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"
53 /* currently organized to interpolate full float[4] attributes even
54 * when some elements are unused. Later, can pack vertex data more
61 /* Function arguments:
66 LLVMValueRef facing
; /* boolean */
73 LLVMValueRef x0_center
;
74 LLVMValueRef y0_center
;
75 LLVMValueRef dy20_ooa
;
76 LLVMValueRef dy01_ooa
;
77 LLVMValueRef dx20_ooa
;
78 LLVMValueRef dx01_ooa
;
79 struct lp_build_context bld
;
84 store_coef(struct gallivm_state
*gallivm
,
85 struct lp_setup_args
*args
,
91 LLVMBuilderRef builder
= gallivm
->builder
;
92 LLVMValueRef idx
= lp_build_const_int32(gallivm
, slot
);
94 LLVMBuildStore(builder
,
96 LLVMBuildGEP(builder
, args
->a0
, &idx
, 1, ""));
98 LLVMBuildStore(builder
,
100 LLVMBuildGEP(builder
, args
->dadx
, &idx
, 1, ""));
102 LLVMBuildStore(builder
,
104 LLVMBuildGEP(builder
, args
->dady
, &idx
, 1, ""));
110 emit_constant_coef4(struct gallivm_state
*gallivm
,
111 struct lp_setup_args
*args
,
115 store_coef(gallivm
, args
, slot
, vert
, args
->bld
.zero
, args
->bld
.zero
);
121 * Setup the fragment input attribute with the front-facing value.
122 * \param frontface is the triangle front facing?
125 emit_facing_coef(struct gallivm_state
*gallivm
,
126 struct lp_setup_args
*args
,
129 LLVMBuilderRef builder
= gallivm
->builder
;
130 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
131 LLVMValueRef a0_0
= args
->facing
;
132 LLVMValueRef a0_0f
= LLVMBuildSIToFP(builder
, a0_0
, float_type
, "");
133 LLVMValueRef a0
, face_val
;
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
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 face_val
= lp_build_broadcast_scalar(&args
->bld
, face_val
);
148 a0
= lp_build_swizzle_aos(&args
->bld
, face_val
, swizzles
);
150 store_coef(gallivm
, args
, slot
, a0
, args
->bld
.zero
, args
->bld
.zero
);
155 vert_attrib(struct gallivm_state
*gallivm
,
161 LLVMBuilderRef b
= gallivm
->builder
;
163 idx
[0] = lp_build_const_int32(gallivm
, attr
);
164 idx
[1] = lp_build_const_int32(gallivm
, elem
);
165 return LLVMBuildLoad(b
, LLVMBuildGEP(b
, vert
, idx
, 2, ""), name
);
170 lp_twoside(struct gallivm_state
*gallivm
,
171 struct lp_setup_args
*args
,
172 const struct lp_setup_variant_key
*key
,
174 LLVMValueRef attribv
[3])
176 LLVMBuilderRef b
= gallivm
->builder
;
177 LLVMValueRef a0_back
, a1_back
, a2_back
;
178 LLVMValueRef idx2
= lp_build_const_int32(gallivm
, bcolor_slot
);
180 LLVMValueRef facing
= args
->facing
;
181 LLVMValueRef front_facing
= LLVMBuildICmp(b
, LLVMIntEQ
, facing
,
182 lp_build_const_int32(gallivm
, 0), ""); /** need i1 for if condition */
184 a0_back
= LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v0
, &idx2
, 1, ""), "v0a_back");
185 a1_back
= LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v1
, &idx2
, 1, ""), "v1a_back");
186 a2_back
= LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v2
, &idx2
, 1, ""), "v2a_back");
188 /* Possibly swap the front and back attrib values,
190 * Prefer select to if so we don't have to worry about phis or
193 attribv
[0] = LLVMBuildSelect(b
, front_facing
, a0_back
, attribv
[0], "");
194 attribv
[1] = LLVMBuildSelect(b
, front_facing
, a1_back
, attribv
[1], "");
195 attribv
[2] = LLVMBuildSelect(b
, front_facing
, a2_back
, attribv
[2], "");
200 lp_do_offset_tri(struct gallivm_state
*gallivm
,
201 struct lp_setup_args
*args
,
202 const struct lp_setup_variant_key
*key
,
203 LLVMValueRef inv_det
,
