xref: /linux/drivers/gpu/drm/vkms/vkms_composer.c (revision b50ecc5aca4d18f1f0c4942f5c797bc85edef144)
1 // SPDX-License-Identifier: GPL-2.0+
2 
3 #include <linux/crc32.h>
4 
5 #include <drm/drm_atomic.h>
6 #include <drm/drm_atomic_helper.h>
7 #include <drm/drm_blend.h>
8 #include <drm/drm_fourcc.h>
9 #include <drm/drm_fixed.h>
10 #include <drm/drm_gem_framebuffer_helper.h>
11 #include <drm/drm_vblank.h>
12 #include <linux/minmax.h>
13 
14 #include "vkms_drv.h"
15 
16 static u16 pre_mul_blend_channel(u16 src, u16 dst, u16 alpha)
17 {
18 	u32 new_color;
19 
20 	new_color = (src * 0xffff + dst * (0xffff - alpha));
21 
22 	return DIV_ROUND_CLOSEST(new_color, 0xffff);
23 }
24 
25 /**
26  * pre_mul_alpha_blend - alpha blending equation
27  * @frame_info: Source framebuffer's metadata
28  * @stage_buffer: The line with the pixels from src_plane
29  * @output_buffer: A line buffer that receives all the blends output
30  *
31  * Using the information from the `frame_info`, this blends only the
32  * necessary pixels from the `stage_buffer` to the `output_buffer`
33  * using premultiplied blend formula.
34  *
35  * The current DRM assumption is that pixel color values have been already
36  * pre-multiplied with the alpha channel values. See more
37  * drm_plane_create_blend_mode_property(). Also, this formula assumes a
38  * completely opaque background.
39  */
40 static void pre_mul_alpha_blend(struct vkms_frame_info *frame_info,
41 				struct line_buffer *stage_buffer,
42 				struct line_buffer *output_buffer)
43 {
44 	int x_dst = frame_info->dst.x1;
45 	struct pixel_argb_u16 *out = output_buffer->pixels + x_dst;
46 	struct pixel_argb_u16 *in = stage_buffer->pixels;
47 	int x_limit = min_t(size_t, drm_rect_width(&frame_info->dst),
48 			    stage_buffer->n_pixels);
49 
50 	for (int x = 0; x < x_limit; x++) {
51 		out[x].a = (u16)0xffff;
52 		out[x].r = pre_mul_blend_channel(in[x].r, out[x].r, in[x].a);
53 		out[x].g = pre_mul_blend_channel(in[x].g, out[x].g, in[x].a);
54 		out[x].b = pre_mul_blend_channel(in[x].b, out[x].b, in[x].a);
55 	}
56 }
57 
58 static int get_y_pos(struct vkms_frame_info *frame_info, int y)
59 {
60 	if (frame_info->rotation & DRM_MODE_REFLECT_Y)
61 		return drm_rect_height(&frame_info->rotated) - y - 1;
62 
63 	switch (frame_info->rotation & DRM_MODE_ROTATE_MASK) {
64 	case DRM_MODE_ROTATE_90:
65 		return frame_info->rotated.x2 - y - 1;
66 	case DRM_MODE_ROTATE_270:
67 		return y + frame_info->rotated.x1;
68 	default:
69 		return y;
70 	}
71 }
72 
73 static bool check_limit(struct vkms_frame_info *frame_info, int pos)
74 {
75 	if (drm_rotation_90_or_270(frame_info->rotation)) {
76 		if (pos >= 0 && pos < drm_rect_width(&frame_info->rotated))
77 			return true;
78 	} else {
79 		if (pos >= frame_info->rotated.y1 && pos < frame_info->rotated.y2)
80 			return true;
81 	}
82 
83 	return false;
84 }
85 
86 static void fill_background(const struct pixel_argb_u16 *background_color,
87 			    struct line_buffer *output_buffer)
88 {
89 	for (size_t i = 0; i < output_buffer->n_pixels; i++)
90 		output_buffer->pixels[i] = *background_color;
91 }
92 
93 // lerp(a, b, t) = a + (b - a) * t
94 static u16 lerp_u16(u16 a, u16 b, s64 t)
95 {
96 	s64 a_fp = drm_int2fixp(a);
97 	s64 b_fp = drm_int2fixp(b);
98 
99 	s64 delta = drm_fixp_mul(b_fp - a_fp,  t);
100 
101 	return drm_fixp2int(a_fp + delta);
102 }
103 
104 static s64 get_lut_index(const struct vkms_color_lut *lut, u16 channel_value)
105 {
106 	s64 color_channel_fp = drm_int2fixp(channel_value);
107 
108 	return drm_fixp_mul(color_channel_fp, lut->channel_value2index_ratio);
109 }
110 
111 /*
112  * This enum is related to the positions of the variables inside
113  * `struct drm_color_lut`, so the order of both needs to be the same.
