1 | use super::pixel_format::blend; |
2 | use super::PixelFormat; |
3 | use crate::BitMapBackend; |
4 | use plotters_backend::DrawingBackend; |
5 | |
6 | /// The marker type that indicates we are currently using a RGB888 pixel format |
7 | pub struct RGBPixel; |
8 | |
9 | impl PixelFormat for RGBPixel { |
10 | const PIXEL_SIZE: usize = 3; |
11 | const EFFECTIVE_PIXEL_SIZE: usize = 3; |
12 | |
13 | #[inline (always)] |
14 | fn byte_at(r: u8, g: u8, b: u8, _a: u64, idx: usize) -> u8 { |
15 | match idx { |
16 | 0 => r, |
17 | 1 => g, |
18 | 2 => b, |
19 | _ => 0xff, |
20 | } |
21 | } |
22 | |
23 | #[inline (always)] |
24 | fn decode_pixel(data: &[u8]) -> (u8, u8, u8, u64) { |
25 | (data[0], data[1], data[2], 0x255) |
26 | } |
27 | |
28 | fn can_be_saved() -> bool { |
29 | true |
30 | } |
31 | |
32 | #[allow (clippy::many_single_char_names, clippy::cast_ptr_alignment)] |
33 | fn blend_rect_fast( |
34 | target: &mut BitMapBackend<'_, Self>, |
35 | upper_left: (i32, i32), |
36 | bottom_right: (i32, i32), |
37 | r: u8, |
38 | g: u8, |
39 | b: u8, |
40 | a: f64, |
41 | ) { |
42 | let (w, h) = target.get_size(); |
43 | let a = a.min(1.0).max(0.0); |
44 | if a == 0.0 { |
45 | return; |
46 | } |
47 | |
48 | let (x0, y0) = ( |
49 | upper_left.0.min(bottom_right.0).max(0), |
50 | upper_left.1.min(bottom_right.1).max(0), |
51 | ); |
52 | let (x1, y1) = ( |
53 | upper_left.0.max(bottom_right.0).min(w as i32), |
54 | upper_left.1.max(bottom_right.1).min(h as i32), |
55 | ); |
56 | |
57 | // This may happen when the minimal value is larger than the limit. |
58 | // Thus we just have something that is completely out-of-range |
59 | if x0 >= x1 || y0 >= y1 { |
60 | return; |
61 | } |
62 | |
63 | let dst = target.get_raw_pixel_buffer(); |
64 | |
65 | let a = (256.0 * a).floor() as u64; |
66 | |
67 | // Since we should always make sure the RGB payload occupies the logic lower bits |
68 | // thus, this type purning should work for both LE and BE CPUs |
69 | #[rustfmt::skip] |
70 | let [p1, p2, p3]: [u64; 3] = unsafe { |
71 | std::mem::transmute([ |
72 | u16::from(r), u16::from(b), u16::from(g), u16::from(r), // QW1 |
73 | u16::from(b), u16::from(g), u16::from(r), u16::from(b), // QW2 |
74 | u16::from(g), u16::from(r), u16::from(b), u16::from(g), // QW3 |
75 | ]) |
76 | }; |
77 | |
78 | #[rustfmt::skip] |
79 | let [q1, q2, q3]: [u64; 3] = unsafe { |
80 | std::mem::transmute([ |
81 | u16::from(g), u16::from(r), u16::from(b), u16::from(g), // QW1 |
82 | u16::from(r), u16::from(b), u16::from(g), u16::from(r), // QW2 |
83 | u16::from(b), u16::from(g), u16::from(r), u16::from(b), // QW3 |
84 | ]) |
85 | }; |
86 | |
87 | const N: u64 = 0xff00_ff00_ff00_ff00; |
88 | const M: u64 = 0x00ff_00ff_00ff_00ff; |
89 | |
90 | for y in y0..y1 { |
91 | let start = (y * w as i32 + x0) as usize; |
92 | let count = (x1 - x0) as usize; |
93 | |
94 | let start_ptr = &mut dst[start * Self::PIXEL_SIZE] as *mut u8 as *mut [u8; 24]; |
95 | let slice = unsafe { std::slice::from_raw_parts_mut(start_ptr, (count - 1) / 8) }; |
96 | for p in slice.iter_mut() { |
97 | let ptr = p as *mut [u8; 24] as *mut [u64; 3]; |
98 | let [d1, d2, d3] = unsafe { ptr.read_unaligned() }; |
99 | let (mut h1, mut h2, mut h3) = ((d1 >> 8) & M, (d2 >> 8) & M, (d3 >> 8) & M); |
100 | let (mut l1, mut l2, mut l3) = (d1 & M, d2 & M, d3 & M); |
101 | |
102 | #[cfg (target_endian = "little" )] |
103 | { |
104 | h1 = (h1 * (256 - a) + q1 * a) & N; |
105 | h2 = (h2 * (256 - a) + q2 * a) & N; |
106 | h3 = (h3 * (256 - a) + q3 * a) & N; |
107 | l1 = ((l1 * (256 - a) + p1 * a) & N) >> 8; |
108 | l2 = ((l2 * (256 - a) + p2 * a) & N) >> 8; |
109 | l3 = ((l3 * (256 - a) + p3 * a) & N) >> 8; |
110 | } |
111 | |
112 | #[cfg (target_endian = "big" )] |
113 | { |
114 | h1 = (h1 * (256 - a) + p1 * a) & N; |
115 | h2 = (h2 * (256 - a) + p2 * a) & N; |
116 | h3 = (h3 * (256 - a) + p3 * a) & N; |
