1 /**************************************************************************
3 * Copyright 2011-2012 Advanced Micro Devices, Inc.
4 * Copyright 2009 VMware, Inc.
5 * Copyright 2007-2008 Tungsten Graphics, Inc., Cedar Park, Texas.
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the
10 * "Software"), to deal in the Software without restriction, including
11 * without limitation the rights to use, copy, modify, merge, publish,
12 * distribute, sub license, and/or sell copies of the Software, and to
13 * permit persons to whom the Software is furnished to do so, subject to
14 * the following conditions:
16 * The above copyright notice and this permission notice (including the
17 * next paragraph) shall be included in all copies or substantial portions
20 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
21 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
22 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
23 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
24 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
25 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
26 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
28 **************************************************************************/
32 * TGSI to LLVM IR translation.
34 * @author Jose Fonseca <jfonseca@vmware.com>
35 * @author Tom Stellard <thomas.stellard@amd.com>
37 * Based on tgsi_sse2.c code written by Michal Krol, Keith Whitwell,
38 * Brian Paul, and others.
42 #include "lp_bld_tgsi_action.h"
44 #include "lp_bld_tgsi.h"
45 #include "lp_bld_arit.h"
46 #include "lp_bld_bitarit.h"
47 #include "lp_bld_const.h"
48 #include "lp_bld_gather.h"
49 #include "lp_bld_logic.h"
51 #include "tgsi/tgsi_exec.h"
53 /* XXX: The CPU only defaults should be repaced by generic ones. In most
54 * cases, the CPU defaults are just wrappers around a function in
55 * lp_build_arit.c and these functions should be inlined here and the CPU
56 * generic code should be removed and placed elsewhere.
61 /* Generic fetch_arg functions */
63 static void scalar_unary_fetch_args(
64 struct lp_build_tgsi_context
* bld_base
,
65 struct lp_build_emit_data
* emit_data
)
68 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, 0);
69 emit_data
->arg_count
= 1;
70 emit_data
->dst_type
= LLVMTypeOf(emit_data
->args
[0]);
73 static void scalar_binary_fetch_args(
74 struct lp_build_tgsi_context
* bld_base
,
75 struct lp_build_emit_data
* emit_data
)
78 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
81 emit_data
->args
[1] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
83 emit_data
->arg_count
= 2;
84 emit_data
->dst_type
= LLVMTypeOf(emit_data
->args
[0]);
90 const struct lp_build_tgsi_action
* action
,
91 struct lp_build_tgsi_context
* bld_base
,
92 struct lp_build_emit_data
* emit_data
)
94 emit_data
->output
[emit_data
->chan
] = LLVMBuildFAdd(
95 bld_base
->base
.gallivm
->builder
,
96 emit_data
->args
[0], emit_data
->args
[1], "");
102 const struct lp_build_tgsi_action
* action
,
103 struct lp_build_tgsi_context
* bld_base
,
104 struct lp_build_emit_data
* emit_data
)
106 LLVMValueRef tmp
= lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_ROUND
, emit_data
->args
[0]);
107 emit_data
->output
[emit_data
->chan
] = LLVMBuildFPToSI(bld_base
->base
.gallivm
->builder
, tmp
,
108 bld_base
->uint_bld
.vec_type
, "");
111 /* TGSI_OPCODE_CLAMP */
114 const struct lp_build_tgsi_action
* action
,
115 struct lp_build_tgsi_context
* bld_base
,
116 struct lp_build_emit_data
* emit_data
)
119 tmp
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MAX
,
122 emit_data
->output
[emit_data
->chan
] = lp_build_emit_llvm_binary(bld_base
,
123 TGSI_OPCODE_MIN
, tmp
, emit_data
->args
[2]);
130 struct lp_build_tgsi_context
* bld_base
,
131 struct lp_build_emit_data
* emit_data
,
132 unsigned dp_components
)
135 for (src
= 0; src
< 2; src
++) {
136 for (chan
= 0; chan
< dp_components
; chan
++) {
137 emit_data
->args
[(src
* dp_components
) + chan
] =
138 lp_build_emit_fetch(bld_base
, emit_data
->inst
, src
, chan
);
141 emit_data
->dst_type
= bld_base
->base
.elem_type
;
144 /* TGSI_OPCODE_DP2 */
147 struct lp_build_tgsi_context
* bld_base
,
148 struct lp_build_emit_data
* emit_data
)
150 dp_fetch_args(bld_base
, emit_data
, 2);
155 const struct lp_build_tgsi_action
* action
,
156 struct lp_build_tgsi_context
* bld_base
,
157 struct lp_build_emit_data
* emit_data
)
159 LLVMValueRef tmp0
, tmp1
;
160 tmp0
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MUL
,
161 emit_data
->args
[0] /* src0.x */,
162 emit_data
->args
[2] /* src1.x */);
163 tmp1
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MUL
,
164 emit_data
->args
[1] /* src0.y */,
165 emit_data
->args
[3] /* src1.y */);
166 emit_data
->output
[emit_data
->chan
] = lp_build_emit_llvm_binary(bld_base
,
167 TGSI_OPCODE_ADD
, tmp0
, tmp1
);
170 static struct lp_build_tgsi_action dp2_action
= {
171 dp2_fetch_args
, /* fetch_args */
175 /* TGSI_OPCODE_DP2A */
178 struct lp_build_tgsi_context
* bld_base
,
179 struct lp_build_emit_data
* emit_data
)
181 dp_fetch_args(bld_base
, emit_data
, 2);
182 emit_data
->args
[5] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
188 const struct lp_build_tgsi_action
* action
,
189 struct lp_build_tgsi_context
* bld_base
,
190 struct lp_build_emit_data
* emit_data
)
193 tmp
= lp_build_emit_llvm(bld_base
, TGSI_OPCODE_DP2
, emit_data
);
194 emit_data
->output
[emit_data
->chan
] = lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_ADD
,
195 emit_data
->args
[5], tmp
);
198 static struct lp_build_tgsi_action dp2a_action
= {
199 dp2a_fetch_args
, /* fetch_args */
203 /* TGSI_OPCODE_DP3 */
206 struct lp_build_tgsi_context
* bld_base
,
207 struct lp_build_emit_data
* emit_data
)
209 dp_fetch_args(bld_base
, emit_data
, 3);
214 const struct lp_build_tgsi_action
* action
,
215 struct lp_build_tgsi_context
* bld_base
,
216 struct lp_build_emit_data
* emit_data
)
218 LLVMValueRef tmp0
, tmp1
;
219 tmp0
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MUL
,
220 emit_data
->args
[0] /* src0.x */,
221 emit_data
->args
[3] /* src1.x */);
222 tmp1
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MUL
,
223 emit_data
->args
[1] /* src0.y */,
224 emit_data
->args
[4] /* src1.y */);
225 tmp0
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_ADD
, tmp1
, tmp0
);
226 tmp1
