rsCpuIntrinsicBlur.cpp revision 8b7117dfbcca8a74dfd83aaff7f82d2d7e3abd8f 
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    uint16_t 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    // TODO Define these constants only once
67    float e = 2.718281828459045f;
68    float pi = 3.1415926535897932f;
69    // g(x) = (1 / (sqrt(2 * pi) * sigma)) * e ^ (-x^2 / (2 * sigma^2))
70    // x is of the form [-radius .. 0 .. radius]
71    // and sigma varies with the radius.
72    // Based on some experimental radius values and sigmas,
73    // we approximately fit sigma = f(radius) as
74    // sigma = radius * 0.4  + 0.6
75    // The larger the radius gets, the more our gaussian blur
76    // will resemble a box blur since with large sigma
77    // the gaussian curve begins to lose its shape
78    float sigma = 0.4f * mRadius + 0.6f;
79
80    // Now compute the coefficients. We will store some redundant values to save
81    // some math during the blur calculations precompute some values
82    float coeff1 = 1.0f / (sqrtf(2.0f * pi) * sigma);
83    float coeff2 = - 1.0f / (2.0f * sigma * sigma);
84
85    float normalizeFactor = 0.0f;
86    float floatR = 0.0f;
87    int r;
88    mIradius = (float)ceil(mRadius) + 0.5f;
89    for (r = -mIradius; r <= mIradius; r ++) {
90        floatR = (float)r;
91        mFp[r + mIradius] = coeff1 * powf(e, floatR * floatR * coeff2);
92        normalizeFactor += mFp[r + mIradius];
93    }
94
95    // Now we need to normalize the weights because all our coefficients need to add up to one
96    normalizeFactor = 1.0f / normalizeFactor;
97    for (r = -mIradius; r <= mIradius; r ++) {
98        mFp[r + mIradius] *= normalizeFactor;
99        mIp[r + mIradius] = (uint16_t)(mFp[r + mIradius] * 65536.0f + 0.5f);
100    }
101}
102
103void RsdCpuScriptIntrinsicBlur::setGlobalObj(uint32_t slot, ObjectBase *data) {
104    rsAssert(slot == 1);
105    mAlloc.set(static_cast<Allocation *>(data));
106}
107
108void RsdCpuScriptIntrinsicBlur::setGlobalVar(uint32_t slot, const void *data, size_t dataLength) {
109    rsAssert(slot == 0);
110    mRadius = ((const float *)data)[0];
111    ComputeGaussianWeights();
112}
113
114
115
116static void OneVU4(const RsForEachStubParamStruct *p, float4 *out, int32_t x, int32_t y,
117                   const uchar *ptrIn, int iStride, const float* gPtr, int iradius) {
118
119    const uchar *pi = ptrIn + x*4;
120
121    float4 blurredPixel = 0;
122    for (int r = -iradius; r <= iradius; r ++) {
123        int validY = rsMax((y + r), 0);
124        validY = rsMin(validY, (int)(p->dimY - 1));
125        const uchar4 *pvy = (const uchar4 *)&pi[validY * iStride];
126        float4 pf = convert_float4(pvy[0]);
127        blurredPixel += pf * gPtr[0];
128        gPtr++;
129    }
130
131    out->xyzw = blurredPixel;
132}
133
134static void OneVU1(const RsForEachStubParamStruct *p, float *out, int32_t x, int32_t y,
135                   const uchar *ptrIn, int iStride, const float* gPtr, int iradius) {
136
137    const uchar *pi = ptrIn + x;
138
139    float blurredPixel = 0;
140    for (int r = -iradius; r <= iradius; r ++) {
141        int validY = rsMax((y + r), 0);
142        validY = rsMin(validY, (int)(p->dimY - 1));
143        float pf = (float)pi[validY * iStride];
144        blurredPixel += pf * gPtr[0];
145        gPtr++;
146    }
147
148    out[0] = blurredPixel;
149}
150
151
152extern "C" void rsdIntrinsicBlurU1_K(uchar *out, uchar const *in, size_t w, size_t h,
153                 size_t p, size_t x, size_t y, size_t count, size_t r, uint16_t const *tab);
