rsCpuIntrinsicBlur.cpp revision 2913f381a554c28abb44f49eddd1ee4c68a72578 
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#include "rsCpuIntrinsic.h"
18#include "rsCpuIntrinsicInlines.h"
19
20using namespace android;
21using namespace android::renderscript;
22
23namespace android {
24namespace renderscript {
25
26
27class RsdCpuScriptIntrinsicBlur : public RsdCpuScriptIntrinsic {
28public:
29    virtual void populateScript(Script *);
30    virtual void invokeFreeChildren();
31
32    virtual void setGlobalVar(uint32_t slot, const void *data, size_t dataLength);
33    virtual void setGlobalObj(uint32_t slot, ObjectBase *data);
34
35    virtual ~RsdCpuScriptIntrinsicBlur();
36    RsdCpuScriptIntrinsicBlur(RsdCpuReferenceImpl *ctx, const Script *s, const Element *e);
37
38protected:
39    float mFp[104];
40    short mIp[104];
41    void **mScratch;
42    size_t *mScratchSize;
43    float mRadius;
44    int mIradius;
45    ObjectBaseRef<Allocation> mAlloc;
46
47    static void kernelU4(const RsForEachStubParamStruct *p,
48                         uint32_t xstart, uint32_t xend,
49                         uint32_t instep, uint32_t outstep);
50    static void kernelU1(const RsForEachStubParamStruct *p,
51                         uint32_t xstart, uint32_t xend,
52                         uint32_t instep, uint32_t outstep);
53    void ComputeGaussianWeights();
54};
55
56}
57}
58
59
60void RsdCpuScriptIntrinsicBlur::ComputeGaussianWeights() {
61    memset(mFp, 0, sizeof(mFp));
62    memset(mIp, 0, sizeof(mIp));
63
64    // Compute gaussian weights for the blur
65    // e is the euler's number
66    float e = 2.718281828459045f;
67    float pi = 3.1415926535897932f;
68    // g(x) = ( 1 / sqrt( 2 * pi ) * sigma) * e ^ ( -x^2 / 2 * sigma^2 )
69    // x is of the form [-radius .. 0 .. radius]
70    // and sigma varies with radius.
71    // Based on some experimental radius values and sigma's
72    // we approximately fit sigma = f(radius) as
73    // sigma = radius * 0.4  + 0.6
74    // The larger the radius gets, the more our gaussian blur
75    // will resemble a box blur since with large sigma
76    // the gaussian curve begins to lose its shape
77    float sigma = 0.4f * mRadius + 0.6f;
78
79    // Now compute the coefficients. We will store some redundant values to save
80    // some math during the blur calculations precompute some values
81    float coeff1 = 1.0f / (sqrtf(2.0f * pi) * sigma);
82    float coeff2 = - 1.0f / (2.0f * sigma * sigma);
83
84    float normalizeFactor = 0.0f;
85    float floatR = 0.0f;
86    int r;
87    mIradius = (float)ceil(mRadius) + 0.5f;
88    for (r = -mIradius; r <= mIradius; r ++) {
89        floatR = (float)r;
90        mFp[r + mIradius] = coeff1 * powf(e, floatR * floatR * coeff2);
91        normalizeFactor += mFp[r + mIradius];
92    }
93
94    //Now we need to normalize the weights because all our coefficients need to add up to one
95    normalizeFactor = 1.0f / normalizeFactor;
96    for (r = -mIradius; r <= mIradius; r ++) {
97        mFp[r + mIradius] *= normalizeFactor;
98        mIp[r + mIradius] = (short)(mIp[r + mIradius] * 32768);
99    }
100}
101
102void RsdCpuScriptIntrinsicBlur::setGlobalObj(uint32_t slot, ObjectBase *data) {
103    rsAssert(slot == 1);
104    mAlloc.set(static_cast<Allocation *>(data));
105}
106
107void RsdCpuScriptIntrinsicBlur::setGlobalVar(uint32_t slot, const void *data, size_t dataLength) {
108    rsAssert(slot == 0);
109    mRadius = ((const float *)data)[0];
