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>
33 #include "brw_context.h"
34 #include "brw_defines.h"
37 #include "util/ralloc.h"
40 * Prior to Sandybridge, the SEND instruction accepted non-MRF source
41 * registers, implicitly moving the operand to a message register.
43 * On Sandybridge, this is no longer the case. This function performs the
44 * explicit move; it should be called before emitting a SEND instruction.
47 gen6_resolve_implied_move(struct brw_codegen
*p
,
51 const struct brw_device_info
*devinfo
= p
->devinfo
;
55 if (src
->file
== BRW_MESSAGE_REGISTER_FILE
)
58 if (src
->file
!= BRW_ARCHITECTURE_REGISTER_FILE
|| src
->nr
!= BRW_ARF_NULL
) {
59 brw_push_insn_state(p
);
60 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
61 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
62 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
63 brw_MOV(p
, retype(brw_message_reg(msg_reg_nr
), BRW_REGISTER_TYPE_UD
),
64 retype(*src
, BRW_REGISTER_TYPE_UD
));
65 brw_pop_insn_state(p
);
67 *src
= brw_message_reg(msg_reg_nr
);
71 gen7_convert_mrf_to_grf(struct brw_codegen
*p
, struct brw_reg
*reg
)
73 /* From the Ivybridge PRM, Volume 4 Part 3, page 218 ("send"):
74 * "The send with EOT should use register space R112-R127 for <src>. This is
75 * to enable loading of a new thread into the same slot while the message
76 * with EOT for current thread is pending dispatch."
78 * Since we're pretending to have 16 MRFs anyway, we may as well use the
79 * registers required for messages with EOT.
81 const struct brw_device_info
*devinfo
= p
->devinfo
;
82 if (devinfo
->gen
>= 7 && reg
->file
== BRW_MESSAGE_REGISTER_FILE
) {
83 reg
->file
= BRW_GENERAL_REGISTER_FILE
;
84 reg
->nr
+= GEN7_MRF_HACK_START
;
89 * Convert a brw_reg_type enumeration value into the hardware representation.
91 * The hardware encoding may depend on whether the value is an immediate.
94 brw_reg_type_to_hw_type(const struct brw_device_info
*devinfo
,
95 enum brw_reg_type type
, enum brw_reg_file file
)
97 if (file
== BRW_IMMEDIATE_VALUE
) {
98 static const int imm_hw_types
[] = {
99 [BRW_REGISTER_TYPE_UD
] = BRW_HW_REG_TYPE_UD
,
100 [BRW_REGISTER_TYPE_D
] = BRW_HW_REG_TYPE_D
,
101 [BRW_REGISTER_TYPE_UW
] = BRW_HW_REG_TYPE_UW
,
102 [BRW_REGISTER_TYPE_W
] = BRW_HW_REG_TYPE_W
,
103 [BRW_REGISTER_TYPE_F
] = BRW_HW_REG_TYPE_F
,
104 [BRW_REGISTER_TYPE_UB
] = -1,
105 [BRW_REGISTER_TYPE_B
] = -1,
106 [BRW_REGISTER_TYPE_UV
] = BRW_HW_REG_IMM_TYPE_UV
,
107 [BRW_REGISTER_TYPE_VF
] = BRW_HW_REG_IMM_TYPE_VF
,
108 [BRW_REGISTER_TYPE_V
] = BRW_HW_REG_IMM_TYPE_V
,
109 [BRW_REGISTER_TYPE_DF
] = GEN8_HW_REG_IMM_TYPE_DF
,
110 [BRW_REGISTER_TYPE_HF
] = GEN8_HW_REG_IMM_TYPE_HF
,
111 [BRW_REGISTER_TYPE_UQ
] = GEN8_HW_REG_TYPE_UQ
,
112 [BRW_REGISTER_TYPE_Q
] = GEN8_HW_REG_TYPE_Q
,
114 assert(type
< ARRAY_SIZE(imm_hw_types
));
115 assert(imm_hw_types
[type
] != -1);
116 assert(devinfo
->gen
>= 8 || type
< BRW_REGISTER_TYPE_DF
);
117 return imm_hw_types
[type
];
119 /* Non-immediate registers */
120 static const int hw_types
[] = {
121 [BRW_REGISTER_TYPE_UD
] = BRW_HW_REG_TYPE_UD
,
122 [BRW_REGISTER_TYPE_D
] = BRW_HW_REG_TYPE_D
,
123 [BRW_REGISTER_TYPE_UW
] = BRW_HW_REG_TYPE_UW
,
124 [BRW_REGISTER_TYPE_W
] = BRW_HW_REG_TYPE_W
,
125 [BRW_REGISTER_TYPE_UB
] = BRW_HW_REG_NON_IMM_TYPE_UB
,
126 [BRW_REGISTER_TYPE_B
] = BRW_HW_REG_NON_IMM_TYPE_B
,
127 [BRW_REGISTER_TYPE_F
] = BRW_HW_REG_TYPE_F
,
128 [BRW_REGISTER_TYPE_UV
] = -1,
129 [BRW_REGISTER_TYPE_VF
] = -1,
130 [BRW_REGISTER_TYPE_V
] = -1,
131 [BRW_REGISTER_TYPE_DF
] = GEN7_HW_REG_NON_IMM_TYPE_DF
,
132 [BRW_REGISTER_TYPE_HF
] = GEN8_HW_REG_NON_IMM_TYPE_HF
,
133 [BRW_REGISTER_TYPE_UQ
] = GEN8_HW_REG_TYPE_UQ
,
134 [BRW_REGISTER_TYPE_Q
] = GEN8_HW_REG_TYPE_Q
,
136 assert(type
< ARRAY_SIZE(hw_types
));
137 assert(hw_types
[type
] != -1);
138 assert(devinfo
->gen
>= 7 || type
< BRW_REGISTER_TYPE_DF
);
139 assert(devinfo
->gen
>= 8 || type
< BRW_REGISTER_TYPE_HF
);
140 return hw_types
[type
];
145 brw_set_dest(struct brw_codegen
*p
, brw_inst
*inst
, struct brw_reg dest
)
147 const struct brw_device_info
*devinfo
= p
->devinfo
;
149 if (dest
.file
== BRW_MESSAGE_REGISTER_FILE
)
150 assert((dest
.nr
& ~BRW_MRF_COMPR4
) < BRW_MAX_MRF(devinfo
->gen
));
151 else if (dest
.file
!= BRW_ARCHITECTURE_REGISTER_FILE
)
152 assert(dest
.nr
< 128);
154 gen7_convert_mrf_to_grf(p
, &dest
);
156 brw_inst_set_dst_reg_file(devinfo
, inst
, dest
.file
);
157 brw_inst_set_dst_reg_type(devinfo
, inst
,
158 brw_reg_type_to_hw_type(devinfo
, dest
.type
,
160 brw_inst_set_dst_address_mode(devinfo
, inst
, dest
.address_mode
);
162 if (dest
.address_mode
== BRW_ADDRESS_DIRECT
) {
163 brw_inst_set_dst_da_reg_nr(devinfo
, inst
, dest
.nr
);
165 if (brw_inst_access_mode(devinfo
, inst
) == BRW_ALIGN_1
) {
166 brw_inst_set_dst_da1_subreg_nr(devinfo
, inst
, dest
.subnr
);
167 if (dest
.hstride
== BRW_HORIZONTAL_STRIDE_0
)
168 dest
.hstride
= BRW_HORIZONTAL_STRIDE_1
;
169 brw_inst_set_dst_hstride(devinfo
, inst
, dest
.hstride
);
171 brw_inst_set_dst_da16_subreg_nr(devinfo
, inst
, dest
.subnr
/ 16);
172 brw_inst_set_da16_writemask(devinfo
, inst
, dest
.writemask
);
173 if (dest
.file
== BRW_GENERAL_REGISTER_FILE
||
174 dest
.file
== BRW_MESSAGE_REGISTER_FILE
) {
175 assert(dest
.writemask
!= 0);
177 /* From the Ivybridge PRM, Vol 4, Part 3, Section 5.2.4.1:
178 * Although Dst.HorzStride is a don't care for Align16, HW needs
179 * this to be programmed as "01".
181 brw_inst_set_dst_hstride(devinfo
, inst
, 1);
184 brw_inst_set_dst_ia_subreg_nr(devinfo
, inst
, dest
.subnr
);
186 /* These are different sizes in align1 vs align16:
188 if (brw_inst_access_mode(devinfo
, inst
) == BRW_ALIGN_1
) {
189 brw_inst_set_dst_ia1_addr_imm(devinfo
, inst
,
190 dest
.indirect_offset
);
191 if (dest
.hstride
== BRW_HORIZONTAL_STRIDE_0
)
192 dest
.hstride
= BRW_HORIZONTAL_STRIDE_1
;
193 brw_inst_set_dst_hstride(devinfo
, inst
, dest
.hstride
);
195 brw_inst_set_dst_ia16_addr_imm(devinfo
, inst
,
196 dest
.indirect_offset
);
197 /* even ignored in da16, still need to set as '01' */
198 brw_inst_set_dst_hstride(devinfo
, inst
, 1);
202 /* Generators should set a default exec_size of either 8 (SIMD4x2 or SIMD8)
203 * or 16 (SIMD16), as that's normally correct. However, when dealing with
204 * small registers, we automatically reduce it to match the register size.
206 if (dest
.width
< BRW_EXECUTE_8
)
207 brw_inst_set_exec_size(devinfo
, inst
, dest
.width
);
210 extern int reg_type_size
[];
213 validate_reg(const struct brw_device_info
*devinfo
,
214 brw_inst
*inst
, struct brw_reg reg
)
216 const int hstride_for_reg
[] = {0, 1, 2, 4};
217 const int vstride_for_reg
[] = {0, 1, 2, 4, 8, 16, 32};
218 const int width_for_reg
[] = {1, 2, 4, 8, 16};
219 const int execsize_for_reg
[] = {1, 2, 4, 8, 16, 32};
220 int width
, hstride
, vstride
, execsize
;
222 if (reg
.file
== BRW_IMMEDIATE_VALUE
) {
223 /* 3.3.6: Region Parameters. Restriction: Immediate vectors
224 * mean the destination has to be 128-bit aligned and the
225 * destination horiz stride has to be a word.
227 if (reg
.type
== BRW_REGISTER_TYPE_V
) {
228 assert(hstride_for_reg
[brw_inst_dst_hstride(devinfo
, inst
)] *
229 reg_type_size
[brw_inst_dst_reg_type(devinfo
, inst
)] == 2);
235 if (reg
.file
== BRW_ARCHITECTURE_REGISTER_FILE
&&
236 reg
.file
== BRW_ARF_NULL
)
239 /* From the IVB PRM Vol. 4, Pt. 3, Section 3.3.3.5:
241 * "Swizzling is not allowed when an accumulator is used as an implicit
242 * source or an explicit source in an instruction."
244 if (reg
.file
== BRW_ARCHITECTURE_REGISTER_FILE
&&
245 reg
.nr
== BRW_ARF_ACCUMULATOR
)
246 assert(reg
.swizzle
== BRW_SWIZZLE_XYZW
);
248 assert(reg
.hstride
>= 0 && reg
.hstride
< ARRAY_SIZE(hstride_for_reg
));
249 hstride
= hstride_for_reg
[reg
.hstride
];
251 if (reg
.vstride
== 0xf) {
254 assert(reg
.vstride
>= 0 && reg
.vstride
< ARRAY_SIZE(vstride_for_reg
));
255 vstride
= vstride_for_reg
[reg
.vstride
];
258 assert(reg
.width
>= 0 && reg
.width
< ARRAY_SIZE(width_for_reg
));
259 width
= width_for_reg
[reg
.width
];
261 assert(brw_inst_exec_size(devinfo
, inst
) >= 0 &&
262 brw_inst_exec_size(devinfo
, inst
) < ARRAY_SIZE(execsize_for_reg
));
263 execsize
= execsize_for_reg
[brw_inst_exec_size(devinfo
, inst
)];
265 /* Restrictions from 3.3.10: Register Region Restrictions. */
267 assert(execsize
>= width
);
270 if (execsize
== width
&& hstride
!= 0) {
271 assert(vstride
== -1 || vstride
== width
* hstride
);
275 if (execsize
== width
&& hstride
== 0) {
276 /* no restriction on vstride. */
281 assert(hstride
== 0);
285 if (execsize
== 1 && width
== 1) {
286 assert(hstride
== 0);
287 assert(vstride
== 0);
291 if (vstride
== 0 && hstride
== 0) {
295 /* 10. Check destination issues. */
299 is_compactable_immediate(unsigned imm
)
301 /* We get the low 12 bits as-is. */
304 /* We get one bit replicated through the top 20 bits. */
305 return imm
== 0 || imm
== 0xfffff000;
309 brw_set_src0(struct brw_codegen
*p
, brw_inst
*inst
, struct brw_reg reg
)
311 const struct brw_device_info
*devinfo
= p
->devinfo
;
313 if (reg
.file
== BRW_MESSAGE_REGISTER_FILE
)
314 assert((reg
.nr
& ~BRW_MRF_COMPR4
) < BRW_MAX_MRF(devinfo
->gen
));
315 else if (reg
.file
!= BRW_ARCHITECTURE_REGISTER_FILE
)
316 assert(reg
.nr
< 128);
318 gen7_convert_mrf_to_grf(p
, ®
);
320 if (devinfo
->gen
>= 6 && (brw_inst_opcode(devinfo
, inst
) == BRW_OPCODE_SEND
||
321 brw_inst_opcode(devinfo
, inst
) == BRW_OPCODE_SENDC
)) {
322 /* Any source modifiers or regions will be ignored, since this just
323 * identifies the MRF/GRF to start reading the message contents from.
324 * Check for some likely failures.
328 assert(reg
.address_mode
== BRW_ADDRESS_DIRECT
);
331 validate_reg(devinfo
, inst
, reg
);
333 brw_inst_set_src0_reg_file(devinfo
, inst
, reg
.file
);
334 brw_inst_set_src0_reg_type(devinfo
, inst
,
335 brw_reg_type_to_hw_type(devinfo
, reg
.type
, reg
.file
));
336 brw_inst_set_src0_abs(devinfo
, inst
, reg
.abs
);
337 brw_inst_set_src0_negate(devinfo
, inst
, reg
.negate
);
338 brw_inst_set_src0_address_mode(devinfo
, inst
, reg
.address_mode
);
340 if (reg
.file
== BRW_IMMEDIATE_VALUE
) {
341 brw_inst_set_imm_ud(devinfo
, inst
, reg
.ud
);
343 /* The Bspec's section titled "Non-present Operands" claims that if src0
344 * is an immediate that src1's type must be the same as that of src0.
346 * The SNB+ DataTypeIndex instruction compaction tables contain mappings
347 * that do not follow this rule. E.g., from the IVB/HSW table:
349 * DataTypeIndex 18-Bit Mapping Mapped Meaning
350 * 3 001000001011111101 r:f | i:vf | a:ud | <1> | dir |
352 * And from the SNB table:
354 * DataTypeIndex 18-Bit Mapping Mapped Meaning
355 * 8 001000000111101100 a:w | i:w | a:ud | <1> | dir |
357 * Neither of these cause warnings from the simulator when used,
358 * compacted or otherwise. In fact, all compaction mappings that have an
359 * immediate in src0 use a:ud for src1.
361 * The GM45 instruction compaction tables do not contain mapped meanings
362 * so it's not clear whether it has the restriction. We'll assume it was
363 * lifted on SNB. (FINISHME: decode the GM45 tables and check.)
365 brw_inst_set_src1_reg_file(devinfo
, inst
, BRW_ARCHITECTURE_REGISTER_FILE
);
366 if (devinfo
->gen
< 6) {
367 brw_inst_set_src1_reg_type(devinfo
, inst
,
368 brw_inst_src0_reg_type(devinfo
, inst
));
370 brw_inst_set_src1_reg_type(devinfo
, inst
, BRW_HW_REG_TYPE_UD
);
373 /* Compacted instructions only have 12-bits (plus 1 for the other 20)
374 * for immediate values. Presumably the hardware engineers realized
375 * that the only useful floating-point value that could be represented
376 * in this format is 0.0, which can also be represented as a VF-typed
377 * immediate, so they gave us the previously mentioned mapping on IVB+.
379 * Strangely, we do have a mapping for imm:f in src1, so we don't need
382 * If we see a 0.0:F, change the type to VF so that it can be compacted.
384 if (brw_inst_imm_ud(devinfo
, inst
) == 0x0 &&
385 brw_inst_src0_reg_type(devinfo
, inst
) == BRW_HW_REG_TYPE_F
) {
386 brw_inst_set_src0_reg_type(devinfo
, inst
, BRW_HW_REG_IMM_TYPE_VF
);
389 /* There are no mappings for dst:d | i:d, so if the immediate is suitable
390 * set the types to :UD so the instruction can be compacted.
