1 /**************************************************************************
3 * Copyright 2009 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 * We are free to use a different pixel layout though. Since our basic
40 * processing unit is a quad (2x2 pixel block) we store the depth/stencil
41 * values tiled, a quad at time. That is, a depth buffer containing
49 * will actually be stored in memory as
51 * Z11 Z12 Z21 Z22 Z13 Z14 Z23 Z24 ...
52 * Z31 Z32 Z41 Z42 Z33 Z34 Z43 Z44 ...
53 * ... ... ... ... ... ... ... ... ...
57 * Two-sided stencil test is supported but probably not as efficient as
58 * it could be. Currently, we use if/then/else constructs to do the
59 * operations for front vs. back-facing polygons. We could probably do
60 * both the front and back arithmetic then use a Select() instruction to
61 * choose the result depending on polyon orientation. We'd have to
62 * measure performance both ways and see which is better.
64 * @author Jose Fonseca <jfonseca@vmware.com>
67 #include "pipe/p_state.h"
68 #include "util/u_format.h"
70 #include "lp_bld_type.h"
71 #include "lp_bld_arit.h"
72 #include "lp_bld_const.h"
73 #include "lp_bld_logic.h"
74 #include "lp_bld_flow.h"
75 #include "lp_bld_debug.h"
76 #include "lp_bld_depth.h"
77 #include "lp_bld_swizzle.h"
80 /** Used to select fields from pipe_stencil_state */
90 * Do the stencil test comparison (compare FB stencil values against ref value).
91 * This will be used twice when generating two-sided stencil code.
92 * \param stencil the front/back stencil state
93 * \param stencilRef the stencil reference value, replicated as a vector
94 * \param stencilVals vector of stencil values from framebuffer
95 * \return vector mask of pass/fail values (~0 or 0)
98 lp_build_stencil_test_single(struct lp_build_context
*bld
,
99 const struct pipe_stencil_state
*stencil
,
100 LLVMValueRef stencilRef
,
101 LLVMValueRef stencilVals
)
103 const unsigned stencilMax
= 255; /* XXX fix */
104 struct lp_type type
= bld
->type
;
109 assert(stencil
->enabled
);
111 if (stencil
->valuemask
!= stencilMax
) {
112 /* compute stencilRef = stencilRef & valuemask */
113 LLVMValueRef valuemask
= lp_build_const_int_vec(type
, stencil
->valuemask
);
114 stencilRef
= LLVMBuildAnd(bld
->builder
, stencilRef
, valuemask
, "");
115 /* compute stencilVals = stencilVals & valuemask */
116 stencilVals
= LLVMBuildAnd(bld
->builder
, stencilVals
, valuemask
, "");
119 res
= lp_build_cmp(bld
, stencil
->func
, stencilVals
, stencilRef
);
126 * Do the one or two-sided stencil test comparison.
127 * \sa lp_build_stencil_test_single
128 * \param face an integer indicating front (+) or back (-) facing polygon.
129 * If NULL, assume front-facing.
132 lp_build_stencil_test(struct lp_build_context
*bld
,
133 const struct pipe_stencil_state stencil
[2],
134 LLVMValueRef stencilRefs
[2],
135 LLVMValueRef stencilVals
,
140 assert(stencil
[0].enabled
);
142 if (stencil
[1].enabled
&& face
) {
143 /* do two-sided test */
144 struct lp_build_flow_context
*flow_ctx
;
145 struct lp_build_if_state if_ctx
;
146 LLVMValueRef front_facing
;