206 LLVMValueRef attribv
[3])
208 LLVMBuilderRef b
= gallivm
->builder
;
209 struct lp_build_context flt_scalar_bld
;
210 struct lp_build_context int_scalar_bld
;
211 struct lp_build_context
*bld
= &args
->bld
;
212 LLVMValueRef zoffset
, mult
;
213 LLVMValueRef z0_new
, z1_new
, z2_new
;
214 LLVMValueRef dzdxdzdy
, dzdx
, dzdy
, dzxyz20
, dyzzx01
, dyzzx01_dzxyz20
, dzx01_dyz20
;
215 LLVMValueRef z0z1
, z0z1z2
;
216 LLVMValueRef max
, max_value
, res12
;
217 LLVMValueRef shuffles
[4];
218 LLVMTypeRef shuf_type
= LLVMInt32TypeInContext(gallivm
->context
);
219 LLVMValueRef onei
= lp_build_const_int32(gallivm
, 1);
220 LLVMValueRef zeroi
= lp_build_const_int32(gallivm
, 0);
221 LLVMValueRef twoi
= lp_build_const_int32(gallivm
, 2);
222 LLVMValueRef threei
= lp_build_const_int32(gallivm
, 3);
224 /* (res12) = cross(e,f).xy */
229 dzxyz20
= LLVMBuildShuffleVector(b
, dxyz20
, dxyz20
, LLVMConstVector(shuffles
, 4), "");
235 dyzzx01
= LLVMBuildShuffleVector(b
, dxyz01
, dxyz01
, LLVMConstVector(shuffles
, 4), "");
237 dyzzx01_dzxyz20
= LLVMBuildFMul(b
, dzxyz20
, dyzzx01
, "dyzzx01_dzxyz20");
240 shuffles
[1] = threei
;
241 shuffles
[2] = LLVMGetUndef(shuf_type
);
242 shuffles
[3] = LLVMGetUndef(shuf_type
);
243 dzx01_dyz20
= LLVMBuildShuffleVector(b
, dyzzx01_dzxyz20
, dyzzx01_dzxyz20
,
244 LLVMConstVector(shuffles
, 4), "");
246 res12
= LLVMBuildFSub(b
, dyzzx01_dzxyz20
, dzx01_dyz20
, "res12");
248 /* dzdx = fabsf(res1 * inv_det), dydx = fabsf(res2 * inv_det)*/
249 dzdxdzdy
= LLVMBuildFMul(b
, res12
, inv_det
, "dzdxdzdy");
250 dzdxdzdy
= lp_build_abs(bld
, dzdxdzdy
);
252 dzdx
= LLVMBuildExtractElement(b
, dzdxdzdy
, zeroi
, "");
253 dzdy
= LLVMBuildExtractElement(b
, dzdxdzdy
, onei
, "");
255 /* mult = MAX2(dzdx, dzdy) * pgon_offset_scale */
256 max
= LLVMBuildFCmp(b
, LLVMRealUGT
, dzdx
, dzdy
, "");
257 max_value
= LLVMBuildSelect(b
, max
, dzdx
, dzdy
, "max");
259 mult
= LLVMBuildFMul(b
, max_value
,
260 lp_build_const_float(gallivm
, key
->pgon_offset_scale
), "");
262 lp_build_context_init(&flt_scalar_bld
, gallivm
, lp_type_float_vec(32, 32));
264 if (key
->floating_point_depth
) {
266 * bias = pgon_offset_units * 2^(exponent(max(z0, z1, z2)) - mantissa_bits) +
267 * MAX2(dzdx, dzdy) * pgon_offset_scale
269 * NOTE: Assumes IEEE float32.
271 LLVMValueRef c23_shifted
, exp_mask
, bias
, exp
;
272 LLVMValueRef maxz_value
, maxz0z1_value
;
274 lp_build_context_init(&int_scalar_bld
, gallivm
, lp_type_int_vec(32, 32));
276 c23_shifted
= lp_build_const_int32(gallivm
, 23 << 23);
277 exp_mask
= lp_build_const_int32(gallivm
, 0xff << 23);
279 maxz0z1_value
= lp_build_max(&flt_scalar_bld
,
280 LLVMBuildExtractElement(b
, attribv
[0], twoi
, ""),
281 LLVMBuildExtractElement(b
, attribv
[1], twoi
, ""));
283 maxz_value
= lp_build_max(&flt_scalar_bld
,
284 LLVMBuildExtractElement(b
, attribv
[2], twoi
, ""),
287 exp
= LLVMBuildBitCast(b
, maxz_value
, int_scalar_bld
.vec_type
, "");
288 exp
= lp_build_and(&int_scalar_bld
, exp
, exp_mask
);
289 exp
= lp_build_sub(&int_scalar_bld
, exp
, c23_shifted
);
290 /* Clamping to zero means mrd will be zero for very small numbers,
291 * but specs do not indicate this should be prevented by clamping
292 * mrd to smallest normal number instead. */
293 exp
= lp_build_max(&int_scalar_bld
, exp
, int_scalar_bld
.zero
);
294 exp
= LLVMBuildBitCast(b
, exp
, flt_scalar_bld
.vec_type
, "");
296 bias
= LLVMBuildFMul(b
, exp
,
297 lp_build_const_float(gallivm
, key
