12282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams/*
22282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams * Copyright (C) 2013 The Android Open Source Project
32282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams *
42282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams * Licensed under the Apache License, Version 2.0 (the "License");
52282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams * you may not use this file except in compliance with the License.
62282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams * You may obtain a copy of the License at
72282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams *
82282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams *      http://www.apache.org/licenses/LICENSE-2.0
92282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams *
102282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams * Unless required by applicable law or agreed to in writing, software
112282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams * distributed under the License is distributed on an "AS IS" BASIS,
122282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
132282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams * See the License for the specific language governing permissions and
142282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams * limitations under the License.
152282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams */
162282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
172282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams#include "rsCpuIntrinsic.h"
182282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams#include "rsCpuIntrinsicInlines.h"
192282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
202282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsnamespace android {
212282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsnamespace renderscript {
222282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
232282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
242282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsclass RsdCpuScriptIntrinsicHistogram : public RsdCpuScriptIntrinsic {
252282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samspublic:
26c060f1435e7b9405f3be8974417fa6f410f03753Stephen Hines    void populateScript(Script *) override;
27c060f1435e7b9405f3be8974417fa6f410f03753Stephen Hines    void invokeFreeChildren() override;
282282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
29c060f1435e7b9405f3be8974417fa6f410f03753Stephen Hines    void setGlobalVar(uint32_t slot, const void *data, size_t dataLength) override;
30c060f1435e7b9405f3be8974417fa6f410f03753Stephen Hines    void setGlobalObj(uint32_t slot, ObjectBase *data) override;
312282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
32c060f1435e7b9405f3be8974417fa6f410f03753Stephen Hines    ~RsdCpuScriptIntrinsicHistogram() override;
332282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    RsdCpuScriptIntrinsicHistogram(RsdCpuReferenceImpl *ctx, const Script *s, const Element *e);
342282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
352282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsprotected:
36f37121300217d3b39ab66dd9c8881bcbcad932dfChris Wailes    void preLaunch(uint32_t slot, const Allocation ** ains, uint32_t inLen,
372282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                   Allocation * aout, const void * usr,
382282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                   uint32_t usrLen, const RsScriptCall *sc);
39f37121300217d3b39ab66dd9c8881bcbcad932dfChris Wailes    void postLaunch(uint32_t slot, const Allocation ** ains, uint32_t inLen,
402282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                    Allocation * aout, const void * usr,
412282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                    uint32_t usrLen, const RsScriptCall *sc);
422282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
432282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
442282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    float mDot[4];
452282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    int mDotI[4];
462282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    int *mSums;
472282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    ObjectBaseRef<Allocation> mAllocOut;
482282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
49b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    static void kernelP1U4(const RsExpandKernelDriverInfo *info,
509ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                           uint32_t xstart, uint32_t xend,
519ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                           uint32_t outstep);
52b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    static void kernelP1U3(const RsExpandKernelDriverInfo *info,
539ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                           uint32_t xstart, uint32_t xend,
549ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                           uint32_t outstep);
55b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    static void kernelP1U2(const RsExpandKernelDriverInfo *info,
569ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                           uint32_t xstart, uint32_t xend,
579ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                           uint32_t outstep);
58b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    static void kernelP1U1(const RsExpandKernelDriverInfo *info,
599ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                           uint32_t xstart, uint32_t xend,
609ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                           uint32_t outstep);
612282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
62b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    static void kernelP1L4(const RsExpandKernelDriverInfo *info,
6375adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams                           uint32_t xstart, uint32_t xend,
649ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                           uint32_t outstep);
65b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    static void kernelP1L3(const RsExpandKernelDriverInfo *info,
6675adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams                           uint32_t xstart, uint32_t xend,
679ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                           uint32_t outstep);
68b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    static void kernelP1L2(const RsExpandKernelDriverInfo *info,
6975adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams                           uint32_t xstart, uint32_t xend,
709ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                           uint32_t outstep);
71b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    static void kernelP1L1(const RsExpandKernelDriverInfo *info,
7275adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams                           uint32_t xstart, uint32_t xend,
739ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                           uint32_t outstep);
7475adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams
752282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams};
762282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
772282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsvoid RsdCpuScriptIntrinsicHistogram::setGlobalObj(uint32_t slot, ObjectBase *data) {
782282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    rsAssert(slot == 1);
