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 LLVMValueRef a0
, face_val
;
135 const unsigned char swizzles
[4] = { PIPE_SWIZZLE_RED
, PIPE_SWIZZLE_ZERO
,
136 PIPE_SWIZZLE_ZERO
, PIPE_SWIZZLE_ZERO
};
137 /* Our face val is either 1 or 0 so we do
138 * face = (val * 2) - 1
142 LLVMBuildFAdd(builder
,
143 LLVMBuildFMul(builder
, a0_0f
,
144 lp_build_const_float(gallivm
, 2.0),
146 lp_build_const_float(gallivm
, -1.0),
148 face_val
= lp_build_broadcast_scalar(&args
->bld
, face_val
);
149 a0
= lp_build_swizzle_aos(&args
->bld
, face_val
, swizzles
);
151 store_coef(gallivm
, args
, slot
, a0
, args
->bld
.zero
, args
->bld
.zero
);
156 vert_attrib(struct gallivm_state
*gallivm
,
162 LLVMBuilderRef b
= gallivm
->builder
;
164 idx
[0] = lp_build_const_int32(gallivm
, attr
);
165 idx
[1] = lp_build_const_int32(gallivm
, elem
);
166 return LLVMBuildLoad(b
, LLVMBuildGEP(b
, vert
, idx
, 2, ""), name
);
171 lp_twoside(struct gallivm_state
*gallivm
,
172 struct lp_setup_args
*args
,
173 const struct lp_setup_variant_key
*key
,
175 LLVMValueRef attribv
[3])
177 LLVMBuilderRef b
= gallivm
->builder
;
178 LLVMValueRef a0_back
, a1_back
, a2_back
;
179 LLVMValueRef idx2
= lp_build_const_int32(gallivm
, bcolor_slot
);
181 LLVMValueRef facing
= args
->facing
;
182 LLVMValueRef front_facing
= LLVMBuildICmp(b
, LLVMIntEQ
, facing
,
183 lp_build_const_int32(gallivm
, 0), ""); /** need i1 for if condition */
185 a0_back
= LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v0
, &idx2
, 1, ""), "v0a_back");
186 a1_back
= LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v1
, &idx2
, 1, ""), "v1a_back");
187 a2_back
= LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v2
, &idx2
, 1, ""), "v2a_back");
189 /* Possibly swap the front and back attrib values,
191 * Prefer select to if so we don't have to worry about phis or
194 attribv
[0] = LLVMBuildSelect(b
, front_facing
, a0_back
, attribv
[0], "");
195 attribv
[1] = LLVMBuildSelect(b
, front_facing
, a1_back
, attribv
[1], "");
196 attribv
[2] = LLVMBuildSelect(b
, front_facing
, a2_back
, attribv
[2], "");
201 lp_do_offset_tri(struct gallivm_state
*gallivm
,
202 struct lp_setup_args
*args
,
203 const struct lp_setup_variant_key
*key
,
204 LLVMValueRef inv_det
,
207 LLVMValueRef attribv
[3])
209 LLVMBuilderRef b
= gallivm
->builder
;
210 struct lp_build_context flt_scalar_bld
;
211 struct lp_build_context int_scalar_bld
;
212 struct lp_build_context
*bld
= &args
->bld
;
213 LLVMValueRef zoffset
, mult
;
214 LLVMValueRef z0_new
, z1_new
, z2_new
;
215 LLVMValueRef dzdxdzdy
, dzdx
, dzdy
, dzxyz20
, dyzzx01
, dyzzx01_dzxyz20
, dzx01_dyz20
;
216 LLVMValueRef z0z1
, z0z1z2
;
217 LLVMValueRef max
, max_value
, res12
;
218 LLVMValueRef shuffles
[4];
219 LLVMTypeRef shuf_type
= LLVMInt32TypeInContext(gallivm
->context
);
220 LLVMValueRef onei
= lp_build_const_int32(gallivm
, 1);
221 LLVMValueRef zeroi
= lp_build_const_int32(gallivm
, 0);
222 LLVMValueRef twoi
= lp_build_const_int32(gallivm
, 2);
223 LLVMValueRef threei
= lp_build_const_int32(gallivm
, 3);
225 /* (res12) = cross(e,f).xy */
230 dzxyz20
= LLVMBuildShuffleVector(b
, dxyz20
, dxyz20
, LLVMConstVector(shuffles
, 4), "");
236 dyzzx01
= LLVMBuildShuffleVector(b
, dxyz01
, dxyz01
, LLVMConstVector(shuffles
, 4), "");
238 dyzzx01_dzxyz20
= LLVMBuildFMul(b
, dzxyz20
, dyzzx01
, "dyzzx01_dzxyz20");
241 shuffles
[1] = threei
;
242 shuffles
[2] = LLVMGetUndef(shuf_type
);
243 shuffles