->pgon_offset_units
),
300 zoffset
= LLVMBuildFAdd(b
, bias
, mult
, "zoffset");
303 * bias = pgon_offset_units + MAX2(dzdx, dzdy) * pgon_offset_scale
305 zoffset
= LLVMBuildFAdd(b
,
306 lp_build_const_float(gallivm
, key
->pgon_offset_units
),
310 if (key
->pgon_offset_clamp
> 0) {
311 zoffset
= lp_build_min(&flt_scalar_bld
,
312 lp_build_const_float(gallivm
, key
->pgon_offset_clamp
),
315 else if (key
->pgon_offset_clamp
< 0) {
316 zoffset
= lp_build_max(&flt_scalar_bld
,
317 lp_build_const_float(gallivm
, key
->pgon_offset_clamp
),
323 shuffles
[1] = lp_build_const_int32(gallivm
, 6);
324 shuffles
[2] = LLVMGetUndef(shuf_type
);
325 shuffles
[3] = LLVMGetUndef(shuf_type
);
326 z0z1
= LLVMBuildShuffleVector(b
, attribv
[0], attribv
[1], LLVMConstVector(shuffles
, 4), "");
329 shuffles
[2] = lp_build_const_int32(gallivm
, 6);
330 shuffles
[3] = LLVMGetUndef(shuf_type
);
331 z0z1z2
= LLVMBuildShuffleVector(b
, z0z1
, attribv
[2], LLVMConstVector(shuffles
, 4), "");
332 zoffset
= lp_build_broadcast_scalar(bld
, zoffset
);
334 /* clamp and do offset */
336 * FIXME I suspect the clamp (is that even right to always clamp to fixed
337 * 0.0/1.0?) should really be per fragment?
339 z0z1z2
= lp_build_clamp(bld
, LLVMBuildFAdd(b
, z0z1z2
, zoffset
, ""), bld
->zero
, bld
->one
);
341 /* insert into args->a0.z, a1.z, a2.z:
343 z0_new
= LLVMBuildExtractElement(b
, z0z1z2
, zeroi
, "");
344 z1_new
= LLVMBuildExtractElement(b
, z0z1z2
, onei
, "");
345 z2_new
= LLVMBuildExtractElement(b
, z0z1z2
, twoi
, "");
346 attribv
[0] = LLVMBuildInsertElement(b
, attribv
[0], z0_new
, twoi
, "");
347 attribv
[1] = LLVMBuildInsertElement(b
, attribv
[1], z1_new
, twoi
, "");
348 attribv
[2] = LLVMBuildInsertElement(b
, attribv
[2], z2_new
, twoi
, "");
352 load_attribute(struct gallivm_state
*gallivm
,
353 struct lp_setup_args
*args
,
354 const struct lp_setup_variant_key
*key
,
356 LLVMValueRef attribv
[3])
358 LLVMBuilderRef b
= gallivm
->builder
;
359 LLVMValueRef idx
= lp_build_const_int32(gallivm
, vert_attr
);
361 /* Load the vertex data
363 attribv
[0] = LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v0
, &idx
, 1, ""), "v0a");
364 attribv
[1] = LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v1
, &idx
, 1, ""), "v1a");
365 attribv
[2] = LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v2
, &idx
, 1, ""), "v2a");
368 /* Potentially modify it according to twoside, etc:
371 if (vert_attr
== key
->color_slot
&& key
->bcolor_slot
>= 0)
372 lp_twoside(gallivm
, args
, key
, key
->bcolor_slot
, attribv
);
373 else if (vert_attr
== key
->spec_slot
&& key
->bspec_slot
>= 0)
374 lp_twoside(gallivm
, args
, key
, key
->bspec_slot
, attribv
);
379 * FIXME: interpolation is always done wrt fb origin (0/0).
380 * However, if some (small) tri is far away from the origin and gradients
381 * are large, this can lead to HUGE errors, since the a0 value calculated
382 * here can get very large (with the actual values inside the triangle way
383 * smaller), leading to complete loss of accuracy. This could be prevented
384 * by using some point inside (or at corner) of the tri as interpolation
385 * origin, or just use barycentric interpolation (which GL suggests and is
386 * what real hw does - you can get the barycentric coordinates from the
387 * edge functions in rasterization in principle (though we skip these
388 * sometimes completely in case of tris covering a block fully,
389 * which obviously wouldn't work)).
