1 /**************************************************************************
3 * Copyright 2010 VMware, Inc.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
30 * TGSI to LLVM IR translation -- AoS.
33 * - No control flow support: the existing control flow code should be factored
34 * out into from the SoA code into a common module and shared.
35 * - No derivatives. Derivate logic should be pluggable, just like the samplers.
37 * @author Jose Fonseca <jfonseca@vmware.com>
40 #include "pipe/p_config.h"
41 #include "pipe/p_shader_tokens.h"
42 #include "util/u_debug.h"
43 #include "util/u_math.h"
44 #include "util/u_memory.h"
45 #include "tgsi/tgsi_dump.h"
46 #include "tgsi/tgsi_info.h"
47 #include "tgsi/tgsi_parse.h"
48 #include "tgsi/tgsi_util.h"
49 #include "tgsi/tgsi_scan.h"
50 #include "lp_bld_type.h"
51 #include "lp_bld_const.h"
52 #include "lp_bld_arit.h"
53 #include "lp_bld_logic.h"
54 #include "lp_bld_swizzle.h"
55 #include "lp_bld_flow.h"
56 #include "lp_bld_quad.h"
57 #include "lp_bld_tgsi.h"
58 #include "lp_bld_debug.h"
59 #include "lp_bld_sample.h"
63 * Wrapper around lp_build_swizzle_aos which translates swizzles to another
67 swizzle_aos(struct lp_build_tgsi_context
*bld_base
,
74 unsigned char swizzles
[4];
75 struct lp_build_tgsi_aos_context
*bld
= lp_aos_context(bld_base
);
77 assert(swizzle_x
< 4);
78 assert(swizzle_y
< 4);
79 assert(swizzle_z
< 4);
80 assert(swizzle_w
< 4);
82 swizzles
[bld
->inv_swizzles
[0]] = bld
->swizzles
[swizzle_x
];
83 swizzles
[bld
->inv_swizzles
[1]] = bld
->swizzles
[swizzle_y
];
84 swizzles
[bld
->inv_swizzles
[2]] = bld
->swizzles
[swizzle_z
];
85 swizzles
[bld
->inv_swizzles
[3]] = bld
->swizzles
[swizzle_w
];
87 return lp_build_swizzle_aos(&bld
->bld_base
.base
, a
, swizzles
);
92 swizzle_scalar_aos(struct lp_build_tgsi_aos_context
*bld
,
96 chan
= bld
->swizzles
[chan
];
97 return lp_build_swizzle_scalar_aos(&bld
->bld_base
.base
, a
, chan
);
103 struct lp_build_tgsi_context
* bld_base
,
104 const struct tgsi_full_src_register
* reg
,
105 enum tgsi_opcode_type stype
,
108 struct lp_build_tgsi_aos_context
* bld
= lp_aos_context(bld_base
);
109 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
110 struct lp_type type
= bld_base
->base
.type
;
114 assert(!reg
->Register
.Indirect
);
117 * Get the constants components
120 res
= bld
->bld_base
.base
.undef
;
121 for (chan
= 0; chan
< 4; ++chan
) {
123 LLVMValueRef scalar_ptr
;
125 LLVMValueRef swizzle
;
127 index
= lp_build_const_int32(bld
->bld_base
.base
.gallivm
,
128 reg
->Register
.Index
* 4 + chan
);
130 scalar_ptr
= LLVMBuildGEP(builder
, bld
->consts_ptr
, &index
, 1, "");
132 scalar
= LLVMBuildLoad(builder
, scalar_ptr
, "");
134 lp_build_name(scalar
, "const[%u].%c", reg
->Register
.Index
, "xyzw"[chan
]);
137 * NOTE: constants array is always assumed to be RGBA
140 swizzle
= lp_build_const_int32(bld
->bld_base
.base
.gallivm
,
141 bld
->swizzles
[chan
]);
143 res
= LLVMBuildInsertElement(builder
, res
, scalar
, swizzle
, "");
147 * Broadcast the first quaternion to all others.
149 * XXX: could be factored into a reusable function.