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MUL
,
227 emit_data
->args
[2] /* src0.z */,
228 emit_data
->args
[5] /* src1.z */);
229 emit_data
->output
[emit_data
->chan
] = lp_build_emit_llvm_binary(bld_base
,
230 TGSI_OPCODE_ADD
, tmp0
, tmp1
);
233 static struct lp_build_tgsi_action dp3_action
= {
234 dp3_fetch_args
, /* fetch_args */
238 /* TGSI_OPCODDE_DP4 */
242 struct lp_build_tgsi_context
* bld_base
,
243 struct lp_build_emit_data
* emit_data
)
245 dp_fetch_args(bld_base
, emit_data
, 4);
250 const struct lp_build_tgsi_action
* action
,
251 struct lp_build_tgsi_context
* bld_base
,
252 struct lp_build_emit_data
* emit_data
)
254 LLVMValueRef tmp0
, tmp1
;
255 tmp0
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MUL
,
256 emit_data
->args
[0] /* src0.x */,
257 emit_data
->args
[4] /* src1.x */);
258 tmp1
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MUL
,
259 emit_data
->args
[1] /* src0.y */,
260 emit_data
->args
[5] /* src1.y */);
261 tmp0
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_ADD
, tmp0
, tmp1
);
262 tmp1
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MUL
,
263 emit_data
->args
[2] /* src0.z */,
264 emit_data
->args
[6] /* src1.z */);
265 tmp0
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_ADD
, tmp0
, tmp1
);
266 tmp1
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MUL
,
267 emit_data
->args
[3] /* src0.w */,
268 emit_data
->args
[7] /* src1.w */);
269 emit_data
->output
[emit_data
->chan
] = lp_build_emit_llvm_binary(bld_base
,
270 TGSI_OPCODE_ADD
, tmp0
, tmp1
);
273 static struct lp_build_tgsi_action dp4_action
= {
274 dp4_fetch_args
, /* fetch_args */
278 /* TGSI_OPCODE_DPH */
281 struct lp_build_tgsi_context
* bld_base
,
282 struct lp_build_emit_data
* emit_data
)
284 dp_fetch_args(bld_base
, emit_data
, 4);
286 emit_data
->args
[3] = bld_base
->base
.one
;
289 const struct lp_build_tgsi_action dph_action
= {
290 dph_fetch_args
, /* fetch_args */
294 /* TGSI_OPCODE_DST */
297 struct lp_build_tgsi_context
* bld_base
,
298 struct lp_build_emit_data
* emit_data
)
301 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
304 emit_data
->args
[1] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
307 emit_data
->args
[2] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
310 emit_data
->args
[3] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
316 const struct lp_build_tgsi_action
* action
,
317 struct lp_build_tgsi_context
* bld_base
,
318 struct lp_build_emit_data
* emit_data
)
321 emit_data
->output
[TGSI_CHAN_X
] = bld_base
->base
.one
;
324 emit_data
->output
[TGSI_CHAN_Y
] = lp_build_emit_llvm_binary(bld_base
,
326 emit_data
->args
[0] /* src0.y */,
327 emit_data
->args
[2] /* src1.y */);
329 emit_data
->output
[TGSI_CHAN_Z
] = emit_data
->args
[1]; /* src0.z */
332 emit_data
->output
[TGSI_CHAN_W
] = emit_data
->args
[3]; /* src1.w */
335 static struct lp_build_tgsi_action dst_action
= {
336 dst_fetch_args
, /* fetch_args */
340 /* TGSI_OPCODE_END */
343 const struct lp_build_tgsi_action
* action
,
344 struct lp_build_tgsi_context
* bld_base
,
345 struct lp_build_emit_data
* emit_data
)
350 /* TGSI_OPCODE_EXP */
354 const struct lp_build_tgsi_action
* action
,
355 struct lp_build_tgsi_context
* bld_base
,
356 struct lp_build_emit_data
* emit_data
)
358 LLVMValueRef floor_x
;
360 /* floor( src0.x ) */
361 floor_x
= lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_FLR
,
364 /* 2 ^ floor( src0.x ) */
365 emit_data
->output
[TGSI_CHAN_X
] = lp_build_emit_llvm_unary(bld_base
,
366 TGSI_OPCODE_EX2
, floor_x
);
368 /* src0.x - floor( src0.x ) */
369 emit_data
->output
[TGSI_CHAN_Y
] = lp_build_emit_llvm_binary(bld_base
,
370 TGSI_OPCODE_SUB
, emit_data
->args
[0] /* src0.x */, floor_x
);
373 emit_data
->output
[TGSI_CHAN_Z
] = lp_build_emit_llvm_unary(bld_base
,
374 TGSI_OPCODE_EX2
, emit_data
->args
[0] /* src0.x */);
376 emit_data
->output
[TGSI_CHAN_W
] = bld_base
->base
.one
;
379 const struct lp_build_tgsi_action exp_action
= {
380 scalar_unary_fetch_args
, /* fetch_args */
384 /* TGSI_OPCODE_FRC */
388 const struct lp_build_tgsi_action
* action
,
389 struct lp_build_tgsi_context
* bld_base
,
390 struct lp_build_emit_data
* emit_data
)
393 tmp
= lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_FLR
,
395 emit_data
->output
[emit_data
->chan
] = lp_build_emit_llvm_binary(bld_base
,
396 TGSI_OPCODE_SUB
, emit_data
->args
[0], tmp
);
399 /* TGSI_OPCODE_KIL */
403 struct lp_build_tgsi_context
* bld_base
,
404 struct lp_build_emit_data
* emit_data
)
407 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
410 emit_data
->args
[1] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
413 emit_data
->args
[2] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
416 emit_data
->args
[3] = lp_build_emit_fetch(bld_base
, emit_data
->inst
,
418 emit_data
->arg_count
= 4;
419 emit_data
->dst_type
= LLVMVoidTypeInContext(bld_base
->base
.gallivm
->context
);
422 /* TGSI_OPCODE_KILP */
426 struct lp_build_tgsi_context
* bld_base
,
427 struct lp_build_emit_data
* emit_data
)
429 emit_data
->dst_type
= LLVMVoidTypeInContext(bld_base
->base
.gallivm
->context
);
432 /* TGSI_OPCODE_LIT */
436 struct lp_build_tgsi_context
* bld_base
,
437 struct lp_build_emit_data
* emit_data
)
440 emit_data
->args
[0] = lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_X
);
442 emit_data
->args
[1] = lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_Y
);
444 emit_data
->args
[2] = lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_W
);
445 emit_data
->arg_count
= 3;
450 const struct lp_build_tgsi_action
* action
,
451 struct lp_build_tgsi_context
* bld_base
,
452 struct lp_build_emit_data
* emit_data
)
454 LLVMValueRef tmp0
, tmp1
, tmp2
;
457 emit_data
->output
[TGSI_CHAN_X
] = bld_base
->base
.one
;
460 emit_data
->output
[TGSI_CHAN_Y
] = lp_build_emit_llvm_binary(bld_base
,
462 emit_data
->args
[0] /* src0.x */,
463 bld_base
->base
.zero
);
466 /* XMM[1] = SrcReg[0].yyyy */
467 tmp1
= emit_data
->args
[1];
468 /* XMM[1] = max(XMM[1], 0) */
469 tmp1
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MAX
,
470 tmp1
, bld_base
->base
.zero
);
471 /* XMM[2] = SrcReg[0].wwww */
472 tmp2
= emit_data
->args
[2];
473 tmp1
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_POW
,
475 tmp0
= emit_data
->args
[0];
476 emit_data
->output
[TGSI_CHAN_Z
