10ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams/* 20ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams * Copyright (C) 2014 The Android Open Source Project 30ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams * 40ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams * Licensed under the Apache License, Version 2.0 (the "License"); 50ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams * you may not use this file except in compliance with the License. 60ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams * You may obtain a copy of the License at 70ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams * 80ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams * http://www.apache.org/licenses/LICENSE-2.0 90ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams * 100ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams * Unless required by applicable law or agreed to in writing, software 110ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams * distributed under the License is distributed on an "AS IS" BASIS, 120ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 130ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams * See the License for the specific language governing permissions and 140ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams * limitations under the License. 150ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams */ 160ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams 170ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams#include "ip.rsh" 180ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams#pragma rs_fp_relaxed 190ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams 200ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Samsint32_t gWidthIn; 210ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Samsint32_t gHeightIn; 220ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Samsrs_allocation gIn; 230ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Samsfloat scale; 240ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams 250ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams 260ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Samsuchar4 __attribute__((kernel)) nearest(uint32_t x, uint32_t y) { 270ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float xf = clamp(x * scale, 0.f, (float)gWidthIn - 1.f); 280ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float yf = clamp(y * scale, 0.f, (float)gHeightIn - 1.f); 290ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams uint32_t ix = xf; 300ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams uint32_t iy = yf; 310ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams 320ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams uchar4 tmp = rsGetElementAt_uchar4(gIn, ix, iy); 330ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams tmp.a = 0xff; 340ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams return tmp; 350ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams} 360ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams 370ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams 380ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Samsstatic float4 cubicInterpolate (float4 p0,float4 p1,float4 p2,float4 p3 , float x) { 390ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams return p1 + 0.5f * x * (p2 - p0 + x * (2.f * p0 - 5.f * p1 + 4.f * p2 - p3 400ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams + x * (3.f * (p1 - p2) + p3 - p0))); 410ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams} 420ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams 430ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Samsuchar4 __attribute__((kernel)) bicubic(uint32_t x, uint32_t y) { 440ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float xf = x * scale; 450ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float yf = y * scale; 460ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams 470ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams int startx = (int) floor(xf - 2); 480ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams int starty = (int) floor(yf - 2); 490ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams xf = xf - floor(xf); 500ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams yf = yf - floor(yf); 510ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams int maxx = gWidthIn - 1; 520ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams int maxy = gHeightIn - 1; 530ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams 540ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams uint32_t xs0 = (uint32_t) max(0, startx + 0); 550ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams uint32_t xs1 = (uint32_t) max(0, startx + 1); 560ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams uint32_t xs2 = (uint32_t) min(maxx, startx + 2); 570ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams uint32_t xs3 = (uint32_t) min(maxx, startx + 3); 580ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams 590ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams uint32_t ys0 = (uint32_t) max(0, starty + 0); 600ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams uint32_t ys1 = (uint32_t) max(0, starty + 1); 610ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams uint32_t ys2 = (uint32_t) min(maxy, starty + 2); 620ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams uint32_t ys3 = (uint32_t) min(maxy, starty + 3); 630ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams 640ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p00 = convert_float4(rsGetElementAt_uchar4(gIn, xs0, ys0)); 650ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p01 = convert_float4(rsGetElementAt_uchar4(gIn, xs1, ys0)); 660ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p02 = convert_float4(rsGetElementAt_uchar4(gIn, xs2, ys0)); 670ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p03 = convert_float4(rsGetElementAt_uchar4(gIn, xs3, ys0)); 680ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p0 = cubicInterpolate(p00, p01, p02, p03, xf); 690ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams 700ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p10 = convert_float4(rsGetElementAt_uchar4(gIn, xs0, ys1)); 710ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p11 = convert_float4(rsGetElementAt_uchar4(gIn, xs1, ys1)); 720ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p12 = convert_float4(rsGetElementAt_uchar4(gIn, xs2, ys1)); 730ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p13 = convert_float4(rsGetElementAt_uchar4(gIn, xs3, ys1)); 740ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p1 = cubicInterpolate(p10, p11, p12, p13, xf); 750ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams 760ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p20 = convert_float4(rsGetElementAt_uchar4(gIn, xs0, ys2)); 770ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p21 = convert_float4(rsGetElementAt_uchar4(gIn, xs1, ys2)); 780ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p22 = convert_float4(rsGetElementAt_uchar4(gIn, xs2, ys2)); 790ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p23 = convert_float4(rsGetElementAt_uchar4(gIn, xs3, ys2)); 800ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p2 = cubicInterpolate(p20, p21, p22, p23, xf); 810ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams 820ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p30 = convert_float4(rsGetElementAt_uchar4(gIn, xs0, ys3)); 830ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p31 = convert_float4(rsGetElementAt_uchar4(gIn, xs1, ys3)); 840ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p32 = convert_float4(rsGetElementAt_uchar4(gIn, xs2, ys3)); 850ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p33 = convert_float4(rsGetElementAt_uchar4(gIn, xs3, ys3)); 860ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p3 = cubicInterpolate(p30, p31, p32, p33, xf); 870ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams 880ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams float4 p = cubicInterpolate(p0, p1, p2, p3, yf); 890ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams p = clamp(p, 0.f, 255.f); 900ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams return convert_uchar4(p); 910ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams} 920ef64c5373a119eb73cbf7b1f7cf7d1da12d97d3Jason Sams 93