152 if (type
.length
> 4) {
153 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
156 for (chan
= 0; chan
< 4; ++chan
) {
157 shuffles
[chan
] = lp_build_const_int32(bld
->bld_base
.base
.gallivm
, chan
);
160 for (i
= 4; i
< type
.length
; ++i
) {
161 shuffles
[i
] = shuffles
[i
% 4];
164 res
= LLVMBuildShuffleVector(builder
,
165 res
, bld
->bld_base
.base
.undef
,
166 LLVMConstVector(shuffles
, type
.length
),
173 emit_fetch_immediate(
174 struct lp_build_tgsi_context
* bld_base
,
175 const struct tgsi_full_src_register
* reg
,
176 enum tgsi_opcode_type stype
,
179 struct lp_build_tgsi_aos_context
* bld
= lp_aos_context(bld_base
);
180 LLVMValueRef res
= bld
->immediates
[reg
->Register
.Index
];
187 struct lp_build_tgsi_context
* bld_base
,
188 const struct tgsi_full_src_register
* reg
,
189 enum tgsi_opcode_type stype
,
192 struct lp_build_tgsi_aos_context
* bld
= lp_aos_context(bld_base
);
193 LLVMValueRef res
= bld
->inputs
[reg
->Register
.Index
];
194 assert(!reg
->Register
.Indirect
);
200 emit_fetch_temporary(
201 struct lp_build_tgsi_context
* bld_base
,
202 const struct tgsi_full_src_register
* reg
,
203 enum tgsi_opcode_type stype
,
206 struct lp_build_tgsi_aos_context
* bld
= lp_aos_context(bld_base
);
207 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
208 LLVMValueRef temp_ptr
= bld
->temps
[reg
->Register
.Index
];
209 LLVMValueRef res
= LLVMBuildLoad(builder
, temp_ptr
, "");
210 assert(!reg
->Register
.Indirect
);
212 return bld
->bld_base
.base
.undef
;
222 struct lp_build_tgsi_aos_context
*bld
,
223 const struct tgsi_full_instruction
*inst
,
227 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
228 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[index
];
229 LLVMValueRef mask
= NULL
;
236 switch (inst
->Instruction
.Saturate
) {
240 case TGSI_SAT_ZERO_ONE
:
241 value
= lp_build_max(&bld
->bld_base
.base
, value
, bld
->bld_base
.base
.zero
);
242 value
= lp_build_min(&bld
->bld_base
.base
, value
, bld
->bld_base
.base
.one
);
245 case TGSI_SAT_MINUS_PLUS_ONE
:
246 value
= lp_build_max(&bld
->bld_base
.base
, value
, lp_build_const_vec(bld
->bld_base
.base
.gallivm
, bld
->bld_base
.base
.type
, -1.0));
247 value
= lp_build_min(&bld
->bld_base
.base
, value
, bld
->bld_base
.base
.one
);
255 * Translate the register file
258 assert(!reg
->Register
.Indirect
);
260 switch (reg
->Register
.File
) {
261 case TGSI_FILE_OUTPUT
:
262 ptr
= bld
->outputs
[reg
->Register
.Index
];
265 case TGSI_FILE_TEMPORARY
:
266 ptr
= bld
->temps
[reg
->Register
.Index
];
269 case TGSI_FILE_ADDRESS
:
270 ptr
= bld
->addr
[reg
->Indirect
.Index
];
273 case TGSI_FILE_PREDICATE
:
274 ptr
= bld
->preds
[reg
->Register
.Index
];
288 if (inst
->Instruction
.Predicate
) {
291 assert(inst
->Predicate
.Index
< LP_MAX_TGSI_PREDS
);
293 pred
= LLVMBuildLoad(builder
,
294 bld
->preds
[inst
->Predicate
.Index
], "");
297 * Convert the value to an integer mask.
