Home | History | Annotate | Download | only in cpu_ref
      1 /*
      2  * Copyright (C) 2012 The Android Open Source Project
      3  *
      4  * Licensed under the Apache License, Version 2.0 (the "License");
      5  * you may not use this file except in compliance with the License.
      6  * You may obtain a copy of the License at
      7  *
      8  *      http://www.apache.org/licenses/LICENSE-2.0
      9  *
     10  * Unless required by applicable law or agreed to in writing, software
     11  * distributed under the License is distributed on an "AS IS" BASIS,
     12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
     13  * See the License for the specific language governing permissions and
     14  * limitations under the License.
     15  */
     16 
     17 
     18 #include "rsCpuIntrinsic.h"
     19 #include "rsCpuIntrinsicInlines.h"
     20 
     21 using namespace android;
     22 using namespace android::renderscript;
     23 
     24 namespace android {
     25 namespace renderscript {
     26 
     27 
     28 class RsdCpuScriptIntrinsicConvolve5x5 : public RsdCpuScriptIntrinsic {
     29 public:
     30     virtual void populateScript(Script *);
     31     virtual void invokeFreeChildren();
     32 
     33     virtual void setGlobalVar(uint32_t slot, const void *data, size_t dataLength);
     34     virtual void setGlobalObj(uint32_t slot, ObjectBase *data);
     35 
     36     virtual ~RsdCpuScriptIntrinsicConvolve5x5();
     37     RsdCpuScriptIntrinsicConvolve5x5(RsdCpuReferenceImpl *ctx, const Script *s, const Element *e);
     38 
     39 protected:
     40     float mFp[28];
     41     short mIp[28];
     42     ObjectBaseRef<Allocation> alloc;
     43 
     44 
     45     static void kernelU1(const RsForEachStubParamStruct *p,
     46                          uint32_t xstart, uint32_t xend,
     47                          uint32_t instep, uint32_t outstep);
     48     static void kernelU2(const RsForEachStubParamStruct *p,
     49                          uint32_t xstart, uint32_t xend,
     50                          uint32_t instep, uint32_t outstep);
     51     static void kernelU4(const RsForEachStubParamStruct *p,
     52                          uint32_t xstart, uint32_t xend,
     53                          uint32_t instep, uint32_t outstep);
     54     static void kernelF1(const RsForEachStubParamStruct *p,
     55                          uint32_t xstart, uint32_t xend,
     56                          uint32_t instep, uint32_t outstep);
     57     static void kernelF2(const RsForEachStubParamStruct *p,
     58                          uint32_t xstart, uint32_t xend,
     59                          uint32_t instep, uint32_t outstep);
     60     static void kernelF4(const RsForEachStubParamStruct *p,
     61                          uint32_t xstart, uint32_t xend,
     62                          uint32_t instep, uint32_t outstep);
     63 
     64 
     65 };
     66 
     67 }
     68 }
     69 
     70 void RsdCpuScriptIntrinsicConvolve5x5::setGlobalObj(uint32_t slot, ObjectBase *data) {
     71     rsAssert(slot == 1);
     72     alloc.set(static_cast<Allocation *>(data));
     73 }
     74 
     75 void RsdCpuScriptIntrinsicConvolve5x5::setGlobalVar(uint32_t slot,
     76                                                     const void *data, size_t dataLength) {
     77     rsAssert(slot == 0);
     78     memcpy (&mFp, data, dataLength);
     79     for(int ct=0; ct < 25; ct++) {
     80         if (mFp[ct] >= 0) {
     81             mIp[ct] = (short)(mFp[ct] * 256.f + 0.5f);
     82         } else {
     83             mIp[ct] = (short)(mFp[ct] * 256.f - 0.5f);
     84         }
     85     }
     86 }
     87 
     88 
     89 static void OneU4(const RsForEachStubParamStruct *p, uint32_t x, uchar4 *out,
