1*e1eccf28SAndroid Build Coastguard Worker /*
2*e1eccf28SAndroid Build Coastguard Worker * Copyright (C) 2013 The Android Open Source Project
3*e1eccf28SAndroid Build Coastguard Worker *
4*e1eccf28SAndroid Build Coastguard Worker * Licensed under the Apache License, Version 2.0 (the "License");
5*e1eccf28SAndroid Build Coastguard Worker * you may not use this file except in compliance with the License.
6*e1eccf28SAndroid Build Coastguard Worker * You may obtain a copy of the License at
7*e1eccf28SAndroid Build Coastguard Worker *
8*e1eccf28SAndroid Build Coastguard Worker * http://www.apache.org/licenses/LICENSE-2.0
9*e1eccf28SAndroid Build Coastguard Worker *
10*e1eccf28SAndroid Build Coastguard Worker * Unless required by applicable law or agreed to in writing, software
11*e1eccf28SAndroid Build Coastguard Worker * distributed under the License is distributed on an "AS IS" BASIS,
12*e1eccf28SAndroid Build Coastguard Worker * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13*e1eccf28SAndroid Build Coastguard Worker * See the License for the specific language governing permissions and
14*e1eccf28SAndroid Build Coastguard Worker * limitations under the License.
15*e1eccf28SAndroid Build Coastguard Worker */
16*e1eccf28SAndroid Build Coastguard Worker
17*e1eccf28SAndroid Build Coastguard Worker #include "rsCpuIntrinsic.h"
18*e1eccf28SAndroid Build Coastguard Worker #include "rsCpuIntrinsicInlines.h"
19*e1eccf28SAndroid Build Coastguard Worker
20*e1eccf28SAndroid Build Coastguard Worker namespace android {
21*e1eccf28SAndroid Build Coastguard Worker namespace renderscript {
22*e1eccf28SAndroid Build Coastguard Worker
23*e1eccf28SAndroid Build Coastguard Worker
24*e1eccf28SAndroid Build Coastguard Worker class RsdCpuScriptIntrinsicHistogram : public RsdCpuScriptIntrinsic {
25*e1eccf28SAndroid Build Coastguard Worker public:
26*e1eccf28SAndroid Build Coastguard Worker void populateScript(Script *) override;
27*e1eccf28SAndroid Build Coastguard Worker void invokeFreeChildren() override;
28*e1eccf28SAndroid Build Coastguard Worker
29*e1eccf28SAndroid Build Coastguard Worker void setGlobalVar(uint32_t slot, const void *data, size_t dataLength) override;
30*e1eccf28SAndroid Build Coastguard Worker void setGlobalObj(uint32_t slot, ObjectBase *data) override;
31*e1eccf28SAndroid Build Coastguard Worker
32*e1eccf28SAndroid Build Coastguard Worker ~RsdCpuScriptIntrinsicHistogram() override;
33*e1eccf28SAndroid Build Coastguard Worker RsdCpuScriptIntrinsicHistogram(RsdCpuReferenceImpl *ctx, const Script *s, const Element *e);
34*e1eccf28SAndroid Build Coastguard Worker
35*e1eccf28SAndroid Build Coastguard Worker protected:
36*e1eccf28SAndroid Build Coastguard Worker void preLaunch(uint32_t slot, const Allocation ** ains, uint32_t inLen,
37*e1eccf28SAndroid Build Coastguard Worker Allocation * aout, const void * usr,
38*e1eccf28SAndroid Build Coastguard Worker uint32_t usrLen, const RsScriptCall *sc);
39*e1eccf28SAndroid Build Coastguard Worker void postLaunch(uint32_t slot, const Allocation ** ains, uint32_t inLen,
