Home | History | Annotate | Download | only in codeflinger
      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 #ifdef __ARM_ARCH__
     29 #include <machine/cpu-features.h>
     30 #endif
     31 
     32 namespace android {
     33 
     34 // ---------------------------------------------------------------------------
     35 
     36 // iterators are initialized like this:
     37 // (intToFixedCenter(x) * dx)>>16 + x0
     38 // ((x<<16 + 0x8000) * dx)>>16 + x0
     39 // ((x<<16)*dx + (0x8000*dx))>>16 + x0
     40 // ( (x*dx) + dx>>1 ) + x0
     41 // (x*dx) + (dx>>1 + x0)
     42 
     43 void GGLAssembler::init_iterated_color(fragment_parts_t& parts, const reg_t& x)
     44 {
     45     context_t const* c = mBuilderContext.c;
     46     const needs_t& needs = mBuilderContext.needs;
     47 
     48     if (mSmooth) {
     49         // NOTE: we could take this case in the mDithering + !mSmooth case,
     50         // but this would use up to 4 more registers for the color components
     51         // for only a little added quality.
     52         // Currently, this causes the system to run out of registers in
     53         // some case (see issue #719496)
     54 
     55         comment("compute initial iterated color (smooth and/or dither case)");
     56 
     57         parts.iterated_packed = 0;
     58         parts.packed = 0;
     59 
     60         // 0x1: color component
     61         // 0x2: iterators
     62         const int optReload = mOptLevel >> 1;
     63         if (optReload >= 3)         parts.reload = 0; // reload nothing
     64         else if (optReload == 2)    parts.reload = 2; // reload iterators
     65         else if (optReload == 1)    parts.reload = 1; // reload colors
     66         else if (optReload <= 0)    parts.reload = 3; // reload both
     67 
     68         if (!mSmooth) {
     69             // we're not smoothing (just dithering), we never have to
     70             // reload the iterators
     71             parts.reload &= ~2;
     72         }
     73 
     74         Scratch scratches(registerFile());
     75         const int t0 = (parts.reload & 1) ? scratches.obtain() : 0;
     76         const int t1 = (parts.reload & 2) ? scratches.obtain() : 0;
     77         for (int i=0 ; i<4 ; i++) {
     78             if (!mInfo[i].iterated)
     79                 continue;
     80 
     81             // this component exists in the destination and is not replaced
     82             // by a texture unit.
     83             const int c = (parts.reload & 1) ? t0 : obtainReg();
     84             if (i==0) CONTEXT_LOAD(c, iterators.ydady);
     85             if (i==1) CONTEXT_LOAD(c, iterators.ydrdy);
     86             if (i==2) CONTEXT_LOAD(c, iterators.ydgdy);
     87             if (i==3) CONTEXT_LOAD(c, iterators.ydbdy);
     88             parts.argb[i].reg = c;
     89 
     90             if (mInfo[i].smooth) {
     91                 parts.argb_dx[i].reg = (parts.reload & 2) ? t1 : obtainReg();
     92                 const int dvdx = parts.argb_dx[i].reg;
     93                 CONTEXT_LOAD(dvdx, generated_vars.argb[i].dx);
     94                 MLA(AL, 0, c, x.reg, dvdx, c);
     95 
     96                 // adjust the color iterator to make sure it won't overflow
     97                 if (!mAA) {
     98                     // this is not needed when we're using anti-aliasing
     99                     // because we will (have to) clamp the components
    100                     // anyway.
    101                     int end = scratches.obtain();
    102                     MOV(AL, 0, end, reg_imm(parts.count.reg, LSR, 16));
    103                     MLA(AL, 1, end, dvdx, end, c);
    104                     SUB(MI, 0, c, c, end);
    105                     BIC(AL, 0, c, c, reg_imm(c, ASR, 31));
    106                     scratches.recycle(end);
    107                 }
    108             }
    109 
    110             if (parts.reload & 1) {
    111                 CONTEXT_STORE(c, generated_vars.argb[i].c);
    112             }
    113         }
    114     } else {
    115         // We're not smoothed, so we can
    116         // just use a packed version of the color and extract the
    117         // components as needed (or not at all if we don't blend)
    118 
    119         // figure out if we need the iterated color
    120         int load = 0;
    121         for (int i=0 ; i<4 ; i++) {
    122             component_info_t& info = mInfo[i];
    123             if ((info.inDest || info.needed) && !info.replaced)
    124                 load |= 1;
    125         }
    126 
    127         parts.iterated_packed = 1;
    128         parts.packed = (!mTextureMachine.mask && !mBlending
    129                 && !mFog && !mDithering);
    130         parts.reload = 0;
    131         if (load || parts.packed) {
    132             if (mBlending || mDithering || mInfo[GGLFormat::ALPHA].needed) {
    133                 comment("load initial iterated color (8888 packed)");
    134                 parts.iterated.setTo(obtainReg(),
    135                         &(c->formats[GGL_PIXEL_FORMAT_RGBA_8888]));
    136                 CONTEXT_LOAD(parts.iterated.reg, packed8888);
    137             } else {
    138                 comment("load initial iterated color (dest format packed)");
    139 
    140                 parts.iterated.setTo(obtainReg(), &mCbFormat);
    141 
    142                 // pre-mask the iterated color
    143                 const int bits = parts.iterated.size();
    144                 const uint32_t size = ((bits>=32) ? 0 : (1LU << bits)) - 1;
    145                 uint32_t mask = 0;
    146                 if (mMasking) {
    147                     for (int i=0 ; i<4 ; i++) {
    148                         const int component_mask = 1<<i;
    149                         const int h = parts.iterated.format.c[i].h;
    150                         const int l = parts.iterated.format.c[i].l;
    151                         if (h && (!(mMasking & component_mask))) {
    152                             mask |= ((1<<(h-l))-1) << l;
    153                         }
    154                     }
    155                 }
    156 
    157                 if (mMasking && ((mask & size)==0)) {
    158                     // none of the components are present in the mask
    159                 } else {
    160                     CONTEXT_LOAD(parts.iterated.reg, packed);
    161                     if (mCbFormat.size == 1) {
    162                         AND(AL, 0, parts.iterated.reg,
    163                                 parts.iterated.reg, imm(0xFF));
    164                     } else if (mCbFormat.size == 2) {
    165                         MOV(AL, 0, parts.iterated.reg,
    166                                 reg_imm(parts.iterated.reg, LSR, 16));
    167                     }
    168                 }
    169 
    170                 // pre-mask the iterated color
    171                 if (mMasking) {
    172                     build_and_immediate(parts.iterated.reg, parts.iterated.reg,
    173                             mask, bits);
    174                 }
    175             }
    176         }
    177     }
    178 }
    179 
    180 void GGLAssembler::build_iterated_color(
    181         component_t& fragment,
    182         const fragment_parts_t& parts,
    183         int component,
    184         Scratch& regs)
    185 {
    186     fragment.setTo( regs.obtain(), 0, 32, CORRUPTIBLE);
    187 
    188     if (!mInfo[component].iterated)
    189         return;
    190 
    191     if (parts.iterated_packed) {
    192         // iterated colors are packed, extract the one we need
    193         extract(fragment, parts.iterated, component);
    194     } else {
    195         fragment.h = GGL_COLOR_BITS;
    196         fragment.l = GGL_COLOR_BITS - 8;
    197         fragment.flags |= CLEAR_LO;
    198         // iterated colors are held in their own register,
    199         // (smooth and/or dithering case)
    200         if (parts.reload==3) {
    201             // this implies mSmooth
    202             Scratch scratches(registerFile());
    203             int dx = scratches.obtain();
    204             CONTEXT_LOAD(fragment.reg, generated_vars.argb[component].c);
    205             CONTEXT_LOAD(dx, generated_vars.argb[component].dx);
    206             ADD(AL, 0, dx, fragment.reg, dx);
    207             CONTEXT_STORE(dx, generated_vars.argb[component].c);
    208         } else if (parts.reload & 1) {
    209             CONTEXT_LOAD(fragment.reg, generated_vars.argb[component].c);
    210         } else {
    211             // we don't reload, so simply rename the register and mark as
    212             // non CORRUPTIBLE so that the texture env or blending code
    213             // won't modify this (renamed) register
    214             regs.recycle(fragment.reg);
    215             fragment.reg = parts.argb[component].reg;
    216             fragment.flags &= ~CORRUPTIBLE;
    217         }
    218         if (mInfo[component].smooth && mAA) {
    219             // when using smooth shading AND anti-aliasing, we need to clamp
    220             // the iterators because there is always an extra pixel on the
    221             // edges, which most of the time will cause an overflow
    222             // (since technically its outside of the domain).
