1 /* libs/pixelflinger/codeflinger/texturing.cpp
2 **
3 ** Copyright 2006, The Android Open Source Project
4 **
5 ** Licensed under the Apache License, Version 2.0 (the "License");
6 ** you may not use this file except in compliance with the License.
7 ** You may obtain a copy of the License at
8 **
9 ** http://www.apache.org/licenses/LICENSE-2.0
10 **
11 ** Unless required by applicable law or agreed to in writing, software
12 ** distributed under the License is distributed on an "AS IS" BASIS,
13 ** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14 ** See the License for the specific language governing permissions and
15 ** limitations under the License.
16 */
17
18 #include <assert.h>
19 #include <stdint.h>
20 #include <stdlib.h>
21 #include <stdio.h>
22 #include <sys/types.h>
23
24 #include <cutils/log.h>
25
26 #include "codeflinger/GGLAssembler.h"
27
28
29 namespace android {
30
31 // ---------------------------------------------------------------------------
32
33 // iterators are initialized like this:
34 // (intToFixedCenter(x) * dx)>>16 + x0
35 // ((x<<16 + 0x8000) * dx)>>16 + x0
36 // ((x<<16)*dx + (0x8000*dx))>>16 + x0
37 // ( (x*dx) + dx>>1 ) + x0
38 // (x*dx) + (dx>>1 + x0)
39
init_iterated_color(fragment_parts_t & parts,const reg_t & x)40 void GGLAssembler::init_iterated_color(fragment_parts_t& parts, const reg_t& x)
41 {
42 context_t const* c = mBuilderContext.c;
43 const needs_t& needs = mBuilderContext.needs;
44
45 if (mSmooth) {
46 // NOTE: we could take this case in the mDithering + !mSmooth case,
47 // but this would use up to 4 more registers for the color components
48 // for only a little added quality.
49 // Currently, this causes the system to run out of registers in
50 // some case (see issue #719496)
51
52 comment("compute initial iterated color (smooth and/or dither case)");
53
54 parts.iterated_packed = 0;
55 parts.packed = 0;
56
57 // 0x1: color component
58 // 0x2: iterators
59 const int optReload = mOptLevel >> 1;
60 if (optReload >= 3) parts.reload = 0; // reload nothing
61 else if (optReload == 2) parts.reload = 2; // reload iterators
62 else if (optReload == 1) parts.reload = 1; // reload colors
63 else if (optReload <= 0) parts.reload = 3; // reload both
64
65 if (!mSmooth) {
66 // we're not smoothing (just dithering), we never have to
67 // reload the iterators
68 parts.reload &= ~2;
69 }
70
71 Scratch scratches(registerFile());
72 const int t0 = (parts.reload & 1) ? scratches.obtain() : 0;
73 const int t1 = (parts.reload & 2) ? scratches.obtain() : 0;
74 for (int i=0 ; i<4 ; i++) {
75 if (!mInfo[i].iterated)
76 continue;
77
78 // this component exists in the destination and is not replaced
79 // by a texture unit.
80 const int c = (parts.reload & 1) ? t0 : obtainReg();
81 if (i==0) CONTEXT_LOAD(c, iterators.ydady);
82 if (i==1) CONTEXT_LOAD(c, iterators.ydrdy);
83 if (i==2) CONTEXT_LOAD(c, iterators.ydgdy);
84 if (i==3) CONTEXT_LOAD(c, iterators.ydbdy);
85 parts.argb[i].reg = c;
86
87 if (mInfo[i].smooth) {
88 parts.argb_dx[i].reg = (parts.reload & 2) ? t1 : obtainReg();
89 const int dvdx = parts.argb_dx[i].reg;
90 CONTEXT_LOAD(dvdx, generated_vars.argb[i].dx);
91 MLA(AL, 0, c, x.reg, dvdx, c);
92
93 // adjust the color iterator to make sure it won't overflow
94 if (!mAA) {
95 // this is not needed when we're using anti-aliasing
96 // because we will (have to) clamp the components
97 // anyway.
98 int end = scratches.obtain();
99 MOV(AL, 0, end, reg_imm(parts.count.reg, LSR, 16));
100 MLA(AL, 1, end, dvdx, end, c);
101 SUB(MI, 0, c, c, end);
102 BIC(AL, 0, c, c, reg_imm(c, ASR, 31));
103 scratches.recycle(end);
104 }
105 }
106
107 if (parts.reload & 1) {
108 CONTEXT_STORE(c, generated_vars.argb[i].c);
109 }
110 }
111 } else {
112 // We're not smoothed, so we can
113 // just use a packed version of the color and extract the
114 // components as needed (or not at all if we don't blend)
115
116 // figure out if we need the iterated color
117 int load = 0;
118 for (int i=0 ; i<4 ; i++) {
119 component_info_t& info = mInfo[i];
120 if ((info.inDest || info.needed) && !info.replaced)
121 load |= 1;
122 }
123
124 parts.iterated_packed = 1;
125 parts.packed = (!mTextureMachine.mask && !mBlending
126 && !mFog && !mDithering);
127 parts.reload = 0;
128 if (load || parts.packed) {
129 if (mBlending || mDithering || mInfo[GGLFormat::ALPHA].needed) {
130 comment("load initial iterated color (8888 packed)");
131 parts.iterated.setTo(obtainReg(),
132 &(c->formats[GGL_PIXEL_FORMAT_RGBA_8888]));
133 CONTEXT_LOAD(parts.iterated.reg, packed8888);
134 } else {
135 comment("load initial iterated color (dest format packed)");
136
137 parts.iterated.setTo(obtainReg(), &mCbFormat);
138
139 // pre-mask the iterated color
140 const int bits = parts.iterated.size();
141 const uint32_t size = ((bits>=32) ? 0 : (1LU << bits)) - 1;
142 uint32_t mask = 0;
143 if (mMasking) {
144 for (int i=0 ; i<4 ; i++) {
145 const int component_mask = 1<<i;
146 const int h = parts.iterated.format.c[i].h;
147 const int l = parts.iterated.format.c[i].l;
148 if (h && (!(mMasking & component_mask))) {
149 mask |= ((1<<(h-l))-1) << l;
150 }
151 }
152 }
153
154 if (mMasking && ((mask & size)==0)) {
155 // none of the components are present in the mask
156 } else {
157 CONTEXT_LOAD(parts.iterated.reg, packed);
158 if (mCbFormat.size == 1) {
159 AND(AL, 0, parts.iterated.reg,
160 parts.iterated.reg, imm(0xFF));
161 } else if (mCbFormat.size == 2) {
162 MOV(AL, 0, parts.iterated.reg,
163 reg_imm(parts.iterated.reg, LSR, 16));
164 }
165 }
166
167 // pre-mask the iterated color
168 if (mMasking) {
169 build_and_immediate(parts.iterated.reg, parts.iterated.reg,
170 mask, bits);
171 }
172 }
173 }
174 }
175 }
176
build_iterated_color(component_t & fragment,const fragment_parts_t & parts,int component,Scratch & regs)177 void GGLAssembler::build_iterated_color(
178 component_t& fragment,
179 const fragment_parts_t& parts,
180 int component,
181 Scratch& regs)
182 {
183 fragment.setTo( regs.obtain(), 0, 32, CORRUPTIBLE);
184
185 if (!mInfo[component].iterated)
186 return;
187
188 if (parts.iterated_packed) {
189 // iterated colors are packed, extract the one we need
190 extract(fragment, parts.iterated, component);
191 } else {
192 fragment.h = GGL_COLOR_BITS;
193 fragment.l = GGL_COLOR_BITS - 8;
194 fragment.flags |= CLEAR_LO;
195 // iterated colors are held in their own register,
196 // (smooth and/or dithering case)
197 if (parts.reload==3) {
198 // this implies mSmooth
199 Scratch scratches(registerFile());
200 int dx = scratches.obtain();
201 CONTEXT_LOAD(fragment.reg, generated_vars.argb[component].c);
202 CONTEXT_LOAD(dx, generated_vars.argb[component].dx);
203 ADD(AL, 0, dx, fragment.reg, dx);
204 CONTEXT_STORE(dx, generated_vars.argb[component].c);
205 } else if (parts.reload & 1) {
206 CONTEXT_LOAD(fragment.reg, generated_vars.argb[component].c);
207 } else {
208 // we don't reload, so simply rename the register and mark as
209 // non CORRUPTIBLE so that the texture env or blending code
210 // won't modify this (renamed) register
211 regs.recycle(fragment.reg);
212 fragment.reg = parts.argb[component].reg;
213 fragment.flags &= ~CORRUPTIBLE;
214 }
215 if (mInfo[component].smooth && mAA) {
216 // when using smooth shading AND anti-aliasing, we need to clamp
217 // the iterators because there is always an extra pixel on the
218 // edges, which most of the time will cause an overflow
219 // (since technically its outside of the domain).
