1/**************************************************************************
2 *
3 * Copyright 2010 VMware.
4 * All Rights Reserved.
5 *
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:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
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.
25 *
26 **************************************************************************/
27
28
29#include "util/u_math.h"
30#include "util/u_memory.h"
31#include "util/simple_list.h"
32#include "util/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"
42
43#include "lp_perf.h"
44#include "lp_debug.h"
45#include "lp_flush.h"
46#include "lp_screen.h"
47#include "lp_context.h"
48#include "lp_state.h"
49#include "lp_state_fs.h"
50#include "lp_state_setup.h"
51
52
53/** Setup shader number (for debugging) */
54static unsigned setup_no = 0;
55
56
57/* currently organized to interpolate full float[4] attributes even
58 * when some elements are unused.  Later, can pack vertex data more
59 * closely.
60 */
61
62
63struct lp_setup_args
64{
65   /* Function arguments:
66    */
67   LLVMValueRef v0;
68   LLVMValueRef v1;
69   LLVMValueRef v2;
70   LLVMValueRef facing;		/* boolean */
71   LLVMValueRef a0;
72   LLVMValueRef dadx;
73   LLVMValueRef dady;
74   LLVMValueRef key;
75
76   /* Derived:
77    */
78   LLVMValueRef x0_center;
79   LLVMValueRef y0_center;
80   LLVMValueRef dy20_ooa;
81   LLVMValueRef dy01_ooa;
82   LLVMValueRef dx20_ooa;
83   LLVMValueRef dx01_ooa;
84   struct lp_build_context bld;
85};
86
87
88static void
89store_coef(struct gallivm_state *gallivm,
90           struct lp_setup_args *args,
91           unsigned slot,
92           LLVMValueRef a0,
93           LLVMValueRef dadx,
94           LLVMValueRef dady)
95{
96   LLVMBuilderRef builder = gallivm->builder;
97   LLVMValueRef idx = lp_build_const_int32(gallivm, slot);
98
99   LLVMBuildStore(builder,
100                  a0,
101                  LLVMBuildGEP(builder, args->a0, &idx, 1, ""));
102
103   LLVMBuildStore(builder,
104                  dadx,
105                  LLVMBuildGEP(builder, args->dadx, &idx, 1, ""));
106
107   LLVMBuildStore(builder,
108                  dady,
109                  LLVMBuildGEP(builder, args->dady, &idx, 1, ""));
110}
111
112
113
114static void
115emit_constant_coef4(struct gallivm_state *gallivm,
116                    struct lp_setup_args *args,
117                    unsigned slot,
118                    LLVMValueRef vert)
119{
120   store_coef(gallivm, args, slot, vert, args->bld.zero, args->bld.zero);
121}
122
123
124
125/**
126 * Setup the fragment input attribute with the front-facing value.
127 * \param frontface  is the triangle front facing?
128 */
129static void
130emit_facing_coef(struct gallivm_state *gallivm,
131                 struct lp_setup_args *args,
132                 unsigned slot )
133{
134   LLVMBuilderRef builder = gallivm->builder;
135   LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
136   LLVMValueRef a0_0 = args->facing;
137   LLVMValueRef a0_0f = LLVMBuildSIToFP(builder, a0_0, float_type, "");
138   LLVMValueRef a0, face_val;
139   const unsigned char swizzles[4] = { PIPE_SWIZZLE_X, PIPE_SWIZZLE_0,
140                                       PIPE_SWIZZLE_0, PIPE_SWIZZLE_0 };
141   /* Our face val is either 1 or 0 so we do
142    * face = (val * 2) - 1
143    * to make it 1 or -1
144    */
145   face_val =
146      LLVMBuildFAdd(builder,
147                    LLVMBuildFMul(builder, a0_0f,
148                                  lp_build_const_float(gallivm, 2.0),
149                                  ""),
150                    lp_build_const_float(gallivm, -1.0),
151                    "facing");
152   face_val = lp_build_broadcast_scalar(&args->bld, face_val);
153   a0 = lp_build_swizzle_aos(&args->bld, face_val, swizzles);
154
155   store_coef(gallivm, args, slot, a0, args->bld.zero, args->bld.zero);
156}
157
158
159static LLVMValueRef
160vert_attrib(struct gallivm_state *gallivm,
161            LLVMValueRef vert,
162            int attr,
163            int elem,
164            const char *name)
165{
166   LLVMBuilderRef b = gallivm->builder;
167   LLVMValueRef idx[2];
168   idx[0] = lp_build_const_int32(gallivm, attr);
169   idx[1] = lp_build_const_int32(gallivm, elem);
170   return LLVMBuildLoad(b, LLVMBuildGEP(b, vert, idx, 2, ""), name);
171}
172
173
174static void
175lp_twoside(struct gallivm_state *gallivm,
176           struct lp_setup_args *args,
177           const struct lp_setup_variant_key *key,
178           int bcolor_slot,
179           LLVMValueRef attribv[3])
180{
181   LLVMBuilderRef b = gallivm->builder;
182   LLVMValueRef a0_back, a1_back, a2_back;
183   LLVMValueRef idx2 = lp_build_const_int32(gallivm, bcolor_slot);
184
185   LLVMValueRef facing = args->facing;
186   LLVMValueRef front_facing = LLVMBuildICmp(b, LLVMIntEQ, facing,
187                                             lp_build_const_int32(gallivm, 0), ""); /** need i1 for if condition */
188
189   a0_back = LLVMBuildLoad(b, LLVMBuildGEP(b, args->v0, &idx2, 1, ""), "v0a_back");
190   a1_back = LLVMBuildLoad(b, LLVMBuildGEP(b, args->v1, &idx2, 1, ""), "v1a_back");
191   a2_back = LLVMBuildLoad(b, LLVMBuildGEP(b, args->v2, &idx2, 1, ""), "v2a_back");
192
193   /* Possibly swap the front and back attrib values,
194    *
195    * Prefer select to if so we don't have to worry about phis or
196    * allocas.
