1 /**************************************************************************
3 * Copyright 2010 VMware.
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 **************************************************************************/
29 #include "util/u_math.h"
30 #include "util/u_memory.h"
31 #include "util/u_simple_list.h"
32 #include "os/os_time.h"
33 #include "gallivm/lp_bld_arit.h"
34 #include "gallivm/lp_bld_bitarit.h"
35 #include "gallivm/lp_bld_const.h"
36 #include "gallivm/lp_bld_debug.h"
37 #include "gallivm/lp_bld_init.h"
38 #include "gallivm/lp_bld_logic.h"
39 #include "gallivm/lp_bld_intr.h"
40 #include "gallivm/lp_bld_flow.h"
41 #include "gallivm/lp_bld_type.h"
46 #include "lp_screen.h"
47 #include "lp_context.h"
49 #include "lp_state_fs.h"
50 #include "lp_state_setup.h"
54 /* currently organized to interpolate full float[4] attributes even
55 * when some elements are unused. Later, can pack vertex data more
62 /* Function arguments:
67 LLVMValueRef facing
; /* boolean */
74 LLVMValueRef x0_center
;
75 LLVMValueRef y0_center
;
76 LLVMValueRef dy20_ooa
;
77 LLVMValueRef dy01_ooa
;
78 LLVMValueRef dx20_ooa
;
79 LLVMValueRef dx01_ooa
;
80 struct lp_build_context bld
;
85 store_coef(struct gallivm_state
*gallivm
,
86 struct lp_setup_args
*args
,
92 LLVMBuilderRef builder
= gallivm
->builder
;
93 LLVMValueRef idx
= lp_build_const_int32(gallivm
, slot
);
95 LLVMBuildStore(builder
,
97 LLVMBuildGEP(builder
, args
->a0
, &idx
, 1, ""));
99 LLVMBuildStore(builder
,
101 LLVMBuildGEP(builder
, args
->dadx
, &idx
, 1, ""));
103 LLVMBuildStore(builder
,
105 LLVMBuildGEP(builder
, args
->dady
, &idx
, 1, ""));
111 emit_constant_coef4(struct gallivm_state
*gallivm
,
112 struct lp_setup_args
*args
,
116 store_coef(gallivm
, args
, slot
, vert
, args
->bld
.zero
, args
->bld
.zero
);
122 * Setup the fragment input attribute with the front-facing value.
123 * \param frontface is the triangle front facing?
126 emit_facing_coef(struct gallivm_state
*gallivm
,
127 struct lp_setup_args
*args
,
130 LLVMBuilderRef builder
= gallivm
->builder
;
131 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
132 LLVMValueRef a0_0
= args
->facing
;
133 LLVMValueRef a0_0f
= LLVMBuildSIToFP(builder
, a0_0
, float_type
, "");
134 const unsigned char swizzles
[4] = { PIPE_SWIZZLE_RED
, PIPE_SWIZZLE_ZERO
,
135 PIPE_SWIZZLE_ZERO
, PIPE_SWIZZLE_ZERO
};
136 /* Our face val is either 1 or 0 so we do
137 * face = (val * 2) - 1
140 LLVMValueRef face_val
=
141 LLVMBuildFAdd(builder
,
142 LLVMBuildFMul(builder
, a0_0f
,
143 lp_build_const_float(gallivm
, 2.0),
145 lp_build_const_float(gallivm
, -1.0),
147 LLVMValueRef a0
= lp_build_swizzle_aos(&args
->bld
, face_val
, swizzles
);
149 store_coef(gallivm
, args
, slot
, a0
, args
->bld
.zero
, args
->bld
.zero
);
154 vert_attrib(struct gallivm_state
*gallivm
,
160 LLVMBuilderRef b
= gallivm
->builder
;
162 idx
[0] = lp_build_const_int32(gallivm
, attr
);
163 idx
[1] = lp_build_const_int32(gallivm
, elem
);
164 return LLVMBuildLoad(b
, LLVMBuildGEP(b
, vert
, idx
, 2, ""), name
);
169 lp_twoside(struct gallivm_state
*gallivm
,
170 struct lp_setup_args
*args
,
171 const struct lp_setup_variant_key
*key
,
173 LLVMValueRef attribv
[3])
175 LLVMBuilderRef b
= gallivm
->builder
;
176 LLVMValueRef a0_back
, a1_back
, a2_back
;
177 LLVMValueRef idx2