220 BIC(AL, 0, fragment.reg, fragment.reg,
221 reg_imm(fragment.reg, ASR, 31));
222 component_sat(fragment);
223 }
224 }
225 }
226
227 // ---------------------------------------------------------------------------
228
decodeLogicOpNeeds(const needs_t & needs)229 void GGLAssembler::decodeLogicOpNeeds(const needs_t& needs)
230 {
231 // gather some informations about the components we need to process...
232 const int opcode = GGL_READ_NEEDS(LOGIC_OP, needs.n) | GGL_CLEAR;
233 switch(opcode) {
234 case GGL_COPY:
235 mLogicOp = 0;
236 break;
237 case GGL_CLEAR:
238 case GGL_SET:
239 mLogicOp = LOGIC_OP;
240 break;
241 case GGL_AND:
242 case GGL_AND_REVERSE:
243 case GGL_AND_INVERTED:
244 case GGL_XOR:
245 case GGL_OR:
246 case GGL_NOR:
247 case GGL_EQUIV:
248 case GGL_OR_REVERSE:
249 case GGL_OR_INVERTED:
250 case GGL_NAND:
251 mLogicOp = LOGIC_OP|LOGIC_OP_SRC|LOGIC_OP_DST;
252 break;
253 case GGL_NOOP:
254 case GGL_INVERT:
255 mLogicOp = LOGIC_OP|LOGIC_OP_DST;
256 break;
257 case GGL_COPY_INVERTED:
258 mLogicOp = LOGIC_OP|LOGIC_OP_SRC;
259 break;
260 };
261 }
262
decodeTMUNeeds(const needs_t & needs,context_t const * c)263 void GGLAssembler::decodeTMUNeeds(const needs_t& needs, context_t const* c)
264 {
265 uint8_t replaced=0;
266 mTextureMachine.mask = 0;
267 mTextureMachine.activeUnits = 0;
268 for (int i=GGL_TEXTURE_UNIT_COUNT-1 ; i>=0 ; i--) {
269 texture_unit_t& tmu = mTextureMachine.tmu[i];
270 if (replaced == 0xF) {
271 // all components are replaced, skip this TMU.
272 tmu.format_idx = 0;
273 tmu.mask = 0;
274 tmu.replaced = replaced;
275 continue;
276 }
277 tmu.format_idx = GGL_READ_NEEDS(T_FORMAT, needs.t[i]);
278 tmu.format = c->formats[tmu.format_idx];
279 tmu.bits = tmu.format.size*8;
280 tmu.swrap = GGL_READ_NEEDS(T_S_WRAP, needs.t[i]);
281 tmu.twrap = GGL_READ_NEEDS(T_T_WRAP, needs.t[i]);
282 tmu.env = ggl_needs_to_env(GGL_READ_NEEDS(T_ENV, needs.t[i]));
283 tmu.pot = GGL_READ_NEEDS(T_POT, needs.t[i]);
284 tmu.linear = GGL_READ_NEEDS(T_LINEAR, needs.t[i])
285 && tmu.format.size!=3; // XXX: only 8, 16 and 32 modes for now
286
287 // 5551 linear filtering is not supported
288 if (tmu.format_idx == GGL_PIXEL_FORMAT_RGBA_5551)
289 tmu.linear = 0;
290
291 tmu.mask = 0;
292 tmu.replaced = replaced;
293
294 if (tmu.format_idx) {
295 mTextureMachine.activeUnits++;
296 if (tmu.format.c[0].h) tmu.mask |= 0x1;
297 if (tmu.format.c[1].h) tmu.mask |= 0x2;
298 if (tmu.format.c[2].h) tmu.mask |= 0x4;
299 if (tmu.format.c[3].h) tmu.mask |= 0x8;
300 if (tmu.env == GGL_REPLACE) {
301 replaced |= tmu.mask;
302 } else if (tmu.env == GGL_DECAL) {
303 if (!tmu.format.c[GGLFormat::ALPHA].h) {
304 // if we don't have alpha, decal does nothing
305 tmu.mask = 0;
306 } else {
307 // decal always ignores At
308 tmu.mask &= ~(1<<GGLFormat::ALPHA);
309 }
310 }
311 }
312 mTextureMachine.mask |= tmu.mask;
313 //printf("%d: mask=%08lx, replaced=%08lx\n",
314 // i, int(tmu.mask), int(tmu.replaced));
315 }
316 mTextureMachine.replaced = replaced;
317 mTextureMachine.directTexture = 0;
318 //printf("replaced=%08lx\n", mTextureMachine.replaced);
319 }
320
321
init_textures(tex_coord_t * coords,const reg_t & x,const reg_t & y)322 void GGLAssembler::init_textures(
323 tex_coord_t* coords,
324 const reg_t& x, const reg_t& y)
325 {
326 context_t const* c = mBuilderContext.c;
327 const needs_t& needs = mBuilderContext.needs;
328 int Rctx = mBuilderContext.Rctx;
329 int Rx = x.reg;
330 int Ry = y.reg;
331
332 if (mTextureMachine.mask) {
333 comment("compute texture coordinates");
334 }
335
336 // init texture coordinates for each tmu
337 const int cb_format_idx = GGL_READ_NEEDS(CB_FORMAT, needs.n);
338 const bool multiTexture = mTextureMachine.activeUnits > 1;
339 for (int i=0 ; i<GGL_TEXTURE_UNIT_COUNT; i++) {
340 const texture_unit_t& tmu = mTextureMachine.tmu[i];
341 if (tmu.format_idx == 0)
342 continue;
343 if ((tmu.swrap == GGL_NEEDS_WRAP_11) &&
344 (tmu.twrap == GGL_NEEDS_WRAP_11))
345 {
346 // 1:1 texture
347 pointer_t& txPtr = coords[i].ptr;
348 txPtr.setTo(obtainReg(), tmu.bits);
349 CONTEXT_LOAD(txPtr.reg, state.texture[i].iterators.ydsdy);
350 ADD(AL, 0, Rx, Rx, reg_imm(txPtr.reg, ASR, 16)); // x += (s>>16)
351 CONTEXT_LOAD(txPtr.reg, state.texture[i].iterators.ydtdy);
352 ADD(AL, 0, Ry, Ry, reg_imm(txPtr.reg, ASR, 16)); // y += (t>>16)
353 // merge base & offset
354 CONTEXT_LOAD(txPtr.reg, generated_vars.texture[i].stride);
355 SMLABB(AL, Rx, Ry, txPtr.reg, Rx); // x+y*stride
356 CONTEXT_LOAD(txPtr.reg, generated_vars.texture[i].data);
357 base_offset(txPtr, txPtr, Rx);
358 } else {
359 Scratch scratches(registerFile());
360 reg_t& s = coords[i].s;
