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_const.h"
35 #include "gallivm/lp_bld_debug.h"
36 #include "gallivm/lp_bld_init.h"
37 #include "gallivm/lp_bld_intr.h"
38 #include "gallivm/lp_bld_flow.h"
39 #include "gallivm/lp_bld_type.h"
40 #include <llvm-c/Analysis.h> /* for LLVMVerifyFunction */
45 #include "lp_screen.h"
46 #include "lp_context.h"
48 #include "lp_state_fs.h"
49 #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
;
80 /* Temporary, per-attribute:
90 type4f(struct gallivm_state
*gallivm
)
92 return LLVMVectorType(LLVMFloatTypeInContext(gallivm
->context
), 4);
96 /* Equivalent of _mm_setr_ps(a,b,c,d)
99 vec4f(struct gallivm_state
*gallivm
,
100 LLVMValueRef a
, LLVMValueRef b
, LLVMValueRef c
, LLVMValueRef d
,
103 LLVMBuilderRef bld
= gallivm
->builder
;
104 LLVMValueRef i0
= lp_build_const_int32(gallivm
, 0);
105 LLVMValueRef i1
= lp_build_const_int32(gallivm
, 1);
106 LLVMValueRef i2
= lp_build_const_int32(gallivm
, 2);
107 LLVMValueRef i3
= lp_build_const_int32(gallivm
, 3);
109 LLVMValueRef res
= LLVMGetUndef(type4f(gallivm
));
111 res
= LLVMBuildInsertElement(bld
, res
, a
, i0
, "");
112 res
= LLVMBuildInsertElement(bld
, res
, b
, i1
, "");
113 res
= LLVMBuildInsertElement(bld
, res
, c
, i2
, "");
114 res
= LLVMBuildInsertElement(bld
, res
, d
, i3
, name
);
119 /* Equivalent of _mm_set1_ps(a)
122 vec4f_from_scalar(struct gallivm_state
*gallivm
,
126 LLVMBuilderRef bld
= gallivm
->builder
;
127 LLVMValueRef res
= LLVMGetUndef(type4f(gallivm
));
130 for(i
= 0; i
< 4; ++i
) {
131 LLVMValueRef index
= lp_build_const_int32(gallivm
, i
);
132 res
= LLVMBuildInsertElement(bld
, res
, a
, index
, i
== 3 ? name
: "");
139 store_coef(struct gallivm_state
*gallivm
,
140 struct lp_setup_args
*args
,
146 LLVMBuilderRef builder
= gallivm
->builder
;
147 LLVMValueRef idx
= lp_build_const_int32(gallivm
, slot
);
149 LLVMBuildStore(builder
,
151 LLVMBuildGEP(builder
, args
->a0
, &idx
, 1, ""));
153 LLVMBuildStore(builder
,
155 LLVMBuildGEP(builder
, args
->dadx
, &idx
, 1, ""));
157 LLVMBuildStore(builder
,
159 LLVMBuildGEP(builder
, args
->dady
, &idx
, 1, ""));
165 emit_constant_coef4(struct gallivm_state
*gallivm
,
166 struct lp_setup_args
*args
,
170 LLVMValueRef zero
= lp_build_const_float(gallivm
, 0.0);
171 LLVMValueRef zerovec
= vec4f_from_scalar(gallivm
, zero
, "zero");
172 store_coef(gallivm
, args
, slot
, vert
, zerovec
, zerovec
);
178 * Setup the fragment input attribute with the front-facing value.
179 * \param frontface is the triangle front facing?