40*e1eccf28SAndroid Build Coastguard Worker Allocation * aout, const void * usr,
41*e1eccf28SAndroid Build Coastguard Worker uint32_t usrLen, const RsScriptCall *sc);
42*e1eccf28SAndroid Build Coastguard Worker
43*e1eccf28SAndroid Build Coastguard Worker
44*e1eccf28SAndroid Build Coastguard Worker float mDot[4];
45*e1eccf28SAndroid Build Coastguard Worker int mDotI[4];
46*e1eccf28SAndroid Build Coastguard Worker int *mSums;
47*e1eccf28SAndroid Build Coastguard Worker ObjectBaseRef<Allocation> mAllocOut;
48*e1eccf28SAndroid Build Coastguard Worker
49*e1eccf28SAndroid Build Coastguard Worker static void kernelP1U4(const RsExpandKernelDriverInfo *info,
50*e1eccf28SAndroid Build Coastguard Worker uint32_t xstart, uint32_t xend,
51*e1eccf28SAndroid Build Coastguard Worker uint32_t outstep);
52*e1eccf28SAndroid Build Coastguard Worker static void kernelP1U3(const RsExpandKernelDriverInfo *info,
53*e1eccf28SAndroid Build Coastguard Worker uint32_t xstart, uint32_t xend,
54*e1eccf28SAndroid Build Coastguard Worker uint32_t outstep);
55*e1eccf28SAndroid Build Coastguard Worker static void kernelP1U2(const RsExpandKernelDriverInfo *info,
56*e1eccf28SAndroid Build Coastguard Worker uint32_t xstart, uint32_t xend,
57*e1eccf28SAndroid Build Coastguard Worker uint32_t outstep);
58*e1eccf28SAndroid Build Coastguard Worker static void kernelP1U1(const RsExpandKernelDriverInfo *info,
59*e1eccf28SAndroid Build Coastguard Worker uint32_t xstart, uint32_t xend,
60*e1eccf28SAndroid Build Coastguard Worker uint32_t outstep);
61*e1eccf28SAndroid Build Coastguard Worker
62*e1eccf28SAndroid Build Coastguard Worker static void kernelP1L4(const RsExpandKernelDriverInfo *info,
63*e1eccf28SAndroid Build Coastguard Worker uint32_t xstart, uint32_t xend,
64*e1eccf28SAndroid Build Coastguard Worker uint32_t outstep);
65*e1eccf28SAndroid Build Coastguard Worker static void kernelP1L3(const RsExpandKernelDriverInfo *info,
66*e1eccf28SAndroid Build Coastguard Worker uint32_t xstart, uint32_t xend,
67*e1eccf28SAndroid Build Coastguard Worker uint32_t outstep);
68*e1eccf28SAndroid Build Coastguard Worker static void kernelP1L2(const RsExpandKernelDriverInfo *info,
69*e1eccf28SAndroid Build Coastguard Worker uint32_t xstart, uint32_t xend,
70*e1eccf28SAndroid Build Coastguard Worker uint32_t outstep);
71*e1eccf28SAndroid Build Coastguard Worker static void kernelP1L1(const RsExpandKernelDriverInfo *info,
72*e1eccf28SAndroid Build Coastguard Worker uint32_t xstart, uint32_t xend,
73*e1eccf28SAndroid Build Coastguard Worker uint32_t outstep);
74*e1eccf28SAndroid Build Coastguard Worker
75*e1eccf28SAndroid Build Coastguard Worker };
76*e1eccf28SAndroid Build Coastguard Worker
setGlobalObj(uint32_t slot,ObjectBase * data)77*e1eccf28SAndroid Build Coastguard Worker void RsdCpuScriptIntrinsicHistogram::setGlobalObj(uint32_t slot, ObjectBase *data) {
78*e1eccf28SAndroid Build Coastguard Worker rsAssert(slot == 1);
79*e1eccf28SAndroid Build Coastguard Worker mAllocOut.set(static_cast<Allocation *>(data));
80*e1eccf28SAndroid Build Coastguard Worker }
81*e1eccf28SAndroid Build Coastguard Worker