361 reg_t& t = coords[i].t;
362 // s = (x * dsdx)>>16 + ydsdy
363 // s = (x * dsdx)>>16 + (y*dsdy)>>16 + s0
364 // t = (x * dtdx)>>16 + ydtdy
365 // t = (x * dtdx)>>16 + (y*dtdy)>>16 + t0
366 s.setTo(obtainReg());
367 t.setTo(obtainReg());
368 const int need_w = GGL_READ_NEEDS(W, needs.n);
369 if (need_w) {
370 CONTEXT_LOAD(s.reg, state.texture[i].iterators.ydsdy);
371 CONTEXT_LOAD(t.reg, state.texture[i].iterators.ydtdy);
372 } else {
373 int ydsdy = scratches.obtain();
374 int ydtdy = scratches.obtain();
375 CONTEXT_LOAD(s.reg, generated_vars.texture[i].dsdx);
376 CONTEXT_LOAD(ydsdy, state.texture[i].iterators.ydsdy);
377 CONTEXT_LOAD(t.reg, generated_vars.texture[i].dtdx);
378 CONTEXT_LOAD(ydtdy, state.texture[i].iterators.ydtdy);
379 MLA(AL, 0, s.reg, Rx, s.reg, ydsdy);
380 MLA(AL, 0, t.reg, Rx, t.reg, ydtdy);
381 }
382
383 if ((mOptLevel&1)==0) {
384 CONTEXT_STORE(s.reg, generated_vars.texture[i].spill[0]);
385 CONTEXT_STORE(t.reg, generated_vars.texture[i].spill[1]);
386 recycleReg(s.reg);
387 recycleReg(t.reg);
388 }
389 }
390
391 // direct texture?
392 if (!multiTexture && !mBlending && !mDithering && !mFog &&
393 cb_format_idx == tmu.format_idx && !tmu.linear &&
394 mTextureMachine.replaced == tmu.mask)
395 {
396 mTextureMachine.directTexture = i + 1;
397 }
398 }
399 }
400
build_textures(fragment_parts_t & parts,Scratch & regs)401 void GGLAssembler::build_textures( fragment_parts_t& parts,
402 Scratch& regs)
403 {
404 context_t const* c = mBuilderContext.c;
405 const needs_t& needs = mBuilderContext.needs;
406 int Rctx = mBuilderContext.Rctx;
407
408 // We don't have a way to spill registers automatically
409 // spill depth and AA regs, when we know we may have to.
410 // build the spill list...
411 uint32_t spill_list = 0;
412 for (int i=0 ; i<GGL_TEXTURE_UNIT_COUNT; i++) {
413 const texture_unit_t& tmu = mTextureMachine.tmu[i];
414 if (tmu.format_idx == 0)
415 continue;
416 if (tmu.linear) {
417 // we may run out of register if we have linear filtering
418 // at 1 or 4 bytes / pixel on any texture unit.
419 if (tmu.format.size == 1) {
420 // if depth and AA enabled, we'll run out of 1 register
421 if (parts.z.reg > 0 && parts.covPtr.reg > 0)
422 spill_list |= 1<<parts.covPtr.reg;
423 }
424 if (tmu.format.size == 4) {
425 // if depth or AA enabled, we'll run out of 1 or 2 registers
426 if (parts.z.reg > 0)
427 spill_list |= 1<<parts.z.reg;
428 if (parts.covPtr.reg > 0)
429 spill_list |= 1<<parts.covPtr.reg;
430 }
431 }
432 }
433
434 Spill spill(registerFile(), *this, spill_list);
435
436 const bool multiTexture = mTextureMachine.activeUnits > 1;
437 for (int i=0 ; i<GGL_TEXTURE_UNIT_COUNT; i++) {
438 const texture_unit_t& tmu = mTextureMachine.tmu[i];
439 if (tmu.format_idx == 0)
440 continue;
441
442 pointer_t& txPtr = parts.coords[i].ptr;
443 pixel_t& texel = parts.texel[i];
444
445 // repeat...
446 if ((tmu.swrap == GGL_NEEDS_WRAP_11) &&
447 (tmu.twrap == GGL_NEEDS_WRAP_11))
448 { // 1:1 textures
449 comment("fetch texel");
450 texel.setTo(regs.obtain(), &tmu.format);
451 load(txPtr, texel, WRITE_BACK);
452 } else {
453 Scratch scratches(registerFile());
454 reg_t& s = parts.coords[i].s;
455 reg_t& t = parts.coords[i].t;
456 if ((mOptLevel&1)==0) {
457 comment("reload s/t (multitexture or linear filtering)");
458 s.reg = scratches.obtain();
459 t.reg = scratches.obtain();
460 CONTEXT_LOAD(s.reg, generated_vars.texture[i].spill[0]);
461 CONTEXT_LOAD(t.reg, generated_vars.texture[i].spill[1]);
462 }
463
464 comment("compute repeat/clamp");
465 int u = scratches.obtain();
466 int v = scratches.obtain();
467 int width = scratches.obtain();
468 int height = scratches.obtain();
469 int U = 0;
470 int V = 0;
471
472 CONTEXT_LOAD(width, generated_vars.texture[i].width);
473 CONTEXT_LOAD(height, generated_vars.texture[i].height);
474
475 int FRAC_BITS = 0;
476 if (tmu.linear) {
477 // linear interpolation
478 if (tmu.format.size == 1) {
479 // for 8-bits textures, we can afford
480 // 7 bits of fractional precision at no
481 // additional cost (we can't do 8 bits
482 // because filter8 uses signed 16 bits muls)
483 FRAC_BITS = 7;
484 } else if (tmu.format.size == 2) {
485 // filter16() is internally limited to 4 bits, so:
486 // FRAC_BITS=2 generates less instructions,
487 // FRAC_BITS=3,4,5 creates unpleasant artifacts,
488 // FRAC_BITS=6+ looks good
489 FRAC_BITS = 6;
490 } else if (tmu.format.size == 4) {
491 // filter32() is internally limited to 8 bits, so:
492 // FRAC_BITS=4 looks good
493 // FRAC_BITS=5+ looks better, but generates 3 extra ipp
494 FRAC_BITS = 6;
495 } else {
496 // for all other cases we use 4 bits.