792282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mAllocOut.set(static_cast<Allocation *>(data));
802282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
812282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
822282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsvoid RsdCpuScriptIntrinsicHistogram::setGlobalVar(uint32_t slot, const void *data, size_t dataLength) {
832282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    rsAssert(slot == 0);
842282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    rsAssert(dataLength == 16);
852282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    memcpy(mDot, data, 16);
862282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mDotI[0] = (int)((mDot[0] * 256.f) + 0.5f);
872282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mDotI[1] = (int)((mDot[1] * 256.f) + 0.5f);
882282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mDotI[2] = (int)((mDot[2] * 256.f) + 0.5f);
892282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mDotI[3] = (int)((mDot[3] * 256.f) + 0.5f);
902282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
912282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
922282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
932282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
94f37121300217d3b39ab66dd9c8881bcbcad932dfChris Wailesvoid
95f37121300217d3b39ab66dd9c8881bcbcad932dfChris WailesRsdCpuScriptIntrinsicHistogram::preLaunch(uint32_t slot,
96f37121300217d3b39ab66dd9c8881bcbcad932dfChris Wailes                                          const Allocation ** ains,
97f37121300217d3b39ab66dd9c8881bcbcad932dfChris Wailes                                          uint32_t inLen, Allocation * aout,
98f37121300217d3b39ab66dd9c8881bcbcad932dfChris Wailes                                          const void * usr, uint32_t usrLen,
99f37121300217d3b39ab66dd9c8881bcbcad932dfChris Wailes                                          const RsScriptCall *sc) {
1002282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
1012282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    const uint32_t threads = mCtx->getThreadCount();
10275adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams    uint32_t vSize = mAllocOut->getType()->getElement()->getVectorSize();
1032282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
1042282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    switch (slot) {
1052282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    case 0:
10675adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams        switch(vSize) {
107b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams        case 1:
1082282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams            mRootPtr = &kernelP1U1;
109b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams            break;
110b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams        case 2:
111b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams            mRootPtr = &kernelP1U2;
112b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams            break;
113b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams        case 3:
114b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams            mRootPtr = &kernelP1U3;
11575adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams            vSize = 4;
116b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams            break;
117b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams        case 4:
1182282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams            mRootPtr = &kernelP1U4;
119b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams            break;
1202282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams        }
1212282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams        break;
1222282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    case 1:
123f37121300217d3b39ab66dd9c8881bcbcad932dfChris Wailes        switch(ains[0]->getType()->getElement()->getVectorSize()) {
12475adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams        case 1:
12575adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams            mRootPtr = &kernelP1L1;
12675adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams            break;
12775adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams        case 2:
12875adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams            mRootPtr = &kernelP1L2;
12975adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams            break;
13075adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams        case 3:
13175adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams            mRootPtr = &kernelP1L3;
13275adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams            break;
13375adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams        case 4:
13475adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams            mRootPtr = &kernelP1L4;
13575adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams            break;
13675adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams        }
1372282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams        break;
1382282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    }
13975adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams    memset(mSums, 0, 256 * sizeof(int32_t) * threads * vSize);
1402282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
1412282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
142f37121300217d3b39ab66dd9c8881bcbcad932dfChris Wailesvoid
143f37121300217d3b39ab66dd9c8881bcbcad932dfChris WailesRsdCpuScriptIntrinsicHistogram::postLaunch(uint32_t slot,
144f37121300217d3b39ab66dd9c8881bcbcad932dfChris Wailes                                           const Allocation ** ains,
145f37121300217d3b39ab66dd9c8881bcbcad932dfChris Wailes                                           uint32_t inLen,  Allocation * aout,
146f37121300217d3b39ab66dd9c8881bcbcad932dfChris Wailes                                           const void * usr, uint32_t usrLen,
147f37121300217d3b39ab66dd9c8881bcbcad932dfChris Wailes                                           const RsScriptCall *sc) {
1482282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
1492282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    unsigned int *o = (unsigned int *)mAllocOut->mHal.drvState.lod[0].mallocPtr;
1502282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    uint32_t threads = mCtx->getThreadCount();
1512282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    uint32_t vSize = mAllocOut->getType()->getElement()->getVectorSize();
1522282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
15375adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams    if (vSize == 3) vSize = 4;
15475adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams
1552282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    for (uint32_t ct=0; ct < (256 * vSize); ct++) {
1562282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams        o[ct] = mSums[ct];
1572282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams        for (uint32_t t=1; t < threads; t++) {
1586de1d8375526502b468232d77fce3e957c705137Tim Murray            o[ct] += mSums[ct + (256 * vSize * t)];
1592282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams        }
1602282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    }
1612282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
1622282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