392 emit_coef4( struct gallivm_state
*gallivm
,
393 struct lp_setup_args
*args
,
399 LLVMBuilderRef b
= gallivm
->builder
;
401 LLVMValueRef dy20_ooa
= args
->dy20_ooa
;
402 LLVMValueRef dy01_ooa
= args
->dy01_ooa
;
403 LLVMValueRef dx20_ooa
= args
->dx20_ooa
;
404 LLVMValueRef dx01_ooa
= args
->dx01_ooa
;
405 LLVMValueRef x0_center
= args
->x0_center
;
406 LLVMValueRef y0_center
= args
->y0_center
;
407 LLVMValueRef da01
= LLVMBuildFSub(b
, a0
, a1
, "da01");
408 LLVMValueRef da20
= LLVMBuildFSub(b
, a2
, a0
, "da20");
410 /* Calculate dadx (vec4f)
412 LLVMValueRef da01_dy20_ooa
= LLVMBuildFMul(b
, da01
, dy20_ooa
, "da01_dy20_ooa");
413 LLVMValueRef da20_dy01_ooa
= LLVMBuildFMul(b
, da20
, dy01_ooa
, "da20_dy01_ooa");
414 LLVMValueRef dadx
= LLVMBuildFSub(b
, da01_dy20_ooa
, da20_dy01_ooa
, "dadx");
416 /* Calculate dady (vec4f)
418 LLVMValueRef da01_dx20_ooa
= LLVMBuildFMul(b
, da01
, dx20_ooa
, "da01_dx20_ooa");
419 LLVMValueRef da20_dx01_ooa
= LLVMBuildFMul(b
, da20
, dx01_ooa
, "da20_dx01_ooa");
420 LLVMValueRef dady
= LLVMBuildFSub(b
, da20_dx01_ooa
, da01_dx20_ooa
, "dady");
422 /* Calculate a0 - the attribute value at the origin
424 LLVMValueRef dadx_x0
= LLVMBuildFMul(b
, dadx
, x0_center
, "dadx_x0");
425 LLVMValueRef dady_y0
= LLVMBuildFMul(b
, dady
, y0_center
, "dady_y0");
426 LLVMValueRef attr_v0
= LLVMBuildFAdd(b
, dadx_x0
, dady_y0
, "attr_v0");
427 attr_0
= LLVMBuildFSub(b
, a0
, attr_v0
, "attr_0");
429 store_coef(gallivm
, args
, slot
, attr_0
, dadx
, dady
);
434 emit_linear_coef( struct gallivm_state
*gallivm
,
435 struct lp_setup_args
*args
,
437 LLVMValueRef attribv
[3])
439 /* nothing to do anymore */
449 * Compute a0, dadx and dady for a perspective-corrected interpolant,
451 * We basically multiply the vertex value by 1/w before computing
452 * the plane coefficients (a0, dadx, dady).
453 * Later, when we compute the value at a particular fragment position we'll
454 * divide the interpolated value by the interpolated W at that fragment.
457 apply_perspective_corr( struct gallivm_state
*gallivm
,
458 struct lp_setup_args
*args
,
460 LLVMValueRef attribv
[3])
462 LLVMBuilderRef b
= gallivm
->builder
;
464 /* premultiply by 1/w (v[0][3] is always 1/w):
466 LLVMValueRef v0_oow
= lp_build_broadcast_scalar(&args
->bld
,
467 vert_attrib(gallivm
, args
->v0
, 0, 3, "v0_oow"));
468 LLVMValueRef v1_oow
= lp_build_broadcast_scalar(&args
->bld
,
469 vert_attrib(gallivm
, args
->v1
, 0, 3, "v1_oow"));
470 LLVMValueRef v2_oow
= lp_build_broadcast_scalar(&args
->bld
,
471 vert_attrib(gallivm
, args
->v2
, 0, 3, "v2_oow"));
473 attribv
[0] = LLVMBuildFMul(b
, attribv
[0], v0_oow
, "v0_oow_v0a");
474 attribv
[1] = LLVMBuildFMul(b
, attribv
[1], v1_oow
, "v1_oow_v1a");
475 attribv
[2] = LLVMBuildFMul(b
, attribv
[2], v2_oow
, "v2_oow_v2a");
480 * Applys cylindrical wrapping to vertex attributes if enabled.
481 * Input coordinates must be in [0, 1] range, otherwise results are undefined.