182 emit_facing_coef(struct gallivm_state
*gallivm
,
183 struct lp_setup_args
*args
,
186 LLVMBuilderRef builder
= gallivm
->builder
;
187 LLVMTypeRef float_type
= LLVMFloatTypeInContext(gallivm
->context
);
188 LLVMValueRef a0_0
= args
->facing
;
189 LLVMValueRef a0_0f
= LLVMBuildSIToFP(builder
, a0_0
, float_type
, "");
190 LLVMValueRef zero
= lp_build_const_float(gallivm
, 0.0);
191 LLVMValueRef a0
= vec4f(gallivm
, a0_0f
, zero
, zero
, zero
, "facing");
192 LLVMValueRef zerovec
= vec4f_from_scalar(gallivm
, zero
, "zero");
194 store_coef(gallivm
, args
, slot
, a0
, zerovec
, zerovec
);
199 vert_attrib(struct gallivm_state
*gallivm
,
205 LLVMBuilderRef b
= gallivm
->builder
;
207 idx
[0] = lp_build_const_int32(gallivm
, attr
);
208 idx
[1] = lp_build_const_int32(gallivm
, elem
);
209 return LLVMBuildLoad(b
, LLVMBuildGEP(b
, vert
, idx
, 2, ""), name
);
213 vert_clamp(LLVMBuilderRef b
,
218 LLVMValueRef min_result
= LLVMBuildFCmp(b
, LLVMRealUGT
, min
, x
, "");
219 LLVMValueRef max_result
= LLVMBuildFCmp(b
, LLVMRealUGT
, x
, max
, "");
220 LLVMValueRef clamp_value
;
222 clamp_value
= LLVMBuildSelect(b
, min_result
, min
, x
, "");
223 clamp_value
= LLVMBuildSelect(b
, max_result
, max
, x
, "");
229 lp_twoside(struct gallivm_state
*gallivm
,
230 struct lp_setup_args
*args
,
231 const struct lp_setup_variant_key
*key
,
234 LLVMBuilderRef b
= gallivm
->builder
;
235 LLVMValueRef a0_back
, a1_back
, a2_back
;
236 LLVMValueRef idx2
= lp_build_const_int32(gallivm
, bcolor_slot
);
238 LLVMValueRef facing
= args
->facing
;
239 LLVMValueRef front_facing
= LLVMBuildICmp(b
, LLVMIntEQ
, facing
, lp_build_const_int32(gallivm
, 0), ""); /** need i1 for if condition */
241 a0_back
= LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v0
, &idx2
, 1, ""), "v0a_back");
242 a1_back
= LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v1
, &idx2
, 1, ""), "v1a_back");
243 a2_back
= LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v2
, &idx2
, 1, ""), "v2a_back");
245 /* Possibly swap the front and back attrib values,
247 * Prefer select to if so we don't have to worry about phis or
250 args
->v0a
= LLVMBuildSelect(b
, front_facing
, a0_back
, args
->v0a
, "");
251 args
->v1a
= LLVMBuildSelect(b
, front_facing
, a1_back
, args
->v1a
, "");
252 args
->v2a
= LLVMBuildSelect(b
, front_facing
, a2_back
, args
->v2a
, "");
257 lp_do_offset_tri(struct gallivm_state
*gallivm
,
258 struct lp_setup_args
*args
,
259 const struct lp_setup_variant_key
*key
)
261 LLVMBuilderRef b
= gallivm
->builder
;
262 struct lp_build_context bld
;
263 LLVMValueRef zoffset
, mult
;
264 LLVMValueRef z0_new
, z1_new
, z2_new
;
265 LLVMValueRef dzdx0
, dzdx
, dzdy0
, dzdy
;
266 LLVMValueRef max
, max_value
;
268 LLVMValueRef one
= lp_build_const_float(gallivm
, 1.0);
269 LLVMValueRef zero
= lp_build_const_float(gallivm
, 0.0);
270 LLVMValueRef two
= lp_build_const_int32(gallivm
, 2);
272 /* edge vectors: e = v0 - v2, f = v1 - v2 */
273 LLVMValueRef v0_x
= vert_attrib(gallivm
, args
->v0
, 0, 0, "v0_x");
274 LLVMValueRef v1_x
