rsCpuIntrinsicConvolve3x3.cpp revision 3b35d775a777c36a178ce3fc97ff1e169aab3f1e 
1/*
2 * Copyright (C) 2012 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 *      http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17
18#include "rsCpuIntrinsic.h"
19#include "rsCpuIntrinsicInlines.h"
20
21using namespace android;
22using namespace android::renderscript;
23
24namespace android {
25namespace renderscript {
26
27
28class RsdCpuScriptIntrinsicConvolve3x3 : public RsdCpuScriptIntrinsic {
29public:
30    virtual void populateScript(Script *);
31    virtual void invokeFreeChildren();
32
33    virtual void setGlobalVar(uint32_t slot, const void *data, size_t dataLength);
34    virtual void setGlobalObj(uint32_t slot, ObjectBase *data);
35
36    virtual ~RsdCpuScriptIntrinsicConvolve3x3();
37    RsdCpuScriptIntrinsicConvolve3x3(RsdCpuReferenceImpl *ctx, const Script *s, const Element *);
38
39protected:
40    float mFp[16];
41    short mIp[16];
42    ObjectBaseRef<const Allocation> mAlloc;
43    ObjectBaseRef<const Element> mElement;
44
45    static void kernelU1(const RsForEachStubParamStruct *p,
46                         uint32_t xstart, uint32_t xend,
47                         uint32_t instep, uint32_t outstep);
48    static void kernelU2(const RsForEachStubParamStruct *p,
49                         uint32_t xstart, uint32_t xend,
50                         uint32_t instep, uint32_t outstep);
51    static void kernelU4(const RsForEachStubParamStruct *p,
52                         uint32_t xstart, uint32_t xend,
53                         uint32_t instep, uint32_t outstep);
54    static void kernelF1(const RsForEachStubParamStruct *p,
55                         uint32_t xstart, uint32_t xend,
56                         uint32_t instep, uint32_t outstep);
57    static void kernelF2(const RsForEachStubParamStruct *p,
58                         uint32_t xstart, uint32_t xend,
59                         uint32_t instep, uint32_t outstep);
60    static void kernelF4(const RsForEachStubParamStruct *p,
61                         uint32_t xstart, uint32_t xend,
62                         uint32_t instep, uint32_t outstep);
63};
64
65}
66}
67
68
69void RsdCpuScriptIntrinsicConvolve3x3::setGlobalObj(uint32_t slot, ObjectBase *data) {
70    rsAssert(slot == 1);
71    mAlloc.set(static_cast<Allocation *>(data));
72}
73
74void RsdCpuScriptIntrinsicConvolve3x3::setGlobalVar(uint32_t slot, const void *data,
75                                                    size_t dataLength) {
76    rsAssert(slot == 0);
77    memcpy (&mFp, data, dataLength);
78    for(int ct=0; ct < 9; ct++) {
79        if (mFp[ct] >= 0) {
80            mIp[ct] = (short)(mFp[ct] * 256.f + 0.5f);
81        } else {
82            mIp[ct] = (short)(mFp[ct] * 256.f - 0.5f);
83        }
84    }
85}
86
87extern "C" void rsdIntrinsicConvolve3x3_K(void *dst, const void *y0, const void *y1,
88                                          const void *y2, const short *coef, uint32_t count);
89
90
91static void ConvolveOneU4(const RsForEachStubParamStruct *p, uint32_t x, uchar4 *out,
92                          const uchar4 *py0, const uchar4 *py1, const uchar4 *py2,