114  */
115 enum lut_channel {
116 	LUT_RED = 0,
117 	LUT_GREEN,
118 	LUT_BLUE,
119 	LUT_RESERVED
120 };
121 
122 static u16 apply_lut_to_channel_value(const struct vkms_color_lut *lut, u16 channel_value,
123 				      enum lut_channel channel)
124 {
125 	s64 lut_index = get_lut_index(lut, channel_value);
126 	u16 *floor_lut_value, *ceil_lut_value;
127 	u16 floor_channel_value, ceil_channel_value;
128 
129 	/*
130 	 * This checks if `struct drm_color_lut` has any gap added by the compiler
131 	 * between the struct fields.
132 	 */
133 	static_assert(sizeof(struct drm_color_lut) == sizeof(__u16) * 4);
134 
135 	floor_lut_value = (__u16 *)&lut->base[drm_fixp2int(lut_index)];
136 	if (drm_fixp2int(lut_index) == (lut->lut_length - 1))
137 		/* We're at the end of the LUT array, use same value for ceil and floor */
138 		ceil_lut_value = floor_lut_value;
139 	else
140 		ceil_lut_value = (__u16 *)&lut->base[drm_fixp2int_ceil(lut_index)];
141 
142 	floor_channel_value = floor_lut_value[channel];
143 	ceil_channel_value = ceil_lut_value[channel];
144 
145 	return lerp_u16(floor_channel_value, ceil_channel_value,
146 			lut_index & DRM_FIXED_DECIMAL_MASK);
147 }
148 
149 static void apply_lut(const struct vkms_crtc_state *crtc_state, struct line_buffer *output_buffer)
150 {
151 	if (!crtc_state->gamma_lut.base)
152 		return;
153 
154 	if (!crtc_state->gamma_lut.lut_length)
155 		return;
156 
157 	for (size_t x = 0; x < output_buffer->n_pixels; x++) {
158 		struct pixel_argb_u16 *pixel = &output_buffer->pixels[x];
159 
160 		pixel->r = apply_lut_to_channel_value(&crtc_state->gamma_lut, pixel->r, LUT_RED);
161 		pixel->g = apply_lut_to_channel_value(&crtc_state->gamma_lut, pixel->g, LUT_GREEN);
162 		pixel->b = apply_lut_to_channel_value(&crtc_state->gamma_lut, pixel->b, LUT_BLUE);
163 	}
164 }
165 
166 /**
167  * blend - blend the pixels from all planes and compute crc
168  * @wb: The writeback frame buffer metadata
169  * @crtc_state: The crtc state
170  * @crc32: The crc output of the final frame
171  * @output_buffer: A buffer of a row that will receive the result of the blend(s)
172  * @stage_buffer: The line with the pixels from plane being blend to the output
173  * @row_size: The size, in bytes, of a single row
174  *
175  * This function blends the pixels (Using the `pre_mul_alpha_blend`)
176  * from all planes, calculates the crc32 of the output from the former step,
177  * and, if necessary, convert and store the output to the writeback buffer.
178  */
179 static void blend(struct vkms_writeback_job *wb,
180 		  struct vkms_crtc_state *crtc_state,
181 		  u32 *crc32, struct line_buffer *stage_buffer,
182 		  struct line_buffer *output_buffer, size_t row_size)
183 {
184 	struct vkms_plane_state **plane = crtc_state->active_planes;
185 	u32 n_active_planes = crtc_state->num_active_planes;
186 	int y_pos;
187 
188 	const struct pixel_argb_u16 background_color = { .a = 0xffff };
189 
190 	size_t crtc_y_limit = crtc_state->base.mode.vdisplay;
191 
192 	/*
193 	 * The planes are composed line-by-line to avoid heavy memory usage. It is a necessary
194 	 * complexity to avoid poor blending performance.