setGlobalVar(uint32_t slot,const void * data,size_t dataLength)82*e1eccf28SAndroid Build Coastguard Worker void RsdCpuScriptIntrinsicHistogram::setGlobalVar(uint32_t slot, const void *data, size_t dataLength) {
83*e1eccf28SAndroid Build Coastguard Worker rsAssert(slot == 0);
84*e1eccf28SAndroid Build Coastguard Worker rsAssert(dataLength == 16);
85*e1eccf28SAndroid Build Coastguard Worker memcpy(mDot, data, 16);
86*e1eccf28SAndroid Build Coastguard Worker mDotI[0] = (int)((mDot[0] * 256.f) + 0.5f);
87*e1eccf28SAndroid Build Coastguard Worker mDotI[1] = (int)((mDot[1] * 256.f) + 0.5f);
88*e1eccf28SAndroid Build Coastguard Worker mDotI[2] = (int)((mDot[2] * 256.f) + 0.5f);
89*e1eccf28SAndroid Build Coastguard Worker mDotI[3] = (int)((mDot[3] * 256.f) + 0.5f);
90*e1eccf28SAndroid Build Coastguard Worker }
91*e1eccf28SAndroid Build Coastguard Worker
92*e1eccf28SAndroid Build Coastguard Worker
93*e1eccf28SAndroid Build Coastguard Worker
94*e1eccf28SAndroid Build Coastguard Worker void
preLaunch(uint32_t slot,const Allocation ** ains,uint32_t inLen,Allocation * aout,const void * usr,uint32_t usrLen,const RsScriptCall * sc)95*e1eccf28SAndroid Build Coastguard Worker RsdCpuScriptIntrinsicHistogram::preLaunch(uint32_t slot,
96*e1eccf28SAndroid Build Coastguard Worker const Allocation ** ains,
97*e1eccf28SAndroid Build Coastguard Worker uint32_t inLen, Allocation * aout,
98*e1eccf28SAndroid Build Coastguard Worker const void * usr, uint32_t usrLen,
99*e1eccf28SAndroid Build Coastguard Worker const RsScriptCall *sc) {
100*e1eccf28SAndroid Build Coastguard Worker
101*e1eccf28SAndroid Build Coastguard Worker const uint32_t threads = mCtx->getThreadCount();
102*e1eccf28SAndroid Build Coastguard Worker uint32_t vSize = mAllocOut->getType()->getElement()->getVectorSize();
103*e1eccf28SAndroid Build Coastguard Worker
104*e1eccf28SAndroid Build Coastguard Worker switch (slot) {
105*e1eccf28SAndroid Build Coastguard Worker case 0:
106*e1eccf28SAndroid Build Coastguard Worker switch(vSize) {
107*e1eccf28SAndroid Build Coastguard Worker case 1:
108*e1eccf28SAndroid Build Coastguard Worker mRootPtr = &kernelP1U1;
109*e1eccf28SAndroid Build Coastguard Worker break;
110*e1eccf28SAndroid Build Coastguard Worker case 2:
111*e1eccf28SAndroid Build Coastguard Worker mRootPtr = &kernelP1U2;
112*e1eccf28SAndroid Build Coastguard Worker break;
113*e1eccf28SAndroid Build Coastguard Worker case 3:
114*e1eccf28SAndroid Build Coastguard Worker mRootPtr = &kernelP1U3;
115*e1eccf28SAndroid Build Coastguard Worker vSize = 4;
116*e1eccf28SAndroid Build Coastguard Worker break;
117*e1eccf28SAndroid Build Coastguard Worker case 4:
118*e1eccf28SAndroid Build Coastguard Worker mRootPtr = &kernelP1U4;
119*e1eccf28SAndroid Build Coastguard Worker break;
120*e1eccf28SAndroid Build Coastguard Worker }
121*e1eccf28SAndroid Build Coastguard Worker break;
122*e1eccf28SAndroid Build Coastguard Worker case 1:
123*e1eccf28SAndroid Build Coastguard Worker switch(ains[0]->getType()->getElement()->getVectorSize()) {
124*e1eccf28SAndroid Build Coastguard Worker case 1:
125*e1eccf28SAndroid Build Coastguard Worker mRootPtr = &kernelP1L1;