     90                   const uchar4 *py0, const uchar4 *py1, const uchar4 *py2, const uchar4 *py3, const uchar4 *py4,
     91                   const float* coeff) {
     92 
     93     uint32_t x0 = rsMax((int32_t)x-2, 0);
     94     uint32_t x1 = rsMax((int32_t)x-1, 0);
     95     uint32_t x2 = x;
     96     uint32_t x3 = rsMin((int32_t)x+1, (int32_t)(p->dimX-1));
     97     uint32_t x4 = rsMin((int32_t)x+2, (int32_t)(p->dimX-1));
     98 
     99     float4 px = convert_float4(py0[x0]) * coeff[0] +
    100                 convert_float4(py0[x1]) * coeff[1] +
    101                 convert_float4(py0[x2]) * coeff[2] +
    102                 convert_float4(py0[x3]) * coeff[3] +
    103                 convert_float4(py0[x4]) * coeff[4] +
    104 
    105                 convert_float4(py1[x0]) * coeff[5] +
    106                 convert_float4(py1[x1]) * coeff[6] +
    107                 convert_float4(py1[x2]) * coeff[7] +
    108                 convert_float4(py1[x3]) * coeff[8] +
    109                 convert_float4(py1[x4]) * coeff[9] +
    110 
    111                 convert_float4(py2[x0]) * coeff[10] +
    112                 convert_float4(py2[x1]) * coeff[11] +
    113                 convert_float4(py2[x2]) * coeff[12] +
    114                 convert_float4(py2[x3]) * coeff[13] +
    115                 convert_float4(py2[x4]) * coeff[14] +
    116 
    117                 convert_float4(py3[x0]) * coeff[15] +
    118                 convert_float4(py3[x1]) * coeff[16] +
    119                 convert_float4(py3[x2]) * coeff[17] +
    120                 convert_float4(py3[x3]) * coeff[18] +
    121                 convert_float4(py3[x4]) * coeff[19] +
    122 
    123                 convert_float4(py4[x0]) * coeff[20] +
    124                 convert_float4(py4[x1]) * coeff[21] +
    125                 convert_float4(py4[x2]) * coeff[22] +
    126                 convert_float4(py4[x3]) * coeff[23] +
    127                 convert_float4(py4[x4]) * coeff[24];
    128     px = clamp(px + 0.5f, 0.f, 255.f);
    129     *out = convert_uchar4(px);
    130 }
    131 
    132 static void OneU2(const RsForEachStubParamStruct *p, uint32_t x, uchar2 *out,
    133                   const uchar2 *py0, const uchar2 *py1, const uchar2 *py2, const uchar2 *py3, const uchar2 *py4,
    134                   const float* coeff) {
    135 
    136     uint32_t x0 = rsMax((int32_t)x-2, 0);
    137     uint32_t x1 = rsMax((int32_t)x-1, 0);
    138     uint32_t x2 = x;
    139     uint32_t x3 = rsMin((int32_t)x+1, (int32_t)(p->dimX-1));
    140     uint32_t x4 = rsMin((int32_t)x+2, (int32_t)(p->dimX-1));
    141 
    142     float2 px = convert_float2(py0[x0]) * coeff[0] +
    143                 convert_float2(py0[x1]) * coeff[1] +
    144                 convert_float2(py0[x2]) * coeff[2] +
    145                 convert_float2(py0[x3]) * coeff[3] +
    146                 convert_float2(py0[x4]) * coeff[4] +
    147 
    148                 convert_float2(py1[x0]) * coeff[5] +
    149                 convert_float2(py1[x1]) * coeff[6] +
    150                 convert_float2(py1[x2]) * coeff[7] +
    151                 convert_float2(py1[x3]) * coeff[8] +
    152                 convert_float2(py1[x4]) * coeff[9] +
    153 
    154                 convert_float2(py2[x0]) * coeff[10] +
    155                 convert_float2(py2[x1]) * coeff[11] +
    156                 convert_float2(py2[x2]) * coeff[12] +
    157                 convert_float2(py2[x3]) * coeff[13] +
    158                 convert_float2(py2[x4]) * coeff[14] +
    159 
    160                 convert_float2(py3[x0]) * coeff[15] +
    161                 convert_float2(py3[x1]) * coeff[16] +
    162                 convert_float2(py3[x2]) * coeff[17] +
    163                 convert_float2(py3[x3]) * coeff[18] +