197    */
198   attribv[0] = LLVMBuildSelect(b, front_facing, a0_back, attribv[0], "");
199   attribv[1] = LLVMBuildSelect(b, front_facing, a1_back, attribv[1], "");
200   attribv[2] = LLVMBuildSelect(b, front_facing, a2_back, attribv[2], "");
201
202}
203
204static LLVMValueRef
205lp_do_offset_tri(struct gallivm_state *gallivm,
206                 struct lp_setup_args *args,
207                 const struct lp_setup_variant_key *key,
208                 LLVMValueRef inv_det,
209                 LLVMValueRef dxyz01,
210                 LLVMValueRef dxyz20,
211                 LLVMValueRef attribv[3])
212{
213   LLVMBuilderRef b = gallivm->builder;
214   struct lp_build_context flt_scalar_bld;
215   struct lp_build_context int_scalar_bld;
216   struct lp_build_context *bld = &args->bld;
217   LLVMValueRef zoffset, mult;
218   LLVMValueRef dzdxdzdy, dzdx, dzdy, dzxyz20, dyzzx01, dyzzx01_dzxyz20, dzx01_dyz20;
219   LLVMValueRef max, max_value, res12;
220   LLVMValueRef shuffles[4];
221   LLVMTypeRef shuf_type = LLVMInt32TypeInContext(gallivm->context);
222   LLVMValueRef onei = lp_build_const_int32(gallivm, 1);
223   LLVMValueRef zeroi = lp_build_const_int32(gallivm, 0);
224   LLVMValueRef twoi = lp_build_const_int32(gallivm, 2);
225   LLVMValueRef threei  = lp_build_const_int32(gallivm, 3);
226
227   /* (res12) = cross(e,f).xy */
228   shuffles[0] = twoi;
229   shuffles[1] = zeroi;
230   shuffles[2] = onei;
231   shuffles[3] = twoi;
232   dzxyz20 = LLVMBuildShuffleVector(b, dxyz20, dxyz20, LLVMConstVector(shuffles, 4), "");
233
234   shuffles[0] = onei;
235   shuffles[1] = twoi;
236   shuffles[2] = twoi;
237   shuffles[3] = zeroi;
238   dyzzx01 = LLVMBuildShuffleVector(b, dxyz01, dxyz01, LLVMConstVector(shuffles, 4), "");
239
240   dyzzx01_dzxyz20 = LLVMBuildFMul(b, dzxyz20, dyzzx01, "dyzzx01_dzxyz20");
241
242   shuffles[0] = twoi;
243   shuffles[1] = threei;
244   shuffles[2] = LLVMGetUndef(shuf_type);
245   shuffles[3] = LLVMGetUndef(shuf_type);
246   dzx01_dyz20 = LLVMBuildShuffleVector(b, dyzzx01_dzxyz20, dyzzx01_dzxyz20,
247                                        LLVMConstVector(shuffles, 4), "");
248
249   res12 = LLVMBuildFSub(b, dyzzx01_dzxyz20, dzx01_dyz20, "res12");
250
251   /* dzdx = fabsf(res1 * inv_det), dydx = fabsf(res2 * inv_det)*/
252   dzdxdzdy = LLVMBuildFMul(b, res12, inv_det, "dzdxdzdy");
253   dzdxdzdy = lp_build_abs(bld, dzdxdzdy);
254
255   dzdx = LLVMBuildExtractElement(b, dzdxdzdy, zeroi, "");
256   dzdy = LLVMBuildExtractElement(b, dzdxdzdy, onei, "");
257
258   /* mult = MAX2(dzdx, dzdy) * pgon_offset_scale */
259   max = LLVMBuildFCmp(b, LLVMRealUGT, dzdx, dzdy, "");
260   max_value = LLVMBuildSelect(b, max, dzdx, dzdy, "max");
261
262   mult = LLVMBuildFMul(b, max_value,
263                        lp_build_const_float(gallivm, key->pgon_offset_scale), "");
264
265   lp_build_context_init(&flt_scalar_bld, gallivm, lp_type_float_vec(32, 32));
266
267   if (key->floating_point_depth) {
268      /*
269       * bias = pgon_offset_units * 2^(exponent(max(abs(z0), abs(z1), abs(z2))) -
270       *           mantissa_bits) + MAX2(dzdx, dzdy) * pgon_offset_scale
271       *
272       * NOTE: Assumes IEEE float32.