[3] = LLVMGetUndef(shuf_type
);
244 dzx01_dyz20
= LLVMBuildShuffleVector(b
, dyzzx01_dzxyz20
, dyzzx01_dzxyz20
,
245 LLVMConstVector(shuffles
, 4), "");
247 res12
= LLVMBuildFSub(b
, dyzzx01_dzxyz20
, dzx01_dyz20
, "res12");
249 /* dzdx = fabsf(res1 * inv_det), dydx = fabsf(res2 * inv_det)*/
250 dzdxdzdy
= LLVMBuildFMul(b
, res12
, inv_det
, "dzdxdzdy");
251 dzdxdzdy
= lp_build_abs(bld
, dzdxdzdy
);
253 dzdx
= LLVMBuildExtractElement(b
, dzdxdzdy
, zeroi
, "");
254 dzdy
= LLVMBuildExtractElement(b
, dzdxdzdy
, onei
, "");
256 /* mult = MAX2(dzdx, dzdy) * pgon_offset_scale */
257 max
= LLVMBuildFCmp(b
, LLVMRealUGT
, dzdx
, dzdy
, "");
258 max_value
= LLVMBuildSelect(b
, max
, dzdx
, dzdy
, "max");
260 mult
= LLVMBuildFMul(b
, max_value
,
261 lp_build_const_float(gallivm
, key
->pgon_offset_scale
), "");
263 lp_build_context_init(&flt_scalar_bld
, gallivm
, lp_type_float_vec(32, 32));
265 if (key
->floating_point_depth
) {
267 * bias = pgon_offset_units * 2^(exponent(max(z0, z1, z2)) - mantissa_bits) +
268 * MAX2(dzdx, dzdy) * pgon_offset_scale
270 * NOTE: Assumes IEEE float32.
272 LLVMValueRef c23_shifted
, exp_mask
, bias
, exp
;
273 LLVMValueRef maxz_value
, maxz0z1_value
;
275 lp_build_context_init(&int_scalar_bld
, gallivm
, lp_type_int_vec(32, 32));
277 c23_shifted
= lp_build_const_int32(gallivm
, 23 << 23);
278 exp_mask
= lp_build_const_int32(gallivm
, 0xff << 23);
280 maxz0z1_value
= lp_build_max(&flt_scalar_bld
,
281 LLVMBuildExtractElement(b
, attribv
[0], twoi
, ""),
282 LLVMBuildExtractElement(b
, attribv
[1], twoi
, ""));
284 maxz_value
= lp_build_max(&flt_scalar_bld
,
285 LLVMBuildExtractElement(b
, attribv
[2], twoi
, ""),
288 exp
= LLVMBuildBitCast(b
, maxz_value
, int_scalar_bld
.vec_type
, "");
289 exp
= lp_build_and(&int_scalar_bld
, exp
, exp_mask
);
290 exp
= lp_build_sub(&int_scalar_bld
, exp
, c23_shifted
);
291 /* Clamping to zero means mrd will be zero for very small numbers,
292 * but specs do not indicate this should be prevented by clamping
293 * mrd to smallest normal number instead. */
294 exp
= lp_build_max(&int_scalar_bld
, exp
, int_scalar_bld
.zero
);
295 exp
= LLVMBuildBitCast(b
, exp
, flt_scalar_bld
.vec_type
, "");
297 bias
= LLVMBuildFMul(b
, exp
,
298 lp_build_const_float(gallivm
, key
->pgon_offset_units
),
301 zoffset
= LLVMBuildFAdd(b
, bias
, mult
, "zoffset");
304 * bias = pgon_offset_units + MAX2(dzdx, dzdy) * pgon_offset_scale
306 zoffset
= LLVMBuildFAdd(b
,
307 lp_build_const_float(gallivm
, key
->pgon_offset_units
),
311 if (key
->pgon_offset_clamp
> 0) {
312 zoffset
= lp_build_min(&flt_scalar_bld
,
313 lp_build_const_float(gallivm
, key
->pgon_offset_clamp
),
316 else if (key
->pgon_offset_clamp
< 0) {
317 zoffset
= lp_build_max(&flt_scalar_bld
,
318 lp_build_const_float(gallivm
, key
->pgon_offset_clamp
),
324 shuffles
[1] = lp_build_const_int32(gallivm
, 6);
325 shuffles
[2] = LLVMGetUndef(shuf_type
);
326 shuffles
[3] = LLVMGetUndef(shuf_type
);
327 z0z1
= LLVMBuildShuffleVector(b
, attribv
[0], attribv
[1], LLVMConstVector(shuffles
, 4), "");
330 shuffles
[2] = lp_build_const_int32(gallivm
, 6);
331 shuffles
[3] = LLVMGetUndef(shuf_type
);
332 z0z1z2
= LLVMBuildShuffleVector(b
, z0z1
, attribv
[2], LLVMConstVector(shuffles
, 4), "");
333 zoffset
= lp_build_broadcast_scalar(bld
, zoffset
);
335 /* clamp and do offset */
337 * FIXME I suspect the clamp (is that even right to always clamp to fixed
338 * 0.0/1.0?) should really be per fragment?
