1 /* libs/pixelflinger/codeflinger/GGLAssembler.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 #define LOG_TAG "GGLAssembler"
19
20 #include <assert.h>
21 #include <stdint.h>
22 #include <stdlib.h>
23 #include <stdio.h>
24 #include <sys/types.h>
25 #include <cutils/log.h>
26
27 #include "GGLAssembler.h"
28
29 namespace android {
30
31 // ----------------------------------------------------------------------------
32
GGLAssembler(ARMAssemblerInterface * target)33 GGLAssembler::GGLAssembler(ARMAssemblerInterface* target)
34 : ARMAssemblerProxy(target),
35 RegisterAllocator(ARMAssemblerProxy::getCodegenArch()), mOptLevel(7)
36 {
37 }
38
~GGLAssembler()39 GGLAssembler::~GGLAssembler()
40 {
41 }
42
prolog()43 void GGLAssembler::prolog()
44 {
45 ARMAssemblerProxy::prolog();
46 }
47
epilog(uint32_t touched)48 void GGLAssembler::epilog(uint32_t touched)
49 {
50 ARMAssemblerProxy::epilog(touched);
51 }
52
reset(int opt_level)53 void GGLAssembler::reset(int opt_level)
54 {
55 ARMAssemblerProxy::reset();
56 RegisterAllocator::reset();
57 mOptLevel = opt_level;
58 }
59
60 // ---------------------------------------------------------------------------
61
scanline(const needs_t & needs,context_t const * c)62 int GGLAssembler::scanline(const needs_t& needs, context_t const* c)
63 {
64 int err = 0;
65 int opt_level = mOptLevel;
66 while (opt_level >= 0) {
67 reset(opt_level);
68 err = scanline_core(needs, c);
69 if (err == 0)
70 break;
71 opt_level--;
72 }
73
74 // XXX: in theory, pcForLabel is not valid before generate()
75 uint32_t* fragment_start_pc = pcForLabel("fragment_loop");
76 uint32_t* fragment_end_pc = pcForLabel("epilog");
77 const int per_fragment_ops = int(fragment_end_pc - fragment_start_pc);
78
79 // build a name for our pipeline
80 char name[64];
81 sprintf(name,
82 "scanline__%08X:%08X_%08X_%08X [%3d ipp]",
83 needs.p, needs.n, needs.t[0], needs.t[1], per_fragment_ops);
84
85 if (err) {
86 ALOGE("Error while generating ""%s""\n", name);
87 disassemble(name);
88 return -1;
89 }
90
91 return generate(name);
92 }
93
scanline_core(const needs_t & needs,context_t const * c)94 int GGLAssembler::scanline_core(const needs_t& needs, context_t const* c)
95 {
96 int64_t duration = ggl_system_time();
97
98 mBlendFactorCached = 0;
99 mBlending = 0;
100 mMasking = 0;
101 mAA = GGL_READ_NEEDS(P_AA, needs.p);
102 mDithering = GGL_READ_NEEDS(P_DITHER, needs.p);
103 mAlphaTest = GGL_READ_NEEDS(P_ALPHA_TEST, needs.p) + GGL_NEVER;
104 mDepthTest = GGL_READ_NEEDS(P_DEPTH_TEST, needs.p) + GGL_NEVER;
105 mFog = GGL_READ_NEEDS(P_FOG, needs.p) != 0;
106 mSmooth = GGL_READ_NEEDS(SHADE, needs.n) != 0;
107 mBuilderContext.needs = needs;
108 mBuilderContext.c = c;
109 mBuilderContext.Rctx = reserveReg(R0); // context always in R0
110 mCbFormat = c->formats[ GGL_READ_NEEDS(CB_FORMAT, needs.n) ];
111
112 // ------------------------------------------------------------------------
113
114 decodeLogicOpNeeds(needs);
115
116 decodeTMUNeeds(needs, c);
117
118 mBlendSrc = ggl_needs_to_blendfactor(GGL_READ_NEEDS(BLEND_SRC, needs.n));
119 mBlendDst = ggl_needs_to_blendfactor(GGL_READ_NEEDS(BLEND_DST, needs.n));
120 mBlendSrcA = ggl_needs_to_blendfactor(GGL_READ_NEEDS(BLEND_SRCA, needs.n));
121 mBlendDstA = ggl_needs_to_blendfactor(GGL_READ_NEEDS(BLEND_DSTA, needs.n));
122
123 if (!mCbFormat.c[GGLFormat::ALPHA].h) {
124 if ((mBlendSrc == GGL_ONE_MINUS_DST_ALPHA) ||
125 (mBlendSrc == GGL_DST_ALPHA)) {
126 mBlendSrc = GGL_ONE;
127 }
128 if ((mBlendSrcA == GGL_ONE_MINUS_DST_ALPHA) ||
129 (mBlendSrcA == GGL_DST_ALPHA)) {
130 mBlendSrcA = GGL_ONE;
131 }
132 if ((mBlendDst == GGL_ONE_MINUS_DST_ALPHA) ||
133 (mBlendDst == GGL_DST_ALPHA)) {
134 mBlendDst = GGL_ONE;
135 }
136 if ((mBlendDstA == GGL_ONE_MINUS_DST_ALPHA) ||
137 (mBlendDstA == GGL_DST_ALPHA)) {
138 mBlendDstA = GGL_ONE;
139 }
140 }
141
142 // if we need the framebuffer, read it now
143 const int blending = blending_codes(mBlendSrc, mBlendDst) |
144 blending_codes(mBlendSrcA, mBlendDstA);
145
146 // XXX: handle special cases, destination not modified...