299 pred
= lp_build_compare(bld
->bld_base
.base
.gallivm
,
300 bld
->bld_base
.base
.type
,
303 bld
->bld_base
.base
.zero
);
305 if (inst
->Predicate
.Negate
) {
306 pred
= LLVMBuildNot(builder
, pred
, "");
309 pred
= bld
->bld_base
.emit_swizzle(&bld
->bld_base
, pred
,
310 inst
->Predicate
.SwizzleX
,
311 inst
->Predicate
.SwizzleY
,
312 inst
->Predicate
.SwizzleZ
,
313 inst
->Predicate
.SwizzleW
);
316 mask
= LLVMBuildAnd(builder
, mask
, pred
, "");
326 if (reg
->Register
.WriteMask
!= TGSI_WRITEMASK_XYZW
) {
327 LLVMValueRef writemask
;
329 writemask
= lp_build_const_mask_aos_swizzled(bld
->bld_base
.base
.gallivm
,
330 bld
->bld_base
.base
.type
,
331 reg
->Register
.WriteMask
,
335 mask
= LLVMBuildAnd(builder
, mask
, writemask
, "");
342 LLVMValueRef orig_value
;
344 orig_value
= LLVMBuildLoad(builder
, ptr
, "");
345 value
= lp_build_select(&bld
->bld_base
.base
,
346 mask
, value
, orig_value
);
349 LLVMBuildStore(builder
, value
, ptr
);
354 * High-level instruction translators.
358 emit_tex(struct lp_build_tgsi_aos_context
*bld
,
359 const struct tgsi_full_instruction
*inst
,
360 enum lp_build_tex_modifier modifier
)
365 struct lp_derivatives derivs
;
368 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
369 return bld
->bld_base
.base
.undef
;
372 target
= inst
->Texture
.Texture
;
374 coords
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0 , LP_CHAN_ALL
);
376 if (0 && modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
) {
377 lp_build_emit_fetch( &bld
->bld_base
, inst
, 1 , LP_CHAN_ALL
);
378 lp_build_emit_fetch( &bld
->bld_base
, inst
, 2 , LP_CHAN_ALL
);
379 unit
= inst
->Src
[3].Register
.Index
;
382 ddx
= lp_build_ddx( &bld
->bld_base
.base
, coords
);
383 ddy
= lp_build_ddy( &bld
->bld_base
.base
, coords
);
386 derivs
.ddx_ddy
[0] = bld
->bld_base
.base
.one
;
387 derivs
.ddx_ddy
[1] = bld
->bld_base
.base
.one
;
389 unit
= inst
->Src
[1].Register
.Index
;
392 return bld
->sampler
->emit_fetch_texel(bld
->sampler
,
401 lp_emit_declaration_aos(
402 struct lp_build_tgsi_aos_context
*bld
,
403 const struct tgsi_full_declaration
*decl
)
405 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
406 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->bld_base
.base
.gallivm
, bld
->bld_base
.base
.type
);
408 unsigned first
= decl
->Range
.First
;
409 unsigned last
= decl
->Range
.Last
;
412 for (idx
= first
; idx
<= last
; ++idx
) {
413 switch (decl
->Declaration
.File
) {
414 case TGSI_FILE_TEMPORARY
:
415 assert(idx
< LP_MAX_TGSI_TEMPS
);
416 if (bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
)) {
417 LLVMValueRef array_size
= lp_build_const_int32(gallivm
, last
+ 1);
418 bld
->temps_array
= lp_build_array_alloca(bld
->bld_base
.base
.gallivm
,
419 vec_type
, array_size
, "");
421 bld
->temps
[idx
] = lp_build_alloca(gallivm
, vec_type
, "");
425 case TGSI_FILE_OUTPUT
:
426 bld
->outputs
[idx
] = lp_build_alloca(gallivm
, vec_type
, "");
429 case TGSI_FILE_ADDRESS
:
430 assert(idx
< LP_MAX_TGSI_ADDRS
);
431 bld
->addr
[idx
] = lp_build_alloca(gallivm
, vec_type
, "");
434 case TGSI_FILE_PREDICATE
:
435 assert(idx
< LP_MAX_TGSI_PREDS
);
436 bld
->preds
[idx
] = lp_build_alloca(gallivm
, vec_type
, "");
440 /* don't need to declare other vars */
448 * Emit LLVM for one TGSI instruction.
449 * \param return TRUE for success, FALSE otherwise
452 lp_emit_instruction_aos(
453 struct lp_build_tgsi_aos_context
*bld
,
454 const struct tgsi_full_instruction
*inst
,
455 const struct tgsi_opcode_info
*info
,
458 LLVMValueRef src0
, src1
, src2
;
459 LLVMValueRef tmp0
, tmp1
;
460 LLVMValueRef dst0
= NULL
;
463 * Stores and write masks are handled in a general fashion after the long
464 * instruction opcode switch statement.
