression_control
= BRW_COMPRESSION_NONE
;
985 insn
->header
.predicate_control
= BRW_PREDICATE_NORMAL
;
986 insn
->header
.mask_control
= BRW_MASK_ENABLE
;
987 if (!p
->single_program_flow
)
988 insn
->header
.thread_control
= BRW_THREAD_SWITCH
;
990 p
->current
->header
.predicate_control
= BRW_PREDICATE_NONE
;
992 push_if_stack(p
, insn
);
993 p
->if_depth_in_loop
[p
->loop_stack_depth
]++;
997 /* This function is only used for gen6-style IF instructions with an
998 * embedded comparison (conditional modifier). It is not used on gen7.
1000 struct brw_instruction
*
1001 gen6_IF(struct brw_compile
*p
, uint32_t conditional
,
1002 struct brw_reg src0
, struct brw_reg src1
)
1004 struct brw_instruction
*insn
;
1006 insn
= next_insn(p
, BRW_OPCODE_IF
);
1008 brw_set_dest(p
, insn
, brw_imm_w(0));
1009 if (p
->compressed
) {
1010 insn
->header
.execution_size
= BRW_EXECUTE_16
;
1012 insn
->header
.execution_size
= BRW_EXECUTE_8
;
1014 insn
->bits1
.branch_gen6
.jump_count
= 0;
1015 brw_set_src0(p
, insn
, src0
);
1016 brw_set_src1(p
, insn
, src1
);
1018 assert(insn
->header
.compression_control
== BRW_COMPRESSION_NONE
);
1019 assert(insn
->header
.predicate_control
== BRW_PREDICATE_NONE
);
1020 insn
->header
.destreg__conditionalmod
= conditional
;
1022 if (!p
->single_program_flow
)
1023 insn
->header
.thread_control
= BRW_THREAD_SWITCH
;
1025 push_if_stack(p
, insn
);
1030 * In single-program-flow (SPF) mode, convert IF and ELSE into ADDs.
1033 convert_IF_ELSE_to_ADD(struct brw_compile
*p
,
1034 struct brw_instruction
*if_inst
,
1035 struct brw_instruction
*else_inst
)
1037 /* The next instruction (where the ENDIF would be, if it existed) */
1038 struct brw_instruction
*next_inst
= &p
->store
[p
->nr_insn
];
1040 assert(p
->single_program_flow
);
1041 assert(if_inst
!= NULL
&& if_inst
->header
.opcode
== BRW_OPCODE_IF
);
1042 assert(else_inst
== NULL
|| else_inst
->header
.opcode
== BRW_OPCODE_ELSE
);
1043 assert(if_inst
->header
.execution_size
== BRW_EXECUTE_1
);
1045 /* Convert IF to an ADD instruction that moves the instruction pointer
1046 * to the first instruction of the ELSE block. If there is no ELSE
1047 * block, point to where ENDIF would be. Reverse the predicate.
1049 * There's no need to execute an ENDIF since we don't need to do any
1050 * stack operations, and if we're currently executing, we just want to
1051 * continue normally.
1053 if_inst
->header
.opcode
= BRW_OPCODE_ADD
;
1054 if_inst
->header
.predicate_inverse
= 1;
1056 if (else_inst
!= NULL
) {
1057 /* Convert ELSE to an ADD instruction that points where the ENDIF
1060 else_inst
->header
.opcode
= BRW_OPCODE_ADD
;
1062 if_inst
->bits3
.ud
= (else_inst
- if_inst
+ 1) * 16;
1063 else_inst
->bits3
.ud
= (next_inst
- else_inst
) * 16;
1065 if_inst
->bits3
.ud
= (next_inst
- if_inst
) * 16;
1070 * Patch IF and ELSE instructions with appropriate jump targets.
1073 patch_IF_ELSE(struct brw_compile
*p
,
1074 struct brw_instruction
*if_inst
,
1075 struct brw_instruction
*else_inst
,
1076 struct brw_instruction
*endif_inst
)
1078 struct intel_context
*intel
= &p
->brw
->intel
;
1080 /* We shouldn't be patching IF and ELSE instructions in single program flow
1081 * mode when gen < 6, because in single program flow mode on those
1082 * platforms, we convert flow control instructions to conditional ADDs that
1083 * operate on IP (see brw_ENDIF).
1085 * However, on Gen6, writing to IP doesn't work in single program flow mode
1086 * (see the SandyBridge PRM, Volume 4 part 2, p79: "When SPF is ON, IP may
1087 * not be updated by non-flow control instructions."). And on later
1088 * platforms, there is no significant benefit to converting control flow
1089 * instructions to conditional ADDs. So we do patch IF and ELSE
1090 * instructions in single program flow mode on those platforms.
1093 assert(!p
->single_program_flow
);
1095 assert(if_inst
!= NULL
&& if_inst
->header
.opcode
== BRW_OPCODE_IF
);
1096 assert(endif_inst
!= NULL
);
1097 assert(else_inst
== NULL
|| else_inst
->header
.opcode
== BRW_OPCODE_ELSE
);
1100 /* Jump count is for 64bit data chunk each, so one 128bit instruction
1101 * requires 2 chunks.
1103 if (intel
->gen
>= 5)
1106 assert(endif_inst
->header
.opcode
== BRW_OPCODE_ENDIF
);
1107 endif_inst
->header
.execution_size
= if_inst
->header
.execution_size
;
1109 if (else_inst
== NULL
) {
1110 /* Patch IF -> ENDIF */
1111 if (intel
->gen
< 6) {
1112 /* Turn it into an IFF, which means no mask stack operations for
1113 * all-false and jumping past the ENDIF.