497 FRAC_BITS = 4;
498 }
499 }
500 wrapping(u, s.reg, width, tmu.swrap, FRAC_BITS);
501 wrapping(v, t.reg, height, tmu.twrap, FRAC_BITS);
502
503 if (tmu.linear) {
504 comment("compute linear filtering offsets");
505 // pixel size scale
506 const int shift = 31 - gglClz(tmu.format.size);
507 U = scratches.obtain();
508 V = scratches.obtain();
509
510 // sample the texel center
511 SUB(AL, 0, u, u, imm(1<<(FRAC_BITS-1)));
512 SUB(AL, 0, v, v, imm(1<<(FRAC_BITS-1)));
513
514 // get the fractionnal part of U,V
515 AND(AL, 0, U, u, imm((1<<FRAC_BITS)-1));
516 AND(AL, 0, V, v, imm((1<<FRAC_BITS)-1));
517
518 // compute width-1 and height-1
519 SUB(AL, 0, width, width, imm(1));
520 SUB(AL, 0, height, height, imm(1));
521
522 // get the integer part of U,V and clamp/wrap
523 // and compute offset to the next texel
524 if (tmu.swrap == GGL_NEEDS_WRAP_REPEAT) {
525 // u has already been REPEATed
526 MOV(AL, 1, u, reg_imm(u, ASR, FRAC_BITS));
527 MOV(MI, 0, u, width);
528 CMP(AL, u, width);
529 MOV(LT, 0, width, imm(1 << shift));
530 if (shift)
531 MOV(GE, 0, width, reg_imm(width, LSL, shift));
532 RSB(GE, 0, width, width, imm(0));
533 } else {
534 // u has not been CLAMPed yet
535 // algorithm:
536 // if ((u>>4) >= width)
537 // u = width<<4
538 // width = 0
539 // else
540 // width = 1<<shift
541 // u = u>>4; // get integer part
542 // if (u<0)
543 // u = 0
544 // width = 0
545 // generated_vars.rt = width
546
547 CMP(AL, width, reg_imm(u, ASR, FRAC_BITS));
548 MOV(LE, 0, u, reg_imm(width, LSL, FRAC_BITS));
549 MOV(LE, 0, width, imm(0));
550 MOV(GT, 0, width, imm(1 << shift));
551 MOV(AL, 1, u, reg_imm(u, ASR, FRAC_BITS));
552 MOV(MI, 0, u, imm(0));
553 MOV(MI, 0, width, imm(0));
554 }
555 CONTEXT_STORE(width, generated_vars.rt);
556
557 const int stride = width;
558 CONTEXT_LOAD(stride, generated_vars.texture[i].stride);
559 if (tmu.twrap == GGL_NEEDS_WRAP_REPEAT) {
560 // v has already been REPEATed
561 MOV(AL, 1, v, reg_imm(v, ASR, FRAC_BITS));
562 MOV(MI, 0, v, height);
563 CMP(AL, v, height);
564 MOV(LT, 0, height, imm(1 << shift));
565 if (shift)
566 MOV(GE, 0, height, reg_imm(height, LSL, shift));
567 RSB(GE, 0, height, height, imm(0));
568 MUL(AL, 0, height, stride, height);
569 } else {
570 // u has not been CLAMPed yet
571 CMP(AL, height, reg_imm(v, ASR, FRAC_BITS));
572 MOV(LE, 0, v, reg_imm(height, LSL, FRAC_BITS));
573 MOV(LE, 0, height, imm(0));
574 if (shift) {
575 MOV(GT, 0, height, reg_imm(stride, LSL, shift));
576 } else {
577 MOV(GT, 0, height, stride);
578 }
579 MOV(AL, 1, v, reg_imm(v, ASR, FRAC_BITS));
580 MOV(MI, 0, v, imm(0));
581 MOV(MI, 0, height, imm(0));
582 }
583 CONTEXT_STORE(height, generated_vars.lb);
584 }
585
586 scratches.recycle(width);
587 scratches.recycle(height);
588
589 // iterate texture coordinates...
590 comment("iterate s,t");
591 int dsdx = scratches.obtain();
592 int dtdx = scratches.obtain();
593 CONTEXT_LOAD(dsdx, generated_vars.texture[i].dsdx);
594 CONTEXT_LOAD(dtdx, generated_vars.texture[i].dtdx);
595 ADD(AL, 0, s.reg, s.reg, dsdx);
596 ADD(AL, 0, t.reg, t.reg, dtdx);
597 if ((mOptLevel&1)==0) {
598 CONTEXT_STORE(s.reg, generated_vars.texture[i].spill[0]);
599 CONTEXT_STORE(t.reg, generated_vars.texture[i].spill[1]);
600 scratches.recycle(s.reg);
601 scratches.recycle(t.reg);
602 }
603 scratches.recycle(dsdx);
604 scratches.recycle(dtdx);
605
606 // merge base & offset...