273       */
274      LLVMValueRef c23_shifted, exp_mask, bias, exp;
275      LLVMValueRef maxz_value, maxz0z1_value;
276
277      lp_build_context_init(&int_scalar_bld, gallivm, lp_type_int_vec(32, 32));
278
279      c23_shifted = lp_build_const_int32(gallivm, 23 << 23);
280      exp_mask = lp_build_const_int32(gallivm, 0xff << 23);
281
282      maxz0z1_value = lp_build_max(&flt_scalar_bld,
283                         lp_build_abs(&flt_scalar_bld,
284                            LLVMBuildExtractElement(b, attribv[0], twoi, "")),
285                         lp_build_abs(&flt_scalar_bld,
286                            LLVMBuildExtractElement(b, attribv[1], twoi, "")));
287
288      maxz_value = lp_build_max(&flt_scalar_bld,
289                      lp_build_abs(&flt_scalar_bld,
290                         LLVMBuildExtractElement(b, attribv[2], twoi, "")),
291                      maxz0z1_value);
292
293      exp = LLVMBuildBitCast(b, maxz_value, int_scalar_bld.vec_type, "");
294      exp = lp_build_and(&int_scalar_bld, exp, exp_mask);
295      exp = lp_build_sub(&int_scalar_bld, exp, c23_shifted);
296      /* Clamping to zero means mrd will be zero for very small numbers,
297       * but specs do not indicate this should be prevented by clamping
298       * mrd to smallest normal number instead. */
299      exp = lp_build_max(&int_scalar_bld, exp, int_scalar_bld.zero);
300      exp = LLVMBuildBitCast(b, exp, flt_scalar_bld.vec_type, "");
301
302      bias = LLVMBuildFMul(b, exp,
303                           lp_build_const_float(gallivm, key->pgon_offset_units),
304                           "bias");
305
306      zoffset = LLVMBuildFAdd(b, bias, mult, "zoffset");
307   } else {
308      /*
309       * bias = pgon_offset_units + MAX2(dzdx, dzdy) * pgon_offset_scale
310       */
311      zoffset = LLVMBuildFAdd(b,
312                              lp_build_const_float(gallivm, key->pgon_offset_units),
313                              mult, "zoffset");
314   }
315
316   if (key->pgon_offset_clamp > 0) {
317      zoffset = lp_build_min(&flt_scalar_bld,
318                             lp_build_const_float(gallivm, key->pgon_offset_clamp),
319                             zoffset);
320   }
321   else if (key->pgon_offset_clamp < 0) {
322      zoffset = lp_build_max(&flt_scalar_bld,
323                             lp_build_const_float(gallivm, key->pgon_offset_clamp),
324                             zoffset);
325   }
326
327   return zoffset;
328}
329
330static void
331load_attribute(struct gallivm_state *gallivm,
332               struct lp_setup_args *args,
333               const struct lp_setup_variant_key *key,
334               unsigned vert_attr,
335               LLVMValueRef attribv[3])
336{
337   LLVMBuilderRef b = gallivm->builder;
338   LLVMValueRef idx = lp_build_const_int32(gallivm, vert_attr);
339
340   /* Load the vertex data
341    */
342   attribv[0] = LLVMBuildLoad(b, LLVMBuildGEP(b, args->v0, &idx, 1, ""), "v0a");
343   attribv[1] = LLVMBuildLoad(b, LLVMBuildGEP(b, args->v1, &idx, 1, ""), "v1a");
344   attribv[2] = LLVMBuildLoad(b, LLVMBuildGEP(b, args->v2, &idx, 1, ""), "v2a");
345
346
347   /* Potentially modify it according to twoside, etc:
348    */
349   if (key->twoside) {
350      if (vert_attr == key->color_slot && key->bcolor_slot >= 0)
351         lp_twoside(gallivm, args, key, key->bcolor_slot, attribv);
352      else if (vert_attr == key->spec_slot && key->bspec_slot >= 0)
353         lp_twoside(gallivm, args, key, key->bspec_slot, attribv);
354   }
355}
356
357/*
358 * FIXME: interpolation is always done wrt fb origin (0/0).
