• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
2 Licensed under the Apache License, Version 2.0 (the "License");
3 you may not use this file except in compliance with the License.
4 You may obtain a copy of the License at
5 
6     http://www.apache.org/licenses/LICENSE-2.0
7 
8 Unless required by applicable law or agreed to in writing, software
9 distributed under the License is distributed on an "AS IS" BASIS,
10 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
11 See the License for the specific language governing permissions and
12 limitations under the License.
13 ==============================================================================*/
14 #ifndef TENSORFLOW_CORE_KERNELS_IMAGEADJUST_HSV_GPU_CU_H_
15 #define TENSORFLOW_CORE_KERNELS_IMAGEADJUST_HSV_GPU_CU_H_
16 
17 #if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
18 
19 #define EIGEN_USE_GPU
20 
21 #include "tensorflow/core/framework/register_types.h"
22 #include "tensorflow/core/framework/types.h"
23 
24 namespace tensorflow {
25 namespace internal {
26 
27 typedef struct RgbTuple {
28   float r;
29   float g;
30   float b;
31 } RgbTuple;
32 
33 typedef struct HsvTuple {
34   float h;
35   float s;
36   float v;
37 } HsvTuple;
38 
rgb2hsv_cuda(const float r,const float g,const float b)39 inline __device__ HsvTuple rgb2hsv_cuda(const float r, const float g,
40                                         const float b) {
41   HsvTuple tuple;
42   const float M = fmaxf(r, fmaxf(g, b));
43   const float m = fminf(r, fminf(g, b));
44   const float chroma = M - m;
45   float h = 0.0f, s = 0.0f;
46   // hue
47   if (chroma > 0.0f) {
48     if (M == r) {
49       const float num = (g - b) / chroma;
50       const float sign = copysignf(1.0f, num);
51       h = ((sign < 0.0f) * 6.0f + sign * fmodf(sign * num, 6.0f)) / 6.0f;
52     } else if (M == g) {
53       h = ((b - r) / chroma + 2.0f) / 6.0f;
54     } else {
55       h = ((r - g) / chroma + 4.0f) / 6.0f;
56     }
57   } else {
58     h = 0.0f;
59   }
60   // saturation
61   if (M > 0.0) {
62     s = chroma / M;
63   } else {
64     s = 0.0f;
65   }
66   tuple.h = h;
67   tuple.s = s;
68   tuple.v = M;
69   return tuple;
70 }
71 
hsv2rgb_cuda(const float h,const float s,const float v)72 inline __device__ RgbTuple hsv2rgb_cuda(const float h, const float s,
73                                         const float v) {
74   RgbTuple tuple;
75   const float new_h = h * 6.0f;
76   const float chroma = v * s;
77   const float x = chroma * (1.0f - fabsf(fmodf(new_h, 2.0f) - 1.0f));
78   const float new_m = v - chroma;
79   const bool between_0_and_1 = new_h >= 0.0f && new_h < 1.0f;
80   const bool between_1_and_2 = new_h >= 1.0f && new_h < 2.0f;
81   const bool between_2_and_3 = new_h >= 2.0f && new_h < 3.0f;
82   const bool between_3_and_4 = new_h >= 3.0f && new_h < 4.0f;
83   const bool between_4_and_5 = new_h >= 4.0f && new_h < 5.0f;
84   const bool between_5_and_6 = new_h >= 5.0f && new_h < 6.0f;
85   tuple.r = chroma * (between_0_and_1 || between_5_and_6) +
86             x * (between_1_and_2 || between_4_and_5) + new_m;
87   tuple.g = chroma * (between_1_and_2 || between_2_and_3) +
88             x * (between_0_and_1 || between_3_and_4) + new_m;
89   tuple.b = chroma * (between_3_and_4 || between_4_and_5) +
90             x * (between_2_and_3 || between_5_and_6) + new_m;
91   return tuple;
92 }
93 
94 template <bool AdjustHue, bool AdjustSaturation, bool AdjustV, typename T>
adjust_hsv_nhwc(const int64 number_elements,const T * const __restrict__ input,T * const __restrict__ output,const float * const __restrict__ hue_delta,const float * const __restrict__ saturation_scale,const float * const __restrict__ value_scale)95 __global__ void adjust_hsv_nhwc(
96     const int64 number_elements, const T* const __restrict__ input,
97     T* const __restrict__ output, const float* const __restrict__ hue_delta,
98     const float* const __restrict__ saturation_scale,
99     const float* const __restrict__ value_scale) {
100   // multiply by 3 since we're dealing with contiguous RGB bytes for each pixel
101   // (NHWC)
102   for (int64 idx = (blockDim.x * blockIdx.x + threadIdx.x) * 3;
103        idx < number_elements; idx += blockDim.x * gridDim.x * 3) {
104     if (!AdjustHue && !AdjustSaturation && !AdjustV) {
105       output[idx] = input[idx];
106       output[idx + 1] = input[idx + 1];
107       output[idx + 2] = input[idx + 2];
108       continue;
109     }
110     const HsvTuple hsv = rgb2hsv_cuda(static_cast<float>(input[idx]),
111                                       static_cast<float>(input[idx + 1]),
112                                       static_cast<float>(input[idx + 2]));
113     float new_h = hsv.h;
114     float new_s = hsv.s;
115     float new_v = hsv.v;
116     // hue adjustment
117     if (AdjustHue) {
118       const float delta = *hue_delta;
119       new_h = fmodf(hsv.h + delta, 1.0f);
120       if (new_h < 0.0f) {
121         new_h = fmodf(1.0f + new_h, 1.0f);
122       }
123     }
124     // saturation adjustment
125     if (AdjustSaturation && saturation_scale != nullptr) {
126       const float scale = *saturation_scale;
127       new_s = fminf(1.0f, fmaxf(0.0f, hsv.s * scale));
128     }
129     // value adjustment
130     if (AdjustV && value_scale != nullptr) {
131       const float scale = *value_scale;
132       new_v = hsv.v * scale;
133     }
134     const RgbTuple rgb = hsv2rgb_cuda(new_h, new_s, new_v);
135     output[idx] = static_cast<T>(rgb.r);
136     output[idx + 1] = static_cast<T>(rgb.g);
137     output[idx + 2] = static_cast<T>(rgb.b);
138   }
139 }
140 
141 }  // namespace internal
142 }  // namespace tensorflow
143 
144 #endif  // GOOGLE_CUDA || TENSORFLOW_USE_ROCM
145 #endif  // TENSORFLOW_CORE_KERNELS_IMAGEADJUST_HSV_GPU_CU_H_
146