1 /**************************************************************************
3 * Copyright 2009-2010 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 * Depth/stencil testing to LLVM IR translation.
32 * To be done accurately/efficiently the depth/stencil test must be done with
33 * the same type/format of the depth/stencil buffer, which implies massaging
34 * the incoming depths to fit into place. Using a more straightforward
35 * type/format for depth/stencil values internally and only convert when
36 * flushing would avoid this, but it would most likely result in depth fighting
39 * Since we're using linear layout for everything, but we need to deal with
40 * 2x2 quads, we need to load/store multiple values and swizzle them into
41 * place (we could avoid this by doing depth/stencil testing in linear format,
42 * which would be easy for late depth/stencil test as we could do that after
43 * the fragment shader loop just as we do for color buffers, but more tricky
44 * for early depth test as we'd need both masks and interpolated depth in
48 * @author Jose Fonseca <jfonseca@vmware.com>
49 * @author Brian Paul <jfonseca@vmware.com>
52 #include "pipe/p_state.h"
53 #include "util/u_format.h"
54 #include "util/u_cpu_detect.h"
56 #include "gallivm/lp_bld_type.h"
57 #include "gallivm/lp_bld_arit.h"
58 #include "gallivm/lp_bld_bitarit.h"
59 #include "gallivm/lp_bld_const.h"
60 #include "gallivm/lp_bld_conv.h"
61 #include "gallivm/lp_bld_logic.h"
62 #include "gallivm/lp_bld_flow.h"
63 #include "gallivm/lp_bld_intr.h"
64 #include "gallivm/lp_bld_debug.h"
65 #include "gallivm/lp_bld_swizzle.h"
66 #include "gallivm/lp_bld_pack.h"
68 #include "lp_bld_depth.h"
71 /** Used to select fields from pipe_stencil_state */
81 * Do the stencil test comparison (compare FB stencil values against ref value).
82 * This will be used twice when generating two-sided stencil code.
83 * \param stencil the front/back stencil state
84 * \param stencilRef the stencil reference value, replicated as a vector
85 * \param stencilVals vector of stencil values from framebuffer
86 * \return vector mask of pass/fail values (~0 or 0)
89 lp_build_stencil_test_single(struct lp_build_context
*bld
,
90 const struct pipe_stencil_state
*stencil
,
91 LLVMValueRef stencilRef
,
92 LLVMValueRef stencilVals
)
94 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
95 const unsigned stencilMax
= 255; /* XXX fix */
96 struct lp_type type
= bld
->type
;
100 * SSE2 has intrinsics for signed comparisons, but not unsigned ones. Values
101 * are between 0..255 so ensure we generate the fastest comparisons for
104 if (type
.width
<= 8) {
110 assert(stencil
->enabled
);
112 if (stencil
->valuemask
!= stencilMax
) {
113 /* compute stencilRef = stencilRef & valuemask */
114 LLVMValueRef valuemask
= lp_build_const_int_vec(bld
->gallivm
, type
, stencil
->valuemask
);
115 stencilRef
= LLVMBuildAnd(builder
, stencilRef
, valuemask
, "");
116 /* compute stencilVals = stencilVals & valuemask */
117 stencilVals
= LLVMBuildAnd(builder
, stencilVals
, valuemask
, "");
120 res
= lp_build_cmp(bld
, stencil
->func
, stencilRef
, stencilVals
);
127 * Do the one or two-sided stencil test comparison.
128 * \sa lp_build_stencil_test_single
129 * \param front_facing an integer vector mask, indicating front (~0) or back
130 * (0) facing polygon. If NULL, assume front-facing.
