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_KILL_IF */
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_KILL */
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 if (bld_base
->rsq_action
.emit
) {
637 bld_base
->rsq_action
.emit(&bld_base
->rsq_action
, bld_base
, emit_data
);
639 emit_data
->output
[emit_data
->chan
] = bld_base
->base
.undef
;
643 const struct lp_build_tgsi_action rsq_action
= {
644 scalar_unary_fetch_args
, /* fetch_args */
649 /* TGSI_OPCODE_SQRT */
653 const struct lp_build_tgsi_action
* action
,
654 struct lp_build_tgsi_context
* bld_base
,
655 struct lp_build_emit_data
* emit_data
)
657 if (bld_base
->sqrt_action
.emit
) {
658 bld_base
->sqrt_action
.emit(&bld_base
->sqrt_action
, bld_base
, emit_data
);
660 emit_data
->output
[emit_data
->chan
] = bld_base
->base
.undef
;
664 const struct lp_build_tgsi_action sqrt_action
= {
665 scalar_unary_fetch_args
, /* fetch_args */
669 /* TGSI_OPCODE_SCS */
672 const struct lp_build_tgsi_action
* action
,
673 struct lp_build_tgsi_context
* bld_base
,
674 struct lp_build_emit_data
* emit_data
)
677 emit_data
->output
[TGSI_CHAN_X
] = lp_build_emit_llvm_unary(bld_base
,
678 TGSI_OPCODE_COS
, emit_data
->args
[0]);
680 emit_data
->output
[TGSI_CHAN_Y
] = lp_build_emit_llvm_unary(bld_base
,
681 TGSI_OPCODE_SIN
, emit_data
->args
[0]);
683 emit_data
->output
[TGSI_CHAN_Z
] = bld_base
->base
.zero
;
686 emit_data
->output
[TGSI_CHAN_W
] = bld_base
->base
.one
;
689 const struct lp_build_tgsi_action scs_action
= {
690 scalar_unary_fetch_args
, /* fetch_args */
694 /* TGSI_OPCODE_SFL */
698 const struct lp_build_tgsi_action
* action
,
699 struct lp_build_tgsi_context
* bld_base
,
700 struct lp_build_emit_data
* emit_data
)
702 emit_data
->output
[emit_data
->chan
] = bld_base
->base
.zero
;
705 /* TGSI_OPCODE_STR */
709 const struct lp_build_tgsi_action
* action
,
710 struct lp_build_tgsi_context
* bld_base
,
711 struct lp_build_emit_data
* emit_data
)
713 emit_data
->output
[emit_data
->chan
] = bld_base
->base
.one
;
716 /* TGSI_OPCODE_SUB */
719 const struct lp_build_tgsi_action
* action
,
720 struct lp_build_tgsi_context
* bld_base
,
721 struct lp_build_emit_data
* emit_data
)
723 emit_data
->output
[emit_data
->chan
] = LLVMBuildFSub(
724 bld_base
->base
.gallivm
->builder
,
726 emit_data
->args
[1], "");
729 /* TGSI_OPCODE_U2F */
732 const struct lp_build_tgsi_action
* action
,
733 struct lp_build_tgsi_context
* bld_base
,
734 struct lp_build_emit_data
* emit_data
)
736 emit_data
->output
[emit_data
->chan
] = LLVMBuildUIToFP(bld_base
->base
.gallivm
->builder
,
738 bld_base
->base
.vec_type
, "");
743 const struct lp_build_tgsi_action
* action
,
744 struct lp_build_tgsi_context
* bld_base
,
745 struct lp_build_emit_data
* emit_data
)
748 tmp
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_UMUL
,
751 emit_data
->output
[emit_data
->chan
] = lp_build_emit_llvm_binary(bld_base
,
752 TGSI_OPCODE_UADD
, tmp
, emit_data
->args
[2]);
755 /* TGSI_OPCODE_UMUL */
758 const struct lp_build_tgsi_action
* action
,
759 struct lp_build_tgsi_context
* bld_base
,
760 struct lp_build_emit_data
* emit_data
)
762 emit_data
->output
[emit_data
->chan
] = lp_build_mul(&bld_base
->uint_bld
,
763 emit_data
->args
[0], emit_data
->args
[1]);
766 /* TGSI_OPCODE_MAX */
767 static void fmax_emit(
768 const struct lp_build_tgsi_action
* action
,
769 struct lp_build_tgsi_context
* bld_base
,
770 struct lp_build_emit_data
* emit_data
)
772 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
773 emit_data
->output
[emit_data
->chan
] = LLVMBuildSelect(builder
,
774 LLVMBuildFCmp(builder
, LLVMRealUGE
,
775 emit_data
->args
[0], emit_data
->args
[1], ""),
776 emit_data
->args
[0], emit_data
->args
[1], "");
779 /* TGSI_OPCODE_MIN */
780 static void fmin_emit(
781 const struct lp_build_tgsi_action
* action
,
782 struct lp_build_tgsi_context
* bld_base
,
783 struct lp_build_emit_data
* emit_data
)
785 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
786 emit_data
->output
[emit_data
->chan
] = LLVMBuildSelect(builder
,
787 LLVMBuildFCmp(builder
, LLVMRealUGE
,
788 emit_data
->args
[0], emit_data
->args
[1], ""),
789 emit_data
->args
[1], emit_data
->args
[0], "");
792 /* TGSI_OPCODE_XPD */
796 struct lp_build_tgsi_context
* bld_base
,
797 struct lp_build_emit_data
* emit_data
)
799 dp_fetch_args(bld_base
, emit_data
, 3);
807 struct lp_build_tgsi_context
* bld_base
,
813 LLVMValueRef tmp0
, tmp1
;
815 tmp0
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MUL
