1ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams/*
2ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams * Copyright (C) 2012 The Android Open Source Project
3ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams *
4ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams * Licensed under the Apache License, Version 2.0 (the "License");
5ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams * you may not use this file except in compliance with the License.
6ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams * You may obtain a copy of the License at
7ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams *
8ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams *      http://www.apache.org/licenses/LICENSE-2.0
9ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams *
10ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams * Unless required by applicable law or agreed to in writing, software
11ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams * distributed under the License is distributed on an "AS IS" BASIS,
12ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams * See the License for the specific language governing permissions and
14ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams * limitations under the License.
15ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams */
16ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
17ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
18ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams#include "rsdCore.h"
19ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams#include "rsdIntrinsics.h"
20ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams#include "rsdAllocation.h"
21ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
22ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams#include "rsdIntrinsicInlines.h"
23ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
24ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Samsusing namespace android;
25ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Samsusing namespace android::renderscript;
26ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
27ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Samsstruct ConvolveParams {
28ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    float fp[16];
29ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    short ip[16];
30ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    bool use3x3;
31ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    bool useDot;
32ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams};
33ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
34ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Samsstatic void ColorMatrix_SetVar(const Context *dc, const Script *script, void * intrinsicData,
35ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams                               uint32_t slot, void *data, size_t dataLength) {
36ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    ConvolveParams *cp = (ConvolveParams *)intrinsicData;
37ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
38ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    rsAssert(slot == 0);
39ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    memcpy (cp->fp, data, dataLength);
40ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    for(int ct=0; ct < 16; ct++) {
41ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams        cp->ip[ct] = (short)(cp->fp[ct] * 255.f + 0.5f);
42ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    }
43ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
44ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    if ((cp->ip[3] == 0) && (cp->ip[7] == 0) && (cp->ip[11] == 0) &&
45ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams        (cp->ip[12] == 0) && (cp->ip[13] == 0) && (cp->ip[14] == 0) &&
46ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams        (cp->ip[15] == 255)) {
47ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams        cp->use3x3 = true;
48ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
49ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams        if ((cp->ip[0] == cp->ip[1]) && (cp->ip[0] == cp->ip[2]) &&
50ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            (cp->ip[4] == cp->ip[5]) && (cp->ip[4] == cp->ip[6]) &&
51ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            (cp->ip[8] == cp->ip[9]) && (cp->ip[8] == cp->ip[10])) {
52ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            cp->useDot = true;
53ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams        }
54ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    }
55ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams}
56ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
57ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Samsextern "C" void rsdIntrinsicColorMatrix4x4_K(void *dst, const void *src, const short *coef, uint32_t count);
58ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Samsextern "C" void rsdIntrinsicColorMatrix3x3_K(void *dst, const void *src, const short *coef, uint32_t count);
59ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Samsextern "C" void rsdIntrinsicColorMatrixDot_K(void *dst, const void *src, const short *coef, uint32_t count);
60ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
61ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Samsstatic void One(const RsForEachStubParamStruct *p, uchar4 *out,
62ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams                const uchar4 *py, const float* coeff) {
63ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    float4 i = convert_float4(py[0]);
64ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
65ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    float4 sum;
66ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    sum.x = i.x * coeff[0] +
67ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            i.y * coeff[4] +
68ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            i.z * coeff[8] +
69ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            i.w * coeff[12];
70ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    sum.y = i.x * coeff[1] +
71ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            i.y * coeff[5] +
72ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            i.z * coeff[9] +
73ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            i.w * coeff[13];
74ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    sum.z = i.x * coeff[2] +
75ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            i.y * coeff[6] +
76ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            i.z * coeff[10] +
77ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            i.w * coeff[14];
78ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    sum.w = i.x * coeff[3] +
79ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            i.y * coeff[7] +
80ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            i.z * coeff[11] +
81ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            i.w * coeff[15];
82ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
83ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    sum.x = sum.x < 0 ? 0 : (sum.x > 255 ? 255 : sum.x);
84ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    sum.y = sum.y < 0 ? 0 : (sum.y > 255 ? 255 : sum.y);
85ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    sum.z = sum.z < 0 ? 0 : (sum.z > 255 ? 255 : sum.z);
86ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    sum.w = sum.w < 0 ? 0 : (sum.w > 255 ? 255 : sum.w);
87ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
88ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    *out = convert_uchar4(sum);
89ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams}
90ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
91ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Samsstatic void ColorMatrix_uchar4(const RsForEachStubParamStruct *p,
92ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams                                    uint32_t xstart, uint32_t xend,
93ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams                                    uint32_t instep, uint32_t outstep) {
94ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    ConvolveParams *cp = (ConvolveParams *)p->usr;
95ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    uchar4 *out = (uchar4 *)p->out;
96ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    uchar4 *in = (uchar4 *)p->in;
97ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    uint32_t x1 = xstart;
98ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    uint32_t x2 = xend;
99ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
100ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    in += xstart;
101ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    out += xstart;
102ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
103ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    if(x2 > x1) {
104ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams#if defined(ARCH_ARM_HAVE_NEON)
105ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams        int32_t len = (x2 - x1) >> 2;
106ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams        if(len > 0) {
107ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            if (cp->use3x3) {
108ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams                if (cp->useDot) {
109ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams                    rsdIntrinsicColorMatrixDot_K(out, in, cp->ip, len);
110ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams                } else {
111ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams                    rsdIntrinsicColorMatrix3x3_K(out, in, cp->ip, len);
112ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams                }
113ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            } else {
114ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams                rsdIntrinsicColorMatrix4x4_K(out, in, cp->ip, len);
115ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            }
116ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            x1 += len << 2;
117ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            out += len << 2;
118ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            in += len << 2;
119ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams        }
120ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams#endif
121ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
122ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams        while(x1 != x2) {
123ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            One(p, out++, in++, cp->fp);
124ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams            x1++;
125ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams        }
126ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    }
127ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams}
128ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
129ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Samsvoid * rsdIntrinsic_InitColorMatrix(const android::renderscript::Context *dc,
130ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams                                    android::renderscript::Script *script,
131ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams                                    RsdIntriniscFuncs_t *funcs) {
132ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
133ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    script->mHal.info.exportedVariableCount = 1;
134ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    funcs->setVar = ColorMatrix_SetVar;
135ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    funcs->root = ColorMatrix_uchar4;
136ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
137ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    ConvolveParams *cp = (ConvolveParams *)calloc(1, sizeof(ConvolveParams));
138ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    cp->fp[0] = 1.f;
139ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    cp->fp[5] = 1.f;
140ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    cp->fp[10] = 1.f;
141ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    cp->fp[15] = 1.f;
142ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    for(int ct=0; ct < 16; ct++) {
143ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams        cp->ip[ct] = (short)(cp->fp[ct] * 255.f + 0.5f);
144ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    }
145ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams    return cp;
146ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams}
147ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
148ca29b8caf56fa4866752f9cea4ec02b2a271dceeJason Sams
149