392 if (is_compactable_immediate(brw_inst_imm_ud(devinfo
, inst
)) &&
393 brw_inst_cond_modifier(devinfo
, inst
) == BRW_CONDITIONAL_NONE
&&
394 brw_inst_src0_reg_type(devinfo
, inst
) == BRW_HW_REG_TYPE_D
&&
395 brw_inst_dst_reg_type(devinfo
, inst
) == BRW_HW_REG_TYPE_D
) {
396 brw_inst_set_src0_reg_type(devinfo
, inst
, BRW_HW_REG_TYPE_UD
);
397 brw_inst_set_dst_reg_type(devinfo
, inst
, BRW_HW_REG_TYPE_UD
);
400 if (reg
.address_mode
== BRW_ADDRESS_DIRECT
) {
401 brw_inst_set_src0_da_reg_nr(devinfo
, inst
, reg
.nr
);
402 if (brw_inst_access_mode(devinfo
, inst
) == BRW_ALIGN_1
) {
403 brw_inst_set_src0_da1_subreg_nr(devinfo
, inst
, reg
.subnr
);
405 brw_inst_set_src0_da16_subreg_nr(devinfo
, inst
, reg
.subnr
/ 16);
408 brw_inst_set_src0_ia_subreg_nr(devinfo
, inst
, reg
.subnr
);
410 if (brw_inst_access_mode(devinfo
, inst
) == BRW_ALIGN_1
) {
411 brw_inst_set_src0_ia1_addr_imm(devinfo
, inst
, reg
.indirect_offset
);
413 brw_inst_set_src0_ia16_addr_imm(devinfo
, inst
, reg
.indirect_offset
);
417 if (brw_inst_access_mode(devinfo
, inst
) == BRW_ALIGN_1
) {
418 if (reg
.width
== BRW_WIDTH_1
&&
419 brw_inst_exec_size(devinfo
, inst
) == BRW_EXECUTE_1
) {
420 brw_inst_set_src0_hstride(devinfo
, inst
, BRW_HORIZONTAL_STRIDE_0
);
421 brw_inst_set_src0_width(devinfo
, inst
, BRW_WIDTH_1
);
422 brw_inst_set_src0_vstride(devinfo
, inst
, BRW_VERTICAL_STRIDE_0
);
424 brw_inst_set_src0_hstride(devinfo
, inst
, reg
.hstride
);
425 brw_inst_set_src0_width(devinfo
, inst
, reg
.width
);
426 brw_inst_set_src0_vstride(devinfo
, inst
, reg
.vstride
);
429 brw_inst_set_src0_da16_swiz_x(devinfo
, inst
,
430 BRW_GET_SWZ(reg
.swizzle
, BRW_CHANNEL_X
));
431 brw_inst_set_src0_da16_swiz_y(devinfo
, inst
,
432 BRW_GET_SWZ(reg
.swizzle
, BRW_CHANNEL_Y
));
433 brw_inst_set_src0_da16_swiz_z(devinfo
, inst
,
434 BRW_GET_SWZ(reg
.swizzle
, BRW_CHANNEL_Z
));
435 brw_inst_set_src0_da16_swiz_w(devinfo
, inst
,
436 BRW_GET_SWZ(reg
.swizzle
, BRW_CHANNEL_W
));
438 /* This is an oddity of the fact we're using the same
439 * descriptions for registers in align_16 as align_1:
441 if (reg
.vstride
== BRW_VERTICAL_STRIDE_8
)
442 brw_inst_set_src0_vstride(devinfo
, inst
, BRW_VERTICAL_STRIDE_4
);
444 brw_inst_set_src0_vstride(devinfo
, inst
, reg
.vstride
);
451 brw_set_src1(struct brw_codegen
*p
, brw_inst
*inst
, struct brw_reg reg
)
453 const struct brw_device_info
*devinfo
= p
->devinfo
;
455 if (reg
.file
!= BRW_ARCHITECTURE_REGISTER_FILE
)
456 assert(reg
.nr
< 128);
458 /* From the IVB PRM Vol. 4, Pt. 3, Section 3.3.3.5:
460 * "Accumulator registers may be accessed explicitly as src0
463 assert(reg
.file
!= BRW_ARCHITECTURE_REGISTER_FILE
||
464 reg
.nr
!= BRW_ARF_ACCUMULATOR
);
466 gen7_convert_mrf_to_grf(p
, ®
);
467 assert(reg
.file
!= BRW_MESSAGE_REGISTER_FILE
);
469 validate_reg(devinfo
, inst
, reg
);
471 brw_inst_set_src1_reg_file(devinfo
, inst
, reg
.file
);
472 brw_inst_set_src1_reg_type(devinfo
, inst
,
473 brw_reg_type_to_hw_type(devinfo
, reg
.type
, reg
.file
));
474 brw_inst_set_src1_abs(devinfo
, inst
, reg
.abs
);
475 brw_inst_set_src1_negate(devinfo
, inst
, reg
.negate
);
477 /* Only src1 can be immediate in two-argument instructions.
479 assert(brw_inst_src0_reg_file(devinfo
, inst
) != BRW_IMMEDIATE_VALUE
);
481 if (reg
.file
== BRW_IMMEDIATE_VALUE
) {
482 brw_inst_set_imm_ud(devinfo
, inst
, reg
.ud
);
484 /* This is a hardware restriction, which may or may not be lifted
487 assert (reg
.address_mode
== BRW_ADDRESS_DIRECT
);
488 /* assert (reg.file == BRW_GENERAL_REGISTER_FILE); */
490 brw_inst_set_src1_da_reg_nr(devinfo
, inst
, reg
.nr
);
491 if (brw_inst_access_mode(devinfo
, inst
) == BRW_ALIGN_1
) {
492 brw_inst_set_src1_da1_subreg_nr(devinfo
, inst
, reg
.subnr
);
494 brw_inst_set_src1_da16_subreg_nr(devinfo
, inst
, reg
.subnr
/ 16);
497 if (brw_inst_access_mode(devinfo
, inst
) == BRW_ALIGN_1
) {
498 if (reg
.width
== BRW_WIDTH_1
&&
499 brw_inst_exec_size(devinfo
, inst
) == BRW_EXECUTE_1
) {
500 brw_inst_set_src1_hstride(devinfo
, inst
, BRW_HORIZONTAL_STRIDE_0
);
501 brw_inst_set_src1_width(devinfo
, inst
, BRW_WIDTH_1
);
502 brw_inst_set_src1_vstride(devinfo
, inst
, BRW_VERTICAL_STRIDE_0
);
504 brw_inst_set_src1_hstride(devinfo
, inst
, reg
.hstride
);
505 brw_inst_set_src1_width(devinfo
, inst
, reg
.width
);
506 brw_inst_set_src1_vstride(devinfo
, inst
, reg
.vstride
);
509 brw_inst_set_src1_da16_swiz_x(devinfo
, inst
,
510 BRW_GET_SWZ(reg
.swizzle
, BRW_CHANNEL_X
));
511 brw_inst_set_src1_da16_swiz_y(devinfo
, inst
,
512 BRW_GET_SWZ(reg
.swizzle
, BRW_CHANNEL_Y
));
513 brw_inst_set_src1_da16_swiz_z(devinfo
, inst
,
514 BRW_GET_SWZ(reg
.swizzle
, BRW_CHANNEL_Z
));
515 brw_inst_set_src1_da16_swiz_w(devinfo
, inst
,
516 BRW_GET_SWZ(reg
.swizzle
, BRW_CHANNEL_W
));
518 /* This is an oddity of the fact we're using the same
519 * descriptions for registers in align_16 as align_1:
521 if (reg
.vstride
== BRW_VERTICAL_STRIDE_8
)
522 brw_inst_set_src1_vstride(devinfo
, inst
, BRW_VERTICAL_STRIDE_4
);
524 brw_inst_set_src1_vstride(devinfo
, inst
, reg
.vstride
);
530 * Set the Message Descriptor and Extended Message Descriptor fields
533 * \note This zeroes out the Function Control bits, so it must be called
534 * \b before filling out any message-specific data. Callers can
535 * choose not to fill in irrelevant bits; they will be zero.
538 brw_set_message_descriptor(struct brw_codegen
*p
,
540 enum brw_message_target sfid
,
542 unsigned response_length
,
546 const struct brw_device_info
*devinfo
= p
->devinfo
;
548 brw_set_src1(p
, inst
, brw_imm_d(0));
550 /* For indirect sends, `inst` will not be the SEND/SENDC instruction
551 * itself; instead, it will be a MOV/OR into the address register.
553 * In this case, we avoid setting the extended message descriptor bits,
554 * since they go on the later SEND/SENDC instead and if set here would
555 * instead clobber the conditionalmod bits.
557 unsigned opcode
= brw_inst_opcode(devinfo
, inst
);
558 if (opcode
== BRW_OPCODE_SEND
|| opcode
== BRW_OPCODE_SENDC
) {
559 brw_inst_set_sfid(devinfo
, inst
, sfid
);
562 brw_inst_set_mlen(devinfo
, inst
, msg_length
);
563 brw_inst_set_rlen(devinfo
, inst
, response_length
);
564 brw_inst_set_eot(devinfo
, inst
, end_of_thread
);
566 if (devinfo
->gen
>= 5) {
567 brw_inst_set_header_present(devinfo
, inst
, header_present
);
571 static void brw_set_math_message( struct brw_codegen
*p
,
574 unsigned integer_type
,
578 const struct brw_device_info
*devinfo
= p
->devinfo
;
580 unsigned response_length
;
582 /* Infer message length from the function */
584 case BRW_MATH_FUNCTION_POW
:
585 case BRW_MATH_FUNCTION_INT_DIV_QUOTIENT
:
586 case BRW_MATH_FUNCTION_INT_DIV_REMAINDER
:
587 case BRW_MATH_FUNCTION_INT_DIV_QUOTIENT_AND_REMAINDER
:
595 /* Infer response length from the function */
597 case BRW_MATH_FUNCTION_SINCOS
:
598 case BRW_MATH_FUNCTION_INT_DIV_QUOTIENT_AND_REMAINDER
:
607 brw_set_message_descriptor(p
, inst
, BRW_SFID_MATH
,
608 msg_length
, response_length
, false, false);
609 brw_inst_set_math_msg_function(devinfo
, inst
, function
);
610 brw_inst_set_math_msg_signed_int(devinfo
, inst
, integer_type
);
611 brw_inst_set_math_msg_precision(devinfo
, inst
, low_precision
);
612 brw_inst_set_math_msg_saturate(devinfo
, inst
, brw_inst_saturate(devinfo
, inst
));
613 brw_inst_set_math_msg_data_type(devinfo
, inst
, dataType
);
614 brw_inst_set_saturate(devinfo
, inst
, 0);
618 static void brw_set_ff_sync_message(struct brw_codegen
*p
,
621 unsigned response_length
,
624 const struct brw_device_info
*devinfo
= p
->devinfo
;
626 brw_set_message_descriptor(p
, insn
, BRW_SFID_URB
,
627 1, response_length
, true, end_of_thread
);
628 brw_inst_set_urb_opcode(devinfo
, insn
, 1); /* FF_SYNC */
629 brw_inst_set_urb_allocate(devinfo
, insn
, allocate
);
630 /* The following fields are not used by FF_SYNC: */
631 brw_inst_set_urb_global_offset(devinfo
, insn
, 0);
632 brw_inst_set_urb_swizzle_control(devinfo
, insn
, 0);
633 brw_inst_set_urb_used(devinfo
, insn
, 0);
634 brw_inst_set_urb_complete(devinfo
, insn
, 0);
637 static void brw_set_urb_message( struct brw_codegen
*p
,
639 enum brw_urb_write_flags flags
,
641 unsigned response_length
,
643 unsigned swizzle_control
)
645 const struct brw_device_info
*devinfo
= p
->devinfo
;
647 assert(devinfo
->gen
< 7 || swizzle_control
!= BRW_URB_SWIZZLE_TRANSPOSE
);
648 assert(devinfo
->gen
< 7 || !(flags
& BRW_URB_WRITE_ALLOCATE
));
649 assert(devinfo
->gen
>= 7 || !(flags
& BRW_URB_WRITE_PER_SLOT_OFFSET
));
651 brw_set_message_descriptor(p
, insn
, BRW_SFID_URB
,
652 msg_length
, response_length
, true,
653 flags
& BRW_URB_WRITE_EOT
);
655 if (flags
& BRW_URB_WRITE_OWORD
) {
656 assert(msg_length
== 2); /* header + one OWORD of data */
657 brw_inst_set_urb_opcode(devinfo
, insn
, BRW_URB_OPCODE_WRITE_OWORD
);
659 brw_inst_set_urb_opcode(devinfo
, insn
, BRW_URB_OPCODE_WRITE_HWORD
);
662 brw_inst_set_urb_global_offset(devinfo
, insn
, offset
);
663 brw_inst_set_urb_swizzle_control(devinfo
, insn
, swizzle_control
);
665 if (devinfo
->gen
< 8) {
666 brw_inst_set_urb_complete(devinfo
, insn
, !!(flags
& BRW_URB_WRITE_COMPLETE
));
669 if (devinfo
->gen
< 7) {
670 brw_inst_set_urb_allocate(devinfo
, insn
, !!(flags
& BRW_URB_WRITE_ALLOCATE
));
671 brw_inst_set_urb_used(devinfo
, insn
, !(flags
& BRW_URB_WRITE_UNUSED
));
673 brw_inst_set_urb_per_slot_offset(devinfo
, insn
,
674 !!(flags
& BRW_URB_WRITE_PER_SLOT_OFFSET
));
679 brw_set_dp_write_message(struct brw_codegen
*p
,
681 unsigned binding_table_index
,
682 unsigned msg_control
,
686 unsigned last_render_target
,
687 unsigned response_length
,
688 unsigned end_of_thread
,
689 unsigned send_commit_msg
)
691 const struct brw_device_info
*devinfo
= p
->devinfo
;
694 if (devinfo
->gen
>= 7) {
695 /* Use the Render Cache for RT writes; otherwise use the Data Cache */
696 if (msg_type
== GEN6_DATAPORT_WRITE_MESSAGE_RENDER_TARGET_WRITE
)
697 sfid
= GEN6_SFID_DATAPORT_RENDER_CACHE
;
699 sfid
= GEN7_SFID_DATAPORT_DATA_CACHE
;
700 } else if (devinfo
->gen
== 6) {
701 /* Use the render cache for all write messages. */
702 sfid
= GEN6_SFID_DATAPORT_RENDER_CACHE
;
704 sfid
= BRW_SFID_DATAPORT_WRITE
;
707 brw_set_message_descriptor(p
, insn
, sfid
, msg_length
, response_length
,
708 header_present
, end_of_thread
);
710 brw_inst_set_binding_table_index(devinfo
, insn
, binding_table_index
);
711 brw_inst_set_dp_write_msg_type(devinfo
, insn
, msg_type
);
712 brw_inst_set_dp_write_msg_control(devinfo
, insn
, msg_control
);
713 brw_inst_set_rt_last(devinfo
, insn
, last_render_target
);
714 if (devinfo
->gen
< 7) {
715 brw_inst_set_dp_write_commit(devinfo
, insn
, send_commit_msg
);
720 brw_set_dp_read_message(struct brw_codegen
*p
,
722 unsigned binding_table_index
,
723 unsigned msg_control
,
725 unsigned target_cache
,
728 unsigned response_length
)
730 const struct brw_device_info
*devinfo
= p
->devinfo
;
733 if (devinfo
->gen
>= 7) {
734 sfid
= GEN7_SFID_DATAPORT_DATA_CACHE
;
735 } else if (devinfo
->gen
== 6) {
736 if (target_cache
== BRW_DATAPORT_READ_TARGET_RENDER_CACHE
)
737 sfid
= GEN6_SFID_DATAPORT_RENDER_CACHE
;
739 sfid
= GEN6_SFID_DATAPORT_SAMPLER_CACHE
;
741 sfid
= BRW_SFID_DATAPORT_READ
;
744 brw_set_message_descriptor(p
, insn
, sfid
, msg_length
, response_length
,
745 header_present
, false);
747 brw_inst_set_binding_table_index(devinfo
, insn
, binding_table_index
);
748 brw_inst_set_dp_read_msg_type(devinfo
, insn
, msg_type
);
749 brw_inst_set_dp_read_msg_control(devinfo
, insn
, msg_control
);
750 if (devinfo
->gen
< 6)
751 brw_inst_set_dp_read_target_cache(devinfo
, insn
, target_cache
);
755 brw_set_sampler_message(struct brw_codegen
*p
,
757 unsigned binding_table_index
,
760 unsigned response_length
,
762 unsigned header_present
,
764 unsigned return_format
)
766 const struct brw_device_info
*devinfo
= p
->devinfo
;
768 brw_set_message_descriptor(p
, inst
, BRW_SFID_SAMPLER
, msg_length
,
769 response_length
, header_present
, false);
771 brw_inst_set_binding_table_index(devinfo
, inst
, binding_table_index
);
772 brw_inst_set_sampler(devinfo
, inst
, sampler
);
773 brw_inst_set_sampler_msg_type(devinfo
, inst
, msg_type
);
774 if (devinfo
->gen
>= 5) {
775 brw_inst_set_sampler_simd_mode(devinfo
, inst
, simd_mode
);
776 } else if (devinfo
->gen
== 4 && !devinfo
->is_g4x
) {
777 brw_inst_set_sampler_return_format(devinfo
, inst
, return_format
);
782 gen7_set_dp_scratch_message(struct brw_codegen
*p
,
786 bool invalidate_after_read