126*e1eccf28SAndroid Build Coastguard Worker break;
127*e1eccf28SAndroid Build Coastguard Worker case 2:
128*e1eccf28SAndroid Build Coastguard Worker mRootPtr = &kernelP1L2;
129*e1eccf28SAndroid Build Coastguard Worker break;
130*e1eccf28SAndroid Build Coastguard Worker case 3:
131*e1eccf28SAndroid Build Coastguard Worker mRootPtr = &kernelP1L3;
132*e1eccf28SAndroid Build Coastguard Worker break;
133*e1eccf28SAndroid Build Coastguard Worker case 4:
134*e1eccf28SAndroid Build Coastguard Worker mRootPtr = &kernelP1L4;
135*e1eccf28SAndroid Build Coastguard Worker break;
136*e1eccf28SAndroid Build Coastguard Worker }
137*e1eccf28SAndroid Build Coastguard Worker break;
138*e1eccf28SAndroid Build Coastguard Worker }
139*e1eccf28SAndroid Build Coastguard Worker memset(mSums, 0, 256 * sizeof(int32_t) * threads * vSize);
140*e1eccf28SAndroid Build Coastguard Worker }
141*e1eccf28SAndroid Build Coastguard Worker
142*e1eccf28SAndroid Build Coastguard Worker void
postLaunch(uint32_t slot,const Allocation ** ains,uint32_t inLen,Allocation * aout,const void * usr,uint32_t usrLen,const RsScriptCall * sc)143*e1eccf28SAndroid Build Coastguard Worker RsdCpuScriptIntrinsicHistogram::postLaunch(uint32_t slot,
144*e1eccf28SAndroid Build Coastguard Worker const Allocation ** ains,
145*e1eccf28SAndroid Build Coastguard Worker uint32_t inLen, Allocation * aout,
146*e1eccf28SAndroid Build Coastguard Worker const void * usr, uint32_t usrLen,
147*e1eccf28SAndroid Build Coastguard Worker const RsScriptCall *sc) {
148*e1eccf28SAndroid Build Coastguard Worker
149*e1eccf28SAndroid Build Coastguard Worker unsigned int *o = (unsigned int *)mAllocOut->mHal.drvState.lod[0].mallocPtr;
150*e1eccf28SAndroid Build Coastguard Worker uint32_t threads = mCtx->getThreadCount();
151*e1eccf28SAndroid Build Coastguard Worker uint32_t vSize = mAllocOut->getType()->getElement()->getVectorSize();
152*e1eccf28SAndroid Build Coastguard Worker
153*e1eccf28SAndroid Build Coastguard Worker if (vSize == 3) vSize = 4;
154*e1eccf28SAndroid Build Coastguard Worker
155*e1eccf28SAndroid Build Coastguard Worker for (uint32_t ct=0; ct < (256 * vSize); ct++) {
156*e1eccf28SAndroid Build Coastguard Worker o[ct] = mSums[ct];
157*e1eccf28SAndroid Build Coastguard Worker for (uint32_t t=1; t < threads; t++) {
158*e1eccf28SAndroid Build Coastguard Worker o[ct] += mSums[ct + (256 * vSize * t)];
159*e1eccf28SAndroid Build Coastguard Worker }
160*e1eccf28SAndroid Build Coastguard Worker }
161*e1eccf28SAndroid Build Coastguard Worker }
162*e1eccf28SAndroid Build Coastguard Worker
kernelP1U4(const RsExpandKernelDriverInfo * info,uint32_t xstart,uint32_t xend,uint32_t outstep)163*e1eccf28SAndroid Build Coastguard Worker void RsdCpuScriptIntrinsicHistogram::kernelP1U4(const RsExpandKernelDriverInfo *info,
164*e1eccf28SAndroid Build Coastguard Worker uint32_t xstart, uint32_t xend,
165*e1eccf28SAndroid Build Coastguard Worker uint32_t outstep) {
166*e1eccf28SAndroid Build Coastguard Worker
167*e1eccf28SAndroid Build Coastguard Worker RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)info->usr;