, a
, b
);
816 tmp1
= lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_MUL
, c
, d
);
818 return lp_build_emit_llvm_binary(bld_base
, TGSI_OPCODE_SUB
, tmp0
, tmp1
);
823 const struct lp_build_tgsi_action
* action
,
824 struct lp_build_tgsi_context
* bld_base
,
825 struct lp_build_emit_data
* emit_data
)
827 emit_data
->output
[TGSI_CHAN_X
] = xpd_helper(bld_base
,
828 emit_data
->args
[1] /* src0.y */, emit_data
->args
[5] /* src1.z */,
829 emit_data
->args
[4] /* src1.y */, emit_data
->args
[2] /* src0.z */);
831 emit_data
->output
[TGSI_CHAN_Y
] = xpd_helper(bld_base
,
832 emit_data
->args
[2] /* src0.z */, emit_data
->args
[3] /* src1.x */,
833 emit_data
->args
[5] /* src1.z */, emit_data
->args
[0] /* src0.x */);
835 emit_data
->output
[TGSI_CHAN_Z
] = xpd_helper(bld_base
,
836 emit_data
->args
[0] /* src0.x */, emit_data
->args
[4] /* src1.y */,
837 emit_data
->args
[3] /* src1.x */, emit_data
->args
[1] /* src0.y */);
839 emit_data
->output
[TGSI_CHAN_W
] = bld_base
->base
.one
;
842 const struct lp_build_tgsi_action xpd_action
= {
843 xpd_fetch_args
, /* fetch_args */
848 lp_set_default_actions(struct lp_build_tgsi_context
* bld_base
)
850 bld_base
->op_actions
[TGSI_OPCODE_DP2
] = dp2_action
;
851 bld_base
->op_actions
[TGSI_OPCODE_DP3
] = dp3_action
;
852 bld_base
->op_actions
[TGSI_OPCODE_DP4
] = dp4_action
;
853 bld_base
->op_actions
[TGSI_OPCODE_DP2A
] = dp2a_action
;
854 bld_base
->op_actions
[TGSI_OPCODE_DPH
] = dph_action
;
855 bld_base
->op_actions
[TGSI_OPCODE_DST
] = dst_action
;
856 bld_base
->op_actions
[TGSI_OPCODE_EXP
] = exp_action
;
857 bld_base
->op_actions
[TGSI_OPCODE_LIT
] = lit_action
;
858 bld_base
->op_actions
[TGSI_OPCODE_LOG
] = log_action
;
859 bld_base
->op_actions
[TGSI_OPCODE_RSQ
] = rsq_action
;
860 bld_base
->op_actions
[TGSI_OPCODE_SQRT
] = sqrt_action
;
861 bld_base
->op_actions
[TGSI_OPCODE_POW
] = pow_action
;
862 bld_base
->op_actions
[TGSI_OPCODE_SCS
] = scs_action
;
863 bld_base
->op_actions
[TGSI_OPCODE_XPD
] = xpd_action
;
865 bld_base
->op_actions
[TGSI_OPCODE_BREAKC
].fetch_args
= scalar_unary_fetch_args
;
866 bld_base
->op_actions
[TGSI_OPCODE_SWITCH
].fetch_args
= scalar_unary_fetch_args
;
867 bld_base
->op_actions
[TGSI_OPCODE_CASE
].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_UIF
].fetch_args
= scalar_unary_fetch_args
;
872 bld_base
->op_actions
[TGSI_OPCODE_KILL_IF
].fetch_args
= kil_fetch_args
;
873 bld_base
->op_actions
[TGSI_OPCODE_KILL
].fetch_args
= kilp_fetch_args
;
874 bld_base
->op_actions
[TGSI_OPCODE_RCP
].fetch_args
= scalar_unary_fetch_args
;
875 bld_base
->op_actions
[TGSI_OPCODE_SIN
].fetch_args
= scalar_unary_fetch_args
;
876 bld_base
->op_actions
[TGSI_OPCODE_LG2
].fetch_args
= scalar_unary_fetch_args
;
878 bld_base
->op_actions
[TGSI_OPCODE_ADD
].emit
= add_emit
;
879 bld_base
->op_actions
[TGSI_OPCODE_ARR
].emit
= arr_emit
;
880 bld_base
->op_actions
[TGSI_OPCODE_CLAMP
].emit
= clamp_emit
;
881 bld_base
->op_actions
[TGSI_OPCODE_END
].emit
= end_emit
;
882 bld_base
->op_actions
[TGSI_OPCODE_FRC
].emit
= frc_emit
;
883 bld_base
->op_actions
[TGSI_OPCODE_LRP
].emit
= lrp_emit
;
884 bld_base
->op_actions
[TGSI_OPCODE_MAD
].emit
= mad_emit
;
885 bld_base
->op_actions
[TGSI_OPCODE_MOV
].emit
= mov_emit
;
886 bld_base
->op_actions
[TGSI_OPCODE_MUL
].emit
= mul_emit
;
887 bld_base
->op_actions
[TGSI_OPCODE_DIV
].emit
= fdiv_emit
;
888 bld_base
->op_actions
[TGSI_OPCODE_RCP
].emit
= rcp_emit
;
889 bld_base
->op_actions
[TGSI_OPCODE_SFL
].emit
= sfl_emit
;
890 bld_base
->op_actions
[TGSI_OPCODE_STR
].emit
= str_emit
;
891 bld_base
->op_actions
[TGSI_OPCODE_SUB
].emit
= sub_emit
;
893 bld_base
->op_actions
[TGSI_OPCODE_UARL
].emit
= mov_emit
;
894 bld_base
->op_actions
[TGSI_OPCODE_U2F
].emit
= u2f_emit
;
895 bld_base
->op_actions
[TGSI_OPCODE_UMAD
].emit
= umad_emit
;
896 bld_base
->op_actions
[TGSI_OPCODE_UMUL
].emit
= umul_emit
;
898 bld_base
->op_actions
[TGSI_OPCODE_MAX
].emit
= fmax_emit
;
899 bld_base
->op_actions
[TGSI_OPCODE_MIN
].emit
= fmin_emit
;
902 /* CPU Only default actions */
904 /* These actions are CPU only, because they could potentially output SSE
908 /* TGSI_OPCODE_ABS (CPU Only)*/
912 const struct lp_build_tgsi_action
* action
,
913 struct lp_build_tgsi_context
* bld_base
,
914 struct lp_build_emit_data
* emit_data
)
916 emit_data
->output
[emit_data
->chan
] = lp_build_abs(&bld_base
->base
,
920 /* TGSI_OPCODE_ADD (CPU Only) */
923 const struct lp_build_tgsi_action
* action
,
924 struct lp_build_tgsi_context