    223             BIC(AL, 0, fragment.reg, fragment.reg,
    224                     reg_imm(fragment.reg, ASR, 31));
    225             component_sat(fragment);
    226         }
    227     }
    228 }
    229 
    230 // ---------------------------------------------------------------------------
    231 
    232 void GGLAssembler::decodeLogicOpNeeds(const needs_t& needs)
    233 {
    234     // gather some informations about the components we need to process...
    235     const int opcode = GGL_READ_NEEDS(LOGIC_OP, needs.n) | GGL_CLEAR;
    236     switch(opcode) {
    237     case GGL_COPY:
    238         mLogicOp = 0;
    239         break;
    240     case GGL_CLEAR:
    241     case GGL_SET:
    242         mLogicOp = LOGIC_OP;
    243         break;
    244     case GGL_AND:
    245     case GGL_AND_REVERSE:
    246     case GGL_AND_INVERTED:
    247     case GGL_XOR:
    248     case GGL_OR:
    249     case GGL_NOR:
    250     case GGL_EQUIV:
    251     case GGL_OR_REVERSE:
    252     case GGL_OR_INVERTED:
    253     case GGL_NAND:
    254         mLogicOp = LOGIC_OP|LOGIC_OP_SRC|LOGIC_OP_DST;
    255         break;
    256     case GGL_NOOP:
    257     case GGL_INVERT:
    258         mLogicOp = LOGIC_OP|LOGIC_OP_DST;
    259         break;
    260     case GGL_COPY_INVERTED:
    261         mLogicOp = LOGIC_OP|LOGIC_OP_SRC;
    262         break;
    263     };
    264 }
    265 
    266 void GGLAssembler::decodeTMUNeeds(const needs_t& needs, context_t const* c)
    267 {
    268     uint8_t replaced=0;
    269     mTextureMachine.mask = 0;
    270     mTextureMachine.activeUnits = 0;
    271     for (int i=GGL_TEXTURE_UNIT_COUNT-1 ; i>=0 ; i--) {
    272         texture_unit_t& tmu = mTextureMachine.tmu[i];
    273         if (replaced == 0xF) {
    274             // all components are replaced, skip this TMU.
    275             tmu.format_idx = 0;
    276             tmu.mask = 0;
    277             tmu.replaced = replaced;
    278             continue;
    279         }
    280         tmu.format_idx = GGL_READ_NEEDS(T_FORMAT, needs.t[i]);
    281         tmu.format = c->formats[tmu.format_idx];
    282         tmu.bits = tmu.format.size*8;
    283         tmu.swrap = GGL_READ_NEEDS(T_S_WRAP, needs.t[i]);
    284         tmu.twrap = GGL_READ_NEEDS(T_T_WRAP, needs.t[i]);
    285         tmu.env = ggl_needs_to_env(GGL_READ_NEEDS(T_ENV, needs.t[i]));
    286         tmu.pot = GGL_READ_NEEDS(T_POT, needs.t[i]);
    287         tmu.linear = GGL_READ_NEEDS(T_LINEAR, needs.t[i])
    288                 && tmu.format.size!=3; // XXX: only 8, 16 and 32 modes for now
    289 
    290         // 5551 linear filtering is not supported
    291         if (tmu.format_idx == GGL_PIXEL_FORMAT_RGBA_5551)
    292             tmu.linear = 0;
    293 
    294         tmu.mask = 0;
    295         tmu.replaced = replaced;
    296 
    297         if (tmu.format_idx) {
    298             mTextureMachine.activeUnits++;
    299             if (tmu.format.c[0].h)    tmu.mask |= 0x1;
    300             if (tmu.format.c[1].h)    tmu.mask |= 0x2;
    301             if (tmu.format.c[2].h)    tmu.mask |= 0x4;
    302             if (tmu.format.c[3].h)    tmu.mask |= 0x8;
    303             if (tmu.env == GGL_REPLACE) {
    304                 replaced |= tmu.mask;
    305             } else if (tmu.env == GGL_DECAL) {
    306                 if (!tmu.format.c[GGLFormat::ALPHA].h) {
    307                     // if we don't have alpha, decal does nothing
    308                     tmu.mask = 0;
    309                 } else {
    310                     // decal always ignores At
    311                     tmu.mask &= ~(1<<GGLFormat::ALPHA);
    312                 }
    313             }
    314         }
    315         mTextureMachine.mask |= tmu.mask;
    316         //printf("%d: mask=%08lx, replaced=%08lx\n",
    317         //    i, int(tmu.mask), int(tmu.replaced));
    318     }
    319     mTextureMachine.replaced = replaced;
    320     mTextureMachine.directTexture = 0;
    321     //printf("replaced=%08lx\n", mTextureMachine.replaced);
    322 }
    323 
    324 
    325 void GGLAssembler::init_textures(
    326         tex_coord_t* coords,
    327         const reg_t& x, const reg_t& y)
    328 {
    329     context_t const* c = mBuilderContext.c;
    330     const needs_t& needs = mBuilderContext.needs;
    331     int Rctx = mBuilderContext.Rctx;
    332     int Rx = x.reg;
    333     int Ry = y.reg;
    334 
    335     if (mTextureMachine.mask) {
    336         comment("compute texture coordinates");
    337     }
    338 
    339     // init texture coordinates for each tmu
    340     const int cb_format_idx = GGL_READ_NEEDS(CB_FORMAT, needs.n);
    341     const bool multiTexture = mTextureMachine.activeUnits > 1;
    342     for (int i=0 ; i<GGL_TEXTURE_UNIT_COUNT; i++) {
    343         const texture_unit_t& tmu = mTextureMachine.tmu[i];
    344         if (tmu.format_idx == 0)
    345             continue;
    346         if ((tmu.swrap == GGL_NEEDS_WRAP_11) &&
    347             (tmu.twrap == GGL_NEEDS_WRAP_11))
    348         {
    349             // 1:1 texture
    350             pointer_t& txPtr = coords[i].ptr;
    351             txPtr.setTo(obtainReg(), tmu.bits);
    352             CONTEXT_LOAD(txPtr.reg, state.texture[i].iterators.ydsdy);
    353             ADD(AL, 0, Rx, Rx, reg_imm(txPtr.reg, ASR, 16));    // x += (s>>16)
    354             CONTEXT_LOAD(txPtr.reg, state.texture[i].iterators.ydtdy);
    355             ADD(AL, 0, Ry, Ry, reg_imm(txPtr.reg, ASR, 16));    // y += (t>>16)
    356             // merge base & offset
    357             CONTEXT_LOAD(txPtr.reg, generated_vars.texture[i].stride);
    358             SMLABB(AL, Rx, Ry, txPtr.reg, Rx);               // x+y*stride
    359             CONTEXT_LOAD(txPtr.reg, generated_vars.texture[i].data);
    360             base_offset(txPtr, txPtr, Rx);
    361         } else {
    362             Scratch scratches(registerFile());
    363             reg_t& s = coords[i].s;
    364             reg_t& t = coords[i].t;
    365             // s = (x * dsdx)>>16 + ydsdy
    366             // s = (x * dsdx)>>16 + (y*dsdy)>>16 + s0
    367             // t = (x * dtdx)>>16 + ydtdy
    368             // t = (x * dtdx)>>16 + (y*dtdy)>>16 + t0
    369             s.setTo(obtainReg());
    370             t.setTo(obtainReg());
    371             const int need_w = GGL_READ_NEEDS(W, needs.n);
    372             if (need_w) {
    373                 CONTEXT_LOAD(s.reg, state.texture[i].iterators.ydsdy);
    374                 CONTEXT_LOAD(t.reg, state.texture[i].iterators.ydtdy);
    375             } else {
    376                 int ydsdy = scratches.obtain();
    377                 int ydtdy = scratches.obtain();
    378                 CONTEXT_LOAD(s.reg, generated_vars.texture[i].dsdx);
    379                 CONTEXT_LOAD(ydsdy, state.texture[i].iterators.ydsdy);
    380                 CONTEXT_LOAD(t.reg, generated_vars.texture[i].dtdx);
    381                 CONTEXT_LOAD(ydtdy, state.texture[i].iterators.ydtdy);
    382                 MLA(AL, 0, s.reg, Rx, s.reg, ydsdy);
    383                 MLA(AL, 0, t.reg, Rx, t.reg, ydtdy);
    384             }
    385 
    386             if ((mOptLevel&1)==0) {
    387                 CONTEXT_STORE(s.reg, generated_vars.texture[i].spill[0]);
    388                 CONTEXT_STORE(t.reg, generated_vars.texture[i].spill[1]);
    389                 recycleReg(s.reg);
    390                 recycleReg(t.reg);
    391             }
    392         }
    393 
    394         // direct texture?