133 lp_build_stencil_test(struct lp_build_context
*bld
,
134 const struct pipe_stencil_state stencil
[2],
135 LLVMValueRef stencilRefs
[2],
136 LLVMValueRef stencilVals
,
137 LLVMValueRef front_facing
)
141 assert(stencil
[0].enabled
);
143 /* do front face test */
144 res
= lp_build_stencil_test_single(bld
, &stencil
[0],
145 stencilRefs
[0], stencilVals
);
147 if (stencil
[1].enabled
&& front_facing
!= NULL
) {
148 /* do back face test */
149 LLVMValueRef back_res
;
151 back_res
= lp_build_stencil_test_single(bld
, &stencil
[1],
152 stencilRefs
[1], stencilVals
);
154 res
= lp_build_select(bld
, front_facing
, res
, back_res
);
162 * Apply the stencil operator (add/sub/keep/etc) to the given vector
164 * \return new stencil values vector
167 lp_build_stencil_op_single(struct lp_build_context
*bld
,
168 const struct pipe_stencil_state
*stencil
,
170 LLVMValueRef stencilRef
,
171 LLVMValueRef stencilVals
)
174 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
175 struct lp_type type
= bld
->type
;
177 LLVMValueRef max
= lp_build_const_int_vec(bld
->gallivm
, type
, 0xff);
184 stencil_op
= stencil
->fail_op
;
187 stencil_op
= stencil
->zfail_op
;
190 stencil_op
= stencil
->zpass_op
;
193 assert(0 && "Invalid stencil_op mode");
194 stencil_op
= PIPE_STENCIL_OP_KEEP
;
197 switch (stencil_op
) {
198 case PIPE_STENCIL_OP_KEEP
:
200 /* we can return early for this case */
202 case PIPE_STENCIL_OP_ZERO
:
205 case PIPE_STENCIL_OP_REPLACE
:
208 case PIPE_STENCIL_OP_INCR
:
209 res
= lp_build_add(bld
, stencilVals
, bld
->one
);
210 res
= lp_build_min(bld
, res
, max
);
212 case PIPE_STENCIL_OP_DECR
:
213 res
= lp_build_sub(bld
, stencilVals
, bld
->one
);
214 res
= lp_build_max(bld
, res
, bld
->zero
);
216 case PIPE_STENCIL_OP_INCR_WRAP
:
217 res
= lp_build_add(bld
, stencilVals
, bld
->one
);
218 res
= LLVMBuildAnd(builder
, res
, max
, "");
220 case PIPE_STENCIL_OP_DECR_WRAP
:
221 res
= lp_build_sub(bld
, stencilVals
, bld
->one
);
222 res
= LLVMBuildAnd(builder
, res
, max
, "");
224 case PIPE_STENCIL_OP_INVERT
:
225 res
= LLVMBuildNot(builder
, stencilVals
, "");
226 res
= LLVMBuildAnd(builder
, res
, max
, "");
229 assert(0 && "bad stencil op mode");
238 * Do the one or two-sided stencil test op/update.
241 lp_build_stencil_op(struct lp_build_context
*bld
,
242 const struct pipe_stencil_state stencil
[2],
244 LLVMValueRef stencilRefs
[2],
245 LLVMValueRef stencilVals
,
247 LLVMValueRef front_facing
)
250 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
253 assert(stencil
[0].enabled
);
255 /* do front face op */
256 res
= lp_build_stencil_op_single(bld
, &stencil
[0], op
,
257 stencilRefs
[0], stencilVals
);
259 if (stencil
[1].enabled
&& front_facing
!= NULL
) {
260 /* do back face op */
261 LLVMValueRef back_res
;
263 back_res
= lp_build_stencil_op_single(bld
, &stencil
[1], op
,
264 stencilRefs
[1], stencilVals
);
266 res
= lp_build_select(bld
, front_facing
, res
, back_res
);
269 if (stencil
[0].writemask
!= 0xff ||
270 (stencil
[1].enabled
&& front_facing
!= NULL
&& stencil
[1].writemask
!= 0xff)) {
271 /* mask &= stencil[0].writemask */
272 LLVMValueRef writemask
= lp_build_const_int_vec(bld
->gallivm
, bld
->type
,
273 stencil
[0].writemask
);
274 if (stencil
[1].enabled
&& stencil
[1].writemask
!= stencil
[0].writemask
&& front_facing
!= NULL
) {
275 LLVMValueRef back_writemask
= lp_build_const_int_vec(bld
->gallivm
, bld
->type
,
276 stencil
[1].writemask
);
277 writemask
= lp_build_select(bld
, front_facing
, writemask
, back_writemask
);
280 mask
= LLVMBuildAnd(builder
, mask
, writemask
, "");
281 /* res = (res & mask) | (stencilVals & ~mask) */
282 res
= lp_build_select_bitwise(bld
, mask
, res
, stencilVals
);
285 /* res = mask ? res : stencilVals */
286 res
= lp_build_select(bld
, mask
, res
, stencilVals
);
295 * Return a type that matches the depth/stencil format.
298 lp_depth_type(const struct util_format_description
*format_desc
,
304 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_ZS
);
305 assert(format_desc
->block
.width
== 1);
306 assert(format_desc
->block
.height
== 1);
308 memset(&type
, 0, sizeof type
);
309 type
.width
= format_desc
->block
.bits
;
311 z_swizzle
= format_desc
->swizzle
[0];
313 if (format_desc
->channel
[z_swizzle
].type
== UTIL_FORMAT_TYPE_FLOAT
) {
314 type
.floating
= TRUE
;
315 assert(z_swizzle
== 0);
316 assert(format_desc
->channel
[z_swizzle
].size
== format_desc
->block
.bits
);
318 else if(format_desc
->channel
[z_swizzle
].type
== UTIL_FORMAT_TYPE_UNSIGNED
) {
319 assert(format_desc
->block
.bits
<= 32);
320 assert(format_desc
->channel
[z_swizzle
].normalized
);
321 if (format_desc
->channel
[z_swizzle
].size
< format_desc
->block
.bits
) {
322 /* Prefer signed integers when possible, as SSE has less support
323 * for unsigned comparison;
332 type
.length
= length
;
339 * Compute bitmask and bit shift to apply to the incoming fragment Z values
340 * and the Z buffer values needed before doing the Z comparison.