466 * Although not stricitly necessary, we avoid generating instructions for
467 * channels which won't be stored, in cases where's that easy. For some
468 * complex instructions, like texture sampling, it is more convenient to
469 * assume a full writemask and then let LLVM optimization passes eliminate
475 assert(info
->num_dst
<= 1);
477 dst0
= bld
->bld_base
.base
.undef
;
480 switch (inst
->Instruction
.Opcode
) {
481 case TGSI_OPCODE_ARL
:
482 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
483 dst0
= lp_build_floor(&bld
->bld_base
.base
, src0
);
486 case TGSI_OPCODE_MOV
:
487 dst0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
490 case TGSI_OPCODE_LIT
:
493 case TGSI_OPCODE_RCP
:
494 /* TGSI_OPCODE_RECIP */
495 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
496 dst0
= lp_build_rcp(&bld
->bld_base
.base
, src0
);
499 case TGSI_OPCODE_RSQ
:
500 /* TGSI_OPCODE_RECIPSQRT */
501 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
502 tmp0
= lp_build_emit_llvm_unary(&bld
->bld_base
, TGSI_OPCODE_ABS
, src0
);
503 dst0
= lp_build_rsqrt(&bld
->bld_base
.base
, tmp0
);
506 case TGSI_OPCODE_EXP
:
509 case TGSI_OPCODE_LOG
:
512 case TGSI_OPCODE_MUL
:
513 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
514 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
515 dst0
= lp_build_mul(&bld
->bld_base
.base
, src0
, src1
);
518 case TGSI_OPCODE_ADD
:
519 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
520 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
521 dst0
= lp_build_add(&bld
->bld_base
.base
, src0
, src1
);
524 case TGSI_OPCODE_DP3
:
525 /* TGSI_OPCODE_DOT3 */
528 case TGSI_OPCODE_DP4
:
529 /* TGSI_OPCODE_DOT4 */
532 case TGSI_OPCODE_DST
:
535 case TGSI_OPCODE_MIN
:
536 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
537 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
538 dst0
= lp_build_max(&bld
->bld_base
.base
, src0
, src1
);
541 case TGSI_OPCODE_MAX
:
542 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
543 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
544 dst0
= lp_build_max(&bld
->bld_base
.base
, src0
, src1
);
547 case TGSI_OPCODE_SLT
:
548 /* TGSI_OPCODE_SETLT */
549 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
550 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
551 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_LESS
, src0
, src1
);
552 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, bld
->bld_base
.base
.one
, bld
->bld_base
.base
.zero
);
555 case TGSI_OPCODE_SGE
:
556 /* TGSI_OPCODE_SETGE */
557 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
558 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
559 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_GEQUAL
, src0
, src1
);
560 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, bld
->bld_base
.base
.one
, bld
->bld_base
.base
.zero
);
563 case TGSI_OPCODE_MAD
:
564 /* TGSI_OPCODE_MADD */
565 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
566 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
567 src2
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 2, LP_CHAN_ALL
);
568 tmp0
= lp_build_mul(&bld
->bld_base
.base
, src0
, src1
);
569 dst0
= lp_build_add(&bld
->bld_base
.base
, tmp0
, src2
);
572 case TGSI_OPCODE_SUB
:
573 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
574 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
575 dst0
= lp_build_sub(&bld
->bld_base
.base
, src0
, src1
);
578 case TGSI_OPCODE_LRP
:
579 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
580 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
581 src2
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 2, LP_CHAN_ALL