    395         if (!multiTexture && !mBlending && !mDithering && !mFog &&
    396             cb_format_idx == tmu.format_idx && !tmu.linear &&
    397             mTextureMachine.replaced == tmu.mask)
    398         {
    399                 mTextureMachine.directTexture = i + 1;
    400         }
    401     }
    402 }
    403 
    404 void GGLAssembler::build_textures(  fragment_parts_t& parts,
    405                                     Scratch& regs)
    406 {
    407     context_t const* c = mBuilderContext.c;
    408     const needs_t& needs = mBuilderContext.needs;
    409     int Rctx = mBuilderContext.Rctx;
    410 
    411     // We don't have a way to spill registers automatically
    412     // spill depth and AA regs, when we know we may have to.
    413     // build the spill list...
    414     uint32_t spill_list = 0;
    415     for (int i=0 ; i<GGL_TEXTURE_UNIT_COUNT; i++) {
    416         const texture_unit_t& tmu = mTextureMachine.tmu[i];
    417         if (tmu.format_idx == 0)
    418             continue;
    419         if (tmu.linear) {
    420             // we may run out of register if we have linear filtering
    421             // at 1 or 4 bytes / pixel on any texture unit.
    422             if (tmu.format.size == 1) {
    423                 // if depth and AA enabled, we'll run out of 1 register
    424                 if (parts.z.reg > 0 && parts.covPtr.reg > 0)
    425                     spill_list |= 1<<parts.covPtr.reg;
    426             }
    427             if (tmu.format.size == 4) {
    428                 // if depth or AA enabled, we'll run out of 1 or 2 registers
    429                 if (parts.z.reg > 0)
    430                     spill_list |= 1<<parts.z.reg;
    431                 if (parts.covPtr.reg > 0)
    432                     spill_list |= 1<<parts.covPtr.reg;
    433             }
    434         }
    435     }
    436 
    437     Spill spill(registerFile(), *this, spill_list);
    438 
    439     const bool multiTexture = mTextureMachine.activeUnits > 1;
    440     for (int i=0 ; i<GGL_TEXTURE_UNIT_COUNT; i++) {
    441         const texture_unit_t& tmu = mTextureMachine.tmu[i];
    442         if (tmu.format_idx == 0)
    443             continue;
    444 
    445         pointer_t& txPtr = parts.coords[i].ptr;
    446         pixel_t& texel = parts.texel[i];
    447 
    448         // repeat...
    449         if ((tmu.swrap == GGL_NEEDS_WRAP_11) &&
    450             (tmu.twrap == GGL_NEEDS_WRAP_11))
    451         { // 1:1 textures
    452             comment("fetch texel");
    453             texel.setTo(regs.obtain(), &tmu.format);
    454             load(txPtr, texel, WRITE_BACK);
    455         } else {
    456             Scratch scratches(registerFile());
    457             reg_t& s = parts.coords[i].s;
    458             reg_t& t = parts.coords[i].t;
    459             if ((mOptLevel&1)==0) {
    460                 comment("reload s/t (multitexture or linear filtering)");
    461                 s.reg = scratches.obtain();
    462                 t.reg = scratches.obtain();
    463                 CONTEXT_LOAD(s.reg, generated_vars.texture[i].spill[0]);
    464                 CONTEXT_LOAD(t.reg, generated_vars.texture[i].spill[1]);
    465             }
    466 
    467             comment("compute repeat/clamp");
    468             int u       = scratches.obtain();
    469             int v       = scratches.obtain();
    470             int width   = scratches.obtain();
    471             int height  = scratches.obtain();
    472             int U = 0;
    473             int V = 0;
    474 
    475             CONTEXT_LOAD(width,  generated_vars.texture[i].width);
    476             CONTEXT_LOAD(height, generated_vars.texture[i].height);
    477 
    478             int FRAC_BITS = 0;
    479             if (tmu.linear) {
    480                 // linear interpolation
    481                 if (tmu.format.size == 1) {
    482                     // for 8-bits textures, we can afford
    483                     // 7 bits of fractional precision at no
    484                     // additional cost (we can't do 8 bits
    485                     // because filter8 uses signed 16 bits muls)
    486                     FRAC_BITS = 7;
    487                 } else if (tmu.format.size == 2) {
    488                     // filter16() is internally limited to 4 bits, so:
    489                     // FRAC_BITS=2 generates less instructions,
    490                     // FRAC_BITS=3,4,5 creates unpleasant artifacts,
    491                     // FRAC_BITS=6+ looks good
    492                     FRAC_BITS = 6;
    493                 } else if (tmu.format.size == 4) {
    494                     // filter32() is internally limited to 8 bits, so:
    495                     // FRAC_BITS=4 looks good
    496                     // FRAC_BITS=5+ looks better, but generates 3 extra ipp
    497                     FRAC_BITS = 6;
    498                 } else {
    499                     // for all other cases we use 4 bits.
