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/format/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
== 32);
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
)
353 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_ZS
);
354 assert(format_desc
->block
.width
== 1);
355 assert(format_desc
->block
.height
== 1);
357 /* 64bit d/s format is special already extracted 32 bits */
358 total_bits
= format_desc
->block
.bits
> 32 ? 32 : format_desc
->block
.bits
;
360 z_swizzle
= format_desc
->swizzle
[0];
362 if (z_swizzle
== PIPE_SWIZZLE_NONE
)
365 *width
= format_desc
->channel
[z_swizzle
].size
;
366 /* & 31 is for the same reason as the 32-bit limit above */
367 *shift
= format_desc
->channel
[z_swizzle
].shift
& 31;
369 if (*width
== total_bits
) {
372 *mask
= ((1 << *width
) - 1) << *shift
;
380 * Compute bitmask and bit shift to apply to the framebuffer pixel values
381 * to put the stencil bits in the least significant position.
385 get_s_shift_and_mask(const struct util_format_description
*format_desc
,
386 unsigned *shift
, unsigned *mask
)
391 s_swizzle
= format_desc
->swizzle
[1];
393 if (s_swizzle
== PIPE_SWIZZLE_NONE
)
396 /* just special case 64bit d/s format */
397 if (format_desc
->block
.bits
> 32) {
398 /* XXX big-endian? */
399 assert(format_desc
->format
== PIPE_FORMAT_Z32_FLOAT_S8X24_UINT
);
405 *shift
= format_desc
->channel
[s_swizzle
].shift
;
406 sz
= format_desc
->channel
[s_swizzle
].size
;
407 *mask
= (1U << sz
) - 1U;
414 * Perform the occlusion test and increase the counter.
415 * Test the depth mask. Add the number of channel which has none zero mask
416 * into the occlusion counter. e.g. maskvalue is {-1, -1, -1, -1}.
417 * The counter will add 4.
418 * TODO: could get that out of the fs loop.
420 * \param type holds element type of the mask vector.
421 * \param maskvalue is the depth test mask.
422 * \param counter is a pointer of the uint32 counter.
425 lp_build_occlusion_count(struct gallivm_state
*gallivm
,
427 LLVMValueRef maskvalue
,
428 LLVMValueRef counter
)
430 LLVMBuilderRef builder
= gallivm
->builder
;
431 LLVMContextRef context
= gallivm
->context
;
432 LLVMValueRef countmask
= lp_build_const_int_vec(gallivm
, type
, 1);
433 LLVMValueRef count
, newcount
;
435 assert(type
.length
<= 16);
436 assert(type
.floating
);
438 if(util_cpu_caps
.has_sse
&& type
.length
== 4) {
439 const char *movmskintr
= "llvm.x86.sse.movmsk.ps";
440 const char *popcntintr
= "llvm.ctpop.i32";
441 LLVMValueRef bits
= LLVMBuildBitCast(builder
, maskvalue
,
442 lp_build_vec_type(gallivm
, type
), "");
443 bits
= lp_build_intrinsic_unary(builder
, movmskintr
,
444 LLVMInt32TypeInContext(context
), bits
);
445 count
= lp_build_intrinsic_unary(builder
, popcntintr
,
446 LLVMInt32TypeInContext(context
), bits
);
447 count
= LLVMBuildZExt(builder
, count
, LLVMIntTypeInContext(context
, 64), "");
449 else if(util_cpu_caps
.has_avx
&& type
.length
== 8) {
450 const char *movmskintr
= "llvm.x86.avx.movmsk.ps.256";
451 const char *popcntintr
= "llvm.ctpop.i32";
452 LLVMValueRef bits
= LLVMBuildBitCast(builder
, maskvalue
,
453 lp_build_vec_type(gallivm
, type
), "");
454 bits
= lp_build_intrinsic_unary(builder
, movmskintr
,
455 LLVMInt32TypeInContext(context
), bits
);
456 count
= lp_build_intrinsic_unary(builder
, popcntintr
,
457 LLVMInt32TypeInContext(context
), bits
);
458 count
= LLVMBuildZExt(builder
, count
, LLVMIntTypeInContext(context
, 64), "");
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
> 8) {
502 count
= LLVMBuildTrunc(builder
, count
, LLVMIntTypeInContext(context
, 64), "");
504 else if (type
.length
< 8) {
505 count
= LLVMBuildZExt(builder
, count
, LLVMIntTypeInContext(context
, 64), "");
508 newcount
= LLVMBuildLoad(builder
, counter
, "origcount");
509 newcount
= LLVMBuildAdd(builder
, newcount
, count
, "newcount");
510 LLVMBuildStore(builder
, newcount
, counter
);
515 * Load depth/stencil values.