] = lp_build_emit_llvm_ternary(bld_base
,
478 tmp0
, bld_base
->base
.zero
, tmp1
);
480 emit_data
->output
[TGSI_CHAN_W
] = bld_base
->base
.one
;
483 static struct lp_build_tgsi_action lit_action
= {
484 lit_fetch_args
, /* fetch_args */
488 /* TGSI_OPCODE_LOG */
492 const struct lp_build_tgsi_action
* action
,
493 struct lp_build_tgsi_context
* bld_base
,
494 struct lp_build_emit_data
* emit_data
)
497 LLVMValueRef abs_x
, log_abs_x
, flr_log_abs_x
, ex2_flr_log_abs_x
;
500 abs_x
= lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_ABS
,
501 emit_data
->args
[0] /* src0.x */);
503 /* log( abs( src0.x ) ) */
504 log_abs_x
= lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_LG2
,
507 /* floor( log( abs( src0.x ) ) ) */
508 flr_log_abs_x
= lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_FLR
,
511 emit_data
->output
[TGSI_CHAN_X
] = flr_log_abs_x
;
514 ex2_flr_log_abs_x
= lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_EX2
,
517 /* abs( src0.x ) / 2^( floor( lg2( abs( src0.x ) ) ) ) */
518 emit_data
->output
[TGSI_CHAN_Y
] = lp_build_emit_llvm_binary(bld_base
,
519 TGSI_OPCODE_DIV
, abs_x
, ex2_flr_log_abs_x
);
522 emit_data
->output
[TGSI_CHAN_Z
] = log_abs_x
;
525 emit_data
->output
[TGSI_CHAN_W
] = bld_base
->base
.one
;
528 static struct lp_build_tgsi_action log_action
= {
529 scalar_unary_fetch_args
, /* fetch_args */
533 /* TGSI_OPCODE_LRP */
537 const struct lp_build_tgsi_action
* action
,
538 struct lp_build_tgsi_context
* bld_base
,
539 struct lp_build_emit_data
* emit_data
)
542 tmp
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_SUB
,
545 emit_data
->output
[emit_data
->chan
] = lp_build_emit_llvm_ternary(bld_base
,
546 TGSI_OPCODE_MAD
, emit_data
->args
[0], tmp
, emit_data
->args
[2]);
549 /* TGSI_OPCODE_MAD */
553 const struct lp_build_tgsi_action
* action
,
554 struct lp_build_tgsi_context
* bld_base
,
555 struct lp_build_emit_data
* emit_data
)
558 tmp
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MUL
,
561 emit_data
->output
[emit_data
->chan
] = lp_build_emit_llvm_binary(bld_base
,
562 TGSI_OPCODE_ADD
, tmp
, emit_data
->args
[2]);
565 /* TGSI_OPCODE_MOV */
569 const struct lp_build_tgsi_action
* action
,
570 struct lp_build_tgsi_context
* bld_base
,
571 struct lp_build_emit_data
* emit_data
)
573 emit_data
->output
[emit_data
->chan
] = emit_data
->args
[0];
576 /* TGSI_OPCODE_MUL */
579 const struct lp_build_tgsi_action
* action
,
580 struct lp_build_tgsi_context
* bld_base
,
581 struct lp_build_emit_data
* emit_data
)
583 emit_data
->output
[emit_data
->chan
] = LLVMBuildFMul(
584 bld_base
->base
.gallivm
->builder
,
585 emit_data
->args
[0], emit_data
->args
[1], "");
588 /*.TGSI_OPCODE_DIV.*/
589 static void fdiv_emit(
590 const struct lp_build_tgsi_action
* action
,
591 struct lp_build_tgsi_context
* bld_base
,
592 struct lp_build_emit_data
* emit_data
)
594 emit_data
->output
[emit_data
->chan
] = LLVMBuildFDiv(
595 bld_base
->base
.gallivm
->builder
,
596 emit_data
->args
[0], emit_data
->args
[1], "");
599 /*.TGSI_OPCODE_RCP.*/
600 static void rcp_emit(
601 const struct lp_build_tgsi_action
* action
,
602 struct lp_build_tgsi_context
* bld_base
,
603 struct lp_build_emit_data
* emit_data
)
606 one
= lp_build_const_float(bld_base
->base
.gallivm
, 1.0f
);
607 emit_data
->output
[emit_data
->chan
] = lp_build_emit_llvm_binary(bld_base
,
608 TGSI_OPCODE_DIV
, one
, emit_data
->args
[0]);
611 /* TGSI_OPCODE_POW */
615 const struct lp_build_tgsi_action
* action
,
616 struct lp_build_tgsi_context
* bld_base
,
617 struct lp_build_emit_data
* emit_data
)
619 emit_data
->output
[emit_data
->chan
] = lp_build_pow(&bld_base
->base
,
620 emit_data
->args
[0], emit_data
->args
[1]);
623 static struct lp_build_tgsi_action pow_action
= {
624 scalar_binary_fetch_args
, /* fetch_args */
628 /* TGSI_OPCODE_RSQ */
632 const struct lp_build_tgsi_action
* action
,
633 struct lp_build_tgsi_context
* bld_base
,
634 struct lp_build_emit_data
* emit_data
)
636 emit_data
->args
[0] = lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_ABS
,
638 if (bld_base
->rsq_action
.emit
) {
639 bld_base
->rsq_action
.emit(&bld_base
->rsq_action
, bld_base
, emit_data
);
641 emit_data
->output
[emit_data
->chan
] = bld_base
->base
.undef
;
645 const struct lp_build_tgsi_action rsq_action
= {
646 scalar_unary_fetch_args
, /* fetch_args */
651 /* TGSI_OPCODE_SQRT */
655 const struct lp_build_tgsi_action
* action
,
656 struct lp_build_tgsi_context
* bld_base
,
657 struct lp_build_emit_data
* emit_data
)
659 if (bld_base
->sqrt_action
.emit
) {
660 bld_base
->sqrt_action
.emit(&bld_base
->sqrt_action
, bld_base
, emit_data
);
662 emit_data
->output
[emit_data
->chan
] = bld_base
->base
.undef
;
666 const struct lp_build_tgsi_action sqrt_action
= {
667 scalar_unary_fetch_args
, /* fetch_args */
671 /* TGSI_OPCODE_SCS */
674 const struct lp_build_tgsi_action
* action
,
675 struct lp_build_tgsi_context
* bld_base
,
676 struct lp_build_emit_data
* emit_data
)
679 emit_data
->output
[TGSI_CHAN_X
] = lp_build_emit_llvm_unary(bld_base
,
680 TGSI_OPCODE_COS
, emit_data
->args
[0]);
682 emit_data
->output
[TGSI_CHAN_Y
] = lp_build_emit_llvm_unary(bld_base
,
683 TGSI_OPCODE_SIN
, emit_data
->args
[0]);
685 emit_data
->output
[TGSI_CHAN_Z
] = bld_base
->base
.zero
;
688 emit_data
->output
[TGSI_CHAN_W
] = bld_base
->base
.one
;
691 const struct lp_build_tgsi_action scs_action
= {
692 scalar_unary_fetch_args
, /* fetch_args */
696 /* TGSI_OPCODE_SFL */
700 const struct lp_build_tgsi_action
* action
,
701 struct lp_build_tgsi_context
* bld_base
,
702 struct lp_build_emit_data
* emit_data
)
704 emit_data
->output
[emit_data
->chan
] = bld_base
->base
.zero
;
707 /* TGSI_OPCODE_STR */
711 const struct lp_build_tgsi_action
* action
,
712 struct lp_build_tgsi_context
* bld_base
,
713 struct lp_build_emit_data
* emit_data
)
715 emit_data
->output
[emit_data
->chan
] = bld_base
->base
.one
;
718 /* TGSI_OPCODE_SUB */
721 const struct lp_build_tgsi_action
* action
,
722 struct lp_build_tgsi_context
* bld_base
,
723 struct lp_build_emit_data
* emit_data
)
725 emit_data
->output
[emit_data
->chan
] = LLVMBuildFSub(
726 bld_base
->base
.gallivm
->builder
,
728 emit_data
->args
[1], "");
731 /* TGSI_OPCODE_U2F */
734 const struct lp_build_tgsi_action
* action
,
735 struct lp_build_tgsi_context
* bld_base
,