147 if ((mBlendSrc==GGL_ZERO) && (mBlendSrcA==GGL_ZERO) &&
148 (mBlendDst==GGL_ONE) && (mBlendDstA==GGL_ONE)) {
149 // Destination unmodified (beware of logic ops)
150 } else if ((mBlendSrc==GGL_ZERO) && (mBlendSrcA==GGL_ZERO) &&
151 (mBlendDst==GGL_ZERO) && (mBlendDstA==GGL_ZERO)) {
152 // Destination is zero (beware of logic ops)
153 }
154
155 int fbComponents = 0;
156 const int masking = GGL_READ_NEEDS(MASK_ARGB, needs.n);
157 for (int i=0 ; i<4 ; i++) {
158 const int mask = 1<<i;
159 component_info_t& info = mInfo[i];
160 int fs = i==GGLFormat::ALPHA ? mBlendSrcA : mBlendSrc;
161 int fd = i==GGLFormat::ALPHA ? mBlendDstA : mBlendDst;
162 if (fs==GGL_SRC_ALPHA_SATURATE && i==GGLFormat::ALPHA)
163 fs = GGL_ONE;
164 info.masked = !!(masking & mask);
165 info.inDest = !info.masked && mCbFormat.c[i].h &&
166 ((mLogicOp & LOGIC_OP_SRC) || (!mLogicOp));
167 if (mCbFormat.components >= GGL_LUMINANCE &&
168 (i==GGLFormat::GREEN || i==GGLFormat::BLUE)) {
169 info.inDest = false;
170 }
171 info.needed = (i==GGLFormat::ALPHA) &&
172 (isAlphaSourceNeeded() || mAlphaTest != GGL_ALWAYS);
173 info.replaced = !!(mTextureMachine.replaced & mask);
174 info.iterated = (!info.replaced && (info.inDest || info.needed));
175 info.smooth = mSmooth && info.iterated;
176 info.fog = mFog && info.inDest && (i != GGLFormat::ALPHA);
177 info.blend = (fs != int(GGL_ONE)) || (fd > int(GGL_ZERO));
178
179 mBlending |= (info.blend ? mask : 0);
180 mMasking |= (mCbFormat.c[i].h && info.masked) ? mask : 0;
181 fbComponents |= mCbFormat.c[i].h ? mask : 0;
182 }
183
184 mAllMasked = (mMasking == fbComponents);
185 if (mAllMasked) {
186 mDithering = 0;
187 }
188
189 fragment_parts_t parts;
190
191 // ------------------------------------------------------------------------
192 prolog();
193 // ------------------------------------------------------------------------
194
195 build_scanline_prolog(parts, needs);
196
197 if (registerFile().status())
198 return registerFile().status();
199
200 // ------------------------------------------------------------------------
201 label("fragment_loop");
202 // ------------------------------------------------------------------------
203 {
204 Scratch regs(registerFile());
205
206 if (mDithering) {
207 // update the dither index.
208 MOV(AL, 0, parts.count.reg,
209 reg_imm(parts.count.reg, ROR, GGL_DITHER_ORDER_SHIFT));
210 ADD(AL, 0, parts.count.reg, parts.count.reg,
211 imm( 1 << (32 - GGL_DITHER_ORDER_SHIFT)));
212 MOV(AL, 0, parts.count.reg,
213 reg_imm(parts.count.reg, ROR, 32 - GGL_DITHER_ORDER_SHIFT));
214 }
215
216 // XXX: could we do an early alpha-test here in some cases?
217 // It would probaly be used only with smooth-alpha and no texture
218 // (or no alpha component in the texture).
219
220 // Early z-test
221 if (mAlphaTest==GGL_ALWAYS) {
222 build_depth_test(parts, Z_TEST|Z_WRITE);
223 } else {
224 // we cannot do the z-write here, because
225 // it might be killed by the alpha-test later
226 build_depth_test(parts, Z_TEST);
227 }
228
229 { // texture coordinates
230 Scratch scratches(registerFile());
231
232 // texel generation
233 build_textures(parts, regs);
234 if (registerFile().status())
235 return registerFile().status();
236 }
237
238 if ((blending & (FACTOR_DST|BLEND_DST)) ||
239 (mMasking && !mAllMasked) ||
240 (mLogicOp & LOGIC_OP_DST))
241 {
242 // blending / logic_op / masking need the framebuffer
243 mDstPixel.setTo(regs.obtain(), &mCbFormat);
244
245 // load the framebuffer pixel
246 comment("fetch color-buffer");
247 load(parts.cbPtr, mDstPixel);
248 }
249
250 if (registerFile().status())
251 return registerFile().status();
252
253 pixel_t pixel;
254 int directTex = mTextureMachine.directTexture;
255 if (directTex | parts.packed) {
256 // note: we can't have both here
257 // iterated color or direct texture
258 pixel = directTex ? parts.texel[directTex-1] : parts.iterated;
259 pixel.flags &= ~CORRUPTIBLE;
260 } else {
261 if (mDithering) {
262 const int ctxtReg = mBuilderContext.Rctx;
263 const int mask = GGL_DITHER_SIZE-1;
264 parts.dither = reg_t(regs.obtain());
265 AND(AL, 0, parts.dither.reg, parts.count.reg, imm(mask));
266 ADDR_ADD(AL, 0, parts.dither.reg, ctxtReg, parts.dither.reg);
267 LDRB(AL, parts.dither.reg, parts.dither.reg,
268 immed12_pre(GGL_OFFSETOF(ditherMatrix)));
269 }
270
271 // allocate a register for the resulting pixel
272 pixel.setTo(regs.obtain(), &mCbFormat, FIRST);
273
274 build_component(pixel, parts, GGLFormat::ALPHA, regs);
275
276 if (mAlphaTest!=GGL_ALWAYS) {
277 // only handle the z-write part here. We know z-test
278 // was successful, as well as alpha-test.
279 build_depth_test(parts, Z_WRITE);
280 }
281
282 build_component(pixel, parts, GGLFormat::RED, regs);
283 build_component(pixel, parts, GGLFormat::GREEN, regs);
284 build_component(pixel, parts, GGLFormat::BLUE, regs);
285
286 pixel.flags |= CORRUPTIBLE;
287 }
288
289 if (registerFile().status())
290 return registerFile().status();
291
292 if (pixel.reg == -1) {
293 // be defensive here. if we're here it's probably
294 // that this whole fragment is a no-op.
295 pixel = mDstPixel;
296 }
297
298 if (!mAllMasked) {
299 // logic operation
300 build_logic_op(pixel, regs);
301
302 // masking
303 build_masking(pixel, regs);
304
305 comment("store");
306 store(parts.cbPtr, pixel, WRITE_BACK);
307 }
308 }
309
310 if (registerFile().status())
311 return registerFile().status();
312
313 // update the iterated color...