516 * The stored values are linear, swizzle them.
518 * \param type the data type of the fragment depth/stencil values
519 * \param format_desc description of the depth/stencil surface
520 * \param is_1d whether this resource has only one dimension
521 * \param loop_counter the current loop iteration
522 * \param depth_ptr pointer to the depth/stencil values of this 4x4 block
523 * \param depth_stride stride of the depth/stencil buffer
524 * \param z_fb contains z values loaded from fb (may include padding)
525 * \param s_fb contains s values loaded from fb (may include padding)
528 lp_build_depth_stencil_load_swizzled(struct gallivm_state
*gallivm
,
529 struct lp_type z_src_type
,
530 const struct util_format_description
*format_desc
,
532 LLVMValueRef depth_ptr
,
533 LLVMValueRef depth_stride
,
536 LLVMValueRef loop_counter
)
538 LLVMBuilderRef builder
= gallivm
->builder
;
539 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
/ 4];
540 LLVMValueRef zs_dst1
, zs_dst2
;
541 LLVMValueRef zs_dst_ptr
;
542 LLVMValueRef depth_offset1
, depth_offset2
;
543 LLVMTypeRef load_ptr_type
;
544 unsigned depth_bytes
= format_desc
->block
.bits
/ 8;
545 struct lp_type zs_type
= lp_depth_type(format_desc
, z_src_type
.length
);
546 struct lp_type zs_load_type
= zs_type
;
548 zs_load_type
.length
= zs_load_type
.length
/ 2;
549 load_ptr_type
= LLVMPointerType(lp_build_vec_type(gallivm
, zs_load_type
), 0);
551 if (z_src_type
.length
== 4) {
553 LLVMValueRef looplsb
= LLVMBuildAnd(builder
, loop_counter
,
554 lp_build_const_int32(gallivm
, 1), "");
555 LLVMValueRef loopmsb
= LLVMBuildAnd(builder
, loop_counter
,
556 lp_build_const_int32(gallivm
, 2), "");
557 LLVMValueRef offset2
= LLVMBuildMul(builder
, loopmsb
,
559 depth_offset1
= LLVMBuildMul(builder
, looplsb
,
560 lp_build_const_int32(gallivm
, depth_bytes
* 2), "");
561 depth_offset1
= LLVMBuildAdd(builder
, depth_offset1
, offset2
, "");
563 /* just concatenate the loaded 2x2 values into 4-wide vector */
564 for (i
= 0; i
< 4; i
++) {
565 shuffles
[i
] = lp_build_const_int32(gallivm
, i
);
570 LLVMValueRef loopx2
= LLVMBuildShl(builder
, loop_counter
,
571 lp_build_const_int32(gallivm
, 1), "");
572 assert(z_src_type
.length
== 8);
573 depth_offset1
= LLVMBuildMul(builder
, loopx2
, depth_stride
, "");
575 * We load 2x4 values, and need to swizzle them (order
576 * 0,1,4,5,2,3,6,7) - not so hot with avx unfortunately.
