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 "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             if (registerFile().status() & RegisterFile::OUT_OF_REGISTERS)
    468                 return;
    469 
    470             comment("compute repeat/clamp");
    471             int u       = scratches.obtain();
    472             int v       = scratches.obtain();
    473             int width   = scratches.obtain();
    474             int height  = scratches.obtain();
    475             int U = 0;
    476             int V = 0;
    477 
    478             if (registerFile().status() & RegisterFile::OUT_OF_REGISTERS)
    479                 return;
    480 
    481             CONTEXT_LOAD(width,  generated_vars.texture[i].width);
    482             CONTEXT_LOAD(height, generated_vars.texture[i].height);
    483 
    484             int FRAC_BITS = 0;
    485             if (tmu.linear) {
    486                 // linear interpolation
    487                 if (tmu.format.size == 1) {
    488                     // for 8-bits textures, we can afford
    489                     // 7 bits of fractional precision at no
    490                     // additional cost (we can't do 8 bits
    491                     // because filter8 uses signed 16 bits muls)
    492                     FRAC_BITS = 7;
    493                 } else if (tmu.format.size == 2) {
    494                     // filter16() is internally limited to 4 bits, so:
    495                     // FRAC_BITS=2 generates less instructions,
    496                     // FRAC_BITS=3,4,5 creates unpleasant artifacts,
    497                     // FRAC_BITS=6+ looks good
    498                     FRAC_BITS = 6;
    499                 } else if (tmu.format.size == 4) {
    500                     // filter32() is internally limited to 8 bits, so:
    501                     // FRAC_BITS=4 looks good
    502                     // FRAC_BITS=5+ looks better, but generates 3 extra ipp
    503                     FRAC_BITS = 6;
    504                 } else {
    505                     // for all other cases we use 4 bits.
    506                     FRAC_BITS = 4;
    507                 }
    508             }
    509             wrapping(u, s.reg, width,  tmu.swrap, FRAC_BITS);
    510             wrapping(v, t.reg, height, tmu.twrap, FRAC_BITS);
    511 
    512             if (tmu.linear) {
    513                 comment("compute linear filtering offsets");
    514                 // pixel size scale
    515                 const int shift = 31 - gglClz(tmu.format.size);
    516                 U = scratches.obtain();
    517                 V = scratches.obtain();
    518 
    519                 if (registerFile().status() & RegisterFile::OUT_OF_REGISTERS)
    520                     return;
    521 
    522                 // sample the texel center
    523                 SUB(AL, 0, u, u, imm(1<<(FRAC_BITS-1)));
    524                 SUB(AL, 0, v, v, imm(1<<(FRAC_BITS-1)));
    525 
    526                 // get the fractionnal part of U,V
    527                 AND(AL, 0, U, u, imm((1<<FRAC_BITS)-1));
    528                 AND(AL, 0, V, v, imm((1<<FRAC_BITS)-1));
    529 
    530                 // compute width-1 and height-1
    531                 SUB(AL, 0, width,  width,  imm(1));
    532                 SUB(AL, 0, height, height, imm(1));
    533 
    534                 // get the integer part of U,V and clamp/wrap
    535                 // and compute offset to the next texel
    536                 if (tmu.swrap == GGL_NEEDS_WRAP_REPEAT) {
    537                     // u has already been REPEATed
    538                     MOV(AL, 1, u, reg_imm(u, ASR, FRAC_BITS));
    539                     MOV(MI, 0, u, width);
    540                     CMP(AL, u, width);
    541                     MOV(LT, 0, width, imm(1 << shift));
    542                     if (shift)
    543                         MOV(GE, 0, width, reg_imm(width, LSL, shift));
    544                     RSB(GE, 0, width, width, imm(0));
    545                 } else {
    546                     // u has not been CLAMPed yet
    547                     // algorithm:
    548                     // if ((u>>4) >= width)
    549                     //      u = width<<4
    550                     //      width = 0
    551                     // else
    552                     //      width = 1<<shift
    553                     // u = u>>4; // get integer part
    554                     // if (u<0)
    555                     //      u = 0
    556                     //      width = 0
    557                     // generated_vars.rt = width
    558 
    559                     CMP(AL, width, reg_imm(u, ASR, FRAC_BITS));
    560                     MOV(LE, 0, u, reg_imm(width, LSL, FRAC_BITS));
    561                     MOV(LE, 0, width, imm(0));
    562                     MOV(GT, 0, width, imm(1 << shift));
    563                     MOV(AL, 1, u, reg_imm(u, ASR, FRAC_BITS));
    564                     MOV(MI, 0, u, imm(0));
    565                     MOV(MI, 0, width, imm(0));
    566                 }
    567                 CONTEXT_STORE(width, generated_vars.rt);
    568 
    569                 const int stride = width;
    570                 CONTEXT_LOAD(stride, generated_vars.texture[i].stride);
    571                 if (tmu.twrap == GGL_NEEDS_WRAP_REPEAT) {
    572                     // v has already been REPEATed
    573                     MOV(AL, 1, v, reg_imm(v, ASR, FRAC_BITS));
    574                     MOV(MI, 0, v, height);
    575                     CMP(AL, v, height);
    576                     MOV(LT, 0, height, imm(1 << shift));
    577                     if (shift)
    578                         MOV(GE, 0, height, reg_imm(height, LSL, shift));
    579                     RSB(GE, 0, height, height, imm(0));
    580                     MUL(AL, 0, height, stride, height);
    581                 } else {
    582                     // v has not been CLAMPed yet
    583                     CMP(AL, height, reg_imm(v, ASR, FRAC_BITS));
    584                     MOV(LE, 0, v, reg_imm(height, LSL, FRAC_BITS));
    585                     MOV(LE, 0, height, imm(0));
    586                     if (shift) {
    587                         MOV(GT, 0, height, reg_imm(stride, LSL, shift));
    588                     } else {
    589                         MOV(GT, 0, height, stride);
    590                     }
    591                     MOV(AL, 1, v, reg_imm(v, ASR, FRAC_BITS));
    592                     MOV(MI, 0, v, imm(0));
    593                     MOV(MI, 0, height, imm(0));
    594                 }
    595                 CONTEXT_STORE(height, generated_vars.lb);
    596             }
    597 
    598             scratches.recycle(width);
    599             scratches.recycle(height);
    600 
    601             // iterate texture coordinates...