342 * Note that we leave the Z bits in the position that we find them
343 * in the Z buffer (typically 0xffffff00 or 0x00ffffff). That lets us
344 * get by with fewer bit twiddling steps.
347 get_z_shift_and_mask(const struct util_format_description
*format_desc
,
348 unsigned *shift
, unsigned *width
, unsigned *mask
)
350 const unsigned total_bits
= format_desc
->block
.bits
;
353 unsigned padding_left
, padding_right
;
355 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_ZS
);
356 assert(format_desc
->block
.width
== 1);
357 assert(format_desc
->block
.height
== 1);
359 z_swizzle
= format_desc
->swizzle
[0];
361 if (z_swizzle
== UTIL_FORMAT_SWIZZLE_NONE
)
364 *width
= format_desc
->channel
[z_swizzle
].size
;
367 for (chan
= 0; chan
< z_swizzle
; ++chan
)
368 padding_right
+= format_desc
->channel
[chan
].size
;
371 total_bits
- (padding_right
+ *width
);
373 if (padding_left
|| padding_right
) {
374 unsigned long long mask_left
= (1ULL << (total_bits
- padding_left
)) - 1;
375 unsigned long long mask_right
= (1ULL << (padding_right
)) - 1;
376 *mask
= mask_left
^ mask_right
;
382 *shift
= padding_right
;
389 * Compute bitmask and bit shift to apply to the framebuffer pixel values
390 * to put the stencil bits in the least significant position.
394 get_s_shift_and_mask(const struct util_format_description
*format_desc
,
395 unsigned *shift
, unsigned *mask
)
400 s_swizzle
= format_desc
->swizzle
[1];
402 if (s_swizzle
== UTIL_FORMAT_SWIZZLE_NONE
)
406 for (chan
= 0; chan
< s_swizzle
; chan
++)
407 *shift
+= format_desc
->channel
[chan
].size
;
409 sz
= format_desc
->channel
[s_swizzle
].size
;
410 *mask
= (1U << sz
) - 1U;
417 * Perform the occlusion test and increase the counter.
418 * Test the depth mask. Add the number of channel which has none zero mask
419 * into the occlusion counter. e.g. maskvalue is {-1, -1, -1, -1}.
420 * The counter will add 4.
422 * \param type holds element type of the mask vector.
423 * \param maskvalue is the depth test mask.
424 * \param counter is a pointer of the uint32 counter.
427 lp_build_occlusion_count(struct gallivm_state
*gallivm
,
429 LLVMValueRef maskvalue
,
430 LLVMValueRef counter
)
432 LLVMBuilderRef builder
= gallivm
->builder
;
433 LLVMContextRef context
= gallivm
->context
;
434 LLVMValueRef countmask
= lp_build_const_int_vec(gallivm
, type
, 1);
435 LLVMValueRef count
, newcount
;
437 assert(type
.length
<= 16);
438 assert(type
.floating
);
440 if(util_cpu_caps
.has_sse
&& type
.length
== 4) {
441 const char *movmskintr
= "llvm.x86.sse.movmsk.ps";
442 const char *popcntintr
= "llvm.ctpop.i32";
443 LLVMValueRef bits
= LLVMBuildBitCast(builder
, maskvalue
,
444 lp_build_vec_type(gallivm
, type
), "");
445 bits
= lp_build_intrinsic_unary(builder
, movmskintr
,
446 LLVMInt32TypeInContext(context
), bits
);
447 count
= lp_build_intrinsic_unary(builder
, popcntintr
,
448 LLVMInt32TypeInContext(context
), bits
);
450 else if(util_cpu_caps
.has_avx
&& type
.length
== 8) {
451 const char *movmskintr
= "llvm.x86.avx.movmsk.ps.256";
452 const char *popcntintr
= "llvm.ctpop.i32";
453 LLVMValueRef bits
= LLVMBuildBitCast(builder
, maskvalue
,
454 lp_build_vec_type(gallivm
, type
), "");
455 bits
= lp_build_intrinsic_unary(builder
, movmskintr
,
456 LLVMInt32TypeInContext(context
), bits
);
457 count
= lp_build_intrinsic_unary(builder
, popcntintr
,
458 LLVMInt32TypeInContext(context
), bits
);
462 LLVMValueRef countv
= LLVMBuildAnd(builder
, maskvalue
, countmask
, "countv");
463 LLVMTypeRef counttype
= LLVMIntTypeInContext(context
, type
.length
* 8);
464 LLVMTypeRef i8vntype
= LLVMVectorType(LLVMInt8TypeInContext(context
), type
.length
* 4);
465 LLVMValueRef shufflev
, countd
;
466 LLVMValueRef shuffles
[16];
467 const char *popcntintr
= NULL
;
469 countv
= LLVMBuildBitCast(builder
, countv
, i8vntype
, "");
471 for (i
= 0; i
< type
.length
; i
++) {
472 shuffles
[i
] = lp_build_const_int32(gallivm
, 4*i
);
475 shufflev
= LLVMConstVector(shuffles
, type
.length
);
476 countd
= LLVMBuildShuffleVector(builder
, countv
, LLVMGetUndef(i8vntype
), shufflev
, "");
477 countd
= LLVMBuildBitCast(builder
, countd
, counttype
, "countd");
481 * this is bad on cpus without popcount (on x86 supported by intel
482 * nehalem, amd barcelona, and up - not tied to sse42).