1115 if_inst
->header
.opcode
= BRW_OPCODE_IFF
;
1116 if_inst
->bits3
.if_else
.jump_count
= br
* (endif_inst
- if_inst
+ 1);
1117 if_inst
->bits3
.if_else
.pop_count
= 0;
1118 if_inst
->bits3
.if_else
.pad0
= 0;
1119 } else if (intel
->gen
== 6) {
1120 /* As of gen6, there is no IFF and IF must point to the ENDIF. */
1121 if_inst
->bits1
.branch_gen6
.jump_count
= br
* (endif_inst
- if_inst
);
1123 if_inst
->bits3
.break_cont
.uip
= br
* (endif_inst
- if_inst
);
1124 if_inst
->bits3
.break_cont
.jip
= br
* (endif_inst
- if_inst
);
1127 else_inst
->header
.execution_size
= if_inst
->header
.execution_size
;
1129 /* Patch IF -> ELSE */
1130 if (intel
->gen
< 6) {
1131 if_inst
->bits3
.if_else
.jump_count
= br
* (else_inst
- if_inst
);
1132 if_inst
->bits3
.if_else
.pop_count
= 0;
1133 if_inst
->bits3
.if_else
.pad0
= 0;
1134 } else if (intel
->gen
== 6) {
1135 if_inst
->bits1
.branch_gen6
.jump_count
= br
* (else_inst
- if_inst
+ 1);
1138 /* Patch ELSE -> ENDIF */
1139 if (intel
->gen
< 6) {
1140 /* BRW_OPCODE_ELSE pre-gen6 should point just past the
1143 else_inst
->bits3
.if_else
.jump_count
= br
*(endif_inst
- else_inst
+ 1);
1144 else_inst
->bits3
.if_else
.pop_count
= 1;
1145 else_inst
->bits3
.if_else
.pad0
= 0;
1146 } else if (intel
->gen
== 6) {
1147 /* BRW_OPCODE_ELSE on gen6 should point to the matching ENDIF. */
1148 else_inst
->bits1
.branch_gen6
.jump_count
= br
*(endif_inst
- else_inst
);
1150 /* The IF instruction's JIP should point just past the ELSE */
1151 if_inst
->bits3
.break_cont
.jip
= br
* (else_inst
- if_inst
+ 1);
1152 /* The IF instruction's UIP and ELSE's JIP should point to ENDIF */
1153 if_inst
->bits3
.break_cont
.uip
= br
* (endif_inst
- if_inst
);
1154 else_inst
->bits3
.break_cont
.jip
= br
* (endif_inst
- else_inst
);
1160 brw_ELSE(struct brw_compile
*p
)
1162 struct intel_context
*intel
= &p
->brw
->intel
;
1163 struct brw_instruction
*insn
;
1165 insn
= next_insn(p
, BRW_OPCODE_ELSE
);
1167 if (intel
->gen
< 6) {
1168 brw_set_dest(p
, insn
, brw_ip_reg());
1169 brw_set_src0(p
, insn
, brw_ip_reg());
1170 brw_set_src1(p
, insn
, brw_imm_d(0x0));
1171 } else if (intel
->gen
== 6) {
1172 brw_set_dest(p
, insn
, brw_imm_w(0));
1173 insn
->bits1
.branch_gen6
.jump_count
= 0;
1174 brw_set_src0(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1175 brw_set_src1(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1177 brw_set_dest(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1178 brw_set_src0(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1179 brw_set_src1(p
, insn
, brw_imm_ud(0));
1180 insn
->bits3
.break_cont
.jip
= 0;
1181 insn
->bits3
.break_cont
.uip
= 0;
1184 insn
->header
.compression_control
= BRW_COMPRESSION_NONE
;
1185 insn
->header
.mask_control
= BRW_MASK_ENABLE
;
1186 if (!p
->single_program_flow
)
1187 insn
->header
.thread_control
= BRW_THREAD_SWITCH
;
1189 push_if_stack(p
, insn
);
1193 brw_ENDIF(struct brw_compile
*p
)
1195 struct intel_context
*intel
= &p
->brw
->intel
;
1196 struct brw_instruction
*insn
;
1197 struct brw_instruction
*else_inst
= NULL
;
1198 struct brw_instruction
*if_inst
= NULL
;
1199 struct brw_instruction
*tmp
;
1200 bool emit_endif
= true;
1202 /* In single program flow mode, we can express IF and ELSE instructions
1203 * equivalently as ADD instructions that operate on IP. On platforms prior
1204 * to Gen6, flow control instructions cause an implied thread switch, so
1205 * this is a significant savings.
1207 * However, on Gen6, writing to IP doesn't work in single program flow mode
1208 * (see the SandyBridge PRM, Volume 4 part 2, p79: "When SPF is ON, IP may
1209 * not be updated by non-flow control instructions."). And on later
1210 * platforms, there is no significant benefit to converting control flow
1211 * instructions to conditional ADDs. So we only do this trick on Gen4 and
1214 if (intel
->gen
< 6 && p
->single_program_flow
)
1218 * A single next_insn() may change the base adress of instruction store
1219 * memory(p->store), so call it first before referencing the instruction
1220 * store pointer from an index
1223 insn
= next_insn(p
, BRW_OPCODE_ENDIF
);
1225 /* Pop the IF and (optional) ELSE instructions from the stack */
1226 p
->if_depth_in_loop
[p
->loop_stack_depth
]--;
1227 tmp
= pop_if_stack(p
);
1228 if (tmp
->header
.opcode
== BRW_OPCODE_ELSE
) {
1230 tmp
= pop_if_stack(p
);
1235 /* ENDIF is useless; don't bother emitting it. */
1236 convert_IF_ELSE_to_ADD(p
, if_inst
, else_inst
);
1240 if (intel
->gen
< 6) {
1241 brw_set_dest(p
, insn
, retype(brw_vec4_grf(0,0), BRW_REGISTER_TYPE_UD
));
1242 brw_set_src0(p
, insn
, retype(brw_vec4_grf(0,0), BRW_REGISTER_TYPE_UD
));
1243 brw_set_src1(p
, insn
, brw_imm_d(0x0));
1244 } else if (intel
->gen
== 6) {
1245 brw_set_dest(p
, insn
, brw_imm_w(0));
1246 brw_set_src0(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1247 brw_set_src1(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1249 brw_set_dest(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1250 brw_set_src0(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1251 brw_set_src1(p
, insn
, brw_imm_ud(0));
1254 insn
->header
.compression_control
= BRW_COMPRESSION_NONE
;
1255 insn
->header
.mask_control
= BRW_MASK_ENABLE
;
1256 insn
->header
.thread_control
= BRW_THREAD_SWITCH
;
1258 /* Also pop item off the stack in the endif instruction: */
1259 if (intel
->gen
< 6) {
1260 insn
->bits3
.if_else
.jump_count
= 0;
1261 insn
->bits3
.if_else
.pop_count
= 1;
1262 insn
->bits3
.if_else
.pad0
= 0;
1263 } else if (intel
->gen
== 6) {
1264 insn
->bits1
.branch_gen6
.jump_count
= 2;
1266 insn
->bits3
.break_cont
.jip
= 2;
1268 patch_IF_ELSE(p
, if_inst
, else_inst
, insn
);
1271 struct brw_instruction
*brw_BREAK(struct brw_compile
*p
)
1273 struct intel_context
*intel
= &p
->brw
->intel
;
1274 struct brw_instruction
*insn
;
1276 insn
= next_insn(p
, BRW_OPCODE_BREAK
);
1277 if (intel
->gen
>= 6) {
1278 brw_set_dest(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1279 brw_set_src0(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1280 brw_set_src1(p
, insn
, brw_imm_d(0x0));
1282 brw_set_dest(p
, insn
, brw_ip_reg());
1283 brw_set_src0(p
, insn
, brw_ip_reg());
1284 brw_set_src1(p
, insn
, brw_imm_d(0x0));
1285 insn
->bits3
.if_else
.pad0
= 0;
1286 insn
->bits3
.if_else
.pop_count
= p
->if_depth_in_loop
[p
->loop_stack_depth
];
1288 insn
->header
.compression_control
= BRW_COMPRESSION_NONE
;
1289 insn
->header
.execution_size
= BRW_EXECUTE_8
;
1294 struct brw_instruction
*gen6_CONT(struct brw_compile
*p
)
1296 struct brw_instruction
*insn
;
1298 insn
= next_insn(p
, BRW_OPCODE_CONTINUE
);
1299 brw_set_dest(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1300 brw_set_src0(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1301 brw_set_dest(p
, insn
, brw_ip_reg());
1302 brw_set_src0(p
, insn
, brw_ip_reg());
1303 brw_set_src1(p
, insn
, brw_imm_d(0x0));
1305 insn
->header
.compression_control
= BRW_COMPRESSION_NONE
;
1306 insn
->header
.execution_size
= BRW_EXECUTE_8
;
1310 struct brw_instruction
*brw_CONT(struct brw_compile
*p
)
1312 struct brw_instruction
*insn
;
1313 insn
= next_insn(p
, BRW_OPCODE_CONTINUE
);
1314 brw_set_dest(p
, insn
, brw_ip_reg());
1315 brw_set_src0(p
, insn
, brw_ip_reg());
1316 brw_set_src1(p
, insn
, brw_imm_d(0x0));
1317 insn
->header
.compression_control
= BRW_COMPRESSION_NONE
;
1318 insn
->header
.execution_size
= BRW_EXECUTE_8
;
1319 /* insn->header.mask_control = BRW_MASK_DISABLE; */
1320 insn
->bits3
.if_else
.pad0
= 0;
1321 insn
->bits3
.if_else
.pop_count
= p
->if_depth_in_loop
[p
->loop_stack_depth
];
1327 * The DO/WHILE is just an unterminated loop -- break or continue are
1328 * used for control within the loop. We have a few ways they can be
1331 * For uniform control flow, the WHILE is just a jump, so ADD ip, ip,
1332 * jip and no DO instruction.