,
788 unsigned addr_offset
,
793 const struct brw_device_info
*devinfo
= p
->devinfo
;
794 assert(num_regs
== 1 || num_regs
== 2 || num_regs
== 4 ||
795 (devinfo
->gen
>= 8 && num_regs
== 8));
796 brw_set_message_descriptor(p
, inst
, GEN7_SFID_DATAPORT_DATA_CACHE
,
797 mlen
, rlen
, header_present
, false);
798 brw_inst_set_dp_category(devinfo
, inst
, 1); /* Scratch Block Read/Write msgs */
799 brw_inst_set_scratch_read_write(devinfo
, inst
, write
);
800 brw_inst_set_scratch_type(devinfo
, inst
, dword
);
801 brw_inst_set_scratch_invalidate_after_read(devinfo
, inst
, invalidate_after_read
);
802 brw_inst_set_scratch_block_size(devinfo
, inst
, ffs(num_regs
) - 1);
803 brw_inst_set_scratch_addr_offset(devinfo
, inst
, addr_offset
);
806 #define next_insn brw_next_insn
808 brw_next_insn(struct brw_codegen
*p
, unsigned opcode
)
810 const struct brw_device_info
*devinfo
= p
->devinfo
;
813 if (p
->nr_insn
+ 1 > p
->store_size
) {
815 p
->store
= reralloc(p
->mem_ctx
, p
->store
, brw_inst
, p
->store_size
);
818 p
->next_insn_offset
+= 16;
819 insn
= &p
->store
[p
->nr_insn
++];
820 memcpy(insn
, p
->current
, sizeof(*insn
));
822 brw_inst_set_opcode(devinfo
, insn
, opcode
);
827 brw_alu1(struct brw_codegen
*p
, unsigned opcode
,
828 struct brw_reg dest
, struct brw_reg src
)
830 brw_inst
*insn
= next_insn(p
, opcode
);
831 brw_set_dest(p
, insn
, dest
);
832 brw_set_src0(p
, insn
, src
);
837 brw_alu2(struct brw_codegen
*p
, unsigned opcode
,
838 struct brw_reg dest
, struct brw_reg src0
, struct brw_reg src1
)
840 brw_inst
*insn
= next_insn(p
, opcode
);
841 brw_set_dest(p
, insn
, dest
);
842 brw_set_src0(p
, insn
, src0
);
843 brw_set_src1(p
, insn
, src1
);
848 get_3src_subreg_nr(struct brw_reg reg
)
850 if (reg
.vstride
== BRW_VERTICAL_STRIDE_0
) {
851 assert(brw_is_single_value_swizzle(reg
.swizzle
));
852 return reg
.subnr
/ 4 + BRW_GET_SWZ(reg
.swizzle
, 0);
854 return reg
.subnr
/ 4;
859 brw_alu3(struct brw_codegen
*p
, unsigned opcode
, struct brw_reg dest
,
860 struct brw_reg src0
, struct brw_reg src1
, struct brw_reg src2
)
862 const struct brw_device_info
*devinfo
= p
->devinfo
;
863 brw_inst
*inst
= next_insn(p
, opcode
);
865 gen7_convert_mrf_to_grf(p
, &dest
);
867 assert(brw_inst_access_mode(devinfo
, inst
) == BRW_ALIGN_16
);
869 assert(dest
.file
== BRW_GENERAL_REGISTER_FILE
||
870 dest
.file
== BRW_MESSAGE_REGISTER_FILE
);
871 assert(dest
.nr
< 128);
872 assert(dest
.address_mode
== BRW_ADDRESS_DIRECT
);
873 assert(dest
.type
== BRW_REGISTER_TYPE_F
||
874 dest
.type
== BRW_REGISTER_TYPE_D
||
875 dest
.type
== BRW_REGISTER_TYPE_UD
);
876 if (devinfo
->gen
== 6) {
877 brw_inst_set_3src_dst_reg_file(devinfo
, inst
,
878 dest
.file
== BRW_MESSAGE_REGISTER_FILE
);
880 brw_inst_set_3src_dst_reg_nr(devinfo
, inst
, dest
.nr
);
881 brw_inst_set_3src_dst_subreg_nr(devinfo
, inst
, dest
.subnr
/ 16);
882 brw_inst_set_3src_dst_writemask(devinfo
, inst
, dest
.writemask
);
884 assert(src0
.file
== BRW_GENERAL_REGISTER_FILE
);
885 assert(src0
.address_mode
== BRW_ADDRESS_DIRECT
);
886 assert(src0
.nr
< 128);
887 brw_inst_set_3src_src0_swizzle(devinfo
, inst
, src0
.swizzle
);
888 brw_inst_set_3src_src0_subreg_nr(devinfo
, inst
, get_3src_subreg_nr(src0
));
889 brw_inst_set_3src_src0_reg_nr(devinfo
, inst
, src0
.nr
);
890 brw_inst_set_3src_src0_abs(devinfo
, inst
, src0
.abs
);
891 brw_inst_set_3src_src0_negate(devinfo
, inst
, src0
.negate
);
892 brw_inst_set_3src_src0_rep_ctrl(devinfo
, inst
,
893 src0
.vstride
== BRW_VERTICAL_STRIDE_0
);
895 assert(src1
.file
== BRW_GENERAL_REGISTER_FILE
);
896 assert(src1
.address_mode
== BRW_ADDRESS_DIRECT
);
897 assert(src1
.nr
< 128);
898 brw_inst_set_3src_src1_swizzle(devinfo
, inst
, src1
.swizzle
);
899 brw_inst_set_3src_src1_subreg_nr(devinfo
, inst
, get_3src_subreg_nr(src1
));
900 brw_inst_set_3src_src1_reg_nr(devinfo
, inst
, src1
.nr
);
901 brw_inst_set_3src_src1_abs(devinfo
, inst
, src1
.abs
);
902 brw_inst_set_3src_src1_negate(devinfo
, inst
, src1
.negate
);
903 brw_inst_set_3src_src1_rep_ctrl(devinfo
, inst
,
904 src1
.vstride
== BRW_VERTICAL_STRIDE_0
);
906 assert(src2
.file
== BRW_GENERAL_REGISTER_FILE
);
907 assert(src2
.address_mode
== BRW_ADDRESS_DIRECT
);
908 assert(src2
.nr
< 128);
909 brw_inst_set_3src_src2_swizzle(devinfo
, inst
, src2
.swizzle
);
910 brw_inst_set_3src_src2_subreg_nr(devinfo
, inst
, get_3src_subreg_nr(src2
));
911 brw_inst_set_3src_src2_reg_nr(devinfo
, inst
, src2
.nr
);
912 brw_inst_set_3src_src2_abs(devinfo
, inst
, src2
.abs
);
913 brw_inst_set_3src_src2_negate(devinfo
, inst
, src2
.negate
);
914 brw_inst_set_3src_src2_rep_ctrl(devinfo
, inst
,
915 src2
.vstride
== BRW_VERTICAL_STRIDE_0
);
917 if (devinfo
->gen
>= 7) {
918 /* Set both the source and destination types based on dest.type,
919 * ignoring the source register types. The MAD and LRP emitters ensure
920 * that all four types are float. The BFE and BFI2 emitters, however,
921 * may send us mixed D and UD types and want us to ignore that and use
922 * the destination type.
925 case BRW_REGISTER_TYPE_F
:
926 brw_inst_set_3src_src_type(devinfo
, inst
, BRW_3SRC_TYPE_F
);
927 brw_inst_set_3src_dst_type(devinfo
, inst
, BRW_3SRC_TYPE_F
);
929 case BRW_REGISTER_TYPE_D
:
930 brw_inst_set_3src_src_type(devinfo
, inst
, BRW_3SRC_TYPE_D
);
931 brw_inst_set_3src_dst_type(devinfo
, inst
, BRW_3SRC_TYPE_D
);
933 case BRW_REGISTER_TYPE_UD
:
934 brw_inst_set_3src_src_type(devinfo
, inst
, BRW_3SRC_TYPE_UD
);
935 brw_inst_set_3src_dst_type(devinfo
, inst
, BRW_3SRC_TYPE_UD
);
938 unreachable("not reached");
946 /***********************************************************************
947 * Convenience routines.
950 brw_inst *brw_##OP(struct brw_codegen *p, \
951 struct brw_reg dest, \
952 struct brw_reg src0) \
954 return brw_alu1(p, BRW_OPCODE_##OP, dest, src0); \
958 brw_inst *brw_##OP(struct brw_codegen *p, \
959 struct brw_reg dest, \
960 struct brw_reg src0, \
961 struct brw_reg src1) \
963 return brw_alu2(p, BRW_OPCODE_##OP, dest, src0, src1); \
967 brw_inst *brw_##OP(struct brw_codegen *p, \
968 struct brw_reg dest, \
969 struct brw_reg src0, \
970 struct brw_reg src1, \
971 struct brw_reg src2) \
973 return brw_alu3(p, BRW_OPCODE_##OP, dest, src0, src1, src2); \
977 brw_inst *brw_##OP(struct brw_codegen *p, \
978 struct brw_reg dest, \
979 struct brw_reg src0, \
980 struct brw_reg src1, \
981 struct brw_reg src2) \
983 assert(dest.type == BRW_REGISTER_TYPE_F); \
984 assert(src0.type == BRW_REGISTER_TYPE_F); \
985 assert(src1.type == BRW_REGISTER_TYPE_F); \
986 assert(src2.type == BRW_REGISTER_TYPE_F); \
987 return brw_alu3(p, BRW_OPCODE_##OP, dest, src0, src1, src2); \
990 /* Rounding operations (other than RNDD) require two instructions - the first
991 * stores a rounded value (possibly the wrong way) in the dest register, but
992 * also sets a per-channel "increment bit" in the flag register. A predicated
993 * add of 1.0 fixes dest to contain the desired result.
995 * Sandybridge and later appear to round correctly without an ADD.
998 void brw_##OP(struct brw_codegen *p, \
999 struct brw_reg dest, \
1000 struct brw_reg src) \
1002 const struct brw_device_info *devinfo = p->devinfo; \
1003 brw_inst *rnd, *add; \
1004 rnd = next_insn(p, BRW_OPCODE_##OP); \
1005 brw_set_dest(p, rnd, dest); \
1006 brw_set_src0(p, rnd, src); \
1008 if (devinfo->gen < 6) { \
1009 /* turn on round-increments */ \
1010 brw_inst_set_cond_modifier(devinfo, rnd, BRW_CONDITIONAL_R); \
1011 add = brw_ADD(p, dest, dest, brw_imm_f(1.0f)); \
1012 brw_inst_set_pred_control(devinfo, add, BRW_PREDICATE_NORMAL); \
1052 brw_ADD(struct brw_codegen
*p
, struct brw_reg dest
,
1053 struct brw_reg src0
, struct brw_reg src1
)
1056 if (src0
.type
== BRW_REGISTER_TYPE_F
||
1057 (src0
.file
== BRW_IMMEDIATE_VALUE
&&
1058 src0
.type
== BRW_REGISTER_TYPE_VF
)) {
1059 assert(src1
.type
!= BRW_REGISTER_TYPE_UD
);
1060 assert(src1
.type
!= BRW_REGISTER_TYPE_D
);
1063 if (src1
.type
== BRW_REGISTER_TYPE_F
||
1064 (src1
.file
== BRW_IMMEDIATE_VALUE
&&
1065 src1
.type
== BRW_REGISTER_TYPE_VF
)) {
1066 assert(src0
.type
!= BRW_REGISTER_TYPE_UD
);
1067 assert(src0
.type
!= BRW_REGISTER_TYPE_D
);
1070 return brw_alu2(p
, BRW_OPCODE_ADD
, dest
, src0
, src1
);
1074 brw_AVG(struct brw_codegen
*p
, struct brw_reg dest
,
1075 struct brw_reg src0
, struct brw_reg src1
)
1077 assert(dest
.type
== src0
.type
);
1078 assert(src0
.type
== src1
.type
);
1079 switch (src0
.type
) {
1080 case BRW_REGISTER_TYPE_B
:
1081 case BRW_REGISTER_TYPE_UB
:
1082 case BRW_REGISTER_TYPE_W
:
1083 case BRW_REGISTER_TYPE_UW
:
1084 case BRW_REGISTER_TYPE_D
:
1085 case BRW_REGISTER_TYPE_UD
:
1088 unreachable("Bad type for brw_AVG");
1091 return brw_alu2(p
, BRW_OPCODE_AVG
, dest
, src0
, src1
);
1095 brw_MUL(struct brw_codegen
*p
, struct brw_reg dest
,
1096 struct brw_reg src0
, struct brw_reg src1
)
1099 if (src0
.type
== BRW_REGISTER_TYPE_D
||
1100 src0
.type
== BRW_REGISTER_TYPE_UD
||
1101 src1
.type
== BRW_REGISTER_TYPE_D
||
1102 src1
.type
== BRW_REGISTER_TYPE_UD
) {
1103 assert(dest
.type
!= BRW_REGISTER_TYPE_F
);
1106 if (src0
.type
== BRW_REGISTER_TYPE_F
||
1107 (src0
.file
== BRW_IMMEDIATE_VALUE
&&
1108 src0
.type
== BRW_REGISTER_TYPE_VF
)) {
1109 assert(src1
.type
!= BRW_REGISTER_TYPE_UD
);
1110 assert(src1
.type
!= BRW_REGISTER_TYPE_D
);
1113 if (src1
.type
== BRW_REGISTER_TYPE_F
||
1114 (src1
.file
== BRW_IMMEDIATE_VALUE
&&
1115 src1
.type
== BRW_REGISTER_TYPE_VF
)) {
1116 assert(src0
.type
!= BRW_REGISTER_TYPE_UD
);
1117 assert(src0
.type
!= BRW_REGISTER_TYPE_D
);
1120 assert(src0
.file
!= BRW_ARCHITECTURE_REGISTER_FILE
||
1121 src0
.nr
!= BRW_ARF_ACCUMULATOR
);
1122 assert(src1
.file
!= BRW_ARCHITECTURE_REGISTER_FILE
||
1123 src1
.nr
!= BRW_ARF_ACCUMULATOR
);
1125 return brw_alu2(p
, BRW_OPCODE_MUL
, dest
, src0
, src1
);
1129 brw_LINE(struct brw_codegen
*p
, struct brw_reg dest
,
1130 struct brw_reg src0
, struct brw_reg src1
)
1132 src0
.vstride
= BRW_VERTICAL_STRIDE_0
;
1133 src0
.width
= BRW_WIDTH_1
;
1134 src0
.hstride
= BRW_HORIZONTAL_STRIDE_0
;
1135 return brw_alu2(p
, BRW_OPCODE_LINE
, dest
, src0
, src1
);
1139 brw_PLN(struct brw_codegen
*p
, struct brw_reg dest
,
1140 struct brw_reg src0
, struct brw_reg src1
)
1142 src0
.vstride
= BRW_VERTICAL_STRIDE_0
;
1143 src0
.width
= BRW_WIDTH_1
;
1144 src0
.hstride
= BRW_HORIZONTAL_STRIDE_0
;
1145 src1
.vstride
= BRW_VERTICAL_STRIDE_8
;
1146 src1
.width
= BRW_WIDTH_8
;
1147 src1
.hstride
= BRW_HORIZONTAL_STRIDE_1
;
1148 return brw_alu2(p
, BRW_OPCODE_PLN
, dest
, src0
, src1
);
1152 brw_F32TO16(struct brw_codegen
*p
, struct brw_reg dst
, struct brw_reg src
)
1154 const struct brw_device_info
*devinfo
= p
->devinfo
;
1155 const bool align16
= brw_inst_access_mode(devinfo
, p
->current
) == BRW_ALIGN_16
;
1156 /* The F32TO16 instruction doesn't support 32-bit destination types in
1157 * Align1 mode, and neither does the Gen8 implementation in terms of a
1158 * converting MOV. Gen7 does zero out the high 16 bits in Align16 mode as
1159 * an undocumented feature.
1161 const bool needs_zero_fill
= (dst
.type
== BRW_REGISTER_TYPE_UD
&&
1162 (!align16
|| devinfo
->gen
>= 8));
1166 assert(dst
.type
== BRW_REGISTER_TYPE_UD
);
1168 assert(dst
.type
== BRW_REGISTER_TYPE_UD
||
1169 dst
.type
== BRW_REGISTER_TYPE_W
||
1170 dst
.type
== BRW_REGISTER_TYPE_UW
||
1171 dst
.type
== BRW_REGISTER_TYPE_HF
);
1174 brw_push_insn_state(p
);
1176 if (needs_zero_fill
) {
1177 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
1178 dst
= spread(retype(dst
, BRW_REGISTER_TYPE_W
), 2);
1181 if (devinfo
->gen
>= 8) {
1182 inst
= brw_MOV(p
, retype(dst
, BRW_REGISTER_TYPE_HF
), src
);
1184 assert(devinfo
->gen
== 7);
1185 inst
= brw_alu1(p
, BRW_OPCODE_F32TO16
, dst
, src
);
1188 if (needs_zero_fill
) {
1189 brw_inst_set_no_dd_clear(devinfo
, inst
, true);
1190 inst
= brw_MOV(p
, suboffset(dst
, 1), brw_imm_ud(0u));
1191 brw_inst_set_no_dd_check(devinfo
, inst
, true);
1194 brw_pop_insn_state(p
);
1199 brw_F16TO32(struct brw_codegen
*p
, struct brw_reg dst
, struct brw_reg src
)
1201 const struct brw_device_info
*devinfo
= p
->devinfo
;
1202 bool align16
= brw_inst_access_mode(devinfo
, p
->current
) == BRW_ALIGN_16
;
1205 assert(src
.type
== BRW_REGISTER_TYPE_UD
);
1207 /* From the Ivybridge PRM, Vol4, Part3, Section 6.26 f16to32:
1209 * Because this instruction does not have a 16-bit floating-point
1210 * type, the source data type must be Word (W). The destination type
1211 * must be F (Float).