195 	 *
196 	 * The function vkms_compose_row() is used to read a line, pixel-by-pixel, into the staging
197 	 * buffer.
198 	 */
199 	for (size_t y = 0; y < crtc_y_limit; y++) {
200 		fill_background(&background_color, output_buffer);
201 
202 		/* The active planes are composed associatively in z-order. */
203 		for (size_t i = 0; i < n_active_planes; i++) {
204 			y_pos = get_y_pos(plane[i]->frame_info, y);
205 
206 			if (!check_limit(plane[i]->frame_info, y_pos))
207 				continue;
208 
209 			vkms_compose_row(stage_buffer, plane[i], y_pos);
210 			pre_mul_alpha_blend(plane[i]->frame_info, stage_buffer,
211 					    output_buffer);
212 		}
213 
214 		apply_lut(crtc_state, output_buffer);
215 
216 		*crc32 = crc32_le(*crc32, (void *)output_buffer->pixels, row_size);
217 
218 		if (wb)
219 			vkms_writeback_row(wb, output_buffer, y_pos);
220 	}
221 }
222 
223 static int check_format_funcs(struct vkms_crtc_state *crtc_state,
224 			      struct vkms_writeback_job *active_wb)
225 {
226 	struct vkms_plane_state **planes = crtc_state->active_planes;
227 	u32 n_active_planes = crtc_state->num_active_planes;
228 
229 	for (size_t i = 0; i < n_active_planes; i++)
230 		if (!planes[i]->pixel_read)
231 			return -1;
232 
233 	if (active_wb && !active_wb->pixel_write)
234 		return -1;
235 
236 	return 0;
237 }
238 
239 static int check_iosys_map(struct vkms_crtc_state *crtc_state)
240 {
241 	struct vkms_plane_state **plane_state = crtc_state->active_planes;
242 	u32 n_active_planes = crtc_state->num_active_planes;
243 
244 	for (size_t i = 0; i < n_active_planes; i++)
245 		if (iosys_map_is_null(&plane_state[i]->frame_info->map[0]))
246 			return -1;
247 
248 	return 0;
249 }
250 
251 static int compose_active_planes(struct vkms_writeback_job *active_wb,
252 				 struct vkms_crtc_state *crtc_state,
253 				 u32 *crc32)
254 {
255 	size_t line_width, pixel_size = sizeof(struct pixel_argb_u16);
256 	struct line_buffer output_buffer, stage_buffer;
257 	int ret = 0;
258 
259 	/*
260 	 * This check exists so we can call `crc32_le` for the entire line
261 	 * instead doing it for each channel of each pixel in case
262 	 * `struct `pixel_argb_u16` had any gap added by the compiler
263 	 * between the struct fields.
264 	 */
265 	static_assert(sizeof(struct pixel_argb_u16) == 8);
266 
267 	if (WARN_ON(check_iosys_map(crtc_state)))
268 		return -EINVAL;
269 
270 	if (WARN_ON(check_format_funcs(crtc_state, active_wb)))
271 		return -EINVAL;
272 
273 	line_width = crtc_state->base.mode.hdisplay;
274 	stage_buffer.n_pixels = line_width;
275 	output_buffer.n_pixels = line_width;
276 
277 	stage_buffer.pixels = kvmalloc(line_width * pixel_size, GFP_KERNEL);
278 	if (!stage_buffer.pixels) {
279 		DRM_ERROR("Cannot allocate memory for the output line buffer");
280 		return -ENOMEM;
281 	}
282 
283 	output_buffer.pixels = kvmalloc(line_width * pixel_size, GFP_KERNEL);
284 	if (!output_buffer.pixels) {
285 		DRM_ERROR("Cannot allocate memory for intermediate line buffer");
286 		ret = -ENOMEM;
287 		goto free_stage_buffer;
288 	}
289 
290 	blend(active_wb, crtc_state, crc32, &stage_buffer,
291 	      &output_buffer, line_width * pixel_size);
292 
293 	kvfree(output_buffer.pixels);
294 free_stage_buffer:
295 	kvfree(stage_buffer.pixels);
296 
297 	return ret;
298 }
299 
300 /**
301  * vkms_composer_worker - ordered work_struct to compute CRC
302  *
303  * @work: work_struct
304  *
305  * Work handler for composing and computing CRCs. work_struct scheduled in
306  * an ordered workqueue that's periodically scheduled to run by
307  * vkms_vblank_simulate() and flushed at vkms_atomic_commit_tail().