117 | l1 = ((l1 * (256 - a) + q1 * a) & N) >> 8; |
118 | l2 = ((l2 * (256 - a) + q2 * a) & N) >> 8; |
119 | l3 = ((l3 * (256 - a) + q3 * a) & N) >> 8; |
120 | } |
121 | |
122 | unsafe { |
123 | ptr.write_unaligned([h1 | l1, h2 | l2, h3 | l3]); |
124 | } |
125 | } |
126 | |
127 | let mut iter = dst[((start + slice.len() * 8) * Self::PIXEL_SIZE) |
128 | ..((start + count) * Self::PIXEL_SIZE)] |
129 | .iter_mut(); |
130 | for _ in (slice.len() * 8)..count { |
131 | blend(iter.next().unwrap(), r, a); |
132 | blend(iter.next().unwrap(), g, a); |
133 | blend(iter.next().unwrap(), b, a); |
134 | } |
135 | } |
136 | } |
137 | |
138 | #[allow (clippy::many_single_char_names, clippy::cast_ptr_alignment)] |
139 | fn fill_rect_fast( |
140 | target: &mut BitMapBackend<'_, Self>, |
141 | upper_left: (i32, i32), |
142 | bottom_right: (i32, i32), |
143 | r: u8, |
144 | g: u8, |
145 | b: u8, |
146 | ) { |
147 | let (w, h) = target.get_size(); |
148 | let (x0, y0) = ( |
149 | upper_left.0.min(bottom_right.0).max(0), |
150 | upper_left.1.min(bottom_right.1).max(0), |
151 | ); |
152 | let (x1, y1) = ( |
153 | upper_left.0.max(bottom_right.0).min(w as i32), |
154 | upper_left.1.max(bottom_right.1).min(h as i32), |
155 | ); |
156 | |
157 | // This may happen when the minimal value is larger than the limit. |
158 | // Thus we just have something that is completely out-of-range |
159 | if x0 >= x1 || y0 >= y1 { |
160 | return; |
161 | } |
162 | |
163 | let dst = target.get_raw_pixel_buffer(); |
164 | |
165 | if r == g && g == b { |
166 | // If r == g == b, then we can use memset |
167 | if x0 != 0 || x1 != w as i32 { |
168 | // If it's not the entire row is filled, we can only do |
169 | // memset per row |
170 | for y in y0..y1 { |
171 | let start = (y * w as i32 + x0) as usize; |
172 | let count = (x1 - x0) as usize; |
173 | dst[(start * Self::PIXEL_SIZE)..((start + count) * Self::PIXEL_SIZE)] |
174 | .iter_mut() |
175 | .for_each(|e| *e = r); |
176 | } |
177 | } else { |
178 | // If the entire memory block is going to be filled, just use single memset |
179 | dst[Self::PIXEL_SIZE * (y0 * w as i32) as usize |
180 | ..(y1 * w as i32) as usize * Self::PIXEL_SIZE] |
181 | .iter_mut() |
182 | .for_each(|e| *e = r); |
183 | } |
184 | } else { |
185 | let count = (x1 - x0) as usize; |
186 | if count < 8 { |
187 | for y in y0..y1 { |
188 | let start = (y * w as i32 + x0) as usize; |
189 | let mut iter = dst |
190 | [(start * Self::PIXEL_SIZE)..((start + count) * Self::PIXEL_SIZE)] |
191 | .iter_mut(); |
192 | for _ in 0..count { |
193 | *iter.next().unwrap() = r; |
194 | *iter.next().unwrap() = g; |
195 | *iter.next().unwrap() = b; |
196 | } |
197 | } |
198 | } else { |
199 | for y in y0..y1 { |
200 | let start = (y * w as i32 + x0) as usize; |
201 | let start_ptr = &mut dst[start * Self::PIXEL_SIZE] as *mut u8 as *mut [u8; 24]; |
202 | let slice = |
203 | unsafe { std::slice::from_raw_parts_mut(start_ptr, (count - 1) / 8) }; |
204 | for p in slice.iter_mut() { |
205 | // In this case, we can actually fill 8 pixels in one iteration with |
206 | // only 3 movq instructions. |
207 | // TODO: Consider using AVX instructions when possible |
208 | let ptr = p as *mut [u8; 24] as *mut u64; |
209 | unsafe { |
210 | let [d1, d2, d3]: [u64; 3] = std::mem::transmute([ |
211 | r, g, b, r, g, b, r, g, // QW1 |
212 | b, r, g, b, r, g, b, r, // QW2 |
213 | g, b, r, g, b, r, g, b, // QW3 |
214 | ]); |
215 | ptr.write_unaligned(d1); |
216 | ptr.offset(1).write_unaligned(d2); |
217 | ptr.offset(2).write_unaligned(d3); |
218 | } |
219 | } |
220 | |
221 | for idx in (slice.len() * 8)..count { |
222 | dst[start * Self::PIXEL_SIZE + idx * Self::PIXEL_SIZE] = r; |
223 | dst[start * Self::PIXEL_SIZE + idx * Self::PIXEL_SIZE + 1] = g; |
224 | dst[start * Self::PIXEL_SIZE + idx * Self::PIXEL_SIZE + 2] = b; |
225 | } |
226 | } |
227 | } |
228 | } |
229 | } |
230 | } |
231 | |