1213 if (src
.type
== BRW_REGISTER_TYPE_UD
)
1214 src
= spread(retype(src
, BRW_REGISTER_TYPE_W
), 2);
1216 assert(src
.type
== BRW_REGISTER_TYPE_W
||
1217 src
.type
== BRW_REGISTER_TYPE_UW
||
1218 src
.type
== BRW_REGISTER_TYPE_HF
);
1221 if (devinfo
->gen
>= 8) {
1222 return brw_MOV(p
, dst
, retype(src
, BRW_REGISTER_TYPE_HF
));
1224 assert(devinfo
->gen
== 7);
1225 return brw_alu1(p
, BRW_OPCODE_F16TO32
, dst
, src
);
1230 void brw_NOP(struct brw_codegen
*p
)
1232 brw_inst
*insn
= next_insn(p
, BRW_OPCODE_NOP
);
1233 brw_set_dest(p
, insn
, retype(brw_vec4_grf(0,0), BRW_REGISTER_TYPE_UD
));
1234 brw_set_src0(p
, insn
, retype(brw_vec4_grf(0,0), BRW_REGISTER_TYPE_UD
));
1235 brw_set_src1(p
, insn
, brw_imm_ud(0x0));
1242 /***********************************************************************
1243 * Comparisons, if/else/endif
1247 brw_JMPI(struct brw_codegen
*p
, struct brw_reg index
,
1248 unsigned predicate_control
)
1250 const struct brw_device_info
*devinfo
= p
->devinfo
;
1251 struct brw_reg ip
= brw_ip_reg();
1252 brw_inst
*inst
= brw_alu2(p
, BRW_OPCODE_JMPI
, ip
, ip
, index
);
1254 brw_inst_set_exec_size(devinfo
, inst
, BRW_EXECUTE_2
);
1255 brw_inst_set_qtr_control(devinfo
, inst
, BRW_COMPRESSION_NONE
);
1256 brw_inst_set_mask_control(devinfo
, inst
, BRW_MASK_DISABLE
);
1257 brw_inst_set_pred_control(devinfo
, inst
, predicate_control
);
1263 push_if_stack(struct brw_codegen
*p
, brw_inst
*inst
)
1265 p
->if_stack
[p
->if_stack_depth
] = inst
- p
->store
;
1267 p
->if_stack_depth
++;
1268 if (p
->if_stack_array_size
<= p
->if_stack_depth
) {
1269 p
->if_stack_array_size
*= 2;
1270 p
->if_stack
= reralloc(p
->mem_ctx
, p
->if_stack
, int,
1271 p
->if_stack_array_size
);
1276 pop_if_stack(struct brw_codegen
*p
)
1278 p
->if_stack_depth
--;
1279 return &p
->store
[p
->if_stack
[p
->if_stack_depth
]];
1283 push_loop_stack(struct brw_codegen
*p
, brw_inst
*inst
)
1285 if (p
->loop_stack_array_size
< p
->loop_stack_depth
) {
1286 p
->loop_stack_array_size
*= 2;
1287 p
->loop_stack
= reralloc(p
->mem_ctx
, p
->loop_stack
, int,
1288 p
->loop_stack_array_size
);
1289 p
->if_depth_in_loop
= reralloc(p
->mem_ctx
, p
->if_depth_in_loop
, int,
1290 p
->loop_stack_array_size
);
1293 p
->loop_stack
[p
->loop_stack_depth
] = inst
- p
->store
;
1294 p
->loop_stack_depth
++;
1295 p
->if_depth_in_loop
[p
->loop_stack_depth
] = 0;
1299 get_inner_do_insn(struct brw_codegen
*p
)
1301 return &p
->store
[p
->loop_stack
[p
->loop_stack_depth
- 1]];
1304 /* EU takes the value from the flag register and pushes it onto some
1305 * sort of a stack (presumably merging with any flag value already on
1306 * the stack). Within an if block, the flags at the top of the stack
1307 * control execution on each channel of the unit, eg. on each of the
1308 * 16 pixel values in our wm programs.
1310 * When the matching 'else' instruction is reached (presumably by
1311 * countdown of the instruction count patched in by our ELSE/ENDIF
1312 * functions), the relevant flags are inverted.
1314 * When the matching 'endif' instruction is reached, the flags are
1315 * popped off. If the stack is now empty, normal execution resumes.
1318 brw_IF(struct brw_codegen
*p
, unsigned execute_size
)
1320 const struct brw_device_info
*devinfo
= p
->devinfo
;
1323 insn
= next_insn(p
, BRW_OPCODE_IF
);
1325 /* Override the defaults for this instruction:
1327 if (devinfo
->gen
< 6) {
1328 brw_set_dest(p
, insn
, brw_ip_reg());
1329 brw_set_src0(p
, insn
, brw_ip_reg());
1330 brw_set_src1(p
, insn
, brw_imm_d(0x0));
1331 } else if (devinfo
->gen
== 6) {
1332 brw_set_dest(p
, insn
, brw_imm_w(0));
1333 brw_inst_set_gen6_jump_count(devinfo
, insn
, 0);
1334 brw_set_src0(p
, insn
, vec1(retype(brw_null_reg(), BRW_REGISTER_TYPE_D
)));
1335 brw_set_src1(p
, insn
, vec1(retype(brw_null_reg(), BRW_REGISTER_TYPE_D
)));
1336 } else if (devinfo
->gen
== 7) {
1337 brw_set_dest(p
, insn
, vec1(retype(brw_null_reg(), BRW_REGISTER_TYPE_D
)));
1338 brw_set_src0(p
, insn
, vec1(retype(brw_null_reg(), BRW_REGISTER_TYPE_D
)));
1339 brw_set_src1(p
, insn
, brw_imm_w(0));
1340 brw_inst_set_jip(devinfo
, insn
, 0);
1341 brw_inst_set_uip(devinfo
, insn
, 0);
1343 brw_set_dest(p
, insn
, vec1(retype(brw_null_reg(), BRW_REGISTER_TYPE_D
)));
1344 brw_set_src0(p
, insn
, brw_imm_d(0));
1345 brw_inst_set_jip(devinfo
, insn
, 0);
1346 brw_inst_set_uip(devinfo
, insn
, 0);
1349 brw_inst_set_exec_size(devinfo
, insn
, execute_size
);
1350 brw_inst_set_qtr_control(devinfo
, insn
, BRW_COMPRESSION_NONE
);
1351 brw_inst_set_pred_control(devinfo
, insn
, BRW_PREDICATE_NORMAL
);
1352 brw_inst_set_mask_control(devinfo
, insn
, BRW_MASK_ENABLE
);
1353 if (!p
->single_program_flow
&& devinfo
->gen
< 6)
1354 brw_inst_set_thread_control(devinfo
, insn
, BRW_THREAD_SWITCH
);
1356 push_if_stack(p
, insn
);
1357 p
->if_depth_in_loop
[p
->loop_stack_depth
]++;
1361 /* This function is only used for gen6-style IF instructions with an
1362 * embedded comparison (conditional modifier). It is not used on gen7.
1365 gen6_IF(struct brw_codegen
*p
, enum brw_conditional_mod conditional
,
1366 struct brw_reg src0
, struct brw_reg src1
)
1368 const struct brw_device_info
*devinfo
= p
->devinfo
;
1371 insn
= next_insn(p
, BRW_OPCODE_IF
);
1373 brw_set_dest(p
, insn
, brw_imm_w(0));
1374 brw_inst_set_exec_size(devinfo
, insn
, p
->compressed
? BRW_EXECUTE_16
1376 brw_inst_set_gen6_jump_count(devinfo
, insn
, 0);
1377 brw_set_src0(p
, insn
, src0
);
1378 brw_set_src1(p
, insn
, src1
);
1380 assert(brw_inst_qtr_control(devinfo
, insn
) == BRW_COMPRESSION_NONE
);
1381 assert(brw_inst_pred_control(devinfo
, insn
) == BRW_PREDICATE_NONE
);
1382 brw_inst_set_cond_modifier(devinfo
, insn
, conditional
);
1384 push_if_stack(p
, insn
);
1389 * In single-program-flow (SPF) mode, convert IF and ELSE into ADDs.
1392 convert_IF_ELSE_to_ADD(struct brw_codegen
*p
,
1393 brw_inst
*if_inst
, brw_inst
*else_inst
)
1395 const struct brw_device_info
*devinfo
= p
->devinfo
;
1397 /* The next instruction (where the ENDIF would be, if it existed) */
1398 brw_inst
*next_inst
= &p
->store
[p
->nr_insn
];
1400 assert(p
->single_program_flow
);
1401 assert(if_inst
!= NULL
&& brw_inst_opcode(devinfo
, if_inst
) == BRW_OPCODE_IF
);
1402 assert(else_inst
== NULL
|| brw_inst_opcode(devinfo
, else_inst
) == BRW_OPCODE_ELSE
);
1403 assert(brw_inst_exec_size(devinfo
, if_inst
) == BRW_EXECUTE_1
);
1405 /* Convert IF to an ADD instruction that moves the instruction pointer
1406 * to the first instruction of the ELSE block. If there is no ELSE
1407 * block, point to where ENDIF would be. Reverse the predicate.
1409 * There's no need to execute an ENDIF since we don't need to do any
1410 * stack operations, and if we're currently executing, we just want to
1411 * continue normally.
1413 brw_inst_set_opcode(devinfo
, if_inst
, BRW_OPCODE_ADD
);
1414 brw_inst_set_pred_inv(devinfo
, if_inst
, true);
1416 if (else_inst
!= NULL
) {
1417 /* Convert ELSE to an ADD instruction that points where the ENDIF
1420 brw_inst_set_opcode(devinfo
, else_inst
, BRW_OPCODE_ADD
);
1422 brw_inst_set_imm_ud(devinfo
, if_inst
, (else_inst
- if_inst
+ 1) * 16);
1423 brw_inst_set_imm_ud(devinfo
, else_inst
, (next_inst
- else_inst
) * 16);
1425 brw_inst_set_imm_ud(devinfo
, if_inst
, (next_inst
- if_inst
) * 16);
1430 * Patch IF and ELSE instructions with appropriate jump targets.
1433 patch_IF_ELSE(struct brw_codegen
*p
,
1434 brw_inst
*if_inst
, brw_inst
*else_inst
, brw_inst
*endif_inst
)
1436 const struct brw_device_info
*devinfo
= p
->devinfo
;
1438 /* We shouldn't be patching IF and ELSE instructions in single program flow
1439 * mode when gen < 6, because in single program flow mode on those
1440 * platforms, we convert flow control instructions to conditional ADDs that
1441 * operate on IP (see brw_ENDIF).
1443 * However, on Gen6, writing to IP doesn't work in single program flow mode
1444 * (see the SandyBridge PRM, Volume 4 part 2, p79: "When SPF is ON, IP may
1445 * not be updated by non-flow control instructions."). And on later
1446 * platforms, there is no significant benefit to converting control flow
1447 * instructions to conditional ADDs. So we do patch IF and ELSE
1448 * instructions in single program flow mode on those platforms.
1450 if (devinfo
->gen
< 6)
1451 assert(!p
->single_program_flow
);
1453 assert(if_inst
!= NULL
&& brw_inst_opcode(devinfo
, if_inst
) == BRW_OPCODE_IF
);
1454 assert(endif_inst
!= NULL
);
1455 assert(else_inst
== NULL
|| brw_inst_opcode(devinfo
, else_inst
) == BRW_OPCODE_ELSE
);
1457 unsigned br
= brw_jump_scale(devinfo
);
1459 assert(brw_inst_opcode(devinfo
, endif_inst
) == BRW_OPCODE_ENDIF
);
1460 brw_inst_set_exec_size(devinfo
, endif_inst
, brw_inst_exec_size(devinfo
, if_inst
));
1462 if (else_inst
== NULL
) {
1463 /* Patch IF -> ENDIF */
1464 if (devinfo
->gen
< 6) {
1465 /* Turn it into an IFF, which means no mask stack operations for
1466 * all-false and jumping past the ENDIF.
1468 brw_inst_set_opcode(devinfo
, if_inst
, BRW_OPCODE_IFF
);
1469 brw_inst_set_gen4_jump_count(devinfo
, if_inst
,
1470 br
* (endif_inst
- if_inst
+ 1));
1471 brw_inst_set_gen4_pop_count(devinfo
, if_inst
, 0);
1472 } else if (devinfo
->gen
== 6) {
1473 /* As of gen6, there is no IFF and IF must point to the ENDIF. */
1474 brw_inst_set_gen6_jump_count(devinfo
, if_inst
, br
*(endif_inst
- if_inst
));
1476 brw_inst_set_uip(devinfo
, if_inst
, br
* (endif_inst
- if_inst
));
1477 brw_inst_set_jip(devinfo
, if_inst
, br
* (endif_inst
- if_inst
));
1480 brw_inst_set_exec_size(devinfo
, else_inst
, brw_inst_exec_size(devinfo
, if_inst
));
1482 /* Patch IF -> ELSE */
1483 if (devinfo
->gen
< 6) {
1484 brw_inst_set_gen4_jump_count(devinfo
, if_inst
,
1485 br
* (else_inst
- if_inst
));
1486 brw_inst_set_gen4_pop_count(devinfo
, if_inst
, 0);
1487 } else if (devinfo
->gen
== 6) {
1488 brw_inst_set_gen6_jump_count(devinfo
, if_inst
,
1489 br
* (else_inst
- if_inst
+ 1));
1492 /* Patch ELSE -> ENDIF */
1493 if (devinfo
->gen
< 6) {
1494 /* BRW_OPCODE_ELSE pre-gen6 should point just past the
1497 brw_inst_set_gen4_jump_count(devinfo
, else_inst
,
1498 br
* (endif_inst
- else_inst
+ 1));
1499 brw_inst_set_gen4_pop_count(devinfo
, else_inst
, 1);
1500 } else if (devinfo
->gen
== 6) {
1501 /* BRW_OPCODE_ELSE on gen6 should point to the matching ENDIF. */
1502 brw_inst_set_gen6_jump_count(devinfo
, else_inst
,
1503 br
* (endif_inst
- else_inst
));
1505 /* The IF instruction's JIP should point just past the ELSE */
1506 brw_inst_set_jip(devinfo
, if_inst
, br
* (else_inst
- if_inst
+ 1));
1507 /* The IF instruction's UIP and ELSE's JIP should point to ENDIF */
1508 brw_inst_set_uip(devinfo
, if_inst
, br
* (endif_inst
- if_inst
));
1509 brw_inst_set_jip(devinfo
, else_inst
, br
* (endif_inst
- else_inst
));
1510 if (devinfo
->gen
>= 8) {
1511 /* Since we don't set branch_ctrl, the ELSE's JIP and UIP both
1512 * should point to ENDIF.
1514 brw_inst_set_uip(devinfo
, else_inst
, br
* (endif_inst
- else_inst
));
1521 brw_ELSE(struct brw_codegen
*p
)
1523 const struct brw_device_info
*devinfo
= p
->devinfo
;
1526 insn
= next_insn(p
, BRW_OPCODE_ELSE
);
1528 if (devinfo
->gen
< 6) {
1529 brw_set_dest(p
, insn
, brw_ip_reg());
1530 brw_set_src0(p
, insn
, brw_ip_reg());
1531 brw_set_src1(p
, insn
, brw_imm_d(0x0));
1532 } else if (devinfo
->gen
== 6) {
1533 brw_set_dest(p
, insn
, brw_imm_w(0));
1534 brw_inst_set_gen6_jump_count(devinfo
, insn
, 0);
1535 brw_set_src0(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1536 brw_set_src1(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1537 } else if (devinfo
->gen
== 7) {
1538 brw_set_dest(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1539 brw_set_src0(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1540 brw_set_src1(p
, insn
, brw_imm_w(0));
1541 brw_inst_set_jip(devinfo
, insn
, 0);
1542 brw_inst_set_uip(devinfo
, insn
, 0);
1544 brw_set_dest(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1545 brw_set_src0(p
, insn
, brw_imm_d(0));
1546 brw_inst_set_jip(devinfo
, insn
, 0);
1547 brw_inst_set_uip(devinfo
, insn
, 0);
1550 brw_inst_set_qtr_control(devinfo
, insn
, BRW_COMPRESSION_NONE
);
1551 brw_inst_set_mask_control(devinfo
, insn
, BRW_MASK_ENABLE
);
1552 if (!p
->single_program_flow
&& devinfo
->gen
< 6)
1553 brw_inst_set_thread_control(devinfo
, insn
, BRW_THREAD_SWITCH
);
1555 push_if_stack(p
, insn
);
1559 brw_ENDIF(struct brw_codegen
*p
)
1561 const struct brw_device_info
*devinfo
= p
->devinfo
;
1562 brw_inst
*insn
= NULL
;
1563 brw_inst
*else_inst
= NULL
;
1564 brw_inst
*if_inst
= NULL
;
1566 bool emit_endif
= true;
1568 /* In single program flow mode, we can express IF and ELSE instructions
1569 * equivalently as ADD instructions that operate on IP. On platforms prior
1570 * to Gen6, flow control instructions cause an implied thread switch, so
1571 * this is a significant savings.