578 for (i
= 0; i
< 8; i
++) {
579 shuffles
[i
] = lp_build_const_int32(gallivm
, (i
&1) + (i
&2) * 2 + (i
&4) / 2);
583 depth_offset2
= LLVMBuildAdd(builder
, depth_offset1
, depth_stride
, "");
585 /* Load current z/stencil values from z/stencil buffer */
586 zs_dst_ptr
= LLVMBuildGEP(builder
, depth_ptr
, &depth_offset1
, 1, "");
587 zs_dst_ptr
= LLVMBuildBitCast(builder
, zs_dst_ptr
, load_ptr_type
, "");
588 zs_dst1
= LLVMBuildLoad(builder
, zs_dst_ptr
, "");
590 zs_dst2
= lp_build_undef(gallivm
, zs_load_type
);
593 zs_dst_ptr
= LLVMBuildGEP(builder
, depth_ptr
, &depth_offset2
, 1, "");
594 zs_dst_ptr
= LLVMBuildBitCast(builder
, zs_dst_ptr
, load_ptr_type
, "");
595 zs_dst2
= LLVMBuildLoad(builder
, zs_dst_ptr
, "");
598 *z_fb
= LLVMBuildShuffleVector(builder
, zs_dst1
, zs_dst2
,
599 LLVMConstVector(shuffles
, zs_type
.length
), "");
602 if (format_desc
->block
.bits
== 8) {
603 /* Extend stencil-only 8 bit values (S8_UINT) */
604 *s_fb
= LLVMBuildZExt(builder
, *s_fb
,
605 lp_build_int_vec_type(gallivm
, z_src_type
), "");
608 if (format_desc
->block
.bits
< z_src_type
.width
) {
609 /* Extend destination ZS values (e.g., when reading from Z16_UNORM) */
610 *z_fb
= LLVMBuildZExt(builder
, *z_fb
,
611 lp_build_int_vec_type(gallivm
, z_src_type
), "");
614 else if (format_desc
->block
.bits
> 32) {
615 /* rely on llvm to handle too wide vector we have here nicely */
617 struct lp_type typex2
= zs_type
;
618 struct lp_type s_type
= zs_type
;
619 LLVMValueRef shuffles1
[LP_MAX_VECTOR_LENGTH
/ 4];
620 LLVMValueRef shuffles2
[LP_MAX_VECTOR_LENGTH
/ 4];
623 typex2
.width
= typex2
.width
/ 2;
624 typex2
.length
= typex2
.length
* 2;
625 s_type
.width
= s_type
.width
/ 2;
628 tmp
= LLVMBuildBitCast(builder
, *z_fb
,
629 lp_build_vec_type(gallivm
, typex2
), "");
631 for (i
= 0; i
< zs_type
.length
; i
++) {
632 shuffles1
[i
] = lp_build_const_int32(gallivm
, i
* 2);
633 shuffles2
[i
] = lp_build_const_int32(gallivm
, i
* 2 + 1);
635 *z_fb
= LLVMBuildShuffleVector(builder
, tmp
, tmp
,
636 LLVMConstVector(shuffles1
, zs_type
.length
), "");
637 *s_fb
= LLVMBuildShuffleVector(builder
, tmp
, tmp
,
638 LLVMConstVector(shuffles2
, zs_type
.length
), "");
639 *s_fb
= LLVMBuildBitCast(builder
, *s_fb
,
640 lp_build_vec_type(gallivm
, s_type
), "");
641 lp_build_name(*s_fb
, "s_dst");
644 lp_build_name(*z_fb
, "z_dst");
645 lp_build_name(*s_fb
, "s_dst");
646 lp_build_name(*z_fb
, "z_dst");
650 * Store depth/stencil values.
651 * Incoming values are swizzled (typically n 2x2 quads), stored linear.
652 * If there's a mask it will do select/store otherwise just store.