= vert_attrib(gallivm
, args
->v1
, 0, 0, "v1_x");
275 LLVMValueRef v2_x
= vert_attrib(gallivm
, args
->v2
, 0, 0, "v2_x");
276 LLVMValueRef v0_y
= vert_attrib(gallivm
, args
->v0
, 0, 1, "v0_y");
277 LLVMValueRef v1_y
= vert_attrib(gallivm
, args
->v1
, 0, 1, "v1_y");
278 LLVMValueRef v2_y
= vert_attrib(gallivm
, args
->v2
, 0, 1, "v2_y");
279 LLVMValueRef v0_z
= vert_attrib(gallivm
, args
->v0
, 0, 2, "v0_z");
280 LLVMValueRef v1_z
= vert_attrib(gallivm
, args
->v1
, 0, 2, "v1_z");
281 LLVMValueRef v2_z
= vert_attrib(gallivm
, args
->v2
, 0, 2, "v2_z");
283 /* edge vectors: e = v0 - v2, f = v1 - v2 */
284 LLVMValueRef dx02
= LLVMBuildFSub(b
, v0_x
, v2_x
, "dx02");
285 LLVMValueRef dy02
= LLVMBuildFSub(b
, v0_y
, v2_y
, "dy02");
286 LLVMValueRef dz02
= LLVMBuildFSub(b
, v0_z
, v2_z
, "dz02");
287 LLVMValueRef dx12
= LLVMBuildFSub(b
, v1_x
, v2_x
, "dx12");
288 LLVMValueRef dy12
= LLVMBuildFSub(b
, v1_y
, v2_y
, "dy12");
289 LLVMValueRef dz12
= LLVMBuildFSub(b
, v1_z
, v2_z
, "dz12");
291 /* det = cross(e,f).z */
292 LLVMValueRef dx02_dy12
= LLVMBuildFMul(b
, dx02
, dy12
, "dx02_dy12");
293 LLVMValueRef dy02_dx12
= LLVMBuildFMul(b
, dy02
, dx12
, "dy02_dx12");
294 LLVMValueRef det
= LLVMBuildFSub(b
, dx02_dy12
, dy02_dx12
, "det");
295 LLVMValueRef inv_det
= LLVMBuildFDiv(b
, one
, det
, "inv_det");
297 /* (res1,res2) = cross(e,f).xy */
298 LLVMValueRef dy02_dz12
= LLVMBuildFMul(b
, dy02
, dz12
, "dy02_dz12");
299 LLVMValueRef dz02_dy12
= LLVMBuildFMul(b
, dz02
, dy12
, "dz02_dy12");
300 LLVMValueRef dz02_dx12
= LLVMBuildFMul(b
, dz02
, dx12
, "dz02_dx12");
301 LLVMValueRef dx02_dz12
= LLVMBuildFMul(b
, dx02
, dz12
, "dx02_dz12");
302 LLVMValueRef res1
= LLVMBuildFSub(b
, dy02_dz12
, dz02_dy12
, "res1");
303 LLVMValueRef res2
= LLVMBuildFSub(b
, dz02_dx12
, dx02_dz12
, "res2");
305 /* dzdx = fabsf(res1 * inv_det), dydx = fabsf(res2 * inv_det)*/
306 lp_build_context_init(&bld
, gallivm
, lp_type_float(32));
307 dzdx0
= LLVMBuildFMul(b
, res1
, inv_det
, "dzdx");
308 dzdx
= lp_build_abs(&bld
, dzdx0
);
309 dzdy0
= LLVMBuildFMul(b
, res2
, inv_det
, "dzdy");
310 dzdy
= lp_build_abs(&bld
, dzdy0
);
312 /* zoffset = offset->units + MAX2(dzdx, dzdy) * offset->scale */
313 max
= LLVMBuildFCmp(b
, LLVMRealUGT
, dzdx
, dzdy
, "");
314 max_value
= LLVMBuildSelect(b
, max
, dzdx
, dzdy
, "max");
316 mult
= LLVMBuildFMul(b
, max_value
, lp_build_const_float(gallivm
, key
->scale
), "");
317 zoffset
= LLVMBuildFAdd(b
, lp_build_const_float(gallivm
, key
->units
), mult
, "zoffset");
319 /* clamp and do offset */
320 z0_new
= vert_clamp(b
, LLVMBuildFAdd(b
, v0_z
, zoffset
, ""), zero
, one
);
321 z1_new
= vert_clamp(b
, LLVMBuildFAdd(b
, v1_z
, zoffset
, ""), zero
, one
);
322 z2_new
= vert_clamp(b
, LLVMBuildFAdd(b
, v2_z
, zoffset
, ""), zero
, one
);
324 /* insert into args->a0.z, a1.z, a2.z:
326 args
->v0a
= LLVMBuildInsertElement(b
, args
->v0a
, z0_new
, two
, "");
327 args
->v1a
= LLVMBuildInsertElement(b
, args
->v1a
, z1_new
, two
, "");
328 args
->v2a
= LLVMBuildInsertElement(b
, args
->v2a
, z2_new