163b0abb140ac51b93d1a85aadaa63fe057f2d29850David Grossvoid RsdCpuScriptIntrinsicHistogram::kernelP1U4(const RsExpandKernelDriverInfo *info,
1642282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                                                uint32_t xstart, uint32_t xend,
1659ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                                                uint32_t outstep) {
1662282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
167b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)info->usr;
168b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    uchar *in = (uchar *)info->inPtr[0];
169b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    int * sums = &cp->mSums[256 * 4 * info->lid];
1702282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
1712282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    for (uint32_t x = xstart; x < xend; x++) {
172b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams        sums[(in[0] << 2)    ] ++;
173b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams        sums[(in[1] << 2) + 1] ++;
174b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams        sums[(in[2] << 2) + 2] ++;
175b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams        sums[(in[3] << 2) + 3] ++;
176b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross        in += info->inStride[0];
177b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams    }
178b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams}
179b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams
180b0abb140ac51b93d1a85aadaa63fe057f2d29850David Grossvoid RsdCpuScriptIntrinsicHistogram::kernelP1U3(const RsExpandKernelDriverInfo *info,
181b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams                                                uint32_t xstart, uint32_t xend,
1829ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                                                uint32_t outstep) {
183b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams
184b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)info->usr;
185b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    uchar *in = (uchar *)info->inPtr[0];
186b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    int * sums = &cp->mSums[256 * 4 * info->lid];
187b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams
188b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams    for (uint32_t x = xstart; x < xend; x++) {
189b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams        sums[(in[0] << 2)    ] ++;
190b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams        sums[(in[1] << 2) + 1] ++;
191b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams        sums[(in[2] << 2) + 2] ++;
192b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross        in += info->inStride[0];
193b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams    }
194b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams}
195b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams
196b0abb140ac51b93d1a85aadaa63fe057f2d29850David Grossvoid RsdCpuScriptIntrinsicHistogram::kernelP1U2(const RsExpandKernelDriverInfo *info,
197b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams                                                uint32_t xstart, uint32_t xend,
1989ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                                                uint32_t outstep) {
199b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams
200b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)info->usr;
201b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    uchar *in = (uchar *)info->inPtr[0];
202b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    int * sums = &cp->mSums[256 * 2 * info->lid];
203b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams
204b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams    for (uint32_t x = xstart; x < xend; x++) {
20575adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams        sums[(in[0] << 1)    ] ++;
20675adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams        sums[(in[1] << 1) + 1] ++;
207b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross        in += info->inStride[0];
2082282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    }
2092282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
2102282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
211b0abb140ac51b93d1a85aadaa63fe057f2d29850David Grossvoid RsdCpuScriptIntrinsicHistogram::kernelP1L4(const RsExpandKernelDriverInfo *info,
21275adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams                                                uint32_t xstart, uint32_t xend,
2139ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                                                uint32_t outstep) {
2142282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
215b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)info->usr;
216b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    uchar *in = (uchar *)info->inPtr[0];
217b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    int * sums = &cp->mSums[256 * info->lid];
2182282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
2192282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    for (uint32_t x = xstart; x < xend; x++) {
220b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams        int t = (cp->mDotI[0] * in[0]) +
221b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams                (cp->mDotI[1] * in[1]) +
222b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams                (cp->mDotI[2] * in[2]) +
223b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams                (cp->mDotI[3] * in[3]);
22475adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams        sums[(t + 0x7f) >> 8] ++;
225b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross        in += info->inStride[0];
22675adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams    }
22775adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams}
22875adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams
229b0abb140ac51b93d1a85aadaa63fe057f2d29850David Grossvoid RsdCpuScriptIntrinsicHistogram::kernelP1L3(const RsExpandKernelDriverInfo *info,
23075adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams                                                uint32_t xstart, uint32_t xend,
2319ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                                                uint32_t outstep) {
23275adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams
233b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)info->usr;
234b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    uchar *in = (uchar *)info->inPtr[0];
235b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    int * sums = &cp->mSums[256 * info->lid];
23675adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams
23775adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams    for (uint32_t x = xstart; x < xend; x++) {
23875adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams        int t = (cp->mDotI[0] * in[0]) +
23975adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams                (cp->mDotI[1] * in[1]) +
24075adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams                (cp->mDotI[2] * in[2]);
24175adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams        sums[(t + 0x7f) >> 8] ++;
242b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross        in += info->inStride[0];
24375adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams    }