654 * \param type the data type of the fragment depth/stencil values
655 * \param format_desc description of the depth/stencil surface
656 * \param is_1d whether this resource has only one dimension
657 * \param mask_value the alive/dead pixel mask for the quad (vector)
658 * \param z_fb z values read from fb (with padding)
659 * \param s_fb s values read from fb (with padding)
660 * \param loop_counter the current loop iteration
661 * \param depth_ptr pointer to the depth/stencil values of this 4x4 block
662 * \param depth_stride stride of the depth/stencil buffer
663 * \param z_value the depth values to store (with padding)
664 * \param s_value the stencil values to store (with padding)
667 lp_build_depth_stencil_write_swizzled(struct gallivm_state
*gallivm
,
668 struct lp_type z_src_type
,
669 const struct util_format_description
*format_desc
,
671 LLVMValueRef mask_value
,
674 LLVMValueRef loop_counter
,
675 LLVMValueRef depth_ptr
,
676 LLVMValueRef depth_stride
,
677 LLVMValueRef z_value
,
678 LLVMValueRef s_value
)
680 struct lp_build_context z_bld
;
681 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
/ 4];
682 LLVMBuilderRef builder
= gallivm
->builder
;
683 LLVMValueRef zs_dst1
, zs_dst2
;
684 LLVMValueRef zs_dst_ptr1
, zs_dst_ptr2
;
685 LLVMValueRef depth_offset1
, depth_offset2
;
686 LLVMTypeRef load_ptr_type
;
687 unsigned depth_bytes
= format_desc
->block
.bits
/ 8;
688 struct lp_type zs_type
= lp_depth_type(format_desc
, z_src_type
.length
);
689 struct lp_type z_type
= zs_type
;
690 struct lp_type zs_load_type
= zs_type
;
692 zs_load_type
.length
= zs_load_type
.length
/ 2;
693 load_ptr_type
= LLVMPointerType(lp_build_vec_type(gallivm
, zs_load_type
), 0);
695 z_type
.width
= z_src_type
.width
;
697 lp_build_context_init(&z_bld
, gallivm
, z_type
);
700 * This is far from ideal, at least for late depth write we should do this
701 * outside the fs loop to avoid all the swizzle stuff.
703 if (z_src_type
.length
== 4) {
704 LLVMValueRef looplsb
= LLVMBuildAnd(builder
, loop_counter
,
705 lp_build_const_int32(gallivm
, 1), "");
706 LLVMValueRef loopmsb
= LLVMBuildAnd(builder
, loop_counter
,
707 lp_build_const_int32(gallivm
, 2), "");
708 LLVMValueRef offset2
= LLVMBuildMul(builder
, loopmsb
,
710 depth_offset1
= LLVMBuildMul(builder
, looplsb
,
711 lp_build_const_int32(gallivm
, depth_bytes
* 2), "");
712 depth_offset1
= LLVMBuildAdd(builder
, depth_offset1
, offset2
, "");
716 LLVMValueRef loopx2
= LLVMBuildShl(builder
, loop_counter
,
717 lp_build_const_int32(gallivm
, 1), "");
718 assert(z_src_type
.length
== 8);
719 depth_offset1
= LLVMBuildMul(builder
, loopx2
, depth_stride
, "");
721 * We load 2x4 values, and need to swizzle them (order
722 * 0,1,4,5,2,3,6,7) - not so hot with avx unfortunately.
724 for (i
= 0; i
< 8; i
++) {
725 shuffles
[i
] = lp_build_const_int32(gallivm
, (i
&1) + (i
&2) * 2 + (i
&4) / 2);
729 depth_offset2
= LLVMBuildAdd(builder
, depth_offset1
, depth_stride
, "");
731 zs_dst_ptr1
= LLVMBuildGEP(builder
, depth_ptr
, &depth_offset1
, 1, "");
732 zs_dst_ptr1
= LLVMBuildBitCast(builder
, zs_dst_ptr1
, load_ptr_type
, "");
733 zs_dst_ptr2
= LLVMBuildGEP(builder
, depth_ptr
, &depth_offset2
, 1, "");
734 zs_dst_ptr2
= LLVMBuildBitCast(builder
, zs_dst_ptr2
, load_ptr_type
, "");
736 if (format_desc
->block
.bits
> 32) {
737 s_value
= LLVMBuildBitCast(builder
, s_value
, z_bld
.vec_type
, "");
741 z_value
= lp_build_select(&z_bld
, mask_value
, z_value
, z_fb
);
742 if (format_desc
->block
.bits
> 32) {
743 s_fb
= LLVMBuildBitCast(builder
, s_fb
, z_bld
.vec_type
, "");
744 s_value
= lp_build_select(&z_bld
, mask_value
, s_value
, s_fb
);
748 if (zs_type
.width
< z_src_type
.width
) {