* bld_base
,
925 struct lp_build_emit_data
* emit_data
)
927 emit_data
->output
[emit_data
->chan
] = lp_build_add(&bld_base
->base
,
928 emit_data
->args
[0], emit_data
->args
[1]);
931 /* TGSI_OPCODE_AND (CPU Only) */
934 const struct lp_build_tgsi_action
* action
,
935 struct lp_build_tgsi_context
* bld_base
,
936 struct lp_build_emit_data
* emit_data
)
938 emit_data
->output
[emit_data
->chan
] = lp_build_and(&bld_base
->uint_bld
,
939 emit_data
->args
[0], emit_data
->args
[1]);
942 /* TGSI_OPCODE_ARL (CPU Only) */
945 const struct lp_build_tgsi_action
* action
,
946 struct lp_build_tgsi_context
* bld_base
,
947 struct lp_build_emit_data
* emit_data
)
950 tmp
= lp_build_floor(&bld_base
->base
,
952 emit_data
->output
[emit_data
->chan
] = LLVMBuildFPToSI(bld_base
->base
.gallivm
->builder
, tmp
,
953 bld_base
->uint_bld
.vec_type
, "");
956 /* TGSI_OPCODE_ARR (CPU Only) */
959 const struct lp_build_tgsi_action
* action
,
960 struct lp_build_tgsi_context
* bld_base
,
961 struct lp_build_emit_data
* emit_data
)
963 emit_data
->output
[emit_data
->chan
] = lp_build_iround(&bld_base
->base
, emit_data
->args
[0]);
966 /* TGSI_OPCODE_CEIL (CPU Only) */
969 const struct lp_build_tgsi_action
* action
,
970 struct lp_build_tgsi_context
* bld_base
,
971 struct lp_build_emit_data
* emit_data
)
973 emit_data
->output
[emit_data
->chan
] = lp_build_ceil(&bld_base
->base
,
977 /* TGSI_OPCODE_CMP (CPU Only) */
980 const struct lp_build_tgsi_action
* action
,
981 struct lp_build_tgsi_context
* bld_base
,
982 struct lp_build_emit_data
* emit_data
)
984 LLVMValueRef cond
= lp_build_cmp(&bld_base
->base
, PIPE_FUNC_LESS
,
985 emit_data
->args
[0], bld_base
->base
.zero
);
986 emit_data
->output
[emit_data
->chan
] = lp_build_select(&bld_base
->base
,
987 cond
, emit_data
->args
[1], emit_data
->args
[2]);
990 /* TGSI_OPCODE_UCMP (CPU Only) */
993 const struct lp_build_tgsi_action
* action
,
994 struct lp_build_tgsi_context
* bld_base
,
995 struct lp_build_emit_data
* emit_data
)
997 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
998 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
999 LLVMValueRef unsigned_cond
=
1000 LLVMBuildBitCast(builder
, emit_data
->args
[0], uint_bld
->vec_type
, "");
1001 LLVMValueRef cond
= lp_build_cmp(uint_bld
, PIPE_FUNC_NOTEQUAL
,
1004 emit_data
->output
[emit_data
->chan
] =
1005 lp_build_select(&bld_base
->base
,
1006 cond
, emit_data
->args
[1], emit_data
->args
[2]);
1010 /* TGSI_OPCODE_CND (CPU Only) */
1013 const struct lp_build_tgsi_action
* action
,
1014 struct lp_build_tgsi_context
* bld_base
,
1015 struct lp_build_emit_data
* emit_data
)
1017 LLVMValueRef half
, tmp
;
1018 half
= lp_build_const_vec(bld_base
->base
.gallivm
, bld_base
->base
.type
, 0.5);
1019 tmp
= lp_build_cmp(&bld_base
->base
, PIPE_FUNC_GREATER
,
1020 emit_data
->args
[2], half
);
1021 emit_data
->output
[emit_data
->chan
] = lp_build_select(&bld_base
->base
,
1024 emit_data
->args
[1]);
1027 /* TGSI_OPCODE_COS (CPU Only) */
1030 const struct lp_build_tgsi_action
* action
,
1031 struct lp_build_tgsi_context
* bld_base
,
1032 struct lp_build_emit_data
* emit_data
)
1034 emit_data
->output
[emit_data
->chan
] = lp_build_cos(&bld_base
->base
,
1035 emit_data
->args
[0]);
1038 /* TGSI_OPCODE_DIV (CPU Only) */
1041 const struct lp_build_tgsi_action
* action
,
1042 struct lp_build_tgsi_context
* bld_base
,
1043 struct lp_build_emit_data
* emit_data
)
1045 emit_data
->output
[emit_data
->chan
] = lp_build_div(&bld_base
->base
,
1046 emit_data
->args
[0], emit_data
->args
[1]);
1049 /* TGSI_OPCODE_EX2 (CPU Only) */
1052 const struct lp_build_tgsi_action
* action
,
1053 struct lp_build_tgsi_context
* bld_base
,
1054 struct lp_build_emit_data
* emit_data
)
1056 emit_data
->output
[emit_data
->chan
] = lp_build_exp2(&bld_base
->base
,
1057 emit_data
->args
[0]);
1060 /* TGSI_OPCODE_EXP (CPU Only) */
1063 const struct lp_build_tgsi_action
* action
,
1064 struct lp_build_tgsi_context
* bld_base
,
1065 struct lp_build_emit_data
* emit_data
)
1067 lp_build_exp2_approx(&bld_base
->base
, emit_data
->args
[0],
1068 &emit_data
->output
[TGSI_CHAN_X
],
1069 &emit_data
->output
[TGSI_CHAN_Y
],
1070 &emit_data
->output
[TGSI_CHAN_Z
]);
1071 emit_data
->output
[TGSI_CHAN_W
] = bld_base
->base
.one
;
1074 /* TGSI_OPCODE_F2I (CPU Only) */
1077 const struct lp_build_tgsi_action
* action
,
1078 struct lp_build_tgsi_context
* bld_base
,
1079 struct lp_build_emit_data
* emit_data
)
1081 emit_data
->output
[emit_data
->chan
