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 Samsusing namespace android;
212282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsusing namespace android::renderscript;
222282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
232282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsnamespace android {
242282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsnamespace renderscript {
252282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
262282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
272282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsclass RsdCpuScriptIntrinsicHistogram : public RsdCpuScriptIntrinsic {
282282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samspublic:
292282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    virtual void populateScript(Script *);
302282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    virtual void invokeFreeChildren();
312282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
322282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    virtual void setGlobalVar(uint32_t slot, const void *data, size_t dataLength);
332282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    virtual void setGlobalObj(uint32_t slot, ObjectBase *data);
342282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
352282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    virtual ~RsdCpuScriptIntrinsicHistogram();
362282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    RsdCpuScriptIntrinsicHistogram(RsdCpuReferenceImpl *ctx, const Script *s, const Element *e);
372282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
382282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsprotected:
392282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    void preLaunch(uint32_t slot, const Allocation * ain,
402282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                   Allocation * aout, const void * usr,
412282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                   uint32_t usrLen, const RsScriptCall *sc);
422282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    void postLaunch(uint32_t slot, const Allocation * ain,
432282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                    Allocation * aout, const void * usr,
442282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                    uint32_t usrLen, const RsScriptCall *sc);
452282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
462282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
472282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    float mDot[4];
482282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    int mDotI[4];
492282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    int *mSums;
502282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    ObjectBaseRef<Allocation> mAllocOut;
512282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
522282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    static void kernelP1U4(const RsForEachStubParamStruct *p,
532282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                          uint32_t xstart, uint32_t xend,
542282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                          uint32_t instep, uint32_t outstep);
55b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams    static void kernelP1U3(const RsForEachStubParamStruct *p,
56b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams                          uint32_t xstart, uint32_t xend,
57b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams                          uint32_t instep, uint32_t outstep);
58b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams    static void kernelP1U2(const RsForEachStubParamStruct *p,
59b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams                          uint32_t xstart, uint32_t xend,
60b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams                          uint32_t instep, uint32_t outstep);
612282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    static void kernelP1U1(const RsForEachStubParamStruct *p,
622282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                          uint32_t xstart, uint32_t xend,
632282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                          uint32_t instep, uint32_t outstep);
642282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
65a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    static void kernelP1L4(const RsForEachStubParamStruct *p,
66a6f119256f29405bd04414d0fd915c023fad5518Jason Sams                           uint32_t xstart, uint32_t xend,
67a6f119256f29405bd04414d0fd915c023fad5518Jason Sams                           uint32_t instep, uint32_t outstep);
68a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    static void kernelP1L3(const RsForEachStubParamStruct *p,
69a6f119256f29405bd04414d0fd915c023fad5518Jason Sams                           uint32_t xstart, uint32_t xend,
70a6f119256f29405bd04414d0fd915c023fad5518Jason Sams                           uint32_t instep, uint32_t outstep);
71a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    static void kernelP1L2(const RsForEachStubParamStruct *p,
72a6f119256f29405bd04414d0fd915c023fad5518Jason Sams                           uint32_t xstart, uint32_t xend,
73a6f119256f29405bd04414d0fd915c023fad5518Jason Sams                           uint32_t instep, uint32_t outstep);
74a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    static void kernelP1L1(const RsForEachStubParamStruct *p,
75a6f119256f29405bd04414d0fd915c023fad5518Jason Sams                           uint32_t xstart, uint32_t xend,
76a6f119256f29405bd04414d0fd915c023fad5518Jason Sams                           uint32_t instep, uint32_t outstep);
77a6f119256f29405bd04414d0fd915c023fad5518Jason Sams
782282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams};
792282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
802282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
812282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
822282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
832282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsvoid RsdCpuScriptIntrinsicHistogram::setGlobalObj(uint32_t slot, ObjectBase *data) {
842282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    rsAssert(slot == 1);
852282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mAllocOut.set(static_cast<Allocation *>(data));
862282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
872282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
882282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsvoid RsdCpuScriptIntrinsicHistogram::setGlobalVar(uint32_t slot, const void *data, size_t dataLength) {
892282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    rsAssert(slot == 0);
902282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    rsAssert(dataLength == 16);
912282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    memcpy(mDot, data, 16);
922282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mDotI[0] = (int)((mDot[0] * 256.f) + 0.5f);
932282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mDotI[1] = (int)((mDot[1] * 256.f) + 0.5f);
942282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mDotI[2] = (int)((mDot[2] * 256.f) + 0.5f);
952282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mDotI[3] = (int)((mDot[3] * 256.f) + 0.5f);
962282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
972282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
982282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
992282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
1002282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsvoid RsdCpuScriptIntrinsicHistogram::preLaunch(uint32_t slot, const Allocation * ain,
1012282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                                      Allocation * aout, const void * usr,