    500                     FRAC_BITS = 4;
    501                 }
    502             }
    503             wrapping(u, s.reg, width,  tmu.swrap, FRAC_BITS);
    504             wrapping(v, t.reg, height, tmu.twrap, FRAC_BITS);
    505 
    506             if (tmu.linear) {
    507                 comment("compute linear filtering offsets");
    508                 // pixel size scale
    509                 const int shift = 31 - gglClz(tmu.format.size);
    510                 U = scratches.obtain();
    511                 V = scratches.obtain();
    512 
    513                 // sample the texel center
    514                 SUB(AL, 0, u, u, imm(1<<(FRAC_BITS-1)));
    515                 SUB(AL, 0, v, v, imm(1<<(FRAC_BITS-1)));
    516 
    517                 // get the fractionnal part of U,V
    518                 AND(AL, 0, U, u, imm((1<<FRAC_BITS)-1));
    519                 AND(AL, 0, V, v, imm((1<<FRAC_BITS)-1));
    520 
    521                 // compute width-1 and height-1
    522                 SUB(AL, 0, width,  width,  imm(1));
    523                 SUB(AL, 0, height, height, imm(1));
    524 
    525                 // get the integer part of U,V and clamp/wrap
    526                 // and compute offset to the next texel
    527                 if (tmu.swrap == GGL_NEEDS_WRAP_REPEAT) {
    528                     // u has already been REPEATed
    529                     MOV(AL, 1, u, reg_imm(u, ASR, FRAC_BITS));
    530                     MOV(MI, 0, u, width);
    531                     CMP(AL, u, width);
    532                     MOV(LT, 0, width, imm(1 << shift));
    533                     if (shift)
    534                         MOV(GE, 0, width, reg_imm(width, LSL, shift));
    535                     RSB(GE, 0, width, width, imm(0));
    536                 } else {
    537                     // u has not been CLAMPed yet
    538                     // algorithm:
    539                     // if ((u>>4) >= width)
    540                     //      u = width<<4
    541                     //      width = 0
    542                     // else
    543                     //      width = 1<<shift
    544                     // u = u>>4; // get integer part
    545                     // if (u<0)
    546                     //      u = 0
    547                     //      width = 0
    548                     // generated_vars.rt = width
    549 
    550                     CMP(AL, width, reg_imm(u, ASR, FRAC_BITS));
    551                     MOV(LE, 0, u, reg_imm(width, LSL, FRAC_BITS));
    552                     MOV(LE, 0, width, imm(0));
    553                     MOV(GT, 0, width, imm(1 << shift));
    554                     MOV(AL, 1, u, reg_imm(u, ASR, FRAC_BITS));
    555                     MOV(MI, 0, u, imm(0));
    556                     MOV(MI, 0, width, imm(0));
    557                 }
    558                 CONTEXT_STORE(width, generated_vars.rt);
    559 
    560                 const int stride = width;
    561                 CONTEXT_LOAD(stride, generated_vars.texture[i].stride);
    562                 if (tmu.twrap == GGL_NEEDS_WRAP_REPEAT) {
    563                     // v has already been REPEATed
    564                     MOV(AL, 1, v, reg_imm(v, ASR, FRAC_BITS));
    565                     MOV(MI, 0, v, height);
    566                     CMP(AL, v, height);
    567                     MOV(LT, 0, height, imm(1 << shift));
    568                     if (shift)
    569                         MOV(GE, 0, height, reg_imm(height, LSL, shift));
    570                     RSB(GE, 0, height, height, imm(0));
    571                     MUL(AL, 0, height, stride, height);
    572                 } else {
    573                     // v has not been CLAMPed yet
    574                     CMP(AL, height, reg_imm(v, ASR, FRAC_BITS));
    575                     MOV(LE, 0, v, reg_imm(height, LSL, FRAC_BITS));
    576                     MOV(LE, 0, height, imm(0));
    577                     if (shift) {
    578                         MOV(GT, 0, height, reg_imm(stride, LSL, shift));
    579                     } else {
    580                         MOV(GT, 0, height, stride);
    581                     }
    582                     MOV(AL, 1, v, reg_imm(v, ASR, FRAC_BITS));
    583                     MOV(MI, 0, v, imm(0));
    584                     MOV(MI, 0, height, imm(0));
    585                 }
    586                 CONTEXT_STORE(height, generated_vars.lb);
    587             }
    588 
    589             scratches.recycle(width);
    590             scratches.recycle(height);
    591 
    592             // iterate texture coordinates...
    593             comment("iterate s,t");
    594             int dsdx = scratches.obtain();
    595             int dtdx = scratches.obtain();
    596             CONTEXT_LOAD(dsdx, generated_vars.texture[i].dsdx);
    597             CONTEXT_LOAD(dtdx, generated_vars.texture[i].dtdx);
    598             ADD(AL, 0, s.reg, s.reg, dsdx);
    599             ADD(AL, 0, t.reg, t.reg, dtdx);
    600             if ((mOptLevel&1)==0) {
    601                 CONTEXT_STORE(s.reg, generated_vars.texture[i].spill[0]);
    602                 CONTEXT_STORE(t.reg, generated_vars.texture[i].spill[1]);
    603                 scratches.recycle(s.reg);
    604                 scratches.recycle(t.reg);
    605             }
    606             scratches.recycle(dsdx);
    607             scratches.recycle(dtdx);
    608 
    609             // merge base & offset...
    610             comment("merge base & offset");
    611             texel.setTo(regs.obtain(), &tmu.format);
    612             txPtr.setTo(texel.reg, tmu.bits);
    613             int stride = scratches.obtain();
    614             CONTEXT_LOAD(stride,    generated_vars.texture[i].stride);
    615             CONTEXT_LOAD(txPtr.reg, generated_vars.texture[i].data);
    616             SMLABB(AL, u, v, stride, u);    // u+v*stride
    617             base_offset(txPtr, txPtr, u);
    618 
    619             // load texel
    620             if (!tmu.linear) {
    621                 comment("fetch texel");
    622                 load(txPtr, texel, 0);
    623             } else {
    624                 // recycle registers we don't need anymore
    625                 scratches.recycle(u);
    626                 scratches.recycle(v);
    627                 scratches.recycle(stride);
    628 
    629                 comment("fetch texel, bilinear");
    630                 switch (tmu.format.size) {
    631                 case 1:  filter8(parts, texel, tmu, U, V, txPtr, FRAC_BITS); break;
    632                 case 2: filter16(parts, texel, tmu, U, V, txPtr, FRAC_BITS); break;
    633                 case 3: filter24(parts, texel, tmu, U, V, txPtr, FRAC_BITS); break;
    634                 case 4: filter32(parts, texel, tmu, U, V, txPtr, FRAC_BITS); break;
    635                 }
    636             }
    637         }
    638     }
    639 }
    640 
    641 void GGLAssembler::build_iterate_texture_coordinates(
    642     const fragment_parts_t& parts)
    643 {
    644     const bool multiTexture = mTextureMachine.activeUnits > 1;
    645     for (int i=0 ; i<GGL_TEXTURE_UNIT_COUNT; i++) {