147 LLVMValueRef zero
= LLVMConstReal(LLVMFloatType(), 0.0);
148 LLVMValueRef result
= bld
->undef
;
150 flow_ctx
= lp_build_flow_create(bld
->builder
);
151 lp_build_flow_scope_begin(flow_ctx
);
153 lp_build_flow_scope_declare(flow_ctx
, &result
);
155 /* front_facing = face > 0.0 */
156 front_facing
= LLVMBuildFCmp(bld
->builder
, LLVMRealUGT
, face
, zero
, "");
158 lp_build_if(&if_ctx
, flow_ctx
, bld
->builder
, front_facing
);
160 result
= lp_build_stencil_test_single(bld
, &stencil
[0],
161 stencilRefs
[0], stencilVals
);
163 lp_build_else(&if_ctx
);
165 result
= lp_build_stencil_test_single(bld
, &stencil
[1],
166 stencilRefs
[1], stencilVals
);
168 lp_build_endif(&if_ctx
);
170 lp_build_flow_scope_end(flow_ctx
);
171 lp_build_flow_destroy(flow_ctx
);
176 /* do single-side test */
177 res
= lp_build_stencil_test_single(bld
, &stencil
[0],
178 stencilRefs
[0], stencilVals
);
186 * Apply the stencil operator (add/sub/keep/etc) to the given vector
188 * \return new stencil values vector
191 lp_build_stencil_op_single(struct lp_build_context
*bld
,
192 const struct pipe_stencil_state
*stencil
,
194 LLVMValueRef stencilRef
,
195 LLVMValueRef stencilVals
,
199 const unsigned stencilMax
= 255; /* XXX fix */
200 struct lp_type type
= bld
->type
;
202 LLVMValueRef max
= lp_build_const_int_vec(type
, stencilMax
);
209 stencil_op
= stencil
->fail_op
;
212 stencil_op
= stencil
->zfail_op
;
215 stencil_op
= stencil
->zpass_op
;
218 assert(0 && "Invalid stencil_op mode");
219 stencil_op
= PIPE_STENCIL_OP_KEEP
;
222 switch (stencil_op
) {
223 case PIPE_STENCIL_OP_KEEP
:
225 /* we can return early for this case */
227 case PIPE_STENCIL_OP_ZERO
:
230 case PIPE_STENCIL_OP_REPLACE
:
233 case PIPE_STENCIL_OP_INCR
:
234 res
= lp_build_add(bld
, stencilVals
, bld
->one
);
235 res
= lp_build_min(bld
, res
, max
);
237 case PIPE_STENCIL_OP_DECR
:
238 res
= lp_build_sub(bld
, stencilVals
, bld
->one
);
239 res
= lp_build_max(bld
, res
, bld
->zero
);
241 case PIPE_STENCIL_OP_INCR_WRAP
:
242 res
= lp_build_add(bld
, stencilVals
, bld
->one
);
243 res
= LLVMBuildAnd(bld
->builder
, res
, max
, "");
245 case PIPE_STENCIL_OP_DECR_WRAP
:
246 res
= lp_build_sub(bld
, stencilVals
, bld
->one
);
247 res
= LLVMBuildAnd(bld
->builder
, res
, max
, "");
249 case PIPE_STENCIL_OP_INVERT
:
250 res
= LLVMBuildNot(bld
->builder
, stencilVals
, "");
251 res
= LLVMBuildAnd(bld
->builder
, res
, max
, "");
254 assert(0 && "bad stencil op mode");
258 if (stencil
->writemask
!= stencilMax
) {
259 /* compute res = (res & mask) | (stencilVals & ~mask) */
260 LLVMValueRef mask
= lp_build_const_int_vec(type
, stencil
->writemask
);
261 LLVMValueRef cmask
= LLVMBuildNot(bld
->builder
, mask
, "notWritemask");
262 LLVMValueRef t1
= LLVMBuildAnd(bld
->builder
, res
, mask
, "t1");
263 LLVMValueRef t2
= LLVMBuildAnd(bld
->builder
, stencilVals
, cmask
, "t2");
264 res
= LLVMBuildOr(bld
->builder
, t1
, t2
, "t1_or_t2");
267 /* only the update the vector elements enabled by 'mask' */
268 res
= lp_build_select(bld
, mask
, res
, stencilVals
);
275 * Do the one or two-sided stencil test op/update.