1334 * For non-uniform control flow pre-gen6, there's a DO instruction to
1335 * push the mask, and a WHILE to jump back, and BREAK to get out and
1338 * For gen6, there's no more mask stack, so no need for DO. WHILE
1339 * just points back to the first instruction of the loop.
1341 struct brw_instruction
*brw_DO(struct brw_compile
*p
, GLuint execute_size
)
1343 struct intel_context
*intel
= &p
->brw
->intel
;
1345 if (intel
->gen
>= 6 || p
->single_program_flow
) {
1346 push_loop_stack(p
, &p
->store
[p
->nr_insn
]);
1347 return &p
->store
[p
->nr_insn
];
1349 struct brw_instruction
*insn
= next_insn(p
, BRW_OPCODE_DO
);
1351 push_loop_stack(p
, insn
);
1353 /* Override the defaults for this instruction:
1355 brw_set_dest(p
, insn
, brw_null_reg());
1356 brw_set_src0(p
, insn
, brw_null_reg());
1357 brw_set_src1(p
, insn
, brw_null_reg());
1359 insn
->header
.compression_control
= BRW_COMPRESSION_NONE
;
1360 insn
->header
.execution_size
= execute_size
;
1361 insn
->header
.predicate_control
= BRW_PREDICATE_NONE
;
1362 /* insn->header.mask_control = BRW_MASK_ENABLE; */
1363 /* insn->header.mask_control = BRW_MASK_DISABLE; */
1370 * For pre-gen6, we patch BREAK/CONT instructions to point at the WHILE
1373 * For gen6+, see brw_set_uip_jip(), which doesn't care so much about the loop
1374 * nesting, since it can always just point to the end of the block/current loop.
1377 brw_patch_break_cont(struct brw_compile
*p
, struct brw_instruction
*while_inst
)
1379 struct intel_context
*intel
= &p
->brw
->intel
;
1380 struct brw_instruction
*do_inst
= get_inner_do_insn(p
);
1381 struct brw_instruction
*inst
;
1382 int br
= (intel
->gen
== 5) ? 2 : 1;
1384 for (inst
= while_inst
- 1; inst
!= do_inst
; inst
--) {
1385 /* If the jump count is != 0, that means that this instruction has already
1386 * been patched because it's part of a loop inside of the one we're
1389 if (inst
->header
.opcode
== BRW_OPCODE_BREAK
&&
1390 inst
->bits3
.if_else
.jump_count
== 0) {
1391 inst
->bits3
.if_else
.jump_count
= br
* ((while_inst
- inst
) + 1);
1392 } else if (inst
->header
.opcode
== BRW_OPCODE_CONTINUE
&&
1393 inst
->bits3
.if_else
.jump_count
== 0) {
1394 inst
->bits3
.if_else
.jump_count
= br
* (while_inst
- inst
);
1399 struct brw_instruction
*brw_WHILE(struct brw_compile
*p
)
1401 struct intel_context
*intel
= &p
->brw
->intel
;
1402 struct brw_instruction
*insn
, *do_insn
;
1405 if (intel
->gen
>= 5)
1408 if (intel
->gen
>= 7) {
1409 insn
= next_insn(p
, BRW_OPCODE_WHILE
);
1410 do_insn
= get_inner_do_insn(p
);
1412 brw_set_dest(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1413 brw_set_src0(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1414 brw_set_src1(p
, insn
, brw_imm_ud(0));
1415 insn
->bits3
.break_cont
.jip
= br
* (do_insn
- insn
);
1417 insn
->header
.execution_size
= BRW_EXECUTE_8
;
1418 } else if (intel
->gen
== 6) {
1419 insn
= next_insn(p
, BRW_OPCODE_WHILE
);
1420 do_insn
= get_inner_do_insn(p
);
1422 brw_set_dest(p
, insn
, brw_imm_w(0));
1423 insn
->bits1
.branch_gen6
.jump_count
= br
* (do_insn
- insn
);
1424 brw_set_src0(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1425 brw_set_src1(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1427 insn
->header
.execution_size
= BRW_EXECUTE_8
;
1429 if (p
->single_program_flow
) {
1430 insn
= next_insn(p
, BRW_OPCODE_ADD
);
1431 do_insn
= get_inner_do_insn(p
);
1433 brw_set_dest(p
, insn
, brw_ip_reg());
1434 brw_set_src0(p
, insn
, brw_ip_reg());
1435 brw_set_src1(p
, insn
, brw_imm_d((do_insn
- insn
) * 16));
1436 insn
->header
.execution_size
= BRW_EXECUTE_1
;
1438 insn
= next_insn(p
, BRW_OPCODE_WHILE
);
1439 do_insn
= get_inner_do_insn(p
);
1441 assert(do_insn
->header
.opcode
== BRW_OPCODE_DO
);
1443 brw_set_dest(p
, insn
, brw_ip_reg());
1444 brw_set_src0(p
, insn
, brw_ip_reg());
1445 brw_set_src1(p
, insn
, brw_imm_d(0));
1447 insn
->header
.execution_size
= do_insn
->header
.execution_size
;
1448 insn
->bits3
.if_else
.jump_count
= br
* (do_insn
- insn
+ 1);
1449 insn
->bits3
.if_else
.pop_count
= 0;
1450 insn
->bits3
.if_else
.pad0
= 0;
1452 brw_patch_break_cont(p
, insn
);
1455 insn
->header
.compression_control
= BRW_COMPRESSION_NONE
;
1456 p
->current
->header
.predicate_control
= BRW_PREDICATE_NONE
;
1458 p
->loop_stack_depth
--;
1466 void brw_land_fwd_jump(struct brw_compile
*p
, int jmp_insn_idx
)
1468 struct intel_context
*intel
= &p
->brw
->intel
;
1469 struct brw_instruction
*jmp_insn
= &p
->store
[jmp_insn_idx
];
1472 if (intel
->gen
>= 5)
1475 assert(jmp_insn
->header
.opcode
== BRW_OPCODE_JMPI
);
1476 assert(jmp_insn
->bits1
.da1
.src1_reg_file
== BRW_IMMEDIATE_VALUE
);
1478 jmp_insn
->bits3
.ud
= jmpi
* (p
->nr_insn
- jmp_insn_idx
- 1);
1483 /* To integrate with the above, it makes sense that the comparison
1484 * instruction should populate the flag register. It might be simpler
1485 * just to use the flag reg for most WM tasks?