    602             comment("iterate s,t");
    603             int dsdx = scratches.obtain();
    604             int dtdx = scratches.obtain();
    605 
    606             if (registerFile().status() & RegisterFile::OUT_OF_REGISTERS)
    607                 return;
    608 
    609             CONTEXT_LOAD(dsdx, generated_vars.texture[i].dsdx);
    610             CONTEXT_LOAD(dtdx, generated_vars.texture[i].dtdx);
    611             ADD(AL, 0, s.reg, s.reg, dsdx);
    612             ADD(AL, 0, t.reg, t.reg, dtdx);
    613             if ((mOptLevel&1)==0) {
    614                 CONTEXT_STORE(s.reg, generated_vars.texture[i].spill[0]);
    615                 CONTEXT_STORE(t.reg, generated_vars.texture[i].spill[1]);
    616                 scratches.recycle(s.reg);
    617                 scratches.recycle(t.reg);
    618             }
    619             scratches.recycle(dsdx);
    620             scratches.recycle(dtdx);
    621 
    622             // merge base & offset...
    623             comment("merge base & offset");
    624             texel.setTo(regs.obtain(), &tmu.format);
    625             txPtr.setTo(texel.reg, tmu.bits);
    626             int stride = scratches.obtain();
    627 
    628             if (registerFile().status() & RegisterFile::OUT_OF_REGISTERS)
    629                 return;
    630 
    631             CONTEXT_LOAD(stride,    generated_vars.texture[i].stride);
    632             CONTEXT_LOAD(txPtr.reg, generated_vars.texture[i].data);
    633             SMLABB(AL, u, v, stride, u);    // u+v*stride
    634             base_offset(txPtr, txPtr, u);
    635 
    636             // load texel
    637             if (!tmu.linear) {
    638                 comment("fetch texel");
    639                 load(txPtr, texel, 0);
    640             } else {
    641                 // recycle registers we don't need anymore
    642                 scratches.recycle(u);
    643                 scratches.recycle(v);
    644                 scratches.recycle(stride);
    645 
    646                 comment("fetch texel, bilinear");
    647                 switch (tmu.format.size) {
    648                 case 1:  filter8(parts, texel, tmu, U, V, txPtr, FRAC_BITS); break;
    649                 case 2: filter16(parts, texel, tmu, U, V, txPtr, FRAC_BITS); break;
    650                 case 3: filter24(parts, texel, tmu, U, V, txPtr, FRAC_BITS); break;
    651                 case 4: filter32(parts, texel, tmu, U, V, txPtr, FRAC_BITS); break;
    652                 }
    653             }
    654         }
    655     }
    656 }
    657 
    658 void GGLAssembler::build_iterate_texture_coordinates(
    659     const fragment_parts_t& parts)
    660 {
    661     const bool multiTexture = mTextureMachine.activeUnits > 1;
    662     for (int i=0 ; i<GGL_TEXTURE_UNIT_COUNT; i++) {
    663         const texture_unit_t& tmu = mTextureMachine.tmu[i];
    664         if (tmu.format_idx == 0)
    665             continue;
    666 
    667         if ((tmu.swrap == GGL_NEEDS_WRAP_11) &&
    668             (tmu.twrap == GGL_NEEDS_WRAP_11))
    669         { // 1:1 textures
    670             const pointer_t& txPtr = parts.coords[i].ptr;
    671             ADD(AL, 0, txPtr.reg, txPtr.reg, imm(txPtr.size>>3));
    672         } else {
    673             Scratch scratches(registerFile());
    674             int s = parts.coords[i].s.reg;
    675             int t = parts.coords[i].t.reg;
    676             if ((mOptLevel&1)==0) {
    677                 s = scratches.obtain();
    678                 t = scratches.obtain();
    679                 CONTEXT_LOAD(s, generated_vars.texture[i].spill[0]);
    680                 CONTEXT_LOAD(t, generated_vars.texture[i].spill[1]);
    681             }
    682             int dsdx = scratches.obtain();
    683             int dtdx = scratches.obtain();
    684             CONTEXT_LOAD(dsdx, generated_vars.texture[i].dsdx);
    685             CONTEXT_LOAD(dtdx, generated_vars.texture[i].dtdx);
    686             ADD(AL, 0, s, s, dsdx);
    687             ADD(AL, 0, t, t, dtdx);