278 lp_build_stencil_op(struct lp_build_context
*bld
,
279 const struct pipe_stencil_state stencil
[2],
281 LLVMValueRef stencilRefs
[2],
282 LLVMValueRef stencilVals
,
287 assert(stencil
[0].enabled
);
289 if (stencil
[1].enabled
&& face
) {
290 /* do two-sided op */
291 struct lp_build_flow_context
*flow_ctx
;
292 struct lp_build_if_state if_ctx
;
293 LLVMValueRef front_facing
;
294 LLVMValueRef zero
= LLVMConstReal(LLVMFloatType(), 0.0);
295 LLVMValueRef result
= bld
->undef
;
297 flow_ctx
= lp_build_flow_create(bld
->builder
);
298 lp_build_flow_scope_begin(flow_ctx
);
300 lp_build_flow_scope_declare(flow_ctx
, &result
);
302 /* front_facing = face > 0.0 */
303 front_facing
= LLVMBuildFCmp(bld
->builder
, LLVMRealUGT
, face
, zero
, "");
305 lp_build_if(&if_ctx
, flow_ctx
, bld
->builder
, front_facing
);
307 result
= lp_build_stencil_op_single(bld
, &stencil
[0], op
,
308 stencilRefs
[0], stencilVals
, mask
);
310 lp_build_else(&if_ctx
);
312 result
= lp_build_stencil_op_single(bld
, &stencil
[1], op
,
313 stencilRefs
[1], stencilVals
, mask
);
315 lp_build_endif(&if_ctx
);
317 lp_build_flow_scope_end(flow_ctx
);
318 lp_build_flow_destroy(flow_ctx
);
323 /* do single-sided op */
324 return lp_build_stencil_op_single(bld
, &stencil
[0], op
,
325 stencilRefs
[0], stencilVals
, mask
);
332 * Return a type appropriate for depth/stencil testing.
335 lp_depth_type(const struct util_format_description
*format_desc
,
341 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_ZS
);
342 assert(format_desc
->block
.width
== 1);
343 assert(format_desc
->block
.height
== 1);
345 swizzle
= format_desc
->swizzle
[0];
348 memset(&type
, 0, sizeof type
);
349 type
.width
= format_desc
->block
.bits
;
351 if(format_desc
->channel
[swizzle
].type
== UTIL_FORMAT_TYPE_FLOAT
) {
352 type
.floating
= TRUE
;
353 assert(swizzle
== 0);
354 assert(format_desc
->channel
[swizzle
].size
== format_desc
->block
.bits
);
356 else if(format_desc
->channel
[swizzle
].type
== UTIL_FORMAT_TYPE_UNSIGNED
) {
357 assert(format_desc
->block
.bits
<= 32);
358 if(format_desc
->channel
[swizzle
].normalized
)
364 assert(type
.width
<= length
);
365 type
.length
= length
/ type
.width
;
372 * Compute bitmask and bit shift to apply to the incoming fragment Z values
373 * and the Z buffer values needed before doing the Z comparison.
375 * Note that we leave the Z bits in the position that we find them
376 * in the Z buffer (typically 0xffffff00 or 0x00ffffff). That lets us
377 * get by with fewer bit twiddling steps.
380 get_z_shift_and_mask(const struct util_format_description
*format_desc
,
381 unsigned *shift
, unsigned *mask
)
383 const unsigned total_bits
= format_desc
->block
.bits
;
386 unsigned padding_left
, padding_right
;
388 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_ZS
);
389 assert(format_desc
->block
.width
== 1);
390 assert(format_desc
->block
.height
== 1);
392 z_swizzle
= format_desc
->swizzle
[0];
394 if (z_swizzle
== UTIL_FORMAT_SWIZZLE_NONE
)
398 for (chan
= 0; chan
< z_swizzle
; ++chan
)
399 padding_right
+= format_desc
->channel
[chan
].size
;
402 total_bits
- (padding_right
+ format_desc
->channel
[z_swizzle
].size
);
404 if (padding_left
|| padding_right
) {
405 unsigned long long mask_left
= (1ULL << (total_bits
- padding_left
)) - 1;
406 unsigned long long mask_right
= (1ULL << (padding_right
)) - 1;
407 *mask
= mask_left
^ mask_right
;
413 *shift
= padding_left
;
420 * Compute bitmask and bit shift to apply to the framebuffer pixel values
421 * to put the stencil bits in the least significant position.
425 get_s_shift_and_mask(const struct util_format_description
*format_desc
,
426 unsigned *shift
, unsigned *mask
)
431 s_swizzle
= format_desc
->swizzle
[1];
433 if (s_swizzle
== UTIL_FORMAT_SWIZZLE_NONE
)
437 for (chan
= 0; chan
< s_swizzle
; chan
++)
438 *shift
+= format_desc
->channel
[chan
].size
;
440 sz
= format_desc
->channel
[s_swizzle
].size
;
441 *mask
= (1U << sz
) - 1U;
449 * Generate code for performing depth and/or stencil tests.
450 * We operate on a vector of values (typically a 2x2 quad).