1487 void brw_CMP(struct brw_compile
*p
,
1488 struct brw_reg dest
,
1490 struct brw_reg src0
,
1491 struct brw_reg src1
)
1493 struct brw_instruction
*insn
= next_insn(p
, BRW_OPCODE_CMP
);
1495 insn
->header
.destreg__conditionalmod
= conditional
;
1496 brw_set_dest(p
, insn
, dest
);
1497 brw_set_src0(p
, insn
, src0
);
1498 brw_set_src1(p
, insn
, src1
);
1500 /* guess_execution_size(insn, src0); */
1503 /* Make it so that future instructions will use the computed flag
1504 * value until brw_set_predicate_control_flag_value() is called
1507 if (dest
.file
== BRW_ARCHITECTURE_REGISTER_FILE
&&
1509 p
->current
->header
.predicate_control
= BRW_PREDICATE_NORMAL
;
1510 p
->flag_value
= 0xff;
1514 /* Issue 'wait' instruction for n1, host could program MMIO
1515 to wake up thread. */
1516 void brw_WAIT (struct brw_compile
*p
)
1518 struct brw_instruction
*insn
= next_insn(p
, BRW_OPCODE_WAIT
);
1519 struct brw_reg src
= brw_notification_1_reg();
1521 brw_set_dest(p
, insn
, src
);
1522 brw_set_src0(p
, insn
, src
);
1523 brw_set_src1(p
, insn
, brw_null_reg());
1524 insn
->header
.execution_size
= 0; /* must */
1525 insn
->header
.predicate_control
= 0;
1526 insn
->header
.compression_control
= 0;
1530 /***********************************************************************
1531 * Helpers for the various SEND message types:
1534 /** Extended math function, float[8].
1536 void brw_math( struct brw_compile
*p
,
1537 struct brw_reg dest
,
1545 struct intel_context
*intel
= &p
->brw
->intel
;
1547 if (intel
->gen
>= 6) {
1548 struct brw_instruction
*insn
= next_insn(p
, BRW_OPCODE_MATH
);
1550 assert(dest
.file
== BRW_GENERAL_REGISTER_FILE
);
1551 assert(src
.file
== BRW_GENERAL_REGISTER_FILE
);
1553 assert(dest
.hstride
== BRW_HORIZONTAL_STRIDE_1
);
1554 if (intel
->gen
== 6)
1555 assert(src
.hstride
== BRW_HORIZONTAL_STRIDE_1
);
1557 /* Source modifiers are ignored for extended math instructions on Gen6. */
1558 if (intel
->gen
== 6) {
1559 assert(!src
.negate
);
1563 if (function
== BRW_MATH_FUNCTION_INT_DIV_QUOTIENT
||
1564 function
== BRW_MATH_FUNCTION_INT_DIV_REMAINDER
||
1565 function
== BRW_MATH_FUNCTION_INT_DIV_QUOTIENT_AND_REMAINDER
) {
1566 assert(src
.type
!= BRW_REGISTER_TYPE_F
);
1568 assert(src
.type
== BRW_REGISTER_TYPE_F
);
1571 /* Math is the same ISA format as other opcodes, except that CondModifier
1572 * becomes FC[3:0] and ThreadCtrl becomes FC[5:4].
1574 insn
->header
.destreg__conditionalmod
= function
;
1575 insn
->header
.saturate
= saturate
;
1577 brw_set_dest(p
, insn
, dest
);
1578 brw_set_src0(p
, insn
, src
);
1579 brw_set_src1(p
, insn
, brw_null_reg());
1581 struct brw_instruction
*insn
= next_insn(p
, BRW_OPCODE_SEND
);
1583 /* Example code doesn't set predicate_control for send
1586 insn
->header
.predicate_control
= 0;
1587 insn
->header
.destreg__conditionalmod
= msg_reg_nr
;
1589 brw_set_dest(p
, insn
, dest
);
1590 brw_set_src0(p
, insn
, src
);
1591 brw_set_math_message(p
,
1594 src
.type
== BRW_REGISTER_TYPE_D
,
1601 /** Extended math function, float[8].
1603 void brw_math2(struct brw_compile
*p
,
1604 struct brw_reg dest
,
1606 struct brw_reg src0
,
1607 struct brw_reg src1
)
1609 struct intel_context
*intel
= &p
->brw
->intel
;
1610 struct brw_instruction
*insn
= next_insn(p
, BRW_OPCODE_MATH
);
1612 assert(intel
->gen
>= 6);
1616 assert(dest
.file
== BRW_GENERAL_REGISTER_FILE
);
1617 assert(src0
.file
== BRW_GENERAL_REGISTER_FILE
);
1618 assert(src1
.file
== BRW_GENERAL_REGISTER_FILE
);
1620 assert(dest
.hstride
== BRW_HORIZONTAL_STRIDE_1
);
1621 if (intel
->gen
== 6) {
1622 assert(src0
.hstride
== BRW_HORIZONTAL_STRIDE_1
);
1623 assert(src1
.hstride
== BRW_HORIZONTAL_STRIDE_1
);
1626 if (function
== BRW_MATH_FUNCTION_INT_DIV_QUOTIENT
||
1627 function
== BRW_MATH_FUNCTION_INT_DIV_REMAINDER
||
1628 function
== BRW_MATH_FUNCTION_INT_DIV_QUOTIENT_AND_REMAINDER
) {
1629 assert(src0
.type
!= BRW_REGISTER_TYPE_F
);
1630 assert(src1
.type
!= BRW_REGISTER_TYPE_F
);
1632 assert(src0
.type
== BRW_REGISTER_TYPE_F
);
1633 assert(src1
.type
== BRW_REGISTER_TYPE_F
);
1636 /* Source modifiers are ignored for extended math instructions on Gen6. */
1637 if (intel
->gen
== 6) {
1638 assert(!src0
.negate
);
1640 assert(!src1
.negate
);
1644 /* Math is the same ISA format as other opcodes, except that CondModifier
1645 * becomes FC[3:0] and ThreadCtrl becomes FC[5:4].