736 struct lp_build_emit_data
* emit_data
)
738 emit_data
->output
[emit_data
->chan
] = LLVMBuildUIToFP(bld_base
->base
.gallivm
->builder
,
740 bld_base
->base
.vec_type
, "");
745 const struct lp_build_tgsi_action
* action
,
746 struct lp_build_tgsi_context
* bld_base
,
747 struct lp_build_emit_data
* emit_data
)
750 tmp
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_UMUL
,
753 emit_data
->output
[emit_data
->chan
] = lp_build_emit_llvm_binary(bld_base
,
754 TGSI_OPCODE_UADD
, tmp
, emit_data
->args
[2]);
757 /* TGSI_OPCODE_UMUL */
760 const struct lp_build_tgsi_action
* action
,
761 struct lp_build_tgsi_context
* bld_base
,
762 struct lp_build_emit_data
* emit_data
)
764 emit_data
->output
[emit_data
->chan
] = lp_build_mul(&bld_base
->uint_bld
,
765 emit_data
->args
[0], emit_data
->args
[1]);
768 /* TGSI_OPCODE_MAX */
769 static void fmax_emit(
770 const struct lp_build_tgsi_action
* action
,
771 struct lp_build_tgsi_context
* bld_base
,
772 struct lp_build_emit_data
* emit_data
)
774 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
775 emit_data
->output
[emit_data
->chan
] = LLVMBuildSelect(builder
,
776 LLVMBuildFCmp(builder
, LLVMRealUGE
,
777 emit_data
->args
[0], emit_data
->args
[1], ""),
778 emit_data
->args
[0], emit_data
->args
[1], "");
781 /* TGSI_OPCODE_MIN */
782 static void fmin_emit(
783 const struct lp_build_tgsi_action
* action
,
784 struct lp_build_tgsi_context
* bld_base
,
785 struct lp_build_emit_data
* emit_data
)
787 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
788 emit_data
->output
[emit_data
->chan
] = LLVMBuildSelect(builder
,
789 LLVMBuildFCmp(builder
, LLVMRealUGE
,
790 emit_data
->args
[0], emit_data
->args
[1], ""),
791 emit_data
->args
[1], emit_data
->args
[0], "");
794 /* TGSI_OPCODE_XPD */
798 struct lp_build_tgsi_context
* bld_base
,
799 struct lp_build_emit_data
* emit_data
)
801 dp_fetch_args(bld_base
, emit_data
, 3);
809 struct lp_build_tgsi_context
* bld_base
,
815 LLVMValueRef tmp0
, tmp1
;
817 tmp0
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MUL
, a
, b
);
818 tmp1
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MUL
, c
, d
);
820 return lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_SUB
, tmp0
, tmp1
);
825 const struct lp_build_tgsi_action
* action
,
826 struct lp_build_tgsi_context
* bld_base
,
827 struct lp_build_emit_data
* emit_data
)
829 emit_data
->output
[TGSI_CHAN_X
] = xpd_helper(bld_base
,
830 emit_data
->args
[1] /* src0.y */, emit_data
->args
[5] /* src1.z */,
831 emit_data
->args
[4] /* src1.y */, emit_data
->args
[2] /* src0.z */);
833 emit_data
->output
[TGSI_CHAN_Y
] = xpd_helper(bld_base
,
834 emit_data
->args
[2] /* src0.z */, emit_data
->args
[3] /* src1.x */,
835 emit_data
->args
[5] /* src1.z */, emit_data
->args
[0] /* src0.x */);
837 emit_data
->output
[TGSI_CHAN_Z
] = xpd_helper(bld_base
,
838 emit_data
->args
[0] /* src0.x */, emit_data
->args
[4] /* src1.y */,
839 emit_data
->args
[3] /* src1.x */, emit_data
->args
[1] /* src0.y */);
841 emit_data
->output
[TGSI_CHAN_W
] = bld_base
->base
.one
;
844 const struct lp_build_tgsi_action xpd_action
= {
845 xpd_fetch_args
, /* fetch_args */
850 lp_set_default_actions(struct lp_build_tgsi_context
* bld_base
)
852 bld_base
->op_actions
[TGSI_OPCODE_DP2
] = dp2_action
;
853 bld_base
->op_actions
[TGSI_OPCODE_DP3
] = dp3_action
;
854 bld_base
->op_actions
[TGSI_OPCODE_DP4
] = dp4_action
;
855 bld_base
->op_actions
[TGSI_OPCODE_DP2A
] = dp2a_action
;
856 bld_base
->op_actions
[TGSI_OPCODE_DPH
] = dph_action
;
857 bld_base
->op_actions
[TGSI_OPCODE_DST
] = dst_action
;
858 bld_base
->op_actions
[TGSI_OPCODE_EXP
] = exp_action
;
859 bld_base
->op_actions
[TGSI_OPCODE_LIT
] = lit_action
;
860 bld_base
->op_actions
[TGSI_OPCODE_LOG
] = log_action
;
861 bld_base
->op_actions
[TGSI_OPCODE_RSQ
] = rsq_action
;
862 bld_base
->op_actions
[TGSI_OPCODE_SQRT
] = sqrt_action
;
863 bld_base
->op_actions
[TGSI_OPCODE_POW
] = pow_action
;
864 bld_base
->op_actions
[TGSI_OPCODE_SCS
] = scs_action
;
865 bld_base
->op_actions
[TGSI_OPCODE_XPD
] = xpd_action
;
867 bld_base
->op_actions
[TGSI_OPCODE_BREAKC
].fetch_args
= scalar_unary_fetch_args
;
868 bld_base
->op_actions
[TGSI_OPCODE_COS
].fetch_args
= scalar_unary_fetch_args
;
869 bld_base
->op_actions
[TGSI_OPCODE_EX2
].fetch_args
= scalar_unary_fetch_args
;
870 bld_base
->op_actions
[TGSI_OPCODE_IF
].fetch_args
= scalar_unary_fetch_args
;
871 bld_base
->op_actions
[TGSI_OPCODE_KIL
].fetch_args
= kil_fetch_args
;
872 bld_base
->op_actions
[TGSI_OPCODE_KILP
].fetch_args
= kilp_fetch_args
;
873 bld_base
->op_actions
[TGSI_OPCODE_RCP
].fetch_args
= scalar_unary_fetch_args
;
874 bld_base
->op_actions
[TGSI_OPCODE_SIN
].fetch_args
= scalar_unary_fetch_args
;
875 bld_base
->op_actions
[TGSI_OPCODE_LG2
].fetch_args
= scalar_unary_fetch_args
;
877 bld_base
->op_actions
[TGSI_OPCODE_ADD
].emit
= add_emit
;
878 bld_base
->op_actions
[TGSI_OPCODE_ARR
].emit
= arr_emit
;
879 bld_base
->op_actions
[TGSI_OPCODE_CLAMP
].emit
= clamp_emit
;
880 bld_base
->op_actions
[TGSI_OPCODE_END
].emit
= end_emit
;
881 bld_base
->op_actions
[TGSI_OPCODE_FRC
].emit
= frc_emit
;
882 bld_base
->op_actions
[TGSI_OPCODE_LRP
].emit
= lrp_emit
;
883 bld_base
->op_actions
[TGSI_OPCODE_MAD
].emit
= mad_emit
;
884 bld_base
->op_actions
[TGSI_OPCODE_MOV
].emit
= mov_emit
;
885 bld_base
->op_actions
[TGSI_OPCODE_MUL
].emit
= mul_emit
;
886 bld_base
->op_actions
[TGSI_OPCODE_DIV
].emit
= fdiv_emit
;
887 bld_base
->op_actions
[TGSI_OPCODE_RCP
].emit
= rcp_emit
;
888 bld_base
->op_actions
[TGSI_OPCODE_SFL
].emit
= sfl_emit
;
889 bld_base
->op_actions
[TGSI_OPCODE_STR
].emit
= str_emit
;
890 bld_base
->op_actions
[TGSI_OPCODE_SUB
].emit
= sub_emit
;
892 bld_base
->op_actions
[TGSI_OPCODE_UARL
].emit
= mov_emit
;
893 bld_base
->op_actions
[TGSI_OPCODE_U2F
].emit
= u2f_emit
;
894 bld_base
->op_actions
[TGSI_OPCODE_UMAD
].emit
= umad_emit
;
895 bld_base
->op_actions
[TGSI_OPCODE_UMUL
].emit
= umul_emit
;
897 bld_base
->op_actions
[TGSI_OPCODE_MAX
].emit
= fmax_emit
;
898 bld_base
->op_actions
[TGSI_OPCODE_MIN
].emit
= fmin_emit
;
901 /* CPU Only default actions */
903 /* These actions are CPU only, because they could potentially output SSE
907 /* TGSI_OPCODE_ABS (CPU Only)*/
911 const struct lp_build_tgsi_action
* action
,
912 struct lp_build_tgsi_context
* bld_base
,