] = lp_build_itrunc(&bld_base
->base
,
1082 emit_data
->args
[0]);
1085 /* TGSI_OPCODE_F2U (CPU Only) */
1088 const struct lp_build_tgsi_action
* action
,
1089 struct lp_build_tgsi_context
* bld_base
,
1090 struct lp_build_emit_data
* emit_data
)
1092 /* FIXME: implement and use lp_build_utrunc() */
1093 emit_data
->output
[emit_data
->chan
] = lp_build_itrunc(&bld_base
->base
,
1094 emit_data
->args
[0]);
1097 /* TGSI_OPCODE_FSET Helper (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
,
1107 if (pipe_func
!= PIPE_FUNC_NOTEQUAL
) {
1108 cond
= lp_build_cmp_ordered(&bld_base
->base
, pipe_func
,
1109 emit_data
->args
[0], emit_data
->args
[1]);
1112 cond
= lp_build_cmp(&bld_base
->base
, pipe_func
,
1113 emit_data
->args
[0], emit_data
->args
[1]);
1116 emit_data
->output
[emit_data
->chan
] = cond
;
1120 /* TGSI_OPCODE_FSEQ (CPU Only) */
1123 const struct lp_build_tgsi_action
* action
,
1124 struct lp_build_tgsi_context
* bld_base
,
1125 struct lp_build_emit_data
* emit_data
)
1127 fset_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_EQUAL
);
1130 /* TGSI_OPCODE_ISGE (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 fset_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_GEQUAL
);
1140 /* TGSI_OPCODE_ISLT (CPU Only) */
1143 const struct lp_build_tgsi_action
* action
,
1144 struct lp_build_tgsi_context
* bld_base
,
1145 struct lp_build_emit_data
* emit_data
)
1147 fset_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_LESS
);
1150 /* TGSI_OPCODE_USNE (CPU Only) */
1154 const struct lp_build_tgsi_action
* action
,
1155 struct lp_build_tgsi_context
* bld_base
,
1156 struct lp_build_emit_data
* emit_data
)
1158 fset_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_NOTEQUAL
);
1161 /* TGSI_OPCODE_FLR (CPU Only) */
1165 const struct lp_build_tgsi_action
* action
,
1166 struct lp_build_tgsi_context
* bld_base
,
1167 struct lp_build_emit_data
* emit_data
)
1169 emit_data
->output
[emit_data
->chan
] = lp_build_floor(&bld_base
->base
,
1170 emit_data
->args
[0]);
1173 /* TGSI_OPCODE_I2F (CPU Only) */
1176 const struct lp_build_tgsi_action
* action
,
1177 struct lp_build_tgsi_context
* bld_base
,
1178 struct lp_build_emit_data
* emit_data
)
1180 emit_data
->output
[emit_data
->chan
] = lp_build_int_to_float(&bld_base
->base
,
1181 emit_data
->args
[0]);
1184 /* TGSI_OPCODE_IABS (CPU Only) */
1187 const struct lp_build_tgsi_action
* action
,
1188 struct lp_build_tgsi_context
* bld_base
,
1189 struct lp_build_emit_data
* emit_data
)
1191 emit_data
->output
[emit_data
->chan
] = lp_build_abs(&bld_base
->int_bld
,
1192 emit_data
->args
[0]);
1195 /* TGSI_OPCODE_IDIV (CPU Only) */
1198 const struct lp_build_tgsi_action
* action
,
1199 struct lp_build_tgsi_context
* bld_base
,
1200 struct lp_build_emit_data
* emit_data
)
1202 emit_data
->output
[emit_data
->chan
] = lp_build_div(&bld_base
->int_bld
,
1203 emit_data
->args
[0], emit_data
->args
[1]);
1206 /* TGSI_OPCODE_INEG (CPU Only) */
1209 const struct lp_build_tgsi_action
* action
,
1210 struct lp_build_tgsi_context
* bld_base
,
1211 struct lp_build_emit_data
* emit_data
)
1213 emit_data
->output
[emit_data
->chan
] = lp_build_sub(&bld_base
->int_bld
,
1214 bld_base
->int_bld
.zero
,
1215 emit_data
->args
[0]);
1218 /* TGSI_OPCODE_ISET Helper (CPU Only) */
1221 const struct lp_build_tgsi_action
* action
,
1222 struct lp_build_tgsi_context
* bld_base
,
1223 struct lp_build_emit_data
* emit_data
,
1226 LLVMValueRef cond
= lp_build_cmp(&bld_base
->int_bld
, pipe_func
,
1227 emit_data
->args
[0], emit_data
->args
[1]);
1228 emit_data
->output
[emit_data
->chan
] = cond
;
1231 /* TGSI_OPCODE_IMAX (CPU Only) */
1234 const struct lp_build_tgsi_action
* action
,
1235 struct lp_build_tgsi_context
* bld_base
,
1236 struct lp_build_emit_data
* emit_data
)
1238 emit_data
->output
[emit_data
->chan
] = lp_build_max(&bld_base
->int_bld
,
1239 emit_data
->args
[0], emit_data
->args
[1]);
1242 /* TGSI_OPCODE_IMIN (CPU Only) */
1245 const struct lp_build_tgsi_action
* action
,
1246 struct lp_build_tgsi_context
* bld_base
,
1247 struct lp_build_emit_data
* emit_data
)
1249 emit_data
->output
[emit_data
->chan
] = lp_build_min(&bld_base
->int_bld
,
1250 emit_data
->args
[0], emit_data
->args
[1]);
1253 /* TGSI_OPCODE_ISGE (CPU Only) */
1256 const struct lp_build_tgsi_action
* action
,
1257 struct lp_build_tgsi_context
* bld_base
,
1258 struct lp_build_emit_data
* emit_data
)
1260 iset_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_GEQUAL
);
1263 /* TGSI_OPCODE_ISHR (CPU Only) */