483 * @param cyl_wrap TGSI_CYLINDRICAL_WRAP_x flags
486 emit_apply_cyl_wrap(struct gallivm_state
*gallivm
,
487 struct lp_setup_args
*args
,
489 LLVMValueRef attribv
[3])
492 LLVMBuilderRef builder
= gallivm
->builder
;
493 struct lp_type type
= args
->bld
.type
;
494 LLVMTypeRef float_vec_type
= args
->bld
.vec_type
;
495 LLVMValueRef pos_half
;
496 LLVMValueRef neg_half
;
497 LLVMValueRef cyl_mask
;
506 pos_half
= lp_build_const_vec(gallivm
, type
, +0.5f
);
507 neg_half
= lp_build_const_vec(gallivm
, type
, -0.5f
);
508 cyl_mask
= lp_build_const_mask_aos(gallivm
, type
, cyl_wrap
, 4);
510 one
= lp_build_const_vec(gallivm
, type
, 1.0f
);
511 one
= LLVMBuildBitCast(builder
, one
, lp_build_int_vec_type(gallivm
, type
), "");
512 one
= LLVMBuildAnd(builder
, one
, cyl_mask
, "");
515 delta
= LLVMBuildFSub(builder
, attribv
[1], attribv
[0], "");
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
[0] = LLVMBuildFAdd(builder
, attribv
[0], 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
[1] = LLVMBuildFAdd(builder
, attribv
[1], offset
, "");
528 delta
= LLVMBuildFSub(builder
, attribv
[2], attribv
[1], "");
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
[1] = LLVMBuildFAdd(builder
, attribv
[1], 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
[2] = LLVMBuildFAdd(builder
, attribv
[2], offset
, "");
541 delta
= LLVMBuildFSub(builder
, attribv
[0], attribv
[2], "");
543 offset
= lp_build_compare(gallivm
, type
, PIPE_FUNC_GREATER
, delta
, pos_half
);
544 offset
= LLVMBuildAnd(builder
, offset
, one
, "");
545 offset
= LLVMBuildBitCast(builder
, offset
, float_vec_type
, "");
546 attribv
[2] = LLVMBuildFAdd(builder
, attribv
[2], offset
, "");
548 offset
= lp_build_compare(gallivm
, type
, PIPE_FUNC_LESS
, delta
, neg_half
);
549 offset
= LLVMBuildAnd(builder
, offset
, one
, "");
550 offset
= LLVMBuildBitCast(builder
, offset
, float_vec_type
, "");
551 attribv
[0] = LLVMBuildFAdd(builder
, attribv
[0], offset
, "");
556 * Compute the inputs-> dadx, dady, a0 values.
559 emit_tri_coef( struct gallivm_state
*gallivm
,
560 const struct lp_setup_variant_key
*key
,
561 struct lp_setup_args
*args
)
565 LLVMValueRef attribs
[3];
567 /* setup interpolation for all the remaining attributes:
569 for (slot
= 0; slot
< key
->num_inputs
; slot
++) {
570 switch (key
->inputs
[slot
].interp
) {
571 case LP_INTERP_CONSTANT
:
572 load_attribute(gallivm
, args
, key
, key
->inputs
[slot
].src_index
, attribs
);
573 if (key
->flatshade_first
) {
574 emit_constant_coef4(gallivm
, args
, slot
+1, attribs
[0]);
577 emit_constant_coef4(gallivm
, args
, slot
+1, attribs
[2]);
581 case LP_INTERP_LINEAR
:
582 load_attribute(gallivm
, args
, key
, key
->inputs
[slot
].src_index
, attribs
);
583 emit_apply_cyl_wrap(gallivm
, args
, key
->inputs
[slot
].cyl_wrap
, attribs
);
584 emit_linear_coef(gallivm
, args
, slot
+1, attribs
);
587 case LP_INTERP_PERSPECTIVE
:
588 load_attribute(gallivm
, args
, key
, key
->inputs
[slot
].src_index
, attribs
);
589 emit_apply_cyl_wrap(gallivm
, args
, key
->inputs
[slot
].cyl_wrap
, attribs
);
590 apply_perspective_corr(gallivm
, args
, slot
+1, attribs
);
591 emit_linear_coef(gallivm
, args
, slot
+1, attribs
);
594 case LP_INTERP_POSITION
:
596 * The generated pixel interpolators will pick up the coeffs from
601 case LP_INTERP_FACING
:
602 emit_facing_coef(gallivm
, args
, slot
+1);
612 /* XXX: generic code:
615 set_noalias(LLVMBuilderRef builder
,
616 LLVMValueRef function
,
617 const LLVMTypeRef
*arg_types
,
621 for(i
= 0; i
< nr_args
; ++i
)
622 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
623 LLVMAddAttribute(LLVMGetParam(function
, i
),
624 LLVMNoAliasAttribute
);
628 init_args(struct gallivm_state
*gallivm
,
629 const struct lp_setup_variant_key
*key
,
630 struct lp_setup_args
*args
)
632 LLVMBuilderRef b
= gallivm
->builder
;
633 LLVMTypeRef shuf_type
= LLVMInt32TypeInContext(gallivm
->context
);
634 LLVMValueRef onef
= lp_build_const_float(gallivm
, 1.0);
635 LLVMValueRef onei
= lp_build_const_int32(gallivm
, 1);
636 LLVMValueRef zeroi
= lp_build_const_int32(gallivm
, 0);
637 LLVMValueRef pixel_center
, xy0_center
, dxy01
, dxy20
, dyx20
;
638 LLVMValueRef e
, f
, ef
, ooa
;
639 LLVMValueRef shuffles
[4], shuf10
;
640 LLVMValueRef attr_pos
[3];
641 struct lp_type typef4
= lp_type_float_vec(32, 128);
642 struct lp_build_context bld
;
644 lp_build_context_init(&bld
, gallivm
, typef4
);
647 /* The internal position input is in slot zero:
649 load_attribute(gallivm
, args
, key
, 0, attr_pos
);
651 pixel_center
= lp_build_const_vec(gallivm
, typef4
,
652 key
->pixel_center_half
? 0.5 : 0.0);
655 * xy are first two elems in v0a/v1a/v2a but just use vec4 arit
656 * also offset_tri uses actually xyz in them
658 xy0_center
= LLVMBuildFSub(b
, attr_pos
[0], pixel_center
, "xy0_center" );
660 dxy01
= LLVMBuildFSub(b
, attr_pos
[0], attr_pos
[1], "dxy01");
661 dxy20
= LLVMBuildFSub(b
, attr_pos
[2], attr_pos
[0], "dxy20");
665 shuffles
[2] = LLVMGetUndef(shuf_type
);
666 shuffles
[3] = LLVMGetUndef(shuf_type
);
667 shuf10
= LLVMConstVector(shuffles
, 4);
669 dyx20
= LLVMBuildShuffleVector(b
, dxy20
, dxy20
, shuf10
, "");
671 ef
= LLVMBuildFMul(b
, dxy01
, dyx20
, "ef");
672 e
= LLVMBuildExtractElement(b
, ef
, zeroi
, "");
673 f
= LLVMBuildExtractElement(b
, ef
, onei
, "");
675 ooa
= LLVMBuildFDiv(b
, onef
, LLVMBuildFSub(b
, e
, f
, ""), "ooa");
677 ooa
= lp_build_broadcast_scalar(&bld
, ooa
);
679 /* tri offset calc shares a lot of arithmetic, do it here */
680 if (key
->pgon_offset_scale
!= 0.0f
|| key
->pgon_offset_units
!= 0.0f
) {
681 lp_do_offset_tri(gallivm
, args
, key
, ooa
, dxy01
, dxy20
, attr_pos
);
684 dxy20
= LLVMBuildFMul(b
, dxy20
, ooa
, "");
685 dxy01
= LLVMBuildFMul(b
, dxy01
, ooa
, "");
687 args
->dy20_ooa
= lp_build_extract_broadcast(gallivm
, typef4
, typef4
, dxy20
, onei
);
688 args
->dy01_ooa
= lp_build_extract_broadcast(gallivm
, typef4
, typef4
, dxy01
, onei
);
690 args
->dx20_ooa
= lp_build_extract_broadcast(gallivm
, typef4
, typef4
, dxy20
, zeroi
);
691 args
->dx01_ooa
= lp_build_extract_broadcast(gallivm
, typef4
, typef4
, dxy01
, zeroi
);
693 args
->x0_center
= lp_build_extract_broadcast(gallivm
, typef4
, typef4
, xy0_center
, zeroi
);
694 args
->y0_center
= lp_build_extract_broadcast(gallivm
, typef4
, typef4
, xy0_center
, onei
);
696 emit_linear_coef(gallivm
, args
, 0, attr_pos
);
700 * Generate the runtime callable function for the coefficient calculation.