    164                 convert_float2(py3[x4]) * coeff[19] +
    165 
    166                 convert_float2(py4[x0]) * coeff[20] +
    167                 convert_float2(py4[x1]) * coeff[21] +
    168                 convert_float2(py4[x2]) * coeff[22] +
    169                 convert_float2(py4[x3]) * coeff[23] +
    170                 convert_float2(py4[x4]) * coeff[24];
    171     px = clamp(px + 0.5f, 0.f, 255.f);
    172     *out = convert_uchar2(px);
    173 }
    174 
    175 static void OneU1(const RsForEachStubParamStruct *p, uint32_t x, uchar *out,
    176                   const uchar *py0, const uchar *py1, const uchar *py2, const uchar *py3, const uchar *py4,
    177                   const float* coeff) {
    178 
    179     uint32_t x0 = rsMax((int32_t)x-2, 0);
    180     uint32_t x1 = rsMax((int32_t)x-1, 0);
    181     uint32_t x2 = x;
    182     uint32_t x3 = rsMin((int32_t)x+1, (int32_t)(p->dimX-1));
    183     uint32_t x4 = rsMin((int32_t)x+2, (int32_t)(p->dimX-1));
    184 
    185     float px = (float)(py0[x0]) * coeff[0] +
    186                (float)(py0[x1]) * coeff[1] +
    187                (float)(py0[x2]) * coeff[2] +
    188                (float)(py0[x3]) * coeff[3] +
    189                (float)(py0[x4]) * coeff[4] +
    190 
    191                (float)(py1[x0]) * coeff[5] +
    192                (float)(py1[x1]) * coeff[6] +
    193                (float)(py1[x2]) * coeff[7] +
    194                (float)(py1[x3]) * coeff[8] +
    195                (float)(py1[x4]) * coeff[9] +
    196 
    197                (float)(py2[x0]) * coeff[10] +
    198                (float)(py2[x1]) * coeff[11] +
    199                (float)(py2[x2]) * coeff[12] +
    200                (float)(py2[x3]) * coeff[13] +
    201                (float)(py2[x4]) * coeff[14] +
    202 
    203                (float)(py3[x0]) * coeff[15] +
    204                (float)(py3[x1]) * coeff[16] +
    205                (float)(py3[x2]) * coeff[17] +
    206                (float)(py3[x3]) * coeff[18] +
    207                (float)(py3[x4]) * coeff[19] +
    208 
    209                (float)(py4[x0]) * coeff[20] +
    210                (float)(py4[x1]) * coeff[21] +
    211                (float)(py4[x2]) * coeff[22] +
    212                (float)(py4[x3]) * coeff[23] +
    213                (float)(py4[x4]) * coeff[24];
    214     px = clamp(px + 0.5f, 0.f, 255.f);
    215     *out = px;
    216 }
    217 
    218 static void OneF4(const RsForEachStubParamStruct *p, uint32_t x, float4 *out,
    219                   const float4 *py0, const float4 *py1, const float4 *py2, const float4 *py3, const float4 *py4,
    220                   const float* coeff) {
    221 
    222     uint32_t x0 = rsMax((int32_t)x-2, 0);
    223     uint32_t x1 = rsMax((int32_t)x-1, 0);
    224     uint32_t x2 = x;
    225     uint32_t x3 = rsMin((int32_t)x+1, (int32_t)(p->dimX-1));
    226     uint32_t x4 = rsMin((int32_t)x+2, (int32_t)(p->dimX-1));
    227 
    228     float4 px = py0[x0] * coeff[0] +
    229                 py0[x1] * coeff[1] +
    230                 py0[x2] * coeff[2] +
    231                 py0[x3] * coeff[3] +
    232                 py0[x4] * coeff[4] +
    233 
    234                 py1[x0] * coeff[5] +
    235                 py1[x1] * coeff[6] +
    236                 py1[x2] * coeff[7] +
    237                 py1[x3] * coeff[8] +
    238                 py1[x4] * coeff[9] +
    239 
    240                 py2[x0] * coeff[10] +
    241                 py2[x1] * coeff[11] +
    242                 py2[x2] * coeff[12] +
    243                 py2[x3] * coeff[13] +
    244                 py2[x4] * coeff[14] +
    245 
    246                 py3[x0] * coeff[15] +
    247                 py3[x1] * coeff[16] +
    248                 py3[x2] * coeff[17] +
    249                 py3[x3] * coeff[18] +