749 /* Truncate ZS values (e.g., when writing to Z16_UNORM) */
750 z_value
= LLVMBuildTrunc(builder
, z_value
,
751 lp_build_int_vec_type(gallivm
, zs_type
), "");
754 if (format_desc
->block
.bits
<= 32) {
755 if (z_src_type
.length
== 4) {
756 zs_dst1
= lp_build_extract_range(gallivm
, z_value
, 0, 2);
757 zs_dst2
= lp_build_extract_range(gallivm
, z_value
, 2, 2);
760 assert(z_src_type
.length
== 8);
761 zs_dst1
= LLVMBuildShuffleVector(builder
, z_value
, z_value
,
762 LLVMConstVector(&shuffles
[0],
763 zs_load_type
.length
), "");
764 zs_dst2
= LLVMBuildShuffleVector(builder
, z_value
, z_value
,
765 LLVMConstVector(&shuffles
[4],
766 zs_load_type
.length
), "");
770 if (z_src_type
.length
== 4) {
771 zs_dst1
= lp_build_interleave2(gallivm
, z_type
,
772 z_value
, s_value
, 0);
773 zs_dst2
= lp_build_interleave2(gallivm
, z_type
,
774 z_value
, s_value
, 1);
778 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
/ 2];
779 assert(z_src_type
.length
== 8);
780 for (i
= 0; i
< 8; i
++) {
781 shuffles
[i
*2] = lp_build_const_int32(gallivm
, (i
&1) + (i
&2) * 2 + (i
&4) / 2);
782 shuffles
[i
*2+1] = lp_build_const_int32(gallivm
, (i
&1) + (i
&2) * 2 + (i
&4) / 2 +
785 zs_dst1
= LLVMBuildShuffleVector(builder
, z_value
, s_value
,
786 LLVMConstVector(&shuffles
[0],
787 z_src_type
.length
), "");
788 zs_dst2
= LLVMBuildShuffleVector(builder
, z_value
, s_value
,
789 LLVMConstVector(&shuffles
[8],
790 z_src_type
.length
), "");
792 zs_dst1
= LLVMBuildBitCast(builder
, zs_dst1
,
793 lp_build_vec_type(gallivm
, zs_load_type
), "");
794 zs_dst2
= LLVMBuildBitCast(builder
, zs_dst2
,
795 lp_build_vec_type(gallivm
, zs_load_type
), "");
798 LLVMBuildStore(builder
, zs_dst1
, zs_dst_ptr1
);
800 LLVMBuildStore(builder
, zs_dst2
, zs_dst_ptr2
);
805 * Generate code for performing depth and/or stencil tests.
806 * We operate on a vector of values (typically n 2x2 quads).
808 * \param depth the depth test state
809 * \param stencil the front/back stencil state
810 * \param type the data type of the fragment depth/stencil values
811 * \param format_desc description of the depth/stencil surface
812 * \param mask the alive/dead pixel mask for the quad (vector)
813 * \param cov_mask coverage mask
814 * \param stencil_refs the front/back stencil ref values (scalar)
815 * \param z_src the incoming depth/stencil values (n 2x2 quad values, float32)
816 * \param zs_dst the depth/stencil values in framebuffer
817 * \param face contains boolean value indicating front/back facing polygon
820 lp_build_depth_stencil_test(struct gallivm_state
*gallivm
,
821 const struct pipe_depth_state
*depth
,
822 const struct pipe_stencil_state stencil
[2],
823 struct lp_type z_src_type
,
824 const struct util_format_description
*format_desc
,
825 struct lp_build_mask_context
*mask
,
826 LLVMValueRef
*cov_mask
,
827 LLVMValueRef stencil_refs
[2],
832 LLVMValueRef
*z_value
,
833 LLVMValueRef
*s_value
,
836 LLVMBuilderRef builder
= gallivm
->builder
;
837 struct lp_type z_type
;
838 struct lp_build_context z_bld
;
839 struct lp_build_context s_bld
;
840 struct lp_type s_type
;
841 unsigned z_shift
= 0, z_width
= 0, z_mask
= 0;
842 LLVMValueRef z_dst
= NULL
;
843 LLVMValueRef stencil_vals
= NULL
;
844 LLVMValueRef z_bitmask
= NULL
, stencil_shift
= NULL
;
845 LLVMValueRef z_pass
= NULL
, s_pass_mask
= NULL
;
846 LLVMValueRef current_mask
= mask
? lp_build_mask_value(mask
) : *cov_mask
;
847 LLVMValueRef front_facing
= NULL
;
848 boolean have_z
, have_s
;
851 * Depths are expected to be between 0 and 1, even if they are stored in
852 * floats. Setting these bits here will ensure that the lp_build_conv() call
853 * below won't try to unnecessarily clamp the incoming values.