93                          const float* coeff) {
94
95    uint32_t x1 = rsMax((int32_t)x-1, 0);
96    uint32_t x2 = rsMin((int32_t)x+1, (int32_t)p->dimX-1);
97
98    float4 px = convert_float4(py0[x1]) * coeff[0] +
99                convert_float4(py0[x]) * coeff[1] +
100                convert_float4(py0[x2]) * coeff[2] +
101                convert_float4(py1[x1]) * coeff[3] +
102                convert_float4(py1[x]) * coeff[4] +
103                convert_float4(py1[x2]) * coeff[5] +
104                convert_float4(py2[x1]) * coeff[6] +
105                convert_float4(py2[x]) * coeff[7] +
106                convert_float4(py2[x2]) * coeff[8];
107
108    px = clamp(px, 0.f, 255.f);
109    uchar4 o = {(uchar)px.x, (uchar)px.y, (uchar)px.z, (uchar)px.w};
110    *out = o;
111}
112
113static void ConvolveOneU2(const RsForEachStubParamStruct *p, uint32_t x, uchar2 *out,
114                          const uchar2 *py0, const uchar2 *py1, const uchar2 *py2,
115                          const float* coeff) {
116
117    uint32_t x1 = rsMax((int32_t)x-1, 0);
118    uint32_t x2 = rsMin((int32_t)x+1, (int32_t)p->dimX-1);
119
120    float2 px = convert_float2(py0[x1]) * coeff[0] +
121                convert_float2(py0[x]) * coeff[1] +
122                convert_float2(py0[x2]) * coeff[2] +
123                convert_float2(py1[x1]) * coeff[3] +
124                convert_float2(py1[x]) * coeff[4] +
125                convert_float2(py1[x2]) * coeff[5] +
126                convert_float2(py2[x1]) * coeff[6] +
127                convert_float2(py2[x]) * coeff[7] +
128                convert_float2(py2[x2]) * coeff[8];
129
130    px = clamp(px, 0.f, 255.f);
131    *out = convert_uchar2(px);
132}
133
134static void ConvolveOneU1(const RsForEachStubParamStruct *p, uint32_t x, uchar *out,
135                          const uchar *py0, const uchar *py1, const uchar *py2,
136                          const float* coeff) {
137
138    uint32_t x1 = rsMax((int32_t)x-1, 0);
139    uint32_t x2 = rsMin((int32_t)x+1, (int32_t)p->dimX-1);
140
141    float px = ((float)py0[x1]) * coeff[0] +
142               ((float)py0[x]) * coeff[1] +
143               ((float)py0[x2]) * coeff[2] +
144               ((float)py1[x1]) * coeff[3] +
145               ((float)py1[x]) * coeff[4] +
146               ((float)py1[x2]) * coeff[5] +
147               ((float)py2[x1]) * coeff[6] +
148               ((float)py2[x]) * coeff[7] +
149               ((float)py2[x2]) * coeff[8];
150    *out = clamp(px, 0.f, 255.f);
151}
152
153static void ConvolveOneF4(const RsForEachStubParamStruct *p, uint32_t x, float4 *out,
154                          const float4 *py0, const float4 *py1, const float4 *py2,
155                          const float* coeff) {
156
157    uint32_t x1 = rsMax((int32_t)x-1, 0);
158    uint32_t x2 = rsMin((int32_t)x+1, (int32_t)p->dimX-1);
159    *out = (py0[x1] * coeff[0]) + (py0[x] * coeff[1]) + (py0[x2] * coeff[2]) +
160           (py1[x1] * coeff[3]) + (py1[x] * coeff[4]) + (py1[x2] * coeff[5]) +
161           (py2[x1] * coeff[6]) + (py2[x] * coeff[7]) + (py2[x2] * coeff[8]);
162}
163