    688             if ((mOptLevel&1)==0) {
    689                 CONTEXT_STORE(s, generated_vars.texture[i].spill[0]);
    690                 CONTEXT_STORE(t, generated_vars.texture[i].spill[1]);
    691             }
    692         }
    693     }
    694 }
    695 
    696 void GGLAssembler::filter8(
    697         const fragment_parts_t& parts,
    698         pixel_t& texel, const texture_unit_t& tmu,
    699         int U, int V, pointer_t& txPtr,
    700         int FRAC_BITS)
    701 {
    702     if (tmu.format.components != GGL_ALPHA &&
    703         tmu.format.components != GGL_LUMINANCE)
    704     {
    705         // this is a packed format, and we don't support
    706         // linear filtering (it's probably RGB 332)
    707         // Should not happen with OpenGL|ES
    708         LDRB(AL, texel.reg, txPtr.reg);
    709         return;
    710     }
    711 
    712     // ------------------------
    713     // about ~22 cycles / pixel
    714     Scratch scratches(registerFile());
    715 
    716     int pixel= scratches.obtain();
    717     int d    = scratches.obtain();
    718     int u    = scratches.obtain();
    719     int k    = scratches.obtain();
    720     int rt   = scratches.obtain();
    721     int lb   = scratches.obtain();
    722 
    723     // RB -> U * V
    724 
    725     CONTEXT_LOAD(rt, generated_vars.rt);
    726     CONTEXT_LOAD(lb, generated_vars.lb);
    727 
    728     int offset = pixel;
    729     ADD(AL, 0, offset, lb, rt);
    730     LDRB(AL, pixel, txPtr.reg, reg_scale_pre(offset));
    731     SMULBB(AL, u, U, V);
    732     SMULBB(AL, d, pixel, u);
    733     RSB(AL, 0, k, u, imm(1<<(FRAC_BITS*2)));
    734 
    735     // LB -> (1-U) * V
    736     RSB(AL, 0, U, U, imm(1<<FRAC_BITS));
    737     LDRB(AL, pixel, txPtr.reg, reg_scale_pre(lb));
    738     SMULBB(AL, u, U, V);
    739     SMLABB(AL, d, pixel, u, d);
    740     SUB(AL, 0, k, k, u);
    741 
    742     // LT -> (1-U)*(1-V)
    743     RSB(AL, 0, V, V, imm(1<<FRAC_BITS));
    744     LDRB(AL, pixel, txPtr.reg);
    745     SMULBB(AL, u, U, V);
    746     SMLABB(AL, d, pixel, u, d);
    747 
    748     // RT -> U*(1-V)
    749     LDRB(AL, pixel, txPtr.reg, reg_scale_pre(rt));
    750     SUB(AL, 0, u, k, u);
    751     SMLABB(AL, texel.reg, pixel, u, d);
    752 
    753     for (int i=0 ; i<4 ; i++) {
    754         if (!texel.format.c[i].h) continue;
    755         texel.format.c[i].h = FRAC_BITS*2+8;
    756         texel.format.c[i].l = FRAC_BITS*2; // keeping 8 bits in enough
    757     }
    758     texel.format.size = 4;
    759     texel.format.bitsPerPixel = 32;
    760     texel.flags |= CLEAR_LO;
    761 }
    762 
    763 void GGLAssembler::filter16(
    764         const fragment_parts_t& parts,
    765         pixel_t& texel, const texture_unit_t& tmu,
    766         int U, int V, pointer_t& txPtr,
    767         int FRAC_BITS)
    768 {
    769     // compute the mask
    770     // XXX: it would be nice if the mask below could be computed
    771     // automatically.
    772     uint32_t mask = 0;
    773     int shift = 0;
    774     int prec = 0;
    775     switch (tmu.format_idx) {
    776         case GGL_PIXEL_FORMAT_RGB_565:
    777             // source: 00000ggg.ggg00000 | rrrrr000.000bbbbb
    778             // result: gggggggg.gggrrrrr | rrrrr0bb.bbbbbbbb
    779             mask = 0x07E0F81F;
    780             shift = 16;
    781             prec = 5;
    782             break;
    783         case GGL_PIXEL_FORMAT_RGBA_4444:
    784             // 0000,1111,0000,1111 | 0000,1111,0000,1111
    785             mask = 0x0F0F0F0F;
    786             shift = 12;
    787             prec = 4;
    788             break;
    789         case GGL_PIXEL_FORMAT_LA_88:
    790             // 0000,0000,1111,1111 | 0000,0000,1111,1111
    791             // AALL -> 00AA | 00LL
    792             mask = 0x00FF00FF;
    793             shift = 8;
    794             prec = 8;
    795             break;
    796         default:
    797             // unsupported format, do something sensical...
