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
18#include "rsdCore.h"
19#include "rsdIntrinsics.h"
20#include "rsdAllocation.h"
21
22#include "rsdIntrinsicInlines.h"
23
24using namespace android;
25using namespace android::renderscript;
26
27struct ConvolveParams {
28    float fp[28];
29    short ip[28];
30    ObjectBaseRef<Allocation> alloc;
31};
32
33static void Convolve5x5_Bind(const Context *dc, const Script *script,
34                             void * intrinsicData, uint32_t slot, Allocation *data) {
35    ConvolveParams *cp = (ConvolveParams *)intrinsicData;
36    rsAssert(slot == 1);
37    cp->alloc.set(data);
38}
39
40static void Convolve5x5_SetVar(const Context *dc, const Script *script, void * intrinsicData,
41                               uint32_t slot, void *data, size_t dataLength) {
42    ConvolveParams *cp = (ConvolveParams *)intrinsicData;
43
44    rsAssert(slot == 0);
45    memcpy (cp->fp, data, dataLength);
46    for(int ct=0; ct < 25; ct++) {
47        cp->ip[ct] = (short)(cp->fp[ct] * 255.f + 0.5f);
48    }
49}
50
51
52static void One(const RsForEachStubParamStruct *p, uint32_t x, uchar4 *out,
53                const uchar4 *py0, const uchar4 *py1, const uchar4 *py2, const uchar4 *py3, const uchar4 *py4,
54                const float* coeff) {
55
56    uint32_t x0 = rsMax((int32_t)x-2, 0);
57    uint32_t x1 = rsMax((int32_t)x-1, 0);
58    uint32_t x2 = x;
59    uint32_t x3 = rsMin((int32_t)x+1, (int32_t)(p->dimX-1));
60    uint32_t x4 = rsMin((int32_t)x+2, (int32_t)(p->dimX-1));
61
62    float4 px = convert_float4(py0[x0]) * coeff[0] +
63                convert_float4(py0[x1]) * coeff[1] +
64                convert_float4(py0[x2]) * coeff[2] +
65                convert_float4(py0[x3]) * coeff[3] +
66                convert_float4(py0[x4]) * coeff[4] +
67
68                convert_float4(py1[x0]) * coeff[5] +
69                convert_float4(py1[x1]) * coeff[6] +
70                convert_float4(py1[x2]) * coeff[7] +
71                convert_float4(py1[x3]) * coeff[8] +
72                convert_float4(py1[x4]) * coeff[9] +
73
74                convert_float4(py2[x0]) * coeff[10] +
75                convert_float4(py2[x1]) * coeff[11] +
76                convert_float4(py2[x2]) * coeff[12] +
77                convert_float4(py2[x3]) * coeff[13] +
78                convert_float4(py2[x4]) * coeff[14] +
79
80                convert_float4(py3[x0]) * coeff[15] +
81                convert_float4(py3[x1]) * coeff[16] +
82                convert_float4(py3[x2]) * coeff[17] +
83                convert_float4(py3[x3]) * coeff[18] +
84                convert_float4(py3[x4]) * coeff[19] +
85
86                convert_float4(py4[x0]) * coeff[20] +
87                convert_float4(py4[x1]) * coeff[21] +
88                convert_float4(py4[x2]) * coeff[22] +
89                convert_float4(py4[x3]) * coeff[23] +
90                convert_float4(py4[x4]) * coeff[24];
91
92    px = clamp(px, 0.f, 255.f);
93    uchar4 o = {(uchar)px.x, (uchar)px.y, (uchar)px.z, (uchar)px.w};
94    *out = o;
95}
96
97extern "C" void rsdIntrinsicConvolve5x5_K(void *dst, const void *y0, const void *y1,
98                                          const void *y2, const void *y3, const void *y4,
99                                          const short *coef, uint32_t count);
100
101static void Convolve5x5_uchar4(const RsForEachStubParamStruct *p,
102                                    uint32_t xstart, uint32_t xend,
103                                    uint32_t instep, uint32_t outstep) {
104    ConvolveParams *cp = (ConvolveParams *)p->usr;
105    DrvAllocation *din = (DrvAllocation *)cp->alloc->mHal.drv;
106    const uchar *pin = (const uchar *)din->lod[0].mallocPtr;
107
108    uint32_t y0 = rsMax((int32_t)p->y-2, 0);
109    uint32_t y1 = rsMax((int32_t)p->y-1, 0);
110    uint32_t y2 = p->y;
111    uint32_t y3 = rsMin((int32_t)p->y+1, (int32_t)(p->dimY-1));
112    uint32_t y4 = rsMin((int32_t)p->y+2, (int32_t)(p->dimY-1));
113
114    const uchar4 *py0 = (const uchar4 *)(pin + din->lod[0].stride * y0);
115    const uchar4 *py1 = (const uchar4 *)(pin + din->lod[0].stride * y1);
116    const uchar4 *py2 = (const uchar4 *)(pin + din->lod[0].stride * y2);
117    const uchar4 *py3 = (const uchar4 *)(pin + din->lod[0].stride * y3);
118    const uchar4 *py4 = (const uchar4 *)(pin + din->lod[0].stride * y4);
119
120    uchar4 *out = (uchar4 *)p->out;
121    uint32_t x1 = xstart;
122    uint32_t x2 = xend;
123
124    while((x1 < x2) && (x1 < 2)) {
125        One(p, x1, out, py0, py1, py2, py3, py4, cp->fp);
126        out++;
127        x1++;
128    }
129
130#if defined(ARCH_ARM_HAVE_NEON)
131    if((x1 + 3) < x2) {
132        uint32_t len = (x2 - x1 - 3) >> 1;
133        rsdIntrinsicConvolve5x5_K(out, py0, py1, py2, py3, py4, cp->ip, len);
134        out += len << 1;
135        x1 += len << 1;
136    }
137#endif
138
139    while(x1 < x2) {
140        One(p, x1, out, py0, py1, py2, py3, py4, cp->fp);
141        out++;
142        x1++;
143    }
144}
145
146void * rsdIntrinsic_InitConvolve5x5(const android::renderscript::Context *dc,
147                                    android::renderscript::Script *script,
148                                    RsdIntriniscFuncs_t *funcs) {
149
150    script->mHal.info.exportedVariableCount = 2;
151    funcs->bind = Convolve5x5_Bind;
152    funcs->setVar = Convolve5x5_SetVar;
153    funcs->root = Convolve5x5_uchar4;
154
155    ConvolveParams *cp = (ConvolveParams *)calloc(1, sizeof(ConvolveParams));
156    for(int ct=0; ct < 25; ct++) {
157        cp->fp[ct] = 1.f / 25.f;
158        cp->ip[ct] = (short)(cp->fp[ct] * 255.f + 0.5f);
159    }
160    return cp;
161}
162
163
164