    250                 py3[x4] * coeff[19] +
    251 
    252                 py4[x0] * coeff[20] +
    253                 py4[x1] * coeff[21] +
    254                 py4[x2] * coeff[22] +
    255                 py4[x3] * coeff[23] +
    256                 py4[x4] * coeff[24];
    257     *out = px;
    258 }
    259 
    260 static void OneF2(const RsForEachStubParamStruct *p, uint32_t x, float2 *out,
    261                   const float2 *py0, const float2 *py1, const float2 *py2, const float2 *py3, const float2 *py4,
    262                   const float* coeff) {
    263 
    264     uint32_t x0 = rsMax((int32_t)x-2, 0);
    265     uint32_t x1 = rsMax((int32_t)x-1, 0);
    266     uint32_t x2 = x;
    267     uint32_t x3 = rsMin((int32_t)x+1, (int32_t)(p->dimX-1));
    268     uint32_t x4 = rsMin((int32_t)x+2, (int32_t)(p->dimX-1));
    269 
    270     float2 px = py0[x0] * coeff[0] +
    271                 py0[x1] * coeff[1] +
    272                 py0[x2] * coeff[2] +
    273                 py0[x3] * coeff[3] +
    274                 py0[x4] * coeff[4] +
    275 
    276                 py1[x0] * coeff[5] +
    277                 py1[x1] * coeff[6] +
    278                 py1[x2] * coeff[7] +
    279                 py1[x3] * coeff[8] +
    280                 py1[x4] * coeff[9] +
    281 
    282                 py2[x0] * coeff[10] +
    283                 py2[x1] * coeff[11] +
    284                 py2[x2] * coeff[12] +
    285                 py2[x3] * coeff[13] +
    286                 py2[x4] * coeff[14] +
    287 
    288                 py3[x0] * coeff[15] +
    289                 py3[x1] * coeff[16] +
    290                 py3[x2] * coeff[17] +
    291                 py3[x3] * coeff[18] +
    292                 py3[x4] * coeff[19] +
    293 
    294                 py4[x0] * coeff[20] +
    295                 py4[x1] * coeff[21] +
    296                 py4[x2] * coeff[22] +
    297                 py4[x3] * coeff[23] +
    298                 py4[x4] * coeff[24];
    299     *out = px;
    300 }
    301 
    302 static void OneF1(const RsForEachStubParamStruct *p, uint32_t x, float *out,
    303                   const float *py0, const float *py1, const float *py2, const float *py3, const float *py4,
    304                   const float* coeff) {
    305 
    306     uint32_t x0 = rsMax((int32_t)x-2, 0);
    307     uint32_t x1 = rsMax((int32_t)x-1, 0);
    308     uint32_t x2 = x;
    309     uint32_t x3 = rsMin((int32_t)x+1, (int32_t)(p->dimX-1));
    310     uint32_t x4 = rsMin((int32_t)x+2, (int32_t)(p->dimX-1));
    311 
    312     float px = py0[x0] * coeff[0] +
    313                py0[x1] * coeff[1] +
    314                py0[x2] * coeff[2] +
    315                py0[x3] * coeff[3] +
    316                py0[x4] * coeff[4] +
    317 
    318                py1[x0] * coeff[5] +
    319                py1[x1] * coeff[6] +
    320                py1[x2] * coeff[7] +
    321                py1[x3] * coeff[8] +
    322                py1[x4] * coeff[9] +
    323 
    324                py2[x0] * coeff[10] +
    325                py2[x1] * coeff[11] +
    326                py2[x2] * coeff[12] +
    327                py2[x3] * coeff[13] +
    328                py2[x4] * coeff[14] +
    329 
    330                py3[x0] * coeff[15] +
    331                py3[x1] * coeff[16] +
    332                py3[x2] * coeff[17] +
    333                py3[x3] * coeff[18] +
    334                py3[x4] * coeff[19] +
    335 
    336                py4[x0] * coeff[20] +
    337                py4[x1] * coeff[21] +
    338                py4[x2] * coeff[22] +
    339                py4[x3] * coeff[23] +
    340                py4[x4] * coeff[24];
    341     *out = px;
    342 }
    343 
    344 
