rsCpuIntrinsicBlur.cpp revision c859de1556dc9ac5c6f5647298bb5c685c5d2bb0 
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            x1 += t;
165            ptrIn += t << 2;
166            out += t;
167        }
168    }
169#endif
170
171    while(x2 > x1) {
172        const uchar *pi = ptrIn;
173        float4 blurredPixel = 0;
174        const float* gp = gPtr;
175
176        for (int r = 0; r < ct; r++) {
177            float4 pf = convert_float4(((const uchar4 *)pi)[0]);
178            blurredPixel += pf * gp[0];
179            pi += iStride;
180            gp++;
181        }
182        out->xyzw = blurredPixel;
183        x1++;
184        out++;
185        ptrIn+=4;
186    }
187}
188
189static void OneVFU1(float *out,
190                    const uchar *ptrIn, int iStride, const float* gPtr, int ct, int x1, int x2) {
191
192    int len = x2 - x1;
193
194    while((x2 > x1) && (((uintptr_t)ptrIn) & 0x3)) {
195        const uchar *pi = ptrIn;
196        float blurredPixel = 0;
197        const float* gp = gPtr;
198
199        for (int r = 0; r < ct; r++) {
200            float pf = (float)pi[0];
201            blurredPixel += pf * gp[0];
202            pi += iStride;
203            gp++;
204        }
205        out[0] = blurredPixel;
206        x1++;
207        out++;
208        ptrIn++;
209        len--;
210    }
211
212#if defined(ARCH_ARM_HAVE_VFP)
213    if (gArchUseSIMD && (x2 > x1)) {
214        int t = (x2 - x1) >> 2;
215        t &= ~1;
216        if(t) {
217            rsdIntrinsicBlurVFU4_K(out, ptrIn, iStride, gPtr, ct, 0, t );
218            len -= t << 2;
219            ptrIn += t << 2;
220            out += t << 2;
221        }
222    }
223#endif
224
225    while(len > 0) {
226        const uchar *pi = ptrIn;
227        float blurredPixel = 0;
228        const float* gp = gPtr;
229
230        for (int r = 0; r < ct; r++) {
231            float pf = (float)pi[0];
232            blurredPixel += pf * gp[0];
233            pi += iStride;
234            gp++;
235        }
236        out[0] = blurredPixel;
237        len--;
238        out++;
239        ptrIn++;
240    }
241}
242
243static void OneHU4(const RsForEachStubParamStruct *p, uchar4 *out, int32_t x,
244                   const float4 *ptrIn, const float* gPtr, int iradius) {
245
246    float4 blurredPixel = 0;
247    for (int r = -iradius; r <= iradius; r ++) {
248        int validX = rsMax((x + r), 0);
249        validX = rsMin(validX, (int)(p->dimX - 1));
250        float4 pf = ptrIn[validX];
251        blurredPixel += pf * gPtr[0];
252        gPtr++;
253    }
254
255    out->xyzw = convert_uchar4(blurredPixel);
256}
257
258static void OneHU1(const RsForEachStubParamStruct *p, uchar *out, int32_t x,
259                   const float *ptrIn, const float* gPtr, int iradius) {
260
261    float blurredPixel = 0;
262    for (int r = -iradius; r <= iradius; r ++) {
263        int validX = rsMax((x + r), 0);
264        validX = rsMin(validX, (int)(p->dimX - 1));
265        float pf = ptrIn[validX];
266        blurredPixel += pf * gPtr[0];
267        gPtr++;
268    }
269
270    out[0] = (uchar)blurredPixel;
271}
272
273
274void RsdCpuScriptIntrinsicBlur::kernelU4(const RsForEachStubParamStruct *p,
275                                         uint32_t xstart, uint32_t xend,
276                                         uint32_t instep, uint32_t outstep) {
277
278    float4 stackbuf[2048];
279    float4 *buf = &stackbuf[0];
280    RsdCpuScriptIntrinsicBlur *cp = (RsdCpuScriptIntrinsicBlur *)p->usr;
281    if (!cp->mAlloc.get()) {
282        ALOGE("Blur 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    uchar4 *out = (uchar4 *)p->out;
289    uint32_t x1 = xstart;
290    uint32_t x2 = xend;
291
292    if (p->dimX > 2048) {
293        if ((p->dimX > cp->mScratchSize[p->lid]) || !cp->mScratch[p->lid]) {
294            // Pad the side of the allocation by one unit to allow alignment later
295            cp->mScratch[p->lid] = realloc(cp->mScratch[p->lid], (p->dimX + 1) * 16);
296            cp->mScratchSize[p->lid] = p->dimX;
297        }
298        // realloc only aligns to 8 bytes so we manually align to 16.