607 comment("merge base & offset");
608 texel.setTo(regs.obtain(), &tmu.format);
609 txPtr.setTo(texel.reg, tmu.bits);
610 int stride = scratches.obtain();
611 CONTEXT_LOAD(stride, generated_vars.texture[i].stride);
612 CONTEXT_LOAD(txPtr.reg, generated_vars.texture[i].data);
613 SMLABB(AL, u, v, stride, u); // u+v*stride
614 base_offset(txPtr, txPtr, u);
615
616 // load texel
617 if (!tmu.linear) {
618 comment("fetch texel");
619 load(txPtr, texel, 0);
620 } else {
621 // recycle registers we don't need anymore
622 scratches.recycle(u);
623 scratches.recycle(v);
624 scratches.recycle(stride);
625
626 comment("fetch texel, bilinear");
627 switch (tmu.format.size) {
628 case 1: filter8(parts, texel, tmu, U, V, txPtr, FRAC_BITS); break;
629 case 2: filter16(parts, texel, tmu, U, V, txPtr, FRAC_BITS); break;
630 case 3: filter24(parts, texel, tmu, U, V, txPtr, FRAC_BITS); break;
631 case 4: filter32(parts, texel, tmu, U, V, txPtr, FRAC_BITS); break;
632 }
633 }
634 }
635 }
636 }
637
build_iterate_texture_coordinates(const fragment_parts_t & parts)638 void GGLAssembler::build_iterate_texture_coordinates(
639 const fragment_parts_t& parts)
640 {
641 const bool multiTexture = mTextureMachine.activeUnits > 1;
642 for (int i=0 ; i<GGL_TEXTURE_UNIT_COUNT; i++) {
643 const texture_unit_t& tmu = mTextureMachine.tmu[i];
644 if (tmu.format_idx == 0)
645 continue;
646
647 if ((tmu.swrap == GGL_NEEDS_WRAP_11) &&
648 (tmu.twrap == GGL_NEEDS_WRAP_11))
649 { // 1:1 textures
650 const pointer_t& txPtr = parts.coords[i].ptr;
651 ADD(AL, 0, txPtr.reg, txPtr.reg, imm(txPtr.size>>3));
652 } else {
653 Scratch scratches(registerFile());
654 int s = parts.coords[i].s.reg;
655 int t = parts.coords[i].t.reg;
656 if ((mOptLevel&1)==0) {
657 s = scratches.obtain();
658 t = scratches.obtain();
659 CONTEXT_LOAD(s, generated_vars.texture[i].spill[0]);
660 CONTEXT_LOAD(t, generated_vars.texture[i].spill[1]);
661 }
662 int dsdx = scratches.obtain();
663 int dtdx = scratches.obtain();
664 CONTEXT_LOAD(dsdx, generated_vars.texture[i].dsdx);
665 CONTEXT_LOAD(dtdx, generated_vars.texture[i].dtdx);
666 ADD(AL, 0, s, s, dsdx);
667 ADD(AL, 0, t, t, dtdx);
668 if ((mOptLevel&1)==0) {
669 CONTEXT_STORE(s, generated_vars.texture[i].spill[0]);
670 CONTEXT_STORE(t, generated_vars.texture[i].spill[1]);
671 }
672 }
673 }
674 }
675
filter8(const fragment_parts_t & parts,pixel_t & texel,const texture_unit_t & tmu,int U,int V,pointer_t & txPtr,int FRAC_BITS)676 void GGLAssembler::filter8(
677 const fragment_parts_t& parts,
678 pixel_t& texel, const texture_unit_t& tmu,
679 int U, int V, pointer_t& txPtr,
680 int FRAC_BITS)
681 {
682 if (tmu.format.components != GGL_ALPHA &&
683 tmu.format.components != GGL_LUMINANCE)
684 {
685 // this is a packed format, and we don't support
686 // linear filtering (it's probably RGB 332)
687 // Should not happen with OpenGL|ES
688 LDRB(AL, texel.reg, txPtr.reg);
689 return;
690 }
691
692 // ------------------------
693 // about ~22 cycles / pixel
694 Scratch scratches(registerFile());
695
696 int pixel= scratches.obtain();
697 int d = scratches.obtain();
698 int u = scratches.obtain();
699 int k = scratches.obtain();
700 int rt = scratches.obtain();
701 int lb = scratches.obtain();
702
703 // RB -> U * V
704
705 CONTEXT_LOAD(rt, generated_vars.rt);
706 CONTEXT_LOAD(lb, generated_vars.lb);
707
708 int offset = pixel;
709 ADD(AL, 0, offset, lb, rt);
710 LDRB(AL, pixel, txPtr.reg, reg_scale_pre(offset));
711 SMULBB(AL, u, U, V);
712 SMULBB(AL, d, pixel, u);
713 RSB(AL, 0, k, u, imm(1<<(FRAC_BITS*2)));
714
715 // LB -> (1-U) * V
716 RSB(AL, 0, U, U, imm(1<<FRAC_BITS));
717 LDRB(AL, pixel, txPtr.reg, reg_scale_pre(lb));
718 SMULBB(AL, u, U, V);
719 SMLABB(AL, d, pixel, u, d);
720 SUB(AL, 0, k, k, u);
721
722 // LT -> (1-U)*(1-V)
723 RSB(AL, 0, V, V, imm(1<<FRAC_BITS));
724 LDRB(AL, pixel, txPtr.reg);
725 SMULBB(AL, u, U, V);
726 SMLABB(AL, d, pixel, u, d);
727
728 // RT -> U*(1-V)
729 LDRB(AL, pixel, txPtr.reg, reg_scale_pre(rt));
730 SUB(AL, 0, u, k, u);
731 SMLABB(AL, texel.reg, pixel, u, d);
732
733 for (int i=0 ; i<4 ; i++) {
734 if (!texel.format.c[i].h) continue;
735 texel.format.c[i].h = FRAC_BITS*2+8;
736 texel.format.c[i].l = FRAC_BITS*2; // keeping 8 bits in enough
737 }
738 texel.format.size = 4;
739 texel.format.bitsPerPixel = 32;
740 texel.flags |= CLEAR_LO;
741 }
742
filter16(const fragment_parts_t & parts,pixel_t & texel,const texture_unit_t & tmu,int U,int V,pointer_t & txPtr,int FRAC_BITS)743 void GGLAssembler::filter16(
744 const fragment_parts_t& parts,
745 pixel_t& texel, const texture_unit_t& tmu,
746 int U, int V, pointer_t& txPtr,
747 int FRAC_BITS)
748 {
749 // compute the mask
750 // XXX: it would be nice if the mask below could be computed
751 // automatically.
752 uint32_t mask = 0;
753 int shift = 0;
754 int prec = 0;
755 switch (tmu.format_idx) {
756 case GGL_PIXEL_FORMAT_RGB_565:
757 // source: 00000ggg.ggg00000 | rrrrr000.000bbbbb
758 // result: gggggggg.gggrrrrr | rrrrr0bb.bbbbbbbb
759 mask = 0x07E0F81F;
760 shift = 16;
761 prec = 5;
762 break;
763 case GGL_PIXEL_FORMAT_RGBA_4444:
764 // 0000,1111,0000,1111 | 0000,1111,0000,1111
765 mask = 0x0F0F0F0F;
766 shift = 12;
767 prec = 4;
768 break;
769 case GGL_PIXEL_FORMAT_LA_88:
770 // 0000,0000,1111,1111 | 0000,0000,1111,1111
771 // AALL -> 00AA | 00LL
772 mask = 0x00FF00FF;
773 shift = 8;
774 prec = 8;
775 break;
776 default:
777 // unsupported format, do something sensical...