24475adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams}
24575adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams
246b0abb140ac51b93d1a85aadaa63fe057f2d29850David Grossvoid RsdCpuScriptIntrinsicHistogram::kernelP1L2(const RsExpandKernelDriverInfo *info,
24775adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams                                                uint32_t xstart, uint32_t xend,
2489ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                                                uint32_t outstep) {
24975adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams
250b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)info->usr;
251b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    uchar *in = (uchar *)info->inPtr[0];
252b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    int * sums = &cp->mSums[256 * info->lid];
25375adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams
25475adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams    for (uint32_t x = xstart; x < xend; x++) {
25575adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams        int t = (cp->mDotI[0] * in[0]) +
25675adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams                (cp->mDotI[1] * in[1]);
25775adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams        sums[(t + 0x7f) >> 8] ++;
258b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross        in += info->inStride[0];
25975adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams    }
26075adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams}
26175adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams
262b0abb140ac51b93d1a85aadaa63fe057f2d29850David Grossvoid RsdCpuScriptIntrinsicHistogram::kernelP1L1(const RsExpandKernelDriverInfo *info,
26375adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams                                                uint32_t xstart, uint32_t xend,
2649ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                                                uint32_t outstep) {
26575adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams
266b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)info->usr;
267b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    uchar *in = (uchar *)info->inPtr[0];
268b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    int * sums = &cp->mSums[256 * info->lid];
26975adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams
27075adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams    for (uint32_t x = xstart; x < xend; x++) {
27175adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams        int t = (cp->mDotI[0] * in[0]);
27275adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams        sums[(t + 0x7f) >> 8] ++;
273b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross        in += info->inStride[0];
2742282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    }
2752282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
2762282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
277b0abb140ac51b93d1a85aadaa63fe057f2d29850David Grossvoid RsdCpuScriptIntrinsicHistogram::kernelP1U1(const RsExpandKernelDriverInfo *info,
2782282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                                                uint32_t xstart, uint32_t xend,
2799ed79105cc6a8dbfaf959875249f36022cc2c798Chris Wailes                                                uint32_t outstep) {
2802282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
281b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)info->usr;
282b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    uchar *in = (uchar *)info->inPtr[0];
283b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross    int * sums = &cp->mSums[256 * info->lid];
28475adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams
28575adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams    for (uint32_t x = xstart; x < xend; x++) {
28675adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams        sums[in[0]] ++;
287b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross        in += info->inStride[0];
28875adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams    }
2892282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
2902282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
2912282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
2922282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason SamsRsdCpuScriptIntrinsicHistogram::RsdCpuScriptIntrinsicHistogram(RsdCpuReferenceImpl *ctx,
2932282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                                                     const Script *s, const Element *e)
2946de1d8375526502b468232d77fce3e957c705137Tim Murray            : RsdCpuScriptIntrinsic(ctx, s, e, RS_SCRIPT_INTRINSIC_ID_HISTOGRAM) {
2952282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
29644bef6fba6244292b751387f3d6c31cca96c28adChris Wailes    mRootPtr = nullptr;
2972282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mSums = new int[256 * 4 * mCtx->getThreadCount()];
2982282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mDot[0] = 0.299f;
2992282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mDot[1] = 0.587f;
3002282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mDot[2] = 0.114f;
3012282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mDot[3] = 0;
3022282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mDotI[0] = (int)((mDot[0] * 256.f) + 0.5f);
3032282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mDotI[1] = (int)((mDot[1] * 256.f) + 0.5f);
3042282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mDotI[2] = (int)((mDot[2] * 256.f) + 0.5f);
3052282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mDotI[3] = (int)((mDot[3] * 256.f) + 0.5f);
3062282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
3072282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
3082282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason SamsRsdCpuScriptIntrinsicHistogram::~RsdCpuScriptIntrinsicHistogram() {
3092282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    if (mSums) {
3102282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams        delete []mSums;
3112282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    }
3122282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
3132282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
3142282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsvoid RsdCpuScriptIntrinsicHistogram::populateScript(Script *s) {
3152282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    s->mHal.info.exportedVariableCount = 2;
3162282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
3172282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
3182282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsvoid RsdCpuScriptIntrinsicHistogram::invokeFreeChildren() {
3192282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
3202282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
3212282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason SamsRsdCpuScriptImpl * rsdIntrinsic_Histogram(RsdCpuReferenceImpl *ctx, const Script *s, const Element *e) {
3222282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
3232282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    return new RsdCpuScriptIntrinsicHistogram(ctx, s, e);
3242282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
325462de21ac2e1773b99aedee012adb374e476ae36Chih-Hung Hsieh
326462de21ac2e1773b99aedee012adb374e476ae36Chih-Hung Hsieh} // namespace renderscript
327462de21ac2e1773b99aedee012adb374e476ae36Chih-Hung Hsieh} // namespace android
328