913 struct lp_build_emit_data
* emit_data
)
915 emit_data
->output
[emit_data
->chan
] = lp_build_abs(&bld_base
->base
,
919 /* TGSI_OPCODE_ADD (CPU Only) */
922 const struct lp_build_tgsi_action
* action
,
923 struct lp_build_tgsi_context
* bld_base
,
924 struct lp_build_emit_data
* emit_data
)
926 emit_data
->output
[emit_data
->chan
] = lp_build_add(&bld_base
->base
,
927 emit_data
->args
[0], emit_data
->args
[1]);
930 /* TGSI_OPCODE_AND (CPU Only) */
933 const struct lp_build_tgsi_action
* action
,
934 struct lp_build_tgsi_context
* bld_base
,
935 struct lp_build_emit_data
* emit_data
)
937 emit_data
->output
[emit_data
->chan
] = lp_build_and(&bld_base
->uint_bld
,
938 emit_data
->args
[0], emit_data
->args
[1]);
941 /* TGSI_OPCODE_ARL (CPU Only) */
944 const struct lp_build_tgsi_action
* action
,
945 struct lp_build_tgsi_context
* bld_base
,
946 struct lp_build_emit_data
* emit_data
)
949 tmp
= lp_build_floor(&bld_base
->base
,
951 emit_data
->output
[emit_data
->chan
] = LLVMBuildFPToSI(bld_base
->base
.gallivm
->builder
, tmp
,
952 bld_base
->uint_bld
.vec_type
, "");
955 /* TGSI_OPCODE_ARR (CPU Only) */
958 const struct lp_build_tgsi_action
* action
,
959 struct lp_build_tgsi_context
* bld_base
,
960 struct lp_build_emit_data
* emit_data
)
962 emit_data
->output
[emit_data
->chan
] = lp_build_iround(&bld_base
->base
, emit_data
->args
[0]);
965 /* TGSI_OPCODE_CEIL (CPU Only) */
968 const struct lp_build_tgsi_action
* action
,
969 struct lp_build_tgsi_context
* bld_base
,
970 struct lp_build_emit_data
* emit_data
)
972 emit_data
->output
[emit_data
->chan
] = lp_build_ceil(&bld_base
->base
,
976 /* TGSI_OPCODE_CMP (CPU Only) */
979 const struct lp_build_tgsi_action
* action
,
980 struct lp_build_tgsi_context
* bld_base
,
981 struct lp_build_emit_data
* emit_data
)
983 LLVMValueRef cond
= lp_build_cmp(&bld_base
->base
, PIPE_FUNC_LESS
,
984 emit_data
->args
[0], bld_base
->base
.zero
);
985 emit_data
->output
[emit_data
->chan
] = lp_build_select(&bld_base
->base
,
986 cond
, emit_data
->args
[1], emit_data
->args
[2]);
989 /* TGSI_OPCODE_UCMP (CPU Only) */
992 const struct lp_build_tgsi_action
* action
,
993 struct lp_build_tgsi_context
* bld_base
,
994 struct lp_build_emit_data
* emit_data
)
996 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
997 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
998 LLVMValueRef unsigned_cond
=
999 LLVMBuildBitCast(builder
, emit_data
->args
[0], uint_bld
->vec_type
, "");
1000 LLVMValueRef cond
= lp_build_cmp(uint_bld
, PIPE_FUNC_NOTEQUAL
,
1003 emit_data
->output
[emit_data
->chan
] =
1004 lp_build_select(&bld_base
->base
,
1005 cond
, emit_data
->args
[1], emit_data
->args
[2]);
1009 /* TGSI_OPCODE_CND (CPU Only) */
1012 const struct lp_build_tgsi_action
* action
,
1013 struct lp_build_tgsi_context
* bld_base
,
1014 struct lp_build_emit_data
* emit_data
)
1016 LLVMValueRef half
, tmp
;
1017 half
= lp_build_const_vec(bld_base
->base
.gallivm
, bld_base
->base
.type
, 0.5);
1018 tmp
= lp_build_cmp(&bld_base
->base
, PIPE_FUNC_GREATER
,
1019 emit_data
->args
[2], half
);
1020 emit_data
->output
[emit_data
->chan
] = lp_build_select(&bld_base
->base
,
1023 emit_data
->args
[1]);
1026 /* TGSI_OPCODE_COS (CPU Only) */
1029 const struct lp_build_tgsi_action
* action
,
1030 struct lp_build_tgsi_context
* bld_base
,
1031 struct lp_build_emit_data
* emit_data
)
1033 emit_data
->output
[emit_data
->chan
] = lp_build_cos(&bld_base
->base
,
1034 emit_data
->args
[0]);
1037 /* TGSI_OPCODE_DIV (CPU Only) */
1040 const struct lp_build_tgsi_action
* action
,
1041 struct lp_build_tgsi_context
* bld_base
,
1042 struct lp_build_emit_data
* emit_data
)
1044 emit_data
->output
[emit_data
->chan
] = lp_build_div(&bld_base
->base
,
1045 emit_data
->args
[0], emit_data
->args
[1]);
1048 /* TGSI_OPCODE_EX2 (CPU Only) */
1051 const struct lp_build_tgsi_action
* action
,
1052 struct lp_build_tgsi_context
* bld_base
,
1053 struct lp_build_emit_data
* emit_data
)
1055 emit_data
->output
[emit_data
->chan
] = lp_build_exp2(&bld_base
->base
,
1056 emit_data
->args
[0]);
1059 /* TGSI_OPCODE_EXP (CPU Only) */
1062 const struct lp_build_tgsi_action
* action
,
1063 struct lp_build_tgsi_context
* bld_base
,
1064 struct lp_build_emit_data
* emit_data
)
1066 lp_build_exp2_approx(&bld_base
->base
, emit_data
->args
[0],
1067 &emit_data
->output
[TGSI_CHAN_X
],
1068 &emit_data
->output
[TGSI_CHAN_Y
],
1069 &emit_data
->output
[TGSI_CHAN_Z
]);
1070 emit_data
->output
[TGSI_CHAN_W
] = bld_base
->base
.one
;
1073 /* TGSI_OPCODE_F2I (CPU Only) */
1076 const struct lp_build_tgsi_action
* action
,
1077 struct lp_build_tgsi_context
* bld_base
,
1078 struct lp_build_emit_data
* emit_data
)
1080 emit_data
->output
[emit_data
->chan
] = lp_build_itrunc(&bld_base
->base
,
1081 emit_data
->args
[0]);
1084 /* TGSI_OPCODE_F2U (CPU Only) */
1087 const struct lp_build_tgsi_action
* action
,
1088 struct lp_build_tgsi_context
* bld_base
,
1089 struct lp_build_emit_data
* emit_data
)
1091 /* FIXME: implement and use lp_build_utrunc() */
1092 emit_data
->output
[emit_data
->chan
] = lp_build_itrunc(&bld_base
->base
,
1093 emit_data
->args
[0]);
1096 /* TGSI_OPCODE_FLR (CPU Only) */
1100 const struct lp_build_tgsi_action
* action
,
1101 struct lp_build_tgsi_context
* bld_base
,
1102 struct lp_build_emit_data
* emit_data
)
1104 emit_data
->output
[emit_data
->chan
] = lp_build_floor(&bld_base
->base
,
1105 emit_data
->args
[0]);
1108 /* TGSI_OPCODE_I2F (CPU Only) */
1111 const struct lp_build_tgsi_action
* action
,
1112 struct lp_build_tgsi_context
* bld_base
,
1113 struct lp_build_emit_data
* emit_data
)
1115 emit_data
->output
[emit_data
->chan
] = lp_build_int_to_float(&bld_base
->base
,
1116 emit_data
->args
[0]);
1119 /* TGSI_OPCODE_IABS (CPU Only) */
1122 const struct lp_build_tgsi_action
* action
,
1123 struct lp_build_tgsi_context
* bld_base
,
1124 struct lp_build_emit_data
* emit_data
)
1126 emit_data
->output
[emit_data
->chan
] = lp_build_abs(&bld_base
->int_bld
,
1127 emit_data
->args
[0]);
1130 /* TGSI_OPCODE_IDIV (CPU Only) */
1133 const struct lp_build_tgsi_action
* action
,
1134 struct lp_build_tgsi_context
* bld_base
,
1135 struct lp_build_emit_data
* emit_data
)
1137 emit_data
->output
[emit_data
->chan
] = lp_build_div(&bld_base
->int_bld
,
1138 emit_data
->args
[0], emit_data
->args
[1]);
1141 /* TGSI_OPCODE_INEG (CPU Only) */