, two
, "");
332 load_attribute(struct gallivm_state
*gallivm
,
333 struct lp_setup_args
*args
,
334 const struct lp_setup_variant_key
*key
,
337 LLVMBuilderRef b
= gallivm
->builder
;
338 LLVMValueRef idx
= lp_build_const_int32(gallivm
, vert_attr
);
340 /* Load the vertex data
342 args
->v0a
= LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v0
, &idx
, 1, ""), "v0a");
343 args
->v1a
= LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v1
, &idx
, 1, ""), "v1a");
344 args
->v2a
= LLVMBuildLoad(b
, LLVMBuildGEP(b
, args
->v2
, &idx
, 1, ""), "v2a");
347 /* Potentially modify it according to twoside, offset, etc:
349 if (vert_attr
== 0 && (key
->scale
!= 0.0f
|| key
->units
!= 0.0f
)) {
350 lp_do_offset_tri(gallivm
, args
, key
);
354 if (vert_attr
== key
->color_slot
&& key
->bcolor_slot
>= 0)
355 lp_twoside(gallivm
, args
, key
, key
->bcolor_slot
);
356 else if (vert_attr
== key
->spec_slot
&& key
->bspec_slot
>= 0)
357 lp_twoside(gallivm
, args
, key
, key
->bspec_slot
);
362 emit_coef4( struct gallivm_state
*gallivm
,
363 struct lp_setup_args
*args
,
369 LLVMBuilderRef b
= gallivm
->builder
;
370 LLVMValueRef dy20_ooa
= args
->dy20_ooa
;
371 LLVMValueRef dy01_ooa
= args
->dy01_ooa
;
372 LLVMValueRef dx20_ooa
= args
->dx20_ooa
;
373 LLVMValueRef dx01_ooa
= args
->dx01_ooa
;
374 LLVMValueRef x0_center
= args
->x0_center
;
375 LLVMValueRef y0_center
= args
->y0_center
;
377 /* XXX: using fsub, fmul on vector types -- does this work??
379 LLVMValueRef da01
= LLVMBuildFSub(b
, a0
, a1
, "da01");
380 LLVMValueRef da20
= LLVMBuildFSub(b
, a2
, a0
, "da20");
382 /* Calculate dadx (vec4f)
384 LLVMValueRef da01_dy20_ooa
= LLVMBuildFMul(b
, da01
, dy20_ooa
, "da01_dy20_ooa");
385 LLVMValueRef da20_dy01_ooa
= LLVMBuildFMul(b
, da20
, dy01_ooa
, "da20_dy01_ooa");
386 LLVMValueRef dadx
= LLVMBuildFSub(b
, da01_dy20_ooa
, da20_dy01_ooa
, "dadx");
388 /* Calculate dady (vec4f)
390 LLVMValueRef da01_dx20_ooa
= LLVMBuildFMul(b
, da01
, dx20_ooa
, "da01_dx20_ooa");
391 LLVMValueRef da20_dx01_ooa
= LLVMBuildFMul(b
, da20
, dx01_ooa
, "da20_dx01_ooa");
392 LLVMValueRef dady
= LLVMBuildFSub(b
, da20_dx01_ooa
, da01_dx20_ooa
, "dady");
394 /* Calculate a0 - the attribute value at the origin
396 LLVMValueRef dadx_x0
= LLVMBuildFMul(b
, dadx
, x0_center
, "dadx_x0");
397 LLVMValueRef dady_y0
= LLVMBuildFMul(b
, dady
, y0_center
, "dady_y0");
398 LLVMValueRef attr_v0
= LLVMBuildFAdd(b
, dadx_x0
, dady_y0
, "attr_v0");
399 LLVMValueRef attr_0
= LLVMBuildFSub(b
, a0
, attr_v0
, "attr_0");
401 store_coef(gallivm
, args
, slot
, attr_0
, dadx
, dady
);
406 emit_linear_coef( struct gallivm_state
*gallivm
,
407 struct lp_setup_args
*args
,
410 /* nothing to do anymore */
420 * Compute a0, dadx and dady for a perspective-corrected interpolant,
422 * We basically multiply the vertex value by 1/w before computing
423 * the plane coefficients (a0, dadx, dady).
424 * Later, when we compute the value at a particular fragment position we'll
425 * divide the interpolated value by the interpolated W at that fragment.