340 z0z1z2
= lp_build_clamp(bld
, LLVMBuildFAdd(b
, z0z1z2
, zoffset
, ""), bld
->zero
, bld
->one
);
342 /* insert into args->a0.z, a1.z, a2.z:
344 z0_new
= LLVMBuildExtractElement(b
, z0z1z2
, zeroi
, "");
345 z1_new
= LLVMBuildExtractElement(b
, z0z1z2
, onei
, "");
346 z2_new
= LLVMBuildExtractElement(b
, z0z1z2
, twoi
, "");
347 attribv
[0] = LLVMBuildInsertElement(b
, attribv
[0], z0_new
, twoi
, "");
348 attribv
[1] = LLVMBuildInsertElement(b
, attribv
[1], z1_new
, twoi
, "");
349 attribv
[2] = LLVMBuildInsertElement(b
, attribv
[2], z2_new
, twoi
, "");
353 load_attribute(struct gallivm_state
*gallivm
,
354 struct lp_setup_args
*args
,
355 const struct lp_setup_variant_key
*key
,
357 LLVMValueRef attribv
[3])
359 LLVMBuilderRef b
= gallivm
->builder
;
360 LLVMValueRef idx
= lp_build_const_int32(gallivm
, vert_attr
);
362 /* Load the vertex data
364 attribv
[0] = LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v0
, &idx
, 1, ""), "v0a");
365 attribv
[1] = LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v1
, &idx
, 1, ""), "v1a");
366 attribv
[2] = LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v2
, &idx
, 1, ""), "v2a");
369 /* Potentially modify it according to twoside, etc:
372 if (vert_attr
== key
->color_slot
&& key
->bcolor_slot
>= 0)
373 lp_twoside(gallivm
, args
, key
, key
->bcolor_slot
, attribv
);
374 else if (vert_attr
== key
->spec_slot
&& key
->bspec_slot
>= 0)
375 lp_twoside(gallivm
, args
, key
, key
->bspec_slot
, attribv
);
380 emit_coef4( struct gallivm_state
*gallivm
,
381 struct lp_setup_args
*args
,
387 LLVMBuilderRef b
= gallivm
->builder
;
388 LLVMValueRef dy20_ooa
= args
->dy20_ooa
;
389 LLVMValueRef dy01_ooa
= args
->dy01_ooa
;
390 LLVMValueRef dx20_ooa
= args
->dx20_ooa
;
391 LLVMValueRef dx01_ooa
= args
->dx01_ooa
;
392 LLVMValueRef x0_center
= args
->x0_center
;
393 LLVMValueRef y0_center
= args
->y0_center
;
394 LLVMValueRef da01
= LLVMBuildFSub(b
, a0
, a1
, "da01");
395 LLVMValueRef da20
= LLVMBuildFSub(b
, a2
, a0
, "da20");
397 /* Calculate dadx (vec4f)
399 LLVMValueRef da01_dy20_ooa
= LLVMBuildFMul(b
, da01
, dy20_ooa
, "da01_dy20_ooa");
400 LLVMValueRef da20_dy01_ooa
= LLVMBuildFMul(b
, da20
, dy01_ooa
, "da20_dy01_ooa");
401 LLVMValueRef dadx
= LLVMBuildFSub(b
, da01_dy20_ooa
, da20_dy01_ooa
, "dadx");
403 /* Calculate dady (vec4f)
405 LLVMValueRef da01_dx20_ooa
= LLVMBuildFMul(b
, da01
, dx20_ooa
, "da01_dx20_ooa");
406 LLVMValueRef da20_dx01_ooa
= LLVMBuildFMul(b
, da20
, dx01_ooa
, "da20_dx01_ooa");
407 LLVMValueRef dady
= LLVMBuildFSub(b
, da20_dx01_ooa
, da01_dx20_ooa
, "dady");
409 /* Calculate a0 - the attribute value at the origin
411 LLVMValueRef dadx_x0
= LLVMBuildFMul(b
, dadx
, x0_center
, "dadx_x0");
412 LLVMValueRef dady_y0
= LLVMBuildFMul(b
, dady
, y0_center
, "dady_y0");
413 LLVMValueRef attr_v0
= LLVMBuildFAdd(b
, dadx_x0
, dady_y0
, "attr_v0");
414 LLVMValueRef attr_0
= LLVMBuildFSub(b
, a0
, attr_v0
, "attr_0");
416 store_coef(gallivm
, args
, slot
, attr_0
, dadx
, dady
);
421 emit_linear_coef( struct gallivm_state
*gallivm
,
422 struct lp_setup_args
*args
,
424 LLVMValueRef attribv
[3])
426 /* nothing to do anymore */
436 * Compute a0, dadx and dady for a perspective-corrected interpolant,
438 * We basically multiply the vertex value by 1/w before computing
439 * the plane coefficients (a0, dadx, dady).
