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 "rsCpuIntrinsic.h"
19#include "rsCpuIntrinsicInlines.h"
20
21using namespace android;
22using namespace android::renderscript;
23
24namespace android {
25namespace renderscript {
26
27
28class RsdCpuScriptIntrinsic3DLUT : public RsdCpuScriptIntrinsic {
29public:
30    void populateScript(Script *) override;
31    void invokeFreeChildren() override;
32
33    void setGlobalObj(uint32_t slot, ObjectBase *data) override;
34
35    ~RsdCpuScriptIntrinsic3DLUT() override;
36    RsdCpuScriptIntrinsic3DLUT(RsdCpuReferenceImpl *ctx, const Script *s, const Element *e);
37
38protected:
39    ObjectBaseRef<Allocation> mLUT;
40
41    static void kernel(const RsExpandKernelDriverInfo *info,
42                       uint32_t xstart, uint32_t xend,
43                       uint32_t outstep);
44};
45
46}
47}
48
49
50void RsdCpuScriptIntrinsic3DLUT::setGlobalObj(uint32_t slot, ObjectBase *data) {
51    rsAssert(slot == 0);
52    mLUT.set(static_cast<Allocation *>(data));
53}
54
55extern "C" void rsdIntrinsic3DLUT_K(void *dst, void const *in, size_t count,
56                                      void const *lut,
57                                      int32_t pitchy, int32_t pitchz,
58                                      int dimx, int dimy, int dimz);
59
60
61void RsdCpuScriptIntrinsic3DLUT::kernel(const RsExpandKernelDriverInfo *info,
62                                        uint32_t xstart, uint32_t xend,
63                                        uint32_t outstep) {
64    RsdCpuScriptIntrinsic3DLUT *cp = (RsdCpuScriptIntrinsic3DLUT *)info->usr;
65
66    uchar4 *out = (uchar4 *)info->outPtr[0];
67    uchar4 *in = (uchar4 *)info->inPtr[0];
68    uint32_t x1 = xstart;
69    uint32_t x2 = xend;
70
71    const uchar *bp = (const uchar *)cp->mLUT->mHal.drvState.lod[0].mallocPtr;
72
73    int4 dims = {
74        static_cast<int>(cp->mLUT->mHal.drvState.lod[0].dimX - 1),
75        static_cast<int>(cp->mLUT->mHal.drvState.lod[0].dimY - 1),
76        static_cast<int>(cp->mLUT->mHal.drvState.lod[0].dimZ - 1),
77        -1
78    };
79    const float4 m = (float4)(1.f / 255.f) * convert_float4(dims);
80    const int4 coordMul = convert_int4(m * (float4)0x8000);
81    const size_t stride_y = cp->mLUT->mHal.drvState.lod[0].stride;
82    const size_t stride_z = stride_y * cp->mLUT->mHal.drvState.lod[0].dimY;
83
84    //ALOGE("strides %zu %zu", stride_y, stride_z);
85
86#if defined(ARCH_ARM_USE_INTRINSICS)
87    if (gArchUseSIMD) {
88        int32_t len = x2 - x1;
89        if(len > 0) {
90            rsdIntrinsic3DLUT_K(out, in, len,
91                                bp, stride_y, stride_z,
92                                dims.x, dims.y, dims.z);
93            x1 += len;
94            out += len;
95            in += len;
96        }
97    }
98#endif
99
100    while (x1 < x2) {
101        int4 baseCoord = convert_int4(*in) * coordMul;
102        int4 coord1 = baseCoord >> (int4)15;
103        //int4 coord2 = min(coord1 + 1, gDims - 1);
104
105        int4 weight2 = baseCoord & 0x7fff;
106        int4 weight1 = (int4)0x8000 - weight2;
107
108        //ALOGE("coord1      %08x %08x %08x %08x", coord1.x, coord1.y, coord1.z, coord1.w);
109        const uchar *bp2 = bp + (coord1.x * 4) + (coord1.y * stride_y) + (coord1.z * stride_z);