314 if (parts.reload != 3) {
315 build_smooth_shade(parts);
316 }
317
318 // update iterated z
319 build_iterate_z(parts);
320
321 // update iterated fog
322 build_iterate_f(parts);
323
324 SUB(AL, S, parts.count.reg, parts.count.reg, imm(1<<16));
325 B(PL, "fragment_loop");
326 label("epilog");
327 epilog(registerFile().touched());
328
329 if ((mAlphaTest!=GGL_ALWAYS) || (mDepthTest!=GGL_ALWAYS)) {
330 if (mDepthTest!=GGL_ALWAYS) {
331 label("discard_before_textures");
332 build_iterate_texture_coordinates(parts);
333 }
334 label("discard_after_textures");
335 build_smooth_shade(parts);
336 build_iterate_z(parts);
337 build_iterate_f(parts);
338 if (!mAllMasked) {
339 ADDR_ADD(AL, 0, parts.cbPtr.reg, parts.cbPtr.reg, imm(parts.cbPtr.size>>3));
340 }
341 SUB(AL, S, parts.count.reg, parts.count.reg, imm(1<<16));
342 B(PL, "fragment_loop");
343 epilog(registerFile().touched());
344 }
345
346 return registerFile().status();
347 }
348
349 // ---------------------------------------------------------------------------
350
build_scanline_prolog(fragment_parts_t & parts,const needs_t & needs)351 void GGLAssembler::build_scanline_prolog(
352 fragment_parts_t& parts, const needs_t& needs)
353 {
354 Scratch scratches(registerFile());
355 int Rctx = mBuilderContext.Rctx;
356
357 // compute count
358 comment("compute ct (# of pixels to process)");
359 parts.count.setTo(obtainReg());
360 int Rx = scratches.obtain();
361 int Ry = scratches.obtain();
362 CONTEXT_LOAD(Rx, iterators.xl);
363 CONTEXT_LOAD(parts.count.reg, iterators.xr);
364 CONTEXT_LOAD(Ry, iterators.y);
365
366 // parts.count = iterators.xr - Rx
367 SUB(AL, 0, parts.count.reg, parts.count.reg, Rx);
368 SUB(AL, 0, parts.count.reg, parts.count.reg, imm(1));
369
370 if (mDithering) {
371 // parts.count.reg = 0xNNNNXXDD
372 // NNNN = count-1
373 // DD = dither offset
374 // XX = 0xxxxxxx (x = garbage)
375 Scratch scratches(registerFile());
376 int tx = scratches.obtain();
377 int ty = scratches.obtain();
378 AND(AL, 0, tx, Rx, imm(GGL_DITHER_MASK));
379 AND(AL, 0, ty, Ry, imm(GGL_DITHER_MASK));
380 ADD(AL, 0, tx, tx, reg_imm(ty, LSL, GGL_DITHER_ORDER_SHIFT));
381 ORR(AL, 0, parts.count.reg, tx, reg_imm(parts.count.reg, LSL, 16));
382 } else {
383 // parts.count.reg = 0xNNNN0000
384 // NNNN = count-1
385 MOV(AL, 0, parts.count.reg, reg_imm(parts.count.reg, LSL, 16));
386 }
387
388 if (!mAllMasked) {
389 // compute dst ptr
390 comment("compute color-buffer pointer");
391 const int cb_bits = mCbFormat.size*8;
392 int Rs = scratches.obtain();
393 parts.cbPtr.setTo(obtainReg(), cb_bits);
394 CONTEXT_LOAD(Rs, state.buffers.color.stride);
395 CONTEXT_ADDR_LOAD(parts.cbPtr.reg, state.buffers.color.data);
396 SMLABB(AL, Rs, Ry, Rs, Rx); // Rs = Rx + Ry*Rs
397 base_offset(parts.cbPtr, parts.cbPtr, Rs);
398 scratches.recycle(Rs);
399 }
400
401 // init fog
402 const int need_fog = GGL_READ_NEEDS(P_FOG, needs.p);
403 if (need_fog) {
404 comment("compute initial fog coordinate");
405 Scratch scratches(registerFile());
406 int dfdx = scratches.obtain();
407 int ydfdy = scratches.obtain();
408 int f = ydfdy;
409 CONTEXT_LOAD(dfdx, generated_vars.dfdx);
410 CONTEXT_LOAD(ydfdy, iterators.ydfdy);
411 MLA(AL, 0, f, Rx, dfdx, ydfdy);
412 CONTEXT_STORE(f, generated_vars.f);
413 }
414
415 // init Z coordinate
416 if ((mDepthTest != GGL_ALWAYS) || GGL_READ_NEEDS(P_MASK_Z, needs.p)) {
417 parts.z = reg_t(obtainReg());
418 comment("compute initial Z coordinate");
419 Scratch scratches(registerFile());
420 int dzdx = scratches.obtain();
421 int ydzdy = parts.z.reg;
422 CONTEXT_LOAD(dzdx, generated_vars.dzdx); // 1.31 fixed-point
423 CONTEXT_LOAD(ydzdy, iterators.ydzdy); // 1.31 fixed-point
424 MLA(AL, 0, parts.z.reg, Rx, dzdx, ydzdy);
425
426 // we're going to index zbase of parts.count
427 // zbase = base + (xl-count + stride*y)*2
428 int Rs = dzdx;
429 int zbase = scratches.obtain();
430 CONTEXT_LOAD(Rs, state.buffers.depth.stride);
431 CONTEXT_ADDR_LOAD(zbase, state.buffers.depth.data);
432 SMLABB(AL, Rs, Ry, Rs, Rx);
433 ADD(AL, 0, Rs, Rs, reg_imm(parts.count.reg, LSR, 16));
434 ADDR_ADD(AL, 0, zbase, zbase, reg_imm(Rs, LSL, 1));
435 CONTEXT_ADDR_STORE(zbase, generated_vars.zbase);
436 }
437
438 // init texture coordinates
439 init_textures(parts.coords, reg_t(Rx), reg_t(Ry));
440 scratches.recycle(Ry);
441
442 // iterated color
443 init_iterated_color(parts, reg_t(Rx));
444
445 // init coverage factor application (anti-aliasing)
446 if (mAA) {
447 parts.covPtr.setTo(obtainReg(), 16);
448 CONTEXT_ADDR_LOAD(parts.covPtr.reg, state.buffers.coverage);
449 ADDR_ADD(AL, 0, parts.covPtr.reg, parts.covPtr.reg, reg_imm(Rx, LSL, 1));
450 }
451 }
452
453 // ---------------------------------------------------------------------------
454
build_component(pixel_t & pixel,const fragment_parts_t & parts,int component,Scratch & regs)455 void GGLAssembler::build_component( pixel_t& pixel,
456 const fragment_parts_t& parts,
457 int component,
458 Scratch& regs)
459 {
460 static char const * comments[] = {"alpha", "red", "green", "blue"};
461 comment(comments[component]);
462
463 // local register file
464 Scratch scratches(registerFile());
465 const int dst_component_size = pixel.component_size(component);
466
467 component_t temp(-1);
468 build_incoming_component( temp, dst_component_size,
469 parts, component, scratches, regs);
470
471 if (mInfo[component].inDest) {
472
473 // blending...