359 * However, if some (small) tri is far away from the origin and gradients
360 * are large, this can lead to HUGE errors, since the a0 value calculated
361 * here can get very large (with the actual values inside the triangle way
362 * smaller), leading to complete loss of accuracy. This could be prevented
363 * by using some point inside (or at corner) of the tri as interpolation
364 * origin, or just use barycentric interpolation (which GL suggests and is
365 * what real hw does - you can get the barycentric coordinates from the
366 * edge functions in rasterization in principle (though we skip these
367 * sometimes completely in case of tris covering a block fully,
368 * which obviously wouldn't work)).
369 */
370static void
371calc_coef4( struct gallivm_state *gallivm,
372            struct lp_setup_args *args,
373            LLVMValueRef a0,
374            LLVMValueRef a1,
375            LLVMValueRef a2,
376            LLVMValueRef out[3])
377{
378   LLVMBuilderRef b = gallivm->builder;
379   LLVMValueRef attr_0;
380   LLVMValueRef dy20_ooa = args->dy20_ooa;
381   LLVMValueRef dy01_ooa = args->dy01_ooa;
382   LLVMValueRef dx20_ooa = args->dx20_ooa;
383   LLVMValueRef dx01_ooa = args->dx01_ooa;
384   LLVMValueRef x0_center = args->x0_center;
385   LLVMValueRef y0_center = args->y0_center;
386   LLVMValueRef da01 = LLVMBuildFSub(b, a0, a1, "da01");
387   LLVMValueRef da20 = LLVMBuildFSub(b, a2, a0, "da20");
388
389   /* Calculate dadx (vec4f)
390    */
391   LLVMValueRef da01_dy20_ooa = LLVMBuildFMul(b, da01, dy20_ooa, "da01_dy20_ooa");
392   LLVMValueRef da20_dy01_ooa = LLVMBuildFMul(b, da20, dy01_ooa, "da20_dy01_ooa");
393   LLVMValueRef dadx          = LLVMBuildFSub(b, da01_dy20_ooa, da20_dy01_ooa, "dadx");
394
395   /* Calculate dady (vec4f)
396    */
397   LLVMValueRef da01_dx20_ooa = LLVMBuildFMul(b, da01, dx20_ooa, "da01_dx20_ooa");
398   LLVMValueRef da20_dx01_ooa = LLVMBuildFMul(b, da20, dx01_ooa, "da20_dx01_ooa");
399   LLVMValueRef dady          = LLVMBuildFSub(b, da20_dx01_ooa, da01_dx20_ooa, "dady");
400
401   /* Calculate a0 - the attribute value at the origin
402    */
403   LLVMValueRef dadx_x0    = LLVMBuildFMul(b, dadx, x0_center, "dadx_x0");
404   LLVMValueRef dady_y0    = LLVMBuildFMul(b, dady, y0_center, "dady_y0");
405   LLVMValueRef attr_v0    = LLVMBuildFAdd(b, dadx_x0, dady_y0, "attr_v0");
406   attr_0                  = LLVMBuildFSub(b, a0, attr_v0, "attr_0");
407
408   out[0] = attr_0;
409   out[1] = dadx;
410   out[2] = dady;
411}
412
413static void
414emit_coef4( struct gallivm_state *gallivm,
415            struct lp_setup_args *args,
416            unsigned slot,
417            LLVMValueRef a0,
418            LLVMValueRef a1,
419            LLVMValueRef a2)
420{
421   LLVMValueRef coeffs[3];
422   calc_coef4(gallivm, args, a0, a1, a2, coeffs);
423   store_coef(gallivm, args, slot,
424              coeffs[0], coeffs[1], coeffs[2]);
425}
426
427
428static void
429emit_linear_coef( struct gallivm_state *gallivm,
430                  struct lp_setup_args *args,
431                  unsigned slot,
432                  LLVMValueRef attribv[3])
433{
434   /* nothing to do anymore */
435   emit_coef4(gallivm,
436              args, slot,
437              attribv[0],
438              attribv[1],
439              attribv[2]);
440}
441
442
443/**
444 * Compute a0, dadx and dady for a perspective-corrected interpolant,
445 * for a triangle.
446 * We basically multiply the vertex value by 1/w before computing
447 * the plane coefficients (a0, dadx, dady).
448 * Later, when we compute the value at a particular fragment position we'll
449 * divide the interpolated value by the interpolated W at that fragment.