    798             ALOGE("Unsupported 16-bits texture format (%d)", tmu.format_idx);
    799             LDRH(AL, texel.reg, txPtr.reg);
    800             return;
    801     }
    802 
    803     const int adjust = FRAC_BITS*2 - prec;
    804     const int round  = 0;
    805 
    806     // update the texel format
    807     texel.format.size = 4;
    808     texel.format.bitsPerPixel = 32;
    809     texel.flags |= CLEAR_HI|CLEAR_LO;
    810     for (int i=0 ; i<4 ; i++) {
    811         if (!texel.format.c[i].h) continue;
    812         const uint32_t offset = (mask & tmu.format.mask(i)) ? 0 : shift;
    813         texel.format.c[i].h = tmu.format.c[i].h + offset + prec;
    814         texel.format.c[i].l = texel.format.c[i].h - (tmu.format.bits(i) + prec);
    815     }
    816 
    817     // ------------------------
    818     // about ~40 cycles / pixel
    819     Scratch scratches(registerFile());
    820 
    821     int pixel= scratches.obtain();
    822     int d    = scratches.obtain();
    823     int u    = scratches.obtain();
    824     int k    = scratches.obtain();
    825 
    826     // RB -> U * V
    827     int offset = pixel;
    828     CONTEXT_LOAD(offset, generated_vars.rt);
    829     CONTEXT_LOAD(u, generated_vars.lb);
    830     ADD(AL, 0, offset, offset, u);
    831 
    832     LDRH(AL, pixel, txPtr.reg, reg_pre(offset));
    833     SMULBB(AL, u, U, V);
    834     ORR(AL, 0, pixel, pixel, reg_imm(pixel, LSL, shift));
    835     build_and_immediate(pixel, pixel, mask, 32);
    836     if (adjust) {
    837         if (round)
    838             ADD(AL, 0, u, u, imm(1<<(adjust-1)));
    839         MOV(AL, 0, u, reg_imm(u, LSR, adjust));
    840     }
    841     MUL(AL, 0, d, pixel, u);
    842     RSB(AL, 0, k, u, imm(1<<prec));
    843 
    844     // LB -> (1-U) * V
    845     CONTEXT_LOAD(offset, generated_vars.lb);
    846     RSB(AL, 0, U, U, imm(1<<FRAC_BITS));
    847     LDRH(AL, pixel, txPtr.reg, reg_pre(offset));
    848     SMULBB(AL, u, U, V);
    849     ORR(AL, 0, pixel, pixel, reg_imm(pixel, LSL, shift));
    850     build_and_immediate(pixel, pixel, mask, 32);
    851     if (adjust) {
    852         if (round)
    853             ADD(AL, 0, u, u, imm(1<<(adjust-1)));
    854         MOV(AL, 0, u, reg_imm(u, LSR, adjust));
    855     }
    856     MLA(AL, 0, d, pixel, u, d);
    857     SUB(AL, 0, k, k, u);
    858 
    859     // LT -> (1-U)*(1-V)
    860     RSB(AL, 0, V, V, imm(1<<FRAC_BITS));
    861     LDRH(AL, pixel, txPtr.reg);
    862     SMULBB(AL, u, U, V);
    863     ORR(AL, 0, pixel, pixel, reg_imm(pixel, LSL, shift));
    864     build_and_immediate(pixel, pixel, mask, 32);
    865     if (adjust) {
    866         if (round)
    867             ADD(AL, 0, u, u, imm(1<<(adjust-1)));
    868         MOV(AL, 0, u, reg_imm(u, LSR, adjust));
    869     }
    870     MLA(AL, 0, d, pixel, u, d);
    871 
    872     // RT -> U*(1-V)
    873     CONTEXT_LOAD(offset, generated_vars.rt);
    874     LDRH(AL, pixel, txPtr.reg, reg_pre(offset));
    875     SUB(AL, 0, u, k, u);
    876     ORR(AL, 0, pixel, pixel, reg_imm(pixel, LSL, shift));
    877     build_and_immediate(pixel, pixel, mask, 32);
    878     MLA(AL, 0, texel.reg, pixel, u, d);
    879 }
    880 
    881 void GGLAssembler::filter24(
    882         const fragment_parts_t& parts,
    883         pixel_t& texel, const texture_unit_t& tmu,
    884         int U, int V, pointer_t& txPtr,
    885         int FRAC_BITS)
    886 {
    887     // not supported yet (currently disabled)
    888     load(txPtr, texel, 0);
    889 }
    890 
    891 #if __ARM_ARCH__ >= 6
    892 // ARMv6 version, using UXTB16, and scheduled for Cortex-A8 pipeline
    893 void GGLAssembler::filter32(
    894         const fragment_parts_t& parts,
    895         pixel_t& texel, const texture_unit_t& tmu,
    896         int U, int V, pointer_t& txPtr,
    897         int FRAC_BITS)
    898 {
    899     const int adjust = FRAC_BITS*2 - 8;
    900     const int round  = 0;
    901     const int prescale = 16 - adjust;
    902 
    903     Scratch scratches(registerFile());
    904 
    905     int pixel= scratches.obtain();
    906     int dh   = scratches.obtain();
    907     int u    = scratches.obtain();
    908     int k    = scratches.obtain();
    909 
    910     int temp = scratches.obtain();
    911     int dl   = scratches.obtain();
    912 
    913     int offsetrt = scratches.obtain();
    914     int offsetlb = scratches.obtain();