855 if(z_src_type
.floating
) {
856 z_src_type
.sign
= FALSE
;
857 z_src_type
.norm
= TRUE
;
860 assert(!z_src_type
.sign
);
861 assert(z_src_type
.norm
);
864 /* Pick the type matching the depth-stencil format. */
865 z_type
= lp_depth_type(format_desc
, z_src_type
.length
);
867 /* Pick the intermediate type for depth operations. */
868 z_type
.width
= z_src_type
.width
;
869 assert(z_type
.length
== z_src_type
.length
);
871 /* FIXME: for non-float depth/stencil might generate better code
872 * if we'd always split it up to use 128bit operations.
873 * For stencil we'd almost certainly want to pack to 8xi16 values,
874 * for z just run twice.
877 /* Sanity checking */
879 const unsigned z_swizzle
= format_desc
->swizzle
[0];
880 const unsigned s_swizzle
= format_desc
->swizzle
[1];
882 assert(z_swizzle
!= PIPE_SWIZZLE_NONE
||
883 s_swizzle
!= PIPE_SWIZZLE_NONE
);
885 assert(depth
->enabled
|| stencil
[0].enabled
);
887 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_ZS
);
888 assert(format_desc
->block
.width
== 1);
889 assert(format_desc
->block
.height
== 1);
891 if (stencil
[0].enabled
) {
892 assert(s_swizzle
< 4);
893 assert(format_desc
->channel
[s_swizzle
].type
== UTIL_FORMAT_TYPE_UNSIGNED
);
894 assert(format_desc
->channel
[s_swizzle
].pure_integer
);
895 assert(!format_desc
->channel
[s_swizzle
].normalized
);
896 assert(format_desc
->channel
[s_swizzle
].size
== 8);
899 if (depth
->enabled
) {
900 assert(z_swizzle
< 4);
901 if (z_type
.floating
) {
902 assert(z_swizzle
== 0);
903 assert(format_desc
->channel
[z_swizzle
].type
==
904 UTIL_FORMAT_TYPE_FLOAT
);
905 assert(format_desc
->channel
[z_swizzle
].size
== 32);
908 assert(format_desc
->channel
[z_swizzle
].type
==
909 UTIL_FORMAT_TYPE_UNSIGNED
);
910 assert(format_desc
->channel
[z_swizzle
].normalized
);
911 assert(!z_type
.fixed
);
917 /* Setup build context for Z vals */
918 lp_build_context_init(&z_bld
, gallivm
, z_type
);
920 /* Setup build context for stencil vals */
921 s_type
= lp_int_type(z_type
);
922 lp_build_context_init(&s_bld
, gallivm
, s_type
);
924 /* Compute and apply the Z/stencil bitmasks and shifts.
927 unsigned s_shift
, s_mask
;
932 have_z
= get_z_shift_and_mask(format_desc
, &z_shift
, &z_width
, &z_mask
);
933 have_s
= get_s_shift_and_mask(format_desc
, &s_shift
, &s_mask
);
936 if (z_mask
!= 0xffffffff) {
937 z_bitmask
= lp_build_const_int_vec(gallivm
, z_type
, z_mask
);
941 * Align the framebuffer Z 's LSB to the right.