164static void ConvolveOneF2(const RsForEachStubParamStruct *p, uint32_t x, float2 *out,
165                          const float2 *py0, const float2 *py1, const float2 *py2,
166                          const float* coeff) {
167
168    uint32_t x1 = rsMax((int32_t)x-1, 0);
169    uint32_t x2 = rsMin((int32_t)x+1, (int32_t)p->dimX-1);
170    *out = (py0[x1] * coeff[0]) + (py0[x] * coeff[1]) + (py0[x2] * coeff[2]) +
171           (py1[x1] * coeff[3]) + (py1[x] * coeff[4]) + (py1[x2] * coeff[5]) +
172           (py2[x1] * coeff[6]) + (py2[x] * coeff[7]) + (py2[x2] * coeff[8]);
173}
174
175static void ConvolveOneF1(const RsForEachStubParamStruct *p, uint32_t x, float *out,
176                          const float *py0, const float *py1, const float *py2,
177                          const float* coeff) {
178
179    uint32_t x1 = rsMax((int32_t)x-1, 0);
180    uint32_t x2 = rsMin((int32_t)x+1, (int32_t)p->dimX-1);
181    *out = (py0[x1] * coeff[0]) + (py0[x] * coeff[1]) + (py0[x2] * coeff[2]) +
182           (py1[x1] * coeff[3]) + (py1[x] * coeff[4]) + (py1[x2] * coeff[5]) +
183           (py2[x1] * coeff[6]) + (py2[x] * coeff[7]) + (py2[x2] * coeff[8]);
184}
185
186void RsdCpuScriptIntrinsicConvolve3x3::kernelU4(const RsForEachStubParamStruct *p,
187                                                uint32_t xstart, uint32_t xend,
188                                                uint32_t instep, uint32_t outstep) {
189    RsdCpuScriptIntrinsicConvolve3x3 *cp = (RsdCpuScriptIntrinsicConvolve3x3 *)p->usr;
190
191    if (!cp->mAlloc.get()) {
192        ALOGE("Convolve3x3 executed without input, skipping");
193        return;
194    }
195    const uchar *pin = (const uchar *)cp->mAlloc->mHal.drvState.lod[0].mallocPtr;
196    const size_t stride = cp->mAlloc->mHal.drvState.lod[0].stride;
197
198    uint32_t y1 = rsMin((int32_t)p->y + 1, (int32_t)(p->dimY-1));
199    uint32_t y2 = rsMax((int32_t)p->y - 1, 0);
200    const uchar4 *py0 = (const uchar4 *)(pin + stride * y2);
201    const uchar4 *py1 = (const uchar4 *)(pin + stride * p->y);
202    const uchar4 *py2 = (const uchar4 *)(pin + stride * y1);
203
204    uchar4 *out = (uchar4 *)p->out;
205    uint32_t x1 = xstart;
206    uint32_t x2 = xend;
207    if(x1 == 0) {
208        ConvolveOneU4(p, 0, out, py0, py1, py2, cp->mFp);
209        x1 ++;
210        out++;
211    }
212
213    if(x2 > x1) {
214#if defined(ARCH_ARM_HAVE_NEON)
215        int32_t len = (x2 - x1 - 1) >> 1;
216        if(len > 0) {
217            rsdIntrinsicConvolve3x3_K(out, &py0[x1-1], &py1[x1-1], &py2[x1-1], cp->mIp, len);
218            x1 += len << 1;
219            out += len << 1;
220        }
221#endif
222
223        while(x1 != x2) {
224            ConvolveOneU4(p, x1, out, py0, py1, py2, cp->mFp);
225            out++;
226            x1++;
227        }
228    }
229}
230
231void RsdCpuScriptIntrinsicConvolve3x3::kernelU2(const RsForEachStubParamStruct *p,
232                                                uint32_t xstart, uint32_t xend,
233                                                uint32_t instep, uint32_t outstep) {
234    RsdCpuScriptIntrinsicConvolve3x3 *cp = (RsdCpuScriptIntrinsicConvolve3x3 *)p->usr;
235
236    if (!cp->mAlloc.get()) {