    345 extern "C" void rsdIntrinsicConvolve5x5_K(void *dst, const void *y0, const void *y1,
    346                                           const void *y2, const void *y3, const void *y4,
    347                                           const short *coef, uint32_t count);
    348 
    349 void RsdCpuScriptIntrinsicConvolve5x5::kernelU4(const RsForEachStubParamStruct *p,
    350                                                 uint32_t xstart, uint32_t xend,
    351                                                 uint32_t instep, uint32_t outstep) {
    352     RsdCpuScriptIntrinsicConvolve5x5 *cp = (RsdCpuScriptIntrinsicConvolve5x5 *)p->usr;
    353     if (!cp->alloc.get()) {
    354         ALOGE("Convolve5x5 executed without input, skipping");
    355         return;
    356     }
    357     const uchar *pin = (const uchar *)cp->alloc->mHal.drvState.lod[0].mallocPtr;
    358     const size_t stride = cp->alloc->mHal.drvState.lod[0].stride;
    359 
    360     uint32_t y0 = rsMax((int32_t)p->y-2, 0);
    361     uint32_t y1 = rsMax((int32_t)p->y-1, 0);
    362     uint32_t y2 = p->y;
    363     uint32_t y3 = rsMin((int32_t)p->y+1, (int32_t)(p->dimY-1));
    364     uint32_t y4 = rsMin((int32_t)p->y+2, (int32_t)(p->dimY-1));
    365 
    366     const uchar4 *py0 = (const uchar4 *)(pin + stride * y0);
    367     const uchar4 *py1 = (const uchar4 *)(pin + stride * y1);
    368     const uchar4 *py2 = (const uchar4 *)(pin + stride * y2);
    369     const uchar4 *py3 = (const uchar4 *)(pin + stride * y3);
    370     const uchar4 *py4 = (const uchar4 *)(pin + stride * y4);
    371 
    372     uchar4 *out = (uchar4 *)p->out;
    373     uint32_t x1 = xstart;
    374     uint32_t x2 = xend;
    375 
    376     while((x1 < x2) && (x1 < 2)) {
    377         OneU4(p, x1, out, py0, py1, py2, py3, py4, cp->mFp);
    378         out++;
    379         x1++;
    380     }
    381 #if defined(ARCH_X86_HAVE_SSSE3)
    382     // for x86 SIMD, require minimum of 7 elements (4 for SIMD,
    383     // 3 for end boundary where x may hit the end boundary)
    384     if (gArchUseSIMD &&((x1 + 6) < x2)) {
    385         // subtract 3 for end boundary
    386         uint32_t len = (x2 - x1 - 3) >> 2;
    387         rsdIntrinsicConvolve5x5_K(out, py0 + x1 - 2, py1 + x1 - 2, py2 + x1 - 2, py3 + x1 - 2, py4 + x1 - 2, cp->mIp, len);
    388         out += len << 2;
    389         x1 += len << 2;
    390     }
    391 #endif
    392 
    393 #if defined(ARCH_ARM_USE_INTRINSICS)
    394     if(gArchUseSIMD && ((x1 + 3) < x2)) {
    395         uint32_t len = (x2 - x1 - 3) >> 1;
    396         rsdIntrinsicConvolve5x5_K(out, py0 + x1 - 2, py1 + x1 - 2, py2 + x1 - 2, py3 + x1 - 2, py4 + x1 - 2, cp->mIp, len);
    397         out += len << 1;
    398         x1 += len << 1;
    399     }
    400 #endif
    401 
    402     while(x1 < x2) {
    403         OneU4(p, x1, out, py0, py1, py2, py3, py4, cp->mFp);
    404         out++;
    405         x1++;
    406     }
    407 }
    408 
    409 void RsdCpuScriptIntrinsicConvolve5x5::kernelU2(const RsForEachStubParamStruct *p,
    410                                                 uint32_t xstart, uint32_t xend,
    411                                                 uint32_t instep, uint32_t outstep) {
    412     RsdCpuScriptIntrinsicConvolve5x5 *cp = (RsdCpuScriptIntrinsicConvolve5x5 *)p->usr;
    413     if (!cp->alloc.get()) {
    414         ALOGE("Convolve5x5 executed without input, skipping");
    415         return;
    416     }
    417     const uchar *pin = (const uchar *)cp->alloc->mHal.drvState.lod[0].mallocPtr;
    418     const size_t stride = cp->alloc->mHal.drvState.lod[0].stride;
    419 
    420     uint32_t y0 = rsMax((int32_t)p->y-2, 0);
    421     uint32_t y1 = rsMax((int32_t)p->y-1, 0);
    422     uint32_t y2 = p->y;
    423     uint32_t y3 = rsMin((int32_t)p->y+1, (int32_t)(p->dimY-1));