1266 const struct lp_build_tgsi_action
* action
,
1267 struct lp_build_tgsi_context
* bld_base
,
1268 struct lp_build_emit_data
* emit_data
)
1270 struct lp_build_context
*int_bld
= &bld_base
->int_bld
;
1271 LLVMValueRef mask
= lp_build_const_vec(int_bld
->gallivm
, int_bld
->type
,
1272 int_bld
->type
.width
- 1);
1273 LLVMValueRef masked_count
= lp_build_and(int_bld
, emit_data
->args
[1], mask
);
1274 emit_data
->output
[emit_data
->chan
] = lp_build_shr(int_bld
, emit_data
->args
[0],
1278 /* TGSI_OPCODE_ISLT (CPU Only) */
1281 const struct lp_build_tgsi_action
* action
,
1282 struct lp_build_tgsi_context
* bld_base
,
1283 struct lp_build_emit_data
* emit_data
)
1285 iset_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_LESS
);
1289 /* TGSI_OPCODE_ISSG (CPU Only) */
1292 const struct lp_build_tgsi_action
* action
,
1293 struct lp_build_tgsi_context
* bld_base
,
1294 struct lp_build_emit_data
* emit_data
)
1296 emit_data
->output
[emit_data
->chan
] = lp_build_sgn(&bld_base
->int_bld
,
1297 emit_data
->args
[0]);
1300 /* TGSI_OPCODE_LG2 (CPU Only) */
1303 const struct lp_build_tgsi_action
* action
,
1304 struct lp_build_tgsi_context
* bld_base
,
1305 struct lp_build_emit_data
* emit_data
)
1307 emit_data
->output
[emit_data
->chan
] = lp_build_log2_safe(&bld_base
->base
,
1308 emit_data
->args
[0]);
1311 /* TGSI_OPCODE_LOG (CPU Only) */
1314 const struct lp_build_tgsi_action
* action
,
1315 struct lp_build_tgsi_context
* bld_base
,
1316 struct lp_build_emit_data
* emit_data
)
1318 LLVMValueRef p_floor_log2
;
1320 LLVMValueRef p_log2
;
1321 LLVMValueRef src0
= emit_data
->args
[0];
1323 lp_build_log2_approx(&bld_base
->base
, src0
,
1324 &p_exp
, &p_floor_log2
, &p_log2
, FALSE
);
1326 emit_data
->output
[TGSI_CHAN_X
] = p_floor_log2
;
1328 emit_data
->output
[TGSI_CHAN_Y
] = lp_build_emit_llvm_binary(bld_base
,
1331 emit_data
->output
[TGSI_CHAN_Z
] = p_log2
;
1333 emit_data
->output
[TGSI_CHAN_W
] = bld_base
->base
.one
;
1337 /* TGSI_OPCODE_MAX (CPU Only) */
1341 const struct lp_build_tgsi_action
* action
,
1342 struct lp_build_tgsi_context
* bld_base
,
1343 struct lp_build_emit_data
* emit_data
)
1345 emit_data
->output
[emit_data
->chan
] =
1346 lp_build_max_ext(&bld_base
->base
,
1347 emit_data
->args
[0], emit_data
->args
[1],
1348 GALLIVM_NAN_RETURN_OTHER
);
1351 /* TGSI_OPCODE_MIN (CPU Only) */
1354 const struct lp_build_tgsi_action
* action
,
1355 struct lp_build_tgsi_context
* bld_base
,
1356 struct lp_build_emit_data
* emit_data
)
1358 emit_data
->output
[emit_data
->chan
] =
1359 lp_build_min_ext(&bld_base
->base
,
1360 emit_data
->args
[0], emit_data
->args
[1],
1361 GALLIVM_NAN_RETURN_OTHER
);
1364 /* TGSI_OPCODE_MOD (CPU Only) */
1367 const struct lp_build_tgsi_action
* action
,
1368 struct lp_build_tgsi_context
* bld_base
,
1369 struct lp_build_emit_data
* emit_data
)
1371 emit_data
->output
[emit_data
->chan
] = lp_build_mod(&bld_base
->int_bld
,
1372 emit_data
->args
[0], emit_data
->args
[1]);
1375 /* TGSI_OPCODE_NOT */
1378 const struct lp_build_tgsi_action
* action
,
1379 struct lp_build_tgsi_context
* bld_base
,
1380 struct lp_build_emit_data
* emit_data
)
1382 emit_data
->output
[emit_data
->chan
] = lp_build_not(&bld_base
->uint_bld
,
1383 emit_data
->args
[0]);
1386 /* TGSI_OPCODE_OR (CPU Only) */
1389 const struct lp_build_tgsi_action
* action
,
1390 struct lp_build_tgsi_context
* bld_base
,
1391 struct lp_build_emit_data
* emit_data
)
1393 emit_data
->output
[emit_data
->chan
] = lp_build_or(&bld_base
->uint_bld
,
1394 emit_data
->args
[0], emit_data
->args
[1]);
1397 /* TGSI_OPCODE_POW (CPU Only) */
1400 const struct lp_build_tgsi_action
* action
,
1401 struct lp_build_tgsi_context
* bld_base
,
1402 struct lp_build_emit_data
* emit_data
)
1404 emit_data
->output
[emit_data
->chan
] = lp_build_pow(&bld_base
->base
,
1405 emit_data
->args
[0], emit_data
->args
[1]);
1409 /* TGSI_OPCODE_RCP (CPU Only) */
1413 const struct lp_build_tgsi_action
* action
,
1414 struct lp_build_tgsi_context
* bld_base
,
1415 struct lp_build_emit_data
* emit_data
)
1417 emit_data
->output
[emit_data
->chan
] = lp_build_rcp(&bld_base
->base
,
1418 emit_data
->args
[0]);
1421 /* Reciprical squareroot (CPU Only) */
1423 recip_sqrt_emit_cpu(
1424 const struct lp_build_tgsi_action
* action
,
1425 struct lp_build_tgsi_context
* bld_base
,
1426 struct lp_build_emit_data
* emit_data
)
1428 emit_data
->output
[emit_data
->chan
] = lp_build_rsqrt(&bld_base
->base
,
1429 emit_data
->args
[0]);