168*e1eccf28SAndroid Build Coastguard Worker uchar *in = (uchar *)info->inPtr[0];
169*e1eccf28SAndroid Build Coastguard Worker int * sums = &cp->mSums[256 * 4 * info->lid];
170*e1eccf28SAndroid Build Coastguard Worker
171*e1eccf28SAndroid Build Coastguard Worker for (uint32_t x = xstart; x < xend; x++) {
172*e1eccf28SAndroid Build Coastguard Worker sums[(in[0] << 2) ] ++;
173*e1eccf28SAndroid Build Coastguard Worker sums[(in[1] << 2) + 1] ++;
174*e1eccf28SAndroid Build Coastguard Worker sums[(in[2] << 2) + 2] ++;
175*e1eccf28SAndroid Build Coastguard Worker sums[(in[3] << 2) + 3] ++;
176*e1eccf28SAndroid Build Coastguard Worker in += info->inStride[0];
177*e1eccf28SAndroid Build Coastguard Worker }
178*e1eccf28SAndroid Build Coastguard Worker }
179*e1eccf28SAndroid Build Coastguard Worker
kernelP1U3(const RsExpandKernelDriverInfo * info,uint32_t xstart,uint32_t xend,uint32_t outstep)180*e1eccf28SAndroid Build Coastguard Worker void RsdCpuScriptIntrinsicHistogram::kernelP1U3(const RsExpandKernelDriverInfo *info,
181*e1eccf28SAndroid Build Coastguard Worker uint32_t xstart, uint32_t xend,
182*e1eccf28SAndroid Build Coastguard Worker uint32_t outstep) {
183*e1eccf28SAndroid Build Coastguard Worker
184*e1eccf28SAndroid Build Coastguard Worker RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)info->usr;
185*e1eccf28SAndroid Build Coastguard Worker uchar *in = (uchar *)info->inPtr[0];
186*e1eccf28SAndroid Build Coastguard Worker int * sums = &cp->mSums[256 * 4 * info->lid];
187*e1eccf28SAndroid Build Coastguard Worker
188*e1eccf28SAndroid Build Coastguard Worker for (uint32_t x = xstart; x < xend; x++) {
189*e1eccf28SAndroid Build Coastguard Worker sums[(in[0] << 2) ] ++;
190*e1eccf28SAndroid Build Coastguard Worker sums[(in[1] << 2) + 1] ++;
191*e1eccf28SAndroid Build Coastguard Worker sums[(in[2] << 2) + 2] ++;
192*e1eccf28SAndroid Build Coastguard Worker in += info->inStride[0];
193*e1eccf28SAndroid Build Coastguard Worker }
194*e1eccf28SAndroid Build Coastguard Worker }
195*e1eccf28SAndroid Build Coastguard Worker
kernelP1U2(const RsExpandKernelDriverInfo * info,uint32_t xstart,uint32_t xend,uint32_t outstep)196*e1eccf28SAndroid Build Coastguard Worker void RsdCpuScriptIntrinsicHistogram::kernelP1U2(const RsExpandKernelDriverInfo *info,
197*e1eccf28SAndroid Build Coastguard Worker uint32_t xstart, uint32_t xend,
198*e1eccf28SAndroid Build Coastguard Worker uint32_t outstep) {
199*e1eccf28SAndroid Build Coastguard Worker
200*e1eccf28SAndroid Build Coastguard Worker RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)info->usr;
201*e1eccf28SAndroid Build Coastguard Worker uchar *in = (uchar *)info->inPtr[0];
202*e1eccf28SAndroid Build Coastguard Worker int * sums = &cp->mSums[256 * 2 * info->lid];
203*e1eccf28SAndroid Build Coastguard Worker
204*e1eccf28SAndroid Build Coastguard Worker for (uint32_t x = xstart; x < xend; x++) {
205*e1eccf28SAndroid Build Coastguard Worker sums[(in[0] << 1) ] ++;
206*e1eccf28SAndroid Build Coastguard Worker sums[(in[1] << 1) + 1] ++;
207*e1eccf28SAndroid Build Coastguard Worker in += info->inStride[0];