440 * Later, when we compute the value at a particular fragment position we'll
441 * divide the interpolated value by the interpolated W at that fragment.
444 apply_perspective_corr( struct gallivm_state
*gallivm
,
445 struct lp_setup_args
*args
,
447 LLVMValueRef attribv
[3])
449 LLVMBuilderRef b
= gallivm
->builder
;
451 /* premultiply by 1/w (v[0][3] is always 1/w):
453 LLVMValueRef v0_oow
= lp_build_broadcast_scalar(&args
->bld
,
454 vert_attrib(gallivm
, args
->v0
, 0, 3, "v0_oow"));
455 LLVMValueRef v1_oow
= lp_build_broadcast_scalar(&args
->bld
,
456 vert_attrib(gallivm
, args
->v1
, 0, 3, "v1_oow"));
457 LLVMValueRef v2_oow
= lp_build_broadcast_scalar(&args
->bld
,
458 vert_attrib(gallivm
, args
->v2
, 0, 3, "v2_oow"));
460 attribv
[0] = LLVMBuildFMul(b
, attribv
[0], v0_oow
, "v0_oow_v0a");
461 attribv
[1] = LLVMBuildFMul(b
, attribv
[1], v1_oow
, "v1_oow_v1a");
462 attribv
[2] = LLVMBuildFMul(b
, attribv
[2], v2_oow
, "v2_oow_v2a");
467 * Applys cylindrical wrapping to vertex attributes if enabled.
468 * Input coordinates must be in [0, 1] range, otherwise results are undefined.
470 * @param cyl_wrap TGSI_CYLINDRICAL_WRAP_x flags
473 emit_apply_cyl_wrap(struct gallivm_state
*gallivm
,
474 struct lp_setup_args
*args
,
476 LLVMValueRef attribv
[3])
479 LLVMBuilderRef builder
= gallivm
->builder
;
480 struct lp_type type
= args
->bld
.type
;
481 LLVMTypeRef float_vec_type
= args
->bld
.vec_type
;
482 LLVMValueRef pos_half
;
483 LLVMValueRef neg_half
;
484 LLVMValueRef cyl_mask
;
493 pos_half
= lp_build_const_vec(gallivm
, type
, +0.5f
);
494 neg_half
= lp_build_const_vec(gallivm
, type
, -0.5f
);
495 cyl_mask
= lp_build_const_mask_aos(gallivm
, type
, cyl_wrap
, 4);
497 one
= lp_build_const_vec(gallivm
, type
, 1.0f
);
498 one
= LLVMBuildBitCast(builder
, one
, lp_build_int_vec_type(gallivm
, type
), "");
499 one
= LLVMBuildAnd(builder
, one
, cyl_mask
, "");
502 delta
= LLVMBuildFSub(builder
, attribv
[1], attribv
[0], "");
504 offset
= lp_build_compare(gallivm
, type
, PIPE_FUNC_GREATER
, delta
, pos_half
);
505 offset
= LLVMBuildAnd(builder
, offset
, one
, "");
506 offset
= LLVMBuildBitCast(builder
, offset
, float_vec_type
, "");
507 attribv
[0] = LLVMBuildFAdd(builder
, attribv
[0], offset
, "");
509 offset
= lp_build_compare(gallivm
, type
, PIPE_FUNC_LESS
, delta
, neg_half
);
510 offset
= LLVMBuildAnd(builder
, offset
, one
, "");
511 offset
= LLVMBuildBitCast(builder
, offset
, float_vec_type
, "");
512 attribv
[1] = LLVMBuildFAdd(builder
, attribv
[1], offset
, "");
515 delta
= LLVMBuildFSub(builder
, attribv
[2], attribv
[1], "");
517 offset
= lp_build_compare(gallivm
, type
, PIPE_FUNC_GREATER
, delta
, pos_half
);
518 offset
= LLVMBuildAnd(builder
, offset
, one
, "");
519 offset