483 * Would be much faster to just sum the 4 elements of the vector with
484 * some horizontal add (shuffle/add/shuffle/add after the initial and).
486 switch (type
.length
) {
488 popcntintr
= "llvm.ctpop.i32";
491 popcntintr
= "llvm.ctpop.i64";
494 popcntintr
= "llvm.ctpop.i128";
499 count
= lp_build_intrinsic_unary(builder
, popcntintr
, counttype
, countd
);
501 if (type
.length
> 4) {
502 count
= LLVMBuildTrunc(builder
, count
, LLVMIntTypeInContext(context
, 32), "");
505 newcount
= LLVMBuildLoad(builder
, counter
, "origcount");
506 newcount
= LLVMBuildAdd(builder
, newcount
, count
, "newcount");
507 LLVMBuildStore(builder
, newcount
, counter
);
512 * Load depth/stencil values.
513 * The stored values are linear, swizzle them.
515 * \param type the data type of the fragment depth/stencil values
516 * \param format_desc description of the depth/stencil surface
517 * \param loop_counter the current loop iteration
518 * \param depth_ptr pointer to the depth/stencil values of this 4x4 block
519 * \param depth_stride stride of the depth/stencil buffer
522 lp_build_depth_stencil_load_swizzled(struct gallivm_state
*gallivm
,
523 struct lp_type z_src_type
,
524 const struct util_format_description
*format_desc
,
525 LLVMValueRef depth_ptr
,
526 LLVMValueRef depth_stride
,
527 LLVMValueRef loop_counter
)
529 LLVMBuilderRef builder
= gallivm
->builder
;
530 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
/ 4];
531 LLVMValueRef zs_dst
, zs_dst1
, zs_dst2
;
532 LLVMValueRef zs_dst_ptr
;
533 LLVMValueRef depth_offset1
, depth_offset2
;
534 unsigned depth_bits
= format_desc
->block
.bits
/8;
535 struct lp_type zs_type
= lp_depth_type(format_desc
, z_src_type
.length
);
536 struct lp_type zs_load_type
= zs_type
;
537 zs_load_type
.length
= zs_load_type
.length
/ 2;
539 if (z_src_type
.length
== 4) {
541 LLVMValueRef looplsb
= LLVMBuildAnd(builder
, loop_counter
,
542 lp_build_const_int32(gallivm
, 1), "");
543 LLVMValueRef loopmsb
= LLVMBuildAnd(builder
, loop_counter
,
544 lp_build_const_int32(gallivm
, 2), "");
545 LLVMValueRef offset2
= LLVMBuildMul(builder
, loopmsb
,
547 depth_offset1
= LLVMBuildMul(builder
, looplsb
,
548 lp_build_const_int32(gallivm
, depth_bits
* 2), "");
549 depth_offset1
= LLVMBuildAdd(builder
, depth_offset1
, offset2
, "");
551 /* just concatenate the loaded 2x2 values into 4-wide vector */
552 for (i
= 0; i
< 4; i
++) {
553 shuffles
[i
] = lp_build_const_int32(gallivm
, i
);
558 LLVMValueRef loopx2
= LLVMBuildShl(builder
, loop_counter
,
559 lp_build_const_int32(gallivm
, 1), "");
560 assert(z_src_type
.length
== 8);
561 depth_offset1
= LLVMBuildMul(builder
, loopx2
, depth_stride
, "");
563 * We load 2x4 values, and need to swizzle them (order
564 * 0,1,4,5,2,3,6,7) - not so hot with avx unfortunately.