428 emit_perspective_coef( struct gallivm_state
*gallivm
,
429 struct lp_setup_args
*args
,
432 LLVMBuilderRef b
= gallivm
->builder
;
434 /* premultiply by 1/w (v[0][3] is always 1/w):
436 LLVMValueRef v0_oow
= vec4f_from_scalar(gallivm
, vert_attrib(gallivm
, args
->v0
, 0, 3, ""), "v0_oow");
437 LLVMValueRef v1_oow
= vec4f_from_scalar(gallivm
, vert_attrib(gallivm
, args
->v1
, 0, 3, ""), "v1_oow");
438 LLVMValueRef v2_oow
= vec4f_from_scalar(gallivm
, vert_attrib(gallivm
, args
->v2
, 0, 3, ""), "v2_oow");
440 LLVMValueRef v0_oow_v0a
= LLVMBuildFMul(b
, args
->v0a
, v0_oow
, "v0_oow_v0a");
441 LLVMValueRef v1_oow_v1a
= LLVMBuildFMul(b
, args
->v1a
, v1_oow
, "v1_oow_v1a");
442 LLVMValueRef v2_oow_v2a
= LLVMBuildFMul(b
, args
->v2a
, v2_oow
, "v2_oow_v2a");
444 emit_coef4(gallivm
, args
, slot
, v0_oow_v0a
, v1_oow_v1a
, v2_oow_v2a
);
449 emit_position_coef( struct gallivm_state
*gallivm
,
450 struct lp_setup_args
*args
,
453 emit_linear_coef(gallivm
, args
, slot
);
460 * Compute the inputs-> dadx, dady, a0 values.
463 emit_tri_coef( struct gallivm_state
*gallivm
,
464 const struct lp_setup_variant_key
*key
,
465 struct lp_setup_args
*args
)
469 /* The internal position input is in slot zero:
471 load_attribute(gallivm
, args
, key
, 0);
472 emit_position_coef(gallivm
, args
, 0);
474 /* setup interpolation for all the remaining attributes:
476 for (slot
= 0; slot
< key
->num_inputs
; slot
++) {
478 if (key
->inputs
[slot
].interp
== LP_INTERP_CONSTANT
||
479 key
->inputs
[slot
].interp
== LP_INTERP_LINEAR
||
480 key
->inputs
[slot
].interp
== LP_INTERP_PERSPECTIVE
)
481 load_attribute(gallivm
, args
, key
, key
->inputs
[slot
].src_index
);
483 switch (key
->inputs
[slot
].interp
) {
484 case LP_INTERP_CONSTANT
:
485 if (key
->flatshade_first
) {
486 emit_constant_coef4(gallivm
, args
, slot
+1, args
->v0a
);
489 emit_constant_coef4(gallivm
, args
, slot
+1, args
->v2a
);
493 case LP_INTERP_LINEAR
:
494 emit_linear_coef(gallivm
, args
, slot
+1);
497 case LP_INTERP_PERSPECTIVE
:
498 emit_perspective_coef(gallivm
, args
, slot
+1);
501 case LP_INTERP_POSITION
:
503 * The generated pixel interpolators will pick up the coeffs from
508 case LP_INTERP_FACING
:
509 emit_facing_coef(gallivm
, args
, slot
+1);
519 /* XXX: This is generic code, share with fs/vs codegen:
521 static lp_jit_setup_triangle
522 finalize_function(struct gallivm_state
*gallivm
,
523 LLVMBuilderRef builder
,
524 LLVMValueRef function
)
528 /* Verify the LLVM IR. If invalid, dump and abort */
530 if (LLVMVerifyFunction(function
, LLVMPrintMessageAction
)) {
532 lp_debug_dump_value(function
);
537 /* Apply optimizations to LLVM IR */
538 LLVMRunFunctionPassManager(gallivm
->passmgr
, function
);
540 if (gallivm_debug
& GALLIVM_DEBUG_IR
)
542 /* Print the LLVM IR to stderr */
543 lp_debug_dump_value(function
);
548 * Translate the LLVM IR into machine code.