    646         const texture_unit_t& tmu = mTextureMachine.tmu[i];
    647         if (tmu.format_idx == 0)
    648             continue;
    649 
    650         if ((tmu.swrap == GGL_NEEDS_WRAP_11) &&
    651             (tmu.twrap == GGL_NEEDS_WRAP_11))
    652         { // 1:1 textures
    653             const pointer_t& txPtr = parts.coords[i].ptr;
    654             ADD(AL, 0, txPtr.reg, txPtr.reg, imm(txPtr.size>>3));
    655         } else {
    656             Scratch scratches(registerFile());
    657             int s = parts.coords[i].s.reg;
    658             int t = parts.coords[i].t.reg;
    659             if ((mOptLevel&1)==0) {
    660                 s = scratches.obtain();
    661                 t = scratches.obtain();
    662                 CONTEXT_LOAD(s, generated_vars.texture[i].spill[0]);
    663                 CONTEXT_LOAD(t, generated_vars.texture[i].spill[1]);
    664             }
    665             int dsdx = scratches.obtain();
    666             int dtdx = scratches.obtain();
    667             CONTEXT_LOAD(dsdx, generated_vars.texture[i].dsdx);
    668             CONTEXT_LOAD(dtdx, generated_vars.texture[i].dtdx);
    669             ADD(AL, 0, s, s, dsdx);
    670             ADD(AL, 0, t, t, dtdx);
    671             if ((mOptLevel&1)==0) {
    672                 CONTEXT_STORE(s, generated_vars.texture[i].spill[0]);
    673                 CONTEXT_STORE(t, generated_vars.texture[i].spill[1]);
    674             }
    675         }
    676     }
    677 }
    678 
    679 void GGLAssembler::filter8(
    680         const fragment_parts_t& parts,
    681         pixel_t& texel, const texture_unit_t& tmu,
    682         int U, int V, pointer_t& txPtr,
    683         int FRAC_BITS)
    684 {
    685     if (tmu.format.components != GGL_ALPHA &&
    686         tmu.format.components != GGL_LUMINANCE)
    687     {
    688         // this is a packed format, and we don't support
    689         // linear filtering (it's probably RGB 332)
    690         // Should not happen with OpenGL|ES
    691         LDRB(AL, texel.reg, txPtr.reg);
    692         return;
    693     }
    694 
    695     // ------------------------
    696     // about ~22 cycles / pixel
    697     Scratch scratches(registerFile());
    698 
    699     int pixel= scratches.obtain();
    700     int d    = scratches.obtain();
    701     int u    = scratches.obtain();
    702     int k    = scratches.obtain();
    703     int rt   = scratches.obtain();
    704     int lb   = scratches.obtain();
    705 
    706     // RB -> U * V
    707 
    708     CONTEXT_LOAD(rt, generated_vars.rt);
    709     CONTEXT_LOAD(lb, generated_vars.lb);
    710 
    711     int offset = pixel;
    712     ADD(AL, 0, offset, lb, rt);
    713     LDRB(AL, pixel, txPtr.reg, reg_scale_pre(offset));
    714     SMULBB(AL, u, U, V);
    715     SMULBB(AL, d, pixel, u);
    716     RSB(AL, 0, k, u, imm(1<<(FRAC_BITS*2)));
    717 
    718     // LB -> (1-U) * V
    719     RSB(AL, 0, U, U, imm(1<<FRAC_BITS));
    720     LDRB(AL, pixel, txPtr.reg, reg_scale_pre(lb));
    721     SMULBB(AL, u, U, V);
    722     SMLABB(AL, d, pixel, u, d);
    723     SUB(AL, 0, k, k, u);
    724 
    725     // LT -> (1-U)*(1-V)
    726     RSB(AL, 0, V, V, imm(1<<FRAC_BITS));
    727     LDRB(AL, pixel, txPtr.reg);
    728     SMULBB(AL, u, U, V);
    729     SMLABB(AL, d, pixel, u, d);
    730 
    731     // RT -> U*(1-V)
    732     LDRB(AL, pixel, txPtr.reg, reg_scale_pre(rt));
    733     SUB(AL, 0, u, k, u);
    734     SMLABB(AL, texel.reg, pixel, u, d);
    735 
    736     for (int i=0 ; i<4 ; i++) {
    737         if (!texel.format.c[i].h) continue;
    738         texel.format.c[i].h = FRAC_BITS*2+8;
    739         texel.format.c[i].l = FRAC_BITS*2; // keeping 8 bits in enough
    740     }
    741     texel.format.size = 4;
    742     texel.format.bitsPerPixel = 32;
    743     texel.flags |= CLEAR_LO;
    744 }
    745 
    746 void GGLAssembler::filter16(
    747         const fragment_parts_t& parts,
    748         pixel_t& texel, const texture_unit_t& tmu,
    749         int U, int V, pointer_t& txPtr,
    750         int FRAC_BITS)
    751 {
    752     // compute the mask
    753     // XXX: it would be nice if the mask below could be computed
    754     // automatically.
    755     uint32_t mask = 0;
    756     int shift = 0;
    757     int prec = 0;
    758     switch (tmu.format_idx) {
    759         case GGL_PIXEL_FORMAT_RGB_565:
    760             // source: 00000ggg.ggg00000 | rrrrr000.000bbbbb
    761             // result: gggggggg.gggrrrrr | rrrrr0bb.bbbbbbbb
    762             mask = 0x07E0F81F;
    763             shift = 16;
    764             prec = 5;
    765             break;
    766         case GGL_PIXEL_FORMAT_RGBA_4444:
    767             // 0000,1111,0000,1111 | 0000,1111,0000,1111
    768             mask = 0x0F0F0F0F;
    769             shift = 12;
    770             prec = 4;
    771             break;
    772         case GGL_PIXEL_FORMAT_LA_88:
    773             // 0000,0000,1111,1111 | 0000,0000,1111,1111
    774             // AALL -> 00AA | 00LL
    775             mask = 0x00FF00FF;
    776             shift = 8;
    777             prec = 8;
    778             break;
    779         default:
    780             // unsupported format, do something sensical...
    781             ALOGE("Unsupported 16-bits texture format (%d)", tmu.format_idx);
    782             LDRH(AL, texel.reg, txPtr.reg);
    783             return;
    784     }
    785 
    786     const int adjust = FRAC_BITS*2 - prec;
    787     const int round  = 0;
    788 
    789     // update the texel format
    790     texel.format.size = 4;
    791     texel.format.bitsPerPixel = 32;
    792     texel.flags |= CLEAR_HI|CLEAR_LO;
    793     for (int i=0 ; i<4 ; i++) {
    794         if (!texel.format.c[i].h) continue;
    795         const uint32_t offset = (mask & tmu.format.mask(i)) ? 0 : shift;
    796         texel.format.c[i].h = tmu.format.c[i].h + offset + prec;
    797         texel.format.c[i].l = texel.format.c[i].h - (tmu.format.bits(i) + prec);
    798     }
    799 
    800     // ------------------------
    801     // about ~40 cycles / pixel
    802     Scratch scratches(registerFile());
    803 
    804     int pixel= scratches.obtain();
    805     int d    = scratches.obtain();
    806     int u    = scratches.obtain();
    807     int k    = scratches.obtain();
    808 
    809     // RB -> U * V
    810     int offset = pixel;
    811     CONTEXT_LOAD(offset, generated_vars.rt);
    812     CONTEXT_LOAD(u, generated_vars.lb);
    813     ADD(AL, 0, offset, offset, u);
    814 
    815     LDRH(AL, pixel, txPtr.reg, reg_pre(offset));
    816     SMULBB(AL, u, U, V);
    817     ORR(AL, 0, pixel, pixel, reg_imm(pixel, LSL, shift));
    818     build_and_immediate(pixel, pixel, mask, 32);
    819     if (adjust) {
    820         if (round)
    821             ADD(AL, 0, u, u, imm(1<<(adjust-1)));
    822         MOV(AL, 0, u, reg_imm(u, LSR, adjust));
    823     }
    824     MUL(AL, 0, d, pixel, u);
    825     RSB(AL, 0, k, u, imm(1<<prec));
    826 
    827     // LB -> (1-U) * V
    828     CONTEXT_LOAD(offset, generated_vars.lb);