= lp_build_const_int32(gallivm
, bcolor_slot
);
179 LLVMValueRef facing
= args
->facing
;
180 LLVMValueRef front_facing
= LLVMBuildICmp(b
, LLVMIntEQ
, facing
,
181 lp_build_const_int32(gallivm
, 0), ""); /** need i1 for if condition */
183 a0_back
= LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v0
, &idx2
, 1, ""), "v0a_back");
184 a1_back
= LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v1
, &idx2
, 1, ""), "v1a_back");
185 a2_back
= LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v2
, &idx2
, 1, ""), "v2a_back");
187 /* Possibly swap the front and back attrib values,
189 * Prefer select to if so we don't have to worry about phis or
192 attribv
[0] = LLVMBuildSelect(b
, front_facing
, a0_back
, attribv
[0], "");
193 attribv
[1] = LLVMBuildSelect(b
, front_facing
, a1_back
, attribv
[1], "");
194 attribv
[2] = LLVMBuildSelect(b
, front_facing
, a2_back
, attribv
[2], "");
199 lp_do_offset_tri(struct gallivm_state
*gallivm
,
200 struct lp_setup_args
*args
,
201 const struct lp_setup_variant_key
*key
,
202 LLVMValueRef inv_det
,
205 LLVMValueRef attribv
[3])
207 LLVMBuilderRef b
= gallivm
->builder
;
208 struct lp_build_context flt_scalar_bld
;
209 struct lp_build_context int_scalar_bld
;
210 struct lp_build_context
*bld
= &args
->bld
;
211 LLVMValueRef zoffset
, mult
;
212 LLVMValueRef z0_new
, z1_new
, z2_new
;
213 LLVMValueRef dzdxdzdy
, dzdx
, dzdy
, dzxyz20
, dyzzx01
, dyzzx01_dzxyz20
, dzx01_dyz20
;
214 LLVMValueRef z0z1
, z0z1z2
;
215 LLVMValueRef max
, max_value
, res12
;
216 LLVMValueRef shuffles
[4];
217 LLVMTypeRef shuf_type
= LLVMInt32TypeInContext(gallivm
->context
);
218 LLVMValueRef onei
= lp_build_const_int32(gallivm
, 1);
219 LLVMValueRef zeroi
= lp_build_const_int32(gallivm
, 0);
220 LLVMValueRef twoi
= lp_build_const_int32(gallivm
, 2);
221 LLVMValueRef threei
= lp_build_const_int32(gallivm
, 3);
223 /* (res12) = cross(e,f).xy */
228 dzxyz20
= LLVMBuildShuffleVector(b
, dxyz20
, dxyz20
, LLVMConstVector(shuffles
, 4), "");
234 dyzzx01
= LLVMBuildShuffleVector(b
, dxyz01
, dxyz01
, LLVMConstVector(shuffles
, 4), "");
236 dyzzx01_dzxyz20
= LLVMBuildFMul(b
, dzxyz20
, dyzzx01
, "dyzzx01_dzxyz20");
239 shuffles
[1] = threei
;
240 shuffles
[2] = LLVMGetUndef(shuf_type
);
241 shuffles
[3] = LLVMGetUndef(shuf_type
);
242 dzx01_dyz20
= LLVMBuildShuffleVector(b
, dyzzx01_dzxyz20
, dyzzx01_dzxyz20
,
243 LLVMConstVector(shuffles
, 4), "");
245 res12
= LLVMBuildFSub(b
, dyzzx01_dzxyz20
, dzx01_dyz20
, "res12");
247 /* dzdx = fabsf(res1 * inv_det), dydx = fabsf(res2 * inv_det)*/
248 dzdxdzdy
= LLVMBuildFMul(b
, res12
, inv_det
, "dzdxdzdy");
249 dzdxdzdy
= lp_build_abs(bld
, dzdxdzdy
);
251 dzdx
= LLVMBuildExtractElement(b
, dzdxdzdy
, zeroi
, "");
252 dzdy
= LLVMBuildExtractElement(b
, dzdxdzdy
, onei
, "");
254 /* mult = MAX2(dzdx, dzdy) * pgon_offset_scale */
255 max
= LLVMBuildFCmp(b
, LLVMRealUGT
, dzdx
, dzdy
, "");
256 max_value
= LLVMBuildSelect(b
, max
, dzdx
, dzdy
, "max");
258 mult
= LLVMBuildFMul(b
, max_value
,
259 lp_build_const_float(gallivm
, key
->pgon_offset_scale
), "");
261 lp_build_context_init(&flt_scalar_bld
, gallivm
, lp_type_float_vec(32, 32));
263 if (key
->floating_point_depth
) {
265 * bias = pgon_offset_units * 2^(exponent(max(z0, z1, z2)) - mantissa_bits) +
266 * MAX2(dzdx, dzdy) * pgon_offset_scale
268 * NOTE: Assumes IEEE float32.
