rsCpuIntrinsicBlur.cpp revision 75adb8213f045bf3ffbc5deb1350b36d486e228a 
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_VFP)
159    if (gArchUseSIMD) {
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+=4;
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) && (((uintptr_t)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        len--;
208    }
209
210#if defined(ARCH_ARM_HAVE_VFP)
211    if (gArchUseSIMD && (x2 > x1)) {
212        int t = (x2 - x1) >> 2;
213        t &= ~1;
214        if(t) {
215            rsdIntrinsicBlurVFU4_K(out, ptrIn, iStride, gPtr, ct, 0, t );
216            len -= t << 2;
217            ptrIn += t << 2;
218            out += t << 2;
219        }
220    }
221#endif
222
223    while(len > 0) {
224        const uchar *pi = ptrIn;
225        float blurredPixel = 0;
226        const float* gp = gPtr;
227
228        for (int r = 0; r < ct; r++) {
229            float pf = (float)pi[0];
230            blurredPixel += pf * gp[0];
231            pi += iStride;
232            gp++;
233        }
234        out[0] = blurredPixel;
235        len--;
236        out++;
237        ptrIn++;
238    }
239}
240
241static void OneHU4(const RsForEachStubParamStruct *p, uchar4 *out, int32_t x,
242                   const float4 *ptrIn, const float* gPtr, int iradius) {
243
244    float4 blurredPixel = 0;
245    for (int r = -iradius; r <= iradius; r ++) {
246        int validX = rsMax((x + r), 0);
247        validX = rsMin(validX, (int)(p->dimX - 1));
248        float4 pf = ptrIn[validX];
249        blurredPixel += pf * gPtr[0];
250        gPtr++;
251    }
252
253    out->xyzw = convert_uchar4(blurredPixel);
254}
255
256static void OneHU1(const RsForEachStubParamStruct *p, uchar *out, int32_t x,
257                   const float *ptrIn, const float* gPtr, int iradius) {
258
259    float blurredPixel = 0;
260    for (int r = -iradius; r <= iradius; r ++) {
261        int validX = rsMax((x + r), 0);
262        validX = rsMin(validX, (int)(p->dimX - 1));
263        float pf = ptrIn[validX];
264        blurredPixel += pf * gPtr[0];
265        gPtr++;
266    }
267
268    out[0] = (uchar)blurredPixel;
269}
270
271
272void RsdCpuScriptIntrinsicBlur::kernelU4(const RsForEachStubParamStruct *p,
273                                         uint32_t xstart, uint32_t xend,
274                                         uint32_t instep, uint32_t outstep) {
275
276    float4 stackbuf[2048];
277    float4 *buf = &stackbuf[0];
278    RsdCpuScriptIntrinsicBlur *cp = (RsdCpuScriptIntrinsicBlur *)p->usr;
279    if (!cp->mAlloc.get()) {
280        ALOGE("Blur executed without input, skipping");
281        return;
282    }
283    const uchar *pin = (const uchar *)cp->mAlloc->mHal.drvState.lod[0].mallocPtr;
284    const size_t stride = cp->mAlloc->mHal.drvState.lod[0].stride;
285
286    uchar4 *out = (uchar4 *)p->out;
287    uint32_t x1 = xstart;
288    uint32_t x2 = xend;
289
290    if (p->dimX > 2048) {
291        if ((p->dimX > cp->mScratchSize[p->lid]) || !cp->mScratch[p->lid]) {
292            // Pad the side of the allocation by one unit to allow alignment later
293            cp->mScratch[p->lid] = realloc(cp->mScratch[p->lid], (p->dimX + 1) * 16);
294            cp->mScratchSize[p->lid] = p->dimX;
295        }
296        // realloc only aligns to 8 bytes so we manually align to 16.