1144 const struct lp_build_tgsi_action
* action
,
1145 struct lp_build_tgsi_context
* bld_base
,
1146 struct lp_build_emit_data
* emit_data
)
1148 emit_data
->output
[emit_data
->chan
] = lp_build_sub(&bld_base
->int_bld
,
1149 bld_base
->int_bld
.zero
,
1150 emit_data
->args
[0]);
1153 /* TGSI_OPCODE_ISET Helper (CPU Only) */
1156 const struct lp_build_tgsi_action
* action
,
1157 struct lp_build_tgsi_context
* bld_base
,
1158 struct lp_build_emit_data
* emit_data
,
1161 LLVMValueRef nz
= lp_build_const_vec(bld_base
->base
.gallivm
,
1162 bld_base
->int_bld
.type
, ~0U);
1163 LLVMValueRef cond
= lp_build_cmp(&bld_base
->int_bld
, pipe_func
,
1164 emit_data
->args
[0], emit_data
->args
[1]);
1165 emit_data
->output
[emit_data
->chan
] = lp_build_select(&bld_base
->int_bld
,
1168 bld_base
->int_bld
.zero
);
1171 /* TGSI_OPCODE_IMAX (CPU Only) */
1174 const struct lp_build_tgsi_action
* action
,
1175 struct lp_build_tgsi_context
* bld_base
,
1176 struct lp_build_emit_data
* emit_data
)
1178 emit_data
->output
[emit_data
->chan
] = lp_build_max(&bld_base
->int_bld
,
1179 emit_data
->args
[0], emit_data
->args
[1]);
1182 /* TGSI_OPCODE_IMIN (CPU Only) */
1185 const struct lp_build_tgsi_action
* action
,
1186 struct lp_build_tgsi_context
* bld_base
,
1187 struct lp_build_emit_data
* emit_data
)
1189 emit_data
->output
[emit_data
->chan
] = lp_build_min(&bld_base
->int_bld
,
1190 emit_data
->args
[0], emit_data
->args
[1]);
1193 /* TGSI_OPCODE_ISGE (CPU Only) */
1196 const struct lp_build_tgsi_action
* action
,
1197 struct lp_build_tgsi_context
* bld_base
,
1198 struct lp_build_emit_data
* emit_data
)
1200 iset_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_GEQUAL
);
1203 /* TGSI_OPCODE_ISHR (CPU Only) */
1206 const struct lp_build_tgsi_action
* action
,
1207 struct lp_build_tgsi_context
* bld_base
,
1208 struct lp_build_emit_data
* emit_data
)
1210 emit_data
->output
[emit_data
->chan
] = lp_build_shr(&bld_base
->int_bld
,
1211 emit_data
->args
[0], emit_data
->args
[1]);
1214 /* TGSI_OPCODE_ISLT (CPU Only) */
1217 const struct lp_build_tgsi_action
* action
,
1218 struct lp_build_tgsi_context
* bld_base
,
1219 struct lp_build_emit_data
* emit_data
)
1221 iset_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_LESS
);
1225 /* TGSI_OPCODE_ISSG (CPU Only) */
1228 const struct lp_build_tgsi_action
* action
,
1229 struct lp_build_tgsi_context
* bld_base
,
1230 struct lp_build_emit_data
* emit_data
)
1232 emit_data
->output
[emit_data
->chan
] = lp_build_sgn(&bld_base
->int_bld
,
1233 emit_data
->args
[0]);
1236 /* TGSI_OPCODE_LG2 (CPU Only) */
1239 const struct lp_build_tgsi_action
* action
,
1240 struct lp_build_tgsi_context
* bld_base
,
1241 struct lp_build_emit_data
* emit_data
)
1243 emit_data
->output
[emit_data
->chan
] = lp_build_log2(&bld_base
->base
,
1244 emit_data
->args
[0]);
1247 /* TGSI_OPCODE_LOG (CPU Only) */
1250 const struct lp_build_tgsi_action
* action
,
1251 struct lp_build_tgsi_context
* bld_base
,
1252 struct lp_build_emit_data
* emit_data
)
1254 LLVMValueRef p_floor_log2
;
1256 LLVMValueRef p_log2
;
1257 LLVMValueRef src0
= emit_data
->args
[0];
1259 lp_build_log2_approx(&bld_base
->base
, src0
,
1260 &p_exp
, &p_floor_log2
, &p_log2
);
1262 emit_data
->output
[TGSI_CHAN_X
] = p_floor_log2
;
1264 emit_data
->output
[TGSI_CHAN_Y
] = lp_build_emit_llvm_binary(bld_base
,
1267 emit_data
->output
[TGSI_CHAN_Z
] = p_log2
;
1269 emit_data
->output
[TGSI_CHAN_W
] = bld_base
->base
.one
;
1273 /* TGSI_OPCODE_MAX (CPU Only) */
1277 const struct lp_build_tgsi_action
* action
,
1278 struct lp_build_tgsi_context
* bld_base
,
1279 struct lp_build_emit_data
* emit_data
)
1281 emit_data
->output
[emit_data
->chan
] = lp_build_max(&bld_base
->base
,
1282 emit_data
->args
[0], emit_data
->args
[1]);
1285 /* TGSI_OPCODE_MIN (CPU Only) */
1288 const struct lp_build_tgsi_action
* action
,
1289 struct lp_build_tgsi_context
* bld_base
,
1290 struct lp_build_emit_data
* emit_data
)
1292 emit_data
->output
[emit_data
->chan
] = lp_build_min(&bld_base
->base
,
1293 emit_data
->args
[0], emit_data
->args
[1]);
1296 /* TGSI_OPCODE_MOD (CPU Only) */
1299 const struct lp_build_tgsi_action
* action
,
1300 struct lp_build_tgsi_context
* bld_base
,
1301 struct lp_build_emit_data
* emit_data
)
1303 emit_data
->output
[emit_data
->chan
] = lp_build_mod(&bld_base
->int_bld
,
1304 emit_data
->args
[0], emit_data
->args
[1]);
1307 /* TGSI_OPCODE_NOT */
1310 const struct lp_build_tgsi_action
* action
,
1311 struct lp_build_tgsi_context
* bld_base
,
1312 struct lp_build_emit_data
* emit_data
)
1314 emit_data
->output
[emit_data
->chan
] = lp_build_not(&bld_base
->base
,
1315 emit_data
->args
[0]);
1318 /* TGSI_OPCODE_OR (CPU Only) */
1321 const struct lp_build_tgsi_action
* action
,
1322 struct lp_build_tgsi_context
* bld_base
,
1323 struct lp_build_emit_data
* emit_data
)
1325 emit_data
->output
[emit_data
->chan
] = lp_build_or(&bld_base
->uint_bld
,
1326 emit_data
->args
[0], emit_data
->args
[1]);
1329 /* TGSI_OPCODE_POW (CPU Only) */
1332 const struct lp_build_tgsi_action
* action
,
1333 struct lp_build_tgsi_context
* bld_base
,
1334 struct lp_build_emit_data
* emit_data
)
1336 emit_data
->output
[emit_data
->chan
] = lp_build_pow(&bld_base
->base
,
1337 emit_data
->args
[0], emit_data
->args
[1]);
1341 /* TGSI_OPCODE_RCP (CPU Only) */
1345 const struct lp_build_tgsi_action
* action
,
1346 struct lp_build_tgsi_context
* bld_base
,
1347 struct lp_build_emit_data
* emit_data
)
1349 emit_data
->output
[emit_data
->chan
] = lp_build_rcp(&bld_base
->base
,
1350 emit_data
->args
[0]);
1353 /* Reciprical squareroot (CPU Only) */
1355 /* This is not the same as TGSI_OPCODE_RSQ, which requres the argument to be
1356 * greater than or equal to 0 */
1358 recip_sqrt_emit_cpu(
1359 const struct lp_build_tgsi_action
* action
,
1360 struct lp_build_tgsi_context
* bld_base
,
1361 struct lp_build_emit_data
* emit_data
)
1363 emit_data
->output
[emit_data
->chan
] = lp_build_rsqrt(&bld_base
->base
,
1364 emit_data
->args
[0]);
1369 const struct lp_build_tgsi_action
* action
,
1370 struct lp_build_tgsi_context
* bld_base
,
1371 struct lp_build_emit_data
* emit_data
)
1373 emit_data
->output
[emit_data
->chan
] = lp_build_sqrt(&bld_base
->base
,
1374 emit_data
->args
[0]);
1378 /* TGSI_OPCODE_ROUND (CPU Only) */
1381 const struct lp_build_tgsi_action