270 LLVMValueRef c23_shifted
, exp_mask
, bias
, exp
;
271 LLVMValueRef maxz_value
, maxz0z1_value
;
273 lp_build_context_init(&int_scalar_bld
, gallivm
, lp_type_int_vec(32, 32));
275 c23_shifted
= lp_build_const_int32(gallivm
, 23 << 23);
276 exp_mask
= lp_build_const_int32(gallivm
, 0xff << 23);
278 maxz0z1_value
= lp_build_max(&flt_scalar_bld
,
279 LLVMBuildExtractElement(b
, attribv
[0], twoi
, ""),
280 LLVMBuildExtractElement(b
, attribv
[1], twoi
, ""));
282 maxz_value
= lp_build_max(&flt_scalar_bld
,
283 LLVMBuildExtractElement(b
, attribv
[2], twoi
, ""),
286 exp
= LLVMBuildBitCast(b
, maxz_value
, int_scalar_bld
.vec_type
, "");
287 exp
= lp_build_and(&int_scalar_bld
, exp
, exp_mask
);
288 exp
= lp_build_sub(&int_scalar_bld
, exp
, c23_shifted
);
289 /* Clamping to zero means mrd will be zero for very small numbers,
290 * but specs do not indicate this should be prevented by clamping
291 * mrd to smallest normal number instead. */
292 exp
= lp_build_max(&int_scalar_bld
, exp
, int_scalar_bld
.zero
);
293 exp
= LLVMBuildBitCast(b
, exp
, flt_scalar_bld
.vec_type
, "");
295 bias
= LLVMBuildFMul(b
, exp
,
296 lp_build_const_float(gallivm
, key
->pgon_offset_units
),
299 zoffset
= LLVMBuildFAdd(b
, bias
, mult
, "zoffset");
302 * bias = pgon_offset_units + MAX2(dzdx, dzdy) * pgon_offset_scale
304 zoffset
= LLVMBuildFAdd(b
,
305 lp_build_const_float(gallivm
, key
->pgon_offset_units
),
309 if (key
->pgon_offset_clamp
> 0) {
310 zoffset
= lp_build_min(&flt_scalar_bld
,
311 lp_build_const_float(gallivm
, key
->pgon_offset_clamp
),
314 else if (key
->pgon_offset_clamp
< 0) {
315 zoffset
= lp_build_max(&flt_scalar_bld
,
316 lp_build_const_float(gallivm
, key
->pgon_offset_clamp
),
322 shuffles
[1] = lp_build_const_int32(gallivm
, 6);
323 shuffles
[2] = LLVMGetUndef(shuf_type
);
324 shuffles
[3] = LLVMGetUndef(shuf_type
);
325 z0z1
= LLVMBuildShuffleVector(b
, attribv
[0], attribv
[1], LLVMConstVector(shuffles
, 4), "");
328 shuffles
[2] = lp_build_const_int32(gallivm
, 6);
329 shuffles
[3] = LLVMGetUndef(shuf_type
);
330 z0z1z2
= LLVMBuildShuffleVector(b
, z0z1
, attribv
[2], LLVMConstVector(shuffles
, 4), "");
331 zoffset
= lp_build_broadcast_scalar(bld
, zoffset
);
333 /* clamp and do offset */
335 * FIXME I suspect the clamp (is that even right to always clamp to fixed
336 * 0.0/1.0?) should really be per fragment?
338 z0z1z2
= lp_build_clamp(bld
, LLVMBuildFAdd(b
, z0z1z2
, zoffset
, ""), bld
->zero
, bld
->one
);
340 /* insert into args->a0.z, a1.z, a2.z:
342 z0_new
= LLVMBuildExtractElement(b
, z0z1z2
, zeroi
, "");
343 z1_new
= LLVMBuildExtractElement(b
, z0z1z2
, onei
, "");
344 z2_new
= LLVMBuildExtractElement(b
, z0z1z2
, twoi
, "");
345 attribv
[0] = LLVMBuildInsertElement(b
, attribv
[0], z0_new
, twoi
, "");
346 attribv
[1] = LLVMBuildInsertElement(b
, attribv
[1], z1_new
, twoi
, "");
347 attribv
[2] = LLVMBuildInsertElement(b
, attribv
[2], z2_new
, twoi
, "");
351 load_attribute(struct gallivm_state
*gallivm
,
352 struct lp_setup_args
*args
,
353 const struct lp_setup_variant_key
*key
,
355 LLVMValueRef attribv
[3])
357 LLVMBuilderRef b
= gallivm
->builder
;
358 LLVMValueRef idx
= lp_build_const_int32(gallivm
, vert_attr
);
360 /* Load the vertex data
362 attribv
[0] = LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v0
, &idx
, 1, ""), "v0a");
363 attribv
[1] = LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v1
, &idx