703 static struct lp_setup_variant
*
704 generate_setup_variant(struct lp_setup_variant_key
*key
,
705 struct llvmpipe_context
*lp
)
707 struct lp_setup_variant
*variant
= NULL
;
708 struct gallivm_state
*gallivm
;
709 struct lp_setup_args args
;
711 LLVMTypeRef vec4f_type
;
712 LLVMTypeRef func_type
;
713 LLVMTypeRef arg_types
[7];
714 LLVMBasicBlockRef block
;
715 LLVMBuilderRef builder
;
721 variant
= CALLOC_STRUCT(lp_setup_variant
);
725 variant
->gallivm
= gallivm
= gallivm_create();
726 if (!variant
->gallivm
) {
730 builder
= gallivm
->builder
;
732 if (LP_DEBUG
& DEBUG_COUNTERS
) {
736 memcpy(&variant
->key
, key
, key
->size
);
737 variant
->list_item_global
.base
= variant
;
739 util_snprintf(func_name
, sizeof(func_name
), "fs%u_setup%u",
742 /* Currently always deal with full 4-wide vertex attributes from
746 vec4f_type
= LLVMVectorType(LLVMFloatTypeInContext(gallivm
->context
), 4);
748 arg_types
[0] = LLVMPointerType(vec4f_type
, 0); /* v0 */
749 arg_types
[1] = LLVMPointerType(vec4f_type
, 0); /* v1 */
750 arg_types
[2] = LLVMPointerType(vec4f_type
, 0); /* v2 */
751 arg_types
[3] = LLVMInt32TypeInContext(gallivm
->context
); /* facing */
752 arg_types
[4] = LLVMPointerType(vec4f_type
, 0); /* a0, aligned */
753 arg_types
[5] = LLVMPointerType(vec4f_type
, 0); /* dadx, aligned */
754 arg_types
[6] = LLVMPointerType(vec4f_type
, 0); /* dady, aligned */
756 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
757 arg_types
, Elements(arg_types
), 0);
759 variant
->function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
760 if (!variant
->function
)
763 LLVMSetFunctionCallConv(variant
->function
, LLVMCCallConv
);
765 args
.v0
= LLVMGetParam(variant
->function
, 0);
766 args
.v1
= LLVMGetParam(variant
->function
, 1);
767 args
.v2
= LLVMGetParam(variant
->function
, 2);
768 args
.facing
= LLVMGetParam(variant
->function
, 3);
769 args
.a0
= LLVMGetParam(variant
->function
, 4);
770 args
.dadx
= LLVMGetParam(variant
->function
, 5);
771 args
.dady
= LLVMGetParam(variant
->function
, 6);
773 lp_build_name(args
.v0
, "in_v0");
774 lp_build_name(args
.v1
, "in_v1");
775 lp_build_name(args
.v2
, "in_v2");
776 lp_build_name(args
.facing
, "in_facing");
777 lp_build_name(args
.a0
, "out_a0");
778 lp_build_name(args
.dadx
, "out_dadx");
779 lp_build_name(args
.dady
, "out_dady");
784 block
= LLVMAppendBasicBlockInContext(gallivm
->context
,
785 variant
->function
, "entry");
786 LLVMPositionBuilderAtEnd(builder
, block
);
788 set_noalias(builder
, variant
->function
, arg_types
, Elements(arg_types
));
789 init_args(gallivm
, &variant
->key
, &args
);
790 emit_tri_coef(gallivm
, &variant
->key
, &args
);
792 LLVMBuildRetVoid(builder
);
794 gallivm_verify_function(gallivm
, variant
->function
);
796 gallivm_compile_module(gallivm
);
798 variant
->jit_function
= (lp_jit_setup_triangle
)
799 gallivm_jit_function(gallivm
, variant
->function
);
800 if (!variant
->jit_function
)
804 * Update timing information:
806 if (LP_DEBUG
& DEBUG_COUNTERS
) {
808 LP_COUNT_ADD(llvm_compile_time
, t1
- t0
);
809 LP_COUNT_ADD(nr_llvm_compiles
, 1);
816 if (variant
->function
) {
817 gallivm_free_function(gallivm
,
819 variant
->jit_function
);
821 if (variant
->gallivm
) {
822 gallivm_destroy(variant
->gallivm
);
833 lp_make_setup_variant_key(struct llvmpipe_context
*lp
,
834 struct lp_setup_variant_key
*key
)
836 struct lp_fragment_shader
*fs
= lp
->fs
;
839 assert(sizeof key
->inputs
[0] == sizeof(uint
));
841 key
->num_inputs
= fs
->info
.base
.num_inputs
;
842 key
->flatshade_first
= lp
->rasterizer
->flatshade_first
;
843 key
->pixel_center_half
= lp
->rasterizer
->half_pixel_center
;
844 key
->twoside
= lp
->rasterizer
->light_twoside
;
845 key
->size
= Offset(struct lp_setup_variant_key
,
846 inputs
[key
->num_inputs
]);
848 key
->color_slot
= lp
->color_slot
[0];
849 key
->bcolor_slot
= lp
->bcolor_slot
[0];
850 key
->spec_slot
= lp
->color_slot
[1];
851 key
->bspec_slot
= lp
->bcolor_slot
[1];
852 assert(key
->color_slot
== lp
->color_slot
[0]);
853 assert(key
->bcolor_slot
== lp
->bcolor_slot
[0]);
854 assert(key
->spec_slot
== lp
->color_slot
[1]);
855 assert(key
->bspec_slot
== lp
->bcolor_slot
[1]);
858 * If depth is floating point, depth bias is calculated with respect
859 * to the primitive's maximum Z value. Retain the original depth bias
860 * value until that stage.