1434 const struct lp_build_tgsi_action
* action
,
1435 struct lp_build_tgsi_context
* bld_base
,
1436 struct lp_build_emit_data
* emit_data
)
1438 emit_data
->output
[emit_data
->chan
] = lp_build_sqrt(&bld_base
->base
,
1439 emit_data
->args
[0]);
1443 /* TGSI_OPCODE_ROUND (CPU Only) */
1446 const struct lp_build_tgsi_action
* action
,
1447 struct lp_build_tgsi_context
* bld_base
,
1448 struct lp_build_emit_data
* emit_data
)
1450 emit_data
->output
[emit_data
->chan
] = lp_build_round(&bld_base
->base
,
1451 emit_data
->args
[0]);
1454 /* TGSI_OPCODE_SET Helper (CPU Only) */
1458 const struct lp_build_tgsi_action
* action
,
1459 struct lp_build_tgsi_context
* bld_base
,
1460 struct lp_build_emit_data
* emit_data
,
1465 if (pipe_func
!= PIPE_FUNC_NOTEQUAL
) {
1466 cond
= lp_build_cmp_ordered(&bld_base
->base
, pipe_func
,
1467 emit_data
->args
[0], emit_data
->args
[1]);
1470 cond
= lp_build_cmp(&bld_base
->base
, pipe_func
,
1471 emit_data
->args
[0], emit_data
->args
[1]);
1474 emit_data
->output
[emit_data
->chan
] = lp_build_select(&bld_base
->base
,
1477 bld_base
->base
.zero
);
1480 /* TGSI_OPCODE_SEQ (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_EQUAL
);
1491 /* TGSI_OPCODE_SGE (CPU Only) */
1494 const struct lp_build_tgsi_action
* action
,
1495 struct lp_build_tgsi_context
* bld_base
,
1496 struct lp_build_emit_data
* emit_data
)
1498 set_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_GEQUAL
);
1501 /* TGSI_OPCODE_SGT (CPU Only)*/
1505 const struct lp_build_tgsi_action
* action
,
1506 struct lp_build_tgsi_context
* bld_base
,
1507 struct lp_build_emit_data
* emit_data
)
1509 set_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_GREATER
);
1512 /* TGSI_OPCODE_SHL (CPU Only) */
1515 const struct lp_build_tgsi_action
* action
,
1516 struct lp_build_tgsi_context
* bld_base
,
1517 struct lp_build_emit_data
* emit_data
)
1519 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
1520 LLVMValueRef mask
= lp_build_const_vec(uint_bld
->gallivm
, uint_bld
->type
,
1521 uint_bld
->type
.width
- 1);
1522 LLVMValueRef masked_count
= lp_build_and(uint_bld
, emit_data
->args
[1], mask
);
1523 emit_data
->output
[emit_data
->chan
] = lp_build_shl(uint_bld
, emit_data
->args
[0],
1527 /* TGSI_OPCODE_SIN (CPU Only) */
1530 const struct lp_build_tgsi_action
* action
,
1531 struct lp_build_tgsi_context
* bld_base
,
1532 struct lp_build_emit_data
* emit_data
)
1534 emit_data
->output
[emit_data
->chan
] = lp_build_sin(&bld_base
->base
,
1535 emit_data
->args
[0]);
1538 /* TGSI_OPCODE_SLE (CPU Only) */
1541 const struct lp_build_tgsi_action
* action
,
1542 struct lp_build_tgsi_context
* bld_base
,
1543 struct lp_build_emit_data
* emit_data
)
1545 set_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_LEQUAL
);
1548 /* TGSI_OPCODE_SLT (CPU Only) */
1551 const struct lp_build_tgsi_action
* action
,
1552 struct lp_build_tgsi_context
* bld_base
,
1553 struct lp_build_emit_data
* emit_data
)
1555 set_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_LESS
);
1558 /* TGSI_OPCODE_SNE (CPU Only) */
1562 const struct lp_build_tgsi_action
* action
,
1563 struct lp_build_tgsi_context
* bld_base
,
1564 struct lp_build_emit_data
* emit_data
)
1566 set_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_NOTEQUAL
);
1569 /* TGSI_OPCODE_SSG (CPU Only) */
1573 const struct lp_build_tgsi_action
* action
,
1574 struct lp_build_tgsi_context
* bld_base
,
1575 struct lp_build_emit_data
* emit_data
)
1577 emit_data
->output
[emit_data
->chan
] = lp_build_sgn(&bld_base
->base
,
1578 emit_data
->args
[0]);
1581 /* TGSI_OPCODE_SUB (CPU Only) */
1585 const struct lp_build_tgsi_action
* action
,
1586 struct lp_build_tgsi_context
* bld_base
,
1587 struct lp_build_emit_data
* emit_data
)
1589 emit_data
->output
[emit_data
->chan
] = lp_build_sub(&bld_base
->base
,
1591 emit_data
->args
[1]);
1594 /* TGSI_OPCODE_TRUNC (CPU Only) */
1598 const struct lp_build_tgsi_action
* action
,
1599 struct lp_build_tgsi_context
* bld_base
,
1600 struct lp_build_emit_data
* emit_data
)
1602 emit_data
->output
[emit_data
->chan
] = lp_build_trunc(&bld_base
->base
,
1603 emit_data
->args
[0]);
1606 /* TGSI_OPCODE_UADD (CPU Only) */
1609 const struct lp_build_tgsi_action
* action
,
1610 struct lp_build_tgsi_context
* bld_base
,
1611 struct lp_build_emit_data
* emit_data
)
1613 emit_data
->output
[emit_data
->chan
] = lp_build_add(&bld_base
->uint_bld
,
1614 emit_data
->args
[0], emit_data
->args
[1]);
1617 /* TGSI_OPCODE_UDIV (CPU Only) */
1620 const struct lp_build_tgsi_action