= LLVMBuildBitCast(builder
, offset
, float_vec_type
, "");
520 attribv
[1] = LLVMBuildFAdd(builder
, attribv
[1], offset
, "");
522 offset
= lp_build_compare(gallivm
, type
, PIPE_FUNC_LESS
, delta
, neg_half
);
523 offset
= LLVMBuildAnd(builder
, offset
, one
, "");
524 offset
= LLVMBuildBitCast(builder
, offset
, float_vec_type
, "");
525 attribv
[2] = LLVMBuildFAdd(builder
, attribv
[2], offset
, "");
528 delta
= LLVMBuildFSub(builder
, attribv
[0], attribv
[2], "");
530 offset
= lp_build_compare(gallivm
, type
, PIPE_FUNC_GREATER
, delta
, pos_half
);
531 offset
= LLVMBuildAnd(builder
, offset
, one
, "");
532 offset
= LLVMBuildBitCast(builder
, offset
, float_vec_type
, "");
533 attribv
[2] = LLVMBuildFAdd(builder
, attribv
[2], offset
, "");
535 offset
= lp_build_compare(gallivm
, type
, PIPE_FUNC_LESS
, delta
, neg_half
);
536 offset
= LLVMBuildAnd(builder
, offset
, one
, "");
537 offset
= LLVMBuildBitCast(builder
, offset
, float_vec_type
, "");
538 attribv
[0] = LLVMBuildFAdd(builder
, attribv
[0], offset
, "");
543 * Compute the inputs-> dadx, dady, a0 values.
546 emit_tri_coef( struct gallivm_state
*gallivm
,
547 const struct lp_setup_variant_key
*key
,
548 struct lp_setup_args
*args
)
552 LLVMValueRef attribs
[3];
554 /* setup interpolation for all the remaining attributes:
556 for (slot
= 0; slot
< key
->num_inputs
; slot
++) {
557 switch (key
->inputs
[slot
].interp
) {
558 case LP_INTERP_CONSTANT
:
559 load_attribute(gallivm
, args
, key
, key
->inputs
[slot
].src_index
, attribs
);
560 if (key
->flatshade_first
) {
561 emit_constant_coef4(gallivm
, args
, slot
+1, attribs
[0]);
564 emit_constant_coef4(gallivm
, args
, slot
+1, attribs
[2]);
568 case LP_INTERP_LINEAR
:
569 load_attribute(gallivm
, args
, key
, key
->inputs
[slot
].src_index
, attribs
);
570 emit_apply_cyl_wrap(gallivm
, args
, key
->inputs
[slot
].cyl_wrap
, attribs
);
571 emit_linear_coef(gallivm
, args
, slot
+1, attribs
);
574 case LP_INTERP_PERSPECTIVE
:
575 load_attribute(gallivm
, args
, key
, key
->inputs
[slot
].src_index
, attribs
);
576 emit_apply_cyl_wrap(gallivm
, args
, key
->inputs
[slot
].cyl_wrap
, attribs
);
577 apply_perspective_corr(gallivm
, args
, slot
+1, attribs
);
578 emit_linear_coef(gallivm
, args
, slot
+1, attribs
);
581 case LP_INTERP_POSITION
:
583 * The generated pixel interpolators will pick up the coeffs from
588 case LP_INTERP_FACING
:
589 emit_facing_coef(gallivm
, args
, slot
+1);
599 /* XXX: generic code:
602 set_noalias(LLVMBuilderRef builder
,
603 LLVMValueRef function
,
604 const LLVMTypeRef
*arg_types
,
608 for(i
= 0; i
< nr_args
; ++i
)
609 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
610 LLVMAddAttribute(LLVMGetParam(function
, i
),
611 LLVMNoAliasAttribute
);