1573 * However, on Gen6, writing to IP doesn't work in single program flow mode
1574 * (see the SandyBridge PRM, Volume 4 part 2, p79: "When SPF is ON, IP may
1575 * not be updated by non-flow control instructions."). And on later
1576 * platforms, there is no significant benefit to converting control flow
1577 * instructions to conditional ADDs. So we only do this trick on Gen4 and
1580 if (devinfo
->gen
< 6 && p
->single_program_flow
)
1584 * A single next_insn() may change the base address of instruction store
1585 * memory(p->store), so call it first before referencing the instruction
1586 * store pointer from an index
1589 insn
= next_insn(p
, BRW_OPCODE_ENDIF
);
1591 /* Pop the IF and (optional) ELSE instructions from the stack */
1592 p
->if_depth_in_loop
[p
->loop_stack_depth
]--;
1593 tmp
= pop_if_stack(p
);
1594 if (brw_inst_opcode(devinfo
, tmp
) == BRW_OPCODE_ELSE
) {
1596 tmp
= pop_if_stack(p
);
1601 /* ENDIF is useless; don't bother emitting it. */
1602 convert_IF_ELSE_to_ADD(p
, if_inst
, else_inst
);
1606 if (devinfo
->gen
< 6) {
1607 brw_set_dest(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1608 brw_set_src0(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1609 brw_set_src1(p
, insn
, brw_imm_d(0x0));
1610 } else if (devinfo
->gen
== 6) {
1611 brw_set_dest(p
, insn
, brw_imm_w(0));
1612 brw_set_src0(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1613 brw_set_src1(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1614 } else if (devinfo
->gen
== 7) {
1615 brw_set_dest(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1616 brw_set_src0(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1617 brw_set_src1(p
, insn
, brw_imm_w(0));
1619 brw_set_src0(p
, insn
, brw_imm_d(0));
1622 brw_inst_set_qtr_control(devinfo
, insn
, BRW_COMPRESSION_NONE
);
1623 brw_inst_set_mask_control(devinfo
, insn
, BRW_MASK_ENABLE
);
1624 if (devinfo
->gen
< 6)
1625 brw_inst_set_thread_control(devinfo
, insn
, BRW_THREAD_SWITCH
);
1627 /* Also pop item off the stack in the endif instruction: */
1628 if (devinfo
->gen
< 6) {
1629 brw_inst_set_gen4_jump_count(devinfo
, insn
, 0);
1630 brw_inst_set_gen4_pop_count(devinfo
, insn
, 1);
1631 } else if (devinfo
->gen
== 6) {
1632 brw_inst_set_gen6_jump_count(devinfo
, insn
, 2);
1634 brw_inst_set_jip(devinfo
, insn
, 2);
1636 patch_IF_ELSE(p
, if_inst
, else_inst
, insn
);
1640 brw_BREAK(struct brw_codegen
*p
)
1642 const struct brw_device_info
*devinfo
= p
->devinfo
;
1645 insn
= next_insn(p
, BRW_OPCODE_BREAK
);
1646 if (devinfo
->gen
>= 8) {
1647 brw_set_dest(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1648 brw_set_src0(p
, insn
, brw_imm_d(0x0));
1649 } else if (devinfo
->gen
>= 6) {
1650 brw_set_dest(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1651 brw_set_src0(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1652 brw_set_src1(p
, insn
, brw_imm_d(0x0));
1654 brw_set_dest(p
, insn
, brw_ip_reg());
1655 brw_set_src0(p
, insn
, brw_ip_reg());
1656 brw_set_src1(p
, insn
, brw_imm_d(0x0));
1657 brw_inst_set_gen4_pop_count(devinfo
, insn
,
1658 p
->if_depth_in_loop
[p
->loop_stack_depth
]);
1660 brw_inst_set_qtr_control(devinfo
, insn
, BRW_COMPRESSION_NONE
);
1661 brw_inst_set_exec_size(devinfo
, insn
, p
->compressed
? BRW_EXECUTE_16
1668 brw_CONT(struct brw_codegen
*p
)
1670 const struct brw_device_info
*devinfo
= p
->devinfo
;
1673 insn
= next_insn(p
, BRW_OPCODE_CONTINUE
);
1674 brw_set_dest(p
, insn
, brw_ip_reg());
1675 if (devinfo
->gen
>= 8) {
1676 brw_set_src0(p
, insn
, brw_imm_d(0x0));
1678 brw_set_src0(p
, insn
, brw_ip_reg());
1679 brw_set_src1(p
, insn
, brw_imm_d(0x0));
1682 if (devinfo
->gen
< 6) {
1683 brw_inst_set_gen4_pop_count(devinfo
, insn
,
1684 p
->if_depth_in_loop
[p
->loop_stack_depth
]);
1686 brw_inst_set_qtr_control(devinfo
, insn
, BRW_COMPRESSION_NONE
);
1687 brw_inst_set_exec_size(devinfo
, insn
, p
->compressed
? BRW_EXECUTE_16
1693 gen6_HALT(struct brw_codegen
*p
)
1695 const struct brw_device_info
*devinfo
= p
->devinfo
;
1698 insn
= next_insn(p
, BRW_OPCODE_HALT
);
1699 brw_set_dest(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1700 if (devinfo
->gen
>= 8) {
1701 brw_set_src0(p
, insn
, brw_imm_d(0x0));
1703 brw_set_src0(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1704 brw_set_src1(p
, insn
, brw_imm_d(0x0)); /* UIP and JIP, updated later. */
1707 if (p
->compressed
) {
1708 brw_inst_set_exec_size(devinfo
, insn
, BRW_EXECUTE_16
);
1710 brw_inst_set_qtr_control(devinfo
, insn
, BRW_COMPRESSION_NONE
);
1711 brw_inst_set_exec_size(devinfo
, insn
, BRW_EXECUTE_8
);
1718 * The DO/WHILE is just an unterminated loop -- break or continue are
1719 * used for control within the loop. We have a few ways they can be
1722 * For uniform control flow, the WHILE is just a jump, so ADD ip, ip,
1723 * jip and no DO instruction.
1725 * For non-uniform control flow pre-gen6, there's a DO instruction to
1726 * push the mask, and a WHILE to jump back, and BREAK to get out and
1729 * For gen6, there's no more mask stack, so no need for DO. WHILE
1730 * just points back to the first instruction of the loop.
1733 brw_DO(struct brw_codegen
*p
, unsigned execute_size
)
1735 const struct brw_device_info
*devinfo
= p
->devinfo
;
1737 if (devinfo
->gen
>= 6 || p
->single_program_flow
) {
1738 push_loop_stack(p
, &p
->store
[p
->nr_insn
]);
1739 return &p
->store
[p
->nr_insn
];
1741 brw_inst
*insn
= next_insn(p
, BRW_OPCODE_DO
);
1743 push_loop_stack(p
, insn
);
1745 /* Override the defaults for this instruction:
1747 brw_set_dest(p
, insn
, brw_null_reg());
1748 brw_set_src0(p
, insn
, brw_null_reg());
1749 brw_set_src1(p
, insn
, brw_null_reg());
1751 brw_inst_set_qtr_control(devinfo
, insn
, BRW_COMPRESSION_NONE
);
1752 brw_inst_set_exec_size(devinfo
, insn
, execute_size
);
1753 brw_inst_set_pred_control(devinfo
, insn
, BRW_PREDICATE_NONE
);
1760 * For pre-gen6, we patch BREAK/CONT instructions to point at the WHILE
1763 * For gen6+, see brw_set_uip_jip(), which doesn't care so much about the loop
1764 * nesting, since it can always just point to the end of the block/current loop.
1767 brw_patch_break_cont(struct brw_codegen
*p
, brw_inst
*while_inst
)
1769 const struct brw_device_info
*devinfo
= p
->devinfo
;
1770 brw_inst
*do_inst
= get_inner_do_insn(p
);
1772 unsigned br
= brw_jump_scale(devinfo
);
1774 assert(devinfo
->gen
< 6);
1776 for (inst
= while_inst
- 1; inst
!= do_inst
; inst
--) {
1777 /* If the jump count is != 0, that means that this instruction has already
1778 * been patched because it's part of a loop inside of the one we're
1781 if (brw_inst_opcode(devinfo
, inst
) == BRW_OPCODE_BREAK
&&
1782 brw_inst_gen4_jump_count(devinfo
, inst
) == 0) {
1783 brw_inst_set_gen4_jump_count(devinfo
, inst
, br
*((while_inst
- inst
) + 1));
1784 } else if (brw_inst_opcode(devinfo
, inst
) == BRW_OPCODE_CONTINUE
&&
1785 brw_inst_gen4_jump_count(devinfo
, inst
) == 0) {
1786 brw_inst_set_gen4_jump_count(devinfo
, inst
, br
* (while_inst
- inst
));
1792 brw_WHILE(struct brw_codegen
*p
)
1794 const struct brw_device_info
*devinfo
= p
->devinfo
;
1795 brw_inst
*insn
, *do_insn
;
1796 unsigned br
= brw_jump_scale(devinfo
);
1798 if (devinfo
->gen
>= 6) {
1799 insn
= next_insn(p
, BRW_OPCODE_WHILE
);
1800 do_insn
= get_inner_do_insn(p
);
1802 if (devinfo
->gen
>= 8) {
1803 brw_set_dest(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1804 brw_set_src0(p
, insn
, brw_imm_d(0));
1805 brw_inst_set_jip(devinfo
, insn
, br
* (do_insn
- insn
));
1806 } else if (devinfo
->gen
== 7) {
1807 brw_set_dest(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1808 brw_set_src0(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1809 brw_set_src1(p
, insn
, brw_imm_w(0));
1810 brw_inst_set_jip(devinfo
, insn
, br
* (do_insn
- insn
));
1812 brw_set_dest(p
, insn
, brw_imm_w(0));
1813 brw_inst_set_gen6_jump_count(devinfo
, insn
, br
* (do_insn
- insn
));
1814 brw_set_src0(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1815 brw_set_src1(p
, insn
, retype(brw_null_reg(), BRW_REGISTER_TYPE_D
));
1818 brw_inst_set_exec_size(devinfo
, insn
, p
->compressed
? BRW_EXECUTE_16
1821 if (p
->single_program_flow
) {
1822 insn
= next_insn(p
, BRW_OPCODE_ADD
);
1823 do_insn
= get_inner_do_insn(p
);
1825 brw_set_dest(p
, insn
, brw_ip_reg());
1826 brw_set_src0(p
, insn
, brw_ip_reg());
1827 brw_set_src1(p
, insn
, brw_imm_d((do_insn
- insn
) * 16));
1828 brw_inst_set_exec_size(devinfo
, insn
, BRW_EXECUTE_1
);
1830 insn
= next_insn(p
, BRW_OPCODE_WHILE
);
1831 do_insn
= get_inner_do_insn(p
);
1833 assert(brw_inst_opcode(devinfo
, do_insn
) == BRW_OPCODE_DO
);
1835 brw_set_dest(p
, insn
, brw_ip_reg());
1836 brw_set_src0(p
, insn
, brw_ip_reg());
1837 brw_set_src1(p
, insn
, brw_imm_d(0));
1839 brw_inst_set_exec_size(devinfo
, insn
, brw_inst_exec_size(devinfo
, do_insn
));
1840 brw_inst_set_gen4_jump_count(devinfo
, insn
, br
* (do_insn
- insn
+ 1));
1841 brw_inst_set_gen4_pop_count(devinfo
, insn
, 0);
1843 brw_patch_break_cont(p
, insn
);
1846 brw_inst_set_qtr_control(devinfo
, insn
, BRW_COMPRESSION_NONE
);
1848 p
->loop_stack_depth
--;
1855 void brw_land_fwd_jump(struct brw_codegen
*p
, int jmp_insn_idx
)
1857 const struct brw_device_info
*devinfo
= p
->devinfo
;
1858 brw_inst
*jmp_insn
= &p
->store
[jmp_insn_idx
];
1861 if (devinfo
->gen
>= 5)
1864 assert(brw_inst_opcode(devinfo
, jmp_insn
) == BRW_OPCODE_JMPI
);
1865 assert(brw_inst_src1_reg_file(devinfo
, jmp_insn
) == BRW_IMMEDIATE_VALUE
);
1867 brw_inst_set_gen4_jump_count(devinfo
, jmp_insn
,
1868 jmpi
* (p
->nr_insn
- jmp_insn_idx
- 1));
1871 /* To integrate with the above, it makes sense that the comparison
1872 * instruction should populate the flag register. It might be simpler
1873 * just to use the flag reg for most WM tasks?
1875 void brw_CMP(struct brw_codegen
*p
,
1876 struct brw_reg dest
,
1877 unsigned conditional
,
1878 struct brw_reg src0
,
1879 struct brw_reg src1
)
1881 const struct brw_device_info
*devinfo
= p
->devinfo
;
1882 brw_inst
*insn
= next_insn(p
, BRW_OPCODE_CMP
);
1884 brw_inst_set_cond_modifier(devinfo
, insn
, conditional
);
1885 brw_set_dest(p
, insn
, dest
);
1886 brw_set_src0(p
, insn
, src0
);
1887 brw_set_src1(p
, insn
, src1
);
1889 /* Item WaCMPInstNullDstForcesThreadSwitch in the Haswell Bspec workarounds
1891 * "Any CMP instruction with a null destination must use a {switch}."
1893 * It also applies to other Gen7 platforms (IVB, BYT) even though it isn't
1894 * mentioned on their work-arounds pages.
1896 if (devinfo
->gen
== 7) {
1897 if (dest
.file
== BRW_ARCHITECTURE_REGISTER_FILE
&&
1898 dest
.nr
== BRW_ARF_NULL
) {
1899 brw_inst_set_thread_control(devinfo
, insn
, BRW_THREAD_SWITCH
);
1904 /***********************************************************************
1905 * Helpers for the various SEND message types:
1908 /** Extended math function, float[8].
1910 void gen4_math(struct brw_codegen
*p
,
1911 struct brw_reg dest
,
1913 unsigned msg_reg_nr
,
1915 unsigned precision
)
1917 const struct brw_device_info
*devinfo
= p
->devinfo
;
1918 brw_inst
*insn
= next_insn(p
, BRW_OPCODE_SEND
);
1920 if (has_scalar_region(src
)) {
1921 data_type
= BRW_MATH_DATA_SCALAR
;
1923 data_type
= BRW_MATH_DATA_VECTOR
;
1926 assert(devinfo
->gen
< 6);
1928 /* Example code doesn't set predicate_control for send
1931 brw_inst_set_pred_control(devinfo
, insn
, 0);
1932 brw_inst_set_base_mrf(devinfo
, insn
, msg_reg_nr
);
1934 brw_set_dest(p
, insn
, dest
);
1935 brw_set_src0(p
, insn
, src
);
1936 brw_set_math_message(p
,
1939 src
.type
== BRW_REGISTER_TYPE_D
,
1944 void gen6_math(struct brw_codegen
*p
,
1945 struct brw_reg dest
,
1947 struct brw_reg src0
,
1948 struct brw_reg src1
)
1950 const struct brw_device_info
*devinfo
= p
->devinfo
;
1951 brw_inst
*insn
= next_insn(p
, BRW_OPCODE_MATH
);
1953 assert(devinfo
->gen
>= 6);
1955 assert(dest
.file
== BRW_GENERAL_REGISTER_FILE
||
1956 (devinfo
->gen
>= 7 && dest
.file
== BRW_MESSAGE_REGISTER_FILE
));
1957 assert(src0
.file
== BRW_GENERAL_REGISTER_FILE
||
1958 (devinfo
->gen
>= 8 && src0
.file
== BRW_IMMEDIATE_VALUE
));
1960 assert(dest
.hstride
== BRW_HORIZONTAL_STRIDE_1
);
1961 if (devinfo
->gen
== 6) {
1962 assert(src0
.hstride
== BRW_HORIZONTAL_STRIDE_1
);
1963 assert(src1
.hstride
== BRW_HORIZONTAL_STRIDE_1
);
1966 if (function
== BRW_MATH_FUNCTION_INT_DIV_QUOTIENT
||
1967 function
== BRW_MATH_FUNCTION_INT_DIV_REMAINDER
||
1968 function
== BRW_MATH_FUNCTION_INT_DIV_QUOTIENT_AND_REMAINDER
) {
1969 assert(src0
.type
!= BRW_REGISTER_TYPE_F
);
1970 assert(src1
.type
!= BRW_REGISTER_TYPE_F
);
1971 assert(src1
.file
== BRW_GENERAL_REGISTER_FILE
||
1972 (devinfo
->gen
>= 8 && src1
.file
== BRW_IMMEDIATE_VALUE
));
1974 assert(src0
.type
== BRW_REGISTER_TYPE_F
);
1975 assert(src1
.type
== BRW_REGISTER_TYPE_F
);
1976 if (function
== BRW_MATH_FUNCTION_POW
) {
1977 assert(src1
.file
== BRW_GENERAL_REGISTER_FILE
||
1978 (devinfo
->gen
>= 8 && src1
.file
== BRW_IMMEDIATE_VALUE
));
1980 assert(src1
.file
== BRW_ARCHITECTURE_REGISTER_FILE
&&
1981 src1
.nr
== BRW_ARF_NULL
);
1985 /* Source modifiers are ignored for extended math instructions on Gen6. */
1986 if (devinfo
->gen
== 6) {
1987 assert(!src0
.negate
);
1989 assert(!src1
.negate
);
1993 brw_inst_set_math_function(devinfo
, insn
, function
);
1995 brw_set_dest(p
, insn
, dest
);
1996 brw_set_src0(p
, insn
, src0
);
1997 brw_set_src1(p
, insn
, src1
);
2001 * Return the right surface index to access the thread scratch space using
2002 * stateless dataport messages.
2005 brw_scratch_surface_idx(const struct brw_codegen
*p
)
2007 /* The scratch space is thread-local so IA coherency is unnecessary. */
2008 if (p
->devinfo
->gen
>= 8)
2009 return GEN8_BTI_STATELESS_NON_COHERENT
;
2011 return BRW_BTI_STATELESS
;
2015 * Write a block of OWORDs (half a GRF each) from the scratch buffer,
2016 * using a constant offset per channel.
2018 * The offset must be aligned to oword size (16 bytes). Used for
2019 * register spilling.
2021 void brw_oword_block_write_scratch(struct brw_codegen
*p
,
2026 const struct brw_device_info
*devinfo
= p
->devinfo
;
2027 uint32_t msg_control
, msg_type
;
2030 if (devinfo
->gen
>= 6)
2033 mrf
= retype(mrf
, BRW_REGISTER_TYPE_UD
);
2035 if (num_regs
== 1) {
2036 msg_control
= BRW_DATAPORT_OWORD_BLOCK_2_OWORDS
;
2039 msg_control
= BRW_DATAPORT_OWORD_BLOCK_4_OWORDS
;
2043 /* Set up the message header. This is g0, with g0.2 filled with
2044 * the offset. We don't want to leave our offset around in g0 or
2045 * it'll screw up texture samples, so set it up inside the message
2049 brw_push_insn_state(p
);
2050 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
2051 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
2052 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
2054 brw_MOV(p
, mrf
, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
2056 /* set message header global offset field (reg 0, element 2) */
2058 retype(brw_vec1_reg(BRW_MESSAGE_REGISTER_FILE
,
2060 2), BRW_REGISTER_TYPE_UD
),
2061 brw_imm_ud(offset
));
2063 brw_pop_insn_state(p
);
2067 struct brw_reg dest
;
2068 brw_inst
*insn
= next_insn(p
, BRW_OPCODE_SEND
);
2069 int send_commit_msg
;
2070 struct brw_reg src_header
= retype(brw_vec8_grf(0, 0),
2071 BRW_REGISTER_TYPE_UW
);
2073 if (brw_inst_qtr_control(devinfo
, insn
) != BRW_COMPRESSION_NONE
) {
2074 brw_inst_set_qtr_control(devinfo
, insn
, BRW_COMPRESSION_NONE
);
2075 src_header
= vec16(src_header
);
2077 assert(brw_inst_pred_control(devinfo
, insn
) == BRW_PREDICATE_NONE
);
2078 if (devinfo
->gen
< 6)
2079 brw_inst_set_base_mrf(devinfo
, insn
, mrf
.nr
);
2081 /* Until gen6, writes followed by reads from the same location
2082 * are not guaranteed to be ordered unless write_commit is set.
2083 * If set, then a no-op write is issued to the destination
2084 * register to set a dependency, and a read from the destination
2085 * can be used to ensure the ordering.
2087 * For gen6, only writes between different threads need ordering
2088 * protection. Our use of DP writes is all about register
2089 * spilling within a thread.