450 */
451static void
452apply_perspective_corr( struct gallivm_state *gallivm,
453                        struct lp_setup_args *args,
454                        unsigned slot,
455                        LLVMValueRef attribv[3])
456{
457   LLVMBuilderRef b = gallivm->builder;
458
459   /* premultiply by 1/w  (v[0][3] is always 1/w):
460    */
461   LLVMValueRef v0_oow = lp_build_broadcast_scalar(&args->bld,
462                            vert_attrib(gallivm, args->v0, 0, 3, "v0_oow"));
463   LLVMValueRef v1_oow = lp_build_broadcast_scalar(&args->bld,
464                            vert_attrib(gallivm, args->v1, 0, 3, "v1_oow"));
465   LLVMValueRef v2_oow = lp_build_broadcast_scalar(&args->bld,
466                            vert_attrib(gallivm, args->v2, 0, 3, "v2_oow"));
467
468   attribv[0] = LLVMBuildFMul(b, attribv[0], v0_oow, "v0_oow_v0a");
469   attribv[1] = LLVMBuildFMul(b, attribv[1], v1_oow, "v1_oow_v1a");
470   attribv[2] = LLVMBuildFMul(b, attribv[2], v2_oow, "v2_oow_v2a");
471}
472
473
474/**
475 * Compute the inputs-> dadx, dady, a0 values.
476 */
477static void
478emit_tri_coef( struct gallivm_state *gallivm,
479               const struct lp_setup_variant_key *key,
480               struct lp_setup_args *args)
481{
482   unsigned slot;
483
484   LLVMValueRef attribs[3];
485
486  /* setup interpolation for all the remaining attributes:
487    */
488   for (slot = 0; slot < key->num_inputs; slot++) {
489      switch (key->inputs[slot].interp) {
490      case LP_INTERP_CONSTANT:
491         load_attribute(gallivm, args, key, key->inputs[slot].src_index, attribs);
492         if (key->flatshade_first) {
493            emit_constant_coef4(gallivm, args, slot+1, attribs[0]);
494         }
495         else {
496            emit_constant_coef4(gallivm, args, slot+1, attribs[2]);
497         }
498         break;
499
500      case LP_INTERP_LINEAR:
501         load_attribute(gallivm, args, key, key->inputs[slot].src_index, attribs);
502         emit_linear_coef(gallivm, args, slot+1, attribs);
503         break;
504
505      case LP_INTERP_PERSPECTIVE:
506         load_attribute(gallivm, args, key, key->inputs[slot].src_index, attribs);
507         apply_perspective_corr(gallivm, args, slot+1, attribs);
508         emit_linear_coef(gallivm, args, slot+1, attribs);
509         break;
510
511      case LP_INTERP_POSITION:
512         /*
513          * The generated pixel interpolators will pick up the coeffs from
514          * slot 0.
515          */
516         break;
517
518      case LP_INTERP_FACING:
519         emit_facing_coef(gallivm, args, slot+1);
520         break;
521
522      default:
523         assert(0);
524      }
525   }
526}
527
528
529/* XXX: generic code:
530 */
531static void
532set_noalias(LLVMBuilderRef builder,
533            LLVMValueRef function,
534            const LLVMTypeRef *arg_types,
535            int nr_args)
536{
537   int i;
538   for(i = 0; i < nr_args; ++i)
539      if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
540         lp_add_function_attr(function, i + 1, LP_FUNC_ATTR_NOALIAS);
541}
542
543static void
544init_args(struct gallivm_state *gallivm,
545          const struct lp_setup_variant_key *key,
546          struct lp_setup_args *args)
547{
548   LLVMBuilderRef b = gallivm->builder;
549   LLVMTypeRef shuf_type = LLVMInt32TypeInContext(gallivm->context);
550   LLVMValueRef onef = lp_build_const_float(gallivm, 1.0);
551   LLVMValueRef onei = lp_build_const_int32(gallivm, 1);
552   LLVMValueRef zeroi = lp_build_const_int32(gallivm, 0);
553   LLVMValueRef pixel_center, xy0_center, dxy01, dxy20, dyx20;
554   LLVMValueRef e, f, ef, ooa;
555   LLVMValueRef shuffles[4], shuf10;
556   LLVMValueRef attr_pos[3];
557   LLVMValueRef polygon_offset;
558   struct lp_type typef4 = lp_type_float_vec(32, 128);
559   struct lp_build_context bld;
560
561   lp_build_context_init(&bld, gallivm, typef4);
562   args->bld = bld;
563
564   /* The internal position input is in slot zero:
565    */
566   load_attribute(gallivm, args, key, 0, attr_pos);
567