615 init_args(struct gallivm_state
*gallivm
,
616 const struct lp_setup_variant_key
*key
,
617 struct lp_setup_args
*args
)
619 LLVMBuilderRef b
= gallivm
->builder
;
620 LLVMTypeRef shuf_type
= LLVMInt32TypeInContext(gallivm
->context
);
621 LLVMValueRef onef
= lp_build_const_float(gallivm
, 1.0);
622 LLVMValueRef onei
= lp_build_const_int32(gallivm
, 1);
623 LLVMValueRef zeroi
= lp_build_const_int32(gallivm
, 0);
624 LLVMValueRef pixel_center
, xy0_center
, dxy01
, dxy20
, dyx20
;
625 LLVMValueRef e
, f
, ef
, ooa
;
626 LLVMValueRef shuffles
[4];
627 LLVMValueRef attr_pos
[3];
628 struct lp_type typef4
= lp_type_float_vec(32, 128);
629 struct lp_build_context bld
;
631 lp_build_context_init(&bld
, gallivm
, typef4
);
634 /* The internal position input is in slot zero:
636 load_attribute(gallivm
, args
, key
, 0, attr_pos
);
638 pixel_center
= lp_build_const_vec(gallivm
, typef4
,
639 key
->pixel_center_half
? 0.5 : 0.0);
642 * xy are first two elems in v0a/v1a/v2a but just use vec4 arit
643 * also offset_tri uses actually xyz in them
645 xy0_center
= LLVMBuildFSub(b
, attr_pos
[0], pixel_center
, "xy0_center" );
647 dxy01
= LLVMBuildFSub(b
, attr_pos
[0], attr_pos
[1], "dxy01");
648 dxy20
= LLVMBuildFSub(b
, attr_pos
[2], attr_pos
[0], "dxy20");
652 shuffles
[2] = LLVMGetUndef(shuf_type
);
653 shuffles
[3] = LLVMGetUndef(shuf_type
);
655 dyx20
= LLVMBuildShuffleVector(b
, dxy20
, dxy20
, LLVMConstVector(shuffles
, 4), "");
657 ef
= LLVMBuildFMul(b
, dxy01
, dyx20
, "ef");
658 e
= LLVMBuildExtractElement(b
, ef
, zeroi
, "");
659 f
= LLVMBuildExtractElement(b
, ef
, onei
, "");
661 ooa
= LLVMBuildFDiv(b
, onef
, LLVMBuildFSub(b
, e
, f
, ""), "ooa");
663 ooa
= lp_build_broadcast_scalar(&bld
, ooa
);
665 /* tri offset calc shares a lot of arithmetic, do it here */
666 if (key
->pgon_offset_scale
!= 0.0f
|| key
->pgon_offset_units
!= 0.0f
) {
667 lp_do_offset_tri(gallivm
, args
, key
, ooa
, dxy01
, dxy20
, attr_pos
);
670 dxy20
= LLVMBuildFMul(b
, dxy20
, ooa
, "");
671 dxy01
= LLVMBuildFMul(b
, dxy01
, ooa
, "");
673 args
->dy20_ooa
= lp_build_extract_broadcast(gallivm
, typef4
, typef4
, dxy20
, onei
);
674 args
->dy01_ooa
= lp_build_extract_broadcast(gallivm
, typef4
, typef4
, dxy01
, onei
);
676 args
->dx20_ooa
= lp_build_extract_broadcast(gallivm
, typef4
, typef4
, dxy20
, zeroi
);
677 args
->dx01_ooa
= lp_build_extract_broadcast(gallivm
, typef4
, typef4
, dxy01
, zeroi
);
679 args
->x0_center
= lp_build_extract_broadcast(gallivm
, typef4
, typef4
, xy0_center
, zeroi
);
680 args
->y0_center
= lp_build_extract_broadcast(gallivm
, typef4
, typef4
, xy0_center
, onei
);
682 emit_linear_coef(gallivm
, args
, 0, attr_pos
);
686 * Generate the runtime callable function for the coefficient calculation.