237        ALOGE("Convolve3x3 executed without input, skipping");
238        return;
239    }
240    const uchar *pin = (const uchar *)cp->mAlloc->mHal.drvState.lod[0].mallocPtr;
241    const size_t stride = cp->mAlloc->mHal.drvState.lod[0].stride;
242
243    uint32_t y1 = rsMin((int32_t)p->y + 1, (int32_t)(p->dimY-1));
244    uint32_t y2 = rsMax((int32_t)p->y - 1, 0);
245    const uchar2 *py0 = (const uchar2 *)(pin + stride * y2);
246    const uchar2 *py1 = (const uchar2 *)(pin + stride * p->y);
247    const uchar2 *py2 = (const uchar2 *)(pin + stride * y1);
248
249    uchar2 *out = (uchar2 *)p->out;
250    uint32_t x1 = xstart;
251    uint32_t x2 = xend;
252    if(x1 == 0) {
253        ConvolveOneU2(p, 0, out, py0, py1, py2, cp->mFp);
254        x1 ++;
255        out++;
256    }
257
258    if(x2 > x1) {
259#if 0//defined(ARCH_ARM_HAVE_NEON)
260        int32_t len = (x2 - x1 - 1) >> 1;
261        if(len > 0) {
262            rsdIntrinsicConvolve3x3_K(out, &py0[x1-1], &py1[x1-1], &py2[x1-1], cp->mIp, len);
263            x1 += len << 1;
264            out += len << 1;
265        }
266#endif
267
268        while(x1 != x2) {
269            ConvolveOneU2(p, x1, out, py0, py1, py2, cp->mFp);
270            out++;
271            x1++;
272        }
273    }
274}
275
276void RsdCpuScriptIntrinsicConvolve3x3::kernelU1(const RsForEachStubParamStruct *p,
277                                                uint32_t xstart, uint32_t xend,
278                                                uint32_t instep, uint32_t outstep) {
279    RsdCpuScriptIntrinsicConvolve3x3 *cp = (RsdCpuScriptIntrinsicConvolve3x3 *)p->usr;
280
281    if (!cp->mAlloc.get()) {
282        ALOGE("Convolve3x3 executed without input, skipping");
283        return;
284    }
285    const uchar *pin = (const uchar *)cp->mAlloc->mHal.drvState.lod[0].mallocPtr;
286    const size_t stride = cp->mAlloc->mHal.drvState.lod[0].stride;
287
288    uint32_t y1 = rsMin((int32_t)p->y + 1, (int32_t)(p->dimY-1));
289    uint32_t y2 = rsMax((int32_t)p->y - 1, 0);
290    const uchar *py0 = (const uchar *)(pin + stride * y2);
291    const uchar *py1 = (const uchar *)(pin + stride * p->y);
292    const uchar *py2 = (const uchar *)(pin + stride * y1);
293
294    uchar *out = (uchar *)p->out;
295    uint32_t x1 = xstart;
296    uint32_t x2 = xend;
297    if(x1 == 0) {
298        ConvolveOneU1(p, 0, out, py0, py1, py2, cp->mFp);
299        x1 ++;
300        out++;
301    }
302
303    if(x2 > x1) {
304#if 0//defined(ARCH_ARM_HAVE_NEON)
305        int32_t len = (x2 - x1 - 1) >> 1;
306        if(len > 0) {
307            rsdIntrinsicConvolve3x3_K(out, &py0[x1-1], &py1[x1-1], &py2[x1-1], cp->mIp, len);
308            x1 += len << 1;
309            out += len << 1;
310        }
311#endif
312
313        while(x1 != x2) {
314            ConvolveOneU1(p, x1, out, py0, py1, py2, cp->mFp);
315            out++;
316            x1++;
317        }
318    }
319}
320
321void RsdCpuScriptIntrinsicConvolve3x3::kernelF4(const RsForEachStubParamStruct *p,
322                                                uint32_t xstart, uint32_t xend,
323                                                uint32_t instep, uint32_t outstep) {