568   pixel_center = lp_build_const_vec(gallivm, typef4,
569                                     (!key->multisample && key->pixel_center_half) ? 0.5 : 0.0);
570
571   /*
572    * xy are first two elems in v0a/v1a/v2a but just use vec4 arit
573    * also offset_tri uses actually xyz in them
574    */
575   xy0_center = LLVMBuildFSub(b, attr_pos[0], pixel_center, "xy0_center" );
576
577   dxy01 = LLVMBuildFSub(b, attr_pos[0], attr_pos[1], "dxy01");
578   dxy20 = LLVMBuildFSub(b, attr_pos[2], attr_pos[0], "dxy20");
579
580   shuffles[0] = onei;
581   shuffles[1] = zeroi;
582   shuffles[2] = LLVMGetUndef(shuf_type);
583   shuffles[3] = LLVMGetUndef(shuf_type);
584   shuf10 = LLVMConstVector(shuffles, 4);
585
586   dyx20 = LLVMBuildShuffleVector(b, dxy20, dxy20, shuf10, "");
587
588   ef = LLVMBuildFMul(b, dxy01, dyx20, "ef");
589   e = LLVMBuildExtractElement(b, ef, zeroi, "");
590   f = LLVMBuildExtractElement(b, ef, onei, "");
591
592   ooa  = LLVMBuildFDiv(b, onef, LLVMBuildFSub(b, e, f, ""), "ooa");
593
594   ooa = lp_build_broadcast_scalar(&bld, ooa);
595
596   /* tri offset calc shares a lot of arithmetic, do it here */
597   if (key->pgon_offset_scale != 0.0f || key->pgon_offset_units != 0.0f) {
598      polygon_offset = lp_do_offset_tri(gallivm, args, key, ooa, dxy01, dxy20, attr_pos);
599   } else {
600      polygon_offset = lp_build_const_float(gallivm, 0.0f);
601   }
602
603   dxy20 = LLVMBuildFMul(b, dxy20, ooa, "");
604   dxy01 = LLVMBuildFMul(b, dxy01, ooa, "");
605
606   args->dy20_ooa  = lp_build_extract_broadcast(gallivm, typef4, typef4, dxy20, onei);
607   args->dy01_ooa  = lp_build_extract_broadcast(gallivm, typef4, typef4, dxy01, onei);
608
609   args->dx20_ooa  = lp_build_extract_broadcast(gallivm, typef4, typef4, dxy20, zeroi);
610   args->dx01_ooa  = lp_build_extract_broadcast(gallivm, typef4, typef4, dxy01, zeroi);
611
612   args->x0_center = lp_build_extract_broadcast(gallivm, typef4, typef4, xy0_center, zeroi);
613   args->y0_center = lp_build_extract_broadcast(gallivm, typef4, typef4, xy0_center, onei);
614
615   LLVMValueRef coeffs[3];
616   calc_coef4(gallivm, args,
617              attr_pos[0], attr_pos[1], attr_pos[2],
618              coeffs);
619
620   /* This is a bit sneaky:
621    * Because we observe that the X component of A0 is otherwise unused,
622    * we can overwrite it with the computed polygon-offset value, to make
623    * sure it's available in the fragment shader without having to change
624    * the interface (which is error-prone).
625    */
626   coeffs[0] = LLVMBuildInsertElement(b, coeffs[0], polygon_offset,
627                                      lp_build_const_int32(gallivm, 0), "");
628
629   store_coef(gallivm, args, 0,
630              coeffs[0], coeffs[1], coeffs[2]);
631}
632
633/**
634 * Generate the runtime callable function for the coefficient calculation.
635 *
636 */
637static struct lp_setup_variant *
638generate_setup_variant(struct lp_setup_variant_key *key,
639                       struct llvmpipe_context *lp)
640{
641   struct lp_setup_variant *variant = NULL;
642   struct gallivm_state *gallivm;
643   struct lp_setup_args args;
644   char func_name[64];
645   LLVMTypeRef vec4f_type;
646   LLVMTypeRef func_type;
647   LLVMTypeRef arg_types[8];
648   LLVMBasicBlockRef block;
649   LLVMBuilderRef builder;
650   int64_t t0 = 0, t1;
651
652   if (0)
653      goto fail;
654
655   variant = CALLOC_STRUCT(lp_setup_variant);
656   if (!variant)
657      goto fail;
658
659   variant->no = setup_no++;
660
661   snprintf(func_name, sizeof(func_name), "setup_variant_%u",
662            variant->no);
663
664   variant->gallivm = gallivm = gallivm_create(func_name, lp->context, NULL);
665   if (!variant->gallivm) {
666      goto fail;
667   }
668
669   builder = gallivm->builder;
670
671   if (LP_DEBUG & DEBUG_COUNTERS) {
672      t0 = os_time_get();
673   }
674
675   memcpy(&variant->key, key, key->size);
676   variant->list_item_global.base = variant;
677
678   /* Currently always deal with full 4-wide vertex attributes from
679    * the vertices.