* action
,
1621 struct lp_build_tgsi_context
* bld_base
,
1622 struct lp_build_emit_data
* emit_data
)
1625 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1626 LLVMValueRef div_mask
= lp_build_cmp(&bld_base
->uint_bld
,
1627 PIPE_FUNC_EQUAL
, emit_data
->args
[1],
1628 bld_base
->uint_bld
.zero
);
1629 /* We want to make sure that we never divide/mod by zero to not
1630 * generate sigfpe. We don't want to crash just because the
1631 * shader is doing something weird. */
1632 LLVMValueRef divisor
= LLVMBuildOr(builder
,
1634 emit_data
->args
[1], "");
1635 LLVMValueRef result
= lp_build_div(&bld_base
->uint_bld
,
1636 emit_data
->args
[0], divisor
);
1637 /* udiv by zero is guaranteed to return 0xffffffff */
1638 emit_data
->output
[emit_data
->chan
] = LLVMBuildOr(builder
,
1643 /* TGSI_OPCODE_UMAX (CPU Only) */
1646 const struct lp_build_tgsi_action
* action
,
1647 struct lp_build_tgsi_context
* bld_base
,
1648 struct lp_build_emit_data
* emit_data
)
1650 emit_data
->output
[emit_data
->chan
] = lp_build_max(&bld_base
->uint_bld
,
1651 emit_data
->args
[0], emit_data
->args
[1]);
1654 /* TGSI_OPCODE_UMIN (CPU Only) */
1657 const struct lp_build_tgsi_action
* action
,
1658 struct lp_build_tgsi_context
* bld_base
,
1659 struct lp_build_emit_data
* emit_data
)
1661 emit_data
->output
[emit_data
->chan
] = lp_build_min(&bld_base
->uint_bld
,
1662 emit_data
->args
[0], emit_data
->args
[1]);
1665 /* TGSI_OPCODE_UMOD (CPU Only) */
1668 const struct lp_build_tgsi_action
* action
,
1669 struct lp_build_tgsi_context
* bld_base
,
1670 struct lp_build_emit_data
* emit_data
)
1672 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1673 LLVMValueRef div_mask
= lp_build_cmp(&bld_base
->uint_bld
,
1674 PIPE_FUNC_EQUAL
, emit_data
->args
[1],
1675 bld_base
->uint_bld
.zero
);
1676 /* We want to make sure that we never divide/mod by zero to not
1677 * generate sigfpe. We don't want to crash just because the
1678 * shader is doing something weird. */
1679 LLVMValueRef divisor
= LLVMBuildOr(builder
,
1681 emit_data
->args
[1], "");
1682 LLVMValueRef result
= lp_build_mod(&bld_base
->uint_bld
,
1683 emit_data
->args
[0], divisor
);
1684 /* umod by zero is guaranteed to return 0xffffffff */
1685 emit_data
->output
[emit_data
->chan
] = LLVMBuildOr(builder
,
1690 /* TGSI_OPCODE_USET Helper (CPU Only) */
1693 const struct lp_build_tgsi_action
* action
,
1694 struct lp_build_tgsi_context
* bld_base
,
1695 struct lp_build_emit_data
* emit_data
,
1698 LLVMValueRef cond
= lp_build_cmp(&bld_base
->uint_bld
, pipe_func
,
1699 emit_data
->args
[0], emit_data
->args
[1]);
1700 emit_data
->output
[emit_data
->chan
] = cond
;
1704 /* TGSI_OPCODE_USEQ (CPU Only) */
1707 const struct lp_build_tgsi_action
* action
,
1708 struct lp_build_tgsi_context
* bld_base
,
1709 struct lp_build_emit_data
* emit_data
)
1711 uset_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_EQUAL
);
1714 /* TGSI_OPCODE_ISGE (CPU Only) */
1717 const struct lp_build_tgsi_action
* action
,
1718 struct lp_build_tgsi_context
* bld_base
,
1719 struct lp_build_emit_data
* emit_data
)
1721 uset_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_GEQUAL
);
1724 /* TGSI_OPCODE_USHR (CPU Only) */
1727 const struct lp_build_tgsi_action
* action
,
1728 struct lp_build_tgsi_context
* bld_base
,
1729 struct lp_build_emit_data
* emit_data
)
1731 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
1732 LLVMValueRef mask
= lp_build_const_vec(uint_bld
->gallivm
, uint_bld
->type
,
1733 uint_bld
->type
.width
- 1);
1734 LLVMValueRef masked_count
= lp_build_and(uint_bld
, emit_data
->args
[1], mask
);
1735 emit_data
->output
[emit_data
->chan
] = lp_build_shr(uint_bld
, emit_data
->args
[0],
1739 /* TGSI_OPCODE_ISLT (CPU Only) */
1742 const struct lp_build_tgsi_action
* action
,
1743 struct lp_build_tgsi_context
* bld_base
,
1744 struct lp_build_emit_data
* emit_data
)
1746 uset_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_LESS
);
1749 /* TGSI_OPCODE_USNE (CPU Only) */
1753 const struct lp_build_tgsi_action
* action
,
1754 struct lp_build_tgsi_context
* bld_base
,
1755 struct lp_build_emit_data
* emit_data
)
1757 uset_emit_cpu(action
, bld_base
, emit_data
, PIPE_FUNC_NOTEQUAL
);
1760 /* TGSI_OPCODE_XOR */
1763 const struct lp_build_tgsi_action
* action
,
1764 struct lp_build_tgsi_context
* bld_base
,
1765 struct lp_build_emit_data
* emit_data
)
1767 emit_data
->output
[emit_data
->chan
] = lp_build_xor(&bld_base
->uint_bld
,
1769 emit_data
->args
[1]);