689 static struct lp_setup_variant
*
690 generate_setup_variant(struct lp_setup_variant_key
*key
,
691 struct llvmpipe_context
*lp
)
693 struct lp_setup_variant
*variant
= NULL
;
694 struct gallivm_state
*gallivm
;
695 struct lp_setup_args args
;
697 LLVMTypeRef vec4f_type
;
698 LLVMTypeRef func_type
;
699 LLVMTypeRef arg_types
[7];
700 LLVMBasicBlockRef block
;
701 LLVMBuilderRef builder
;
707 variant
= CALLOC_STRUCT(lp_setup_variant
);
711 variant
->gallivm
= gallivm
= gallivm_create();
712 if (!variant
->gallivm
) {
716 builder
= gallivm
->builder
;
718 if (LP_DEBUG
& DEBUG_COUNTERS
) {
722 memcpy(&variant
->key
, key
, key
->size
);
723 variant
->list_item_global
.base
= variant
;
725 util_snprintf(func_name
, sizeof(func_name
), "fs%u_setup%u",
728 /* Currently always deal with full 4-wide vertex attributes from
732 vec4f_type
= LLVMVectorType(LLVMFloatTypeInContext(gallivm
->context
), 4);
734 arg_types
[0] = LLVMPointerType(vec4f_type
, 0); /* v0 */
735 arg_types
[1] = LLVMPointerType(vec4f_type
, 0); /* v1 */
736 arg_types
[2] = LLVMPointerType(vec4f_type
, 0); /* v2 */
737 arg_types
[3] = LLVMInt32TypeInContext(gallivm
->context
); /* facing */
738 arg_types
[4] = LLVMPointerType(vec4f_type
, 0); /* a0, aligned */
739 arg_types
[5] = LLVMPointerType(vec4f_type
, 0); /* dadx, aligned */
740 arg_types
[6] = LLVMPointerType(vec4f_type
, 0); /* dady, aligned */
742 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
743 arg_types
, Elements(arg_types
), 0);
745 variant
->function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
746 if (!variant
->function
)
749 LLVMSetFunctionCallConv(variant
->function
, LLVMCCallConv
);
751 args
.v0
= LLVMGetParam(variant
->function
, 0);
752 args
.v1
= LLVMGetParam(variant
->function
, 1);
753 args
.v2
= LLVMGetParam(variant
->function
, 2);
754 args
.facing
= LLVMGetParam(variant
->function
, 3);
755 args
.a0
= LLVMGetParam(variant
->function
, 4);
756 args
.dadx
= LLVMGetParam(variant
->function
, 5);
757 args
.dady
= LLVMGetParam(variant
->function
, 6);
759 lp_build_name(args
.v0
, "in_v0");
760 lp_build_name(args
.v1
, "in_v1");
761 lp_build_name(args
.v2
, "in_v2");
762 lp_build_name(args
.facing
, "in_facing");
763 lp_build_name(args
.a0
, "out_a0");
764 lp_build_name(args
.dadx
, "out_dadx");
765 lp_build_name(args
.dady
, "out_dady");
770 block
= LLVMAppendBasicBlockInContext(gallivm
->context
,
771 variant
->function
, "entry");
772 LLVMPositionBuilderAtEnd(builder
, block
);
774 set_noalias(builder
, variant
->function
, arg_types
, Elements(arg_types
));
775 init_args(gallivm
, &variant
->key
, &args
);
776 emit_tri_coef(gallivm
, &variant
->key
, &args
);
778 LLVMBuildRetVoid(builder
);
780 gallivm_verify_function(gallivm
, variant
->function
);
782 gallivm_compile_module(gallivm
);
784 variant
->jit_function
= (lp_jit_setup_triangle
)
785 gallivm_jit_function(gallivm
, variant
->function
);
786 if (!variant
->jit_function
)
790 * Update timing information:
792 if (LP_DEBUG
& DEBUG_COUNTERS
) {
794 LP_COUNT_ADD(llvm_compile_time
, t1
- t0
);
795 LP_COUNT_ADD(nr_llvm_compiles
, 1);
802 if (variant
->function
) {
803 gallivm_free_function(gallivm
,
805 variant
->jit_function
);
807 if (variant
->gallivm
) {
808 gallivm_destroy(variant
->gallivm
);
819 lp_make_setup_variant_key(struct llvmpipe_context
*lp
,
820 struct lp_setup_variant_key
*key
)
822 struct lp_fragment_shader
*fs
= lp
->fs
;
825 assert(sizeof key
->inputs
[0] == sizeof(uint
));
827 key
->num_inputs
= fs
->info
.base
.num_inputs
;
828 key
->flatshade_first
= lp
->rasterizer
->flatshade_first
;
829 key
->pixel_center_half
= lp
->rasterizer
->half_pixel_center
;
830 key
->twoside
= lp
->rasterizer
->light_twoside
;
831 key
->size
= Offset(struct lp_setup_variant_key
,
832 inputs
[key
->num_inputs
]);
834 key
->color_slot
= lp
->color_slot
[0];
835 key
->bcolor_slot
= lp
->bcolor_slot
[0];
836 key
->spec_slot
= lp
->color_slot
[1];
837 key
->bspec_slot
= lp
->bcolor_slot
[1];
838 assert(key
->color_slot
== lp
->color_slot
[0]);
839 assert(key
->bcolor_slot
== lp
->bcolor_slot
[0]);
840 assert(key
->spec_slot
== lp
->color_slot
[1]);
841 assert(key
->bspec_slot
== lp
->bcolor_slot
[1]);
844 * If depth is floating point, depth bias is calculated with respect
845 * to the primitive's maximum Z value. Retain the original depth bias
846 * value until that stage.