    915 
    916     int pixellb = offsetlb;
    917 
    918     // RB -> U * V
    919     CONTEXT_LOAD(offsetrt, generated_vars.rt);
    920     CONTEXT_LOAD(offsetlb, generated_vars.lb);
    921     if(!round) {
    922         MOV(AL, 0, U, reg_imm(U, LSL, prescale));
    923     }
    924     ADD(AL, 0, u, offsetrt, offsetlb);
    925 
    926     LDR(AL, pixel, txPtr.reg, reg_scale_pre(u));
    927     if (round) {
    928         SMULBB(AL, u, U, V);
    929         RSB(AL, 0, U, U, imm(1<<FRAC_BITS));
    930     } else {
    931         SMULWB(AL, u, U, V);
    932         RSB(AL, 0, U, U, imm(1<<(FRAC_BITS+prescale)));
    933     }
    934     UXTB16(AL, temp, pixel, 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     LDR(AL, pixellb, txPtr.reg, reg_scale_pre(offsetlb));
    940     MUL(AL, 0, dh, temp, u);
    941     UXTB16(AL, temp, pixel, 8);
    942     MUL(AL, 0, dl, temp, u);
    943     RSB(AL, 0, k, u, imm(0x100));
    944 
    945     // LB -> (1-U) * V
    946     if (round) {
    947         SMULBB(AL, u, U, V);
    948     } else {
    949         SMULWB(AL, u, U, V);
    950     }
    951     UXTB16(AL, temp, pixellb, 0);
    952     if (round) {
    953         ADD(AL, 0, u, u, imm(1<<(adjust-1)));
    954         MOV(AL, 0, u, reg_imm(u, LSR, adjust));
    955     }
    956     MLA(AL, 0, dh, temp, u, dh);
    957     UXTB16(AL, temp, pixellb, 8);
    958     MLA(AL, 0, dl, temp, u, dl);
    959     SUB(AL, 0, k, k, u);
    960 
    961     // LT -> (1-U)*(1-V)
    962     RSB(AL, 0, V, V, imm(1<<FRAC_BITS));
    963     LDR(AL, pixel, txPtr.reg);
    964     if (round) {
    965         SMULBB(AL, u, U, V);
    966     } else {
    967         SMULWB(AL, u, U, V);
    968     }
    969     UXTB16(AL, temp, pixel, 0);
    970     if (round) {
    971         ADD(AL, 0, u, u, imm(1<<(adjust-1)));
    972         MOV(AL, 0, u, reg_imm(u, LSR, adjust));
    973     }
    974     MLA(AL, 0, dh, temp, u, dh);
    975     UXTB16(AL, temp, pixel, 8);
    976     MLA(AL, 0, dl, temp, u, dl);
    977 
    978     // RT -> U*(1-V)
    979     LDR(AL, pixel, txPtr.reg, reg_scale_pre(offsetrt));
    980     SUB(AL, 0, u, k, u);
    981     UXTB16(AL, temp, pixel, 0);
    982     MLA(AL, 0, dh, temp, u, dh);
    983     UXTB16(AL, temp, pixel, 8);
    984     MLA(AL, 0, dl, temp, u, dl);
    985 
    986     UXTB16(AL, dh, dh, 8);
    987     UXTB16(AL, dl, dl, 8);
    988     ORR(AL, 0, texel.reg, dh, reg_imm(dl, LSL, 8));
    989 }
    990 #else
    991 void GGLAssembler::filter32(
    992         const fragment_parts_t& parts,
    993         pixel_t& texel, const texture_unit_t& tmu,
    994         int U, int V, pointer_t& txPtr,
    995         int FRAC_BITS)
    996 {
    997     const int adjust = FRAC_BITS*2 - 8;
    998     const int round  = 0;
    999 
   1000     // ------------------------
   1001     // about ~38 cycles / pixel
   1002     Scratch scratches(registerFile());
   1003 
   1004     int pixel= scratches.obtain();
   1005     int dh   = scratches.obtain();
   1006     int u    = scratches.obtain();
   1007     int k    = scratches.obtain();
   1008 
   1009     int temp = scratches.obtain();
   1010     int dl   = scratches.obtain();
   1011     int mask = scratches.obtain();
   1012 
   1013     MOV(AL, 0, mask, imm(0xFF));
   1014     ORR(AL, 0, mask, mask, imm(0xFF0000));
   1015 
   1016     // RB -> U * V
   1017     int offset = pixel;
   1018     CONTEXT_LOAD(offset, generated_vars.rt);
   1019     CONTEXT_LOAD(u, generated_vars.lb);
   1020     ADD(AL, 0, offset, offset, u);
   1021 
   1022     LDR(AL, pixel, txPtr.reg, reg_scale_pre(offset));
   1023     SMULBB(AL, u, U, V);
   1024     AND(AL, 0, temp, mask, pixel);
   1025     if (adjust) {
   1026         if (round)
   1027             ADD(AL, 0, u, u, imm(1<<(adjust-1)));
   1028         MOV(AL, 0, u, reg_imm(u, LSR, adjust));
   1029     }
   1030     MUL(AL, 0, dh, temp, u);
   1031     AND(AL, 0, temp, mask, reg_imm(pixel, LSR, 8));
   1032     MUL(AL, 0, dl, temp, u);
   1033     RSB(AL, 0, k, u, imm(0x100));
   1034 
   1035     // LB -> (1-U) * V
   1036     CONTEXT_LOAD(offset, generated_vars.lb);
   1037     RSB(AL, 0, U, U, imm(1<<FRAC_BITS));
   1038     LDR(AL, pixel, txPtr.reg, reg_scale_pre(offset));
   1039     SMULBB(AL, u, U, V);
   1040     AND(AL, 0, temp, mask, pixel);
   1041     if (adjust) {
   1042         if (round)
   1043             ADD(AL, 0, u, u, imm(1<<(adjust-1)));
   1044         MOV(AL, 0, u, reg_imm(u, LSR, adjust));