, 1, ""), "v1a");
364 attribv
[2] = LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v2
, &idx
, 1, ""), "v2a");
367 /* Potentially modify it according to twoside, etc:
370 if (vert_attr
== key
->color_slot
&& key
->bcolor_slot
>= 0)
371 lp_twoside(gallivm
, args
, key
, key
->bcolor_slot
, attribv
);
372 else if (vert_attr
== key
->spec_slot
&& key
->bspec_slot
>= 0)
373 lp_twoside(gallivm
, args
, key
, key
->bspec_slot
, attribv
);
378 emit_coef4( struct gallivm_state
*gallivm
,
379 struct lp_setup_args
*args
,
385 LLVMBuilderRef b
= gallivm
->builder
;
386 LLVMValueRef dy20_ooa
= args
->dy20_ooa
;
387 LLVMValueRef dy01_ooa
= args
->dy01_ooa
;
388 LLVMValueRef dx20_ooa
= args
->dx20_ooa
;
389 LLVMValueRef dx01_ooa
= args
->dx01_ooa
;
390 LLVMValueRef x0_center
= args
->x0_center
;
391 LLVMValueRef y0_center
= args
->y0_center
;
392 LLVMValueRef da01
= LLVMBuildFSub(b
, a0
, a1
, "da01");
393 LLVMValueRef da20
= LLVMBuildFSub(b
, a2
, a0
, "da20");
395 /* Calculate dadx (vec4f)
397 LLVMValueRef da01_dy20_ooa
= LLVMBuildFMul(b
, da01
, dy20_ooa
, "da01_dy20_ooa");
398 LLVMValueRef da20_dy01_ooa
= LLVMBuildFMul(b
, da20
, dy01_ooa
, "da20_dy01_ooa");
399 LLVMValueRef dadx
= LLVMBuildFSub(b
, da01_dy20_ooa
, da20_dy01_ooa
, "dadx");
401 /* Calculate dady (vec4f)
403 LLVMValueRef da01_dx20_ooa
= LLVMBuildFMul(b
, da01
, dx20_ooa
, "da01_dx20_ooa");
404 LLVMValueRef da20_dx01_ooa
= LLVMBuildFMul(b
, da20
, dx01_ooa
, "da20_dx01_ooa");
405 LLVMValueRef dady
= LLVMBuildFSub(b
, da20_dx01_ooa
, da01_dx20_ooa
, "dady");
407 /* Calculate a0 - the attribute value at the origin
409 LLVMValueRef dadx_x0
= LLVMBuildFMul(b
, dadx
, x0_center
, "dadx_x0");
410 LLVMValueRef dady_y0
= LLVMBuildFMul(b
, dady
, y0_center
, "dady_y0");
411 LLVMValueRef attr_v0
= LLVMBuildFAdd(b
, dadx_x0
, dady_y0
, "attr_v0");
412 LLVMValueRef attr_0
= LLVMBuildFSub(b
, a0
, attr_v0
, "attr_0");
414 store_coef(gallivm
, args
, slot
, attr_0
, dadx
, dady
);
419 emit_linear_coef( struct gallivm_state
*gallivm
,
420 struct lp_setup_args
*args
,
422 LLVMValueRef attribv
[3])
424 /* nothing to do anymore */
434 * Compute a0, dadx and dady for a perspective-corrected interpolant,
436 * We basically multiply the vertex value by 1/w before computing
437 * the plane coefficients (a0, dadx, dady).
438 * Later, when we compute the value at a particular fragment position we'll
439 * divide the interpolated value by the interpolated W at that fragment.
442 apply_perspective_corr( struct gallivm_state
*gallivm
,
443 struct lp_setup_args
*args
,
445 LLVMValueRef attribv
[3])
447 LLVMBuilderRef b
= gallivm
->builder
;
449 /* premultiply by 1/w (v[0][3] is always 1/w):
451 LLVMValueRef v0_oow
= lp_build_broadcast_scalar(&args
->bld
,
452 vert_attrib(gallivm
, args
->v0
, 0, 3, "v0_oow"));
453 LLVMValueRef v1_oow
= lp_build_broadcast_scalar(&args
->bld
,
454 vert_attrib(gallivm
, args
->v1
, 0, 3, "v1_oow"));
455 LLVMValueRef v2_oow
= lp_build_broadcast_scalar(&args
->bld
,
456 vert_attrib(gallivm
, args
->v2
, 0, 3, "v2_oow"));
458 attribv
[0] = LLVMBuildFMul(b
, attribv
[0], v0_oow
, "v0_oow_v0a");
459 attribv
[1] = LLVMBuildFMul(b
, attribv
[1], v1_oow
, "v1_oow_v1a");
460 attribv
[2] = LLVMBuildFMul(b
, attribv
[2], v2_oow
, "v2_oow_v2a");
465 * Applys cylindrical wrapping to vertex attributes if enabled.
466 * Input coordinates must be in [0, 1] range, otherwise results are undefined.