    424     uint32_t y4 = rsMin((int32_t)p->y+2, (int32_t)(p->dimY-1));
    425 
    426     const uchar2 *py0 = (const uchar2 *)(pin + stride * y0);
    427     const uchar2 *py1 = (const uchar2 *)(pin + stride * y1);
    428     const uchar2 *py2 = (const uchar2 *)(pin + stride * y2);
    429     const uchar2 *py3 = (const uchar2 *)(pin + stride * y3);
    430     const uchar2 *py4 = (const uchar2 *)(pin + stride * y4);
    431 
    432     uchar2 *out = (uchar2 *)p->out;
    433     uint32_t x1 = xstart;
    434     uint32_t x2 = xend;
    435 
    436     while((x1 < x2) && (x1 < 2)) {
    437         OneU2(p, x1, out, py0, py1, py2, py3, py4, cp->mFp);
    438         out++;
    439         x1++;
    440     }
    441 
    442 #if 0//defined(ARCH_ARM_HAVE_NEON)
    443     if((x1 + 3) < x2) {
    444         uint32_t len = (x2 - x1 - 3) >> 1;
    445         rsdIntrinsicConvolve5x5_K(out, py0, py1, py2, py3, py4, cp->ip, len);
    446         out += len << 1;
    447         x1 += len << 1;
    448     }
    449 #endif
    450 
    451     while(x1 < x2) {
    452         OneU2(p, x1, out, py0, py1, py2, py3, py4, cp->mFp);
    453         out++;
    454         x1++;
    455     }
    456 }
    457 
    458 void RsdCpuScriptIntrinsicConvolve5x5::kernelU1(const RsForEachStubParamStruct *p,
    459                                                 uint32_t xstart, uint32_t xend,
    460                                                 uint32_t instep, uint32_t outstep) {
    461     RsdCpuScriptIntrinsicConvolve5x5 *cp = (RsdCpuScriptIntrinsicConvolve5x5 *)p->usr;
    462     if (!cp->alloc.get()) {
    463         ALOGE("Convolve5x5 executed without input, skipping");
    464         return;
    465     }
    466     const uchar *pin = (const uchar *)cp->alloc->mHal.drvState.lod[0].mallocPtr;
    467     const size_t stride = cp->alloc->mHal.drvState.lod[0].stride;
    468 
    469     uint32_t y0 = rsMax((int32_t)p->y-2, 0);
    470     uint32_t y1 = rsMax((int32_t)p->y-1, 0);
    471     uint32_t y2 = p->y;
    472     uint32_t y3 = rsMin((int32_t)p->y+1, (int32_t)(p->dimY-1));
    473     uint32_t y4 = rsMin((int32_t)p->y+2, (int32_t)(p->dimY-1));
    474 
    475     const uchar *py0 = (const uchar *)(pin + stride * y0);
    476     const uchar *py1 = (const uchar *)(pin + stride * y1);
    477     const uchar *py2 = (const uchar *)(pin + stride * y2);
    478     const uchar *py3 = (const uchar *)(pin + stride * y3);
    479     const uchar *py4 = (const uchar *)(pin + stride * y4);
    480 
    481     uchar *out = (uchar *)p->out;
    482     uint32_t x1 = xstart;
    483     uint32_t x2 = xend;
    484 
    485     while((x1 < x2) && (x1 < 2)) {
    486         OneU1(p, x1, out, py0, py1, py2, py3, py4, cp->mFp);
    487         out++;
    488         x1++;
    489     }
    490 
    491 #if 0//defined(ARCH_ARM_HAVE_NEON)
    492     if((x1 + 3) < x2) {
    493         uint32_t len = (x2 - x1 - 3) >> 1;
    494         rsdIntrinsicConvolve5x5_K(out, py0, py1, py2, py3, py4, cp->ip, len);
    495         out += len << 1;
    496         x1 += len << 1;
    497     }
    498 #endif
    499 
    500     while(x1 < x2) {
    501         OneU1(p, x1, out, py0, py1, py2, py3, py4, cp->mFp);
    502         out++;
    503         x1++;
    504     }
    505 }
    506 
    507 void RsdCpuScriptIntrinsicConvolve5x5::kernelF4(const RsForEachStubParamStruct *p,
    508                                                 uint32_t xstart, uint32_t xend,
    509                                                 uint32_t instep, uint32_t outstep) {
    510     RsdCpuScriptIntrinsicConvolve5x5 *cp = (RsdCpuScriptIntrinsicConvolve5x5 *)p->usr;
    511     if (!cp->alloc.get()) {
    512         ALOGE("Convolve5x5 executed without input, skipping");
    513         return;
    514     }