);
582 tmp0
= lp_build_sub(&bld
->bld_base
.base
, src1
, src2
);
583 tmp0
= lp_build_mul(&bld
->bld_base
.base
, src0
, tmp0
);
584 dst0
= lp_build_add(&bld
->bld_base
.base
, tmp0
, src2
);
587 case TGSI_OPCODE_CND
:
588 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
589 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
590 src2
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 2, LP_CHAN_ALL
);
591 tmp1
= lp_build_const_vec(bld
->bld_base
.base
.gallivm
, bld
->bld_base
.base
.type
, 0.5);
592 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_GREATER
, src2
, tmp1
);
593 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, src0
, src1
);
596 case TGSI_OPCODE_DP2A
:
599 case TGSI_OPCODE_FRC
:
600 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
601 tmp0
= lp_build_floor(&bld
->bld_base
.base
, src0
);
602 dst0
= lp_build_sub(&bld
->bld_base
.base
, src0
, tmp0
);
605 case TGSI_OPCODE_CLAMP
:
606 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
607 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
608 src2
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 2, LP_CHAN_ALL
);
609 tmp0
= lp_build_max(&bld
->bld_base
.base
, src0
, src1
);
610 dst0
= lp_build_min(&bld
->bld_base
.base
, tmp0
, src2
);
613 case TGSI_OPCODE_FLR
:
614 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
615 dst0
= lp_build_floor(&bld
->bld_base
.base
, src0
);
618 case TGSI_OPCODE_ROUND
:
619 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
620 dst0
= lp_build_round(&bld
->bld_base
.base
, src0
);
623 case TGSI_OPCODE_EX2
:
624 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
625 tmp0
= lp_build_swizzle_scalar_aos(&bld
->bld_base
.base
, src0
, TGSI_SWIZZLE_X
);
626 dst0
= lp_build_exp2(&bld
->bld_base
.base
, tmp0
);
629 case TGSI_OPCODE_LG2
:
630 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
631 tmp0
= swizzle_scalar_aos(bld
, src0
, TGSI_SWIZZLE_X
);
632 dst0
= lp_build_log2(&bld
->bld_base
.base
, tmp0
);
635 case TGSI_OPCODE_POW
:
636 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
637 src0
= swizzle_scalar_aos(bld
, src0
, TGSI_SWIZZLE_X
);
638 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
639 src1
= swizzle_scalar_aos(bld
, src1
, TGSI_SWIZZLE_X
);
640 dst0
= lp_build_pow(&bld
->bld_base
.base
, src0
, src1
);
643 case TGSI_OPCODE_XPD
:
646 case TGSI_OPCODE_RCC
:
651 case TGSI_OPCODE_DPH
:
654 case TGSI_OPCODE_COS
:
655 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
656 tmp0
= swizzle_scalar_aos(bld
, src0
, TGSI_SWIZZLE_X
);
657 dst0
= lp_build_cos(&bld
->bld_base
.base
, tmp0
);
660 case TGSI_OPCODE_DDX
:
663 case TGSI_OPCODE_DDY
:
666 case TGSI_OPCODE_KILP
:
667 /* predicated kill */
670 case TGSI_OPCODE_KIL
:
671 /* conditional kill */
674 case TGSI_OPCODE_PK2H
:
678 case TGSI_OPCODE_PK2US
:
682 case TGSI_OPCODE_PK4B
:
686 case TGSI_OPCODE_PK4UB
:
689 case TGSI_OPCODE_RFL
:
692 case TGSI_OPCODE_SEQ
:
693 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
694 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
695 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_EQUAL
, src0
, src1
);
696 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, bld
->bld_base
.base
.one
, bld
->bld_base
.base
.zero
);
699 case TGSI_OPCODE_SFL
:
700 dst0
= bld
->bld_base
.base
.zero
;
703 case TGSI_OPCODE_SGT
:
704 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
705 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
706 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_GREATER
, src0
, src1
);
707 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, bld
->bld_base
.base
