295     RsdCpuScriptIntrinsicBlur *cp = (RsdCpuScriptIntrinsicBlur *)info->usr;  in kernelU4()  local
296     if (!cp->mAlloc.get()) {  in kernelU4()
300     const uchar *pin = (const uchar *)cp->mAlloc->mHal.drvState.lod[0].mallocPtr;  in kernelU4()
301     const size_t stride = cp->mAlloc->mHal.drvState.lod[0].stride;  in kernelU4()
311                  stride, x1, info->current.y, x2 - x1, cp->mIradius, cp->mIp + cp->mIradius);  in kernelU4()
317         if ((info->dim.x > cp->mScratchSize[info->lid]) || !cp->mScratch[info->lid]) {  in kernelU4()
319             cp->mScratch[info->lid] = realloc(cp->mScratch[info->lid], (info->dim.x + 1) * 16);  in kernelU4()
320             cp->mScratchSize[info->lid] = info->dim.x;  in kernelU4()
323         buf = (float4 *) ((((intptr_t)cp->mScratch[info->lid]) + 15) & ~0xf);  in kernelU4()
327     if ((y > cp->mIradius) && (y < ((int)info->dim.y - cp->mIradius))) {  in kernelU4()
328         const uchar *pi = pin + (y - cp->mIradius) * stride;  in kernelU4()
329         OneVFU4(fout, pi, stride, cp->mFp, cp->mIradius * 2 + 1, 0, info->dim.x);  in kernelU4()
333             OneVU4(info, fout, x1, y, pin, stride, cp->mFp, cp->mIradius);  in kernelU4()
340     while ((x1 < (uint32_t)cp->mIradius) && (x1 < x2)) {  in kernelU4()
341         OneHU4(info, out, x1, buf, cp->mFp, cp->mIradius);  in kernelU4()
347         if ((x1 + cp->mIradius) < x2) {  in kernelU4()
348             rsdIntrinsicBlurHFU4_K(out, buf - cp->mIradius, cp->mFp,  in kernelU4()
349                                    cp->mIradius * 2 + 1, x1, x2 - cp->mIradius);  in kernelU4()
350             out += (x2 - cp->mIradius) - x1;  in kernelU4()
351             x1 = x2 - cp->mIradius;  in kernelU4()
356         OneHU4(info, out, x1, buf, cp->mFp, cp->mIradius);  in kernelU4()
366     RsdCpuScriptIntrinsicBlur *cp = (RsdCpuScriptIntrinsicBlur *)info->usr;  in kernelU1()  local
367     if (!cp->mAlloc.get()) {  in kernelU1()
371     const uchar *pin = (const uchar *)cp->mAlloc->mHal.drvState.lod[0].mallocPtr;  in kernelU1()
372     const size_t stride = cp->mAlloc->mHal.drvState.lod[0].stride;  in kernelU1()
383         if (cp->mIradius > 8 || (info->dim.x - rsMax(0, (int32_t)x1 - 8)) >= 16) {  in kernelU1()
385                      stride, x1, info->current.y, x2 - x1, cp->mIradius, cp->mIp + cp->mIradius);  in kernelU1()
393     if ((y > cp->mIradius) && (y < ((int)info->dim.y - cp->mIradius -1))) {  in kernelU1()
394         const uchar *pi = pin + (y - cp->mIradius) * stride;  in kernelU1()
395         OneVFU1(fout, pi, stride, cp->mFp, cp->mIradius * 2 + 1, 0, info->dim.x);  in kernelU1()
399             OneVU1(info, fout, x1, y, pin, stride, cp->mFp, cp->mIradius);  in kernelU1()
407            ((x1 < (uint32_t)cp->mIradius) || (((uintptr_t)out) & 0x3))) {  in kernelU1()
408         OneHU1(info, out, x1, buf, cp->mFp, cp->mIradius);  in kernelU1()
414         if ((x1 + cp->mIradius) < x2) {  in kernelU1()
415             uint32_t len = x2 - (x1 + cp->mIradius);  in kernelU1()
423                 rsdIntrinsicBlurHFU1_K(out, ((float *)buf) - cp->mIradius, cp->mFp,  in kernelU1()
424                                        cp->mIradius * 2 + 1, x1, x1 + len);  in kernelU1()
432         OneHU1(info, out, x1, buf, cp->mFp, cp->mIradius);  in kernelU1()