324    RsdCpuScriptIntrinsicConvolve3x3 *cp = (RsdCpuScriptIntrinsicConvolve3x3 *)p->usr;
325
326    if (!cp->mAlloc.get()) {
327        ALOGE("Convolve3x3 executed without input, skipping");
328        return;
329    }
330    const uchar *pin = (const uchar *)cp->mAlloc->mHal.drvState.lod[0].mallocPtr;
331    const size_t stride = cp->mAlloc->mHal.drvState.lod[0].stride;
332
333    uint32_t y1 = rsMin((int32_t)p->y + 1, (int32_t)(p->dimY-1));
334    uint32_t y2 = rsMax((int32_t)p->y - 1, 0);
335    const float4 *py0 = (const float4 *)(pin + stride * y2);
336    const float4 *py1 = (const float4 *)(pin + stride * p->y);
337    const float4 *py2 = (const float4 *)(pin + stride * y1);
338
339    float4 *out = (float4 *)p->out;
340    uint32_t x1 = xstart;
341    uint32_t x2 = xend;
342    if(x1 == 0) {
343        ConvolveOneF4(p, 0, out, py0, py1, py2, cp->mFp);
344        x1 ++;
345        out++;
346    }
347
348    if(x2 > x1) {
349#if 0//defined(ARCH_ARM_HAVE_NEON)
350        int32_t len = (x2 - x1 - 1) >> 1;
351        if(len > 0) {
352            rsdIntrinsicConvolve3x3_K(out, &py0[x1-1], &py1[x1-1], &py2[x1-1], cp->mIp, len);
353            x1 += len << 1;
354            out += len << 1;
355        }
356#endif
357
358        while(x1 != x2) {
359            ConvolveOneF4(p, x1, out, py0, py1, py2, cp->mFp);
360            out++;
361            x1++;
362        }
363    }
364}
365
366void RsdCpuScriptIntrinsicConvolve3x3::kernelF2(const RsForEachStubParamStruct *p,
367                                                uint32_t xstart, uint32_t xend,
368                                                uint32_t instep, uint32_t outstep) {
369    RsdCpuScriptIntrinsicConvolve3x3 *cp = (RsdCpuScriptIntrinsicConvolve3x3 *)p->usr;
370
371    if (!cp->mAlloc.get()) {
372        ALOGE("Convolve3x3 executed without input, skipping");
373        return;
374    }
375    const uchar *pin = (const uchar *)cp->mAlloc->mHal.drvState.lod[0].mallocPtr;
376    const size_t stride = cp->mAlloc->mHal.drvState.lod[0].stride;
377
378    uint32_t y1 = rsMin((int32_t)p->y + 1, (int32_t)(p->dimY-1));
379    uint32_t y2 = rsMax((int32_t)p->y - 1, 0);
380    const float2 *py0 = (const float2 *)(pin + stride * y2);
381    const float2 *py1 = (const float2 *)(pin + stride * p->y);
382    const float2 *py2 = (const float2 *)(pin + stride * y1);
383
384    float2 *out = (float2 *)p->out;
385    uint32_t x1 = xstart;
386    uint32_t x2 = xend;
387    if(x1 == 0) {
388        ConvolveOneF2(p, 0, out, py0, py1, py2, cp->mFp);
389        x1 ++;
390        out++;
391    }
392
393    if(x2 > x1) {
394#if 0//defined(ARCH_ARM_HAVE_NEON)
395        int32_t len = (x2 - x1 - 1) >> 1;
396        if(len > 0) {
397            rsdIntrinsicConvolve3x3_K(out, &py0[x1-1], &py1[x1-1], &py2[x1-1], cp->mIp, len);
398            x1 += len << 1;
399            out += len << 1;
400        }
401#endif
402
403        while(x1 != x2) {
404            ConvolveOneF2(p, x1, out, py0, py1, py2, cp->mFp);
405            out++;
406            x1++;
407        }
408    }
409}
410void RsdCpuScriptIntrinsicConvolve3x3::kernelF1(const RsForEachStubParamStruct *p,