* action
,
1382 struct lp_build_tgsi_context
* bld_base
,
1383 struct lp_build_emit_data
* emit_data
)
1385 emit_data
->output
[emit_data
->chan
] = lp_build_round(&bld_base
->base
,
1386 emit_data
->args
[0]);
1389 /* TGSI_OPCODE_SET Helper (CPU Only) */
1393 const struct lp_build_tgsi_action
* action
,
1394 struct lp_build_tgsi_context
* bld_base
,
1395 struct lp_build_emit_data
* emit_data
,
1398 LLVMValueRef cond
= lp_build_cmp(&bld_base
->base
, pipe_func
,
1399 emit_data
->args
[0], emit_data
->args
[1]);
1400 emit_data
->output
[emit_data
->chan
] = lp_build_select(&bld_base
->base
,
1403 bld_base
->base
.zero
);
1406 /* TGSI_OPCODE_SEQ (CPU Only) */
1410 const struct lp_build_tgsi_action
* action
,
1411 struct lp_build_tgsi_context
* bld_base
,
1412 struct lp_build_emit_data
* emit_data
)
1414 set_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_EQUAL
);
1417 /* TGSI_OPCODE_SGE (CPU Only) */
1420 const struct lp_build_tgsi_action
* action
,
1421 struct lp_build_tgsi_context
* bld_base
,
1422 struct lp_build_emit_data
* emit_data
)
1424 set_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_GEQUAL
);
1427 /* TGSI_OPCODE_SGT (CPU Only)*/
1431 const struct lp_build_tgsi_action
* action
,
1432 struct lp_build_tgsi_context
* bld_base
,
1433 struct lp_build_emit_data
* emit_data
)
1435 set_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_GREATER
);
1438 /* TGSI_OPCODE_SHL (CPU Only) */
1441 const struct lp_build_tgsi_action
* action
,
1442 struct lp_build_tgsi_context
* bld_base
,
1443 struct lp_build_emit_data
* emit_data
)
1445 emit_data
->output
[emit_data
->chan
] = lp_build_shl(&bld_base
->uint_bld
,
1446 emit_data
->args
[0], emit_data
->args
[1]);
1449 /* TGSI_OPCODE_SIN (CPU Only) */
1452 const struct lp_build_tgsi_action
* action
,
1453 struct lp_build_tgsi_context
* bld_base
,
1454 struct lp_build_emit_data
* emit_data
)
1456 emit_data
->output
[emit_data
->chan
] = lp_build_sin(&bld_base
->base
,
1457 emit_data
->args
[0]);
1460 /* TGSI_OPCODE_SLE (CPU Only) */
1463 const struct lp_build_tgsi_action
* action
,
1464 struct lp_build_tgsi_context
* bld_base
,
1465 struct lp_build_emit_data
* emit_data
)
1467 set_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_LEQUAL
);
1470 /* TGSI_OPCODE_SLT (CPU Only) */
1473 const struct lp_build_tgsi_action
* action
,
1474 struct lp_build_tgsi_context
* bld_base
,
1475 struct lp_build_emit_data
* emit_data
)
1477 set_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_LESS
);
1480 /* TGSI_OPCODE_SNE (CPU Only) */
1484 const struct lp_build_tgsi_action
* action
,
1485 struct lp_build_tgsi_context
* bld_base
,
1486 struct lp_build_emit_data
* emit_data
)
1488 set_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_NOTEQUAL
);
1491 /* TGSI_OPCODE_SSG (CPU Only) */
1495 const struct lp_build_tgsi_action
* action
,
1496 struct lp_build_tgsi_context
* bld_base
,
1497 struct lp_build_emit_data
* emit_data
)
1499 emit_data
->output
[emit_data
->chan
] = lp_build_sgn(&bld_base
->base
,
1500 emit_data
->args
[0]);
1503 /* TGSI_OPCODE_SUB (CPU Only) */
1507 const struct lp_build_tgsi_action
* action
,
1508 struct lp_build_tgsi_context
* bld_base
,
1509 struct lp_build_emit_data
* emit_data
)
1511 emit_data
->output
[emit_data
->chan
] = lp_build_sub(&bld_base
->base
,
1513 emit_data
->args
[1]);
1516 /* TGSI_OPCODE_TRUNC (CPU Only) */
1520 const struct lp_build_tgsi_action
* action
,
1521 struct lp_build_tgsi_context
* bld_base
,
1522 struct lp_build_emit_data
* emit_data
)
1524 emit_data
->output
[emit_data
->chan
] = lp_build_trunc(&bld_base
->base
,
1525 emit_data
->args
[0]);
1528 /* TGSI_OPCODE_UADD (CPU Only) */
1531 const struct lp_build_tgsi_action
* action
,
1532 struct lp_build_tgsi_context
* bld_base
,
1533 struct lp_build_emit_data
* emit_data
)
1535 emit_data
->output
[emit_data
->chan
] = lp_build_add(&bld_base
->uint_bld
,
1536 emit_data
->args
[0], emit_data
->args
[1]);
1539 /* TGSI_OPCODE_UDIV (CPU Only) */
1542 const struct lp_build_tgsi_action
* action
,
1543 struct lp_build_tgsi_context
* bld_base
,
1544 struct lp_build_emit_data
* emit_data
)
1547 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1548 LLVMValueRef div_mask
= lp_build_cmp(&bld_base
->uint_bld
,
1549 PIPE_FUNC_EQUAL
, emit_data
->args
[1],
1550 bld_base
->uint_bld
.zero
);
1551 /* We want to make sure that we never divide/mod by zero to not
1552 * generate sigfpe. We don't want to crash just because the
1553 * shader is doing something weird. */
1554 LLVMValueRef divisor
= LLVMBuildOr(builder
,
1556 emit_data
->args
[1], "");
1557 LLVMValueRef result
= lp_build_div(&bld_base
->uint_bld
,
1558 emit_data
->args
[0], divisor
);
1559 /* udiv by zero is guaranteed to return 0xffffffff */
1560 emit_data
->output
[emit_data
->chan
] = LLVMBuildOr(builder
,
1565 /* TGSI_OPCODE_UMAX (CPU Only) */
1568 const struct lp_build_tgsi_action
* action
,
1569 struct lp_build_tgsi_context
* bld_base
,
1570 struct lp_build_emit_data
* emit_data
)
1572 emit_data
->output
[emit_data
->chan
] = lp_build_max(&bld_base
->uint_bld
,
1573 emit_data
->args
[0], emit_data
->args
[1]);
1576 /* TGSI_OPCODE_UMIN (CPU Only) */
1579 const struct lp_build_tgsi_action
* action
,
1580 struct lp_build_tgsi_context
* bld_base
,
1581 struct lp_build_emit_data
* emit_data
)
1583 emit_data
->output
[emit_data
->chan
] = lp_build_min(&bld_base
->uint_bld
,
1584 emit_data
->args
[0], emit_data
->args
[1]);
1587 /* TGSI_OPCODE_UMOD (CPU Only) */
1590 const struct lp_build_tgsi_action
* action
,
1591 struct lp_build_tgsi_context
* bld_base
,
1592 struct lp_build_emit_data
* emit_data
)
1594 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1595 LLVMValueRef div_mask
= lp_build_cmp(&bld_base
->uint_bld
,
1596 PIPE_FUNC_EQUAL
, emit_data
->args
[1],
1597 bld_base
->uint_bld
.zero
);
1598 /* We want to make sure that we never divide/mod by zero to not
1599 * generate sigfpe. We don't want to crash just because the
1600 * shader is doing something weird. */
1601 LLVMValueRef divisor
= LLVMBuildOr(builder
,
1603 emit_data
->args
[1], "");
1604 LLVMValueRef result
= lp_build_mod(&bld_base
->uint_bld
,
1605 emit_data
->args
[0], divisor
);
1606 /* umod by zero is guaranteed to return 0xffffffff */
1607 emit_data
->output
[emit_data
->chan
] = LLVMBuildOr(builder
,
1612 /* TGSI_OPCODE_USET Helper (CPU Only) */
