1e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams/*
2e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams * Copyright (C) 2012 The Android Open Source Project
3e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams *
4e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams * Licensed under the Apache License, Version 2.0 (the "License");
5e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams * you may not use this file except in compliance with the License.
6e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams * You may obtain a copy of the License at
7e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams *
8e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams *      http://www.apache.org/licenses/LICENSE-2.0
9e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams *
10e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams * Unless required by applicable law or agreed to in writing, software
11e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams * distributed under the License is distributed on an "AS IS" BASIS,
12e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams * See the License for the specific language governing permissions and
14e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams * limitations under the License.
15e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams */
16e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
17e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
18e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams#include "rsdCore.h"
19e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams#include "rsdIntrinsics.h"
20e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams#include "rsdAllocation.h"
21e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
22e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams#include "rsdIntrinsicInlines.h"
23e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
24e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Samsusing namespace android;
25e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Samsusing namespace android::renderscript;
26e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
27e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Samsstruct ConvolveParams {
28e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    float fp[16];
29e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    short ip[16];
30e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    ObjectBaseRef<Allocation> alloc;
31e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams};
32e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
33e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Samsstatic void Convolve3x3_Bind(const Context *dc, const Script *script,
34e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams                             void * intrinsicData, uint32_t slot, Allocation *data) {
35e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    ConvolveParams *cp = (ConvolveParams *)intrinsicData;
36e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    rsAssert(slot == 1);
37e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    cp->alloc.set(data);
38e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams}
39e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
40e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Samsstatic void Convolve3x3_SetVar(const Context *dc, const Script *script, void * intrinsicData,
41e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams                               uint32_t slot, void *data, size_t dataLength) {
42e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    ConvolveParams *cp = (ConvolveParams *)intrinsicData;
43e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
44e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    rsAssert(slot == 0);
45e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    memcpy (cp->fp, data, dataLength);
46e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    for(int ct=0; ct < 9; ct++) {
47e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams        cp->ip[ct] = (short)(cp->fp[ct] * 255.f + 0.5f);
48e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    }
49e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams}
50e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
51e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Samsextern "C" void rsdIntrinsicConvolve3x3_K(void *dst, const void *y0, const void *y1, const void *y2, const short *coef, uint32_t count);
52e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
53e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
54e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Samsstatic void ConvolveOne(const RsForEachStubParamStruct *p, uint32_t x, uchar4 *out,
55e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams                        const uchar4 *py0, const uchar4 *py1, const uchar4 *py2,
56e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams                        const float* coeff) {
57e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
58d85e283087ecd00760a0d8d0c9d8482cda845efcJason Sams    uint32_t x1 = rsMax((int32_t)x-1, 0);
598575a2b6c601357cf34deb80b801228c6bc72a66Tim Murray    uint32_t x2 = rsMin((int32_t)x+1, (int32_t)p->dimX-1);
60e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
61e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    float4 px = convert_float4(py0[x1]) * coeff[0] +
62e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams                convert_float4(py0[x]) * coeff[1] +
63e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams                convert_float4(py0[x2]) * coeff[2] +
64e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams                convert_float4(py1[x1]) * coeff[3] +
65e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams                convert_float4(py1[x]) * coeff[4] +
66e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams                convert_float4(py1[x2]) * coeff[5] +
67e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams                convert_float4(py2[x1]) * coeff[6] +
68e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams                convert_float4(py2[x]) * coeff[7] +
69e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams                convert_float4(py2[x2]) * coeff[8];
70e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
71d85e283087ecd00760a0d8d0c9d8482cda845efcJason Sams    px = clamp(px, 0.f, 255.f);
72e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    uchar4 o = {(uchar)px.x, (uchar)px.y, (uchar)px.z, (uchar)px.w};
73e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    *out = o;
74e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams}
75e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
76e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Samsstatic void Convolve3x3_uchar4(const RsForEachStubParamStruct *p,
77e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams                                    uint32_t xstart, uint32_t xend,
78e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams                                    uint32_t instep, uint32_t outstep) {
79e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    ConvolveParams *cp = (ConvolveParams *)p->usr;
80b801b949e286275b5d19a33135235ba68d3a19a9Jason Sams
81b801b949e286275b5d19a33135235ba68d3a19a9Jason Sams    if (!cp->alloc.get()) {
82b801b949e286275b5d19a33135235ba68d3a19a9Jason Sams        ALOGE("Convolve3x3 executed without input, skipping");
83b801b949e286275b5d19a33135235ba68d3a19a9Jason Sams        return;
84b801b949e286275b5d19a33135235ba68d3a19a9Jason Sams    }
85e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    DrvAllocation *din = (DrvAllocation *)cp->alloc->mHal.drv;
86e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    const uchar *pin = (const uchar *)din->lod[0].mallocPtr;
87e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
88e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    uint32_t y1 = rsMin((int32_t)p->y + 1, (int32_t)(p->dimY-1));
89e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    uint32_t y2 = rsMax((int32_t)p->y - 1, 0);
90e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    const uchar4 *py0 = (const uchar4 *)(pin + din->lod[0].stride * y2);
91e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    const uchar4 *py1 = (const uchar4 *)(pin + din->lod[0].stride * p->y);
92e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    const uchar4 *py2 = (const uchar4 *)(pin + din->lod[0].stride * y1);
93e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
94e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    uchar4 *out = (uchar4 *)p->out;
95e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    uint32_t x1 = xstart;
96e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    uint32_t x2 = xend;
97e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    if(x1 == 0) {
98e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams        ConvolveOne(p, 0, out, py0, py1, py2, cp->fp);
99e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams        x1 ++;
100e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams        out++;
101e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    }
102e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
103e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    if(x2 > x1) {
104e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams#if defined(ARCH_ARM_HAVE_NEON)
105e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams        int32_t len = (x2 - x1 - 1) >> 1;
106e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams        if(len > 0) {
10740945e01597adaed9e728a14a17bf4a35452abd5Jason Sams            rsdIntrinsicConvolve3x3_K(out, &py0[x1-1], &py1[x1-1], &py2[x1-1], cp->ip, len);
108e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams            x1 += len << 1;
109e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams            out += len << 1;
110e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams        }
111e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams#endif
112e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
113e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams        while(x1 != x2) {
114e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams            ConvolveOne(p, x1, out, py0, py1, py2, cp->fp);
115e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams            out++;
116e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams            x1++;
117e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams        }
118e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    }
119e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams}
120e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
121e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Samsvoid * rsdIntrinsic_InitConvolve3x3(const android::renderscript::Context *dc,
122e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams                                    android::renderscript::Script *script,
123e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams                                    RsdIntriniscFuncs_t *funcs) {
124e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
125e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    script->mHal.info.exportedVariableCount = 2;
12678b050ea1e13108110ce3b8ead63252e5ebe2468Jason Sams    funcs->setVarObj = Convolve3x3_Bind;
127e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    funcs->setVar = Convolve3x3_SetVar;
128e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    funcs->root = Convolve3x3_uchar4;
129e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
130e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    ConvolveParams *cp = (ConvolveParams *)calloc(1, sizeof(ConvolveParams));
131e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    for(int ct=0; ct < 9; ct++) {
132e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams        cp->fp[ct] = 1.f / 9.f;
133e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams        cp->ip[ct] = (short)(cp->fp[ct] * 255.f + 0.5f);
134e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    }
135e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams    return cp;
136e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams}
137e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
138e1e08b4c9cc80c51224fdaf3aeab0804daf073e6Jason Sams
139