550 f
= LLVMGetPointerToGlobal(gallivm
->engine
, function
);
552 if (gallivm_debug
& GALLIVM_DEBUG_ASM
)
557 lp_func_delete_body(function
);
559 return (lp_jit_setup_triangle
) pointer_to_func(f
);
562 /* XXX: Generic code:
565 lp_emit_emms(struct gallivm_state
*gallivm
)
568 /* Avoid corrupting the FPU stack on 32bit OSes. */
569 lp_build_intrinsic(gallivm
->builder
, "llvm.x86.mmx.emms",
570 LLVMVoidTypeInContext(gallivm
->context
), NULL
, 0);
575 /* XXX: generic code:
578 set_noalias(LLVMBuilderRef builder
,
579 LLVMValueRef function
,
580 const LLVMTypeRef
*arg_types
,
584 for(i
= 0; i
< Elements(arg_types
); ++i
)
585 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
)
586 LLVMAddAttribute(LLVMGetParam(function
, i
),
587 LLVMNoAliasAttribute
);
591 init_args(struct gallivm_state
*gallivm
,
592 struct lp_setup_args
*args
,
593 const struct lp_setup_variant
*variant
)
595 LLVMBuilderRef b
= gallivm
->builder
;
597 LLVMValueRef v0_x
= vert_attrib(gallivm
, args
->v0
, 0, 0, "v0_x");
598 LLVMValueRef v0_y
= vert_attrib(gallivm
, args
->v0
, 0, 1, "v0_y");
600 LLVMValueRef v1_x
= vert_attrib(gallivm
, args
->v1
, 0, 0, "v1_x");
601 LLVMValueRef v1_y
= vert_attrib(gallivm
, args
->v1
, 0, 1, "v1_y");
603 LLVMValueRef v2_x
= vert_attrib(gallivm
, args
->v2
, 0, 0, "v2_x");
604 LLVMValueRef v2_y
= vert_attrib(gallivm
, args
->v2
, 0, 1, "v2_y");
606 LLVMValueRef pixel_center
= lp_build_const_float(gallivm
,
607 variant
->key
.pixel_center_half
? 0.5 : 0);
609 LLVMValueRef x0_center
= LLVMBuildFSub(b
, v0_x
, pixel_center
, "x0_center" );
610 LLVMValueRef y0_center
= LLVMBuildFSub(b
, v0_y
, pixel_center
, "y0_center" );
612 LLVMValueRef dx01
= LLVMBuildFSub(b
, v0_x
, v1_x
, "dx01");
613 LLVMValueRef dy01
= LLVMBuildFSub(b
, v0_y
, v1_y
, "dy01");
614 LLVMValueRef dx20
= LLVMBuildFSub(b
, v2_x
, v0_x
, "dx20");
615 LLVMValueRef dy20
= LLVMBuildFSub(b
, v2_y
, v0_y
, "dy20");
617 LLVMValueRef one
= lp_build_const_float(gallivm
, 1.0);
618 LLVMValueRef e
= LLVMBuildFMul(b
, dx01
, dy20
, "e");
619 LLVMValueRef f
= LLVMBuildFMul(b
, dx20
, dy01
, "f");
620 LLVMValueRef ooa
= LLVMBuildFDiv(b
, one
, LLVMBuildFSub(b
, e
, f
, ""), "ooa");
622 LLVMValueRef dy20_ooa
= LLVMBuildFMul(b
, dy20
, ooa
, "dy20_ooa");
623 LLVMValueRef dy01_ooa
= LLVMBuildFMul(b
, dy01
, ooa
, "dy01_ooa");
624 LLVMValueRef dx20_ooa
= LLVMBuildFMul(b
, dx20
, ooa
, "dx20_ooa");
625 LLVMValueRef dx01_ooa
= LLVMBuildFMul(b
, dx01
, ooa
, "dx01_ooa");
627 args
->dy20_ooa
= vec4f_from_scalar(gallivm
, dy20_ooa
, "dy20_ooa_4f");
628 args
->dy01_ooa
= vec4f_from_scalar(gallivm
, dy01_ooa
, "dy01_ooa_4f");
630 args
->dx20_ooa
= vec4f_from_scalar(gallivm
, dx20_ooa
, "dx20_ooa_4f");
631 args
->dx01_ooa
= vec4f_from_scalar(gallivm
, dx01_ooa
, "dx01_ooa_4f");
633 args
->x0_center
= vec4f_from_scalar(gallivm
, x0_center
, "x0_center_4f");
634 args
->y0_center
= vec4f_from_scalar(gallivm
, y0_center
, "y0_center_4f");
638 * Generate the runtime callable function for the coefficient calculation.