1773 lp_set_default_actions_cpu(
1774 struct lp_build_tgsi_context
* bld_base
)
1776 lp_set_default_actions(bld_base
);
1777 bld_base
->op_actions
[TGSI_OPCODE_ABS
].emit
= abs_emit_cpu
;
1778 bld_base
->op_actions
[TGSI_OPCODE_ADD
].emit
= add_emit_cpu
;
1779 bld_base
->op_actions
[TGSI_OPCODE_AND
].emit
= and_emit_cpu
;
1780 bld_base
->op_actions
[TGSI_OPCODE_ARL
].emit
= arl_emit_cpu
;
1781 bld_base
->op_actions
[TGSI_OPCODE_ARR
].emit
= arr_emit_cpu
;
1782 bld_base
->op_actions
[TGSI_OPCODE_CEIL
].emit
= ceil_emit_cpu
;
1783 bld_base
->op_actions
[TGSI_OPCODE_CND
].emit
= cnd_emit_cpu
;
1784 bld_base
->op_actions
[TGSI_OPCODE_COS
].emit
= cos_emit_cpu
;
1785 bld_base
->op_actions
[TGSI_OPCODE_CMP
].emit
= cmp_emit_cpu
;
1786 bld_base
->op_actions
[TGSI_OPCODE_DIV
].emit
= div_emit_cpu
;
1787 bld_base
->op_actions
[TGSI_OPCODE_EX2
].emit
= ex2_emit_cpu
;
1788 bld_base
->op_actions
[TGSI_OPCODE_EXP
].emit
= exp_emit_cpu
;
1789 bld_base
->op_actions
[TGSI_OPCODE_F2I
].emit
= f2i_emit_cpu
;
1790 bld_base
->op_actions
[TGSI_OPCODE_F2U
].emit
= f2u_emit_cpu
;
1791 bld_base
->op_actions
[TGSI_OPCODE_FLR
].emit
= flr_emit_cpu
;
1792 bld_base
->op_actions
[TGSI_OPCODE_FSEQ
].emit
= fseq_emit_cpu
;
1793 bld_base
->op_actions
[TGSI_OPCODE_FSGE
].emit
= fsge_emit_cpu
;
1794 bld_base
->op_actions
[TGSI_OPCODE_FSLT
].emit
= fslt_emit_cpu
;
1795 bld_base
->op_actions
[TGSI_OPCODE_FSNE
].emit
= fsne_emit_cpu
;
1797 bld_base
->op_actions
[TGSI_OPCODE_I2F
].emit
= i2f_emit_cpu
;
1798 bld_base
->op_actions
[TGSI_OPCODE_IABS
].emit
= iabs_emit_cpu
;
1799 bld_base
->op_actions
[TGSI_OPCODE_IDIV
].emit
= idiv_emit_cpu
;
1800 bld_base
->op_actions
[TGSI_OPCODE_INEG
].emit
= ineg_emit_cpu
;
1801 bld_base
->op_actions
[TGSI_OPCODE_IMAX
].emit
= imax_emit_cpu
;
1802 bld_base
->op_actions
[TGSI_OPCODE_IMIN
].emit
= imin_emit_cpu
;
1803 bld_base
->op_actions
[TGSI_OPCODE_ISGE
].emit
= isge_emit_cpu
;
1804 bld_base
->op_actions
[TGSI_OPCODE_ISHR
].emit
= ishr_emit_cpu
;
1805 bld_base
->op_actions
[TGSI_OPCODE_ISLT
].emit
= islt_emit_cpu
;
1806 bld_base
->op_actions
[TGSI_OPCODE_ISSG
].emit
= issg_emit_cpu
;
1808 bld_base
->op_actions
[TGSI_OPCODE_LG2
].emit
= lg2_emit_cpu
;
1809 bld_base
->op_actions
[TGSI_OPCODE_LOG
].emit
= log_emit_cpu
;
1810 bld_base
->op_actions
[TGSI_OPCODE_MAX
].emit
= max_emit_cpu
;
1811 bld_base
->op_actions
[TGSI_OPCODE_MIN
].emit
= min_emit_cpu
;
1812 bld_base
->op_actions
[TGSI_OPCODE_MOD
].emit
= mod_emit_cpu
;
1813 bld_base
->op_actions
[TGSI_OPCODE_NOT
].emit
= not_emit_cpu
;
1814 bld_base
->op_actions
[TGSI_OPCODE_OR
].emit
= or_emit_cpu
;
1815 bld_base
->op_actions
[TGSI_OPCODE_POW
].emit
= pow_emit_cpu
;
1816 bld_base
->op_actions
[TGSI_OPCODE_RCP
].emit
= rcp_emit_cpu
;
1817 bld_base
->op_actions
[TGSI_OPCODE_ROUND
].emit
= round_emit_cpu
;
1818 bld_base
->op_actions
[TGSI_OPCODE_SEQ
].emit
= seq_emit_cpu
;
1819 bld_base
->op_actions
[TGSI_OPCODE_SGE
].emit
= sge_emit_cpu
;
1820 bld_base
->op_actions
[TGSI_OPCODE_SGT
].emit
= sgt_emit_cpu
;
1821 bld_base
->op_actions
[TGSI_OPCODE_SIN
].emit
= sin_emit_cpu
;
1822 bld_base
->op_actions
[TGSI_OPCODE_SHL
].emit
= shl_emit_cpu
;
1823 bld_base
->op_actions
[TGSI_OPCODE_SLE
].emit
= sle_emit_cpu
;
1824 bld_base
->op_actions
[TGSI_OPCODE_SLT
].emit
= slt_emit_cpu
;
1825 bld_base
->op_actions
[TGSI_OPCODE_SNE
].emit
= sne_emit_cpu
;
1826 bld_base
->op_actions
[TGSI_OPCODE_SSG
].emit
= ssg_emit_cpu
;
1827 bld_base
->op_actions
[TGSI_OPCODE_SUB
].emit
= sub_emit_cpu
;
1828 bld_base
->op_actions
[TGSI_OPCODE_TRUNC
].emit
= trunc_emit_cpu
;
1830 bld_base
->rsq_action
.emit
= recip_sqrt_emit_cpu
;
1831 bld_base
->sqrt_action
.emit
= sqrt_emit_cpu
;
1833 bld_base
->op_actions
[TGSI_OPCODE_UADD
].emit
= uadd_emit_cpu
;
1834 bld_base
->op_actions
[TGSI_OPCODE_UCMP
].emit
= ucmp_emit_cpu
;
1835 bld_base
->op_actions
[TGSI_OPCODE_UDIV
].emit
= udiv_emit_cpu
;
1836 bld_base
->op_actions
[TGSI_OPCODE_UMAX
].emit
= umax_emit_cpu
;
1837 bld_base
->op_actions
[TGSI_OPCODE_UMIN
].emit
= umin_emit_cpu
;
1838 bld_base
->op_actions
[TGSI_OPCODE_UMOD
].emit
= umod_emit_cpu
;
1839 bld_base
->op_actions
[TGSI_OPCODE_USEQ
].emit
= useq_emit_cpu
;
1840 bld_base
->op_actions
[TGSI_OPCODE_USGE
].emit
= usge_emit_cpu
;
1841 bld_base
->op_actions
[TGSI_OPCODE_USHR
].emit
= ushr_emit_cpu
;
1842 bld_base
->op_actions
[TGSI_OPCODE_USLT
].emit
= uslt_emit_cpu
;
1843 bld_base
->op_actions
[TGSI_OPCODE_USNE
].emit
= usne_emit_cpu
;
1845 bld_base
->op_actions
[TGSI_OPCODE_XOR
].emit
= xor_emit_cpu
;