   1045     }
   1046     MLA(AL, 0, dh, temp, u, dh);
   1047     AND(AL, 0, temp, mask, reg_imm(pixel, LSR, 8));
   1048     MLA(AL, 0, dl, temp, u, dl);
   1049     SUB(AL, 0, k, k, u);
   1050 
   1051     // LT -> (1-U)*(1-V)
   1052     RSB(AL, 0, V, V, imm(1<<FRAC_BITS));
   1053     LDR(AL, pixel, txPtr.reg);
   1054     SMULBB(AL, u, U, V);
   1055     AND(AL, 0, temp, mask, pixel);
   1056     if (adjust) {
   1057         if (round)
   1058             ADD(AL, 0, u, u, imm(1<<(adjust-1)));
   1059         MOV(AL, 0, u, reg_imm(u, LSR, adjust));
   1060     }
   1061     MLA(AL, 0, dh, temp, u, dh);
   1062     AND(AL, 0, temp, mask, reg_imm(pixel, LSR, 8));
   1063     MLA(AL, 0, dl, temp, u, dl);
   1064 
   1065     // RT -> U*(1-V)
   1066     CONTEXT_LOAD(offset, generated_vars.rt);
   1067     LDR(AL, pixel, txPtr.reg, reg_scale_pre(offset));
   1068     SUB(AL, 0, u, k, u);
   1069     AND(AL, 0, temp, mask, pixel);
   1070     MLA(AL, 0, dh, temp, u, dh);
   1071     AND(AL, 0, temp, mask, reg_imm(pixel, LSR, 8));
   1072     MLA(AL, 0, dl, temp, u, dl);
   1073 
   1074     AND(AL, 0, dh, mask, reg_imm(dh, LSR, 8));
   1075     AND(AL, 0, dl, dl, reg_imm(mask, LSL, 8));
   1076     ORR(AL, 0, texel.reg, dh, dl);
   1077 }
   1078 #endif
   1079 
   1080 void GGLAssembler::build_texture_environment(
   1081         component_t& fragment,
   1082         const fragment_parts_t& parts,
   1083         int component,
   1084         Scratch& regs)
   1085 {
   1086     const uint32_t component_mask = 1<<component;
   1087     const bool multiTexture = mTextureMachine.activeUnits > 1;
   1088     for (int i=0 ; i<GGL_TEXTURE_UNIT_COUNT ; i++) {
   1089         texture_unit_t& tmu = mTextureMachine.tmu[i];
   1090 
   1091         if (tmu.mask & component_mask) {
   1092             // replace or modulate with this texture
   1093             if ((tmu.replaced & component_mask) == 0) {
   1094                 // not replaced by a later tmu...
   1095 
   1096                 Scratch scratches(registerFile());
   1097                 pixel_t texel(parts.texel[i]);
   1098 
   1099                 if (multiTexture &&
   1100                     tmu.swrap == GGL_NEEDS_WRAP_11 &&
   1101                     tmu.twrap == GGL_NEEDS_WRAP_11)
   1102                 {
   1103                     texel.reg = scratches.obtain();
   1104                     texel.flags |= CORRUPTIBLE;
   1105                     comment("fetch texel (multitexture 1:1)");
   1106                     load(parts.coords[i].ptr, texel, WRITE_BACK);
   1107                  }
   1108 
   1109                 component_t incoming(fragment);
   1110                 modify(fragment, regs);
   1111 
   1112                 switch (tmu.env) {
   1113                 case GGL_REPLACE:
   1114                     extract(fragment, texel, component);
   1115                     break;
   1116                 case GGL_MODULATE:
   1117                     modulate(fragment, incoming, texel, component);
   1118                     break;
   1119                 case GGL_DECAL:
   1120                     decal(fragment, incoming, texel, component);
   1121                     break;
   1122                 case GGL_BLEND:
   1123                     blend(fragment, incoming, texel, component, i);
   1124                     break;
   1125                 case GGL_ADD:
   1126                     add(fragment, incoming, texel, component);
   1127                     break;
   1128                 }
   1129             }
   1130         }
   1131     }
   1132 }
   1133 
   1134 // ---------------------------------------------------------------------------
   1135 
   1136 void GGLAssembler::wrapping(
   1137             int d,
   1138             int coord, int size,
   1139             int tx_wrap, int tx_linear)
   1140 {
   1141     // notes:
   1142     // if tx_linear is set, we need 4 extra bits of precision on the result
   1143     // SMULL/UMULL is 3 cycles
   1144     Scratch scratches(registerFile());
   1145     int c = coord;
   1146     if (tx_wrap == GGL_NEEDS_WRAP_REPEAT) {
   1147         // UMULL takes 4 cycles (interlocked), and we can get away with
   1148         // 2 cycles using SMULWB, but we're loosing 16 bits of precision
   1149         // out of 32 (this is not a problem because the iterator keeps
   1150         // its full precision)
   1151         // UMULL(AL, 0, size, d, c, size);
   1152         // note: we can't use SMULTB because it's signed.