411                                                uint32_t xstart, uint32_t xend,
412                                                uint32_t instep, uint32_t outstep) {
413    RsdCpuScriptIntrinsicConvolve3x3 *cp = (RsdCpuScriptIntrinsicConvolve3x3 *)p->usr;
414
415    if (!cp->mAlloc.get()) {
416        ALOGE("Convolve3x3 executed without input, skipping");
417        return;
418    }
419    const uchar *pin = (const uchar *)cp->mAlloc->mHal.drvState.lod[0].mallocPtr;
420    const size_t stride = cp->mAlloc->mHal.drvState.lod[0].stride;
421
422    uint32_t y1 = rsMin((int32_t)p->y + 1, (int32_t)(p->dimY-1));
423    uint32_t y2 = rsMax((int32_t)p->y - 1, 0);
424    const float *py0 = (const float *)(pin + stride * y2);
425    const float *py1 = (const float *)(pin + stride * p->y);
426    const float *py2 = (const float *)(pin + stride * y1);
427
428    float *out = (float *)p->out;
429    uint32_t x1 = xstart;
430    uint32_t x2 = xend;
431    if(x1 == 0) {
432        ConvolveOneF1(p, 0, out, py0, py1, py2, cp->mFp);
433        x1 ++;
434        out++;
435    }
436
437    if(x2 > x1) {
438#if 0//defined(ARCH_ARM_HAVE_NEON)
439        int32_t len = (x2 - x1 - 1) >> 1;
440        if(len > 0) {
441            rsdIntrinsicConvolve3x3_K(out, &py0[x1-1], &py1[x1-1], &py2[x1-1], cp->mIp, len);
442            x1 += len << 1;
443            out += len << 1;
444        }
445#endif
446
447        while(x1 != x2) {
448            ConvolveOneF1(p, x1, out, py0, py1, py2, cp->mFp);
449            out++;
450            x1++;
451        }
452    }
453}
454
455RsdCpuScriptIntrinsicConvolve3x3::RsdCpuScriptIntrinsicConvolve3x3(
456            RsdCpuReferenceImpl *ctx, const Script *s, const Element *e)
457            : RsdCpuScriptIntrinsic(ctx, s, e, RS_SCRIPT_INTRINSIC_ID_CONVOLVE_3x3) {
458
459    if (e->getType() == RS_TYPE_FLOAT_32) {
460        switch(e->getVectorSize()) {
461        case 1:
462            mRootPtr = &kernelF1;
463            break;
464        case 2:
465            mRootPtr = &kernelF2;
466            break;
467        case 3:
468        case 4:
469            mRootPtr = &kernelF4;
470            break;
471        }
472    } else {
473        switch(e->getVectorSize()) {
474        case 1:
475            mRootPtr = &kernelU1;
476            break;
477        case 2:
478            mRootPtr = &kernelU2;
479            break;
480        case 3:
481        case 4:
482            mRootPtr = &kernelU4;
483            break;
484        }
485    }
486    for(int ct=0; ct < 9; ct++) {
487        mFp[ct] = 1.f / 9.f;
488        mIp[ct] = (short)(mFp[ct] * 256.f + 0.5f);
489    }
490}
491
492RsdCpuScriptIntrinsicConvolve3x3::~RsdCpuScriptIntrinsicConvolve3x3() {
493}
494
495void RsdCpuScriptIntrinsicConvolve3x3::populateScript(Script *s) {
496    s->mHal.info.exportedVariableCount = 2;
497}
498
499void RsdCpuScriptIntrinsicConvolve3x3::invokeFreeChildren() {
500    mAlloc.clear();
501}
502
503
504RsdCpuScriptImpl * rsdIntrinsic_Convolve3x3(RsdCpuReferenceImpl *ctx, const Script *s, const Element *e) {
505
506    return new RsdCpuScriptIntrinsicConvolve3x3(ctx, s, e);
507}
508
509
510