.one
, bld
->bld_base
.base
.zero
);
710 case TGSI_OPCODE_SIN
:
711 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
712 tmp0
= swizzle_scalar_aos(bld
, src0
, TGSI_SWIZZLE_X
);
713 dst0
= lp_build_sin(&bld
->bld_base
.base
, tmp0
);
716 case TGSI_OPCODE_SLE
:
717 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
718 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
719 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_LEQUAL
, src0
, src1
);
720 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, bld
->bld_base
.base
.one
, bld
->bld_base
.base
.zero
);
723 case TGSI_OPCODE_SNE
:
724 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
725 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
726 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_NOTEQUAL
, src0
, src1
);
727 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, bld
->bld_base
.base
.one
, bld
->bld_base
.base
.zero
);
730 case TGSI_OPCODE_STR
:
731 dst0
= bld
->bld_base
.base
.one
;
734 case TGSI_OPCODE_TEX
:
735 dst0
= emit_tex(bld
, inst
, LP_BLD_TEX_MODIFIER_NONE
);
738 case TGSI_OPCODE_TXD
:
739 dst0
= emit_tex(bld
, inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
);
742 case TGSI_OPCODE_UP2H
:
748 case TGSI_OPCODE_UP2US
:
754 case TGSI_OPCODE_UP4B
:
760 case TGSI_OPCODE_UP4UB
:
766 case TGSI_OPCODE_X2D
:
772 case TGSI_OPCODE_ARA
:
778 case TGSI_OPCODE_ARR
:
779 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
780 dst0
= lp_build_round(&bld
->bld_base
.base
, src0
);
783 case TGSI_OPCODE_BRA
:
789 case TGSI_OPCODE_CAL
:
792 case TGSI_OPCODE_RET
:
795 case TGSI_OPCODE_END
:
799 case TGSI_OPCODE_SSG
:
800 /* TGSI_OPCODE_SGN */
801 tmp0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
802 dst0
= lp_build_sgn(&bld
->bld_base
.base
, tmp0
);
805 case TGSI_OPCODE_CMP
:
806 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
807 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
808 src2
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 2, LP_CHAN_ALL
);
809 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_LESS
, src0
, bld
->bld_base
.base
.zero
);
810 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, src1
, src2
);
813 case TGSI_OPCODE_SCS
:
816 case TGSI_OPCODE_TXB
:
817 dst0
= emit_tex(bld
, inst
, LP_BLD_TEX_MODIFIER_LOD_BIAS
);
820 case TGSI_OPCODE_NRM
:
822 case TGSI_OPCODE_NRM4
:
825 case TGSI_OPCODE_DIV
:
831 case TGSI_OPCODE_DP2
:
834 case TGSI_OPCODE_TXL
:
835 dst0
= emit_tex(bld
, inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
);
838 case TGSI_OPCODE_TXP
:
839 dst0
= emit_tex(bld
, inst
, LP_BLD_TEX_MODIFIER_PROJECTED
);
842 case TGSI_OPCODE_BRK
:
848 case TGSI_OPCODE_BGNLOOP
:
851 case TGSI_OPCODE_BGNSUB
:
854 case TGSI_OPCODE_ELSE
:
857 case TGSI_OPCODE_ENDIF
:
860 case TGSI_OPCODE_ENDLOOP
:
863 case TGSI_OPCODE_ENDSUB
:
866 case TGSI_OPCODE_PUSHA
:
872 case TGSI_OPCODE_POPA
:
878 case TGSI_OPCODE_CEIL
:
879 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
880 dst0
= lp_build_ceil(&bld
->bld_base
.base
, src0
);
883 case TGSI_OPCODE_I2F
:
889 case TGSI_OPCODE_NOT
:
895 case TGSI_OPCODE_TRUNC
:
896 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
897 dst0
= lp_build_trunc(&bld
->bld_base
.base
, src0
);
900 case TGSI_OPCODE_SHL
:
906 case TGSI_OPCODE_ISHR
:
912 case TGSI_OPCODE_AND
:
924 case TGSI_OPCODE_MOD
:
930 case TGSI_OPCODE_XOR
:
936 case TGSI_OPCODE_SAD
:
942 case TGSI_OPCODE_TXF
:
948 case TGSI_OPCODE_TXQ
:
954 case TGSI_OPCODE_CONT
:
957 case TGSI_OPCODE_EMIT
:
961 case TGSI_OPCODE_ENDPRIM
:
965 case TGSI_OPCODE_NOP
:
973 lp_emit_store_aos(bld
, inst
, 0, dst0
);
981 lp_build_tgsi_aos(struct gallivm_state