474 build_blending( temp, mDstPixel, component, scratches );
475
476 // downshift component and rebuild pixel...
477 downshift(pixel, component, temp, parts.dither);
478 }
479 }
480
build_incoming_component(component_t & temp,int dst_size,const fragment_parts_t & parts,int component,Scratch & scratches,Scratch & global_regs)481 void GGLAssembler::build_incoming_component(
482 component_t& temp,
483 int dst_size,
484 const fragment_parts_t& parts,
485 int component,
486 Scratch& scratches,
487 Scratch& global_regs)
488 {
489 const uint32_t component_mask = 1<<component;
490
491 // Figure out what we need for the blending stage...
492 int fs = component==GGLFormat::ALPHA ? mBlendSrcA : mBlendSrc;
493 int fd = component==GGLFormat::ALPHA ? mBlendDstA : mBlendDst;
494 if (fs==GGL_SRC_ALPHA_SATURATE && component==GGLFormat::ALPHA) {
495 fs = GGL_ONE;
496 }
497
498 // Figure out what we need to extract and for what reason
499 const int blending = blending_codes(fs, fd);
500
501 // Are we actually going to blend?
502 const int need_blending = (fs != int(GGL_ONE)) || (fd > int(GGL_ZERO));
503
504 // expand the source if the destination has more bits
505 int need_expander = false;
506 for (int i=0 ; i<GGL_TEXTURE_UNIT_COUNT-1 ; i++) {
507 texture_unit_t& tmu = mTextureMachine.tmu[i];
508 if ((tmu.format_idx) &&
509 (parts.texel[i].component_size(component) < dst_size)) {
510 need_expander = true;
511 }
512 }
513
514 // do we need to extract this component?
515 const bool multiTexture = mTextureMachine.activeUnits > 1;
516 const int blend_needs_alpha_source = (component==GGLFormat::ALPHA) &&
517 (isAlphaSourceNeeded());
518 int need_extract = mInfo[component].needed;
519 if (mInfo[component].inDest)
520 {
521 need_extract |= ((need_blending ?
522 (blending & (BLEND_SRC|FACTOR_SRC)) : need_expander));
523 need_extract |= (mTextureMachine.mask != mTextureMachine.replaced);
524 need_extract |= mInfo[component].smooth;
525 need_extract |= mInfo[component].fog;
526 need_extract |= mDithering;
527 need_extract |= multiTexture;
528 }
529
530 if (need_extract) {
531 Scratch& regs = blend_needs_alpha_source ? global_regs : scratches;
532 component_t fragment;
533
534 // iterated color
535 build_iterated_color(fragment, parts, component, regs);
536
537 // texture environement (decal, modulate, replace)
538 build_texture_environment(fragment, parts, component, regs);
539
540 // expand the source if the destination has more bits
541 if (need_expander && (fragment.size() < dst_size)) {
542 // we're here only if we fetched a texel
543 // (so we know for sure fragment is CORRUPTIBLE)
544 expand(fragment, fragment, dst_size);
545 }
546
547 // We have a few specific things to do for the alpha-channel
548 if ((component==GGLFormat::ALPHA) &&
549 (mInfo[component].needed || fragment.size()<dst_size))
550 {
551 // convert to integer_t first and make sure
552 // we don't corrupt a needed register
553 if (fragment.l) {
554 component_t incoming(fragment);
555 modify(fragment, regs);
556 MOV(AL, 0, fragment.reg, reg_imm(incoming.reg, LSR, incoming.l));
557 fragment.h -= fragment.l;
558 fragment.l = 0;
559 }
560
561 // coverage factor application
562 build_coverage_application(fragment, parts, regs);
563
564 // alpha-test
565 build_alpha_test(fragment, parts);
566
567 if (blend_needs_alpha_source) {
568 // We keep only 8 bits for the blending stage
569 const int shift = fragment.h <= 8 ? 0 : fragment.h-8;
570 if (fragment.flags & CORRUPTIBLE) {
571 fragment.flags &= ~CORRUPTIBLE;
572 mAlphaSource.setTo(fragment.reg,
573 fragment.size(), fragment.flags);
574 if (shift) {
575 MOV(AL, 0, mAlphaSource.reg,
576 reg_imm(mAlphaSource.reg, LSR, shift));
577 }
578 } else {
579 // XXX: it would better to do this in build_blend_factor()
580 // so we can avoid the extra MOV below.
581 mAlphaSource.setTo(regs.obtain(),
582 fragment.size(), CORRUPTIBLE);
583 if (shift) {
584 MOV(AL, 0, mAlphaSource.reg,
585 reg_imm(fragment.reg, LSR, shift));
586 } else {
587 MOV(AL, 0, mAlphaSource.reg, fragment.reg);
588 }
589 }
590 mAlphaSource.s -= shift;
591 }
592 }
593
594 // fog...
595 build_fog( fragment, component, regs );
596
597 temp = fragment;
598 } else {
599 if (mInfo[component].inDest) {
600 // extraction not needed and replace
601 // we just select the right component
602 if ((mTextureMachine.replaced & component_mask) == 0) {
603 // component wasn't replaced, so use it!