468 * @param cyl_wrap TGSI_CYLINDRICAL_WRAP_x flags
471 emit_apply_cyl_wrap(struct gallivm_state
*gallivm
,
472 struct lp_setup_args
*args
,
474 LLVMValueRef attribv
[3])
477 LLVMBuilderRef builder
= gallivm
->builder
;
478 struct lp_type type
= args
->bld
.type
;
479 LLVMTypeRef float_vec_type
= args
->bld
.vec_type
;
480 LLVMValueRef pos_half
;
481 LLVMValueRef neg_half
;
482 LLVMValueRef cyl_mask
;
491 pos_half
= lp_build_const_vec(gallivm
, type
, +0.5f
);
492 neg_half
= lp_build_const_vec(gallivm
, type
, -0.5f
);
493 cyl_mask
= lp_build_const_mask_aos(gallivm
, type
, cyl_wrap
, 4);
495 one
= lp_build_const_vec(gallivm
, type
, 1.0f
);
496 one
= LLVMBuildBitCast(builder
, one
, lp_build_int_vec_type(gallivm
, type
), "");
497 one
= LLVMBuildAnd(builder
, one
, cyl_mask
, "");
500 delta
= LLVMBuildFSub(builder
, attribv
[1], attribv
[0], "");
502 offset
= lp_build_compare(gallivm
, type
, PIPE_FUNC_GREATER
, delta
, pos_half
);
503 offset
= LLVMBuildAnd(builder
, offset
, one
, "");
504 offset
= LLVMBuildBitCast(builder
, offset
, float_vec_type
, "");
505 attribv
[0] = LLVMBuildFAdd(builder
, attribv
[0], offset
, "");
507 offset
= lp_build_compare(gallivm
, type
, PIPE_FUNC_LESS
, delta
, neg_half
);
508 offset
= LLVMBuildAnd(builder
, offset
, one
, "");
509 offset
= LLVMBuildBitCast(builder
, offset
, float_vec_type
, "");
510 attribv
[1] = LLVMBuildFAdd(builder
, attribv
[1], offset
, "");
513 delta
= LLVMBuildFSub(builder
, attribv
[2], attribv
[1], "");
515 offset
= lp_build_compare(gallivm
, type
, PIPE_FUNC_GREATER
, delta
, pos_half
);
516 offset
= LLVMBuildAnd(builder
, offset
, one
, "");
517 offset
= LLVMBuildBitCast(builder
, offset
, float_vec_type
, "");
518 attribv
[1] = LLVMBuildFAdd(builder
, attribv
[1], offset
, "");
520 offset
= lp_build_compare(gallivm
, type
, PIPE_FUNC_LESS
, delta
, neg_half
);
521 offset
= LLVMBuildAnd(builder
, offset
, one
, "");
522 offset
= LLVMBuildBitCast(builder
, offset
, float_vec_type
, "");
523 attribv
[2] = LLVMBuildFAdd(builder
, attribv
[2], offset
, "");
526 delta
= LLVMBuildFSub(builder
, attribv
[0], attribv
[2], "");
528 offset
= lp_build_compare(gallivm
, type
, PIPE_FUNC_GREATER
, delta
, pos_half
);
529 offset
= LLVMBuildAnd(builder
, offset
, one
, "");
530 offset
= LLVMBuildBitCast(builder
, offset
, float_vec_type
, "");
531 attribv
[2] = LLVMBuildFAdd(builder
, attribv
[2], offset
, "");
533 offset
= lp_build_compare(gallivm
, type
, PIPE_FUNC_LESS
, delta
, neg_half
);
534 offset
= LLVMBuildAnd(builder
, offset
, one
, "");
535 offset
= LLVMBuildBitCast(builder
, offset
, float_vec_type
, "");
536 attribv
[0] = LLVMBuildFAdd(builder
, attribv
[0], offset
, "");
541 * Compute the inputs-> dadx, dady, a0 values.
544 emit_tri_coef( struct gallivm_state
*gallivm
,
545 const struct lp_setup_variant_key
*key
,
546 struct lp_setup_args
*args
)
550 LLVMValueRef attribs
[3];
552 /* setup interpolation for all the remaining attributes:
554 for (slot
= 0; slot
< key
->num_inputs
; slot
++) {
555 switch (key
->inputs
[slot
].interp
) {
556 case LP_INTERP_CONSTANT
:
557 load_attribute(gallivm
, args
, key
, key
->inputs
[slot
].src_index
, attribs
);
558 if (key
->flatshade_first
) {
559 emit_constant_coef4(gallivm
, args
, slot
+1, attribs
[0]);
562 emit_constant_coef4(gallivm
, args
, slot
+1, attribs
[2]);
566 case LP_INTERP_LINEAR
:
567 load_attribute(gallivm
, args
, key
, key
->inputs
[slot
].src_index
, attribs
);
568 emit_apply_cyl_wrap(gallivm
, args
, key
->inputs
[slot
].cyl_wrap
, attribs
);
569 emit_linear_coef(gallivm
, args
, slot