848 key
->floating_point_depth
= lp
->floating_point_depth
;
850 if (key
->floating_point_depth
) {
851 key
->pgon_offset_units
= (float) lp
->rasterizer
->offset_units
;
853 key
->pgon_offset_units
=
854 (float) (lp
->rasterizer
->offset_units
* lp
->mrd
);
857 key
->pgon_offset_scale
= lp
->rasterizer
->offset_scale
;
858 key
->pgon_offset_clamp
= lp
->rasterizer
->offset_clamp
;
860 memcpy(key
->inputs
, fs
->inputs
, key
->num_inputs
* sizeof key
->inputs
[0]);
861 for (i
= 0; i
< key
->num_inputs
; i
++) {
862 if (key
->inputs
[i
].interp
== LP_INTERP_COLOR
) {
863 if (lp
->rasterizer
->flatshade
)
864 key
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
866 key
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
874 remove_setup_variant(struct llvmpipe_context
*lp
,
875 struct lp_setup_variant
*variant
)
877 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
878 debug_printf("llvmpipe: del setup_variant #%u total %u\n",
879 variant
->no
, lp
->nr_setup_variants
);
882 if (variant
->function
) {
883 gallivm_free_function(variant
->gallivm
,
885 variant
->jit_function
);
888 if (variant
->gallivm
) {
889 gallivm_destroy(variant
->gallivm
);
892 remove_from_list(&variant
->list_item_global
);
893 lp
->nr_setup_variants
--;
899 /* When the number of setup variants exceeds a threshold, cull a
900 * fraction (currently a quarter) of them.
903 cull_setup_variants(struct llvmpipe_context
*lp
)
905 struct pipe_context
*pipe
= &lp
->pipe
;
909 * XXX: we need to flush the context until we have some sort of reference
910 * counting in fragment shaders as they may still be binned
911 * Flushing alone might not be sufficient we need to wait on it too.
913 llvmpipe_finish(pipe
, __FUNCTION__
);
915 for (i
= 0; i
< LP_MAX_SETUP_VARIANTS
/ 4; i
++) {
916 struct lp_setup_variant_list_item
*item
;
917 if (is_empty_list(&lp
->setup_variants_list
)) {
920 item
= last_elem(&lp
->setup_variants_list
);
923 remove_setup_variant(lp
, item
->base
);
929 * Update fragment/vertex shader linkage state. This is called just
930 * prior to drawing something when some fragment-related state has
934 llvmpipe_update_setup(struct llvmpipe_context
*lp
)
936 struct lp_setup_variant_key
*key
= &lp
->setup_variant
.key
;
937 struct lp_setup_variant
*variant
= NULL
;
938 struct lp_setup_variant_list_item
*li
;
940 lp_make_setup_variant_key(lp
, key
);
942 foreach(li
, &lp
->setup_variants_list
) {
943 if(li
->base
->key
.size
== key
->size
&&
944 memcmp(&li
->base
->key
, key
, key
->size
) == 0) {
951 move_to_head(&lp
->setup_variants_list
, &variant
->list_item_global
);
954 if (lp
->nr_setup_variants
>= LP_MAX_SETUP_VARIANTS
) {
955 cull_setup_variants(lp
);
958 variant
= generate_setup_variant(key
, lp
);
960 insert_at_head(&lp
->setup_variants_list
, &variant
->list_item_global
);
961 lp
->nr_setup_variants
++;
962 llvmpipe_variant_count
++;
966 lp_setup_set_setup_variant(lp
->setup
,
971 lp_delete_setup_variants(struct llvmpipe_context
*lp
)
973 struct lp_setup_variant_list_item
*li
;
974 li
= first_elem(&lp
->setup_variants_list
);
975 while(!at_end(&lp
->setup_variants_list
, li
)) {
976 struct lp_setup_variant_list_item
*next
= next_elem(li
);
977 remove_setup_variant(lp
, li
->base
);
983 lp_dump_setup_coef( const struct lp_setup_variant_key
*key
,
984 const float (*sa0
)[4],
985 const float (*sdadx
)[4],
986 const float (*sdady
)[4])
990 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
991 float a0
= sa0
[0][i
];
992 float dadx
= sdadx
[0][i
];
993 float dady
= sdady
[0][i
];
995 debug_printf("POS.%c: a0 = %f, dadx = %f, dady = %f\n",
1000 for (slot
= 0; slot
< key
->num_inputs
; slot
++) {
1001 unsigned usage_mask
= key
->inputs
[slot
].usage_mask
;
1002 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1003 if (usage_mask
& (1 << i
)) {
1004 float a0
= sa0
[1 + slot
][i
];
1005 float dadx
= sdadx
[1 + slot
][i
];
1006 float dady
= sdady
[1 + slot
][i
];
1008 debug_printf("IN[%u].%c: a0 = %f, dadx = %f, dady = %f\n",