1647 insn
->header
.destreg__conditionalmod
= function
;
1649 brw_set_dest(p
, insn
, dest
);
1650 brw_set_src0(p
, insn
, src0
);
1651 brw_set_src1(p
, insn
, src1
);
1655 * Extended math function, float[16].
1656 * Use 2 send instructions.
1658 void brw_math_16( struct brw_compile
*p
,
1659 struct brw_reg dest
,
1666 struct intel_context
*intel
= &p
->brw
->intel
;
1667 struct brw_instruction
*insn
;
1669 if (intel
->gen
>= 6) {
1670 insn
= next_insn(p
, BRW_OPCODE_MATH
);
1672 /* Math is the same ISA format as other opcodes, except that CondModifier
1673 * becomes FC[3:0] and ThreadCtrl becomes FC[5:4].
1675 insn
->header
.destreg__conditionalmod
= function
;
1676 insn
->header
.saturate
= saturate
;
1678 /* Source modifiers are ignored for extended math instructions. */
1679 assert(!src
.negate
);
1682 brw_set_dest(p
, insn
, dest
);
1683 brw_set_src0(p
, insn
, src
);
1684 brw_set_src1(p
, insn
, brw_null_reg());
1688 /* First instruction:
1690 brw_push_insn_state(p
);
1691 brw_set_predicate_control_flag_value(p
, 0xff);
1692 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
1694 insn
= next_insn(p
, BRW_OPCODE_SEND
);
1695 insn
->header
.destreg__conditionalmod
= msg_reg_nr
;
1697 brw_set_dest(p
, insn
, dest
);
1698 brw_set_src0(p
, insn
, src
);
1699 brw_set_math_message(p
,
1702 BRW_MATH_INTEGER_UNSIGNED
,
1705 BRW_MATH_DATA_VECTOR
);
1707 /* Second instruction:
1709 insn
= next_insn(p
, BRW_OPCODE_SEND
);
1710 insn
->header
.compression_control
= BRW_COMPRESSION_2NDHALF
;
1711 insn
->header
.destreg__conditionalmod
= msg_reg_nr
+1;
1713 brw_set_dest(p
, insn
, offset(dest
,1));
1714 brw_set_src0(p
, insn
, src
);
1715 brw_set_math_message(p
,
1718 BRW_MATH_INTEGER_UNSIGNED
,
1721 BRW_MATH_DATA_VECTOR
);
1723 brw_pop_insn_state(p
);
1728 * Write a block of OWORDs (half a GRF each) from the scratch buffer,
1729 * using a constant offset per channel.
1731 * The offset must be aligned to oword size (16 bytes). Used for
1732 * register spilling.
1734 void brw_oword_block_write_scratch(struct brw_compile
*p
,
1739 struct intel_context
*intel
= &p
->brw
->intel
;
1740 uint32_t msg_control
, msg_type
;
1743 if (intel
->gen
>= 6)
1746 mrf
= retype(mrf
, BRW_REGISTER_TYPE_UD
);
1748 if (num_regs
== 1) {
1749 msg_control
= BRW_DATAPORT_OWORD_BLOCK_2_OWORDS
;
1752 msg_control
= BRW_DATAPORT_OWORD_BLOCK_4_OWORDS
;
1756 /* Set up the message header. This is g0, with g0.2 filled with
1757 * the offset. We don't want to leave our offset around in g0 or
1758 * it'll screw up texture samples, so set it up inside the message
1762 brw_push_insn_state(p
);
1763 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
1764 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
1766 brw_MOV(p
, mrf
, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
1768 /* set message header global offset field (reg 0, element 2) */
1770 retype(brw_vec1_reg(BRW_MESSAGE_REGISTER_FILE
,
1772 2), BRW_REGISTER_TYPE_UD
),
1773 brw_imm_ud(offset
));
1775 brw_pop_insn_state(p
);
1779 struct brw_reg dest
;
1780 struct brw_instruction
*insn
= next_insn(p
, BRW_OPCODE_SEND
);
1781 int send_commit_msg
;
1782 struct brw_reg src_header
= retype(brw_vec8_grf(0, 0),
1783 BRW_REGISTER_TYPE_UW
);
1785 if (insn
->header
.compression_control
!= BRW_COMPRESSION_NONE
) {
1786 insn
->header
.compression_control
= BRW_COMPRESSION_NONE
;
1787 src_header
= vec16(src_header
);
1789 assert(insn
->header
.predicate_control
== BRW_PREDICATE_NONE
);
1790 insn
->header
.destreg__conditionalmod
= mrf
.nr
;
1792 /* Until gen6, writes followed by reads from the same location
1793 * are not guaranteed to be ordered unless write_commit is set.
1794 * If set, then a no-op write is issued to the destination
1795 * register to set a dependency, and a read from the destination
1796 * can be used to ensure the ordering.
1798 * For gen6, only writes between different threads need ordering
1799 * protection. Our use of DP writes is all about register
1800 * spilling within a thread.
1802 if (intel
->gen
>= 6) {
1803 dest
= retype(vec16(brw_null_reg()), BRW_REGISTER_TYPE_UW
);
1804 send_commit_msg
= 0;
1807 send_commit_msg
= 1;
1810 brw_set_dest(p
, insn
, dest
);
1811 if (intel
->gen
>= 6) {
1812 brw_set_src0(p
, insn
, mrf
);
1814 brw_set_src0(p
, insn
, brw_null_reg());
1817 if (intel
->gen
>= 6)
1818 msg_type
= GEN6_DATAPORT_WRITE_MESSAGE_OWORD_BLOCK_WRITE
;
1820 msg_type
= BRW_DATAPORT_WRITE_MESSAGE_OWORD_BLOCK_WRITE
;
1822 brw_set_dp_write_message(p
,
1824 255, /* binding table index (255=stateless) */
1828 true, /* header_present */
1829 0, /* not a render target */
1830 send_commit_msg
, /* response_length */
1838 * Read a block of owords (half a GRF each) from the scratch buffer
1839 * using a constant index per channel.