    829     RSB(AL, 0, U, U, imm(1<<FRAC_BITS));
    830     LDRH(AL, pixel, txPtr.reg, reg_pre(offset));
    831     SMULBB(AL, u, U, V);
    832     ORR(AL, 0, pixel, pixel, reg_imm(pixel, LSL, shift));
    833     build_and_immediate(pixel, pixel, mask, 32);
    834     if (adjust) {
    835         if (round)
    836             ADD(AL, 0, u, u, imm(1<<(adjust-1)));
    837         MOV(AL, 0, u, reg_imm(u, LSR, adjust));
    838     }
    839     MLA(AL, 0, d, pixel, u, d);
    840     SUB(AL, 0, k, k, u);
    841 
    842     // LT -> (1-U)*(1-V)
    843     RSB(AL, 0, V, V, imm(1<<FRAC_BITS));
    844     LDRH(AL, pixel, txPtr.reg);
    845     SMULBB(AL, u, U, V);
    846     ORR(AL, 0, pixel, pixel, reg_imm(pixel, LSL, shift));
    847     build_and_immediate(pixel, pixel, mask, 32);
    848     if (adjust) {
    849         if (round)
    850             ADD(AL, 0, u, u, imm(1<<(adjust-1)));
    851         MOV(AL, 0, u, reg_imm(u, LSR, adjust));
    852     }
    853     MLA(AL, 0, d, pixel, u, d);
    854 
    855     // RT -> U*(1-V)
    856     CONTEXT_LOAD(offset, generated_vars.rt);
    857     LDRH(AL, pixel, txPtr.reg, reg_pre(offset));
    858     SUB(AL, 0, u, k, u);
    859     ORR(AL, 0, pixel, pixel, reg_imm(pixel, LSL, shift));
    860     build_and_immediate(pixel, pixel, mask, 32);
    861     MLA(AL, 0, texel.reg, pixel, u, d);
    862 }
    863 
    864 void GGLAssembler::filter24(
    865         const fragment_parts_t& parts,
    866         pixel_t& texel, const texture_unit_t& tmu,
    867         int U, int V, pointer_t& txPtr,
    868         int FRAC_BITS)
    869 {
    870     // not supported yet (currently disabled)
    871     load(txPtr, texel, 0);
    872 }
    873 
    874 #if __ARM_ARCH__ >= 6
    875 // ARMv6 version, using UXTB16, and scheduled for Cortex-A8 pipeline
    876 void GGLAssembler::filter32(
    877         const fragment_parts_t& parts,
    878         pixel_t& texel, const texture_unit_t& tmu,
    879         int U, int V, pointer_t& txPtr,
    880         int FRAC_BITS)
    881 {
    882     const int adjust = FRAC_BITS*2 - 8;
    883     const int round  = 0;
    884     const int prescale = 16 - adjust;
    885 
    886     Scratch scratches(registerFile());
    887 
    888     int pixel= scratches.obtain();
    889     int dh   = scratches.obtain();
    890     int u    = scratches.obtain();
    891     int k    = scratches.obtain();
    892 
    893     int temp = scratches.obtain();
    894     int dl   = scratches.obtain();
    895 
    896     int offsetrt = scratches.obtain();
    897     int offsetlb = scratches.obtain();
    898 
    899     int pixellb = offsetlb;
    900 
    901     // RB -> U * V
    902     CONTEXT_LOAD(offsetrt, generated_vars.rt);
    903     CONTEXT_LOAD(offsetlb, generated_vars.lb);
    904     if(!round) {
    905         MOV(AL, 0, U, reg_imm(U, LSL, prescale));
    906     }
    907     ADD(AL, 0, u, offsetrt, offsetlb);
    908 
    909     LDR(AL, pixel, txPtr.reg, reg_scale_pre(u));
    910     if (round) {
    911         SMULBB(AL, u, U, V);
    912         RSB(AL, 0, U, U, imm(1<<FRAC_BITS));
    913     } else {
    914         SMULWB(AL, u, U, V);
    915         RSB(AL, 0, U, U, imm(1<<(FRAC_BITS+prescale)));
    916     }
    917     UXTB16(AL, temp, pixel, 0);
    918     if (round) {
    919         ADD(AL, 0, u, u, imm(1<<(adjust-1)));
    920         MOV(AL, 0, u, reg_imm(u, LSR, adjust));
    921     }
    922     LDR(AL, pixellb, txPtr.reg, reg_scale_pre(offsetlb));
    923     MUL(AL, 0, dh, temp, u);
    924     UXTB16(AL, temp, pixel, 8);
    925     MUL(AL, 0, dl, temp, u);
    926     RSB(AL, 0, k, u, imm(0x100));
    927 
    928     // LB -> (1-U) * V
    929     if (round) {
    930         SMULBB(AL, u, U, V);
    931     } else {
    932         SMULWB(AL, u, U, V);
    933     }
    934     UXTB16(AL, temp, pixellb, 0);
    935     if (round) {
    936         ADD(AL, 0, u, u, imm(1<<(adjust-1)));
    937         MOV(AL, 0, u, reg_imm(u, LSR, adjust));
    938     }
    939     MLA(AL, 0, dh, temp, u, dh);
    940     UXTB16(AL, temp, pixellb, 8);
    941     MLA(AL, 0, dl, temp, u, dl);
    942     SUB(AL, 0, k, k, u);
    943 
    944     // LT -> (1-U)*(1-V)
    945     RSB(AL, 0, V, V, imm(1<<FRAC_BITS));
    946     LDR(AL, pixel, txPtr.reg);
    947     if (round) {
    948         SMULBB(AL, u, U, V);
    949     } else {
    950         SMULWB(AL, u, U, V);
    951     }
    952     UXTB16(AL, temp, pixel, 0);
    953     if (round) {
    954         ADD(AL, 0, u, u, imm(1<<(adjust-1)));
    955         MOV(AL, 0, u, reg_imm(u, LSR, adjust));
    956     }
    957     MLA(AL, 0, dh, temp, u, dh);
    958     UXTB16(AL, temp, pixel, 8);
    959     MLA(AL, 0, dl, temp, u, dl);
    960 
    961     // RT -> U*(1-V)
    962     LDR(AL, pixel, txPtr.reg, reg_scale_pre(offsetrt));
    963     SUB(AL, 0, u, k, u);
    964     UXTB16(AL, temp, pixel, 0);
    965     MLA(AL, 0, dh, temp, u, dh);
    966     UXTB16(AL, temp, pixel, 8);
    967     MLA(AL, 0, dl, temp, u, dl);
    968 
    969     UXTB16(AL, dh, dh, 8);
    970     UXTB16(AL, dl, dl, 8);
    971     ORR(AL, 0, texel.reg, dh, reg_imm(dl, LSL, 8));
    972 }
    973 #else
    974 void GGLAssembler::filter32(
    975         const fragment_parts_t& parts,
    976         pixel_t& texel, const texture_unit_t& tmu,
    977         int U, int V, pointer_t& txPtr,
    978         int FRAC_BITS)
    979 {
    980     const int adjust = FRAC_BITS*2 - 8;
    981     const int round  = 0;
    982 
    983     // ------------------------
    984     // about ~38 cycles / pixel
    985     Scratch scratches(registerFile());
    986 
    987     int pixel= scratches.obtain();
    988     int dh   = scratches.obtain();
    989     int u    = scratches.obtain();
    990     int k    = scratches.obtain();
    991 
    992     int temp = scratches.obtain();
    993     int dl   = scratches.obtain();
    994     int mask = scratches.obtain();
    995 
    996     MOV(AL, 0, mask, imm(0xFF));
    997     ORR(AL, 0, mask, mask, imm(0xFF0000));
    998 
    999     // RB -> U * V
   1000     int offset = pixel;
   1001     CONTEXT_LOAD(offset, generated_vars.rt);
   1002     CONTEXT_LOAD(u, generated_vars.lb);
   1003     ADD(AL, 0, offset, offset, u);
   1004 
   1005     LDR(AL, pixel, txPtr.reg, reg_scale_pre(offset));
   1006     SMULBB(AL, u, U, V);
   1007     AND(AL, 0, temp, mask, pixel);
   1008     if (adjust) {
   1009         if (round)
   1010             ADD(AL, 0, u, u, imm(1<<(adjust-1)));
   1011         MOV(AL, 0, u, reg_imm(u, LSR, adjust));
   1012     }
   1013     MUL(AL, 0, dh, temp, u);
   1014     AND(AL, 0, temp, mask, reg_imm(pixel, LSR, 8));
   1015     MUL(AL, 0, dl, temp, u);
   1016     RSB(AL, 0, k, u, imm(0x100));
   1017 
   1018     // LB -> (1-U) * V
   1019     CONTEXT_LOAD(offset, generated_vars.lb);
   1020     RSB(AL, 0, U, U, imm(1<<FRAC_BITS));
   1021     LDR(AL, pixel, txPtr.reg, reg_scale_pre(offset));
   1022     SMULBB(AL, u, U, V);
   1023     AND(AL, 0, temp, mask, pixel);
   1024     if (adjust) {
   1025         if (round)
   1026             ADD(AL, 0, u, u, imm(1<<(adjust-1)));
   1027         MOV(AL, 0, u, reg_imm(u, LSR, adjust));
   1028     }