*gallivm
,
982 const struct tgsi_token
*tokens
,
984 const unsigned char swizzles
[4],
985 LLVMValueRef consts_ptr
,
986 const LLVMValueRef
*inputs
,
987 LLVMValueRef
*outputs
,
988 struct lp_build_sampler_aos
*sampler
,
989 const struct tgsi_shader_info
*info
)
991 struct lp_build_tgsi_aos_context bld
;
992 struct tgsi_parse_context parse
;
993 uint num_immediates
= 0;
997 /* Setup build context */
998 memset(&bld
, 0, sizeof bld
);
999 lp_build_context_init(&bld
.bld_base
.base
, gallivm
, type
);
1000 lp_build_context_init(&bld
.bld_base
.uint_bld
, gallivm
, lp_uint_type(type
));
1001 lp_build_context_init(&bld
.bld_base
.int_bld
, gallivm
, lp_int_type(type
));
1002 lp_build_context_init(&bld
.int_bld
, gallivm
, lp_int_type(type
));
1004 for (chan
= 0; chan
< 4; ++chan
) {
1005 bld
.swizzles
[chan
] = swizzles
[chan
];
1006 bld
.inv_swizzles
[swizzles
[chan
]] = chan
;
1009 bld
.inputs
= inputs
;
1010 bld
.outputs
= outputs
;
1011 bld
.consts_ptr
= consts_ptr
;
1012 bld
.sampler
= sampler
;
1013 bld
.indirect_files
= info
->indirect_files
;
1014 bld
.bld_base
.emit_swizzle
= swizzle_aos
;
1015 bld
.bld_base
.info
= info
;
1017 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = emit_fetch_constant
;
1018 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_IMMEDIATE
] = emit_fetch_immediate
;
1019 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_INPUT
] = emit_fetch_input
;
1020 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_TEMPORARY
] = emit_fetch_temporary
;
1022 /* Set opcode actions */
1023 lp_set_default_actions_cpu(&bld
.bld_base
);
1025 if (!lp_bld_tgsi_list_init(&bld
.bld_base
)) {
1029 tgsi_parse_init(&parse
, tokens
);
1031 while (!tgsi_parse_end_of_tokens(&parse
)) {
1032 tgsi_parse_token(&parse
);
1034 switch(parse
.FullToken
.Token
.Type
) {
1035 case TGSI_TOKEN_TYPE_DECLARATION
:
1036 /* Inputs already interpolated */
1037 lp_emit_declaration_aos(&bld
, &parse
.FullToken
.FullDeclaration
);
1040 case TGSI_TOKEN_TYPE_INSTRUCTION
:
1041 /* save expanded instruction */
1042 lp_bld_tgsi_add_instruction(&bld
.bld_base
,
1043 &parse
.FullToken
.FullInstruction
);
1046 case TGSI_TOKEN_TYPE_IMMEDIATE
:
1047 /* simply copy the immediate values into the next immediates[] slot */
1049 const uint size
= parse
.FullToken
.FullImmediate
.Immediate
.NrTokens
- 1;
1052 assert(num_immediates
< LP_MAX_TGSI_IMMEDIATES
);
1053 for (chan
= 0; chan
< 4; ++chan
) {
1056 for (chan
= 0; chan
< size
; ++chan
) {
1057 unsigned swizzle
= bld
.swizzles
[chan
];
1058 imm
[swizzle
] = parse
.FullToken
.FullImmediate
.u
[chan
].Float
;
1060 bld
.immediates
[num_immediates
] =
1061 lp_build_const_aos(gallivm
, type
,
1062 imm
[0], imm
[1], imm
[2], imm
[3],
1068 case TGSI_TOKEN_TYPE_PROPERTY
:
1077 struct tgsi_full_instruction
*instr
= bld
.bld_base
.instructions
+ pc
;
1078 const struct tgsi_opcode_info
*opcode_info
=
1079 tgsi_get_opcode_info(instr
->Instruction
.Opcode
);
1080 if (!lp_emit_instruction_aos(&bld
, instr
, opcode_info
, &pc
))
1081 _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
1082 opcode_info
->mnemonic
);
1086 LLVMBasicBlockRef block
= LLVMGetInsertBlock(gallivm
->builder
);
1087 LLVMValueRef function
= LLVMGetBasicBlockParent(block
);
1088 debug_printf("11111111111111111111111111111 \n");
1089 tgsi_dump(tokens
, 0);
1090 lp_debug_dump_value(function
);
1091 debug_printf("2222222222222222222222222222 \n");
1093 tgsi_parse_free(&parse
);
1094 FREE(bld
.bld_base
.instructions
);
1097 LLVMModuleRef module
= LLVMGetGlobalParent(
1098 LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm
->builder
)));
1099 LLVMDumpModule(module
);