2091 if (devinfo
->gen
>= 6) {
2092 dest
= retype(vec16(brw_null_reg()), BRW_REGISTER_TYPE_UW
);
2093 send_commit_msg
= 0;
2096 send_commit_msg
= 1;
2099 brw_set_dest(p
, insn
, dest
);
2100 if (devinfo
->gen
>= 6) {
2101 brw_set_src0(p
, insn
, mrf
);
2103 brw_set_src0(p
, insn
, brw_null_reg());
2106 if (devinfo
->gen
>= 6)
2107 msg_type
= GEN6_DATAPORT_WRITE_MESSAGE_OWORD_BLOCK_WRITE
;
2109 msg_type
= BRW_DATAPORT_WRITE_MESSAGE_OWORD_BLOCK_WRITE
;
2111 brw_set_dp_write_message(p
,
2113 brw_scratch_surface_idx(p
),
2117 true, /* header_present */
2118 0, /* not a render target */
2119 send_commit_msg
, /* response_length */
2127 * Read a block of owords (half a GRF each) from the scratch buffer
2128 * using a constant index per channel.
2130 * Offset must be aligned to oword size (16 bytes). Used for register
2134 brw_oword_block_read_scratch(struct brw_codegen
*p
,
2135 struct brw_reg dest
,
2140 const struct brw_device_info
*devinfo
= p
->devinfo
;
2141 uint32_t msg_control
;
2144 if (devinfo
->gen
>= 6)
2147 if (p
->devinfo
->gen
>= 7) {
2148 /* On gen 7 and above, we no longer have message registers and we can
2149 * send from any register we want. By using the destination register
2150 * for the message, we guarantee that the implied message write won't
2151 * accidentally overwrite anything. This has been a problem because
2152 * the MRF registers and source for the final FB write are both fixed
2155 mrf
= retype(dest
, BRW_REGISTER_TYPE_UD
);
2157 mrf
= retype(mrf
, BRW_REGISTER_TYPE_UD
);
2159 dest
= retype(dest
, BRW_REGISTER_TYPE_UW
);
2161 if (num_regs
== 1) {
2162 msg_control
= BRW_DATAPORT_OWORD_BLOCK_2_OWORDS
;
2165 msg_control
= BRW_DATAPORT_OWORD_BLOCK_4_OWORDS
;
2170 brw_push_insn_state(p
);
2171 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
2172 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
2173 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
2175 brw_MOV(p
, mrf
, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
2177 /* set message header global offset field (reg 0, element 2) */
2178 brw_MOV(p
, get_element_ud(mrf
, 2), brw_imm_ud(offset
));
2180 brw_pop_insn_state(p
);
2184 brw_inst
*insn
= next_insn(p
, BRW_OPCODE_SEND
);
2186 assert(brw_inst_pred_control(devinfo
, insn
) == 0);
2187 brw_inst_set_qtr_control(devinfo
, insn
, BRW_COMPRESSION_NONE
);
2189 brw_set_dest(p
, insn
, dest
); /* UW? */
2190 if (devinfo
->gen
>= 6) {
2191 brw_set_src0(p
, insn
, mrf
);
2193 brw_set_src0(p
, insn
, brw_null_reg());
2194 brw_inst_set_base_mrf(devinfo
, insn
, mrf
.nr
);
2197 brw_set_dp_read_message(p
,
2199 brw_scratch_surface_idx(p
),
2201 BRW_DATAPORT_READ_MESSAGE_OWORD_BLOCK_READ
, /* msg_type */
2202 BRW_DATAPORT_READ_TARGET_RENDER_CACHE
,
2204 true, /* header_present */
2210 gen7_block_read_scratch(struct brw_codegen
*p
,
2211 struct brw_reg dest
,
2215 const struct brw_device_info
*devinfo
= p
->devinfo
;
2216 brw_inst
*insn
= next_insn(p
, BRW_OPCODE_SEND
);
2217 assert(brw_inst_pred_control(devinfo
, insn
) == BRW_PREDICATE_NONE
);
2219 brw_inst_set_qtr_control(devinfo
, insn
, BRW_COMPRESSION_NONE
);
2220 brw_set_dest(p
, insn
, retype(dest
, BRW_REGISTER_TYPE_UW
));
2222 /* The HW requires that the header is present; this is to get the g0.5
2225 brw_set_src0(p
, insn
, brw_vec8_grf(0, 0));
2227 /* According to the docs, offset is "A 12-bit HWord offset into the memory
2228 * Immediate Memory buffer as specified by binding table 0xFF." An HWORD
2229 * is 32 bytes, which happens to be the size of a register.
2232 assert(offset
< (1 << 12));
2234 gen7_set_dp_scratch_message(p
, insn
,
2235 false, /* scratch read */
2237 false, /* invalidate after read */
2240 1, /* mlen: just g0 */
2241 num_regs
, /* rlen */
2242 true); /* header present */
2246 * Read a float[4] vector from the data port Data Cache (const buffer).
2247 * Location (in buffer) should be a multiple of 16.
2248 * Used for fetching shader constants.
2250 void brw_oword_block_read(struct brw_codegen
*p
,
2251 struct brw_reg dest
,
2254 uint32_t bind_table_index
)
2256 const struct brw_device_info
*devinfo
= p
->devinfo
;
2258 /* On newer hardware, offset is in units of owords. */
2259 if (devinfo
->gen
>= 6)
2262 mrf
= retype(mrf
, BRW_REGISTER_TYPE_UD
);
2264 brw_push_insn_state(p
);
2265 brw_set_default_exec_size(p
, BRW_EXECUTE_8
);
2266 brw_set_default_predicate_control(p
, BRW_PREDICATE_NONE
);
2267 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
2268 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
2270 brw_MOV(p
, mrf
, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
2272 /* set message header global offset field (reg 0, element 2) */
2274 retype(brw_vec1_reg(BRW_MESSAGE_REGISTER_FILE
,
2276 2), BRW_REGISTER_TYPE_UD
),
2277 brw_imm_ud(offset
));
2279 brw_inst
*insn
= next_insn(p
, BRW_OPCODE_SEND
);
2281 /* cast dest to a uword[8] vector */
2282 dest
= retype(vec8(dest
), BRW_REGISTER_TYPE_UW
);
2284 brw_set_dest(p
, insn
, dest
);
2285 if (devinfo
->gen
>= 6) {
2286 brw_set_src0(p
, insn
, mrf
);
2288 brw_set_src0(p
, insn
, brw_null_reg());
2289 brw_inst_set_base_mrf(devinfo
, insn
, mrf
.nr
);
2292 brw_set_dp_read_message(p
,
2295 BRW_DATAPORT_OWORD_BLOCK_1_OWORDLOW
,
2296 BRW_DATAPORT_READ_MESSAGE_OWORD_BLOCK_READ
,
2297 BRW_DATAPORT_READ_TARGET_DATA_CACHE
,
2299 true, /* header_present */
2300 1); /* response_length (1 reg, 2 owords!) */
2302 brw_pop_insn_state(p
);
2306 void brw_fb_WRITE(struct brw_codegen
*p
,
2308 struct brw_reg payload
,
2309 struct brw_reg implied_header
,
2310 unsigned msg_control
,
2311 unsigned binding_table_index
,
2312 unsigned msg_length
,
2313 unsigned response_length
,
2315 bool last_render_target
,
2316 bool header_present
)
2318 const struct brw_device_info
*devinfo
= p
->devinfo
;
2321 struct brw_reg dest
, src0
;
2323 if (dispatch_width
== 16)
2324 dest
= retype(vec16(brw_null_reg()), BRW_REGISTER_TYPE_UW
);
2326 dest
= retype(vec8(brw_null_reg()), BRW_REGISTER_TYPE_UW
);
2328 if (devinfo
->gen
>= 6) {
2329 insn
= next_insn(p
, BRW_OPCODE_SENDC
);
2331 insn
= next_insn(p
, BRW_OPCODE_SEND
);
2333 brw_inst_set_qtr_control(devinfo
, insn
, BRW_COMPRESSION_NONE
);
2335 if (devinfo
->gen
>= 6) {
2336 /* headerless version, just submit color payload */
2339 msg_type
= GEN6_DATAPORT_WRITE_MESSAGE_RENDER_TARGET_WRITE
;
2341 assert(payload
.file
== BRW_MESSAGE_REGISTER_FILE
);
2342 brw_inst_set_base_mrf(devinfo
, insn
, payload
.nr
);
2343 src0
= implied_header
;
2345 msg_type
= BRW_DATAPORT_WRITE_MESSAGE_RENDER_TARGET_WRITE
;
2348 brw_set_dest(p
, insn
, dest
);
2349 brw_set_src0(p
, insn
, src0
);
2350 brw_set_dp_write_message(p
,
2352 binding_table_index
,
2360 0 /* send_commit_msg */);
2365 * Texture sample instruction.
2366 * Note: the msg_type plus msg_length values determine exactly what kind
2367 * of sampling operation is performed. See volume 4, page 161 of docs.
2369 void brw_SAMPLE(struct brw_codegen
*p
,
2370 struct brw_reg dest
,
2371 unsigned msg_reg_nr
,
2372 struct brw_reg src0
,
2373 unsigned binding_table_index
,
2376 unsigned response_length
,
2377 unsigned msg_length
,
2378 unsigned header_present
,
2380 unsigned return_format
)
2382 const struct brw_device_info
*devinfo
= p
->devinfo
;
2385 if (msg_reg_nr
!= -1)
2386 gen6_resolve_implied_move(p
, &src0
, msg_reg_nr
);
2388 insn
= next_insn(p
, BRW_OPCODE_SEND
);
2389 brw_inst_set_pred_control(devinfo
, insn
, BRW_PREDICATE_NONE
); /* XXX */
2391 /* From the 965 PRM (volume 4, part 1, section 14.2.41):
2393 * "Instruction compression is not allowed for this instruction (that
2394 * is, send). The hardware behavior is undefined if this instruction is
2395 * set as compressed. However, compress control can be set to "SecHalf"
2396 * to affect the EMask generation."
2398 * No similar wording is found in later PRMs, but there are examples
2399 * utilizing send with SecHalf. More importantly, SIMD8 sampler messages
2400 * are allowed in SIMD16 mode and they could not work without SecHalf. For
2401 * these reasons, we allow BRW_COMPRESSION_2NDHALF here.
2403 if (brw_inst_qtr_control(devinfo
, insn
) != BRW_COMPRESSION_2NDHALF
)
2404 brw_inst_set_qtr_control(devinfo
, insn
, BRW_COMPRESSION_NONE
);
2406 if (devinfo
->gen
< 6)
2407 brw_inst_set_base_mrf(devinfo
, insn
, msg_reg_nr
);
2409 brw_set_dest(p
, insn
, dest
);
2410 brw_set_src0(p
, insn
, src0
);
2411 brw_set_sampler_message(p
, insn
,
2412 binding_table_index
,
2422 /* Adjust the message header's sampler state pointer to
2423 * select the correct group of 16 samplers.
2425 void brw_adjust_sampler_state_pointer(struct brw_codegen
*p
,
2426 struct brw_reg header
,
2427 struct brw_reg sampler_index
)
2429 /* The "Sampler Index" field can only store values between 0 and 15.
2430 * However, we can add an offset to the "Sampler State Pointer"
2431 * field, effectively selecting a different set of 16 samplers.
2433 * The "Sampler State Pointer" needs to be aligned to a 32-byte
2434 * offset, and each sampler state is only 16-bytes, so we can't
2435 * exclusively use the offset - we have to use both.
2438 const struct brw_device_info
*devinfo
= p
->devinfo
;
2440 if (sampler_index
.file
== BRW_IMMEDIATE_VALUE
) {
2441 const int sampler_state_size
= 16; /* 16 bytes */
2442 uint32_t sampler
= sampler_index
.ud
;
2444 if (sampler
>= 16) {
2445 assert(devinfo
->is_haswell
|| devinfo
->gen
>= 8);
2447 get_element_ud(header
, 3),
2448 get_element_ud(brw_vec8_grf(0, 0), 3),
2449 brw_imm_ud(16 * (sampler
/ 16) * sampler_state_size
));
2452 /* Non-const sampler array indexing case */
2453 if (devinfo
->gen
< 8 && !devinfo
->is_haswell
) {
2457 struct brw_reg temp
= get_element_ud(header
, 3);
2459 brw_AND(p
, temp
, get_element_ud(sampler_index
, 0), brw_imm_ud(0x0f0));
2460 brw_SHL(p
, temp
, temp
, brw_imm_ud(4));
2462 get_element_ud(header
, 3),
2463 get_element_ud(brw_vec8_grf(0, 0), 3),
2468 /* All these variables are pretty confusing - we might be better off
2469 * using bitmasks and macros for this, in the old style. Or perhaps
2470 * just having the caller instantiate the fields in dword3 itself.
2472 void brw_urb_WRITE(struct brw_codegen
*p
,
2473 struct brw_reg dest
,
2474 unsigned msg_reg_nr
,
2475 struct brw_reg src0
,
2476 enum brw_urb_write_flags flags
,
2477 unsigned msg_length
,
2478 unsigned response_length
,
2482 const struct brw_device_info
*devinfo
= p
->devinfo
;
2485 gen6_resolve_implied_move(p
, &src0
, msg_reg_nr
);
2487 if (devinfo
->gen
>= 7 && !(flags
& BRW_URB_WRITE_USE_CHANNEL_MASKS
)) {
2488 /* Enable Channel Masks in the URB_WRITE_HWORD message header */
2489 brw_push_insn_state(p
);
2490 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
2491 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
2492 brw_OR(p
, retype(brw_vec1_reg(BRW_MESSAGE_REGISTER_FILE
, msg_reg_nr
, 5),
2493 BRW_REGISTER_TYPE_UD
),
2494 retype(brw_vec1_grf(0, 5), BRW_REGISTER_TYPE_UD
),
2495 brw_imm_ud(0xff00));
2496 brw_pop_insn_state(p
);
2499 insn
= next_insn(p
, BRW_OPCODE_SEND
);
2501 assert(msg_length
< BRW_MAX_MRF(devinfo
->gen
));
2503 brw_set_dest(p
, insn
, dest
);
2504 brw_set_src0(p
, insn
, src0
);
2505 brw_set_src1(p
, insn
, brw_imm_d(0));
2507 if (devinfo
->gen
< 6)
2508 brw_inst_set_base_mrf(devinfo
, insn
, msg_reg_nr
);
2510 brw_set_urb_message(p
,
2520 brw_send_indirect_message(struct brw_codegen
*p
,
2523 struct brw_reg payload
,
2524 struct brw_reg desc
)
2526 const struct brw_device_info
*devinfo
= p
->devinfo
;
2527 struct brw_inst
*send
;
2530 assert(desc
.type
== BRW_REGISTER_TYPE_UD
);
2532 /* We hold on to the setup instruction (the SEND in the direct case, the OR
2533 * in the indirect case) by its index in the instruction store. The
2534 * pointer returned by next_insn() may become invalid if emitting the SEND
2535 * in the indirect case reallocs the store.
2538 if (desc
.file
== BRW_IMMEDIATE_VALUE
) {
2540 send
= next_insn(p
, BRW_OPCODE_SEND
);
2541 brw_set_src1(p
, send
, desc
);
2544 struct brw_reg addr
= retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD
);
2546 brw_push_insn_state(p
);
2547 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
2548 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
2549 brw_set_default_predicate_control(p
, BRW_PREDICATE_NONE
);
2551 /* Load the indirect descriptor to an address register using OR so the
2552 * caller can specify additional descriptor bits with the usual
2553 * brw_set_*_message() helper functions.
2556 brw_OR(p
, addr
, desc
, brw_imm_ud(0));
2558 brw_pop_insn_state(p
);
2560 send
= next_insn(p
, BRW_OPCODE_SEND
);
2561 brw_set_src1(p
, send
, addr
);
2564 brw_set_dest(p
, send
, dst
);
2565 brw_set_src0(p
, send
, retype(payload
, BRW_REGISTER_TYPE_UD
));
2566 brw_inst_set_sfid(devinfo
, send
, sfid
);
2568 return &p
->store
[setup
];
2571 static struct brw_inst
*
2572 brw_send_indirect_surface_message(struct brw_codegen
*p
,
2575 struct brw_reg payload
,
2576 struct brw_reg surface
,
2577 unsigned message_len
,
2578 unsigned response_len
,
2579 bool header_present
)
2581 const struct brw_device_info
*devinfo
= p
->devinfo
;
2582 struct brw_inst
*insn
;
2584 if (surface
.file
!= BRW_IMMEDIATE_VALUE
) {
2585 struct brw_reg addr
= retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD
);
2587 brw_push_insn_state(p
);
2588 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
2589 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
2590 brw_set_default_predicate_control(p
, BRW_PREDICATE_NONE
);
2592 /* Mask out invalid bits from the surface index to avoid hangs e.g. when
2593 * some surface array is accessed out of bounds.