110    ComputeGaussianWeights();
111}
112
113
114
115static void OneVU4(const RsForEachStubParamStruct *p, float4 *out, int32_t x, int32_t y,
116                   const uchar *ptrIn, int iStride, const float* gPtr, int iradius) {
117
118    const uchar *pi = ptrIn + x*4;
119
120    float4 blurredPixel = 0;
121    for (int r = -iradius; r <= iradius; r ++) {
122        int validY = rsMax((y + r), 0);
123        validY = rsMin(validY, (int)(p->dimY - 1));
124        const uchar4 *pvy = (const uchar4 *)&pi[validY * iStride];
125        float4 pf = convert_float4(pvy[0]);
126        blurredPixel += pf * gPtr[0];
127        gPtr++;
128    }
129
130    out->xyzw = blurredPixel;
131}
132
133static void OneVU1(const RsForEachStubParamStruct *p, float *out, int32_t x, int32_t y,
134                   const uchar *ptrIn, int iStride, const float* gPtr, int iradius) {
135
136    const uchar *pi = ptrIn + x;
137
138    float blurredPixel = 0;
139    for (int r = -iradius; r <= iradius; r ++) {
140        int validY = rsMax((y + r), 0);
141        validY = rsMin(validY, (int)(p->dimY - 1));
142        float pf = (float)pi[validY * iStride];
143        blurredPixel += pf * gPtr[0];
144        gPtr++;
145    }
146
147    out[0] = blurredPixel;
148}
149
150extern "C" void rsdIntrinsicBlurVFU4_K(void *dst, const void *pin, int stride, const void *gptr, int rct, int x1, int ct);
151extern "C" void rsdIntrinsicBlurHFU4_K(void *dst, const void *pin, const void *gptr, int rct, int x1, int ct);
152extern "C" void rsdIntrinsicBlurHFU1_K(void *dst, const void *pin, const void *gptr, int rct, int x1, int ct);
153
154static void OneVFU4(float4 *out,
155                    const uchar *ptrIn, int iStride, const float* gPtr, int ct,
156                    int x1, int x2) {
157
158#if defined(ARCH_ARM_HAVE_NEON)
159    {
160        int t = (x2 - x1);
161        t &= ~1;
162        if(t) {
163            rsdIntrinsicBlurVFU4_K(out, ptrIn, iStride, gPtr, ct, x1, x1 + t);
164        }
165        x1 += t;
166    }
167#endif
168
169    while(x2 > x1) {
170        const uchar *pi = ptrIn;
171        float4 blurredPixel = 0;
172        const float* gp = gPtr;
173
174        for (int r = 0; r < ct; r++) {
175            float4 pf = convert_float4(((const uchar4 *)pi)[0]);
176            blurredPixel += pf * gp[0];
177            pi += iStride;
178            gp++;
179        }
180        out->xyzw = blurredPixel;
181        x1++;
182        out++;
183        ptrIn++;
184    }
185}
186
187static void OneVFU1(float *out,
188                    const uchar *ptrIn, int iStride, const float* gPtr, int ct, int x1, int x2) {
189
190    int len = x2 - x1;
191
192    while((x2 > x1) && (((int)ptrIn) & 0x3)) {
193        const uchar *pi = ptrIn;
194        float blurredPixel = 0;
195        const float* gp = gPtr;
196
197        for (int r = 0; r < ct; r++) {
198            float pf = (float)pi[0];
199            blurredPixel += pf * gp[0];
200            pi += iStride;
201            gp++;
202        }
203        out[0] = blurredPixel;
204        x1++;
205        out++;
206        ptrIn++;
207    }
208
209#if defined(ARCH_ARM_HAVE_NEON)
210    {
211        int t = (x2 - x1) >> 2;
212        t &= ~1;
213        if(t) {
214            rsdIntrinsicBlurVFU4_K(out, ptrIn, iStride, gPtr, ct, 0, t << 2);
215            len -= t << 2;
216            ptrIn += t << 2;
217            out += t << 2;
218        }
219    }
220#endif
221
222    while(len) {
223        const uchar *pi = ptrIn;