   1153         MOV(AL, 0, d, reg_imm(c, LSR, 16-tx_linear));
   1154         SMULWB(AL, d, d, size);
   1155     } else if (tx_wrap == GGL_NEEDS_WRAP_CLAMP_TO_EDGE) {
   1156         if (tx_linear) {
   1157             // 1 cycle
   1158             MOV(AL, 0, d, reg_imm(coord, ASR, 16-tx_linear));
   1159         } else {
   1160             // 4 cycles (common case)
   1161             MOV(AL, 0, d, reg_imm(coord, ASR, 16));
   1162             BIC(AL, 0, d, d, reg_imm(d, ASR, 31));
   1163             CMP(AL, d, size);
   1164             SUB(GE, 0, d, size, imm(1));
   1165         }
   1166     }
   1167 }
   1168 
   1169 // ---------------------------------------------------------------------------
   1170 
   1171 void GGLAssembler::modulate(
   1172         component_t& dest,
   1173         const component_t& incoming,
   1174         const pixel_t& incomingTexel, int component)
   1175 {
   1176     Scratch locals(registerFile());
   1177     integer_t texel(locals.obtain(), 32, CORRUPTIBLE);
   1178     extract(texel, incomingTexel, component);
   1179 
   1180     const int Nt = texel.size();
   1181         // Nt should always be less than 10 bits because it comes
   1182         // from the TMU.
   1183 
   1184     int Ni = incoming.size();
   1185         // Ni could be big because it comes from previous MODULATEs
   1186 
   1187     if (Nt == 1) {
   1188         // texel acts as a bit-mask
   1189         // dest = incoming & ((texel << incoming.h)-texel)
   1190         RSB(AL, 0, dest.reg, texel.reg, reg_imm(texel.reg, LSL, incoming.h));
   1191         AND(AL, 0, dest.reg, dest.reg, incoming.reg);
   1192         dest.l = incoming.l;
   1193         dest.h = incoming.h;
   1194         dest.flags |= (incoming.flags & CLEAR_LO);
   1195     } else if (Ni == 1) {
   1196         MOV(AL, 0, dest.reg, reg_imm(incoming.reg, LSL, 31-incoming.h));
   1197         AND(AL, 0, dest.reg, texel.reg, reg_imm(dest.reg, ASR, 31));
   1198         dest.l = 0;
   1199         dest.h = Nt;
   1200     } else {
   1201         int inReg = incoming.reg;
   1202         int shift = incoming.l;
   1203         if ((Nt + Ni) > 32) {
   1204             // we will overflow, reduce the precision of Ni to 8 bits
   1205             // (Note Nt cannot be more than 10 bits which happens with
   1206             // 565 textures and GGL_LINEAR)
   1207             shift += Ni-8;
   1208             Ni = 8;
   1209         }
   1210 
   1211         // modulate by the component with the lowest precision
   1212         if (Nt >= Ni) {
   1213             if (shift) {
   1214                 // XXX: we should be able to avoid this shift
   1215                 // when shift==16 && Nt<16 && Ni<16, in which
   1216                 // we could use SMULBT below.
   1217                 MOV(AL, 0, dest.reg, reg_imm(inReg, LSR, shift));
   1218                 inReg = dest.reg;
   1219                 shift = 0;
   1220             }
   1221             // operation:           (Cf*Ct)/((1<<Ni)-1)
   1222             // approximated with:   Cf*(Ct + Ct>>(Ni-1))>>Ni
   1223             // this operation doesn't change texel's size
   1224             ADD(AL, 0, dest.reg, inReg, reg_imm(inReg, LSR, Ni-1));
   1225             if (Nt<16 && Ni<16) SMULBB(AL, dest.reg, texel.reg, dest.reg);
   1226             else                MUL(AL, 0, dest.reg, texel.reg, dest.reg);
   1227             dest.l = Ni;
   1228             dest.h = Nt + Ni;
   1229         } else {
   1230             if (shift && (shift != 16)) {
   1231                 // if shift==16, we can use 16-bits mul instructions later
   1232                 MOV(AL, 0, dest.reg, reg_imm(inReg, LSR, shift));
   1233                 inReg = dest.reg;
   1234                 shift = 0;
   1235             }
   1236             // operation:           (Cf*Ct)/((1<<Nt)-1)
   1237             // approximated with:   Ct*(Cf + Cf>>(Nt-1))>>Nt
   1238             // this operation doesn't change incoming's size
   1239             Scratch scratches(registerFile());
   1240             int t = (texel.flags & CORRUPTIBLE) ? texel.reg : dest.reg;
   1241             if (t == inReg)
   1242                 t = scratches.obtain();
   1243             ADD(AL, 0, t, texel.reg, reg_imm(texel.reg, LSR, Nt-1));
   1244             if (Nt<16 && Ni<16) {
   1245                 if (shift==16)  SMULBT(AL, dest.reg, t, inReg);
   1246                 else            SMULBB(AL, dest.reg, t, inReg);
   1247             } else              MUL(AL, 0, dest.reg, t, inReg);
   1248             dest.l = Nt;
   1249             dest.h = Nt + Ni;
   1250         }
   1251 
   1252         // low bits are not valid