   1029     MLA(AL, 0, dh, temp, u, dh);
   1030     AND(AL, 0, temp, mask, reg_imm(pixel, LSR, 8));
   1031     MLA(AL, 0, dl, temp, u, dl);
   1032     SUB(AL, 0, k, k, u);
   1033 
   1034     // LT -> (1-U)*(1-V)
   1035     RSB(AL, 0, V, V, imm(1<<FRAC_BITS));
   1036     LDR(AL, pixel, txPtr.reg);
   1037     SMULBB(AL, u, U, V);
   1038     AND(AL, 0, temp, mask, pixel);
   1039     if (adjust) {
   1040         if (round)
   1041             ADD(AL, 0, u, u, imm(1<<(adjust-1)));
   1042         MOV(AL, 0, u, reg_imm(u, LSR, adjust));
   1043     }
   1044     MLA(AL, 0, dh, temp, u, dh);
   1045     AND(AL, 0, temp, mask, reg_imm(pixel, LSR, 8));
   1046     MLA(AL, 0, dl, temp, u, dl);
   1047 
   1048     // RT -> U*(1-V)
   1049     CONTEXT_LOAD(offset, generated_vars.rt);
   1050     LDR(AL, pixel, txPtr.reg, reg_scale_pre(offset));
   1051     SUB(AL, 0, u, k, u);
   1052     AND(AL, 0, temp, mask, pixel);
   1053     MLA(AL, 0, dh, temp, u, dh);
   1054     AND(AL, 0, temp, mask, reg_imm(pixel, LSR, 8));
   1055     MLA(AL, 0, dl, temp, u, dl);
   1056 
   1057     AND(AL, 0, dh, mask, reg_imm(dh, LSR, 8));
   1058     AND(AL, 0, dl, dl, reg_imm(mask, LSL, 8));
   1059     ORR(AL, 0, texel.reg, dh, dl);
   1060 }
   1061 #endif
   1062 
   1063 void GGLAssembler::build_texture_environment(
   1064         component_t& fragment,
   1065         const fragment_parts_t& parts,
   1066         int component,
   1067         Scratch& regs)
   1068 {
   1069     const uint32_t component_mask = 1<<component;
   1070     const bool multiTexture = mTextureMachine.activeUnits > 1;
   1071     for (int i=0 ; i<GGL_TEXTURE_UNIT_COUNT ; i++) {
   1072         texture_unit_t& tmu = mTextureMachine.tmu[i];
   1073 
   1074         if (tmu.mask & component_mask) {
   1075             // replace or modulate with this texture
   1076             if ((tmu.replaced & component_mask) == 0) {
   1077                 // not replaced by a later tmu...
   1078 
   1079                 Scratch scratches(registerFile());
   1080                 pixel_t texel(parts.texel[i]);
   1081                 if (multiTexture &&
   1082                     tmu.swrap == GGL_NEEDS_WRAP_11 &&
   1083                     tmu.twrap == GGL_NEEDS_WRAP_11)
   1084                 {
   1085                     texel.reg = scratches.obtain();
   1086                     texel.flags |= CORRUPTIBLE;
   1087                     comment("fetch texel (multitexture 1:1)");
   1088                     load(parts.coords[i].ptr, texel, WRITE_BACK);
   1089                  }
   1090 
   1091                 component_t incoming(fragment);
   1092                 modify(fragment, regs);
   1093 
   1094                 switch (tmu.env) {
   1095                 case GGL_REPLACE:
   1096                     extract(fragment, texel, component);
   1097                     break;
   1098                 case GGL_MODULATE:
   1099                     modulate(fragment, incoming, texel, component);
   1100                     break;
   1101                 case GGL_DECAL:
   1102                     decal(fragment, incoming, texel, component);
   1103                     break;
   1104                 case GGL_BLEND:
   1105                     blend(fragment, incoming, texel, component, i);
   1106                     break;
   1107                 case GGL_ADD:
   1108                     add(fragment, incoming, texel, component);
   1109                     break;
   1110                 }
   1111             }
   1112         }
   1113     }
   1114 }
   1115 
   1116 // ---------------------------------------------------------------------------
   1117 
   1118 void GGLAssembler::wrapping(
   1119             int d,
   1120             int coord, int size,
   1121             int tx_wrap, int tx_linear)
   1122 {
   1123     // notes:
   1124     // if tx_linear is set, we need 4 extra bits of precision on the result
   1125     // SMULL/UMULL is 3 cycles
   1126     Scratch scratches(registerFile());
   1127     int c = coord;
   1128     if (tx_wrap == GGL_NEEDS_WRAP_REPEAT) {
   1129         // UMULL takes 4 cycles (interlocked), and we can get away with
   1130         // 2 cycles using SMULWB, but we're loosing 16 bits of precision
   1131         // out of 32 (this is not a problem because the iterator keeps
   1132         // its full precision)
   1133         // UMULL(AL, 0, size, d, c, size);
   1134         // note: we can't use SMULTB because it's signed.
   1135         MOV(AL, 0, d, reg_imm(c, LSR, 16-tx_linear));
   1136         SMULWB(AL, d, d, size);
   1137     } else if (tx_wrap == GGL_NEEDS_WRAP_CLAMP_TO_EDGE) {
   1138         if (tx_linear) {
   1139             // 1 cycle
   1140             MOV(AL, 0, d, reg_imm(coord, ASR, 16-tx_linear));
   1141         } else {
   1142             // 4 cycles (common case)
   1143             MOV(AL, 0, d, reg_imm(coord, ASR, 16));
   1144             BIC(AL, 0, d, d, reg_imm(d, ASR, 31));
   1145             CMP(AL, d, size);
   1146             SUB(GE, 0, d, size, imm(1));
   1147         }
   1148     }
   1149 }
   1150 
   1151 // ---------------------------------------------------------------------------
   1152 
   1153 void GGLAssembler::modulate(
   1154         component_t& dest,
   1155         const component_t& incoming,
   1156         const pixel_t& incomingTexel, int component)
   1157 {
   1158     Scratch locals(registerFile());
   1159     integer_t texel(locals.obtain(), 32, CORRUPTIBLE);
   1160     extract(texel, incomingTexel, component);
   1161 
   1162     const int Nt = texel.size();
   1163         // Nt should always be less than 10 bits because it comes
   1164         // from the TMU.
   1165 
   1166     int Ni = incoming.size();
   1167         // Ni could be big because it comes from previous MODULATEs
   1168 
   1169     if (Nt == 1) {
   1170         // texel acts as a bit-mask
   1171         // dest = incoming & ((texel << incoming.h)-texel)
   1172         RSB(AL, 0, dest.reg, texel.reg, reg_imm(texel.reg, LSL, incoming.h));
   1173         AND(AL, 0, dest.reg, dest.reg, incoming.reg);
   1174         dest.l = incoming.l;
   1175         dest.h = incoming.h;
   1176         dest.flags |= (incoming.flags & CLEAR_LO);
   1177     } else if (Ni == 1) {
   1178         MOV(AL, 0, dest.reg, reg_imm(incoming.reg, LSL, 31-incoming.h));
   1179         AND(AL, 0, dest.reg, texel.reg, reg_imm(dest.reg, ASR, 31));
   1180         dest.l = 0;
   1181         dest.h = Nt;
   1182     } else {
   1183         int inReg = incoming.reg;
   1184         int shift = incoming.l;
   1185         if ((Nt + Ni) > 32) {
   1186             // we will overflow, reduce the precision of Ni to 8 bits
   1187             // (Note Nt cannot be more than 10 bits which happens with
   1188             // 565 textures and GGL_LINEAR)
   1189             shift += Ni-8;
   1190             Ni = 8;
   1191         }
   1192 
   1193         // modulate by the component with the lowest precision
   1194         if (Nt >= Ni) {
   1195             if (shift) {
   1196                 // XXX: we should be able to avoid this shift
   1197                 // when shift==16 && Nt<16 && Ni<16, in which
   1198                 // we could use SMULBT below.