+1, attribs
);
572 case LP_INTERP_PERSPECTIVE
:
573 load_attribute(gallivm
, args
, key
, key
->inputs
[slot
].src_index
, attribs
);
574 emit_apply_cyl_wrap(gallivm
, args
, key
->inputs
[slot
].cyl_wrap
, attribs
);
575 apply_perspective_corr(gallivm
, args
, slot
+1, attribs
);
576 emit_linear_coef(gallivm
, args
, slot
+1, attribs
);
579 case LP_INTERP_POSITION
:
581 * The generated pixel interpolators will pick up the coeffs from
586 case LP_INTERP_FACING
:
587 emit_facing_coef(gallivm
, args
, slot
+1);
597 /* XXX: generic code:
600 set_noalias(LLVMBuilderRef builder
,
601 LLVMValueRef function
,
602 const LLVMTypeRef
*arg_types
,
606 for(i
= 0; i
< nr_args
; ++i
)
607 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
608 LLVMAddAttribute(LLVMGetParam(function
, i
),
609 LLVMNoAliasAttribute
);
613 init_args(struct gallivm_state
*gallivm
,
614 const struct lp_setup_variant_key
*key
,
615 struct lp_setup_args
*args
)
617 LLVMBuilderRef b
= gallivm
->builder
;
618 LLVMTypeRef shuf_type
= LLVMInt32TypeInContext(gallivm
->context
);
619 LLVMValueRef onef
= lp_build_const_float(gallivm
, 1.0);
620 LLVMValueRef onei
= lp_build_const_int32(gallivm
, 1);
621 LLVMValueRef zeroi
= lp_build_const_int32(gallivm
, 0);
622 LLVMValueRef pixel_center
, xy0_center
, dxy01
, dxy20
, dyx20
;
623 LLVMValueRef e
, f
, ef
, ooa
;
624 LLVMValueRef shuffles
[4];
625 LLVMValueRef attr_pos
[3];
626 struct lp_type typef4
= lp_type_float_vec(32, 128);
627 struct lp_build_context bld
;
629 lp_build_context_init(&bld
, gallivm
, typef4
);
632 /* The internal position input is in slot zero:
634 load_attribute(gallivm
, args
, key
, 0, attr_pos
);
636 pixel_center
= lp_build_const_vec(gallivm
, typef4
,
637 key
->pixel_center_half
? 0.5 : 0.0);
640 * xy are first two elems in v0a/v1a/v2a but just use vec4 arit
641 * also offset_tri uses actually xyz in them
643 xy0_center
= LLVMBuildFSub(b
, attr_pos
[0], pixel_center
, "xy0_center" );
645 dxy01
= LLVMBuildFSub(b
, attr_pos
[0], attr_pos
[1], "dxy01");
646 dxy20
= LLVMBuildFSub(b
, attr_pos
[2], attr_pos
[0], "dxy20");
650 shuffles
[2] = LLVMGetUndef(shuf_type
);
651 shuffles
[3] = LLVMGetUndef(shuf_type
);
653 dyx20
= LLVMBuildShuffleVector(b
, dxy20
, dxy20
, LLVMConstVector(shuffles
, 4), "");
655 ef
= LLVMBuildFMul(b
, dxy01
, dyx20
, "ef");
656 e
= LLVMBuildExtractElement(b
, ef
, zeroi
, "");
657 f
= LLVMBuildExtractElement(b
, ef
, onei
, "");
659 ooa
= LLVMBuildFDiv(b
, onef
, LLVMBuildFSub(b
, e
, f
, ""), "ooa");
661 ooa
= lp_build_broadcast_scalar(&bld
, ooa
);
663 /* tri offset calc shares a lot of arithmetic, do it here */
664 if (key
->pgon_offset_scale
!= 0.0f
|| key
->pgon_offset_units
!= 0.0f
) {
665 lp_do_offset_tri(gallivm
, args
, key
, ooa
, dxy01
, dxy20
, attr_pos
);
668 dxy20
= LLVMBuildFMul(b
, dxy20
, ooa
, "");
669 dxy01
= LLVMBuildFMul(b
, dxy01
, ooa
, "");
671 args
->dy20_ooa
= lp_build_extract_broadcast(gallivm
, typef4
, typef4
, dxy20
, onei
);
672 args
->dy01_ooa
= lp_build_extract_broadcast(gallivm
, typef4
, typef4
, dxy01
, onei
);
674 args
->dx20_ooa
= lp_build_extract_broadcast(gallivm
, typef4
, typef4
, dxy20
, zeroi
);
675 args
->dx01_ooa
= lp_build_extract_broadcast(gallivm
, typef4
, typef4
, dxy01
, zeroi
);
677 args
->x0_center
= lp_build_extract_broadcast(gallivm
, typef4
, typef4
, xy0_center
, zeroi
);
678 args
->y0_center
= lp_build_extract_broadcast(gallivm
, typef4
, typef4
, xy0_center
, onei
);
680 emit_linear_coef(gallivm
, args
, 0, attr_pos
);
684 * Generate the runtime callable function for the coefficient calculation.
