rsCpuIntrinsicHistogram.cpp revision 80ef693674f69c0343c41564e30f80e7fb513b60
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 5280ef693674f69c0343c41564e30f80e7fb513b60Chris Wailes static void kernelP1U4(const RsExpandKernelParams *p, 532282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams uint32_t xstart, uint32_t xend, 542282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams uint32_t instep, uint32_t outstep); 5580ef693674f69c0343c41564e30f80e7fb513b60Chris Wailes static void kernelP1U3(const RsExpandKernelParams *p, 56b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams uint32_t xstart, uint32_t xend, 57b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams uint32_t instep, uint32_t outstep); 5880ef693674f69c0343c41564e30f80e7fb513b60Chris Wailes static void kernelP1U2(const RsExpandKernelParams *p, 59b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams uint32_t xstart, uint32_t xend, 60b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams uint32_t instep, uint32_t outstep); 6180ef693674f69c0343c41564e30f80e7fb513b60Chris Wailes static void kernelP1U1(const RsExpandKernelParams *p, 622282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams uint32_t xstart, uint32_t xend, 632282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams uint32_t instep, uint32_t outstep); 642282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams 6580ef693674f69c0343c41564e30f80e7fb513b60Chris Wailes static void kernelP1L4(const RsExpandKernelParams *p, 6675adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uint32_t xstart, uint32_t xend, 6775adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uint32_t instep, uint32_t outstep); 6880ef693674f69c0343c41564e30f80e7fb513b60Chris Wailes static void kernelP1L3(const RsExpandKernelParams *p, 6975adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uint32_t xstart, uint32_t xend, 7075adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uint32_t instep, uint32_t outstep); 7180ef693674f69c0343c41564e30f80e7fb513b60Chris Wailes static void kernelP1L2(const RsExpandKernelParams *p, 7275adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uint32_t xstart, uint32_t xend, 7375adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uint32_t instep, uint32_t outstep); 7480ef693674f69c0343c41564e30f80e7fb513b60Chris Wailes static void kernelP1L1(const RsExpandKernelParams *p, 7575adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uint32_t xstart, uint32_t xend, 7675adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uint32_t instep, uint32_t outstep); 7775adb8213f045bf3ffbc5deb1350b36d486e228aJason 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(); 10575adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uint32_t vSize = mAllocOut->getType()->getElement()->getVectorSize(); 1062282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams 1072282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams switch (slot) { 1082282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams case 0: 10975adb8213f045bf3ffbc5deb1350b36d486e228aJason 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; 11875adb8213f045bf3ffbc5deb1350b36d486e228aJason 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: 12675adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams switch(ain->getType()->getElement()->getVectorSize()) { 12775adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams case 1: 12875adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams mRootPtr = &kernelP1L1; 12975adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams break; 13075adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams case 2: 13175adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams mRootPtr = &kernelP1L2; 13275adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams break; 13375adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams case 3: 13475adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams mRootPtr = &kernelP1L3; 13575adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams break; 13675adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams case 4: 13775adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams mRootPtr = &kernelP1L4; 13875adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams break; 13975adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams } 1402282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams break; 1412282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams } 14275adb8213f045bf3ffbc5deb1350b36d486e228aJason 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 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 16380ef693674f69c0343c41564e30f80e7fb513b60Chris Wailesvoid RsdCpuScriptIntrinsicHistogram::kernelP1U4(const RsExpandKernelParams *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] ++; 17675adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams in += instep; 177b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams } 178b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams} 179b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams 18080ef693674f69c0343c41564e30f80e7fb513b60Chris Wailesvoid RsdCpuScriptIntrinsicHistogram::kernelP1U3(const RsExpandKernelParams *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] ++; 19275adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams in += instep; 193b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams } 194b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams} 195b68ba7e6f3d1e6440c1409e85a6c3af14fc39ee6Jason Sams 19680ef693674f69c0343c41564e30f80e7fb513b60Chris Wailesvoid RsdCpuScriptIntrinsicHistogram::kernelP1U2(const RsExpandKernelParams *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++) { 20575adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams sums[(in[0] << 1) ] ++; 20675adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams sums[(in[1] << 1) + 1] ++; 20775adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams in += instep; 2082282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams } 2092282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams} 2102282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams 21180ef693674f69c0343c41564e30f80e7fb513b60Chris Wailesvoid RsdCpuScriptIntrinsicHistogram::kernelP1L4(const RsExpandKernelParams *p, 21275adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uint32_t xstart, uint32_t xend, 21375adb8213f045bf3ffbc5deb1350b36d486e228aJason 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]); 22475adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams sums[(t + 0x7f) >> 8] ++; 22575adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams in += instep; 22675adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams } 22775adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams} 22875adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams 22980ef693674f69c0343c41564e30f80e7fb513b60Chris Wailesvoid RsdCpuScriptIntrinsicHistogram::kernelP1L3(const RsExpandKernelParams *p, 23075adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uint32_t xstart, uint32_t xend, 23175adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uint32_t instep, uint32_t outstep) { 23275adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams 23375adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)p->usr; 23475adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uchar *in = (uchar *)p->in; 23575adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams int * sums = &cp->mSums[256 * p->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] ++; 24275adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams in += instep; 24375adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams } 24475adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams} 24575adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams 24680ef693674f69c0343c41564e30f80e7fb513b60Chris Wailesvoid RsdCpuScriptIntrinsicHistogram::kernelP1L2(const RsExpandKernelParams *p, 24775adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uint32_t xstart, uint32_t xend, 24875adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uint32_t instep, uint32_t outstep) { 24975adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams 25075adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)p->usr; 25175adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uchar *in = (uchar *)p->in; 25275adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams int * sums = &cp->mSums[256 * p->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] ++; 25875adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams in += instep; 25975adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams } 26075adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams} 26175adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams 26280ef693674f69c0343c41564e30f80e7fb513b60Chris Wailesvoid RsdCpuScriptIntrinsicHistogram::kernelP1L1(const RsExpandKernelParams *p, 26375adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uint32_t xstart, uint32_t xend, 26475adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uint32_t instep, uint32_t outstep) { 26575adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams 26675adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)p->usr; 26775adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uchar *in = (uchar *)p->in; 26875adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams int * sums = &cp->mSums[256 * p->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] ++; 27375adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams in += instep; 2742282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams } 2752282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams} 2762282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams 27780ef693674f69c0343c41564e30f80e7fb513b60Chris Wailesvoid RsdCpuScriptIntrinsicHistogram::kernelP1U1(const RsExpandKernelParams *p, 2782282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams uint32_t xstart, uint32_t xend, 2792282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams uint32_t instep, uint32_t outstep) { 2802282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams 28175adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)p->usr; 28275adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams uchar *in = (uchar *)p->in; 28375adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams int * sums = &cp->mSums[256 * p->lid]; 28475adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams 28575adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams for (uint32_t x = xstart; x < xend; x++) { 28675adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams sums[in[0]] ++; 28775adb8213f045bf3ffbc5deb1350b36d486e228aJason Sams in += instep; 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 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