1841 * Offset must be aligned to oword size (16 bytes). Used for register
1845 brw_oword_block_read_scratch(struct brw_compile
*p
,
1846 struct brw_reg dest
,
1851 struct intel_context
*intel
= &p
->brw
->intel
;
1852 uint32_t msg_control
;
1855 if (intel
->gen
>= 6)
1858 mrf
= retype(mrf
, BRW_REGISTER_TYPE_UD
);
1859 dest
= retype(dest
, BRW_REGISTER_TYPE_UW
);
1861 if (num_regs
== 1) {
1862 msg_control
= BRW_DATAPORT_OWORD_BLOCK_2_OWORDS
;
1865 msg_control
= BRW_DATAPORT_OWORD_BLOCK_4_OWORDS
;
1870 brw_push_insn_state(p
);
1871 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
1872 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
1874 brw_MOV(p
, mrf
, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
1876 /* set message header global offset field (reg 0, element 2) */
1878 retype(brw_vec1_reg(BRW_MESSAGE_REGISTER_FILE
,
1880 2), BRW_REGISTER_TYPE_UD
),
1881 brw_imm_ud(offset
));
1883 brw_pop_insn_state(p
);
1887 struct brw_instruction
*insn
= next_insn(p
, BRW_OPCODE_SEND
);
1889 assert(insn
->header
.predicate_control
== 0);
1890 insn
->header
.compression_control
= BRW_COMPRESSION_NONE
;
1891 insn
->header
.destreg__conditionalmod
= mrf
.nr
;
1893 brw_set_dest(p
, insn
, dest
); /* UW? */
1894 if (intel
->gen
>= 6) {
1895 brw_set_src0(p
, insn
, mrf
);
1897 brw_set_src0(p
, insn
, brw_null_reg());
1900 brw_set_dp_read_message(p
,
1902 255, /* binding table index (255=stateless) */
1904 BRW_DATAPORT_READ_MESSAGE_OWORD_BLOCK_READ
, /* msg_type */
1905 BRW_DATAPORT_READ_TARGET_RENDER_CACHE
,
1912 * Read a float[4] vector from the data port Data Cache (const buffer).
1913 * Location (in buffer) should be a multiple of 16.
1914 * Used for fetching shader constants.
1916 void brw_oword_block_read(struct brw_compile
*p
,
1917 struct brw_reg dest
,
1920 uint32_t bind_table_index
)
1922 struct intel_context
*intel
= &p
->brw
->intel
;
1924 /* On newer hardware, offset is in units of owords. */
1925 if (intel
->gen
>= 6)
1928 mrf
= retype(mrf
, BRW_REGISTER_TYPE_UD
);
1930 brw_push_insn_state(p
);
1931 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
1932 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
1933 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
1935 brw_MOV(p
, mrf
, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
1937 /* set message header global offset field (reg 0, element 2) */
1939 retype(brw_vec1_reg(BRW_MESSAGE_REGISTER_FILE
,
1941 2), BRW_REGISTER_TYPE_UD
),
1942 brw_imm_ud(offset
));
1944 struct brw_instruction
*insn
= next_insn(p
, BRW_OPCODE_SEND
);
1945 insn
->header
.destreg__conditionalmod
= mrf
.nr
;
1947 /* cast dest to a uword[8] vector */
1948 dest
= retype(vec8(dest
), BRW_REGISTER_TYPE_UW
);
1950 brw_set_dest(p
, insn
, dest
);
1951 if (intel
->gen
>= 6) {
1952 brw_set_src0(p
, insn
, mrf
);
1954 brw_set_src0(p
, insn
, brw_null_reg());
1957 brw_set_dp_read_message(p
,
1960 BRW_DATAPORT_OWORD_BLOCK_1_OWORDLOW
,
1961 BRW_DATAPORT_READ_MESSAGE_OWORD_BLOCK_READ
,
1962 BRW_DATAPORT_READ_TARGET_DATA_CACHE
,
1964 1); /* response_length (1 reg, 2 owords!) */
1966 brw_pop_insn_state(p
);
1970 * Read a set of dwords from the data port Data Cache (const buffer).
1972 * Location (in buffer) appears as UD offsets in the register after
1973 * the provided mrf header reg.
1975 void brw_dword_scattered_read(struct brw_compile
*p
,
1976 struct brw_reg dest
,
1978 uint32_t bind_table_index
)
1980 mrf
= retype(mrf
, BRW_REGISTER_TYPE_UD
);
1982 brw_push_insn_state(p
);
1983 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
1984 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
1985 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
1986 brw_MOV(p
, mrf
, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
1987 brw_pop_insn_state(p
);
1989 struct brw_instruction
*insn
= next_insn(p
, BRW_OPCODE_SEND
);
1990 insn
->header
.destreg__conditionalmod
= mrf
.nr
;
1992 /* cast dest to a uword[8] vector */
1993 dest
= retype(vec8(dest
), BRW_REGISTER_TYPE_UW
);
1995 brw_set_dest(p
, insn
, dest
);
1996 brw_set_src0(p
, insn
, brw_null_reg());
1998 brw_set_dp_read_message(p
,
2001 BRW_DATAPORT_DWORD_SCATTERED_BLOCK_8DWORDS
,
2002 BRW_DATAPORT_READ_MESSAGE_DWORD_SCATTERED_READ
,
2003 BRW_DATAPORT_READ_TARGET_DATA_CACHE
,
2005 1); /* response_length */
2011 * Read float[4] constant(s) from VS constant buffer.
2012 * For relative addressing, two float[4] constants will be read into 'dest'.
2013 * Otherwise, one float[4] constant will be read into the lower half of 'dest'.
2015 void brw_dp_READ_4_vs(struct brw_compile
*p
,
2016 struct brw_reg dest
,
2018 GLuint bind_table_index
)
2020 struct intel_context
*intel
= &p
->brw
->intel
;
2021 struct brw_instruction
*insn
;
2022 GLuint msg_reg_nr
= 1;
2024 if (intel
->gen
>= 6)
2027 /* Setup MRF[1] with location/offset into const buffer */
2028 brw_push_insn_state(p
);
2029 brw_set_access_mode(p
, BRW_ALIGN_1
);
2030 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
2031 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
2032 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
2033 brw_MOV(p
, retype(brw_vec1_reg(BRW_MESSAGE_REGISTER_FILE
, msg_reg_nr
, 2),
2034 BRW_REGISTER_TYPE_UD
),
2035 brw_imm_ud(location
));
2036 brw_pop_insn_state(p
);
2038 insn
= next_insn(p
, BRW_OPCODE_SEND
);
2040 insn
->header
.predicate_control
= BRW_PREDICATE_NONE
;
2041 insn
->header
.compression_control
= BRW_COMPRESSION_NONE
;
2042 insn
->header
.destreg__conditionalmod
= msg_reg_nr
;
2043 insn
->header
.mask_control
= BRW_MASK_DISABLE
;
2045 brw_set_dest(p
, insn
, dest
);
2046 if (intel
->gen
>= 6) {
2047 brw_set_src0(p
, insn
, brw_message_reg(msg_reg_nr
));
2049 brw_set_src0(p
, insn
, brw_null_reg());
2052 brw_set_dp_read_message(p
,
2056 BRW_DATAPORT_READ_MESSAGE_OWORD_BLOCK_READ
, /* msg_type */
2057 BRW_DATAPORT_READ_TARGET_DATA_CACHE
,
2059 1); /* response_length (1 Oword) */
2063 * Read a float[4] constant per vertex from VS constant buffer, with
2064 * relative addressing.