687 static struct lp_setup_variant
*
688 generate_setup_variant(struct lp_setup_variant_key
*key
,
689 struct llvmpipe_context
*lp
)
691 struct lp_setup_variant
*variant
= NULL
;
692 struct gallivm_state
*gallivm
;
693 struct lp_setup_args args
;
695 LLVMTypeRef vec4f_type
;
696 LLVMTypeRef func_type
;
697 LLVMTypeRef arg_types
[7];
698 LLVMBasicBlockRef block
;
699 LLVMBuilderRef builder
;
705 variant
= CALLOC_STRUCT(lp_setup_variant
);
709 variant
->gallivm
= gallivm
= gallivm_create();
710 if (!variant
->gallivm
) {
714 builder
= gallivm
->builder
;
716 if (LP_DEBUG
& DEBUG_COUNTERS
) {
720 memcpy(&variant
->key
, key
, key
->size
);
721 variant
->list_item_global
.base
= variant
;
723 util_snprintf(func_name
, sizeof(func_name
), "fs%u_setup%u",
726 /* Currently always deal with full 4-wide vertex attributes from
730 vec4f_type
= LLVMVectorType(LLVMFloatTypeInContext(gallivm
->context
), 4);
732 arg_types
[0] = LLVMPointerType(vec4f_type
, 0); /* v0 */
733 arg_types
[1] = LLVMPointerType(vec4f_type
, 0); /* v1 */
734 arg_types
[2] = LLVMPointerType(vec4f_type
, 0); /* v2 */
735 arg_types
[3] = LLVMInt32TypeInContext(gallivm
->context
); /* facing */
736 arg_types
[4] = LLVMPointerType(vec4f_type
, 0); /* a0, aligned */
737 arg_types
[5] = LLVMPointerType(vec4f_type
, 0); /* dadx, aligned */
738 arg_types
[6] = LLVMPointerType(vec4f_type
, 0); /* dady, aligned */
740 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
741 arg_types
, Elements(arg_types
), 0);
743 variant
->function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
744 if (!variant
->function
)
747 LLVMSetFunctionCallConv(variant
->function
, LLVMCCallConv
);
749 args
.v0
= LLVMGetParam(variant
->function
, 0);
750 args
.v1
= LLVMGetParam(variant
->function
, 1);
751 args
.v2
= LLVMGetParam(variant
->function
, 2);
752 args
.facing
= LLVMGetParam(variant
->function
, 3);
753 args
.a0
= LLVMGetParam(variant
->function
, 4);
754 args
.dadx
= LLVMGetParam(variant
->function
, 5);
755 args
.dady
= LLVMGetParam(variant
->function
, 6);
757 lp_build_name(args
.v0
, "in_v0");
758 lp_build_name(args
.v1
, "in_v1");
759 lp_build_name(args
.v2
, "in_v2");
760 lp_build_name(args
.facing
, "in_facing");
761 lp_build_name(args
.a0
, "out_a0");
762 lp_build_name(args
.dadx
, "out_dadx");
763 lp_build_name(args
.dady
, "out_dady");
768 block
= LLVMAppendBasicBlockInContext(gallivm
->context
,
769 variant
->function
, "entry");
770 LLVMPositionBuilderAtEnd(builder
, block
);
772 set_noalias(builder
, variant
->function
, arg_types
, Elements(arg_types
));
773 init_args(gallivm
, &variant
->key
, &args
);
774 emit_tri_coef(gallivm
, &variant
->key
, &args
);
776 LLVMBuildRetVoid(builder
);
778 gallivm_verify_function(gallivm
, variant
->function
);
780 gallivm_compile_module(gallivm
);
782 variant
->jit_function
= (lp_jit_setup_triangle
)
783 gallivm_jit_function(gallivm
, variant
->function
);
784 if (!variant
->jit_function
)
788 * Update timing information:
790 if (LP_DEBUG
& DEBUG_COUNTERS
) {
792 LP_COUNT_ADD(llvm_compile_time
, t1
- t0
);
793 LP_COUNT_ADD(nr_llvm_compiles
, 1);
800 if (variant
->function
) {
801 gallivm_free_function(gallivm
,
803 variant
->jit_function
);
805 if (variant
->gallivm
) {
806 gallivm_destroy(variant
->gallivm
);
817 lp_make_setup_variant_key(struct llvmpipe_context
*lp
,
818 struct lp_setup_variant_key
*key
)
820 struct lp_fragment_shader
*fs
= lp
->fs
;
823 assert(sizeof key
->inputs
[0] == sizeof(uint
));
825 key
->num_inputs
= fs
->info
.base
.num_inputs
;
826 key
->flatshade_first
= lp
->rasterizer
->flatshade_first
;
827 key
->pixel_center_half
= lp
->rasterizer
->half_pixel_center
;
828 key
->twoside
= lp
->rasterizer
->light_twoside
;
829 key
->size
= Offset(struct lp_setup_variant_key
,
830 inputs
[key
->num_inputs
]);
832 key
->color_slot
= lp
->color_slot
[0];
833 key
->bcolor_slot
= lp
->bcolor_slot
[0];
834 key
->spec_slot
= lp
->color_slot
[1];
835 key
->bspec_slot
= lp
->bcolor_slot
[1];
836 assert(key
->color_slot
== lp
->color_slot
[0]);
837 assert(key
->bcolor_slot
== lp
->bcolor_slot
[0]);
838 assert(key
->spec_slot
== lp
->color_slot
[1]);
839 assert(key
->bspec_slot
== lp
->bcolor_slot
[1]);
842 * If depth is floating point, depth bias is calculated with respect
843 * to the primitive's maximum Z value. Retain the original depth bias
844 * value until that stage.