778 LOGE("Unsupported 16-bits texture format (%d)", tmu.format_idx);
779 LDRH(AL, texel.reg, txPtr.reg);
780 return;
781 }
782
783 const int adjust = FRAC_BITS*2 - prec;
784 const int round = 0;
785
786 // update the texel format
787 texel.format.size = 4;
788 texel.format.bitsPerPixel = 32;
789 texel.flags |= CLEAR_HI|CLEAR_LO;
790 for (int i=0 ; i<4 ; i++) {
791 if (!texel.format.c[i].h) continue;
792 const uint32_t offset = (mask & tmu.format.mask(i)) ? 0 : shift;
793 texel.format.c[i].h = tmu.format.c[i].h + offset + prec;
794 texel.format.c[i].l = texel.format.c[i].h - (tmu.format.bits(i) + prec);
795 }
796
797 // ------------------------
798 // about ~40 cycles / pixel
799 Scratch scratches(registerFile());
800
801 int pixel= scratches.obtain();
802 int d = scratches.obtain();
803 int u = scratches.obtain();
804 int k = scratches.obtain();
805
806 // RB -> U * V
807 int offset = pixel;
808 CONTEXT_LOAD(offset, generated_vars.rt);
809 CONTEXT_LOAD(u, generated_vars.lb);
810 ADD(AL, 0, offset, offset, u);
811
812 LDRH(AL, pixel, txPtr.reg, reg_pre(offset));
813 SMULBB(AL, u, U, V);
814 ORR(AL, 0, pixel, pixel, reg_imm(pixel, LSL, shift));
815 build_and_immediate(pixel, pixel, mask, 32);
816 if (adjust) {
817 if (round)
818 ADD(AL, 0, u, u, imm(1<<(adjust-1)));
819 MOV(AL, 0, u, reg_imm(u, LSR, adjust));
820 }
821 MUL(AL, 0, d, pixel, u);
822 RSB(AL, 0, k, u, imm(1<<prec));
823
824 // LB -> (1-U) * V
825 CONTEXT_LOAD(offset, generated_vars.lb);
826 RSB(AL, 0, U, U, imm(1<<FRAC_BITS));
827 LDRH(AL, pixel, txPtr.reg, reg_pre(offset));
828 SMULBB(AL, u, U, V);
829 ORR(AL, 0, pixel, pixel, reg_imm(pixel, LSL, shift));
830 build_and_immediate(pixel, pixel, mask, 32);
831 if (adjust) {
832 if (round)
833 ADD(AL, 0, u, u, imm(1<<(adjust-1)));
834 MOV(AL, 0, u, reg_imm(u, LSR, adjust));
835 }
836 MLA(AL, 0, d, pixel, u, d);
837 SUB(AL, 0, k, k, u);
838
839 // LT -> (1-U)*(1-V)
840 RSB(AL, 0, V, V, imm(1<<FRAC_BITS));
841 LDRH(AL, pixel, txPtr.reg);
842 SMULBB(AL, u, U, V);
843 ORR(AL, 0, pixel, pixel, reg_imm(pixel, LSL, shift));
844 build_and_immediate(pixel, pixel, mask, 32);
845 if (adjust) {
846 if (round)
847 ADD(AL, 0, u, u, imm(1<<(adjust-1)));
848 MOV(AL, 0, u, reg_imm(u, LSR, adjust));
849 }
850 MLA(AL, 0, d, pixel, u, d);
851
852 // RT -> U*(1-V)
853 CONTEXT_LOAD(offset, generated_vars.rt);
854 LDRH(AL, pixel, txPtr.reg, reg_pre(offset));
855 SUB(AL, 0, u, k, u);
856 ORR(AL, 0, pixel, pixel, reg_imm(pixel, LSL, shift));
857 build_and_immediate(pixel, pixel, mask, 32);
858 MLA(AL, 0, texel.reg, pixel, u, d);
859 }
860
filter24(const fragment_parts_t & parts,pixel_t & texel,const texture_unit_t & tmu,int U,int V,pointer_t & txPtr,int FRAC_BITS)861 void GGLAssembler::filter24(
862 const fragment_parts_t& parts,
863 pixel_t& texel, const texture_unit_t& tmu,
864 int U, int V, pointer_t& txPtr,
865 int FRAC_BITS)
866 {
867 // not supported yet (currently disabled)
868 load(txPtr, texel, 0);
869 }
870
filter32(const fragment_parts_t & parts,pixel_t & texel,const texture_unit_t & tmu,int U,int V,pointer_t & txPtr,int FRAC_BITS)871 void GGLAssembler::filter32(
872 const fragment_parts_t& parts,
873 pixel_t& texel, const texture_unit_t& tmu,
874 int U, int V, pointer_t& txPtr,
875 int FRAC_BITS)
876 {
877 const int adjust = FRAC_BITS*2 - 8;
878 const int round = 0;
879
880 // ------------------------
881 // about ~38 cycles / pixel
882 Scratch scratches(registerFile());
883
884 int pixel= scratches.obtain();
885 int dh = scratches.obtain();
886 int u = scratches.obtain();
887 int k = scratches.obtain();
888
889 int temp = scratches.obtain();
890 int dl = scratches.obtain();
891 int mask = scratches.obtain();
892
893 MOV(AL, 0, mask, imm(0xFF));
894 ORR(AL, 0, mask, mask, imm(0xFF0000));
895
896 // RB -> U * V
897 int offset = pixel;
898 CONTEXT_LOAD(offset, generated_vars.rt);
899 CONTEXT_LOAD(u, generated_vars.lb);
900 ADD(AL, 0, offset, offset, u);
901
902 LDR(AL, pixel, txPtr.reg, reg_scale_pre(offset));
903 SMULBB(AL, u, U, V);
904 AND(AL, 0, temp, mask, pixel);
905 if (adjust) {
906 if (round)
907 ADD(AL, 0, u, u, imm(1<<(adjust-1)));
908 MOV(AL, 0, u, reg_imm(u, LSR, adjust));
909 }
910 MUL(AL, 0, dh, temp, u);
911 AND(AL, 0, temp, mask, reg_imm(pixel, LSR, 8));
912 MUL(AL, 0, dl, temp, u);
913 RSB(AL, 0, k, u, imm(0x100));
914
915 // LB -> (1-U) * V
916 CONTEXT_LOAD(offset, generated_vars.lb);
917 RSB(AL, 0, U, U, imm(1<<FRAC_BITS));
918 LDR(AL, pixel, txPtr.reg, reg_scale_pre(offset));
919 SMULBB(AL, u, U, V);
920 AND(AL, 0, temp, mask, pixel);
921 if (adjust) {
922 if (round)
923 ADD(AL, 0, u, u, imm(1<<(adjust-1)));
924 MOV(AL, 0, u, reg_imm(u, LSR, adjust));
925 }
926 MLA(AL, 0, dh, temp, u, dh);
927 AND(AL, 0, temp, mask, reg_imm(pixel, LSR, 8));
928 MLA(AL, 0, dl, temp, u, dl);
929 SUB(AL, 0, k, k, u);
930
931 // LT -> (1-U)*(1-V)
932 RSB(AL, 0, V, V, imm(1<<FRAC_BITS));
933 LDR(AL, pixel, txPtr.reg);
934 SMULBB(AL, u, U, V);
935 AND(AL, 0, temp, mask, pixel);
936 if (adjust) {
937 if (round)
938 ADD(AL, 0, u, u, imm(1<<(adjust-1)));
939 MOV(AL, 0, u, reg_imm(u, LSR, adjust));
940 }
941 MLA(AL, 0, dh, temp, u, dh);
942 AND(AL, 0, temp, mask, reg_imm(pixel, LSR, 8));
943 MLA(AL, 0, dl, temp, u, dl);
944
945 // RT -> U*(1-V)
946 CONTEXT_LOAD(offset, generated_vars.rt);
947 LDR(AL, pixel, txPtr.reg, reg_scale_pre(offset));
948 SUB(AL, 0, u, k, u);
949 AND(AL, 0, temp, mask, pixel);
950 MLA(AL, 0, dh, temp, u, dh);
951 AND(AL, 0, temp, mask, reg_imm(pixel, LSR, 8));
952 MLA(AL, 0, dl, temp, u, dl);
953
954 AND(AL, 0, dh, mask, reg_imm(dh, LSR, 8));
955 AND(AL, 0, dl, dl, reg_imm(mask, LSL, 8));
956 ORR(AL, 0, texel.reg, dh, dl);
957 }
958
build_texture_environment(component_t & fragment,const fragment_parts_t & parts,int component,Scratch & regs)959 void GGLAssembler::build_texture_environment(
960 component_t& fragment,
961 const fragment_parts_t& parts,
962 int component,
963 Scratch& regs)
964 {
965 const uint32_t component_mask = 1<<component;
966 const bool multiTexture = mTextureMachine.activeUnits > 1;
967 for (int i=0 ; i<GGL_TEXTURE_UNIT_COUNT ; i++) {
968 texture_unit_t& tmu = mTextureMachine.tmu[i];
969
970 if (tmu.mask & component_mask) {
971 // replace or modulate with this texture
972 if ((tmu.replaced & component_mask) == 0) {
973 // not replaced by a later tmu...