680    */
681
682   vec4f_type = LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4);
683
684   arg_types[0] = LLVMPointerType(vec4f_type, 0);        /* v0 */
685   arg_types[1] = LLVMPointerType(vec4f_type, 0);        /* v1 */
686   arg_types[2] = LLVMPointerType(vec4f_type, 0);        /* v2 */
687   arg_types[3] = LLVMInt32TypeInContext(gallivm->context); /* facing */
688   arg_types[4] = LLVMPointerType(vec4f_type, 0);	/* a0, aligned */
689   arg_types[5] = LLVMPointerType(vec4f_type, 0);	/* dadx, aligned */
690   arg_types[6] = LLVMPointerType(vec4f_type, 0);	/* dady, aligned */
691   arg_types[7] = LLVMPointerType(vec4f_type, 0);	/* key (placeholder) */
692
693   func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
694                                arg_types, ARRAY_SIZE(arg_types), 0);
695
696   variant->function = LLVMAddFunction(gallivm->module, func_name, func_type);
697   if (!variant->function)
698      goto fail;
699
700   LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
701
702   args.v0       = LLVMGetParam(variant->function, 0);
703   args.v1       = LLVMGetParam(variant->function, 1);
704   args.v2       = LLVMGetParam(variant->function, 2);
705   args.facing   = LLVMGetParam(variant->function, 3);
706   args.a0       = LLVMGetParam(variant->function, 4);
707   args.dadx     = LLVMGetParam(variant->function, 5);
708   args.dady     = LLVMGetParam(variant->function, 6);
709   args.key      = LLVMGetParam(variant->function, 7);
710
711   lp_build_name(args.v0, "in_v0");
712   lp_build_name(args.v1, "in_v1");
713   lp_build_name(args.v2, "in_v2");
714   lp_build_name(args.facing, "in_facing");
715   lp_build_name(args.a0, "out_a0");
716   lp_build_name(args.dadx, "out_dadx");
717   lp_build_name(args.dady, "out_dady");
718   lp_build_name(args.key, "key");
719
720   /*
721    * Function body
722    */
723   block = LLVMAppendBasicBlockInContext(gallivm->context,
724                                         variant->function, "entry");
725   LLVMPositionBuilderAtEnd(builder, block);
726
727   set_noalias(builder, variant->function, arg_types, ARRAY_SIZE(arg_types));
728   init_args(gallivm, &variant->key, &args);
729   emit_tri_coef(gallivm, &variant->key, &args);
730
731   LLVMBuildRetVoid(builder);
732
733   gallivm_verify_function(gallivm, variant->function);
734
735   gallivm_compile_module(gallivm);
736
737   variant->jit_function = (lp_jit_setup_triangle)
738      gallivm_jit_function(gallivm, variant->function);
739   if (!variant->jit_function)
740      goto fail;
741
742   gallivm_free_ir(variant->gallivm);
743
744   /*
745    * Update timing information:
746    */
747   if (LP_DEBUG & DEBUG_COUNTERS) {
748      t1 = os_time_get();
749      LP_COUNT_ADD(llvm_compile_time, t1 - t0);
750      LP_COUNT_ADD(nr_llvm_compiles, 1);
751   }
752
753   return variant;
754
755fail:
756   if (variant) {
757      if (variant->gallivm) {
758         gallivm_destroy(variant->gallivm);
759      }
760      FREE(variant);
761   }
762
763   return NULL;
764}
765
766
767
768static void
769lp_make_setup_variant_key(struct llvmpipe_context *lp,
770                          struct lp_setup_variant_key *key)
771{
772   struct lp_fragment_shader *fs = lp->fs;
773   unsigned i;
774
775   assert(sizeof key->inputs[0] == sizeof(uint));
776
777   key->num_inputs = fs->info.base.num_inputs;
778   key->flatshade_first = lp->rasterizer->flatshade_first;
779   key->pixel_center_half = lp->rasterizer->half_pixel_center;
780   key->multisample = lp->rasterizer->multisample;
781   key->twoside = lp->rasterizer->light_twoside;
782   key->size = Offset(struct lp_setup_variant_key,
783                      inputs[key->num_inputs]);
784
785   key->color_slot = lp->color_slot[0];
786   key->bcolor_slot = lp->bcolor_slot[0];
787   key->spec_slot = lp->color_slot[1];
788   key->bspec_slot = lp->bcolor_slot[1];
789
790   /*
791    * If depth is floating point, depth bias is calculated with respect
792    * to the primitive's maximum Z value. Retain the original depth bias
793    * value until that stage.