   1199                 MOV(AL, 0, dest.reg, reg_imm(inReg, LSR, shift));
   1200                 inReg = dest.reg;
   1201                 shift = 0;
   1202             }
   1203             // operation:           (Cf*Ct)/((1<<Ni)-1)
   1204             // approximated with:   Cf*(Ct + Ct>>(Ni-1))>>Ni
   1205             // this operation doesn't change texel's size
   1206             ADD(AL, 0, dest.reg, inReg, reg_imm(inReg, LSR, Ni-1));
   1207             if (Nt<16 && Ni<16) SMULBB(AL, dest.reg, texel.reg, dest.reg);
   1208             else                MUL(AL, 0, dest.reg, texel.reg, dest.reg);
   1209             dest.l = Ni;
   1210             dest.h = Nt + Ni;
   1211         } else {
   1212             if (shift && (shift != 16)) {
   1213                 // if shift==16, we can use 16-bits mul instructions later
   1214                 MOV(AL, 0, dest.reg, reg_imm(inReg, LSR, shift));
   1215                 inReg = dest.reg;
   1216                 shift = 0;
   1217             }
   1218             // operation:           (Cf*Ct)/((1<<Nt)-1)
   1219             // approximated with:   Ct*(Cf + Cf>>(Nt-1))>>Nt
   1220             // this operation doesn't change incoming's size
   1221             Scratch scratches(registerFile());
   1222             int t = (texel.flags & CORRUPTIBLE) ? texel.reg : dest.reg;
   1223             if (t == inReg)
   1224                 t = scratches.obtain();
   1225             ADD(AL, 0, t, texel.reg, reg_imm(texel.reg, LSR, Nt-1));
   1226             if (Nt<16 && Ni<16) {
   1227                 if (shift==16)  SMULBT(AL, dest.reg, t, inReg);
   1228                 else            SMULBB(AL, dest.reg, t, inReg);
   1229             } else              MUL(AL, 0, dest.reg, t, inReg);
   1230             dest.l = Nt;
   1231             dest.h = Nt + Ni;
   1232         }
   1233 
   1234         // low bits are not valid
   1235         dest.flags |= CLEAR_LO;
   1236 
   1237         // no need to keep more than 8 bits/component
   1238         if (dest.size() > 8)
   1239             dest.l = dest.h-8;
   1240     }
   1241 }
   1242 
   1243 void GGLAssembler::decal(
   1244         component_t& dest,
   1245         const component_t& incoming,
   1246         const pixel_t& incomingTexel, int component)
   1247 {
   1248     // RGBA:
   1249     // Cv = Cf*(1 - At) + Ct*At = Cf + (Ct - Cf)*At
   1250     // Av = Af
   1251     Scratch locals(registerFile());
   1252     integer_t texel(locals.obtain(), 32, CORRUPTIBLE);
   1253     integer_t factor(locals.obtain(), 32, CORRUPTIBLE);
   1254     extract(texel, incomingTexel, component);
   1255     extract(factor, incomingTexel, GGLFormat::ALPHA);
   1256 
   1257     // no need to keep more than 8-bits for decal
   1258     int Ni = incoming.size();
   1259     int shift = incoming.l;
   1260     if (Ni > 8) {
   1261         shift += Ni-8;
   1262         Ni = 8;
   1263     }
   1264     integer_t incomingNorm(incoming.reg, Ni, incoming.flags);
   1265     if (shift) {
   1266         MOV(AL, 0, dest.reg, reg_imm(incomingNorm.reg, LSR, shift));
   1267         incomingNorm.reg = dest.reg;
   1268         incomingNorm.flags |= CORRUPTIBLE;
   1269     }
   1270     ADD(AL, 0, factor.reg, factor.reg, reg_imm(factor.reg, LSR, factor.s-1));
   1271     build_blendOneMinusFF(dest, factor, incomingNorm, texel);
   1272 }
   1273 
   1274 void GGLAssembler::blend(
   1275         component_t& dest,
   1276         const component_t& incoming,
   1277         const pixel_t& incomingTexel, int component, int tmu)
   1278 {
   1279     // RGBA:
   1280     // Cv = (1 - Ct)*Cf + Ct*Cc = Cf + (Cc - Cf)*Ct
   1281     // Av = At*Af
   1282 
   1283     if (component == GGLFormat::ALPHA) {
   1284         modulate(dest, incoming, incomingTexel, component);
   1285         return;
   1286     }
   1287 
   1288     Scratch locals(registerFile());
   1289     integer_t color(locals.obtain(), 8, CORRUPTIBLE);
   1290     integer_t factor(locals.obtain(), 32, CORRUPTIBLE);
   1291     LDRB(AL, color.reg, mBuilderContext.Rctx,
   1292             immed12_pre(GGL_OFFSETOF(state.texture[tmu].env_color[component])));
   1293     extract(factor, incomingTexel, component);
   1294 
   1295     // no need to keep more than 8-bits for blend
   1296     int Ni = incoming.size();
   1297     int shift = incoming.l;
   1298     if (Ni > 8) {
   1299         shift += Ni-8;
   1300         Ni = 8;
   1301     }
   1302     integer_t incomingNorm(incoming.reg, Ni, incoming.flags);
   1303     if (shift) {
   1304         MOV(AL, 0, dest.reg, reg_imm(incomingNorm.reg, LSR, shift));
   1305         incomingNorm.reg = dest.reg;
   1306         incomingNorm.flags |= CORRUPTIBLE;
   1307     }
   1308     ADD(AL, 0, factor.reg, factor.reg, reg_imm(factor.reg, LSR, factor.s-1));
   1309     build_blendOneMinusFF(dest, factor, incomingNorm, color);
   1310 }
   1311 
   1312 void GGLAssembler::add(
   1313         component_t& dest,
   1314         const component_t& incoming,
   1315         const pixel_t& incomingTexel, int component)
   1316 {
   1317     // RGBA:
   1318     // Cv = Cf + Ct;
   1319     Scratch locals(registerFile());
   1320 
   1321     component_t incomingTemp(incoming);
   1322 
   1323     // use "dest" as a temporary for extracting the texel, unless "dest"
   1324     // overlaps "incoming".
   1325     integer_t texel(dest.reg, 32, CORRUPTIBLE);
   1326     if (dest.reg == incomingTemp.reg)
   1327         texel.reg = locals.obtain();
   1328     extract(texel, incomingTexel, component);
   1329 
   1330     if (texel.s < incomingTemp.size()) {
   1331         expand(texel, texel, incomingTemp.size());
   1332     } else if (texel.s > incomingTemp.size()) {
   1333         if (incomingTemp.flags & CORRUPTIBLE) {
   1334             expand(incomingTemp, incomingTemp, texel.s);
   1335         } else {
   1336             incomingTemp.reg = locals.obtain();
   1337             expand(incomingTemp, incoming, texel.s);
   1338         }
   1339     }
   1340 
   1341     if (incomingTemp.l) {
   1342         ADD(AL, 0, dest.reg, texel.reg,
   1343                 reg_imm(incomingTemp.reg, LSR, incomingTemp.l));
   1344     } else {
   1345         ADD(AL, 0, dest.reg, texel.reg, incomingTemp.reg);
   1346     }
   1347     dest.l = 0;
   1348     dest.h = texel.size();
   1349     component_sat(dest);
   1350 }
   1351 
   1352 // ----------------------------------------------------------------------------
   1353 
   1354 }; // namespace android
   1355 
   1356