604 temp = component_t(parts.iterated, component);
605 }
606 for (int i=0 ; i<GGL_TEXTURE_UNIT_COUNT ; i++) {
607 const texture_unit_t& tmu = mTextureMachine.tmu[i];
608 if ((tmu.mask & component_mask) &&
609 ((tmu.replaced & component_mask) == 0)) {
610 temp = component_t(parts.texel[i], component);
611 }
612 }
613 }
614 }
615 }
616
isAlphaSourceNeeded() const617 bool GGLAssembler::isAlphaSourceNeeded() const
618 {
619 // XXX: also needed for alpha-test
620 const int bs = mBlendSrc;
621 const int bd = mBlendDst;
622 return bs==GGL_SRC_ALPHA_SATURATE ||
623 bs==GGL_SRC_ALPHA || bs==GGL_ONE_MINUS_SRC_ALPHA ||
624 bd==GGL_SRC_ALPHA || bd==GGL_ONE_MINUS_SRC_ALPHA ;
625 }
626
627 // ---------------------------------------------------------------------------
628
build_smooth_shade(const fragment_parts_t & parts)629 void GGLAssembler::build_smooth_shade(const fragment_parts_t& parts)
630 {
631 if (mSmooth && !parts.iterated_packed) {
632 // update the iterated color in a pipelined way...
633 comment("update iterated color");
634 Scratch scratches(registerFile());
635
636 const int reload = parts.reload;
637 for (int i=0 ; i<4 ; i++) {
638 if (!mInfo[i].iterated)
639 continue;
640
641 int c = parts.argb[i].reg;
642 int dx = parts.argb_dx[i].reg;
643
644 if (reload & 1) {
645 c = scratches.obtain();
646 CONTEXT_LOAD(c, generated_vars.argb[i].c);
647 }
648 if (reload & 2) {
649 dx = scratches.obtain();
650 CONTEXT_LOAD(dx, generated_vars.argb[i].dx);
651 }
652
653 if (mSmooth) {
654 ADD(AL, 0, c, c, dx);
655 }
656
657 if (reload & 1) {
658 CONTEXT_STORE(c, generated_vars.argb[i].c);
659 scratches.recycle(c);
660 }
661 if (reload & 2) {
662 scratches.recycle(dx);
663 }
664 }
665 }
666 }
667
668 // ---------------------------------------------------------------------------
669
build_coverage_application(component_t & fragment,const fragment_parts_t & parts,Scratch & regs)670 void GGLAssembler::build_coverage_application(component_t& fragment,
671 const fragment_parts_t& parts, Scratch& regs)
672 {
673 // here fragment.l is guarenteed to be 0
674 if (mAA) {
675 // coverages are 1.15 fixed-point numbers
676 comment("coverage application");
677
678 component_t incoming(fragment);
679 modify(fragment, regs);
680
681 Scratch scratches(registerFile());
682 int cf = scratches.obtain();
683 LDRH(AL, cf, parts.covPtr.reg, immed8_post(2));
684 if (fragment.h > 31) {
685 fragment.h--;
686 SMULWB(AL, fragment.reg, incoming.reg, cf);
687 } else {
688 MOV(AL, 0, fragment.reg, reg_imm(incoming.reg, LSL, 1));
689 SMULWB(AL, fragment.reg, fragment.reg, cf);
690 }
691 }
692 }
693
694 // ---------------------------------------------------------------------------
695
build_alpha_test(component_t & fragment,const fragment_parts_t &)696 void GGLAssembler::build_alpha_test(component_t& fragment,
697 const fragment_parts_t& /*parts*/)
698 {
699 if (mAlphaTest != GGL_ALWAYS) {
700 comment("Alpha Test");
701 Scratch scratches(registerFile());
702 int ref = scratches.obtain();
703 const int shift = GGL_COLOR_BITS-fragment.size();
704 CONTEXT_LOAD(ref, state.alpha_test.ref);
705 if (shift) CMP(AL, fragment.reg, reg_imm(ref, LSR, shift));
706 else CMP(AL, fragment.reg, ref);
707 int cc = NV;
708 switch (mAlphaTest) {
709 case GGL_NEVER: cc = NV; break;
710 case GGL_LESS: cc = LT; break;
711 case GGL_EQUAL: cc = EQ; break;
712 case GGL_LEQUAL: cc = LS; break;
713 case GGL_GREATER: cc = HI; break;
714 case GGL_NOTEQUAL: cc = NE; break;
715 case GGL_GEQUAL: cc = HS; break;
716 }
717 B(cc^1, "discard_after_textures");
718 }
719 }
720
721 // ---------------------------------------------------------------------------
722
build_depth_test(const fragment_parts_t & parts,uint32_t mask)723 void GGLAssembler::build_depth_test(
724 const fragment_parts_t& parts, uint32_t mask)
725 {
726 mask &= Z_TEST|Z_WRITE;
727 const needs_t& needs = mBuilderContext.needs;
728 const int zmask = GGL_READ_NEEDS(P_MASK_Z, needs.p);
729 Scratch scratches(registerFile());
730
731 if (mDepthTest != GGL_ALWAYS || zmask) {
732 int cc=AL, ic=AL;
733 switch (mDepthTest) {
734 case GGL_LESS: ic = HI; break;
735 case GGL_EQUAL: ic = EQ; break;
736 case GGL_LEQUAL: ic = HS; break;
737 case GGL_GREATER: ic = LT; break;
738 case GGL_NOTEQUAL: ic = NE; break;
739 case GGL_GEQUAL: ic = LS; break;
740 case GGL_NEVER:
741 // this never happens, because it's taken care of when
742 // computing the needs. but we keep it for completness.
743 comment("Depth Test (NEVER)");
744 B(AL, "discard_before_textures");
745 return;
746 case GGL_ALWAYS:
747 // we're here because zmask is enabled
748 mask &= ~Z_TEST; // test always passes.