    515     const uchar *pin = (const uchar *)cp->alloc->mHal.drvState.lod[0].mallocPtr;
    516     const size_t stride = cp->alloc->mHal.drvState.lod[0].stride;
    517 
    518     uint32_t y0 = rsMax((int32_t)p->y-2, 0);
    519     uint32_t y1 = rsMax((int32_t)p->y-1, 0);
    520     uint32_t y2 = p->y;
    521     uint32_t y3 = rsMin((int32_t)p->y+1, (int32_t)(p->dimY-1));
    522     uint32_t y4 = rsMin((int32_t)p->y+2, (int32_t)(p->dimY-1));
    523 
    524     const float4 *py0 = (const float4 *)(pin + stride * y0);
    525     const float4 *py1 = (const float4 *)(pin + stride * y1);
    526     const float4 *py2 = (const float4 *)(pin + stride * y2);
    527     const float4 *py3 = (const float4 *)(pin + stride * y3);
    528     const float4 *py4 = (const float4 *)(pin + stride * y4);
    529 
    530     float4 *out = (float4 *)p->out;
    531     uint32_t x1 = xstart;
    532     uint32_t x2 = xend;
    533 
    534     while((x1 < x2) && (x1 < 2)) {
    535         OneF4(p, x1, out, py0, py1, py2, py3, py4, cp->mFp);
    536         out++;
    537         x1++;
    538     }
    539 
    540 #if 0//defined(ARCH_ARM_HAVE_NEON)
    541     if((x1 + 3) < x2) {
    542         uint32_t len = (x2 - x1 - 3) >> 1;
    543         rsdIntrinsicConvolve5x5_K(out, py0, py1, py2, py3, py4, cp->ip, len);
    544         out += len << 1;
    545         x1 += len << 1;
    546     }
    547 #endif
    548 
    549     while(x1 < x2) {
    550         OneF4(p, x1, out, py0, py1, py2, py3, py4, cp->mFp);
    551         out++;
    552         x1++;
    553     }
    554 }
    555 
    556 void RsdCpuScriptIntrinsicConvolve5x5::kernelF2(const RsForEachStubParamStruct *p,
    557                                                 uint32_t xstart, uint32_t xend,
    558                                                 uint32_t instep, uint32_t outstep) {
    559     RsdCpuScriptIntrinsicConvolve5x5 *cp = (RsdCpuScriptIntrinsicConvolve5x5 *)p->usr;
    560     if (!cp->alloc.get()) {
    561         ALOGE("Convolve5x5 executed without input, skipping");
    562         return;
    563     }
    564     const uchar *pin = (const uchar *)cp->alloc->mHal.drvState.lod[0].mallocPtr;
    565     const size_t stride = cp->alloc->mHal.drvState.lod[0].stride;
    566 
    567     uint32_t y0 = rsMax((int32_t)p->y-2, 0);
    568     uint32_t y1 = rsMax((int32_t)p->y-1, 0);
    569     uint32_t y2 = p->y;
    570     uint32_t y3 = rsMin((int32_t)p->y+1, (int32_t)(p->dimY-1));
    571     uint32_t y4 = rsMin((int32_t)p->y+2, (int32_t)(p->dimY-1));
    572 
    573     const float2 *py0 = (const float2 *)(pin + stride * y0);
    574     const float2 *py1 = (const float2 *)(pin + stride * y1);
    575     const float2 *py2 = (const float2 *)(pin + stride * y2);
    576     const float2 *py3 = (const float2 *)(pin + stride * y3);
    577     const float2 *py4 = (const float2 *)(pin + stride * y4);
    578 
    579     float2 *out = (float2 *)p->out;
    580     uint32_t x1 = xstart;
    581     uint32_t x2 = xend;
    582 
    583     while((x1 < x2) && (x1 < 2)) {
    584         OneF2(p, x1, out, py0, py1, py2, py3, py4, cp->mFp);
    585         out++;
    586         x1++;
    587     }
    588 
    589 #if 0//defined(ARCH_ARM_HAVE_NEON)
    590     if((x1 + 3) < x2) {
    591         uint32_t len = (x2 - x1 - 3) >> 1;
    592         rsdIntrinsicConvolve5x5_K(out, py0, py1, py2, py3, py4, cp->ip, len);
    593         out += len << 1;
    594         x1 += len << 1;
    595     }
    596 #endif
    597 
    598     while(x1 < x2) {
    599         OneF2(p, x1, out, py0, py1, py2, py3, py4, cp->mFp);
    600         out++;
    601         x1++;
    602     }
    603 }
    604 
    605 void RsdCpuScriptIntrinsicConvolve5x5::kernelF1(const RsForEachStubParamStruct *p,
    606                                                 uint32_t xstart, uint32_t xend,