224        float blurredPixel = 0;
225        const float* gp = gPtr;
226
227        for (int r = 0; r < ct; r++) {
228            float pf = (float)pi[0];
229            blurredPixel += pf * gp[0];
230            pi += iStride;
231            gp++;
232        }
233        out[0] = blurredPixel;
234        len--;
235        out++;
236        ptrIn++;
237    }
238}
239
240static void OneHU4(const RsForEachStubParamStruct *p, uchar4 *out, int32_t x,
241                   const float4 *ptrIn, const float* gPtr, int iradius) {
242
243    float4 blurredPixel = 0;
244    for (int r = -iradius; r <= iradius; r ++) {
245        int validX = rsMax((x + r), 0);
246        validX = rsMin(validX, (int)(p->dimX - 1));
247        float4 pf = ptrIn[validX];
248        blurredPixel += pf * gPtr[0];
249        gPtr++;
250    }
251
252    out->xyzw = convert_uchar4(blurredPixel);
253}
254
255static void OneHU1(const RsForEachStubParamStruct *p, uchar *out, int32_t x,
256                   const float *ptrIn, const float* gPtr, int iradius) {
257
258    float blurredPixel = 0;
259    for (int r = -iradius; r <= iradius; r ++) {
260        int validX = rsMax((x + r), 0);
261        validX = rsMin(validX, (int)(p->dimX - 1));
262        float pf = ptrIn[validX];
263        blurredPixel += pf * gPtr[0];
264        gPtr++;
265    }
266
267    out[0] = (uchar)blurredPixel;
268}
269
270
271void RsdCpuScriptIntrinsicBlur::kernelU4(const RsForEachStubParamStruct *p,
272                                         uint32_t xstart, uint32_t xend,
273                                         uint32_t instep, uint32_t outstep) {
274
275    float4 stackbuf[2048];
276    float4 *buf = &stackbuf[0];
277    RsdCpuScriptIntrinsicBlur *cp = (RsdCpuScriptIntrinsicBlur *)p->usr;
278    if (!cp->mAlloc.get()) {
279        ALOGE("Blur executed without input, skipping");
280        return;
281    }
282    const uchar *pin = (const uchar *)cp->mAlloc->mHal.drvState.lod[0].mallocPtr;
283    const size_t stride = cp->mAlloc->mHal.drvState.lod[0].stride;
284
285    uchar4 *out = (uchar4 *)p->out;
286    uint32_t x1 = xstart;
287    uint32_t x2 = xend;
288
289    if (p->dimX > 2048) {
290        if ((p->dimX > cp->mScratchSize[p->lid]) || !cp->mScratch[p->lid]) {
291            cp->mScratch[p->lid] = realloc(cp->mScratch[p->lid], p->dimX * 16);
292            cp->mScratchSize[p->lid] = p->dimX;
293        }
294        buf = (float4 *)cp->mScratch[p->lid];
295    }
296    float4 *fout = (float4 *)buf;
297    int y = p->y;
298    if ((y > cp->mIradius) && (y < ((int)p->dimY - cp->mIradius))) {
299        const uchar *pi = pin + (y - cp->mIradius) * stride;
300        OneVFU4(fout, pi, stride, cp->mFp, cp->mIradius * 2 + 1, x1, x2);
301    } else {
302        while(x2 > x1) {
303            OneVU4(p, fout, x1, y, pin, stride, cp->mFp, cp->mIradius);
304            fout++;
305            x1++;
306        }
307    }
308
309    x1 = xstart;
310    while ((x1 < (uint32_t)cp->mIradius) && (x1 < x2)) {
311        OneHU4(p, out, x1, buf, cp->mFp, cp->mIradius);
312        out++;
313        x1++;
314    }
315#if defined(ARCH_ARM_HAVE_NEON)
316    if ((x1 + cp->mIradius) < x2) {
317        rsdIntrinsicBlurHFU4_K(out, buf - cp->mIradius, cp->mFp,
318                               cp->mIradius * 2 + 1, x1, x2 - cp->mIradius);
319        out += (x2 - cp->mIradius) - x1;
320        x1 = x2 - cp->mIradius;
321    }
322#endif
323    while(x2 > x1) {
324        OneHU4(p, out, x1, buf, cp->mFp, cp->mIradius);