862 key
->floating_point_depth
= lp
->floating_point_depth
;
864 if (key
->floating_point_depth
) {
865 key
->pgon_offset_units
= (float) lp
->rasterizer
->offset_units
;
867 key
->pgon_offset_units
=
868 (float) (lp
->rasterizer
->offset_units
* lp
->mrd
);
871 key
->pgon_offset_scale
= lp
->rasterizer
->offset_scale
;
872 key
->pgon_offset_clamp
= lp
->rasterizer
->offset_clamp
;
874 memcpy(key
->inputs
, fs
->inputs
, key
->num_inputs
* sizeof key
->inputs
[0]);
875 for (i
= 0; i
< key
->num_inputs
; i
++) {
876 if (key
->inputs
[i
].interp
== LP_INTERP_COLOR
) {
877 if (lp
->rasterizer
->flatshade
)
878 key
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
880 key
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
888 remove_setup_variant(struct llvmpipe_context
*lp
,
889 struct lp_setup_variant
*variant
)
891 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
892 debug_printf("llvmpipe: del setup_variant #%u total %u\n",
893 variant
->no
, lp
->nr_setup_variants
);
896 if (variant
->function
) {
897 gallivm_free_function(variant
->gallivm
,
899 variant
->jit_function
);
902 if (variant
->gallivm
) {
903 gallivm_destroy(variant
->gallivm
);
906 remove_from_list(&variant
->list_item_global
);
907 lp
->nr_setup_variants
--;
913 /* When the number of setup variants exceeds a threshold, cull a
914 * fraction (currently a quarter) of them.
917 cull_setup_variants(struct llvmpipe_context
*lp
)
919 struct pipe_context
*pipe
= &lp
->pipe
;
923 * XXX: we need to flush the context until we have some sort of reference
924 * counting in fragment shaders as they may still be binned
925 * Flushing alone might not be sufficient we need to wait on it too.
927 llvmpipe_finish(pipe
, __FUNCTION__
);
929 for (i
= 0; i
< LP_MAX_SETUP_VARIANTS
/ 4; i
++) {
930 struct lp_setup_variant_list_item
*item
;
931 if (is_empty_list(&lp
->setup_variants_list
)) {
934 item
= last_elem(&lp
->setup_variants_list
);
937 remove_setup_variant(lp
, item
->base
);
943 * Update fragment/vertex shader linkage state. This is called just
944 * prior to drawing something when some fragment-related state has
948 llvmpipe_update_setup(struct llvmpipe_context
*lp
)
950 struct lp_setup_variant_key
*key
= &lp
->setup_variant
.key
;
951 struct lp_setup_variant
*variant
= NULL
;
952 struct lp_setup_variant_list_item
*li
;
954 lp_make_setup_variant_key(lp
, key
);
956 foreach(li
, &lp
->setup_variants_list
) {
957 if(li
->base
->key
.size
== key
->size
&&
958 memcmp(&li
->base
->key
, key
, key
->size
) == 0) {
965 move_to_head(&lp
->setup_variants_list
, &variant
->list_item_global
);
968 if (lp
->nr_setup_variants
>= LP_MAX_SETUP_VARIANTS
) {
969 cull_setup_variants(lp
);
972 variant
= generate_setup_variant(key
, lp
);
974 insert_at_head(&lp
->setup_variants_list
, &variant
->list_item_global
);
975 lp
->nr_setup_variants
++;
976 llvmpipe_variant_count
++;
980 lp_setup_set_setup_variant(lp
->setup
,
985 lp_delete_setup_variants(struct llvmpipe_context
*lp
)
987 struct lp_setup_variant_list_item
*li
;
988 li
= first_elem(&lp
->setup_variants_list
);
989 while(!at_end(&lp
->setup_variants_list
, li
)) {
990 struct lp_setup_variant_list_item
*next
= next_elem(li
);
991 remove_setup_variant(lp
, li
->base
);
997 lp_dump_setup_coef( const struct lp_setup_variant_key
*key
,
998 const float (*sa0
)[4],
999 const float (*sdadx
)[4],
1000 const float (*sdady
)[4])
1004 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1005 float a0
= sa0
[0][i
];
1006 float dadx
= sdadx
[0][i
];
1007 float dady
= sdady
[0][i
];
1009 debug_printf("POS.%c: a0 = %f, dadx = %f, dady = %f\n",
1014 for (slot
= 0; slot
< key
->num_inputs
; slot
++) {
1015 unsigned usage_mask
= key
->inputs
[slot
].usage_mask
;
1016 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
1017 if (usage_mask
& (1 << i
)) {
1018 float a0
= sa0
[1 + slot
][i
];
1019 float dadx
= sdadx
[1 + slot
][i
];
1020 float dady
= sdady
[1 + slot
][i
];
1022 debug_printf("IN[%u].%c: a0 = %f, dadx = %f, dady = %f\n",