566 for (i
= 0; i
< 8; i
++) {
568 shuffles
[i
] = lp_build_const_int32(gallivm
, (i
&1) + (i
&2) * 2 + (i
&4) / 2);
572 depth_offset2
= LLVMBuildAdd(builder
, depth_offset1
, depth_stride
, "");
574 /* Load current z/stencil values from z/stencil buffer */
575 zs_dst_ptr
= LLVMBuildGEP(builder
, depth_ptr
, &depth_offset1
, 1, "");
576 zs_dst_ptr
= LLVMBuildBitCast(builder
,
578 LLVMPointerType(lp_build_vec_type(gallivm
, zs_load_type
), 0), "");
579 zs_dst1
= LLVMBuildLoad(builder
, zs_dst_ptr
, "");
580 zs_dst_ptr
= LLVMBuildGEP(builder
, depth_ptr
, &depth_offset2
, 1, "");
581 zs_dst_ptr
= LLVMBuildBitCast(builder
,
583 LLVMPointerType(lp_build_vec_type(gallivm
, zs_load_type
), 0), "");
584 zs_dst2
= LLVMBuildLoad(builder
, zs_dst_ptr
, "");
586 zs_dst
= LLVMBuildShuffleVector(builder
, zs_dst1
, zs_dst2
,
587 LLVMConstVector(shuffles
, zs_type
.length
), "");
589 if (format_desc
->block
.bits
< z_src_type
.width
) {
590 /* Extend destination ZS values (e.g., when reading from Z16_UNORM) */
591 zs_dst
= LLVMBuildZExt(builder
, zs_dst
, lp_build_int_vec_type(gallivm
, z_src_type
), "");
594 lp_build_name(zs_dst
, "zs_dst");
600 * Store depth/stencil values.
601 * Incoming values are swizzled (typically n 2x2 quads), stored linear.
602 * If there's a mask it will do reload/select/store otherwise just store.
604 * \param type the data type of the fragment depth/stencil values
605 * \param format_desc description of the depth/stencil surface
606 * \param mask the alive/dead pixel mask for the quad (vector)
607 * \param loop_counter the current loop iteration
608 * \param depth_ptr pointer to the depth/stencil values of this 4x4 block
609 * \param depth_stride stride of the depth/stencil buffer
610 * \param zs_value the depth/stencil values to store
613 lp_build_depth_stencil_write_swizzled(struct gallivm_state
*gallivm
,
614 struct lp_type z_src_type
,
615 const struct util_format_description
*format_desc
,
616 struct lp_build_mask_context
*mask
,
617 LLVMValueRef loop_counter
,
618 LLVMValueRef depth_ptr
,
619 LLVMValueRef depth_stride
,
620 LLVMValueRef zs_value
)
622 struct lp_build_context z_bld
;
623 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
/ 4];
624 LLVMBuilderRef builder
= gallivm
->builder
;
625 LLVMValueRef mask_value
= NULL
;
626 LLVMValueRef zs_dst
= NULL
, zs_dst1
, zs_dst2
;
627 LLVMValueRef zs_dst_ptr1
, zs_dst_ptr2
;
628 LLVMValueRef depth_offset1
, depth_offset2
;
629 unsigned depth_bits
= format_desc
->block
.bits
/8;
630 struct lp_type zs_type
= lp_depth_type(format_desc
, z_src_type
.length
);
631 struct lp_type zs_load_type
= zs_type
;
632 zs_load_type
.length
= zs_load_type
.length
/ 2;
634 lp_build_context_init(&z_bld
, gallivm
, zs_type
);
637 * This is far from ideal, at least for late depth write we should do this
638 * outside the fs loop to avoid all the swizzle stuff.
640 if (z_src_type
.length
== 4) {
642 LLVMValueRef looplsb
= LLVMBuildAnd(builder
, loop_counter
,
643 lp_build_const_int32(gallivm
, 1), "");
644 LLVMValueRef loopmsb
= LLVMBuildAnd(builder
, loop_counter
,
645 lp_build_const_int32(gallivm
, 2), "");
646 LLVMValueRef offset2
= LLVMBuildMul(builder
, loopmsb
,
648 depth_offset1
= LLVMBuildMul(builder
, looplsb
,
649 lp_build_const_int32(gallivm
, depth_bits
* 2), "");
650 depth_offset1
= LLVMBuildAdd(builder
, depth_offset1
, offset2
, "");
652 /* just concatenate the loaded 2x2 values into 4-wide vector */
653 for (i
= 0; i
< 4; i
++) {
654 shuffles
[i
] = lp_build_const_int32(gallivm
, i
);
659 LLVMValueRef loopx2
= LLVMBuildShl(builder
, loop_counter
,
660 lp_build_const_int32(gallivm
, 1), "");
661 assert(z_src_type
.length
== 8);
662 depth_offset1
= LLVMBuildMul(builder
, loopx2
, depth_stride
, "");
664 * We load 2x4 values, and need to swizzle them (order
665 * 0,1,4,5,2,3,6,7) - not so hot with avx unfortunately.