452 * \param depth the depth test state
453 * \param stencil the front/back stencil state
454 * \param type the data type of the fragment depth/stencil values
455 * \param format_desc description of the depth/stencil surface
456 * \param mask the alive/dead pixel mask for the quad (vector)
457 * \param stencil_refs the front/back stencil ref values (scalar)
458 * \param z_src the incoming depth/stencil values (a 2x2 quad)
459 * \param zs_dst_ptr pointer to depth/stencil values in framebuffer
460 * \param facing contains float value indicating front/back facing polygon
463 lp_build_depth_stencil_test(LLVMBuilderRef builder
,
464 const struct pipe_depth_state
*depth
,
465 const struct pipe_stencil_state stencil
[2],
467 const struct util_format_description
*format_desc
,
468 struct lp_build_mask_context
*mask
,
469 LLVMValueRef stencil_refs
[2],
471 LLVMValueRef zs_dst_ptr
,
474 struct lp_build_context bld
;
475 struct lp_build_context sbld
;
476 struct lp_type s_type
;
477 LLVMValueRef zs_dst
, z_dst
= NULL
;
478 LLVMValueRef stencil_vals
= NULL
;
479 LLVMValueRef z_bitmask
= NULL
, stencil_shift
= NULL
;
480 LLVMValueRef z_pass
= NULL
, s_pass_mask
= NULL
;
481 LLVMValueRef orig_mask
= mask
->value
;
483 /* Sanity checking */
485 const unsigned z_swizzle
= format_desc
->swizzle
[0];
486 const unsigned s_swizzle
= format_desc
->swizzle
[1];
488 assert(z_swizzle
!= UTIL_FORMAT_SWIZZLE_NONE
||
489 s_swizzle
!= UTIL_FORMAT_SWIZZLE_NONE
);
491 assert(depth
->enabled
|| stencil
[0].enabled
);
493 assert(format_desc
->colorspace
== UTIL_FORMAT_COLORSPACE_ZS
);
494 assert(format_desc
->block
.width
== 1);
495 assert(format_desc
->block
.height
== 1);
497 if (stencil
[0].enabled
) {
498 assert(format_desc
->format
== PIPE_FORMAT_Z24_UNORM_S8_USCALED
||
499 format_desc
->format
== PIPE_FORMAT_S8_USCALED_Z24_UNORM
);
502 assert(z_swizzle
< 4);
503 assert(format_desc
->block
.bits
== type
.width
);
505 assert(z_swizzle
== 0);
506 assert(format_desc
->channel
[z_swizzle
].type
==
507 UTIL_FORMAT_TYPE_FLOAT
);
508 assert(format_desc
->channel
[z_swizzle
].size
==
509 format_desc
->block
.bits
);
512 assert(format_desc
->channel
[z_swizzle
].type
==
513 UTIL_FORMAT_TYPE_UNSIGNED
);
514 assert(format_desc
->channel
[z_swizzle
].normalized
);
522 /* Setup build context for Z vals */
523 lp_build_context_init(&bld
, builder
, type
);
525 /* Setup build context for stencil vals */
526 s_type
= lp_type_int_vec(type
.width
);
527 lp_build_context_init(&sbld
, builder
, s_type
);
529 /* Load current z/stencil value from z/stencil buffer */
530 zs_dst
= LLVMBuildLoad(builder
, zs_dst_ptr
, "");
532 lp_build_name(zs_dst
, "zsbufval");
535 /* Compute and apply the Z/stencil bitmasks and shifts.