297        buf = (float4 *) ((((intptr_t)cp->mScratch[p->lid]) + 15) & ~0xf);
298    }
299    float4 *fout = (float4 *)buf;
300    int y = p->y;
301    if ((y > cp->mIradius) && (y < ((int)p->dimY - cp->mIradius))) {
302        const uchar *pi = pin + (y - cp->mIradius) * stride;
303        OneVFU4(fout, pi, stride, cp->mFp, cp->mIradius * 2 + 1, x1, x2);
304    } else {
305        while(x2 > x1) {
306            OneVU4(p, fout, x1, y, pin, stride, cp->mFp, cp->mIradius);
307            fout++;
308            x1++;
309        }
310    }
311
312    x1 = xstart;
313    while ((x1 < (uint32_t)cp->mIradius) && (x1 < x2)) {
314        OneHU4(p, out, x1, buf, cp->mFp, cp->mIradius);
315        out++;
316        x1++;
317    }
318#if defined(ARCH_ARM_HAVE_VFP)
319    if (gArchUseSIMD) {
320        if ((x1 + cp->mIradius) < x2) {
321            rsdIntrinsicBlurHFU4_K(out, buf - cp->mIradius, cp->mFp,
322                                   cp->mIradius * 2 + 1, x1, x2 - cp->mIradius);
323            out += (x2 - cp->mIradius) - x1;
324            x1 = x2 - cp->mIradius;
325        }
326    }
327#endif
328    while(x2 > x1) {
329        OneHU4(p, out, x1, buf, cp->mFp, cp->mIradius);
330        out++;
331        x1++;
332    }
333}
334
335void RsdCpuScriptIntrinsicBlur::kernelU1(const RsForEachStubParamStruct *p,
336                                         uint32_t xstart, uint32_t xend,
337                                         uint32_t instep, uint32_t outstep) {
338    float buf[4 * 2048];
339    RsdCpuScriptIntrinsicBlur *cp = (RsdCpuScriptIntrinsicBlur *)p->usr;
340    if (!cp->mAlloc.get()) {
341        ALOGE("Blur executed without input, skipping");
342        return;
343    }
344    const uchar *pin = (const uchar *)cp->mAlloc->mHal.drvState.lod[0].mallocPtr;
345    const size_t stride = cp->mAlloc->mHal.drvState.lod[0].stride;
346
347    uchar *out = (uchar *)p->out;
348    uint32_t x1 = xstart;
349    uint32_t x2 = xend;
350
351    float *fout = (float *)buf;
352    int y = p->y;
353    if ((y > cp->mIradius) && (y < ((int)p->dimY - cp->mIradius -1))) {
354        const uchar *pi = pin + (y - cp->mIradius) * stride;
355        OneVFU1(fout, pi, stride, cp->mFp, cp->mIradius * 2 + 1, x1, x2);
356    } else {
357        while(x2 > x1) {
358            OneVU1(p, fout, x1, y, pin, stride, cp->mFp, cp->mIradius);
359            fout++;
360            x1++;
361        }
362    }
363
364    x1 = xstart;
365    while ((x1 < x2) &&
366           ((x1 < (uint32_t)cp->mIradius) || (((uintptr_t)out) & 0x3))) {
367        OneHU1(p, out, x1, buf, cp->mFp, cp->mIradius);
368        out++;
369        x1++;
370    }
371#if defined(ARCH_ARM_HAVE_VFP)
372    if (gArchUseSIMD) {
373        if ((x1 + cp->mIradius) < x2) {
374            uint32_t len = x2 - (x1 + cp->mIradius);
375            len &= ~3;
376            if (len > 0) {
377                rsdIntrinsicBlurHFU1_K(out, ((float *)buf) - cp->mIradius, cp->mFp,
378                                       cp->mIradius * 2 + 1, x1, x1 + len);
379                out += len;
380                x1 += len;
381            }
382        }
383    }
384#endif
385    while(x2 > x1) {
386        OneHU1(p, out, x1, buf, cp->mFp, cp->mIradius);
387        out++;
388        x1++;
389    }
390}
391
392RsdCpuScriptIntrinsicBlur::RsdCpuScriptIntrinsicBlur(RsdCpuReferenceImpl *ctx,
393                                                     const Script *s, const Element *e)
394            : RsdCpuScriptIntrinsic(ctx, s, e, RS_SCRIPT_INTRINSIC_ID_BLUR) {
395
396    mRootPtr = NULL;
397    if (e->getType() == RS_TYPE_UNSIGNED_8) {
398        switch (e->getVectorSize()) {
399        case 1:
400            mRootPtr = &kernelU1;
401            break;
402        case 4:
403            mRootPtr = &kernelU4;
404            break;
405        }
406    }
407    rsAssert(mRootPtr);
408    mRadius = 5;
409
410    mScratch = new void *[mCtx->getThreadCount()];
411    mScratchSize = new size_t[mCtx->getThreadCount()];
412    memset(mScratch, 0, sizeof(void *) * mCtx->getThreadCount());
413    memset(mScratchSize, 0, sizeof(size_t) * mCtx->getThreadCount());
414
415    ComputeGaussianWeights();
416}
417
418RsdCpuScriptIntrinsicBlur::~RsdCpuScriptIntrinsicBlur() {
419    uint32_t threads = mCtx->getThreadCount();
420    if (mScratch) {
421        for (size_t i = 0; i < threads; i++) {
422            if (mScratch[i]) {
423                free(mScratch[i]);
424            }
425        }
426        delete []mScratch;
427    }
428    if (mScratchSize) {
429        delete []mScratchSize;
430    }
431}
432
433void RsdCpuScriptIntrinsicBlur::populateScript(Script *s) {
434    s->mHal.info.exportedVariableCount = 2;
435}
436
437void RsdCpuScriptIntrinsicBlur::invokeFreeChildren() {
438    mAlloc.clear();
439}
440
441
442RsdCpuScriptImpl * rsdIntrinsic_Blur(RsdCpuReferenceImpl *ctx, const Script *s, const Element *e) {
443
444    return new RsdCpuScriptIntrinsicBlur(ctx, s, e);
445}
446
447
448