667 for (i
= 0; i
< 8; i
++) {
668 shuffles
[i
] = lp_build_const_int32(gallivm
, (i
&1) + (i
&2) * 2 + (i
&4) / 2);
673 depth_offset2
= LLVMBuildAdd(builder
, depth_offset1
, depth_stride
, "");
675 zs_dst_ptr1
= LLVMBuildGEP(builder
, depth_ptr
, &depth_offset1
, 1, "");
676 zs_dst_ptr1
= LLVMBuildBitCast(builder
,
678 LLVMPointerType(lp_build_vec_type(gallivm
, zs_load_type
), 0), "");
679 zs_dst_ptr2
= LLVMBuildGEP(builder
, depth_ptr
, &depth_offset2
, 1, "");
680 zs_dst_ptr2
= LLVMBuildBitCast(builder
,
682 LLVMPointerType(lp_build_vec_type(gallivm
, zs_load_type
), 0), "");
685 zs_dst1
= LLVMBuildLoad(builder
, zs_dst_ptr1
, "");
686 zs_dst2
= LLVMBuildLoad(builder
, zs_dst_ptr2
, "");
687 zs_dst
= LLVMBuildShuffleVector(builder
, zs_dst1
, zs_dst2
,
688 LLVMConstVector(shuffles
, zs_type
.length
),
691 mask_value
= lp_build_mask_value(mask
);
694 if (zs_type
.width
< z_src_type
.width
) {
695 /* Truncate incoming ZS and mask values (e.g., when writing to Z16_UNORM) */
696 zs_value
= LLVMBuildTrunc(builder
, zs_value
, z_bld
.vec_type
, "");
698 mask_value
= LLVMBuildTrunc(builder
, mask_value
, z_bld
.vec_type
, "");
702 zs_value
= lp_build_select(&z_bld
, mask_value
, zs_value
, zs_dst
);
705 if (z_src_type
.length
== 4) {
706 zs_dst1
= lp_build_extract_range(gallivm
, zs_value
, 0, 2);
707 zs_dst2
= lp_build_extract_range(gallivm
, zs_value
, 2, 2);
710 assert(z_src_type
.length
== 8);
711 zs_dst1
= LLVMBuildShuffleVector(builder
, zs_value
, zs_value
,
712 LLVMConstVector(&shuffles
[0],
713 zs_load_type
.length
),
715 zs_dst2
= LLVMBuildShuffleVector(builder
, zs_value
, zs_value
,
716 LLVMConstVector(&shuffles
[4],
717 zs_load_type
.length
),
721 LLVMBuildStore(builder
, zs_dst1
, zs_dst_ptr1
);
722 LLVMBuildStore(builder
, zs_dst2
, zs_dst_ptr2
);
726 * Generate code for performing depth and/or stencil tests.
727 * We operate on a vector of values (typically n 2x2 quads).
729 * \param depth the depth test state
730 * \param stencil the front/back stencil state
731 * \param type the data type of the fragment depth/stencil values
732 * \param format_desc description of the depth/stencil surface
733 * \param mask the alive/dead pixel mask for the quad (vector)
734 * \param stencil_refs the front/back stencil ref values (scalar)
735 * \param z_src the incoming depth/stencil values (n 2x2 quad values, float32)
736 * \param zs_dst the depth/stencil values in framebuffer
737 * \param face contains boolean value indicating front/back facing polygon
740 lp_build_depth_stencil_test(struct gallivm_state
*gallivm
,
741 const struct pipe_depth_state
*depth
,
742 const struct pipe_stencil_state stencil
[2],
743 struct lp_type z_src_type
,
744 const struct util_format_description
*format_desc
,
745 struct lp_build_mask_context
*mask
,
746 LLVMValueRef stencil_refs
[2],
750 LLVMValueRef
*zs_value
,
753 LLVMBuilderRef builder
= gallivm
->builder
;
754 struct lp_type zs_type
;
755 struct lp_type z_type
;
756 struct lp_build_context z_bld
;
757 struct lp_build_context s_bld
;
758 struct lp_type s_type
;
759 unsigned z_shift
= 0, z_width
= 0, z_mask
= 0;
760 LLVMValueRef z_dst
= NULL
;
761 LLVMValueRef stencil_vals
= NULL
;
762 LLVMValueRef z_bitmask
= NULL
, stencil_shift
= NULL
;
763 LLVMValueRef z_pass
= NULL
, s_pass_mask
= NULL
;
764 LLVMValueRef orig_mask
= lp_build_mask_value(mask
);
765 LLVMValueRef front_facing
= NULL
;
769 * Depths are expected to be between 0 and 1, even if they are stored in
770 * floats. Setting these bits here will ensure that the lp_build_conv() call
771 * below won't try to unnecessarily clamp the incoming values.
773 if(z_src_type
.floating
) {
774 z_src_type
.sign
= FALSE
;
775 z_src_type
.norm
= TRUE
;
778 assert(!z_src_type
.sign
);
779 assert(z_src_type
.norm
);
782 /* Pick the type matching the depth-stencil format. */
783 zs_type
= lp_depth_type(format_desc
, z_src_type
.length
);
785 /* Pick the intermediate type for depth operations. */
787 /* FIXME: Cope with a depth test type with higher bit width. */
788 assert(zs_type
.width
<= z_src_type
.width
);
789 z_type
.width
= z_src_type
.width
;
790 assert(z_type
.length
== z_src_type
.length
);
792 /* FIXME: for non-float depth/stencil might generate better code
793 * if we'd always split it up to use 128bit operations.
794 * For stencil we'd almost certainly want to pack to 8xi16 values,
795 * for z just run twice.