110        const uchar4 *pt_00 = (const uchar4 *)&bp2[0];
111        const uchar4 *pt_10 = (const uchar4 *)&bp2[stride_y];
112        const uchar4 *pt_01 = (const uchar4 *)&bp2[stride_z];
113        const uchar4 *pt_11 = (const uchar4 *)&bp2[stride_y + stride_z];
114
115        uint4 v000 = convert_uint4(pt_00[0]);
116        uint4 v100 = convert_uint4(pt_00[1]);
117        uint4 v010 = convert_uint4(pt_10[0]);
118        uint4 v110 = convert_uint4(pt_10[1]);
119        uint4 v001 = convert_uint4(pt_01[0]);
120        uint4 v101 = convert_uint4(pt_01[1]);
121        uint4 v011 = convert_uint4(pt_11[0]);
122        uint4 v111 = convert_uint4(pt_11[1]);
123
124        uint4 yz00 = ((v000 * weight1.x) + (v100 * weight2.x)) >> (int4)7;
125        uint4 yz10 = ((v010 * weight1.x) + (v110 * weight2.x)) >> (int4)7;
126        uint4 yz01 = ((v001 * weight1.x) + (v101 * weight2.x)) >> (int4)7;
127        uint4 yz11 = ((v011 * weight1.x) + (v111 * weight2.x)) >> (int4)7;
128
129        uint4 z0 = ((yz00 * weight1.y) + (yz10 * weight2.y)) >> (int4)15;
130        uint4 z1 = ((yz01 * weight1.y) + (yz11 * weight2.y)) >> (int4)15;
131
132        uint4 v = ((z0 * weight1.z) + (z1 * weight2.z)) >> (int4)15;
133        uint4 v2 = (v + 0x7f) >> (int4)8;
134
135        uchar4 ret = convert_uchar4(v2);
136        ret.w = in->w;
137
138        #if 0
139        if (!x1) {
140            ALOGE("in          %08x %08x %08x %08x", in->r, in->g, in->b, in->a);
141            ALOGE("baseCoord   %08x %08x %08x %08x", baseCoord.x, baseCoord.y, baseCoord.z, baseCoord.w);
142            ALOGE("coord1      %08x %08x %08x %08x", coord1.x, coord1.y, coord1.z, coord1.w);
143            ALOGE("weight1     %08x %08x %08x %08x", weight1.x, weight1.y, weight1.z, weight1.w);
144            ALOGE("weight2     %08x %08x %08x %08x", weight2.x, weight2.y, weight2.z, weight2.w);
145
146            ALOGE("v000        %08x %08x %08x %08x", v000.x, v000.y, v000.z, v000.w);
147            ALOGE("v100        %08x %08x %08x %08x", v100.x, v100.y, v100.z, v100.w);
148            ALOGE("yz00        %08x %08x %08x %08x", yz00.x, yz00.y, yz00.z, yz00.w);
149            ALOGE("z0          %08x %08x %08x %08x", z0.x, z0.y, z0.z, z0.w);
150
151            ALOGE("v           %08x %08x %08x %08x", v.x, v.y, v.z, v.w);
152            ALOGE("v2          %08x %08x %08x %08x", v2.x, v2.y, v2.z, v2.w);
153        }
154        #endif
155        *out = ret;
156
157
158        in++;
159        out++;
160        x1++;
161    }
162}
163
164RsdCpuScriptIntrinsic3DLUT::RsdCpuScriptIntrinsic3DLUT(
165    RsdCpuReferenceImpl *ctx, const Script *s, const Element *e) :
166        RsdCpuScriptIntrinsic(ctx, s, e, RS_SCRIPT_INTRINSIC_ID_3DLUT) {
167
168    mRootPtr = &kernel;
169}
170
171RsdCpuScriptIntrinsic3DLUT::~RsdCpuScriptIntrinsic3DLUT() {
172}
173
174void RsdCpuScriptIntrinsic3DLUT::populateScript(Script *s) {
175    s->mHal.info.exportedVariableCount = 1;
176}
177
178void RsdCpuScriptIntrinsic3DLUT::invokeFreeChildren() {
179    mLUT.clear();
180}
181
182
183RsdCpuScriptImpl * rsdIntrinsic_3DLUT(RsdCpuReferenceImpl *ctx,
184                                    const Script *s, const Element *e) {
185
186    return new RsdCpuScriptIntrinsic3DLUT(ctx, s, e);
187}
188