1022282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                                      uint32_t usrLen, const RsScriptCall *sc) {
1032282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
1042282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    const uint32_t threads = mCtx->getThreadCount();
105a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    uint32_t vSize = mAllocOut->getType()->getElement()->getVectorSize();
1062282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
1072282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    switch (slot) {
1082282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    case 0:
109a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        switch(vSize) {
110b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams        case 1:
1112282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams            mRootPtr = &kernelP1U1;
112b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams            break;
113b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams        case 2:
114b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams            mRootPtr = &kernelP1U2;
115b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams            break;
116b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams        case 3:
117b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams            mRootPtr = &kernelP1U3;
118a6f119256f29405bd04414d0fd915c023fad5518Jason Sams            vSize = 4;
119b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams            break;
120b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams        case 4:
1212282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams            mRootPtr = &kernelP1U4;
122b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams            break;
1232282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams        }
1242282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams        break;
1252282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    case 1:
126a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        switch(ain->getType()->getElement()->getVectorSize()) {
127a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        case 1:
128a6f119256f29405bd04414d0fd915c023fad5518Jason Sams            mRootPtr = &kernelP1L1;
129a6f119256f29405bd04414d0fd915c023fad5518Jason Sams            break;
130a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        case 2:
131a6f119256f29405bd04414d0fd915c023fad5518Jason Sams            mRootPtr = &kernelP1L2;
132a6f119256f29405bd04414d0fd915c023fad5518Jason Sams            break;
133a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        case 3:
134a6f119256f29405bd04414d0fd915c023fad5518Jason Sams            mRootPtr = &kernelP1L3;
135a6f119256f29405bd04414d0fd915c023fad5518Jason Sams            break;
136a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        case 4:
137a6f119256f29405bd04414d0fd915c023fad5518Jason Sams            mRootPtr = &kernelP1L4;
138a6f119256f29405bd04414d0fd915c023fad5518Jason Sams            break;
139a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        }
1402282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams        break;
1412282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    }
142a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    memset(mSums, 0, 256 * sizeof(int32_t) * threads * vSize);
1432282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
1442282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
1452282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsvoid RsdCpuScriptIntrinsicHistogram::postLaunch(uint32_t slot, const Allocation * ain,
1462282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                                       Allocation * aout, const void * usr,
1472282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                                       uint32_t usrLen, 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
153a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    if (vSize == 3) vSize = 4;
154a6f119256f29405bd04414d0fd915c023fad5518Jason 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++) {
158af79a59353ae59be186d849b5a7b6865d1241d4cTim Murray            o[ct] += mSums[ct + (256 * vSize * t)];
1592282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams        }
1602282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    }
1612282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
1622282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
1632282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsvoid RsdCpuScriptIntrinsicHistogram::kernelP1U4(const RsForEachStubParamStruct *p,
1642282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                                                uint32_t xstart, uint32_t xend,
1652282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                                                uint32_t instep, uint32_t outstep) {
1662282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
1672282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)p->usr;
168b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams    uchar *in = (uchar *)p->in;
1692282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    int * sums = &cp->mSums[256 * 4 * p->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] ++;
176a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        in += instep;
177b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams    }
178b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams}
179b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams
180b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Samsvoid RsdCpuScriptIntrinsicHistogram::kernelP1U3(const RsForEachStubParamStruct *p,
181b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams                                                uint32_t xstart, uint32_t xend,
182b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams                                                uint32_t instep, uint32_t outstep) {
183b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams
184b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams    RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)p->usr;
185b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams    uchar *in = (uchar *)p->in;
186b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams    int * sums = &cp->mSums[256 * 4 * p->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] ++;
192a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        in += instep;
193b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams    }
194b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams}
195b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams
196b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Samsvoid RsdCpuScriptIntrinsicHistogram::kernelP1U2(const RsForEachStubParamStruct *p,
197b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams                                                uint32_t xstart, uint32_t xend,
198b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams                                                uint32_t instep, uint32_t outstep) {
199b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams
200b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams    RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)p->usr;
201b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams    uchar *in = (uchar *)p->in;
202b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams    int * sums = &cp->mSums[256 * 2 * p->lid];
203b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams
204b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams    for (uint32_t x = xstart; x < xend; x++) {
205a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        sums[(in[0] << 1)    ] ++;
206a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        sums[(in[1] << 1) + 1] ++;
207a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        in += instep;