1615 const struct lp_build_tgsi_action
* action
,
1616 struct lp_build_tgsi_context
* bld_base
,
1617 struct lp_build_emit_data
* emit_data
,
1620 LLVMValueRef nz
= lp_build_const_vec(bld_base
->base
.gallivm
,
1621 bld_base
->uint_bld
.type
, ~0U);
1622 LLVMValueRef cond
= lp_build_cmp(&bld_base
->uint_bld
, pipe_func
,
1623 emit_data
->args
[0], emit_data
->args
[1]);
1624 emit_data
->output
[emit_data
->chan
] = lp_build_select(&bld_base
->uint_bld
,
1627 bld_base
->uint_bld
.zero
);
1631 /* TGSI_OPCODE_USEQ (CPU Only) */
1634 const struct lp_build_tgsi_action
* action
,
1635 struct lp_build_tgsi_context
* bld_base
,
1636 struct lp_build_emit_data
* emit_data
)
1638 uset_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_EQUAL
);
1641 /* TGSI_OPCODE_ISGE (CPU Only) */
1644 const struct lp_build_tgsi_action
* action
,
1645 struct lp_build_tgsi_context
* bld_base
,
1646 struct lp_build_emit_data
* emit_data
)
1648 uset_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_GEQUAL
);
1651 /* TGSI_OPCODE_USHR (CPU Only) */
1654 const struct lp_build_tgsi_action
* action
,
1655 struct lp_build_tgsi_context
* bld_base
,
1656 struct lp_build_emit_data
* emit_data
)
1658 emit_data
->output
[emit_data
->chan
] = lp_build_shr(&bld_base
->uint_bld
,
1659 emit_data
->args
[0], emit_data
->args
[1]);
1662 /* TGSI_OPCODE_ISLT (CPU Only) */
1665 const struct lp_build_tgsi_action
* action
,
1666 struct lp_build_tgsi_context
* bld_base
,
1667 struct lp_build_emit_data
* emit_data
)
1669 uset_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_LESS
);
1672 /* TGSI_OPCODE_USNE (CPU Only) */
1676 const struct lp_build_tgsi_action
* action
,
1677 struct lp_build_tgsi_context
* bld_base
,
1678 struct lp_build_emit_data
* emit_data
)
1680 uset_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_NOTEQUAL
);
1683 /* TGSI_OPCODE_XOR */
1686 const struct lp_build_tgsi_action
* action
,
1687 struct lp_build_tgsi_context
* bld_base
,
1688 struct lp_build_emit_data
* emit_data
)
1690 emit_data
->output
[emit_data
->chan
] = lp_build_xor(&bld_base
->uint_bld
,
1692 emit_data
->args
[1]);
1696 lp_set_default_actions_cpu(
1697 struct lp_build_tgsi_context
* bld_base
)
1699 lp_set_default_actions(bld_base
);
1700 bld_base
->op_actions
[TGSI_OPCODE_ABS
].emit
= abs_emit_cpu
;
1701 bld_base
->op_actions
[TGSI_OPCODE_ADD
].emit
= add_emit_cpu
;
1702 bld_base
->op_actions
[TGSI_OPCODE_AND
].emit
= and_emit_cpu
;
1703 bld_base
->op_actions
[TGSI_OPCODE_ARL
].emit
= arl_emit_cpu
;
1704 bld_base
->op_actions
[TGSI_OPCODE_ARR
].emit
= arr_emit_cpu
;
1705 bld_base
->op_actions
[TGSI_OPCODE_CEIL
].emit
= ceil_emit_cpu
;
1706 bld_base
->op_actions
[TGSI_OPCODE_CND
].emit
= cnd_emit_cpu
;
1707 bld_base
->op_actions
[TGSI_OPCODE_COS
].emit
= cos_emit_cpu
;
1708 bld_base
->op_actions
[TGSI_OPCODE_CMP
].emit
= cmp_emit_cpu
;
1709 bld_base
->op_actions
[TGSI_OPCODE_DIV
].emit
= div_emit_cpu
;
1710 bld_base
->op_actions
[TGSI_OPCODE_EX2
].emit
= ex2_emit_cpu
;
1711 bld_base
->op_actions
[TGSI_OPCODE_EXP
].emit
= exp_emit_cpu
;
1712 bld_base
->op_actions
[TGSI_OPCODE_F2I
].emit
= f2i_emit_cpu
;
1713 bld_base
->op_actions
[TGSI_OPCODE_F2U
].emit
= f2u_emit_cpu
;
1714 bld_base
->op_actions
[TGSI_OPCODE_FLR
].emit
= flr_emit_cpu
;
1716 bld_base
->op_actions
[TGSI_OPCODE_I2F
].emit
= i2f_emit_cpu
;
1717 bld_base
->op_actions
[TGSI_OPCODE_IABS
].emit
= iabs_emit_cpu
;
1718 bld_base
->op_actions
[TGSI_OPCODE_IDIV
].emit
= idiv_emit_cpu
;
1719 bld_base
->op_actions
[TGSI_OPCODE_INEG
].emit
= ineg_emit_cpu
;
1720 bld_base
->op_actions
[TGSI_OPCODE_IMAX
].emit
= imax_emit_cpu
;
1721 bld_base
->op_actions
[TGSI_OPCODE_IMIN
].emit
= imin_emit_cpu
;
1722 bld_base
->op_actions
[TGSI_OPCODE_ISGE
].emit
= isge_emit_cpu
;
1723 bld_base
->op_actions
[TGSI_OPCODE_ISHR
].emit
= ishr_emit_cpu
;
1724 bld_base
->op_actions
[TGSI_OPCODE_ISLT
].emit
= islt_emit_cpu
;
1725 bld_base
->op_actions
[TGSI_OPCODE_ISSG
].emit
= issg_emit_cpu
;
1727 bld_base
->op_actions
[TGSI_OPCODE_LG2
].emit
= lg2_emit_cpu
;
1728 bld_base
->op_actions
[TGSI_OPCODE_LOG
].emit
= log_emit_cpu
;
1729 bld_base
->op_actions
[TGSI_OPCODE_MAX
].emit
= max_emit_cpu
;
1730 bld_base
->op_actions
[TGSI_OPCODE_MIN
].emit
= min_emit_cpu
;
1731 bld_base
->op_actions
[TGSI_OPCODE_MOD
].emit
= mod_emit_cpu
;
1732 bld_base
->op_actions
[TGSI_OPCODE_NOT
].emit
= not_emit_cpu
;
1733 bld_base
->op_actions
[TGSI_OPCODE_OR
].emit
= or_emit_cpu
;
1734 bld_base
->op_actions
[TGSI_OPCODE_POW
].emit
= pow_emit_cpu
;
1735 bld_base
->op_actions
[TGSI_OPCODE_RCP
].emit
= rcp_emit_cpu
;
1736 bld_base
->op_actions
[TGSI_OPCODE_ROUND
].emit
= round_emit_cpu
;
1737 bld_base
->op_actions
[TGSI_OPCODE_SEQ
].emit
= seq_emit_cpu
;
1738 bld_base
->op_actions
[TGSI_OPCODE_SGE
].emit
= sge_emit_cpu
;
1739 bld_base
->op_actions
[TGSI_OPCODE_SGT
].emit
= sgt_emit_cpu
;
1740 bld_base
->op_actions
[TGSI_OPCODE_SIN
].emit
= sin_emit_cpu
;
1741 bld_base
->op_actions
[TGSI_OPCODE_SHL
].emit
= shl_emit_cpu
;
1742 bld_base
->op_actions
[TGSI_OPCODE_SLE
].emit
= sle_emit_cpu
;
1743 bld_base
->op_actions
[TGSI_OPCODE_SLT
].emit
= slt_emit_cpu
;
1744 bld_base
->op_actions
[TGSI_OPCODE_SNE
].emit
= sne_emit_cpu
;
1745 bld_base
->op_actions
[TGSI_OPCODE_SSG
].emit
= ssg_emit_cpu
;
1746 bld_base
->op_actions
[TGSI_OPCODE_SUB
].emit
= sub_emit_cpu
;
1747 bld_base
->op_actions
[TGSI_OPCODE_TRUNC
].emit
= trunc_emit_cpu
;
1749 bld_base
->rsq_action
.emit
= recip_sqrt_emit_cpu
;
1750 bld_base
->sqrt_action
.emit
= sqrt_emit_cpu
;
1752 bld_base
->op_actions
[TGSI_OPCODE_UADD
].emit
= uadd_emit_cpu
;
1753 bld_base
->op_actions
[TGSI_OPCODE_UCMP
].emit
= ucmp_emit_cpu
;
1754 bld_base
->op_actions
[TGSI_OPCODE_UDIV
].emit
= udiv_emit_cpu
;
1755 bld_base
->op_actions
[TGSI_OPCODE_UMAX
].emit
= umax_emit_cpu
;
1756 bld_base
->op_actions
[TGSI_OPCODE_UMIN
].emit
= umin_emit_cpu
;
1757 bld_base
->op_actions
[TGSI_OPCODE_UMOD
].emit
= umod_emit_cpu
;
1758 bld_base
->op_actions
[TGSI_OPCODE_USEQ
].emit
= useq_emit_cpu
;
1759 bld_base
->op_actions
[TGSI_OPCODE_USGE
].emit
= usge_emit_cpu
;
1760 bld_base
->op_actions
[TGSI_OPCODE_USHR
].emit
= ushr_emit_cpu
;
1761 bld_base
->op_actions
[TGSI_OPCODE_USLT
].emit
= uslt_emit_cpu
;
1762 bld_base
->op_actions
[TGSI_OPCODE_USNE
].emit
= usne_emit_cpu
;
1764 bld_base
->op_actions
[TGSI_OPCODE_XOR
].emit
= xor_emit_cpu
;