974
975 Scratch scratches(registerFile());
976 pixel_t texel(parts.texel[i]);
977 if (multiTexture &&
978 tmu.swrap == GGL_NEEDS_WRAP_11 &&
979 tmu.twrap == GGL_NEEDS_WRAP_11)
980 {
981 texel.reg = scratches.obtain();
982 texel.flags |= CORRUPTIBLE;
983 comment("fetch texel (multitexture 1:1)");
984 load(parts.coords[i].ptr, texel, WRITE_BACK);
985 }
986
987 component_t incoming(fragment);
988 modify(fragment, regs);
989
990 switch (tmu.env) {
991 case GGL_REPLACE:
992 extract(fragment, texel, component);
993 break;
994 case GGL_MODULATE:
995 modulate(fragment, incoming, texel, component);
996 break;
997 case GGL_DECAL:
998 decal(fragment, incoming, texel, component);
999 break;
1000 case GGL_BLEND:
1001 blend(fragment, incoming, texel, component, i);
1002 break;
1003 case GGL_ADD:
1004 add(fragment, incoming, texel, component);
1005 break;
1006 }
1007 }
1008 }
1009 }
1010 }
1011
1012 // ---------------------------------------------------------------------------
1013
wrapping(int d,int coord,int size,int tx_wrap,int tx_linear)1014 void GGLAssembler::wrapping(
1015 int d,
1016 int coord, int size,
1017 int tx_wrap, int tx_linear)
1018 {
1019 // notes:
1020 // if tx_linear is set, we need 4 extra bits of precision on the result
1021 // SMULL/UMULL is 3 cycles
1022 Scratch scratches(registerFile());
1023 int c = coord;
1024 if (tx_wrap == GGL_NEEDS_WRAP_REPEAT) {
1025 // UMULL takes 4 cycles (interlocked), and we can get away with
1026 // 2 cycles using SMULWB, but we're loosing 16 bits of precision
1027 // out of 32 (this is not a problem because the iterator keeps
1028 // its full precision)
1029 // UMULL(AL, 0, size, d, c, size);
1030 // note: we can't use SMULTB because it's signed.
1031 MOV(AL, 0, d, reg_imm(c, LSR, 16-tx_linear));
1032 SMULWB(AL, d, d, size);
1033 } else if (tx_wrap == GGL_NEEDS_WRAP_CLAMP_TO_EDGE) {
1034 if (tx_linear) {
1035 // 1 cycle
1036 MOV(AL, 0, d, reg_imm(coord, ASR, 16-tx_linear));
1037 } else {
1038 // 4 cycles (common case)
1039 MOV(AL, 0, d, reg_imm(coord, ASR, 16));
1040 BIC(AL, 0, d, d, reg_imm(d, ASR, 31));
1041 CMP(AL, d, size);
1042 SUB(GE, 0, d, size, imm(1));
1043 }
1044 }
1045 }
1046
1047 // ---------------------------------------------------------------------------
1048
modulate(component_t & dest,const component_t & incoming,const pixel_t & incomingTexel,int component)1049 void GGLAssembler::modulate(
1050 component_t& dest,
1051 const component_t& incoming,
1052 const pixel_t& incomingTexel, int component)
1053 {
1054 Scratch locals(registerFile());
1055 integer_t texel(locals.obtain(), 32, CORRUPTIBLE);
1056 extract(texel, incomingTexel, component);
1057
1058 const int Nt = texel.size();
1059 // Nt should always be less than 10 bits because it comes
1060 // from the TMU.
1061
1062 int Ni = incoming.size();
1063 // Ni could be big because it comes from previous MODULATEs
1064
1065 if (Nt == 1) {
1066 // texel acts as a bit-mask
1067 // dest = incoming & ((texel << incoming.h)-texel)
1068 RSB(AL, 0, dest.reg, texel.reg, reg_imm(texel.reg, LSL, incoming.h));
1069 AND(AL, 0, dest.reg, dest.reg, incoming.reg);
1070 dest.l = incoming.l;
1071 dest.h = incoming.h;
1072 dest.flags |= (incoming.flags & CLEAR_LO);
1073 } else if (Ni == 1) {
1074 MOV(AL, 0, dest.reg, reg_imm(incoming.reg, LSL, 31-incoming.h));
1075 AND(AL, 0, dest.reg, texel.reg, reg_imm(dest.reg, ASR, 31));
1076 dest.l = 0;
1077 dest.h = Nt;
1078 } else {
1079 int inReg = incoming.reg;
1080 int shift = incoming.l;
1081 if ((Nt + Ni) > 32) {
1082 // we will overflow, reduce the precision of Ni to 8 bits
1083 // (Note Nt cannot be more than 10 bits which happens with
1084 // 565 textures and GGL_LINEAR)
1085 shift += Ni-8;
1086 Ni = 8;
1087 }
1088
1089 // modulate by the component with the lowest precision
1090 if (Nt >= Ni) {
1091 if (shift) {
1092 // XXX: we should be able to avoid this shift
1093 // when shift==16 && Nt<16 && Ni<16, in which
1094 // we could use SMULBT below.