2082282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    }
2092282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
2102282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
211a6f119256f29405bd04414d0fd915c023fad5518Jason Samsvoid RsdCpuScriptIntrinsicHistogram::kernelP1L4(const RsForEachStubParamStruct *p,
212a6f119256f29405bd04414d0fd915c023fad5518Jason Sams                                                uint32_t xstart, uint32_t xend,
213a6f119256f29405bd04414d0fd915c023fad5518Jason Sams                                                uint32_t instep, uint32_t outstep) {
2142282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
2152282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)p->usr;
216b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams    uchar *in = (uchar *)p->in;
2172282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    int * sums = &cp->mSums[256 * p->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]);
224a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        sums[(t + 0x7f) >> 8] ++;
225a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        in += instep;
226a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    }
227a6f119256f29405bd04414d0fd915c023fad5518Jason Sams}
228a6f119256f29405bd04414d0fd915c023fad5518Jason Sams
229a6f119256f29405bd04414d0fd915c023fad5518Jason Samsvoid RsdCpuScriptIntrinsicHistogram::kernelP1L3(const RsForEachStubParamStruct *p,
230a6f119256f29405bd04414d0fd915c023fad5518Jason Sams                                                uint32_t xstart, uint32_t xend,
231a6f119256f29405bd04414d0fd915c023fad5518Jason Sams                                                uint32_t instep, uint32_t outstep) {
232a6f119256f29405bd04414d0fd915c023fad5518Jason Sams
233a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)p->usr;
234a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    uchar *in = (uchar *)p->in;
235a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    int * sums = &cp->mSums[256 * p->lid];
236a6f119256f29405bd04414d0fd915c023fad5518Jason Sams
237a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    for (uint32_t x = xstart; x < xend; x++) {
238a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        int t = (cp->mDotI[0] * in[0]) +
239a6f119256f29405bd04414d0fd915c023fad5518Jason Sams                (cp->mDotI[1] * in[1]) +
240a6f119256f29405bd04414d0fd915c023fad5518Jason Sams                (cp->mDotI[2] * in[2]);
241a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        sums[(t + 0x7f) >> 8] ++;
242a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        in += instep;
243a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    }
244a6f119256f29405bd04414d0fd915c023fad5518Jason Sams}
245a6f119256f29405bd04414d0fd915c023fad5518Jason Sams
246a6f119256f29405bd04414d0fd915c023fad5518Jason Samsvoid RsdCpuScriptIntrinsicHistogram::kernelP1L2(const RsForEachStubParamStruct *p,
247a6f119256f29405bd04414d0fd915c023fad5518Jason Sams                                                uint32_t xstart, uint32_t xend,
248a6f119256f29405bd04414d0fd915c023fad5518Jason Sams                                                uint32_t instep, uint32_t outstep) {
249a6f119256f29405bd04414d0fd915c023fad5518Jason Sams
250a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)p->usr;
251a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    uchar *in = (uchar *)p->in;
252a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    int * sums = &cp->mSums[256 * p->lid];
253a6f119256f29405bd04414d0fd915c023fad5518Jason Sams
254a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    for (uint32_t x = xstart; x < xend; x++) {
255a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        int t = (cp->mDotI[0] * in[0]) +
256a6f119256f29405bd04414d0fd915c023fad5518Jason Sams                (cp->mDotI[1] * in[1]);
257a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        sums[(t + 0x7f) >> 8] ++;
258a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        in += instep;
259a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    }
260a6f119256f29405bd04414d0fd915c023fad5518Jason Sams}
261a6f119256f29405bd04414d0fd915c023fad5518Jason Sams
262a6f119256f29405bd04414d0fd915c023fad5518Jason Samsvoid RsdCpuScriptIntrinsicHistogram::kernelP1L1(const RsForEachStubParamStruct *p,
263a6f119256f29405bd04414d0fd915c023fad5518Jason Sams                                                uint32_t xstart, uint32_t xend,
264a6f119256f29405bd04414d0fd915c023fad5518Jason Sams                                                uint32_t instep, uint32_t outstep) {
265a6f119256f29405bd04414d0fd915c023fad5518Jason Sams
266a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)p->usr;
267a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    uchar *in = (uchar *)p->in;
268a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    int * sums = &cp->mSums[256 * p->lid];
269a6f119256f29405bd04414d0fd915c023fad5518Jason Sams
270a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    for (uint32_t x = xstart; x < xend; x++) {
271a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        int t = (cp->mDotI[0] * in[0]);
272a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        sums[(t + 0x7f) >> 8] ++;
273a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        in += instep;
2742282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    }
2752282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
2762282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
2772282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsvoid RsdCpuScriptIntrinsicHistogram::kernelP1U1(const RsForEachStubParamStruct *p,
2782282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                                                uint32_t xstart, uint32_t xend,
2792282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                                                uint32_t instep, uint32_t outstep) {
2802282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
281a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)p->usr;
282a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    uchar *in = (uchar *)p->in;
283a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    int * sums = &cp->mSums[256 * p->lid];
284a6f119256f29405bd04414d0fd915c023fad5518Jason Sams
285a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    for (uint32_t x = xstart; x < xend; x++) {
286a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        sums[in[0]] ++;
287a6f119256f29405bd04414d0fd915c023fad5518Jason Sams        in += instep;
288a6f119256f29405bd04414d0fd915c023fad5518Jason Sams    }
2892282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
2902282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
2912282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
2922282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason SamsRsdCpuScriptIntrinsicHistogram::RsdCpuScriptIntrinsicHistogram(RsdCpuReferenceImpl *ctx,
2932282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams                                                     const Script *s, const Element *e)
294af79a59353ae59be186d849b5a7b6865d1241d4cTim Murray            : RsdCpuScriptIntrinsic(ctx, s, e, RS_SCRIPT_INTRINSIC_ID_HISTOGRAM) {
2952282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
2962282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    mRootPtr = NULL;
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 Sams
3222282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason SamsRsdCpuScriptImpl * rsdIntrinsic_Histogram(RsdCpuReferenceImpl *ctx, const Script *s, const Element *e) {
3232282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
3242282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams    return new RsdCpuScriptIntrinsicHistogram(ctx, s, e);
3252282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams}
3262282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
3272282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams
328