794    */
795   key->floating_point_depth = lp->floating_point_depth;
796
797   if (key->floating_point_depth) {
798      key->pgon_offset_units = (float) lp->rasterizer->offset_units;
799   } else {
800      key->pgon_offset_units =
801         (float) (lp->rasterizer->offset_units * lp->mrd * 2);
802   }
803
804   key->pgon_offset_scale = lp->rasterizer->offset_scale;
805   key->pgon_offset_clamp = lp->rasterizer->offset_clamp;
806   key->uses_constant_interp = 0;
807   key->pad = 0;
808   memcpy(key->inputs, fs->inputs, key->num_inputs * sizeof key->inputs[0]);
809   for (i = 0; i < key->num_inputs; i++) {
810      if (key->inputs[i].interp == LP_INTERP_COLOR) {
811         if (lp->rasterizer->flatshade)
812            key->inputs[i].interp = LP_INTERP_CONSTANT;
813         else
814            key->inputs[i].interp = LP_INTERP_PERSPECTIVE;
815      }
816      if (key->inputs[i].interp == LP_INTERP_CONSTANT) {
817         key->uses_constant_interp = 1;
818      }
819   }
820}
821
822
823static void
824remove_setup_variant(struct llvmpipe_context *lp,
825                     struct lp_setup_variant *variant)
826{
827   if (gallivm_debug & GALLIVM_DEBUG_IR) {
828      debug_printf("llvmpipe: del setup_variant #%u total %u\n",
829                   variant->no, lp->nr_setup_variants);
830   }
831
832   if (variant->gallivm) {
833      gallivm_destroy(variant->gallivm);
834   }
835
836   remove_from_list(&variant->list_item_global);
837   lp->nr_setup_variants--;
838   FREE(variant);
839}
840
841
842
843/* When the number of setup variants exceeds a threshold, cull a
844 * fraction (currently a quarter) of them.
845 */
846static void
847cull_setup_variants(struct llvmpipe_context *lp)
848{
849   struct pipe_context *pipe = &lp->pipe;
850   int i;
851
852   /*
853    * XXX: we need to flush the context until we have some sort of reference
854    * counting in fragment shaders as they may still be binned
855    * Flushing alone might not be sufficient we need to wait on it too.
856    */
857   llvmpipe_finish(pipe, __FUNCTION__);
858
859   for (i = 0; i < LP_MAX_SETUP_VARIANTS / 4; i++) {
860      struct lp_setup_variant_list_item *item;
861      if (is_empty_list(&lp->setup_variants_list)) {
862         break;
863      }
864      item = last_elem(&lp->setup_variants_list);
865      assert(item);
866      assert(item->base);
867      remove_setup_variant(lp, item->base);
868   }
869}
870
871
872/**
873 * Update fragment/vertex shader linkage state.  This is called just
874 * prior to drawing something when some fragment-related state has
875 * changed.
876 */
877void
878llvmpipe_update_setup(struct llvmpipe_context *lp)
879{
880   struct lp_setup_variant_key *key = &lp->setup_variant.key;
881   struct lp_setup_variant *variant = NULL;
882   struct lp_setup_variant_list_item *li;
883
884   lp_make_setup_variant_key(lp, key);
885
886   foreach(li, &lp->setup_variants_list) {
887      if(li->base->key.size == key->size &&
888         memcmp(&li->base->key, key, key->size) == 0) {
889         variant = li->base;
890         break;
891      }
892   }
893
894   if (variant) {
895      move_to_head(&lp->setup_variants_list, &variant->list_item_global);
896   }
897   else {
898      if (lp->nr_setup_variants >= LP_MAX_SETUP_VARIANTS) {
899         cull_setup_variants(lp);
900      }
901
902      variant = generate_setup_variant(key, lp);
903      if (variant) {
904         insert_at_head(&lp->setup_variants_list, &variant->list_item_global);
905         lp->nr_setup_variants++;
906      }
907   }
908
909   lp_setup_set_setup_variant(lp->setup, variant);
910}
911
912void
913lp_delete_setup_variants(struct llvmpipe_context *lp)
914{
915   struct lp_setup_variant_list_item *li;
916   li = first_elem(&lp->setup_variants_list);
917   while(!at_end(&lp->setup_variants_list, li)) {
918      struct lp_setup_variant_list_item *next = next_elem(li);
919      remove_setup_variant(lp, li->base);
920      li = next;
921   }
922}
923
924void
925lp_dump_setup_coef(const struct lp_setup_variant_key *key,
926                   const float (*sa0)[4],
927                   const float (*sdadx)[4],
928                   const float (*sdady)[4])
929{
930   int i, slot;
931
932   for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
933      float a0   = sa0  [0][i];
934      float dadx = sdadx[0][i];
935      float dady = sdady[0][i];
936
937      debug_printf("POS.%c: a0 = %f, dadx = %f, dady = %f\n",
938                   "xyzw"[i], a0, dadx, dady);
939   }
940
941   for (slot = 0; slot < key->num_inputs; slot++) {
942      unsigned usage_mask = key->inputs[slot].usage_mask;
943      for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
944         if (usage_mask & (1 << i)) {
945            float a0   = sa0  [1 + slot][i];
946            float dadx = sdadx[1 + slot][i];
947            float dady = sdady[1 + slot][i];
948
949            debug_printf("IN[%u].%c: a0 = %f, dadx = %f, dady = %f\n",
950                         slot, "xyzw"[i], a0, dadx, dady);
951         }
952      }
953   }
954}
955