2066 void brw_dp_READ_4_vs_relative(struct brw_compile
*p
,
2067 struct brw_reg dest
,
2068 struct brw_reg addr_reg
,
2070 GLuint bind_table_index
)
2072 struct intel_context
*intel
= &p
->brw
->intel
;
2073 struct brw_reg src
= brw_vec8_grf(0, 0);
2076 /* Setup MRF[1] with offset into const buffer */
2077 brw_push_insn_state(p
);
2078 brw_set_access_mode(p
, BRW_ALIGN_1
);
2079 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
2080 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
2081 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
2083 /* M1.0 is block offset 0, M1.4 is block offset 1, all other
2086 brw_ADD(p
, retype(brw_message_reg(1), BRW_REGISTER_TYPE_D
),
2087 addr_reg
, brw_imm_d(offset
));
2088 brw_pop_insn_state(p
);
2090 gen6_resolve_implied_move(p
, &src
, 0);
2091 struct brw_instruction
*insn
= next_insn(p
, BRW_OPCODE_SEND
);
2093 insn
->header
.predicate_control
= BRW_PREDICATE_NONE
;
2094 insn
->header
.compression_control
= BRW_COMPRESSION_NONE
;
2095 insn
->header
.destreg__conditionalmod
= 0;
2096 insn
->header
.mask_control
= BRW_MASK_DISABLE
;
2098 brw_set_dest(p
, insn
, dest
);
2099 brw_set_src0(p
, insn
, src
);
2101 if (intel
->gen
>= 6)
2102 msg_type
= GEN6_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ
;
2103 else if (intel
->gen
== 5 || intel
->is_g4x
)
2104 msg_type
= G45_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ
;
2106 msg_type
= BRW_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ
;
2108 brw_set_dp_read_message(p
,
2111 BRW_DATAPORT_OWORD_DUAL_BLOCK_1OWORD
,
2113 BRW_DATAPORT_READ_TARGET_DATA_CACHE
,
2115 1); /* response_length */
2120 void brw_fb_WRITE(struct brw_compile
*p
,
2123 struct brw_reg src0
,
2124 GLuint binding_table_index
,
2126 GLuint response_length
,
2128 bool header_present
)
2130 struct intel_context
*intel
= &p
->brw
->intel
;
2131 struct brw_instruction
*insn
;
2132 GLuint msg_control
, msg_type
;
2133 struct brw_reg dest
;
2135 if (dispatch_width
== 16)
2136 dest
= retype(vec16(brw_null_reg()), BRW_REGISTER_TYPE_UW
);
2138 dest
= retype(vec8(brw_null_reg()), BRW_REGISTER_TYPE_UW
);
2140 if (intel
->gen
>= 6 && binding_table_index
== 0) {
2141 insn
= next_insn(p
, BRW_OPCODE_SENDC
);
2143 insn
= next_insn(p
, BRW_OPCODE_SEND
);
2145 /* The execution mask is ignored for render target writes. */
2146 insn
->header
.predicate_control
= 0;
2147 insn
->header
.compression_control
= BRW_COMPRESSION_NONE
;
2149 if (intel
->gen
>= 6) {
2150 /* headerless version, just submit color payload */
2151 src0
= brw_message_reg(msg_reg_nr
);
2153 msg_type
= GEN6_DATAPORT_WRITE_MESSAGE_RENDER_TARGET_WRITE
;
2155 insn
->header
.destreg__conditionalmod
= msg_reg_nr
;
2157 msg_type
= BRW_DATAPORT_WRITE_MESSAGE_RENDER_TARGET_WRITE
;
2160 if (dispatch_width
== 16)
2161 msg_control
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE
;
2163 msg_control
= BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_SINGLE_SOURCE_SUBSPAN01
;
2165 brw_set_dest(p
, insn
, dest
);
2166 brw_set_src0(p
, insn
, src0
);
2167 brw_set_dp_write_message(p
,
2169 binding_table_index
,
2174 1, /* last render target write */
2177 0 /* send_commit_msg */);
2182 * Texture sample instruction.
2183 * Note: the msg_type plus msg_length values determine exactly what kind
2184 * of sampling operation is performed. See volume 4, page 161 of docs.
2186 void brw_SAMPLE(struct brw_compile
*p
,
2187 struct brw_reg dest
,
2189 struct brw_reg src0
,
2190 GLuint binding_table_index
,
2194 GLuint response_length
,
2196 GLuint header_present
,
2198 GLuint return_format
)
2200 struct intel_context
*intel
= &p
->brw
->intel
;
2201 bool need_stall
= 0;
2203 if (writemask
== 0) {
2204 /*printf("%s: zero writemask??\n", __FUNCTION__); */
2208 /* Hardware doesn't do destination dependency checking on send
2209 * instructions properly. Add a workaround which generates the
2210 * dependency by other means. In practice it seems like this bug
2211 * only crops up for texture samples, and only where registers are
2212 * written by the send and then written again later without being
2213 * read in between. Luckily for us, we already track that
2214 * information and use it to modify the writemask for the
2215 * instruction, so that is a guide for whether a workaround is
2218 if (writemask
!= WRITEMASK_XYZW
) {
2219 GLuint dst_offset
= 0;
2220 GLuint i
, newmask
= 0, len
= 0;
2222 for (i
= 0; i
< 4; i
++) {
2223 if (writemask
& (1<<i
))
2227 for (; i
< 4; i
++) {
2228 if (!(writemask
& (1<<i
)))
2234 if (newmask
!= writemask
) {
2236 /* printf("need stall %x %x\n", newmask , writemask); */
2239 bool dispatch_16
= false;
2241 struct brw_reg m1
= brw_message_reg(msg_reg_nr
);
2243 guess_execution_size(p
, p
->current
, dest
);
2244 if (p
->current
->header
.execution_size
== BRW_EXECUTE_16
)
2247 newmask
= ~newmask
& WRITEMASK_XYZW
;
2249 brw_push_insn_state(p
);
2251 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
2252 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
2254 brw_MOV(p
, retype(m1
, BRW_REGISTER_TYPE_UD
),
2255 retype(brw_vec8_grf(0,0), BRW_REGISTER_TYPE_UD
));
2256 brw_MOV(p
, get_element_ud(m1
, 2), brw_imm_ud(newmask
<< 12));
2258 brw_pop_insn_state(p
);
2260 src0
= retype(brw_null_reg(), BRW_REGISTER_TYPE_UW
);
2261 dest
= offset(dest
, dst_offset
);
2263 /* For 16-wide dispatch, masked channels are skipped in the
2264 * response. For 8-wide, masked channels still take up slots,
2265 * and are just not written to.
2268 response_length
= len
* 2;
2273 struct brw_instruction
*insn
;
2275 gen6_resolve_implied_move(p
, &src0
, msg_reg_nr
);
2277 insn
= next_insn(p
, BRW_OPCODE_SEND
);
2278 insn
->header
.predicate_control
= 0; /* XXX */
2279 insn
->header
.compression_control
= BRW_COMPRESSION_NONE
;
2281 insn
->header
.destreg__conditionalmod
= msg_reg_nr
;
2283 brw_set_dest(p
, insn
, dest
);
2284 brw_set_src0(p
, insn
, src0
);
2285 brw_set_sampler_message(p
, insn
,
2286 binding_table_index
,
2297 struct brw_reg reg
= vec8(offset(dest
, response_length
-1));
2299 /* mov (8) r9.0<1>:f r9.0<8;8,1>:f { Align1 }
2301 brw_push_insn_state(p
);
2302 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
2303 brw_MOV(p
, retype(reg
, BRW_REGISTER_TYPE_UD
),
2304 retype(reg
, BRW_REGISTER_TYPE_UD
));
2305 brw_pop_insn_state(p
);
2310 /* All these variables are pretty confusing - we might be better off
2311 * using bitmasks and macros for this, in the old style. Or perhaps
2312 * just having the caller instantiate the fields in dword3 itself.