154extern "C" void rsdIntrinsicBlurU4_K(uchar4 *out, uchar4 const *in, size_t w, size_t h,
155                 size_t p, size_t x, size_t y, size_t count, size_t r, uint16_t const *tab);
156
157#if defined(ARCH_X86_HAVE_SSSE3)
158extern "C" void rsdIntrinsicBlurVFU4_K(void *dst, const void *pin, int stride, const void *gptr, int rct, int x1, int ct);
159extern "C" void rsdIntrinsicBlurHFU4_K(void *dst, const void *pin, const void *gptr, int rct, int x1, int ct);
160extern "C" void rsdIntrinsicBlurHFU1_K(void *dst, const void *pin, const void *gptr, int rct, int x1, int ct);
161#endif
162
163static void OneVFU4(float4 *out,
164                    const uchar *ptrIn, int iStride, const float* gPtr, int ct,
165                    int x1, int x2) {
166#if defined(ARCH_X86_HAVE_SSSE3)
167    if (gArchUseSIMD) {
168        int t = (x2 - x1);
169        t &= ~1;
170        if (t) {
171            rsdIntrinsicBlurVFU4_K(out, ptrIn, iStride, gPtr, ct, x1, x1 + t);
172        }
173        x1 += t;
174    }
175#endif
176    while(x2 > x1) {
177        const uchar *pi = ptrIn;
178        float4 blurredPixel = 0;
179        const float* gp = gPtr;
180
181        for (int r = 0; r < ct; r++) {
182            float4 pf = convert_float4(((const uchar4 *)pi)[0]);
183            blurredPixel += pf * gp[0];
184            pi += iStride;
185            gp++;
186        }
187        out->xyzw = blurredPixel;
188        x1++;
189        out++;
190        ptrIn+=4;
191    }
192}
193
194static void OneVFU1(float *out,
195                    const uchar *ptrIn, int iStride, const float* gPtr, int ct, int x1, int x2) {
196
197    int len = x2 - x1;
198
199    while((x2 > x1) && (((uintptr_t)ptrIn) & 0x3)) {
200        const uchar *pi = ptrIn;
201        float blurredPixel = 0;
202        const float* gp = gPtr;
203
204        for (int r = 0; r < ct; r++) {
205            float pf = (float)pi[0];
206            blurredPixel += pf * gp[0];
207            pi += iStride;
208            gp++;
209        }
210        out[0] = blurredPixel;
211        x1++;
212        out++;
213        ptrIn++;
214        len--;
215    }
216#if defined(ARCH_X86_HAVE_SSSE3)
217    if (gArchUseSIMD && (x2 > x1)) {
218        int t = (x2 - x1) >> 2;
219        t &= ~1;
220        if (t) {
221            rsdIntrinsicBlurVFU4_K(out, ptrIn, iStride, gPtr, ct, 0, t );
222            len -= t << 2;
223            ptrIn += t << 2;
224            out += t << 2;
225        }
226    }
227#endif
228    while(len > 0) {
229        const uchar *pi = ptrIn;
230        float blurredPixel = 0;
231        const float* gp = gPtr;
232
233        for (int r = 0; r < ct; r++) {
234            float pf = (float)pi[0];
235            blurredPixel += pf * gp[0];
236            pi += iStride;
237            gp++;
238        }
239        out[0] = blurredPixel;
240        len--;
241        out++;
242        ptrIn++;
243    }
244}
245
246static void OneHU4(const RsForEachStubParamStruct *p, uchar4 *out, int32_t x,
247                   const float4 *ptrIn, const float* gPtr, int iradius) {
248
249    float4 blurredPixel = 0;
250    for (int r = -iradius; r <= iradius; r ++) {
251        int validX = rsMax((x + r), 0);
252        validX = rsMin(validX, (int)(p->dimX - 1));
253        float4 pf = ptrIn[validX];
254        blurredPixel += pf * gPtr[0];
255        gPtr++;
256    }
257
258    out->xyzw = convert_uchar4(blurredPixel);
259}
260
261static void OneHU1(const RsForEachStubParamStruct *p, uchar *out, int32_t x,