    607                                                 uint32_t instep, uint32_t outstep) {
    608     RsdCpuScriptIntrinsicConvolve5x5 *cp = (RsdCpuScriptIntrinsicConvolve5x5 *)p->usr;
    609     if (!cp->alloc.get()) {
    610         ALOGE("Convolve5x5 executed without input, skipping");
    611         return;
    612     }
    613     const uchar *pin = (const uchar *)cp->alloc->mHal.drvState.lod[0].mallocPtr;
    614     const size_t stride = cp->alloc->mHal.drvState.lod[0].stride;
    615 
    616     uint32_t y0 = rsMax((int32_t)p->y-2, 0);
    617     uint32_t y1 = rsMax((int32_t)p->y-1, 0);
    618     uint32_t y2 = p->y;
    619     uint32_t y3 = rsMin((int32_t)p->y+1, (int32_t)(p->dimY-1));
    620     uint32_t y4 = rsMin((int32_t)p->y+2, (int32_t)(p->dimY-1));
    621 
    622     const float *py0 = (const float *)(pin + stride * y0);
    623     const float *py1 = (const float *)(pin + stride * y1);
    624     const float *py2 = (const float *)(pin + stride * y2);
    625     const float *py3 = (const float *)(pin + stride * y3);
    626     const float *py4 = (const float *)(pin + stride * y4);
    627 
    628     float *out = (float *)p->out;
    629     uint32_t x1 = xstart;
    630     uint32_t x2 = xend;
    631 
    632     while((x1 < x2) && (x1 < 2)) {
    633         OneF1(p, x1, out, py0, py1, py2, py3, py4, cp->mFp);
    634         out++;
    635         x1++;
    636     }
    637 
    638 #if 0//defined(ARCH_ARM_HAVE_NEON)
    639     if((x1 + 3) < x2) {
    640         uint32_t len = (x2 - x1 - 3) >> 1;
    641         rsdIntrinsicConvolve5x5_K(out, py0, py1, py2, py3, py4, cp->ip, len);
    642         out += len << 1;
    643         x1 += len << 1;
    644     }
    645 #endif
    646 
    647     while(x1 < x2) {
    648         OneF1(p, x1, out, py0, py1, py2, py3, py4, cp->mFp);
    649         out++;
    650         x1++;
    651     }
    652 }
    653 
    654 RsdCpuScriptIntrinsicConvolve5x5::RsdCpuScriptIntrinsicConvolve5x5(
    655             RsdCpuReferenceImpl *ctx, const Script *s, const Element *e)
    656             : RsdCpuScriptIntrinsic(ctx, s, e, RS_SCRIPT_INTRINSIC_ID_CONVOLVE_5x5) {
    657 
    658     if (e->getType() == RS_TYPE_FLOAT_32) {
    659         switch(e->getVectorSize()) {
    660         case 1:
    661             mRootPtr = &kernelF1;
    662             break;
    663         case 2:
    664             mRootPtr = &kernelF2;
    665             break;
    666         case 3:
    667         case 4:
    668             mRootPtr = &kernelF4;
    669             break;
    670         }
    671     } else {
    672         switch(e->getVectorSize()) {
    673         case 1:
    674             mRootPtr = &kernelU1;
    675             break;
    676         case 2:
    677             mRootPtr = &kernelU2;
    678             break;
    679         case 3:
    680         case 4:
    681             mRootPtr = &kernelU4;
    682             break;
    683         }
    684     }
    685     for(int ct=0; ct < 25; ct++) {
    686         mFp[ct] = 1.f / 25.f;
    687         mIp[ct] = (short)(mFp[ct] * 256.f);
    688     }
    689 }
    690 
    691 RsdCpuScriptIntrinsicConvolve5x5::~RsdCpuScriptIntrinsicConvolve5x5() {
    692 }
    693 
    694 void RsdCpuScriptIntrinsicConvolve5x5::populateScript(Script *s) {
    695     s->mHal.info.exportedVariableCount = 2;
    696 }
    697 
    698 void RsdCpuScriptIntrinsicConvolve5x5::invokeFreeChildren() {
    699     alloc.clear();
    700 }
    701 
    702 
    703 RsdCpuScriptImpl * rsdIntrinsic_Convolve5x5(RsdCpuReferenceImpl *ctx,
    704                                             const Script *s, const Element *e) {
    705 
    706     return new RsdCpuScriptIntrinsicConvolve5x5(ctx, s, e);
    707 }
    708 
    709 
    710 
    711