749 break;
750 }
751
752 // inverse the condition
753 cc = ic^1;
754
755 if ((mask & Z_WRITE) && !zmask) {
756 mask &= ~Z_WRITE;
757 }
758
759 if (!mask)
760 return;
761
762 comment("Depth Test");
763
764 int zbase = scratches.obtain();
765 int depth = scratches.obtain();
766 int z = parts.z.reg;
767
768 CONTEXT_ADDR_LOAD(zbase, generated_vars.zbase); // stall
769 ADDR_SUB(AL, 0, zbase, zbase, reg_imm(parts.count.reg, LSR, 15));
770 // above does zbase = zbase + ((count >> 16) << 1)
771
772 if (mask & Z_TEST) {
773 LDRH(AL, depth, zbase); // stall
774 CMP(AL, depth, reg_imm(z, LSR, 16));
775 B(cc, "discard_before_textures");
776 }
777 if (mask & Z_WRITE) {
778 if (mask == Z_WRITE) {
779 // only z-write asked, cc is meaningless
780 ic = AL;
781 }
782 MOV(AL, 0, depth, reg_imm(z, LSR, 16));
783 STRH(ic, depth, zbase);
784 }
785 }
786 }
787
build_iterate_z(const fragment_parts_t & parts)788 void GGLAssembler::build_iterate_z(const fragment_parts_t& parts)
789 {
790 const needs_t& needs = mBuilderContext.needs;
791 if ((mDepthTest != GGL_ALWAYS) || GGL_READ_NEEDS(P_MASK_Z, needs.p)) {
792 Scratch scratches(registerFile());
793 int dzdx = scratches.obtain();
794 CONTEXT_LOAD(dzdx, generated_vars.dzdx); // stall
795 ADD(AL, 0, parts.z.reg, parts.z.reg, dzdx);
796 }
797 }
798
build_iterate_f(const fragment_parts_t &)799 void GGLAssembler::build_iterate_f(const fragment_parts_t& /*parts*/)
800 {
801 const needs_t& needs = mBuilderContext.needs;
802 if (GGL_READ_NEEDS(P_FOG, needs.p)) {
803 Scratch scratches(registerFile());
804 int dfdx = scratches.obtain();
805 int f = scratches.obtain();
806 CONTEXT_LOAD(f, generated_vars.f);
807 CONTEXT_LOAD(dfdx, generated_vars.dfdx); // stall
808 ADD(AL, 0, f, f, dfdx);
809 CONTEXT_STORE(f, generated_vars.f);
810 }
811 }
812
813 // ---------------------------------------------------------------------------
814
build_logic_op(pixel_t & pixel,Scratch & regs)815 void GGLAssembler::build_logic_op(pixel_t& pixel, Scratch& regs)
816 {
817 const needs_t& needs = mBuilderContext.needs;
818 const int opcode = GGL_READ_NEEDS(LOGIC_OP, needs.n) | GGL_CLEAR;
819 if (opcode == GGL_COPY)
820 return;
821
822 comment("logic operation");
823
824 pixel_t s(pixel);
825 if (!(pixel.flags & CORRUPTIBLE)) {
826 pixel.reg = regs.obtain();
827 pixel.flags |= CORRUPTIBLE;
828 }
829
830 pixel_t d(mDstPixel);
831 switch(opcode) {
832 case GGL_CLEAR: MOV(AL, 0, pixel.reg, imm(0)); break;
833 case GGL_AND: AND(AL, 0, pixel.reg, s.reg, d.reg); break;
834 case GGL_AND_REVERSE: BIC(AL, 0, pixel.reg, s.reg, d.reg); break;
835 case GGL_COPY: break;
836 case GGL_AND_INVERTED: BIC(AL, 0, pixel.reg, d.reg, s.reg); break;
837 case GGL_NOOP: MOV(AL, 0, pixel.reg, d.reg); break;
838 case GGL_XOR: EOR(AL, 0, pixel.reg, s.reg, d.reg); break;
839 case GGL_OR: ORR(AL, 0, pixel.reg, s.reg, d.reg); break;
840 case GGL_NOR: ORR(AL, 0, pixel.reg, s.reg, d.reg);
841 MVN(AL, 0, pixel.reg, pixel.reg); break;
842 case GGL_EQUIV: EOR(AL, 0, pixel.reg, s.reg, d.reg);
843 MVN(AL, 0, pixel.reg, pixel.reg); break;
844 case GGL_INVERT: MVN(AL, 0, pixel.reg, d.reg); break;
845 case GGL_OR_REVERSE: // s | ~d == ~(~s & d)
846 BIC(AL, 0, pixel.reg, d.reg, s.reg);
847 MVN(AL, 0, pixel.reg, pixel.reg); break;
848 case GGL_COPY_INVERTED: MVN(AL, 0, pixel.reg, s.reg); break;
849 case GGL_OR_INVERTED: // ~s | d == ~(s & ~d)
850 BIC(AL, 0, pixel.reg, s.reg, d.reg);
851 MVN(AL, 0, pixel.reg, pixel.reg); break;
852 case GGL_NAND: AND(AL, 0, pixel.reg, s.reg, d.reg);
853 MVN(AL, 0, pixel.reg, pixel.reg); break;
854 case GGL_SET: MVN(AL, 0, pixel.reg, imm(0)); break;
855 };
856 }
857
858 // ---------------------------------------------------------------------------
859
find_bottom(uint32_t val)860 static uint32_t find_bottom(uint32_t val)
861 {
862 uint32_t i = 0;
863 while (!(val & (3<<i)))
864 i+= 2;
865 return i;
866 }
867
normalize(uint32_t & val,uint32_t & rot)868 static void normalize(uint32_t& val, uint32_t& rot)
869 {
870 rot = 0;
871 while (!(val&3) || (val & 0xFC000000)) {
872 uint32_t newval;
873 newval = val >> 2;
874 newval |= (val&3) << 30;
875 val = newval;
876 rot += 2;
877 if (rot == 32) {
878 rot = 0;
879 break;
880 }
881 }
882 }
883