262                   const float *ptrIn, const float* gPtr, int iradius) {
263
264    float blurredPixel = 0;
265    for (int r = -iradius; r <= iradius; r ++) {
266        int validX = rsMax((x + r), 0);
267        validX = rsMin(validX, (int)(p->dimX - 1));
268        float pf = ptrIn[validX];
269        blurredPixel += pf * gPtr[0];
270        gPtr++;
271    }
272
273    out[0] = (uchar)blurredPixel;
274}
275
276
277void RsdCpuScriptIntrinsicBlur::kernelU4(const RsForEachStubParamStruct *p,
278                                         uint32_t xstart, uint32_t xend,
279                                         uint32_t instep, uint32_t outstep) {
280
281    float4 stackbuf[2048];
282    float4 *buf = &stackbuf[0];
283    RsdCpuScriptIntrinsicBlur *cp = (RsdCpuScriptIntrinsicBlur *)p->usr;
284    if (!cp->mAlloc.get()) {
285        ALOGE("Blur executed without input, skipping");
286        return;
287    }
288    const uchar *pin = (const uchar *)cp->mAlloc->mHal.drvState.lod[0].mallocPtr;
289    const size_t stride = cp->mAlloc->mHal.drvState.lod[0].stride;
290
291    uchar4 *out = (uchar4 *)p->out;
292    uint32_t x1 = xstart;
293    uint32_t x2 = xend;
294
295#if defined(ARCH_ARM_HAVE_VFP)
296    if (gArchUseSIMD) {
297        rsdIntrinsicBlurU4_K(out, (uchar4 const *)(pin + stride * p->y), p->dimX, p->dimY,
298                 stride, x1, p->y, x2 - x1, cp->mIradius, cp->mIp + cp->mIradius);
299        return;
300    }
301#endif
302
303    if (p->dimX > 2048) {
304        if ((p->dimX > cp->mScratchSize[p->lid]) || !cp->mScratch[p->lid]) {
305            // Pad the side of the allocation by one unit to allow alignment later
306            cp->mScratch[p->lid] = realloc(cp->mScratch[p->lid], (p->dimX + 1) * 16);
307            cp->mScratchSize[p->lid] = p->dimX;
308        }
309        // realloc only aligns to 8 bytes so we manually align to 16.
310        buf = (float4 *) ((((intptr_t)cp->mScratch[p->lid]) + 15) & ~0xf);
311    }
312    float4 *fout = (float4 *)buf;
313    int y = p->y;
314    if ((y > cp->mIradius) && (y < ((int)p->dimY - cp->mIradius))) {
315        const uchar *pi = pin + (y - cp->mIradius) * stride;
316        OneVFU4(fout, pi, stride, cp->mFp, cp->mIradius * 2 + 1, x1, x2);
317    } else {
318        while(x2 > x1) {
319            OneVU4(p, fout, x1, y, pin, stride, cp->mFp, cp->mIradius);
320            fout++;
321            x1++;
322        }
323    }
324
325    x1 = xstart;
326    while ((x1 < (uint32_t)cp->mIradius) && (x1 < x2)) {
327        OneHU4(p, out, x1, buf, cp->mFp, cp->mIradius);
328        out++;
329        x1++;
330    }
331#if defined(ARCH_X86_HAVE_SSSE3)
332    if (gArchUseSIMD) {
333        if ((x1 + cp->mIradius) < x2) {
334            rsdIntrinsicBlurHFU4_K(out, buf - cp->mIradius, cp->mFp,
335                                   cp->mIradius * 2 + 1, x1, x2 - cp->mIradius);
336            out += (x2 - cp->mIradius) - x1;
337            x1 = x2 - cp->mIradius;
338        }
339    }
340#endif
341    while(x2 > x1) {
342        OneHU4(p, out, x1, buf, cp->mFp, cp->mIradius);
343        out++;
344        x1++;
345    }
346}
347
348void RsdCpuScriptIntrinsicBlur::kernelU1(const RsForEachStubParamStruct *p,
349                                         uint32_t xstart, uint32_t xend,
350                                         uint32_t instep, uint32_t outstep) {
351    float buf[4 * 2048];
352    RsdCpuScriptIntrinsicBlur *cp = (RsdCpuScriptIntrinsicBlur *)p->usr;
353    if (!cp->mAlloc.get()) {
354        ALOGE("Blur executed without input, skipping");
355        return;