299        buf = (float4 *) ((((intptr_t)cp->mScratch[p->lid]) + 15) & ~0xf);
300    }
301    float4 *fout = (float4 *)buf;
302    int y = p->y;
303    if ((y > cp->mIradius) && (y < ((int)p->dimY - cp->mIradius))) {
304        const uchar *pi = pin + (y - cp->mIradius) * stride;
305        OneVFU4(fout, pi, stride, cp->mFp, cp->mIradius * 2 + 1, x1, x2);
306    } else {
307        while(x2 > x1) {
308            OneVU4(p, fout, x1, y, pin, stride, cp->mFp, cp->mIradius);
309            fout++;
310            x1++;
311        }
312    }
313
314    x1 = xstart;
315    while ((x1 < (uint32_t)cp->mIradius) && (x1 < x2)) {
316        OneHU4(p, out, x1, buf, cp->mFp, cp->mIradius);
317        out++;
318        x1++;
319    }
320#if defined(ARCH_ARM_HAVE_VFP)
321    if (gArchUseSIMD) {
322        if ((x1 + cp->mIradius) < x2) {
323            rsdIntrinsicBlurHFU4_K(out, buf - cp->mIradius, cp->mFp,
324                                   cp->mIradius * 2 + 1, x1, x2 - cp->mIradius);
325            out += (x2 - cp->mIradius) - x1;
326            x1 = x2 - cp->mIradius;
327        }
328    }
329#endif
330    while(x2 > x1) {
331        OneHU4(p, out, x1, buf, cp->mFp, cp->mIradius);
332        out++;
333        x1++;
334    }
335}
336
337void RsdCpuScriptIntrinsicBlur::kernelU1(const RsForEachStubParamStruct *p,
338                                         uint32_t xstart, uint32_t xend,
339                                         uint32_t instep, uint32_t outstep) {
340    float buf[4 * 2048];
341    RsdCpuScriptIntrinsicBlur *cp = (RsdCpuScriptIntrinsicBlur *)p->usr;
342    if (!cp->mAlloc.get()) {
343        ALOGE("Blur executed without input, skipping");
344        return;
345    }
346    const uchar *pin = (const uchar *)cp->mAlloc->mHal.drvState.lod[0].mallocPtr;
347    const size_t stride = cp->mAlloc->mHal.drvState.lod[0].stride;
348
349    uchar *out = (uchar *)p->out;
350    uint32_t x1 = xstart;
351    uint32_t x2 = xend;
352
353    float *fout = (float *)buf;
354    int y = p->y;
355    if ((y > cp->mIradius) && (y < ((int)p->dimY - cp->mIradius -1))) {
356        const uchar *pi = pin + (y - cp->mIradius) * stride;
357        OneVFU1(fout, pi, stride, cp->mFp, cp->mIradius * 2 + 1, x1, x2);
358    } else {
359        while(x2 > x1) {
360            OneVU1(p, fout, x1, y, pin, stride, cp->mFp, cp->mIradius);
361            fout++;
362            x1++;
363        }
364    }
365
366    x1 = xstart;
367    while ((x1 < x2) &&
368           ((x1 < (uint32_t)cp->mIradius) || (((uintptr_t)out) & 0x3))) {
369        OneHU1(p, out, x1, buf, cp->mFp, cp->mIradius);
370        out++;
371        x1++;
372    }
373#if defined(ARCH_ARM_HAVE_VFP)
374    if (gArchUseSIMD) {
375        if ((x1 + cp->mIradius) < x2) {
376            uint32_t len = x2 - (x1 + cp->mIradius);
377            len &= ~3;
378            if (len > 0) {
379                rsdIntrinsicBlurHFU1_K(out, ((float *)buf) - cp->mIradius, cp->mFp,
380                                       cp->mIradius * 2 + 1, x1, x1 + len);
381                out += len;
382                x1 += len;
383            }
384        }
385    }
386#endif
387    while(x2 > x1) {
388        OneHU1(p, out, x1, buf, cp->mFp, cp->mIradius);
389        out++;
390        x1++;
391    }
392}
393
394RsdCpuScriptIntrinsicBlur::RsdCpuScriptIntrinsicBlur(RsdCpuReferenceImpl *ctx,
395                                                     const Script *s, const Element *e)
396            : RsdCpuScriptIntrinsic(ctx, s, e, RS_SCRIPT_INTRINSIC_ID_BLUR) {
397
398    mRootPtr = NULL;
399    if (e->getType() == RS_TYPE_UNSIGNED_8) {
400        switch (e->getVectorSize()) {
401        case 1:
402            mRootPtr = &kernelU1;
403            break;
404        case 4:
405            mRootPtr = &kernelU4;
406            break;
407        }
408    }
409    rsAssert(mRootPtr);
410    mRadius = 5;
411
412    mScratch = new void *[mCtx->getThreadCount()];
413    mScratchSize = new size_t[mCtx->getThreadCount()];
414    memset(mScratch, 0, sizeof(void *) * mCtx->getThreadCount());
415    memset(mScratchSize, 0, sizeof(size_t) * mCtx->getThreadCount());
416
417    ComputeGaussianWeights();
418}
419
420RsdCpuScriptIntrinsicBlur::~RsdCpuScriptIntrinsicBlur() {
421    uint32_t threads = mCtx->getThreadCount();
422    if (mScratch) {
423        for (size_t i = 0; i < threads; i++) {
424            if (mScratch[i]) {
425                free(mScratch[i]);
426            }
427        }
428        delete []mScratch;
429    }
430    if (mScratchSize) {
431        delete []mScratchSize;
432    }
433}
434
435void RsdCpuScriptIntrinsicBlur::populateScript(Script *s) {
436    s->mHal.info.exportedVariableCount = 2;
437}
438
439void RsdCpuScriptIntrinsicBlur::invokeFreeChildren() {
440    mAlloc.clear();
441}
442
443
444RsdCpuScriptImpl * rsdIntrinsic_Blur(RsdCpuReferenceImpl *ctx, const Script *s, const Element *e) {
445
446    return new RsdCpuScriptIntrinsicBlur(ctx, s, e);
447}
448
449
450