641 static struct lp_setup_variant
*
642 generate_setup_variant(struct gallivm_state
*gallivm
,
643 struct lp_setup_variant_key
*key
,
644 struct llvmpipe_context
*lp
)
646 struct lp_setup_variant
*variant
= NULL
;
647 struct lp_setup_args args
;
649 LLVMTypeRef vec4f_type
;
650 LLVMTypeRef func_type
;
651 LLVMTypeRef arg_types
[7];
652 LLVMBasicBlockRef block
;
653 LLVMBuilderRef builder
= gallivm
->builder
;
659 variant
= CALLOC_STRUCT(lp_setup_variant
);
663 if (LP_DEBUG
& DEBUG_COUNTERS
) {
667 memcpy(&variant
->key
, key
, key
->size
);
668 variant
->list_item_global
.base
= variant
;
670 util_snprintf(func_name
, sizeof(func_name
), "fs%u_setup%u",
674 /* Currently always deal with full 4-wide vertex attributes from
678 vec4f_type
= LLVMVectorType(LLVMFloatTypeInContext(gallivm
->context
), 4);
680 arg_types
[0] = LLVMPointerType(vec4f_type
, 0); /* v0 */
681 arg_types
[1] = LLVMPointerType(vec4f_type
, 0); /* v1 */
682 arg_types
[2] = LLVMPointerType(vec4f_type
, 0); /* v2 */
683 arg_types
[3] = LLVMInt32TypeInContext(gallivm
->context
); /* facing */
684 arg_types
[4] = LLVMPointerType(vec4f_type
, 0); /* a0, aligned */
685 arg_types
[5] = LLVMPointerType(vec4f_type
, 0); /* dadx, aligned */
686 arg_types
[6] = LLVMPointerType(vec4f_type
, 0); /* dady, aligned */
688 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
689 arg_types
, Elements(arg_types
), 0);
691 variant
->function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
692 if (!variant
->function
)
695 LLVMSetFunctionCallConv(variant
->function
, LLVMCCallConv
);
697 args
.v0
= LLVMGetParam(variant
->function
, 0);
698 args
.v1
= LLVMGetParam(variant
->function
, 1);
699 args
.v2
= LLVMGetParam(variant
->function
, 2);
700 args
.facing
= LLVMGetParam(variant
->function
, 3);
701 args
.a0
= LLVMGetParam(variant
->function
, 4);
702 args
.dadx
= LLVMGetParam(variant
->function
, 5);
703 args
.dady
= LLVMGetParam(variant
->function
, 6);
705 lp_build_name(args
.v0
, "in_v0");
706 lp_build_name(args
.v1
, "in_v1");
707 lp_build_name(args
.v2
, "in_v2");
708 lp_build_name(args
.facing
, "in_facing");
709 lp_build_name(args
.a0
, "out_a0");
710 lp_build_name(args
.dadx
, "out_dadx");
711 lp_build_name(args
.dady
, "out_dady");
716 block
= LLVMAppendBasicBlockInContext(gallivm
->context
,
717 variant
->function
, "entry");
718 LLVMPositionBuilderAtEnd(builder
, block
);
720 set_noalias(builder
, variant
->function
, arg_types
, Elements(arg_types
));
721 init_args(gallivm
, &args
, variant
);
722 emit_tri_coef(gallivm
, &variant
->key
, &args
);
724 lp_emit_emms(gallivm
);
725 LLVMBuildRetVoid(builder
);
727 variant
->jit_function
= finalize_function(gallivm
, builder
,
729 if (!variant
->jit_function
)
733 * Update timing information:
735 if (LP_DEBUG
& DEBUG_COUNTERS
) {
737 LP_COUNT_ADD(llvm_compile_time
, t1
- t0
);
738 LP_COUNT_ADD(nr_llvm_compiles
, 1);
745 if (variant
->function
) {
746 if (variant
->jit_function
)
747 LLVMFreeMachineCodeForFunction(gallivm
->engine
,
749 LLVMDeleteFunction(variant
->function
);
760 lp_make_setup_variant_key(struct llvmpipe_context
*lp
,
761 struct lp_setup_variant_key
*key
)
763 struct lp_fragment_shader
*fs
= lp
->fs
;
766 assert(sizeof key
->inputs
[0] == sizeof(ushort
));
768 key
->num_inputs
= fs
->info
.base
.num_inputs
;
769 key
->flatshade_first
= lp
->rasterizer
->flatshade_first
;
770 key
->pixel_center_half
= lp
->rasterizer
->gl_rasterization_rules
;
771 key
->twoside
= lp
->rasterizer
->light_twoside
;
772 key
->size
= Offset(struct lp_setup_variant_key
,
773 inputs
[key
->num_inputs
]);
775 key
->color_slot
= lp
->color_slot
[0];
776 key
->bcolor_slot
= lp
->bcolor_slot
[0];
777 key
->spec_slot
= lp
->color_slot
[1];
778 key
->bspec_slot
= lp
->bcolor_slot
[1];
779 assert(key
->color_slot
== lp
->color_slot
[0]);
780 assert(key
->bcolor_slot
== lp
->bcolor_slot
[0]);
781 assert(key
->spec_slot
== lp
->color_slot
[1]);
782 assert(key
->bspec_slot
== lp
->bcolor_slot
[1]);
784 key
->units
= (float) (lp
->rasterizer
->offset_units
* lp
->mrd
);
785 key
->scale
= lp
->rasterizer
->offset_scale
;
787 memcpy(key
->inputs
, fs
->inputs
, key
->num_inputs
* sizeof key
->inputs
[0]);
788 for (i
= 0; i
< key
->num_inputs
; i
++) {
789 if (key
->inputs
[i
].interp
== LP_INTERP_COLOR
) {
790 if (lp
->rasterizer
->flatshade
)
791 key
->inputs
[i
].interp
= LP_INTERP_CONSTANT
;
793 key
->inputs
[i
].interp
= LP_INTERP_PERSPECTIVE
;
801 remove_setup_variant(struct llvmpipe_context
*lp
,
802 struct lp_setup_variant
*variant
)
804 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
805 debug_printf("llvmpipe: del setup_variant #%u total %u\n",
806 variant
->no
, lp
->nr_setup_variants
);
809 if (variant
->function
) {
810 if (variant
->jit_function
)
811 LLVMFreeMachineCodeForFunction(lp
->gallivm
->engine
,
813 LLVMDeleteFunction(variant
->function
);
816 remove_from_list(&variant
->list_item_global
);
817 lp
->nr_setup_variants
--;
823 /* When the number of setup variants exceeds a threshold, cull a
824 * fraction (currently a quarter) of them.