   1253         dest.flags |= CLEAR_LO;
   1254 
   1255         // no need to keep more than 8 bits/component
   1256         if (dest.size() > 8)
   1257             dest.l = dest.h-8;
   1258     }
   1259 }
   1260 
   1261 void GGLAssembler::decal(
   1262         component_t& dest,
   1263         const component_t& incoming,
   1264         const pixel_t& incomingTexel, int component)
   1265 {
   1266     // RGBA:
   1267     // Cv = Cf*(1 - At) + Ct*At = Cf + (Ct - Cf)*At
   1268     // Av = Af
   1269     Scratch locals(registerFile());
   1270     integer_t texel(locals.obtain(), 32, CORRUPTIBLE);
   1271     integer_t factor(locals.obtain(), 32, CORRUPTIBLE);
   1272     extract(texel, incomingTexel, component);
   1273     extract(factor, incomingTexel, GGLFormat::ALPHA);
   1274 
   1275     // no need to keep more than 8-bits for decal
   1276     int Ni = incoming.size();
   1277     int shift = incoming.l;
   1278     if (Ni > 8) {
   1279         shift += Ni-8;
   1280         Ni = 8;
   1281     }
   1282     integer_t incomingNorm(incoming.reg, Ni, incoming.flags);
   1283     if (shift) {
   1284         MOV(AL, 0, dest.reg, reg_imm(incomingNorm.reg, LSR, shift));
   1285         incomingNorm.reg = dest.reg;
   1286         incomingNorm.flags |= CORRUPTIBLE;
   1287     }
   1288     ADD(AL, 0, factor.reg, factor.reg, reg_imm(factor.reg, LSR, factor.s-1));
   1289     build_blendOneMinusFF(dest, factor, incomingNorm, texel);
   1290 }
   1291 
   1292 void GGLAssembler::blend(
   1293         component_t& dest,
   1294         const component_t& incoming,
   1295         const pixel_t& incomingTexel, int component, int tmu)
   1296 {
   1297     // RGBA:
   1298     // Cv = (1 - Ct)*Cf + Ct*Cc = Cf + (Cc - Cf)*Ct
   1299     // Av = At*Af
   1300 
   1301     if (component == GGLFormat::ALPHA) {
   1302         modulate(dest, incoming, incomingTexel, component);
   1303         return;
   1304     }
   1305 
   1306     Scratch locals(registerFile());
   1307     integer_t color(locals.obtain(), 8, CORRUPTIBLE);
   1308     integer_t factor(locals.obtain(), 32, CORRUPTIBLE);
   1309     LDRB(AL, color.reg, mBuilderContext.Rctx,
   1310             immed12_pre(GGL_OFFSETOF(state.texture[tmu].env_color[component])));
   1311     extract(factor, incomingTexel, component);
   1312 
   1313     // no need to keep more than 8-bits for blend
   1314     int Ni = incoming.size();
   1315     int shift = incoming.l;
   1316     if (Ni > 8) {
   1317         shift += Ni-8;
   1318         Ni = 8;
   1319     }
   1320     integer_t incomingNorm(incoming.reg, Ni, incoming.flags);
   1321     if (shift) {
   1322         MOV(AL, 0, dest.reg, reg_imm(incomingNorm.reg, LSR, shift));
   1323         incomingNorm.reg = dest.reg;
   1324         incomingNorm.flags |= CORRUPTIBLE;
   1325     }
   1326     ADD(AL, 0, factor.reg, factor.reg, reg_imm(factor.reg, LSR, factor.s-1));
   1327     build_blendOneMinusFF(dest, factor, incomingNorm, color);
   1328 }
   1329 
   1330 void GGLAssembler::add(
   1331         component_t& dest,
   1332         const component_t& incoming,
   1333         const pixel_t& incomingTexel, int component)
   1334 {
   1335     // RGBA:
   1336     // Cv = Cf + Ct;
   1337     Scratch locals(registerFile());
   1338 
   1339     component_t incomingTemp(incoming);
   1340 
   1341     // use "dest" as a temporary for extracting the texel, unless "dest"
   1342     // overlaps "incoming".
   1343     integer_t texel(dest.reg, 32, CORRUPTIBLE);
   1344     if (dest.reg == incomingTemp.reg)
   1345         texel.reg = locals.obtain();
   1346     extract(texel, incomingTexel, component);
   1347 
   1348     if (texel.s < incomingTemp.size()) {
   1349         expand(texel, texel, incomingTemp.size());
   1350     } else if (texel.s > incomingTemp.size()) {
   1351         if (incomingTemp.flags & CORRUPTIBLE) {
   1352             expand(incomingTemp, incomingTemp, texel.s);
   1353         } else {
   1354             incomingTemp.reg = locals.obtain();
   1355             expand(incomingTemp, incoming, texel.s);
   1356         }
   1357     }
   1358 
   1359     if (incomingTemp.l) {
   1360         ADD(AL, 0, dest.reg, texel.reg,
   1361                 reg_imm(incomingTemp.reg, LSR, incomingTemp.l));
   1362     } else {
   1363         ADD(AL, 0, dest.reg, texel.reg, incomingTemp.reg);
   1364     }
   1365     dest.l = 0;
   1366     dest.h = texel.size();
   1367     component_sat(dest);
   1368 }
   1369 
   1370 // ----------------------------------------------------------------------------
   1371 
   1372 }; // namespace android
   1373 
   1374