944 LLVMValueRef shift
= lp_build_const_int_vec(gallivm
, z_type
, z_shift
);
945 z_dst
= LLVMBuildLShr(builder
, z_dst
, shift
, "z_dst");
946 } else if (z_bitmask
) {
947 z_dst
= LLVMBuildAnd(builder
, z_dst
, z_bitmask
, "z_dst");
949 lp_build_name(z_dst
, "z_dst");
955 LLVMValueRef shift
= lp_build_const_int_vec(gallivm
, s_type
, s_shift
);
956 stencil_vals
= LLVMBuildLShr(builder
, stencil_vals
, shift
, "");
957 stencil_shift
= shift
; /* used below */
960 if (s_mask
!= 0xffffffff) {
961 LLVMValueRef mask
= lp_build_const_int_vec(gallivm
, s_type
, s_mask
);
962 stencil_vals
= LLVMBuildAnd(builder
, stencil_vals
, mask
, "");
965 lp_build_name(stencil_vals
, "s_dst");
969 if (stencil
[0].enabled
) {
974 * XXX: the scalar expansion below produces atrocious code
975 * (basically producing a 64bit scalar value, then moving the 2
976 * 32bit pieces separately to simd, plus 4 shuffles, which is
977 * seriously lame). But the scalar-simd transitions are always
978 * tricky, so no big surprise there.
979 * This here would be way better, however llvm has some serious
980 * trouble later using it in the select, probably because it will
981 * recognize the expression as constant and move the simd value
982 * away (out of the loop) - and then it will suddenly try
983 * constructing i1 high-bit masks out of it later...
984 * (Try piglit stencil-twoside.)
985 * Note this is NOT due to using SExt/Trunc, it fails exactly the
986 * same even when using native compare/select.
987 * I cannot reproduce this problem when using stand-alone compiler
988 * though, suggesting some problem with optimization passes...
989 * (With stand-alone compilation, the construction of this mask
990 * value, no matter if the easy 3 instruction here or the complex
991 * 16+ one below, never gets separated from where it's used.)
992 * The scalar code still has the same problem, but the generated
993 * code looks a bit better at least for some reason, even if
994 * mostly by luck (the fundamental issue clearly is the same).
996 front_facing
= lp_build_broadcast(gallivm
, s_bld
.vec_type
, face
);
997 /* front_facing = face != 0 ? ~0 : 0 */
998 front_facing
= lp_build_compare(gallivm
, s_bld
.type
,
1000 front_facing
, s_bld
.zero
);
1002 LLVMValueRef zero
= lp_build_const_int32(gallivm
, 0);
1004 /* front_facing = face != 0 ? ~0 : 0 */
1005 front_facing
= LLVMBuildICmp(builder
, LLVMIntNE
, face
, zero
, "");
1006 front_facing
= LLVMBuildSExt(builder
, front_facing
,
1007 LLVMIntTypeInContext(gallivm
->context
,
1008 s_bld
.type
.length
*s_bld
.type
.width
),
1010 front_facing
= LLVMBuildBitCast(builder
, front_facing
,
1011 s_bld
.int_vec_type
, "");
1016 s_pass_mask
= lp_build_stencil_test(&s_bld
, stencil
,
1017 stencil_refs
, stencil_vals
,
1020 /* apply stencil-fail operator */
1022 LLVMValueRef s_fail_mask
= lp_build_andnot(&s_bld
, current_mask
, s_pass_mask
);
1023 stencil_vals
= lp_build_stencil_op(&s_bld
, stencil
, S_FAIL_OP
,
1024 stencil_refs
, stencil_vals
,
1025 s_fail_mask
, front_facing
);
1029 if (depth
->enabled
) {
1031 * Convert fragment Z to the desired type, aligning the LSB to the right.
1034 assert(z_type
.width
== z_src_type
.width
);
1035 assert(z_type
.length
== z_src_type
.length
);
1036 assert(lp_check_value(z_src_type
, z_src
));
1037 if (z_src_type
.floating
) {
1039 * Convert from floating point values
1042 if (!z_type
.floating
) {
1043 z_src
= lp_build_clamped_float_to_unsigned_norm(gallivm
,
1050 * Convert from unsigned normalized values.