325        out++;
326        x1++;
327    }
328}
329
330void RsdCpuScriptIntrinsicBlur::kernelU1(const RsForEachStubParamStruct *p,
331                                         uint32_t xstart, uint32_t xend,
332                                         uint32_t instep, uint32_t outstep) {
333    float buf[4 * 2048];
334    RsdCpuScriptIntrinsicBlur *cp = (RsdCpuScriptIntrinsicBlur *)p->usr;
335    if (!cp->mAlloc.get()) {
336        ALOGE("Blur executed without input, skipping");
337        return;
338    }
339    const uchar *pin = (const uchar *)cp->mAlloc->mHal.drvState.lod[0].mallocPtr;
340    const size_t stride = cp->mAlloc->mHal.drvState.lod[0].stride;
341
342    uchar *out = (uchar *)p->out;
343    uint32_t x1 = xstart;
344    uint32_t x2 = xend;
345
346    float *fout = (float *)buf;
347    int y = p->y;
348    if ((y > cp->mIradius) && (y < ((int)p->dimY - cp->mIradius))) {
349        const uchar *pi = pin + (y - cp->mIradius) * stride;
350        OneVFU1(fout, pi, stride, cp->mFp, cp->mIradius * 2 + 1, x1, x2);
351    } else {
352        while(x2 > x1) {
353            OneVU1(p, fout, x1, y, pin, stride, cp->mFp, cp->mIradius);
354            fout++;
355            x1++;
356        }
357    }
358
359    x1 = xstart;
360    while ((x1 < x2) &&
361           ((x1 < (uint32_t)cp->mIradius) || (((int)out) & 0x3))) {
362        OneHU1(p, out, x1, buf, cp->mFp, cp->mIradius);
363        out++;
364        x1++;
365    }
366#if defined(ARCH_ARM_HAVE_NEON)
367    if ((x1 + cp->mIradius) < x2) {
368        uint32_t len = x2 - (x1 + cp->mIradius);
369        len &= ~3;
370        rsdIntrinsicBlurHFU1_K(out, ((float *)buf) - cp->mIradius, cp->mFp,
371                               cp->mIradius * 2 + 1, x1, x1 + len);
372        out += len;
373        x1 += len;
374    }
375#endif
376    while(x2 > x1) {
377        OneHU1(p, out, x1, buf, cp->mFp, cp->mIradius);
378        out++;
379        x1++;
380    }
381}
382
383RsdCpuScriptIntrinsicBlur::RsdCpuScriptIntrinsicBlur(RsdCpuReferenceImpl *ctx,
384                                                     const Script *s, const Element *e)
385            : RsdCpuScriptIntrinsic(ctx, s, e, RS_SCRIPT_INTRINSIC_ID_BLUR) {
386
387    mRootPtr = NULL;
388    if (e->getType() == RS_TYPE_UNSIGNED_8) {
389        switch (e->getVectorSize()) {
390        case 1:
391            mRootPtr = &kernelU1;
392            break;
393        case 4:
394            mRootPtr = &kernelU4;
395            break;
396        }
397    }
398    rsAssert(mRootPtr);
399    mRadius = 5;
400
401    mScratch = new void *[mCtx->getThreadCount()];
402    mScratchSize = new size_t[mCtx->getThreadCount()];
403
404    ComputeGaussianWeights();
405}
406
407RsdCpuScriptIntrinsicBlur::~RsdCpuScriptIntrinsicBlur() {
408    uint32_t threads = mCtx->getThreadCount();
409    if (mScratch) {
410        for (size_t i = 0; i < threads; i++) {
411            if (mScratch[i]) {
412                free(mScratch[i]);
413            }
414        }
415        delete []mScratch;
416    }
417    if (mScratchSize) {
418        delete []mScratchSize;
419    }
420}
421
422void RsdCpuScriptIntrinsicBlur::populateScript(Script *s) {
423    s->mHal.info.exportedVariableCount = 2;
424}
425
426void RsdCpuScriptIntrinsicBlur::invokeFreeChildren() {
427    mAlloc.clear();
428}
429
430
431RsdCpuScriptImpl * rsdIntrinsic_Blur(RsdCpuReferenceImpl *ctx, const Script *s, const Element *e) {
432
433    return new RsdCpuScriptIntrinsicBlur(ctx, s, e);
434}
435
436
437