208*e1eccf28SAndroid Build Coastguard Worker }
209*e1eccf28SAndroid Build Coastguard Worker }
210*e1eccf28SAndroid Build Coastguard Worker
kernelP1L4(const RsExpandKernelDriverInfo * info,uint32_t xstart,uint32_t xend,uint32_t outstep)211*e1eccf28SAndroid Build Coastguard Worker void RsdCpuScriptIntrinsicHistogram::kernelP1L4(const RsExpandKernelDriverInfo *info,
212*e1eccf28SAndroid Build Coastguard Worker uint32_t xstart, uint32_t xend,
213*e1eccf28SAndroid Build Coastguard Worker uint32_t outstep) {
214*e1eccf28SAndroid Build Coastguard Worker
215*e1eccf28SAndroid Build Coastguard Worker RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)info->usr;
216*e1eccf28SAndroid Build Coastguard Worker uchar *in = (uchar *)info->inPtr[0];
217*e1eccf28SAndroid Build Coastguard Worker int * sums = &cp->mSums[256 * info->lid];
218*e1eccf28SAndroid Build Coastguard Worker
219*e1eccf28SAndroid Build Coastguard Worker for (uint32_t x = xstart; x < xend; x++) {
220*e1eccf28SAndroid Build Coastguard Worker int t = (cp->mDotI[0] * in[0]) +
221*e1eccf28SAndroid Build Coastguard Worker (cp->mDotI[1] * in[1]) +
222*e1eccf28SAndroid Build Coastguard Worker (cp->mDotI[2] * in[2]) +
223*e1eccf28SAndroid Build Coastguard Worker (cp->mDotI[3] * in[3]);
224*e1eccf28SAndroid Build Coastguard Worker sums[(t + 0x7f) >> 8] ++;
225*e1eccf28SAndroid Build Coastguard Worker in += info->inStride[0];
226*e1eccf28SAndroid Build Coastguard Worker }
227*e1eccf28SAndroid Build Coastguard Worker }
228*e1eccf28SAndroid Build Coastguard Worker
kernelP1L3(const RsExpandKernelDriverInfo * info,uint32_t xstart,uint32_t xend,uint32_t outstep)229*e1eccf28SAndroid Build Coastguard Worker void RsdCpuScriptIntrinsicHistogram::kernelP1L3(const RsExpandKernelDriverInfo *info,
230*e1eccf28SAndroid Build Coastguard Worker uint32_t xstart, uint32_t xend,
231*e1eccf28SAndroid Build Coastguard Worker uint32_t outstep) {
232*e1eccf28SAndroid Build Coastguard Worker
233*e1eccf28SAndroid Build Coastguard Worker RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)info->usr;
234*e1eccf28SAndroid Build Coastguard Worker uchar *in = (uchar *)info->inPtr[0];
235*e1eccf28SAndroid Build Coastguard Worker int * sums = &cp->mSums[256 * info->lid];
236*e1eccf28SAndroid Build Coastguard Worker
237*e1eccf28SAndroid Build Coastguard Worker for (uint32_t x = xstart; x < xend; x++) {
238*e1eccf28SAndroid Build Coastguard Worker int t = (cp->mDotI[0] * in[0]) +
239*e1eccf28SAndroid Build Coastguard Worker (cp->mDotI[1] * in[1]) +
240*e1eccf28SAndroid Build Coastguard Worker (cp->mDotI[2] * in[2]);
241*e1eccf28SAndroid Build Coastguard Worker sums[(t + 0x7f) >> 8] ++;
242*e1eccf28SAndroid Build Coastguard Worker in += info->inStride[0];
243*e1eccf28SAndroid Build Coastguard Worker }
244*e1eccf28SAndroid Build Coastguard Worker }
245*e1eccf28SAndroid Build Coastguard Worker
kernelP1L2(const RsExpandKernelDriverInfo * info,uint32_t xstart,uint32_t xend,uint32_t outstep)246*e1eccf28SAndroid Build Coastguard Worker void RsdCpuScriptIntrinsicHistogram::kernelP1L2(const RsExpandKernelDriverInfo *info,
247*e1eccf28SAndroid Build Coastguard Worker uint32_t xstart, uint32_t xend,