798 /* Sanity checking */
800 const unsigned z_swizzle
= format_desc
->swizzle
[0];
801 const unsigned s_swizzle
= format_desc
->swizzle
[1];
803 assert(z_swizzle
!= UTIL_FORMAT_SWIZZLE_NONE
||
804 s_swizzle
!= UTIL_FORMAT_SWIZZLE_NONE
);
806 assert(depth
->enabled
|| stencil
[0].enabled
);
808 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_ZS
);
809 assert(format_desc
->block
.width
== 1);
810 assert(format_desc
->block
.height
== 1);
812 if (stencil
[0].enabled
) {
813 assert(s_swizzle
< 4);
814 assert(format_desc
->channel
[s_swizzle
].type
== UTIL_FORMAT_TYPE_UNSIGNED
);
815 assert(format_desc
->channel
[s_swizzle
].pure_integer
);
816 assert(!format_desc
->channel
[s_swizzle
].normalized
);
817 assert(format_desc
->channel
[s_swizzle
].size
== 8);
820 if (depth
->enabled
) {
821 assert(z_swizzle
< 4);
822 assert(format_desc
->block
.bits
<= z_type
.width
);
823 if (z_type
.floating
) {
824 assert(z_swizzle
== 0);
825 assert(format_desc
->channel
[z_swizzle
].type
==
826 UTIL_FORMAT_TYPE_FLOAT
);
827 assert(format_desc
->channel
[z_swizzle
].size
==
828 format_desc
->block
.bits
);
831 assert(format_desc
->channel
[z_swizzle
].type
==
832 UTIL_FORMAT_TYPE_UNSIGNED
);
833 assert(format_desc
->channel
[z_swizzle
].normalized
);
834 assert(!z_type
.fixed
);
840 /* Setup build context for Z vals */
841 lp_build_context_init(&z_bld
, gallivm
, z_type
);
843 /* Setup build context for stencil vals */
844 s_type
= lp_int_type(z_type
);
845 lp_build_context_init(&s_bld
, gallivm
, s_type
);
847 /* Compute and apply the Z/stencil bitmasks and shifts.
850 unsigned s_shift
, s_mask
;
852 if (get_z_shift_and_mask(format_desc
, &z_shift
, &z_width
, &z_mask
)) {
853 if (z_mask
!= 0xffffffff) {
854 z_bitmask
= lp_build_const_int_vec(gallivm
, z_type
, z_mask
);
858 * Align the framebuffer Z 's LSB to the right.
861 LLVMValueRef shift
= lp_build_const_int_vec(gallivm
, z_type
, z_shift
);
862 z_dst
= LLVMBuildLShr(builder
, zs_dst
, shift
, "z_dst");
863 } else if (z_bitmask
) {
864 /* TODO: Instead of loading a mask from memory and ANDing, it's
865 * probably faster to just shake the bits with two shifts. */
866 z_dst
= LLVMBuildAnd(builder
, zs_dst
, z_bitmask
, "z_dst");
869 lp_build_name(z_dst
, "z_dst");
873 if (get_s_shift_and_mask(format_desc
, &s_shift
, &s_mask
)) {
875 LLVMValueRef shift
= lp_build_const_int_vec(gallivm
, s_type
, s_shift
);
876 stencil_vals
= LLVMBuildLShr(builder
, zs_dst
, shift
, "");
877 stencil_shift
= shift
; /* used below */
880 stencil_vals
= zs_dst
;
883 if (s_mask
!= 0xffffffff) {
884 LLVMValueRef mask
= lp_build_const_int_vec(gallivm
, s_type
, s_mask
);
885 stencil_vals
= LLVMBuildAnd(builder
, stencil_vals
, mask
, "");
888 lp_build_name(stencil_vals
, "s_dst");
892 if (stencil
[0].enabled
) {
895 LLVMValueRef zero
= lp_build_const_int32(gallivm
, 0);
897 /* front_facing = face != 0 ? ~0 : 0 */
898 front_facing
= LLVMBuildICmp(builder
, LLVMIntNE
, face
, zero
, "");
899 front_facing
= LLVMBuildSExt(builder
, front_facing
,
900 LLVMIntTypeInContext(gallivm
->context
,
901 s_bld
.type
.length
*s_bld
.type
.width
),
903 front_facing
= LLVMBuildBitCast(builder
, front_facing
,
904 s_bld
.int_vec_type
, "");
907 /* convert scalar stencil refs into vectors */
908 stencil_refs
[0] = lp_build_broadcast_scalar(&s_bld
, stencil_refs
[0]);
909 stencil_refs
[1] = lp_build_broadcast_scalar(&s_bld
, stencil_refs
[1]);
911 s_pass_mask
= lp_build_stencil_test(&s_bld
, stencil
,
912 stencil_refs
, stencil_vals
,
915 /* apply stencil-fail operator */
917 LLVMValueRef s_fail_mask
= lp_build_andnot(&s_bld
, orig_mask
, s_pass_mask
);
918 stencil_vals
= lp_build_stencil_op(&s_bld
, stencil
, S_FAIL_OP
,
919 stencil_refs
, stencil_vals
,
920 s_fail_mask
, front_facing
);
924 if (depth
->enabled
) {
926 * Convert fragment Z to the desired type, aligning the LSB to the right.