1053 assert(!z_src_type
.sign
);
1054 assert(!z_src_type
.fixed
);
1055 assert(z_src_type
.norm
);
1056 assert(!z_type
.floating
);
1057 if (z_src_type
.width
> z_width
) {
1058 LLVMValueRef shift
= lp_build_const_int_vec(gallivm
, z_src_type
,
1059 z_src_type
.width
- z_width
);
1060 z_src
= LLVMBuildLShr(builder
, z_src
, shift
, "");
1063 assert(lp_check_value(z_type
, z_src
));
1065 lp_build_name(z_src
, "z_src");
1067 /* compare src Z to dst Z, returning 'pass' mask */
1068 z_pass
= lp_build_cmp(&z_bld
, depth
->func
, z_src
, z_dst
);
1070 /* mask off bits that failed stencil test */
1072 current_mask
= LLVMBuildAnd(builder
, current_mask
, s_pass_mask
, "");
1075 if (!stencil
[0].enabled
&& mask
) {
1076 /* We can potentially skip all remaining operations here, but only
1077 * if stencil is disabled because we still need to update the stencil
1078 * buffer values. Don't need to update Z buffer values.
1080 lp_build_mask_update(mask
, z_pass
);
1083 lp_build_mask_check(mask
);
1087 if (depth
->writemask
) {
1088 LLVMValueRef z_pass_mask
;
1090 /* mask off bits that failed Z test */
1091 z_pass_mask
= LLVMBuildAnd(builder
, current_mask
, z_pass
, "");
1093 /* Mix the old and new Z buffer values.
1094 * z_dst[i] = zselectmask[i] ? z_src[i] : z_dst[i]
1096 z_dst
= lp_build_select(&z_bld
, z_pass_mask
, z_src
, z_dst
);
1099 if (stencil
[0].enabled
) {
1100 /* update stencil buffer values according to z pass/fail result */
1101 LLVMValueRef z_fail_mask
, z_pass_mask
;
1103 /* apply Z-fail operator */
1104 z_fail_mask
= lp_build_andnot(&s_bld
, current_mask
, z_pass
);
1105 stencil_vals
= lp_build_stencil_op(&s_bld
, stencil
, Z_FAIL_OP
,
1106 stencil_refs
, stencil_vals
,
1107 z_fail_mask
, front_facing
);
1109 /* apply Z-pass operator */
1110 z_pass_mask
= LLVMBuildAnd(builder
, current_mask
, z_pass
, "");
1111 stencil_vals
= lp_build_stencil_op(&s_bld
, stencil
, Z_PASS_OP
,
1112 stencil_refs
, stencil_vals
,
1113 z_pass_mask
, front_facing
);
1117 /* No depth test: apply Z-pass operator to stencil buffer values which
1118 * passed the stencil test.
1120 s_pass_mask
= LLVMBuildAnd(builder
, current_mask
, s_pass_mask
, "");
1121 stencil_vals
= lp_build_stencil_op(&s_bld
, stencil
, Z_PASS_OP
,
1122 stencil_refs
, stencil_vals
,
1123 s_pass_mask
, front_facing
);
1126 /* Put Z and stencil bits in the right place */
1127 if (have_z
&& z_shift
) {
1128 LLVMValueRef shift
= lp_build_const_int_vec(gallivm
, z_type
, z_shift
);
1129 z_dst
= LLVMBuildShl(builder
, z_dst
, shift
, "");
1131 if (stencil_vals
&& stencil_shift
)
1132 stencil_vals
= LLVMBuildShl(builder
, stencil_vals
,
1135 /* Finally, merge the z/stencil values */
1136 if (format_desc
->block
.bits
<= 32) {
1137 if (have_z
&& have_s
)
1138 *z_value
= LLVMBuildOr(builder
, z_dst
, stencil_vals
, "");
1142 *z_value
= stencil_vals
;
1143 *s_value
= *z_value
;
1147 *s_value
= stencil_vals
;
1152 lp_build_mask_update(mask
, s_pass_mask
);
1154 if (depth
->enabled
&& stencil
[0].enabled
)
1155 lp_build_mask_update(mask
, z_pass
);
1157 LLVMValueRef tmp_mask
= *cov_mask
;
1159 tmp_mask
= LLVMBuildAnd(builder
, tmp_mask
, s_pass_mask
, "");
1161 /* for multisample we don't do the stencil optimisation so update always */
1163 tmp_mask
= LLVMBuildAnd(builder
, tmp_mask
, z_pass
, "");
1164 *cov_mask
= tmp_mask
;