248*e1eccf28SAndroid Build Coastguard Worker uint32_t outstep) {
249*e1eccf28SAndroid Build Coastguard Worker
250*e1eccf28SAndroid Build Coastguard Worker RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)info->usr;
251*e1eccf28SAndroid Build Coastguard Worker uchar *in = (uchar *)info->inPtr[0];
252*e1eccf28SAndroid Build Coastguard Worker int * sums = &cp->mSums[256 * info->lid];
253*e1eccf28SAndroid Build Coastguard Worker
254*e1eccf28SAndroid Build Coastguard Worker for (uint32_t x = xstart; x < xend; x++) {
255*e1eccf28SAndroid Build Coastguard Worker int t = (cp->mDotI[0] * in[0]) +
256*e1eccf28SAndroid Build Coastguard Worker (cp->mDotI[1] * in[1]);
257*e1eccf28SAndroid Build Coastguard Worker sums[(t + 0x7f) >> 8] ++;
258*e1eccf28SAndroid Build Coastguard Worker in += info->inStride[0];
259*e1eccf28SAndroid Build Coastguard Worker }
260*e1eccf28SAndroid Build Coastguard Worker }
261*e1eccf28SAndroid Build Coastguard Worker
kernelP1L1(const RsExpandKernelDriverInfo * info,uint32_t xstart,uint32_t xend,uint32_t outstep)262*e1eccf28SAndroid Build Coastguard Worker void RsdCpuScriptIntrinsicHistogram::kernelP1L1(const RsExpandKernelDriverInfo *info,
263*e1eccf28SAndroid Build Coastguard Worker uint32_t xstart, uint32_t xend,
264*e1eccf28SAndroid Build Coastguard Worker uint32_t outstep) {
265*e1eccf28SAndroid Build Coastguard Worker
266*e1eccf28SAndroid Build Coastguard Worker RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)info->usr;
267*e1eccf28SAndroid Build Coastguard Worker uchar *in = (uchar *)info->inPtr[0];
268*e1eccf28SAndroid Build Coastguard Worker int * sums = &cp->mSums[256 * info->lid];
269*e1eccf28SAndroid Build Coastguard Worker
270*e1eccf28SAndroid Build Coastguard Worker for (uint32_t x = xstart; x < xend; x++) {
271*e1eccf28SAndroid Build Coastguard Worker int t = (cp->mDotI[0] * in[0]);
272*e1eccf28SAndroid Build Coastguard Worker sums[(t + 0x7f) >> 8] ++;
273*e1eccf28SAndroid Build Coastguard Worker in += info->inStride[0];
274*e1eccf28SAndroid Build Coastguard Worker }
275*e1eccf28SAndroid Build Coastguard Worker }
276*e1eccf28SAndroid Build Coastguard Worker
kernelP1U1(const RsExpandKernelDriverInfo * info,uint32_t xstart,uint32_t xend,uint32_t outstep)277*e1eccf28SAndroid Build Coastguard Worker void RsdCpuScriptIntrinsicHistogram::kernelP1U1(const RsExpandKernelDriverInfo *info,
278*e1eccf28SAndroid Build Coastguard Worker uint32_t xstart, uint32_t xend,
279*e1eccf28SAndroid Build Coastguard Worker uint32_t outstep) {
280*e1eccf28SAndroid Build Coastguard Worker
281*e1eccf28SAndroid Build Coastguard Worker RsdCpuScriptIntrinsicHistogram *cp = (RsdCpuScriptIntrinsicHistogram *)info->usr;
282*e1eccf28SAndroid Build Coastguard Worker uchar *in = (uchar *)info->inPtr[0];
283*e1eccf28SAndroid Build Coastguard Worker int * sums = &cp->mSums[256 * info->lid];
284*e1eccf28SAndroid Build Coastguard Worker
285*e1eccf28SAndroid Build Coastguard Worker for (uint32_t x = xstart; x < xend; x++) {
286*e1eccf28SAndroid Build Coastguard Worker sums[in[0]] ++;