build_and_immediate(int d,int s,uint32_t mask,int bits)884 void GGLAssembler::build_and_immediate(int d, int s, uint32_t mask, int bits)
885 {
886 uint32_t rot;
887 uint32_t size = ((bits>=32) ? 0 : (1LU << bits)) - 1;
888 mask &= size;
889
890 if (mask == size) {
891 if (d != s)
892 MOV( AL, 0, d, s);
893 return;
894 }
895
896 if ((getCodegenArch() == CODEGEN_ARCH_MIPS) ||
897 (getCodegenArch() == CODEGEN_ARCH_MIPS64)) {
898 // MIPS can do 16-bit imm in 1 instr, 32-bit in 3 instr
899 // the below ' while (mask)' code is buggy on mips
900 // since mips returns true on isValidImmediate()
901 // then we get multiple AND instr (positive logic)
902 AND( AL, 0, d, s, imm(mask) );
903 return;
904 }
905 else if (getCodegenArch() == CODEGEN_ARCH_ARM64) {
906 AND( AL, 0, d, s, imm(mask) );
907 return;
908 }
909
910 int negative_logic = !isValidImmediate(mask);
911 if (negative_logic) {
912 mask = ~mask & size;
913 }
914 normalize(mask, rot);
915
916 if (mask) {
917 while (mask) {
918 uint32_t bitpos = find_bottom(mask);
919 int shift = rot + bitpos;
920 uint32_t m = mask & (0xff << bitpos);
921 mask &= ~m;
922 m >>= bitpos;
923 int32_t newMask = (m<<shift) | (m>>(32-shift));
924 if (!negative_logic) {
925 AND( AL, 0, d, s, imm(newMask) );
926 } else {
927 BIC( AL, 0, d, s, imm(newMask) );
928 }
929 s = d;
930 }
931 } else {
932 MOV( AL, 0, d, imm(0));
933 }
934 }
935
build_masking(pixel_t & pixel,Scratch & regs)936 void GGLAssembler::build_masking(pixel_t& pixel, Scratch& regs)
937 {
938 if (!mMasking || mAllMasked) {
939 return;
940 }
941
942 comment("color mask");
943
944 pixel_t fb(mDstPixel);
945 pixel_t s(pixel);
946 if (!(pixel.flags & CORRUPTIBLE)) {
947 pixel.reg = regs.obtain();
948 pixel.flags |= CORRUPTIBLE;
949 }
950
951 int mask = 0;
952 for (int i=0 ; i<4 ; i++) {
953 const int component_mask = 1<<i;
954 const int h = fb.format.c[i].h;
955 const int l = fb.format.c[i].l;
956 if (h && (!(mMasking & component_mask))) {
957 mask |= ((1<<(h-l))-1) << l;
958 }
959 }
960
961 // There is no need to clear the masked components of the source
962 // (unless we applied a logic op), because they're already zeroed
963 // by construction (masked components are not computed)
964
965 if (mLogicOp) {
966 const needs_t& needs = mBuilderContext.needs;
967 const int opcode = GGL_READ_NEEDS(LOGIC_OP, needs.n) | GGL_CLEAR;
968 if (opcode != GGL_CLEAR) {
969 // clear masked component of source
970 build_and_immediate(pixel.reg, s.reg, mask, fb.size());
971 s = pixel;
972 }
973 }
974
975 // clear non masked components of destination
976 build_and_immediate(fb.reg, fb.reg, ~mask, fb.size());
977
978 // or back the channels that were masked
979 if (s.reg == fb.reg) {
980 // this is in fact a MOV
981 if (s.reg == pixel.reg) {
982 // ugh. this in in fact a nop
983 } else {
984 MOV(AL, 0, pixel.reg, fb.reg);
985 }
986 } else {
987 ORR(AL, 0, pixel.reg, s.reg, fb.reg);
988 }
989 }
990
991 // ---------------------------------------------------------------------------
992
base_offset(const pointer_t & d,const pointer_t & b,const reg_t & o)993 void GGLAssembler::base_offset(
994 const pointer_t& d, const pointer_t& b, const reg_t& o)
995 {
996 switch (b.size) {
997 case 32:
998 ADDR_ADD(AL, 0, d.reg, b.reg, reg_imm(o.reg, LSL, 2));
999 break;
1000 case 24:
1001 if (d.reg == b.reg) {
1002 ADDR_ADD(AL, 0, d.reg, b.reg, reg_imm(o.reg, LSL, 1));
1003 ADDR_ADD(AL, 0, d.reg, d.reg, o.reg);
1004 } else {
1005 ADDR_ADD(AL, 0, d.reg, o.reg, reg_imm(o.reg, LSL, 1));
1006 ADDR_ADD(AL, 0, d.reg, d.reg, b.reg);
1007 }
1008 break;
1009 case 16:
1010 ADDR_ADD(AL, 0, d.reg, b.reg, reg_imm(o.reg, LSL, 1));
1011 break;
1012 case 8:
1013 ADDR_ADD(AL, 0, d.reg, b.reg, o.reg);
1014 break;
1015 }
1016 }
1017
1018 // ----------------------------------------------------------------------------
1019 // cheezy register allocator...
1020 // ----------------------------------------------------------------------------
1021
1022 // Modified to support MIPS processors, in a very simple way. We retain the
1023 // (Arm) limit of 16 total registers, but shift the mapping of those registers
1024 // from 0-15, to 2-17. Register 0 on Mips cannot be used as GP registers, and
1025 // register 1 has a traditional use as a temp).