538 unsigned z_shift
, z_mask
;
539 unsigned s_shift
, s_mask
;
541 if (get_z_shift_and_mask(format_desc
, &z_shift
, &z_mask
)) {
543 LLVMValueRef shift
= lp_build_const_int_vec(type
, z_shift
);
544 z_src
= LLVMBuildLShr(builder
, z_src
, shift
, "");
547 if (z_mask
!= 0xffffffff) {
548 LLVMValueRef mask
= lp_build_const_int_vec(type
, z_mask
);
549 z_src
= LLVMBuildAnd(builder
, z_src
, mask
, "");
550 z_dst
= LLVMBuildAnd(builder
, zs_dst
, mask
, "");
551 z_bitmask
= mask
; /* used below */
557 lp_build_name(z_dst
, "zsbuf.z");
560 if (get_s_shift_and_mask(format_desc
, &s_shift
, &s_mask
)) {
562 LLVMValueRef shift
= lp_build_const_int_vec(type
, s_shift
);
563 stencil_vals
= LLVMBuildLShr(builder
, zs_dst
, shift
, "");
564 stencil_shift
= shift
; /* used below */
567 stencil_vals
= zs_dst
;
570 if (s_mask
!= 0xffffffff) {
571 LLVMValueRef mask
= lp_build_const_int_vec(type
, s_mask
);
572 stencil_vals
= LLVMBuildAnd(builder
, stencil_vals
, mask
, "");
575 lp_build_name(stencil_vals
, "stencil");
580 if (stencil
[0].enabled
) {
581 /* convert scalar stencil refs into vectors */
582 stencil_refs
[0] = lp_build_broadcast_scalar(&bld
, stencil_refs
[0]);
583 stencil_refs
[1] = lp_build_broadcast_scalar(&bld
, stencil_refs
[1]);
585 s_pass_mask
= lp_build_stencil_test(&sbld
, stencil
,
586 stencil_refs
, stencil_vals
, face
);
588 /* apply stencil-fail operator */
590 LLVMValueRef s_fail_mask
= lp_build_andc(&bld
, orig_mask
, s_pass_mask
);
591 stencil_vals
= lp_build_stencil_op(&sbld
, stencil
, S_FAIL_OP
,
592 stencil_refs
, stencil_vals
,
597 if (depth
->enabled
) {
598 /* compare src Z to dst Z, returning 'pass' mask */
599 z_pass
= lp_build_cmp(&bld
, depth
->func
, z_src
, z_dst
);
601 if (!stencil
[0].enabled
) {
602 /* We can potentially skip all remaining operations here, but only
603 * if stencil is disabled because we still need to update the stencil
604 * buffer values. Don't need to update Z buffer values.
606 lp_build_mask_update(mask
, z_pass
);
609 if (depth
->writemask
) {
611 z_bitmask
= LLVMBuildAnd(builder
, mask
->value
, z_bitmask
, "");
613 z_bitmask
= mask
->value
;
615 z_dst
= lp_build_select(&bld
, z_bitmask
, z_src
, z_dst
);
618 if (stencil
[0].enabled
) {
619 /* update stencil buffer values according to z pass/fail result */
620 LLVMValueRef z_fail_mask
, z_pass_mask
;
622 /* apply Z-fail operator */
623 z_fail_mask
= lp_build_andc(&bld
, orig_mask
, z_pass
);
624 stencil_vals
= lp_build_stencil_op(&sbld
, stencil
, Z_FAIL_OP
,
625 stencil_refs
, stencil_vals
,
628 /* apply Z-pass operator */
629 z_pass_mask
= LLVMBuildAnd(bld
.builder
, orig_mask
, z_pass
, "");
630 stencil_vals
= lp_build_stencil_op(&sbld
, stencil
, Z_PASS_OP
,
631 stencil_refs
, stencil_vals
,
636 /* No depth test: apply Z-pass operator to stencil buffer values which
637 * passed the stencil test.
639 s_pass_mask
= LLVMBuildAnd(bld
.builder
, orig_mask
, s_pass_mask
, "");
640 stencil_vals
= lp_build_stencil_op(&sbld
, stencil
, Z_PASS_OP
,
641 stencil_refs
, stencil_vals
,
645 /* The Z bits are already in the right place but we may need to shift the
646 * stencil bits before ORing Z with Stencil to make the final pixel value.
648 if (stencil_vals
&& stencil_shift
)
649 stencil_vals
= LLVMBuildShl(bld
.builder
, stencil_vals
,
652 /* Finally, merge/store the z/stencil values */
653 if ((depth
->enabled
&& depth
->writemask
) ||
654 (stencil
[0].enabled
&& stencil
[0].writemask
)) {
656 if (z_dst
&& stencil_vals
)
657 zs_dst
= LLVMBuildOr(bld
.builder
, z_dst
, stencil_vals
, "");
661 zs_dst
= stencil_vals
;
663 LLVMBuildStore(builder
, zs_dst
, zs_dst_ptr
);
667 lp_build_mask_update(mask
, s_pass_mask
);
669 if (depth
->enabled
&& stencil
[0].enabled
)
670 lp_build_mask_update(mask
, z_pass
);