287*e1eccf28SAndroid Build Coastguard Worker in += info->inStride[0];
288*e1eccf28SAndroid Build Coastguard Worker }
289*e1eccf28SAndroid Build Coastguard Worker }
290*e1eccf28SAndroid Build Coastguard Worker
291*e1eccf28SAndroid Build Coastguard Worker
RsdCpuScriptIntrinsicHistogram(RsdCpuReferenceImpl * ctx,const Script * s,const Element * e)292*e1eccf28SAndroid Build Coastguard Worker RsdCpuScriptIntrinsicHistogram::RsdCpuScriptIntrinsicHistogram(RsdCpuReferenceImpl *ctx,
293*e1eccf28SAndroid Build Coastguard Worker const Script *s, const Element *e)
294*e1eccf28SAndroid Build Coastguard Worker : RsdCpuScriptIntrinsic(ctx, s, e, RS_SCRIPT_INTRINSIC_ID_HISTOGRAM) {
295*e1eccf28SAndroid Build Coastguard Worker
296*e1eccf28SAndroid Build Coastguard Worker mRootPtr = nullptr;
297*e1eccf28SAndroid Build Coastguard Worker mSums = new int[256 * 4 * mCtx->getThreadCount()];
298*e1eccf28SAndroid Build Coastguard Worker mDot[0] = 0.299f;
299*e1eccf28SAndroid Build Coastguard Worker mDot[1] = 0.587f;
300*e1eccf28SAndroid Build Coastguard Worker mDot[2] = 0.114f;
301*e1eccf28SAndroid Build Coastguard Worker mDot[3] = 0;
302*e1eccf28SAndroid Build Coastguard Worker mDotI[0] = (int)((mDot[0] * 256.f) + 0.5f);
303*e1eccf28SAndroid Build Coastguard Worker mDotI[1] = (int)((mDot[1] * 256.f) + 0.5f);
304*e1eccf28SAndroid Build Coastguard Worker mDotI[2] = (int)((mDot[2] * 256.f) + 0.5f);
305*e1eccf28SAndroid Build Coastguard Worker mDotI[3] = (int)((mDot[3] * 256.f) + 0.5f);
306*e1eccf28SAndroid Build Coastguard Worker }
307*e1eccf28SAndroid Build Coastguard Worker
~RsdCpuScriptIntrinsicHistogram()308*e1eccf28SAndroid Build Coastguard Worker RsdCpuScriptIntrinsicHistogram::~RsdCpuScriptIntrinsicHistogram() {
309*e1eccf28SAndroid Build Coastguard Worker if (mSums) {
310*e1eccf28SAndroid Build Coastguard Worker delete []mSums;
311*e1eccf28SAndroid Build Coastguard Worker }
312*e1eccf28SAndroid Build Coastguard Worker }
313*e1eccf28SAndroid Build Coastguard Worker
populateScript(Script * s)314*e1eccf28SAndroid Build Coastguard Worker void RsdCpuScriptIntrinsicHistogram::populateScript(Script *s) {
315*e1eccf28SAndroid Build Coastguard Worker s->mHal.info.exportedVariableCount = 2;
316*e1eccf28SAndroid Build Coastguard Worker }
317*e1eccf28SAndroid Build Coastguard Worker
invokeFreeChildren()318*e1eccf28SAndroid Build Coastguard Worker void RsdCpuScriptIntrinsicHistogram::invokeFreeChildren() {
319*e1eccf28SAndroid Build Coastguard Worker }
320*e1eccf28SAndroid Build Coastguard Worker
rsdIntrinsic_Histogram(RsdCpuReferenceImpl * ctx,const Script * s,const Element * e)321*e1eccf28SAndroid Build Coastguard Worker RsdCpuScriptImpl * rsdIntrinsic_Histogram(RsdCpuReferenceImpl *ctx, const Script *s, const Element *e) {
322*e1eccf28SAndroid Build Coastguard Worker
323*e1eccf28SAndroid Build Coastguard Worker return new RsdCpuScriptIntrinsicHistogram(ctx, s, e);
324*e1eccf28SAndroid Build Coastguard Worker }
325*e1eccf28SAndroid Build Coastguard Worker
326*e1eccf28SAndroid Build Coastguard Worker } // namespace renderscript
327*e1eccf28SAndroid Build Coastguard Worker } // namespace android
328