2595 insn
= brw_AND(p
, addr
,
2596 suboffset(vec1(retype(surface
, BRW_REGISTER_TYPE_UD
)),
2597 BRW_GET_SWZ(surface
.swizzle
, 0)),
2600 brw_pop_insn_state(p
);
2605 insn
= brw_send_indirect_message(p
, sfid
, dst
, payload
, surface
);
2606 brw_inst_set_mlen(devinfo
, insn
, message_len
);
2607 brw_inst_set_rlen(devinfo
, insn
, response_len
);
2608 brw_inst_set_header_present(devinfo
, insn
, header_present
);
2614 brw_find_next_block_end(struct brw_codegen
*p
, int start_offset
)
2617 void *store
= p
->store
;
2618 const struct brw_device_info
*devinfo
= p
->devinfo
;
2622 for (offset
= next_offset(devinfo
, store
, start_offset
);
2623 offset
< p
->next_insn_offset
;
2624 offset
= next_offset(devinfo
, store
, offset
)) {
2625 brw_inst
*insn
= store
+ offset
;
2627 switch (brw_inst_opcode(devinfo
, insn
)) {
2631 case BRW_OPCODE_ENDIF
:
2636 case BRW_OPCODE_ELSE
:
2637 case BRW_OPCODE_WHILE
:
2638 case BRW_OPCODE_HALT
:
2647 /* There is no DO instruction on gen6, so to find the end of the loop
2648 * we have to see if the loop is jumping back before our start
2652 brw_find_loop_end(struct brw_codegen
*p
, int start_offset
)
2654 const struct brw_device_info
*devinfo
= p
->devinfo
;
2656 int scale
= 16 / brw_jump_scale(devinfo
);
2657 void *store
= p
->store
;
2659 assert(devinfo
->gen
>= 6);
2661 /* Always start after the instruction (such as a WHILE) we're trying to fix
2664 for (offset
= next_offset(devinfo
, store
, start_offset
);
2665 offset
< p
->next_insn_offset
;
2666 offset
= next_offset(devinfo
, store
, offset
)) {
2667 brw_inst
*insn
= store
+ offset
;
2669 if (brw_inst_opcode(devinfo
, insn
) == BRW_OPCODE_WHILE
) {
2670 int jip
= devinfo
->gen
== 6 ? brw_inst_gen6_jump_count(devinfo
, insn
)
2671 : brw_inst_jip(devinfo
, insn
);
2672 if (offset
+ jip
* scale
<= start_offset
)
2676 assert(!"not reached");
2677 return start_offset
;
2680 /* After program generation, go back and update the UIP and JIP of
2681 * BREAK, CONT, and HALT instructions to their correct locations.
2684 brw_set_uip_jip(struct brw_codegen
*p
)
2686 const struct brw_device_info
*devinfo
= p
->devinfo
;
2688 int br
= brw_jump_scale(devinfo
);
2689 int scale
= 16 / br
;
2690 void *store
= p
->store
;
2692 if (devinfo
->gen
< 6)
2695 for (offset
= 0; offset
< p
->next_insn_offset
;
2696 offset
= next_offset(devinfo
, store
, offset
)) {
2697 brw_inst
*insn
= store
+ offset
;
2699 if (brw_inst_cmpt_control(devinfo
, insn
)) {
2700 /* Fixups for compacted BREAK/CONTINUE not supported yet. */
2701 assert(brw_inst_opcode(devinfo
, insn
) != BRW_OPCODE_BREAK
&&
2702 brw_inst_opcode(devinfo
, insn
) != BRW_OPCODE_CONTINUE
&&
2703 brw_inst_opcode(devinfo
, insn
) != BRW_OPCODE_HALT
);
2707 int block_end_offset
= brw_find_next_block_end(p
, offset
);
2708 switch (brw_inst_opcode(devinfo
, insn
)) {
2709 case BRW_OPCODE_BREAK
:
2710 assert(block_end_offset
!= 0);
2711 brw_inst_set_jip(devinfo
, insn
, (block_end_offset
- offset
) / scale
);
2712 /* Gen7 UIP points to WHILE; Gen6 points just after it */
2713 brw_inst_set_uip(devinfo
, insn
,
2714 (brw_find_loop_end(p
, offset
) - offset
+
2715 (devinfo
->gen
== 6 ? 16 : 0)) / scale
);
2717 case BRW_OPCODE_CONTINUE
:
2718 assert(block_end_offset
!= 0);
2719 brw_inst_set_jip(devinfo
, insn
, (block_end_offset
- offset
) / scale
);
2720 brw_inst_set_uip(devinfo
, insn
,
2721 (brw_find_loop_end(p
, offset
) - offset
) / scale
);
2723 assert(brw_inst_uip(devinfo
, insn
) != 0);
2724 assert(brw_inst_jip(devinfo
, insn
) != 0);
2727 case BRW_OPCODE_ENDIF
: {
2728 int32_t jump
= (block_end_offset
== 0) ?
2729 1 * br
: (block_end_offset
- offset
) / scale
;
2730 if (devinfo
->gen
>= 7)
2731 brw_inst_set_jip(devinfo
, insn
, jump
);
2733 brw_inst_set_gen6_jump_count(devinfo
, insn
, jump
);
2737 case BRW_OPCODE_HALT
:
2738 /* From the Sandy Bridge PRM (volume 4, part 2, section 8.3.19):
2740 * "In case of the halt instruction not inside any conditional
2741 * code block, the value of <JIP> and <UIP> should be the
2742 * same. In case of the halt instruction inside conditional code
2743 * block, the <UIP> should be the end of the program, and the
2744 * <JIP> should be end of the most inner conditional code block."
2746 * The uip will have already been set by whoever set up the
2749 if (block_end_offset
== 0) {
2750 brw_inst_set_jip(devinfo
, insn
, brw_inst_uip(devinfo
, insn
));
2752 brw_inst_set_jip(devinfo
, insn
, (block_end_offset
- offset
) / scale
);
2754 assert(brw_inst_uip(devinfo
, insn
) != 0);
2755 assert(brw_inst_jip(devinfo
, insn
) != 0);
2761 void brw_ff_sync(struct brw_codegen
*p
,
2762 struct brw_reg dest
,
2763 unsigned msg_reg_nr
,
2764 struct brw_reg src0
,
2766 unsigned response_length
,
2769 const struct brw_device_info
*devinfo
= p
->devinfo
;
2772 gen6_resolve_implied_move(p
, &src0
, msg_reg_nr
);
2774 insn
= next_insn(p
, BRW_OPCODE_SEND
);
2775 brw_set_dest(p
, insn
, dest
);
2776 brw_set_src0(p
, insn
, src0
);
2777 brw_set_src1(p
, insn
, brw_imm_d(0));
2779 if (devinfo
->gen
< 6)
2780 brw_inst_set_base_mrf(devinfo
, insn
, msg_reg_nr
);
2782 brw_set_ff_sync_message(p
,
2790 * Emit the SEND instruction necessary to generate stream output data on Gen6
2791 * (for transform feedback).
2793 * If send_commit_msg is true, this is the last piece of stream output data
2794 * from this thread, so send the data as a committed write. According to the
2795 * Sandy Bridge PRM (volume 2 part 1, section 4.5.1):
2797 * "Prior to End of Thread with a URB_WRITE, the kernel must ensure all
2798 * writes are complete by sending the final write as a committed write."
2801 brw_svb_write(struct brw_codegen
*p
,
2802 struct brw_reg dest
,
2803 unsigned msg_reg_nr
,
2804 struct brw_reg src0
,
2805 unsigned binding_table_index
,
2806 bool send_commit_msg
)
2810 gen6_resolve_implied_move(p
, &src0
, msg_reg_nr
);
2812 insn
= next_insn(p
, BRW_OPCODE_SEND
);
2813 brw_set_dest(p
, insn
, dest
);
2814 brw_set_src0(p
, insn
, src0
);
2815 brw_set_src1(p
, insn
, brw_imm_d(0));
2816 brw_set_dp_write_message(p
, insn
,
2817 binding_table_index
,
2818 0, /* msg_control: ignored */
2819 GEN6_DATAPORT_WRITE_MESSAGE_STREAMED_VB_WRITE
,
2821 true, /* header_present */
2822 0, /* last_render_target: ignored */
2823 send_commit_msg
, /* response_length */
2824 0, /* end_of_thread */
2825 send_commit_msg
); /* send_commit_msg */
2829 brw_surface_payload_size(struct brw_codegen
*p
,
2830 unsigned num_channels
,
2834 if (has_simd4x2
&& brw_inst_access_mode(p
->devinfo
, p
->current
) == BRW_ALIGN_16
)
2836 else if (has_simd16
&& p
->compressed
)
2837 return 2 * num_channels
;
2839 return num_channels
;
2843 brw_set_dp_untyped_atomic_message(struct brw_codegen
*p
,
2846 bool response_expected
)
2848 const struct brw_device_info
*devinfo
= p
->devinfo
;
2849 unsigned msg_control
=
2850 atomic_op
| /* Atomic Operation Type: BRW_AOP_* */
2851 (response_expected
? 1 << 5 : 0); /* Return data expected */
2853 if (devinfo
->gen
>= 8 || devinfo
->is_haswell
) {
2854 if (brw_inst_access_mode(devinfo
, p
->current
) == BRW_ALIGN_1
) {
2856 msg_control
|= 1 << 4; /* SIMD8 mode */
2858 brw_inst_set_dp_msg_type(devinfo
, insn
,
2859 HSW_DATAPORT_DC_PORT1_UNTYPED_ATOMIC_OP
);
2861 brw_inst_set_dp_msg_type(devinfo
, insn
,
2862 HSW_DATAPORT_DC_PORT1_UNTYPED_ATOMIC_OP_SIMD4X2
);
2865 brw_inst_set_dp_msg_type(devinfo
, insn
,
2866 GEN7_DATAPORT_DC_UNTYPED_ATOMIC_OP
);
2869 msg_control
|= 1 << 4; /* SIMD8 mode */
2872 brw_inst_set_dp_msg_control(devinfo
, insn
, msg_control
);
2876 brw_untyped_atomic(struct brw_codegen
*p
,
2878 struct brw_reg payload
,
2879 struct brw_reg surface
,
2881 unsigned msg_length
,
2882 bool response_expected
)
2884 const struct brw_device_info
*devinfo
= p
->devinfo
;
2885 const unsigned sfid
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
?
2886 HSW_SFID_DATAPORT_DATA_CACHE_1
:
2887 GEN7_SFID_DATAPORT_DATA_CACHE
);
2888 const bool align1
= brw_inst_access_mode(devinfo
, p
->current
) == BRW_ALIGN_1
;
2889 /* Mask out unused components -- This is especially important in Align16
2890 * mode on generations that don't have native support for SIMD4x2 atomics,
2891 * because unused but enabled components will cause the dataport to perform
2892 * additional atomic operations on the addresses that happen to be in the
2893 * uninitialized Y, Z and W coordinates of the payload.
2895 const unsigned mask
= align1
? WRITEMASK_XYZW
: WRITEMASK_X
;
2896 struct brw_inst
*insn
= brw_send_indirect_surface_message(
2897 p
, sfid
, brw_writemask(dst
, mask
), payload
, surface
, msg_length
,
2898 brw_surface_payload_size(p
, response_expected
,
2899 devinfo
->gen
>= 8 || devinfo
->is_haswell
, true),
2902 brw_set_dp_untyped_atomic_message(
2903 p
, insn
, atomic_op
, response_expected
);
2907 brw_set_dp_untyped_surface_read_message(struct brw_codegen
*p
,
2908 struct brw_inst
*insn
,
2909 unsigned num_channels
)
2911 const struct brw_device_info
*devinfo
= p
->devinfo
;
2912 /* Set mask of 32-bit channels to drop. */
2913 unsigned msg_control
= 0xf & (0xf << num_channels
);
2915 if (brw_inst_access_mode(devinfo
, p
->current
) == BRW_ALIGN_1
) {
2917 msg_control
|= 1 << 4; /* SIMD16 mode */
2919 msg_control
|= 2 << 4; /* SIMD8 mode */
2922 brw_inst_set_dp_msg_type(devinfo
, insn
,
2923 (devinfo
->gen
>= 8 || devinfo
->is_haswell
?
2924 HSW_DATAPORT_DC_PORT1_UNTYPED_SURFACE_READ
:
2925 GEN7_DATAPORT_DC_UNTYPED_SURFACE_READ
));
2926 brw_inst_set_dp_msg_control(devinfo
, insn
, msg_control
);
2930 brw_untyped_surface_read(struct brw_codegen
*p
,
2932 struct brw_reg payload
,
2933 struct brw_reg surface
,
2934 unsigned msg_length
,
2935 unsigned num_channels
)
2937 const struct brw_device_info
*devinfo
= p
->devinfo
;
2938 const unsigned sfid
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
?
2939 HSW_SFID_DATAPORT_DATA_CACHE_1
:
2940 GEN7_SFID_DATAPORT_DATA_CACHE
);
2941 struct brw_inst
*insn
= brw_send_indirect_surface_message(
2942 p
, sfid
, dst
, payload
, surface
, msg_length
,
2943 brw_surface_payload_size(p
, num_channels
, true, true),
2946 brw_set_dp_untyped_surface_read_message(
2947 p
, insn
, num_channels
);
2951 brw_set_dp_untyped_surface_write_message(struct brw_codegen
*p
,
2952 struct brw_inst
*insn
,
2953 unsigned num_channels
)
2955 const struct brw_device_info
*devinfo
= p
->devinfo
;
2956 /* Set mask of 32-bit channels to drop. */
2957 unsigned msg_control
= 0xf & (0xf << num_channels
);
2959 if (brw_inst_access_mode(devinfo
, p
->current
) == BRW_ALIGN_1
) {
2961 msg_control
|= 1 << 4; /* SIMD16 mode */
2963 msg_control
|= 2 << 4; /* SIMD8 mode */
2965 if (devinfo
->gen
>= 8 || devinfo
->is_haswell
)
2966 msg_control
|= 0 << 4; /* SIMD4x2 mode */
2968 msg_control
|= 2 << 4; /* SIMD8 mode */
2971 brw_inst_set_dp_msg_type(devinfo
, insn
,
2972 devinfo
->gen
>= 8 || devinfo
->is_haswell
?
2973 HSW_DATAPORT_DC_PORT1_UNTYPED_SURFACE_WRITE
:
2974 GEN7_DATAPORT_DC_UNTYPED_SURFACE_WRITE
);
2975 brw_inst_set_dp_msg_control(devinfo
, insn
, msg_control
);
2979 brw_untyped_surface_write(struct brw_codegen
*p
,
2980 struct brw_reg payload
,
2981 struct brw_reg surface
,
2982 unsigned msg_length
,
2983 unsigned num_channels
)
2985 const struct brw_device_info
*devinfo
= p
->devinfo
;
2986 const unsigned sfid
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
?
2987 HSW_SFID_DATAPORT_DATA_CACHE_1
:
2988 GEN7_SFID_DATAPORT_DATA_CACHE
);
2989 const bool align1
= brw_inst_access_mode(devinfo
, p
->current
) == BRW_ALIGN_1
;
2990 /* Mask out unused components -- See comment in brw_untyped_atomic(). */
2991 const unsigned mask
= devinfo
->gen
== 7 && !devinfo
->is_haswell
&& !align1
?
2992 WRITEMASK_X
: WRITEMASK_XYZW
;
2993 struct brw_inst
*insn
= brw_send_indirect_surface_message(
2994 p
, sfid
, brw_writemask(brw_null_reg(), mask
),
2995 payload
, surface
, msg_length
, 0, align1
);
2997 brw_set_dp_untyped_surface_write_message(
2998 p
, insn
, num_channels
);
3002 brw_set_dp_typed_atomic_message(struct brw_codegen
*p
,
3003 struct brw_inst
*insn
,
3005 bool response_expected
)
3007 const struct brw_device_info
*devinfo
= p
->devinfo
;
3008 unsigned msg_control
=
3009 atomic_op
| /* Atomic Operation Type: BRW_AOP_* */
3010 (response_expected
? 1 << 5 : 0); /* Return data expected */
3012 if (devinfo
->gen
>= 8 || devinfo
->is_haswell
) {
3013 if (brw_inst_access_mode(devinfo
, p
->current
) == BRW_ALIGN_1
) {
3014 if (brw_inst_qtr_control(devinfo
, p
->current
) == GEN6_COMPRESSION_2Q
)
3015 msg_control
|= 1 << 4; /* Use high 8 slots of the sample mask */
3017 brw_inst_set_dp_msg_type(devinfo
, insn
,
3018 HSW_DATAPORT_DC_PORT1_TYPED_ATOMIC_OP
);
3020 brw_inst_set_dp_msg_type(devinfo
, insn
,
3021 HSW_DATAPORT_DC_PORT1_TYPED_ATOMIC_OP_SIMD4X2
);
3025 brw_inst_set_dp_msg_type(devinfo
, insn
,
3026 GEN7_DATAPORT_RC_TYPED_ATOMIC_OP
);
3028 if (brw_inst_qtr_control(devinfo
, p
->current
) == GEN6_COMPRESSION_2Q
)
3029 msg_control
|= 1 << 4; /* Use high 8 slots of the sample mask */
3032 brw_inst_set_dp_msg_control(devinfo
, insn
, msg_control
);
3036 brw_typed_atomic(struct brw_codegen
*p
,
3038 struct brw_reg payload
,
3039 struct brw_reg surface
,
3041 unsigned msg_length
,
3042 bool response_expected
) {
3043 const struct brw_device_info
*devinfo
= p
->devinfo
;
3044 const unsigned sfid
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
?
3045 HSW_SFID_DATAPORT_DATA_CACHE_1
:
3046 GEN6_SFID_DATAPORT_RENDER_CACHE
);
3047 const bool align1
= (brw_inst_access_mode(devinfo
, p
->current
) == BRW_ALIGN_1
);
3048 /* Mask out unused components -- See comment in brw_untyped_atomic(). */
3049 const unsigned mask
= align1
? WRITEMASK_XYZW
: WRITEMASK_X
;
3050 struct brw_inst
*insn
= brw_send_indirect_surface_message(
3051 p
, sfid
, brw_writemask(dst
, mask
), payload
, surface
, msg_length
,
3052 brw_surface_payload_size(p
, response_expected
,
3053 devinfo
->gen
>= 8 || devinfo
->is_haswell
, false),
3056 brw_set_dp_typed_atomic_message(
3057 p
, insn
, atomic_op
, response_expected
);
3061 brw_set_dp_typed_surface_read_message(struct brw_codegen
*p
,
3062 struct brw_inst
*insn
,
3063 unsigned num_channels
)
3065 const struct brw_device_info
*devinfo
= p
->devinfo
;
3066 /* Set mask of unused channels. */
3067 unsigned msg_control
= 0xf & (0xf << num_channels
);
3069 if (devinfo
->gen
>= 8 || devinfo
->is_haswell
) {
3070 if (brw_inst_access_mode(devinfo
, p
->current
) == BRW_ALIGN_1
) {
3071 if (brw_inst_qtr_control(devinfo
, p
->current
) == GEN6_COMPRESSION_2Q
)
3072 msg_control
|= 2 << 4; /* Use high 8 slots of the sample mask */
3074 msg_control
|= 1 << 4; /* Use low 8 slots of the sample mask */
3077 brw_inst_set_dp_msg_type(devinfo
, insn
,
3078 HSW_DATAPORT_DC_PORT1_TYPED_SURFACE_READ
);
3080 if (brw_inst_access_mode(devinfo
, p
->current
) == BRW_ALIGN_1
) {
3081 if (brw_inst_qtr_control(devinfo
, p
->current
) == GEN6_COMPRESSION_2Q
)
3082 msg_control
|= 1 << 5; /* Use high 8 slots of the sample mask */
3085 brw_inst_set_dp_msg_type(devinfo
, insn
,
3086 GEN7_DATAPORT_RC_TYPED_SURFACE_READ
);
3089 brw_inst_set_dp_msg_control(devinfo
, insn
, msg_control
);
3093 brw_typed_surface_read(struct brw_codegen
*p
,
3095 struct brw_reg payload
,
3096 struct brw_reg surface
,
3097 unsigned msg_length
,
3098 unsigned num_channels
)
3100 const struct brw_device_info
*devinfo
= p
->devinfo
;
3101 const unsigned sfid
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
?