846 key
->floating_point_depth
= lp
->floating_point_depth
;
848 if (key
->floating_point_depth
) {
849 key
->pgon_offset_units
= (float) lp
->rasterizer
->offset_units
;
851 key
->pgon_offset_units
=
852 (float) (lp
->rasterizer
->offset_units
* lp
->mrd
);
855 key
->pgon_offset_scale
= lp
->rasterizer
->offset_scale
;
856 key
->pgon_offset_clamp
= lp
->rasterizer
->offset_clamp
;
858 memcpy(key
->inputs
, fs
->inputs
, key
->num_inputs
* sizeof key
->inputs
[0]);
859 for (i
= 0; i
< key
->num_inputs
; i
++) {
860 if (key
->inputs
[i
].interp
== LP_INTERP_COLOR
) {
861 if (lp
->rasterizer
->flatshade
)
862 key
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
864 key
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
872 remove_setup_variant(struct llvmpipe_context
*lp
,
873 struct lp_setup_variant
*variant
)
875 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
876 debug_printf("llvmpipe: del setup_variant #%u total %u\n",
877 variant
->no
, lp
->nr_setup_variants
);
880 if (variant
->function
) {
881 gallivm_free_function(variant
->gallivm
,
883 variant
->jit_function
);
886 if (variant
->gallivm
) {
887 gallivm_destroy(variant
->gallivm
);
890 remove_from_list(&variant
->list_item_global
);
891 lp
->nr_setup_variants
--;
897 /* When the number of setup variants exceeds a threshold, cull a
898 * fraction (currently a quarter) of them.
901 cull_setup_variants(struct llvmpipe_context
*lp
)
903 struct pipe_context
*pipe
= &lp
->pipe
;
907 * XXX: we need to flush the context until we have some sort of reference
908 * counting in fragment shaders as they may still be binned
909 * Flushing alone might not be sufficient we need to wait on it too.
911 llvmpipe_finish(pipe
, __FUNCTION__
);
913 for (i
= 0; i
< LP_MAX_SETUP_VARIANTS
/ 4; i
++) {
914 struct lp_setup_variant_list_item
*item
;
915 if (is_empty_list(&lp
->setup_variants_list
)) {
918 item
= last_elem(&lp
->setup_variants_list
);
921 remove_setup_variant(lp
, item
->base
);
927 * Update fragment/vertex shader linkage state. This is called just
928 * prior to drawing something when some fragment-related state has
932 llvmpipe_update_setup(struct llvmpipe_context
*lp
)
934 struct lp_setup_variant_key
*key
= &lp
->setup_variant
.key
;
935 struct lp_setup_variant
*variant
= NULL
;
936 struct lp_setup_variant_list_item
*li
;
938 lp_make_setup_variant_key(lp
, key
);
940 foreach(li
, &lp
->setup_variants_list
) {
941 if(li
->base
->key
.size
== key
->size
&&
942 memcmp(&li
->base
->key
, key
, key
->size
) == 0) {
949 move_to_head(&lp
->setup_variants_list
, &variant
->list_item_global
);
952 if (lp
->nr_setup_variants
>= LP_MAX_SETUP_VARIANTS
) {
953 cull_setup_variants(lp
);
956 variant
= generate_setup_variant(key
, lp
);
958 insert_at_head(&lp
->setup_variants_list
, &variant
->list_item_global
);
959 lp
->nr_setup_variants
++;
960 llvmpipe_variant_count
++;
964 lp_setup_set_setup_variant(lp
->setup
,
969 lp_delete_setup_variants(struct llvmpipe_context
*lp
)
971 struct lp_setup_variant_list_item
*li
;
972 li
= first_elem(&lp
->setup_variants_list
);
973 while(!at_end(&lp
->setup_variants_list
, li
)) {
974 struct lp_setup_variant_list_item
*next
= next_elem(li
);
975 remove_setup_variant(lp
, li
->base
);
981 lp_dump_setup_coef( const struct lp_setup_variant_key
*key
,
982 const float (*sa0
)[4],
983 const float (*sdadx
)[4],
984 const float (*sdady
)[4])
988 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
989 float a0
= sa0
[0][i
];
990 float dadx
= sdadx
[0][i
];
991 float dady
= sdady
[0][i
];
993 debug_printf("POS.%c: a0 = %f, dadx = %f, dady = %f\n",
998 for (slot
= 0; slot
< key
->num_inputs
; slot
++) {
999 unsigned usage_mask
= key
->inputs
[slot
].usage_mask
;
1000 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1001 if (usage_mask
& (1 << i
)) {
1002 float a0
= sa0
[1 + slot
][i
];
1003 float dadx
= sdadx
[1 + slot
][i
];
1004 float dady
= sdady
[1 + slot
][i
];
1006 debug_printf("IN[%u].%c: a0 = %f, dadx = %f, dady = %f\n",