1026
RegisterAllocator(int arch)1027 RegisterAllocator::RegisterAllocator(int arch) : mRegs(arch)
1028 {
1029 }
1030
reset()1031 void RegisterAllocator::reset()
1032 {
1033 mRegs.reset();
1034 }
1035
reserveReg(int reg)1036 int RegisterAllocator::reserveReg(int reg)
1037 {
1038 return mRegs.reserve(reg);
1039 }
1040
obtainReg()1041 int RegisterAllocator::obtainReg()
1042 {
1043 return mRegs.obtain();
1044 }
1045
recycleReg(int reg)1046 void RegisterAllocator::recycleReg(int reg)
1047 {
1048 mRegs.recycle(reg);
1049 }
1050
registerFile()1051 RegisterAllocator::RegisterFile& RegisterAllocator::registerFile()
1052 {
1053 return mRegs;
1054 }
1055
1056 // ----------------------------------------------------------------------------
1057
RegisterFile(int codegen_arch)1058 RegisterAllocator::RegisterFile::RegisterFile(int codegen_arch)
1059 : mRegs(0), mTouched(0), mStatus(0), mArch(codegen_arch), mRegisterOffset(0)
1060 {
1061 if ((mArch == ARMAssemblerInterface::CODEGEN_ARCH_MIPS) ||
1062 (mArch == ARMAssemblerInterface::CODEGEN_ARCH_MIPS64)) {
1063 mRegisterOffset = 2; // ARM has regs 0..15, MIPS offset to 2..17
1064 }
1065 reserve(ARMAssemblerInterface::SP);
1066 reserve(ARMAssemblerInterface::PC);
1067 }
1068
RegisterFile(const RegisterFile & rhs,int codegen_arch)1069 RegisterAllocator::RegisterFile::RegisterFile(const RegisterFile& rhs, int codegen_arch)
1070 : mRegs(rhs.mRegs), mTouched(rhs.mTouched), mArch(codegen_arch), mRegisterOffset(0)
1071 {
1072 if ((mArch == ARMAssemblerInterface::CODEGEN_ARCH_MIPS) ||
1073 (mArch == ARMAssemblerInterface::CODEGEN_ARCH_MIPS64)) {
1074 mRegisterOffset = 2; // ARM has regs 0..15, MIPS offset to 2..17
1075 }
1076 }
1077
~RegisterFile()1078 RegisterAllocator::RegisterFile::~RegisterFile()
1079 {
1080 }
1081
operator ==(const RegisterFile & rhs) const1082 bool RegisterAllocator::RegisterFile::operator == (const RegisterFile& rhs) const
1083 {
1084 return (mRegs == rhs.mRegs);
1085 }
1086
reset()1087 void RegisterAllocator::RegisterFile::reset()
1088 {
1089 mRegs = mTouched = mStatus = 0;
1090 reserve(ARMAssemblerInterface::SP);
1091 reserve(ARMAssemblerInterface::PC);
1092 }
1093
1094 // RegisterFile::reserve() take a register parameter in the
1095 // range 0-15 (Arm compatible), but on a Mips processor, will
1096 // return the actual allocated register in the range 2-17.
reserve(int reg)1097 int RegisterAllocator::RegisterFile::reserve(int reg)
1098 {
1099 reg += mRegisterOffset;
1100 LOG_ALWAYS_FATAL_IF(isUsed(reg),
1101 "reserving register %d, but already in use",
1102 reg);
1103 mRegs |= (1<<reg);
1104 mTouched |= mRegs;
1105 return reg;
1106 }
1107
1108 // This interface uses regMask in range 2-17 on MIPS, no translation.
reserveSeveral(uint32_t regMask)1109 void RegisterAllocator::RegisterFile::reserveSeveral(uint32_t regMask)
1110 {
1111 mRegs |= regMask;
1112 mTouched |= regMask;
1113 }
1114
isUsed(int reg) const1115 int RegisterAllocator::RegisterFile::isUsed(int reg) const
1116 {
1117 LOG_ALWAYS_FATAL_IF(reg>=16+(int)mRegisterOffset, "invalid register %d", reg);
1118 return mRegs & (1<<reg);
1119 }
1120
obtain()1121 int RegisterAllocator::RegisterFile::obtain()
1122 {
1123 const char priorityList[14] = { 0, 1, 2, 3,
1124 12, 14, 4, 5,
1125 6, 7, 8, 9,
1126 10, 11 };
1127 const int nbreg = sizeof(priorityList);
1128 int i, r, reg;
1129 for (i=0 ; i<nbreg ; i++) {
1130 r = priorityList[i];
1131 if (!isUsed(r + mRegisterOffset)) {
1132 break;
1133 }
1134 }
1135 // this is not an error anymore because, we'll try again with
1136 // a lower optimization level.
1137 //ALOGE_IF(i >= nbreg, "pixelflinger ran out of registers\n");
1138 if (i >= nbreg) {
1139 mStatus |= OUT_OF_REGISTERS;
1140 // we return SP so we can more easily debug things
1141 // the code will never be run anyway.
1142 return ARMAssemblerInterface::SP;
1143 }
1144 reg = reserve(r); // Param in Arm range 0-15, returns range 2-17 on Mips.
1145 return reg;
1146 }
1147
hasFreeRegs() const1148 bool RegisterAllocator::RegisterFile::hasFreeRegs() const
1149 {
1150 uint32_t regs = mRegs >> mRegisterOffset; // MIPS fix.
1151 return ((regs & 0xFFFF) == 0xFFFF) ? false : true;
1152 }
1153
countFreeRegs() const1154 int RegisterAllocator::RegisterFile::countFreeRegs() const
1155 {
1156 uint32_t regs = mRegs >> mRegisterOffset; // MIPS fix.
1157 int f = ~regs & 0xFFFF;
1158 // now count number of 1
1159 f = (f & 0x5555) + ((f>>1) & 0x5555);
1160 f = (f & 0x3333) + ((f>>2) & 0x3333);
1161 f = (f & 0x0F0F) + ((f>>4) & 0x0F0F);
1162 f = (f & 0x00FF) + ((f>>8) & 0x00FF);
1163 return f;
1164 }
1165
recycle(int reg)1166 void RegisterAllocator::RegisterFile::recycle(int reg)
1167 {
1168 // commented out, since common failure of running out of regs
1169 // triggers this assertion. Since the code is not execectued
1170 // in that case, it does not matter. No reason to FATAL err.
1171 // LOG_FATAL_IF(!isUsed(reg),
1172 // "recycling unallocated register %d",
1173 // reg);
1174 mRegs &= ~(1<<reg);
1175 }
1176
recycleSeveral(uint32_t regMask)1177 void RegisterAllocator::RegisterFile::recycleSeveral(uint32_t regMask)
1178 {
1179 // commented out, since common failure of running out of regs
1180 // triggers this assertion. Since the code is not execectued
1181 // in that case, it does not matter. No reason to FATAL err.
1182 // LOG_FATAL_IF((mRegs & regMask)!=regMask,
1183 // "recycling unallocated registers "
1184 // "(recycle=%08x, allocated=%08x, unallocated=%08x)",
1185 // regMask, mRegs, mRegs®Mask);
1186 mRegs &= ~regMask;
1187 }
1188
touched() const1189 uint32_t RegisterAllocator::RegisterFile::touched() const
1190 {
1191 return mTouched;
1192 }
1193
1194 // ----------------------------------------------------------------------------
1195
1196 }; // namespace android
1197
1198