356    }
357    const uchar *pin = (const uchar *)cp->mAlloc->mHal.drvState.lod[0].mallocPtr;
358    const size_t stride = cp->mAlloc->mHal.drvState.lod[0].stride;
359
360    uchar *out = (uchar *)p->out;
361    uint32_t x1 = xstart;
362    uint32_t x2 = xend;
363
364#if defined(ARCH_ARM_HAVE_VFP)
365    if (gArchUseSIMD) {
366        rsdIntrinsicBlurU1_K(out, pin + stride * p->y, p->dimX, p->dimY,
367                 stride, x1, p->y, x2 - x1, cp->mIradius, cp->mIp + cp->mIradius);
368        return;
369    }
370#endif
371
372    float *fout = (float *)buf;
373    int y = p->y;
374    if ((y > cp->mIradius) && (y < ((int)p->dimY - cp->mIradius -1))) {
375        const uchar *pi = pin + (y - cp->mIradius) * stride;
376        OneVFU1(fout, pi, stride, cp->mFp, cp->mIradius * 2 + 1, x1, x2);
377    } else {
378        while(x2 > x1) {
379            OneVU1(p, fout, x1, y, pin, stride, cp->mFp, cp->mIradius);
380            fout++;
381            x1++;
382        }
383    }
384
385    x1 = xstart;
386    while ((x1 < x2) &&
387           ((x1 < (uint32_t)cp->mIradius) || (((uintptr_t)out) & 0x3))) {
388        OneHU1(p, out, x1, buf, cp->mFp, cp->mIradius);
389        out++;
390        x1++;
391    }
392#if defined(ARCH_X86_HAVE_SSSE3)
393    if (gArchUseSIMD) {
394        if ((x1 + cp->mIradius) < x2) {
395            uint32_t len = x2 - (x1 + cp->mIradius);
396            len &= ~3;
397            if (len > 0) {
398                rsdIntrinsicBlurHFU1_K(out, ((float *)buf) - cp->mIradius, cp->mFp,
399                                       cp->mIradius * 2 + 1, x1, x1 + len);
400                out += len;
401                x1 += len;
402            }
403        }
404    }
405#endif
406    while(x2 > x1) {
407        OneHU1(p, out, x1, buf, cp->mFp, cp->mIradius);
408        out++;
409        x1++;
410    }
411}
412
413RsdCpuScriptIntrinsicBlur::RsdCpuScriptIntrinsicBlur(RsdCpuReferenceImpl *ctx,
414                                                     const Script *s, const Element *e)
415            : RsdCpuScriptIntrinsic(ctx, s, e, RS_SCRIPT_INTRINSIC_ID_BLUR) {
416
417    mRootPtr = NULL;
418    if (e->getType() == RS_TYPE_UNSIGNED_8) {
419        switch (e->getVectorSize()) {
420        case 1:
421            mRootPtr = &kernelU1;
422            break;
423        case 4:
424            mRootPtr = &kernelU4;
425            break;
426        }
427    }
428    rsAssert(mRootPtr);
429    mRadius = 5;
430
431    mScratch = new void *[mCtx->getThreadCount()];
432    mScratchSize = new size_t[mCtx->getThreadCount()];
433    memset(mScratch, 0, sizeof(void *) * mCtx->getThreadCount());
434    memset(mScratchSize, 0, sizeof(size_t) * mCtx->getThreadCount());
435
436    ComputeGaussianWeights();
437}
438
439RsdCpuScriptIntrinsicBlur::~RsdCpuScriptIntrinsicBlur() {
440    uint32_t threads = mCtx->getThreadCount();
441    if (mScratch) {
442        for (size_t i = 0; i < threads; i++) {
443            if (mScratch[i]) {
444                free(mScratch[i]);
445            }
446        }
447        delete []mScratch;
448    }
449    if (mScratchSize) {
450        delete []mScratchSize;
451    }
452}
453
454void RsdCpuScriptIntrinsicBlur::populateScript(Script *s) {
455    s->mHal.info.exportedVariableCount = 2;
456}
457
458void RsdCpuScriptIntrinsicBlur::invokeFreeChildren() {
459    mAlloc.clear();
460}
461
462
463RsdCpuScriptImpl * rsdIntrinsic_Blur(RsdCpuReferenceImpl *ctx, const Script *s, const Element *e) {
464
465    return new RsdCpuScriptIntrinsicBlur(ctx, s, e);
466}
467
468
469