1095 MOV(AL, 0, dest.reg, reg_imm(inReg, LSR, shift));
1096 inReg = dest.reg;
1097 shift = 0;
1098 }
1099 // operation: (Cf*Ct)/((1<<Ni)-1)
1100 // approximated with: Cf*(Ct + Ct>>(Ni-1))>>Ni
1101 // this operation doesn't change texel's size
1102 ADD(AL, 0, dest.reg, inReg, reg_imm(inReg, LSR, Ni-1));
1103 if (Nt<16 && Ni<16) SMULBB(AL, dest.reg, texel.reg, dest.reg);
1104 else MUL(AL, 0, dest.reg, texel.reg, dest.reg);
1105 dest.l = Ni;
1106 dest.h = Nt + Ni;
1107 } else {
1108 if (shift && (shift != 16)) {
1109 // if shift==16, we can use 16-bits mul instructions later
1110 MOV(AL, 0, dest.reg, reg_imm(inReg, LSR, shift));
1111 inReg = dest.reg;
1112 shift = 0;
1113 }
1114 // operation: (Cf*Ct)/((1<<Nt)-1)
1115 // approximated with: Ct*(Cf + Cf>>(Nt-1))>>Nt
1116 // this operation doesn't change incoming's size
1117 Scratch scratches(registerFile());
1118 int t = (texel.flags & CORRUPTIBLE) ? texel.reg : dest.reg;
1119 if (t == inReg)
1120 t = scratches.obtain();
1121 ADD(AL, 0, t, texel.reg, reg_imm(texel.reg, LSR, Nt-1));
1122 if (Nt<16 && Ni<16) {
1123 if (shift==16) SMULBT(AL, dest.reg, t, inReg);
1124 else SMULBB(AL, dest.reg, t, inReg);
1125 } else MUL(AL, 0, dest.reg, t, inReg);
1126 dest.l = Nt;
1127 dest.h = Nt + Ni;
1128 }
1129
1130 // low bits are not valid
1131 dest.flags |= CLEAR_LO;
1132
1133 // no need to keep more than 8 bits/component
1134 if (dest.size() > 8)
1135 dest.l = dest.h-8;
1136 }
1137 }
1138
decal(component_t & dest,const component_t & incoming,const pixel_t & incomingTexel,int component)1139 void GGLAssembler::decal(
1140 component_t& dest,
1141 const component_t& incoming,
1142 const pixel_t& incomingTexel, int component)
1143 {
1144 // RGBA:
1145 // Cv = Cf*(1 - At) + Ct*At = Cf + (Ct - Cf)*At
1146 // Av = Af
1147 Scratch locals(registerFile());
1148 integer_t texel(locals.obtain(), 32, CORRUPTIBLE);
1149 integer_t factor(locals.obtain(), 32, CORRUPTIBLE);
1150 extract(texel, incomingTexel, component);
1151 extract(factor, incomingTexel, GGLFormat::ALPHA);
1152
1153 // no need to keep more than 8-bits for decal
1154 int Ni = incoming.size();
1155 int shift = incoming.l;
1156 if (Ni > 8) {
1157 shift += Ni-8;
1158 Ni = 8;
1159 }
1160 integer_t incomingNorm(incoming.reg, Ni, incoming.flags);
1161 if (shift) {
1162 MOV(AL, 0, dest.reg, reg_imm(incomingNorm.reg, LSR, shift));
1163 incomingNorm.reg = dest.reg;
1164 incomingNorm.flags |= CORRUPTIBLE;
1165 }
1166 ADD(AL, 0, factor.reg, factor.reg, reg_imm(factor.reg, LSR, factor.s-1));
1167 build_blendOneMinusFF(dest, factor, incomingNorm, texel);
1168 }
1169
blend(component_t & dest,const component_t & incoming,const pixel_t & incomingTexel,int component,int tmu)1170 void GGLAssembler::blend(
1171 component_t& dest,
1172 const component_t& incoming,
1173 const pixel_t& incomingTexel, int component, int tmu)
1174 {
1175 // RGBA:
1176 // Cv = (1 - Ct)*Cf + Ct*Cc = Cf + (Cc - Cf)*Ct
1177 // Av = At*Af
1178
1179 if (component == GGLFormat::ALPHA) {
1180 modulate(dest, incoming, incomingTexel, component);
1181 return;
1182 }
1183
1184 Scratch locals(registerFile());
1185 integer_t color(locals.obtain(), 8, CORRUPTIBLE);
1186 integer_t factor(locals.obtain(), 32, CORRUPTIBLE);
1187 LDRB(AL, color.reg, mBuilderContext.Rctx,
1188 immed12_pre(GGL_OFFSETOF(state.texture[tmu].env_color[component])));
1189 extract(factor, incomingTexel, component);
1190
1191 // no need to keep more than 8-bits for blend
1192 int Ni = incoming.size();
1193 int shift = incoming.l;
1194 if (Ni > 8) {
1195 shift += Ni-8;
1196 Ni = 8;
1197 }
1198 integer_t incomingNorm(incoming.reg, Ni, incoming.flags);
1199 if (shift) {
1200 MOV(AL, 0, dest.reg, reg_imm(incomingNorm.reg, LSR, shift));
1201 incomingNorm.reg = dest.reg;
1202 incomingNorm.flags |= CORRUPTIBLE;
1203 }
1204 ADD(AL, 0, factor.reg, factor.reg, reg_imm(factor.reg, LSR, factor.s-1));
1205 build_blendOneMinusFF(dest, factor, incomingNorm, color);
1206 }
1207
add(component_t & dest,const component_t & incoming,const pixel_t & incomingTexel,int component)1208 void GGLAssembler::add(
1209 component_t& dest,
1210 const component_t& incoming,
1211 const pixel_t& incomingTexel, int component)
1212 {
1213 // RGBA:
1214 // Cv = Cf + Ct;
1215 Scratch locals(registerFile());
1216
1217 component_t incomingTemp(incoming);
1218
1219 // use "dest" as a temporary for extracting the texel, unless "dest"
1220 // overlaps "incoming".
1221 integer_t texel(dest.reg, 32, CORRUPTIBLE);
1222 if (dest.reg == incomingTemp.reg)
1223 texel.reg = locals.obtain();
1224 extract(texel, incomingTexel, component);
1225
1226 if (texel.s < incomingTemp.size()) {
1227 expand(texel, texel, incomingTemp.size());
1228 } else if (texel.s > incomingTemp.size()) {
1229 if (incomingTemp.flags & CORRUPTIBLE) {
1230 expand(incomingTemp, incomingTemp, texel.s);
1231 } else {
1232 incomingTemp.reg = locals.obtain();
1233 expand(incomingTemp, incoming, texel.s);
1234 }
1235 }
1236
1237 if (incomingTemp.l) {
1238 ADD(AL, 0, dest.reg, texel.reg,
1239 reg_imm(incomingTemp.reg, LSR, incomingTemp.l));
1240 } else {
1241 ADD(AL, 0, dest.reg, texel.reg, incomingTemp.reg);
1242 }
1243 dest.l = 0;
1244 dest.h = texel.size();
1245 component_sat(dest);
1246 }
1247
1248 // ----------------------------------------------------------------------------
1249
1250 }; // namespace android
1251
1252