929 assert(z_type
.width
== z_src_type
.width
);
930 assert(z_type
.length
== z_src_type
.length
);
931 assert(lp_check_value(z_src_type
, z_src
));
932 if (z_src_type
.floating
) {
934 * Convert from floating point values
937 if (!z_type
.floating
) {
938 z_src
= lp_build_clamped_float_to_unsigned_norm(gallivm
,
945 * Convert from unsigned normalized values.
948 assert(!z_src_type
.sign
);
949 assert(!z_src_type
.fixed
);
950 assert(z_src_type
.norm
);
951 assert(!z_type
.floating
);
952 if (z_src_type
.width
> z_width
) {
953 LLVMValueRef shift
= lp_build_const_int_vec(gallivm
, z_src_type
,
954 z_src_type
.width
- z_width
);
955 z_src
= LLVMBuildLShr(builder
, z_src
, shift
, "");
958 assert(lp_check_value(z_type
, z_src
));
960 lp_build_name(z_src
, "z_src");
962 /* compare src Z to dst Z, returning 'pass' mask */
963 z_pass
= lp_build_cmp(&z_bld
, depth
->func
, z_src
, z_dst
);
965 if (!stencil
[0].enabled
) {
966 /* We can potentially skip all remaining operations here, but only
967 * if stencil is disabled because we still need to update the stencil
968 * buffer values. Don't need to update Z buffer values.
970 lp_build_mask_update(mask
, z_pass
);
973 lp_build_mask_check(mask
);
978 if (depth
->writemask
) {
979 LLVMValueRef zselectmask
;
981 /* mask off bits that failed Z test */
982 zselectmask
= LLVMBuildAnd(builder
, orig_mask
, z_pass
, "");
984 /* mask off bits that failed stencil test */
986 zselectmask
= LLVMBuildAnd(builder
, zselectmask
, s_pass_mask
, "");
989 /* Mix the old and new Z buffer values.
990 * z_dst[i] = zselectmask[i] ? z_src[i] : z_dst[i]
992 z_dst
= lp_build_select(&z_bld
, zselectmask
, z_src
, z_dst
);
995 if (stencil
[0].enabled
) {
996 /* update stencil buffer values according to z pass/fail result */
997 LLVMValueRef z_fail_mask
, z_pass_mask
;
999 /* apply Z-fail operator */
1000 z_fail_mask
= lp_build_andnot(&z_bld
, orig_mask
, z_pass
);
1001 stencil_vals
= lp_build_stencil_op(&s_bld
, stencil
, Z_FAIL_OP
,
1002 stencil_refs
, stencil_vals
,
1003 z_fail_mask
, front_facing
);
1005 /* apply Z-pass operator */
1006 z_pass_mask
= LLVMBuildAnd(builder
, orig_mask
, z_pass
, "");
1007 stencil_vals
= lp_build_stencil_op(&s_bld
, stencil
, Z_PASS_OP
,
1008 stencil_refs
, stencil_vals
,
1009 z_pass_mask
, front_facing
);
1013 /* No depth test: apply Z-pass operator to stencil buffer values which
1014 * passed the stencil test.
1016 s_pass_mask
= LLVMBuildAnd(builder
, orig_mask
, s_pass_mask
, "");
1017 stencil_vals
= lp_build_stencil_op(&s_bld
, stencil
, Z_PASS_OP
,
1018 stencil_refs
, stencil_vals
,
1019 s_pass_mask
, front_facing
);
1022 /* Put Z and ztencil bits in the right place */
1023 if (z_dst
&& z_shift
) {
1024 LLVMValueRef shift
= lp_build_const_int_vec(gallivm
, z_type
, z_shift
);
1025 z_dst
= LLVMBuildShl(builder
, z_dst
, shift
, "");
1027 if (stencil_vals
&& stencil_shift
)
1028 stencil_vals
= LLVMBuildShl(builder
, stencil_vals
,
1031 /* Finally, merge/store the z/stencil values */
1032 if ((depth
->enabled
&& depth
->writemask
) ||
1033 (stencil
[0].enabled
&& stencil
[0].writemask
)) {
1035 if (z_dst
&& stencil_vals
)
1036 zs_dst
= LLVMBuildOr(builder
, z_dst
, stencil_vals
, "");
1040 zs_dst
= stencil_vals
;
1046 lp_build_mask_update(mask
, s_pass_mask
);
1048 if (depth
->enabled
&& stencil
[0].enabled
)
1049 lp_build_mask_update(mask
, z_pass
);
1052 lp_build_mask_check(mask
);