827 cull_setup_variants(struct llvmpipe_context
*lp
)
829 struct pipe_context
*pipe
= &lp
->pipe
;
833 * XXX: we need to flush the context until we have some sort of reference
834 * counting in fragment shaders as they may still be binned
835 * Flushing alone might not be sufficient we need to wait on it too.
837 llvmpipe_finish(pipe
, __FUNCTION__
);
839 for (i
= 0; i
< LP_MAX_SETUP_VARIANTS
/ 4; i
++) {
840 struct lp_setup_variant_list_item
*item
;
841 if (is_empty_list(&lp
->setup_variants_list
)) {
844 item
= last_elem(&lp
->setup_variants_list
);
847 remove_setup_variant(lp
, item
->base
);
853 * Update fragment/vertex shader linkage state. This is called just
854 * prior to drawing something when some fragment-related state has
858 llvmpipe_update_setup(struct llvmpipe_context
*lp
)
860 struct lp_setup_variant_key
*key
= &lp
->setup_variant
.key
;
861 struct lp_setup_variant
*variant
= NULL
;
862 struct lp_setup_variant_list_item
*li
;
864 lp_make_setup_variant_key(lp
, key
);
866 foreach(li
, &lp
->setup_variants_list
) {
867 if(li
->base
->key
.size
== key
->size
&&
868 memcmp(&li
->base
->key
, key
, key
->size
) == 0) {
875 move_to_head(&lp
->setup_variants_list
, &variant
->list_item_global
);
878 if (lp
->nr_setup_variants
>= LP_MAX_SETUP_VARIANTS
) {
879 cull_setup_variants(lp
);
882 variant
= generate_setup_variant(lp
->gallivm
, key
, lp
);
884 insert_at_head(&lp
->setup_variants_list
, &variant
->list_item_global
);
885 lp
->nr_setup_variants
++;
886 llvmpipe_variant_count
++;
890 lp_setup_set_setup_variant(lp
->setup
,
895 lp_delete_setup_variants(struct llvmpipe_context
*lp
)
897 struct lp_setup_variant_list_item
*li
;
898 li
= first_elem(&lp
->setup_variants_list
);
899 while(!at_end(&lp
->setup_variants_list
, li
)) {
900 struct lp_setup_variant_list_item
*next
= next_elem(li
);
901 remove_setup_variant(lp
, li
->base
);
907 lp_dump_setup_coef( const struct lp_setup_variant_key
*key
,
908 const float (*sa0
)[4],
909 const float (*sdadx
)[4],
910 const float (*sdady
)[4])
914 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
915 float a0
= sa0
[0][i
];
916 float dadx
= sdadx
[0][i
];
917 float dady
= sdady
[0][i
];
919 debug_printf("POS.%c: a0 = %f, dadx = %f, dady = %f\n",
924 for (slot
= 0; slot
< key
->num_inputs
; slot
++) {
925 unsigned usage_mask
= key
->inputs
[slot
].usage_mask
;
926 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
927 if (usage_mask
& (1 << i
)) {
928 float a0
= sa0
[1 + slot
][i
];
929 float dadx
= sdadx
[1 + slot
][i
];
930 float dady
= sdady
[1 + slot
][i
];
932 debug_printf("IN[%u].%c: a0 = %f, dadx = %f, dady = %f\n",