2314 void brw_urb_WRITE(struct brw_compile
*p
,
2315 struct brw_reg dest
,
2317 struct brw_reg src0
,
2321 GLuint response_length
,
2323 bool writes_complete
,
2327 struct intel_context
*intel
= &p
->brw
->intel
;
2328 struct brw_instruction
*insn
;
2330 gen6_resolve_implied_move(p
, &src0
, msg_reg_nr
);
2332 if (intel
->gen
== 7) {
2333 /* Enable Channel Masks in the URB_WRITE_HWORD message header */
2334 brw_push_insn_state(p
);
2335 brw_set_access_mode(p
, BRW_ALIGN_1
);
2336 brw_OR(p
, retype(brw_vec1_reg(BRW_MESSAGE_REGISTER_FILE
, msg_reg_nr
, 5),
2337 BRW_REGISTER_TYPE_UD
),
2338 retype(brw_vec1_grf(0, 5), BRW_REGISTER_TYPE_UD
),
2339 brw_imm_ud(0xff00));
2340 brw_pop_insn_state(p
);
2343 insn
= next_insn(p
, BRW_OPCODE_SEND
);
2345 assert(msg_length
< BRW_MAX_MRF
);
2347 brw_set_dest(p
, insn
, dest
);
2348 brw_set_src0(p
, insn
, src0
);
2349 brw_set_src1(p
, insn
, brw_imm_d(0));
2352 insn
->header
.destreg__conditionalmod
= msg_reg_nr
;
2354 brw_set_urb_message(p
,
2367 brw_find_next_block_end(struct brw_compile
*p
, int start
)
2371 for (ip
= start
+ 1; ip
< p
->nr_insn
; ip
++) {
2372 struct brw_instruction
*insn
= &p
->store
[ip
];
2374 switch (insn
->header
.opcode
) {
2375 case BRW_OPCODE_ENDIF
:
2376 case BRW_OPCODE_ELSE
:
2377 case BRW_OPCODE_WHILE
:
2381 assert(!"not reached");
2385 /* There is no DO instruction on gen6, so to find the end of the loop
2386 * we have to see if the loop is jumping back before our start
2390 brw_find_loop_end(struct brw_compile
*p
, int start
)
2392 struct intel_context
*intel
= &p
->brw
->intel
;
2396 for (ip
= start
+ 1; ip
< p
->nr_insn
; ip
++) {
2397 struct brw_instruction
*insn
= &p
->store
[ip
];
2399 if (insn
->header
.opcode
== BRW_OPCODE_WHILE
) {
2400 int jip
= intel
->gen
== 6 ? insn
->bits1
.branch_gen6
.jump_count
2401 : insn
->bits3
.break_cont
.jip
;
2402 if (ip
+ jip
/ br
<= start
)
2406 assert(!"not reached");
2410 /* After program generation, go back and update the UIP and JIP of
2411 * BREAK and CONT instructions to their correct locations.
2414 brw_set_uip_jip(struct brw_compile
*p
)
2416 struct intel_context
*intel
= &p
->brw
->intel
;
2423 for (ip
= 0; ip
< p
->nr_insn
; ip
++) {
2424 struct brw_instruction
*insn
= &p
->store
[ip
];
2426 switch (insn
->header
.opcode
) {
2427 case BRW_OPCODE_BREAK
:
2428 insn
->bits3
.break_cont
.jip
= br
* (brw_find_next_block_end(p
, ip
) - ip
);
2429 /* Gen7 UIP points to WHILE; Gen6 points just after it */
2430 insn
->bits3
.break_cont
.uip
=
2431 br
* (brw_find_loop_end(p
, ip
) - ip
+ (intel
->gen
== 6 ? 1 : 0));
2433 case BRW_OPCODE_CONTINUE
:
2434 insn
->bits3
.break_cont
.jip
= br
* (brw_find_next_block_end(p
, ip
) - ip
);
2435 insn
->bits3
.break_cont
.uip
= br
* (brw_find_loop_end(p
, ip
) - ip
);
2437 assert(insn
->bits3
.break_cont
.uip
!= 0);
2438 assert(insn
->bits3
.break_cont
.jip
!= 0);
2444 void brw_ff_sync(struct brw_compile
*p
,
2445 struct brw_reg dest
,
2447 struct brw_reg src0
,
2449 GLuint response_length
,
2452 struct intel_context
*intel
= &p
->brw
->intel
;
2453 struct brw_instruction
*insn
;
2455 gen6_resolve_implied_move(p
, &src0
, msg_reg_nr
);
2457 insn
= next_insn(p
, BRW_OPCODE_SEND
);
2458 brw_set_dest(p
, insn
, dest
);
2459 brw_set_src0(p
, insn
, src0
);
2460 brw_set_src1(p
, insn
, brw_imm_d(0));
2463 insn
->header
.destreg__conditionalmod
= msg_reg_nr
;
2465 brw_set_ff_sync_message(p
,
2473 * Emit the SEND instruction necessary to generate stream output data on Gen6
2474 * (for transform feedback).
2476 * If send_commit_msg is true, this is the last piece of stream output data
2477 * from this thread, so send the data as a committed write. According to the
2478 * Sandy Bridge PRM (volume 2 part 1, section 4.5.1):
2480 * "Prior to End of Thread with a URB_WRITE, the kernel must ensure all
2481 * writes are complete by sending the final write as a committed write."
2484 brw_svb_write(struct brw_compile
*p
,
2485 struct brw_reg dest
,
2487 struct brw_reg src0
,
2488 GLuint binding_table_index
,
2489 bool send_commit_msg
)
2491 struct brw_instruction
*insn
;
2493 gen6_resolve_implied_move(p
, &src0
, msg_reg_nr
);
2495 insn
= next_insn(p
, BRW_OPCODE_SEND
);
2496 brw_set_dest(p
, insn
, dest
);
2497 brw_set_src0(p
, insn
, src0
);
2498 brw_set_src1(p
, insn
, brw_imm_d(0));
2499 brw_set_dp_write_message(p
, insn
,
2500 binding_table_index
,
2501 0, /* msg_control: ignored */
2502 GEN6_DATAPORT_WRITE_MESSAGE_STREAMED_VB_WRITE
,
2504 true, /* header_present */
2505 0, /* last_render_target: ignored */
2506 send_commit_msg
, /* response_length */
2507 0, /* end_of_thread */
2508 send_commit_msg
); /* send_commit_msg */