3102 HSW_SFID_DATAPORT_DATA_CACHE_1
:
3103 GEN6_SFID_DATAPORT_RENDER_CACHE
);
3104 struct brw_inst
*insn
= brw_send_indirect_surface_message(
3105 p
, sfid
, dst
, payload
, surface
, msg_length
,
3106 brw_surface_payload_size(p
, num_channels
,
3107 devinfo
->gen
>= 8 || devinfo
->is_haswell
, false),
3110 brw_set_dp_typed_surface_read_message(
3111 p
, insn
, num_channels
);
3115 brw_set_dp_typed_surface_write_message(struct brw_codegen
*p
,
3116 struct brw_inst
*insn
,
3117 unsigned num_channels
)
3119 const struct brw_device_info
*devinfo
= p
->devinfo
;
3120 /* Set mask of unused channels. */
3121 unsigned msg_control
= 0xf & (0xf << num_channels
);
3123 if (devinfo
->gen
>= 8 || devinfo
->is_haswell
) {
3124 if (brw_inst_access_mode(devinfo
, p
->current
) == BRW_ALIGN_1
) {
3125 if (brw_inst_qtr_control(devinfo
, p
->current
) == GEN6_COMPRESSION_2Q
)
3126 msg_control
|= 2 << 4; /* Use high 8 slots of the sample mask */
3128 msg_control
|= 1 << 4; /* Use low 8 slots of the sample mask */
3131 brw_inst_set_dp_msg_type(devinfo
, insn
,
3132 HSW_DATAPORT_DC_PORT1_TYPED_SURFACE_WRITE
);
3135 if (brw_inst_access_mode(devinfo
, p
->current
) == BRW_ALIGN_1
) {
3136 if (brw_inst_qtr_control(devinfo
, p
->current
) == GEN6_COMPRESSION_2Q
)
3137 msg_control
|= 1 << 5; /* Use high 8 slots of the sample mask */
3140 brw_inst_set_dp_msg_type(devinfo
, insn
,
3141 GEN7_DATAPORT_RC_TYPED_SURFACE_WRITE
);
3144 brw_inst_set_dp_msg_control(devinfo
, insn
, msg_control
);
3148 brw_typed_surface_write(struct brw_codegen
*p
,
3149 struct brw_reg payload
,
3150 struct brw_reg surface
,
3151 unsigned msg_length
,
3152 unsigned num_channels
)
3154 const struct brw_device_info
*devinfo
= p
->devinfo
;
3155 const unsigned sfid
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
?
3156 HSW_SFID_DATAPORT_DATA_CACHE_1
:
3157 GEN6_SFID_DATAPORT_RENDER_CACHE
);
3158 const bool align1
= (brw_inst_access_mode(devinfo
, p
->current
) == BRW_ALIGN_1
);
3159 /* Mask out unused components -- See comment in brw_untyped_atomic(). */
3160 const unsigned mask
= (devinfo
->gen
== 7 && !devinfo
->is_haswell
&& !align1
?
3161 WRITEMASK_X
: WRITEMASK_XYZW
);
3162 struct brw_inst
*insn
= brw_send_indirect_surface_message(
3163 p
, sfid
, brw_writemask(brw_null_reg(), mask
),
3164 payload
, surface
, msg_length
, 0, true);
3166 brw_set_dp_typed_surface_write_message(
3167 p
, insn
, num_channels
);
3171 brw_set_memory_fence_message(struct brw_codegen
*p
,
3172 struct brw_inst
*insn
,
3173 enum brw_message_target sfid
,
3176 const struct brw_device_info
*devinfo
= p
->devinfo
;
3178 brw_set_message_descriptor(p
, insn
, sfid
,
3179 1 /* message length */,
3180 (commit_enable
? 1 : 0) /* response length */,
3181 true /* header present */,
3185 case GEN6_SFID_DATAPORT_RENDER_CACHE
:
3186 brw_inst_set_dp_msg_type(devinfo
, insn
, GEN7_DATAPORT_RC_MEMORY_FENCE
);
3188 case GEN7_SFID_DATAPORT_DATA_CACHE
:
3189 brw_inst_set_dp_msg_type(devinfo
, insn
, GEN7_DATAPORT_DC_MEMORY_FENCE
);
3192 unreachable("Not reached");
3196 brw_inst_set_dp_msg_control(devinfo
, insn
, 1 << 5);
3200 brw_memory_fence(struct brw_codegen
*p
,
3203 const struct brw_device_info
*devinfo
= p
->devinfo
;
3204 const bool commit_enable
= devinfo
->gen
== 7 && !devinfo
->is_haswell
;
3205 struct brw_inst
*insn
;
3207 /* Set dst as destination for dependency tracking, the MEMORY_FENCE
3208 * message doesn't write anything back.
3210 insn
= next_insn(p
, BRW_OPCODE_SEND
);
3211 brw_set_dest(p
, insn
, dst
);
3212 brw_set_src0(p
, insn
, dst
);
3213 brw_set_memory_fence_message(p
, insn
, GEN7_SFID_DATAPORT_DATA_CACHE
,
3216 if (devinfo
->gen
== 7 && !devinfo
->is_haswell
) {
3217 /* IVB does typed surface access through the render cache, so we need to
3218 * flush it too. Use a different register so both flushes can be
3219 * pipelined by the hardware.
3221 insn
= next_insn(p
, BRW_OPCODE_SEND
);
3222 brw_set_dest(p
, insn
, offset(dst
, 1));
3223 brw_set_src0(p
, insn
, offset(dst
, 1));
3224 brw_set_memory_fence_message(p
, insn
, GEN6_SFID_DATAPORT_RENDER_CACHE
,
3227 /* Now write the response of the second message into the response of the
3228 * first to trigger a pipeline stall -- This way future render and data
3229 * cache messages will be properly ordered with respect to past data and
3230 * render cache messages.
3232 brw_push_insn_state(p
);
3233 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
3234 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
3235 brw_MOV(p
, dst
, offset(dst
, 1));
3236 brw_pop_insn_state(p
);
3241 brw_pixel_interpolator_query(struct brw_codegen
*p
,
3242 struct brw_reg dest
,
3246 struct brw_reg data
,
3247 unsigned msg_length
,
3248 unsigned response_length
)
3250 const struct brw_device_info
*devinfo
= p
->devinfo
;
3251 struct brw_inst
*insn
;
3252 const uint16_t exec_size
= brw_inst_exec_size(devinfo
, p
->current
);
3254 /* brw_send_indirect_message will automatically use a direct send message
3255 * if data is actually immediate.
3257 insn
= brw_send_indirect_message(p
,
3258 GEN7_SFID_PIXEL_INTERPOLATOR
,
3262 brw_inst_set_mlen(devinfo
, insn
, msg_length
);
3263 brw_inst_set_rlen(devinfo
, insn
, response_length
);
3265 brw_inst_set_pi_simd_mode(devinfo
, insn
, exec_size
== BRW_EXECUTE_16
);
3266 brw_inst_set_pi_slot_group(devinfo
, insn
, 0); /* zero unless 32/64px dispatch */
3267 brw_inst_set_pi_nopersp(devinfo
, insn
, noperspective
);
3268 brw_inst_set_pi_message_type(devinfo
, insn
, mode
);
3272 brw_find_live_channel(struct brw_codegen
*p
, struct brw_reg dst
)
3274 const struct brw_device_info
*devinfo
= p
->devinfo
;
3277 assert(devinfo
->gen
>= 7);
3279 brw_push_insn_state(p
);
3281 if (brw_inst_access_mode(devinfo
, p
->current
) == BRW_ALIGN_1
) {
3282 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
3284 if (devinfo
->gen
>= 8) {
3285 /* Getting the first active channel index is easy on Gen8: Just find
3286 * the first bit set in the mask register. The same register exists
3287 * on HSW already but it reads back as all ones when the current
3288 * instruction has execution masking disabled, so it's kind of
3291 inst
= brw_FBL(p
, vec1(dst
),
3292 retype(brw_mask_reg(0), BRW_REGISTER_TYPE_UD
));
3294 /* Quarter control has the effect of magically shifting the value of
3295 * this register. Make sure it's set to zero.
3297 brw_inst_set_qtr_control(devinfo
, inst
, GEN6_COMPRESSION_1Q
);
3299 const struct brw_reg flag
= retype(brw_flag_reg(1, 0),
3300 BRW_REGISTER_TYPE_UD
);
3302 brw_MOV(p
, flag
, brw_imm_ud(0));
3304 /* Run a 16-wide instruction returning zero with execution masking
3305 * and a conditional modifier enabled in order to get the current
3306 * execution mask in f1.0.
3308 inst
= brw_MOV(p
, brw_null_reg(), brw_imm_ud(0));
3309 brw_inst_set_exec_size(devinfo
, inst
, BRW_EXECUTE_16
);
3310 brw_inst_set_mask_control(devinfo
, inst
, BRW_MASK_ENABLE
);
3311 brw_inst_set_cond_modifier(devinfo
, inst
, BRW_CONDITIONAL_Z
);
3312 brw_inst_set_flag_reg_nr(devinfo
, inst
, 1);
3314 brw_FBL(p
, vec1(dst
), flag
);
3317 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
3319 if (devinfo
->gen
>= 8) {
3320 /* In SIMD4x2 mode the first active channel index is just the
3321 * negation of the first bit of the mask register.
3323 inst
= brw_AND(p
, brw_writemask(dst
, WRITEMASK_X
),
3324 negate(retype(brw_mask_reg(0), BRW_REGISTER_TYPE_UD
)),
3328 /* Overwrite the destination without and with execution masking to
3329 * find out which of the channels is active.
3331 brw_MOV(p
, brw_writemask(vec4(dst
), WRITEMASK_X
),
3334 inst
= brw_MOV(p
, brw_writemask(vec4(dst
), WRITEMASK_X
),
3336 brw_inst_set_mask_control(devinfo
, inst
, BRW_MASK_ENABLE
);
3340 brw_pop_insn_state(p
);
3344 brw_broadcast(struct brw_codegen
*p
,
3349 const struct brw_device_info
*devinfo
= p
->devinfo
;
3350 const bool align1
= brw_inst_access_mode(devinfo
, p
->current
) == BRW_ALIGN_1
;
3353 assert(src
.file
== BRW_GENERAL_REGISTER_FILE
&&
3354 src
.address_mode
== BRW_ADDRESS_DIRECT
);
3356 if ((src
.vstride
== 0 && (src
.hstride
== 0 || !align1
)) ||
3357 idx
.file
== BRW_IMMEDIATE_VALUE
) {
3358 /* Trivial, the source is already uniform or the index is a constant.
3359 * We will typically not get here if the optimizer is doing its job, but
3360 * asserting would be mean.
3362 const unsigned i
= idx
.file
== BRW_IMMEDIATE_VALUE
? idx
.ud
: 0;
3364 (align1
? stride(suboffset(src
, i
), 0, 1, 0) :
3365 stride(suboffset(src
, 4 * i
), 0, 4, 1)));
3368 const struct brw_reg addr
=
3369 retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD
);
3370 const unsigned offset
= src
.nr
* REG_SIZE
+ src
.subnr
;
3371 /* Limit in bytes of the signed indirect addressing immediate. */
3372 const unsigned limit
= 512;
3374 brw_push_insn_state(p
);
3375 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
3376 brw_set_default_predicate_control(p
, BRW_PREDICATE_NONE
);
3378 /* Take into account the component size and horizontal stride. */
3379 assert(src
.vstride
== src
.hstride
+ src
.width
);
3380 brw_SHL(p
, addr
, vec1(idx
),
3381 brw_imm_ud(_mesa_logbase2(type_sz(src
.type
)) +
3384 /* We can only address up to limit bytes using the indirect
3385 * addressing immediate, account for the difference if the source
3386 * register is above this limit.
3388 if (offset
>= limit
)
3389 brw_ADD(p
, addr
, addr
, brw_imm_ud(offset
- offset
% limit
));
3391 brw_pop_insn_state(p
);
3393 /* Use indirect addressing to fetch the specified component. */
3395 retype(brw_vec1_indirect(addr
.subnr
, offset
% limit
),
3398 /* In SIMD4x2 mode the index can be either zero or one, replicate it
3399 * to all bits of a flag register,
3403 stride(brw_swizzle1(idx
, 0), 0, 4, 1));
3404 brw_inst_set_pred_control(devinfo
, inst
, BRW_PREDICATE_NONE
);
3405 brw_inst_set_cond_modifier(devinfo
, inst
, BRW_CONDITIONAL_NZ
);
3406 brw_inst_set_flag_reg_nr(devinfo
, inst
, 1);
3408 /* and use predicated SEL to pick the right channel. */
3409 inst
= brw_SEL(p
, dst
,
3410 stride(suboffset(src
, 4), 0, 4, 1),
3411 stride(src
, 0, 4, 1));
3412 brw_inst_set_pred_control(devinfo
, inst
, BRW_PREDICATE_NORMAL
);
3413 brw_inst_set_flag_reg_nr(devinfo
, inst
, 1);
3419 * This instruction is generated as a single-channel align1 instruction by
3420 * both the VS and FS stages when using INTEL_DEBUG=shader_time.
3422 * We can't use the typed atomic op in the FS because that has the execution
3423 * mask ANDed with the pixel mask, but we just want to write the one dword for
3426 * We don't use the SIMD4x2 atomic ops in the VS because want to just write
3427 * one u32. So we use the same untyped atomic write message as the pixel
3430 * The untyped atomic operation requires a BUFFER surface type with RAW
3431 * format, and is only accessible through the legacy DATA_CACHE dataport
3434 void brw_shader_time_add(struct brw_codegen
*p
,
3435 struct brw_reg payload
,
3436 uint32_t surf_index
)
3438 const unsigned sfid
= (p
->devinfo
->gen
>= 8 || p
->devinfo
->is_haswell
?
3439 HSW_SFID_DATAPORT_DATA_CACHE_1
:
3440 GEN7_SFID_DATAPORT_DATA_CACHE
);
3441 assert(p
->devinfo
->gen
>= 7);
3443 brw_push_insn_state(p
);
3444 brw_set_default_access_mode(p
, BRW_ALIGN_1
);
3445 brw_set_default_mask_control(p
, BRW_MASK_DISABLE
);
3446 brw_set_default_compression_control(p
, BRW_COMPRESSION_NONE
);
3447 brw_inst
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
3449 /* We use brw_vec1_reg and unmasked because we want to increment the given
3452 brw_set_dest(p
, send
, brw_vec1_reg(BRW_ARCHITECTURE_REGISTER_FILE
,
3454 brw_set_src0(p
, send
, brw_vec1_reg(payload
.file
,
3456 brw_set_src1(p
, send
, brw_imm_ud(0));
3457 brw_set_message_descriptor(p
, send
, sfid
, 2, 0, false, false);
3458 brw_inst_set_binding_table_index(p
->devinfo
, send
, surf_index
);
3459 brw_set_dp_untyped_atomic_message(p
, send
, BRW_AOP_ADD
, false);
3461 brw_pop_insn_state(p
);
3466 * Emit the SEND message for a barrier
3469 brw_barrier(struct brw_codegen
*p
, struct brw_reg src
)
3471 const struct brw_device_info
*devinfo
= p
->devinfo
;
3472 struct brw_inst
*inst
;
3474 assert(devinfo
->gen
>= 7);
3476 inst
= next_insn(p
, BRW_OPCODE_SEND
);
3477 brw_set_dest(p
, inst
, brw_null_reg());
3478 brw_set_src0(p
, inst
, src
);
3479 brw_set_src1(p
, inst
, brw_null_reg());
3481 brw_set_message_descriptor(p
, inst
, BRW_SFID_MESSAGE_GATEWAY
,
3483 0 /* response_length */,
3484 false /* header_present */,
3485 false /* end_of_thread */);
3487 brw_inst_set_gateway_notify(devinfo
, inst
, 1);
3488 brw_inst_set_gateway_subfuncid(devinfo
, inst
,
3489 BRW_MESSAGE_GATEWAY_SFID_BARRIER_MSG
);
3491 brw_inst_set_mask_control(devinfo
, inst
, BRW_MASK_DISABLE
);
3496 * Emit the wait instruction for a barrier
3499 brw_WAIT(struct brw_codegen
*p
)
3501 const struct brw_device_info
*devinfo
= p
->devinfo
;
3502 struct brw_inst
*insn
;
3504 struct brw_reg src
= brw_notification_reg();
3506 insn
= next_insn(p
, BRW_OPCODE_WAIT
);
3507 brw_set_dest(p
, insn
, src
);
3508 brw_set_src0(p
, insn
, src
);
3509 brw_set_src1(p
, insn
, brw_null_reg());
3511 brw_inst_set_exec_size(devinfo
, insn
, BRW_EXECUTE_1
);
3512 brw_inst_set_mask_control(devinfo
, insn
, BRW_MASK_DISABLE
);