308  */
309 void vkms_composer_worker(struct work_struct *work)
310 {
311 	struct vkms_crtc_state *crtc_state = container_of(work,
312 						struct vkms_crtc_state,
313 						composer_work);
314 	struct drm_crtc *crtc = crtc_state->base.crtc;
315 	struct vkms_writeback_job *active_wb = crtc_state->active_writeback;
316 	struct vkms_output *out = drm_crtc_to_vkms_output(crtc);
317 	bool crc_pending, wb_pending;
318 	u64 frame_start, frame_end;
319 	u32 crc32 = 0;
320 	int ret;
321 
322 	spin_lock_irq(&out->composer_lock);
323 	frame_start = crtc_state->frame_start;
324 	frame_end = crtc_state->frame_end;
325 	crc_pending = crtc_state->crc_pending;
326 	wb_pending = crtc_state->wb_pending;
327 	crtc_state->frame_start = 0;
328 	crtc_state->frame_end = 0;
329 	crtc_state->crc_pending = false;
330 
331 	if (crtc->state->gamma_lut) {
332 		s64 max_lut_index_fp;
333 		s64 u16_max_fp = drm_int2fixp(0xffff);
334 
335 		crtc_state->gamma_lut.base = (struct drm_color_lut *)crtc->state->gamma_lut->data;
336 		crtc_state->gamma_lut.lut_length =
337 			crtc->state->gamma_lut->length / sizeof(struct drm_color_lut);
338 		max_lut_index_fp = drm_int2fixp(crtc_state->gamma_lut.lut_length  - 1);
339 		crtc_state->gamma_lut.channel_value2index_ratio = drm_fixp_div(max_lut_index_fp,
340 									       u16_max_fp);
341 
342 	} else {
343 		crtc_state->gamma_lut.base = NULL;
344 	}
345 
346 	spin_unlock_irq(&out->composer_lock);
347 
348 	/*
349 	 * We raced with the vblank hrtimer and previous work already computed
350 	 * the crc, nothing to do.
351 	 */
352 	if (!crc_pending)
353 		return;
354 
355 	if (wb_pending)
356 		ret = compose_active_planes(active_wb, crtc_state, &crc32);
357 	else
358 		ret = compose_active_planes(NULL, crtc_state, &crc32);
359 
360 	if (ret)
361 		return;
362 
363 	if (wb_pending) {
364 		drm_writeback_signal_completion(&out->wb_connector, 0);
365 		spin_lock_irq(&out->composer_lock);
366 		crtc_state->wb_pending = false;
367 		spin_unlock_irq(&out->composer_lock);
368 	}
369 
370 	/*
371 	 * The worker can fall behind the vblank hrtimer, make sure we catch up.
372 	 */
373 	while (frame_start <= frame_end)
374 		drm_crtc_add_crc_entry(crtc, true, frame_start++, &crc32);
375 }
376 
377 static const char * const pipe_crc_sources[] = {"auto"};
378 
379 const char *const *vkms_get_crc_sources(struct drm_crtc *crtc,
380 					size_t *count)
381 {
382 	*count = ARRAY_SIZE(pipe_crc_sources);
383 	return pipe_crc_sources;
384 }
385 
386 static int vkms_crc_parse_source(const char *src_name, bool *enabled)
387 {
388 	int ret = 0;
389 
390 	if (!src_name) {
391 		*enabled = false;
392 	} else if (strcmp(src_name, "auto") == 0) {
393 		*enabled = true;
394 	} else {
395 		*enabled = false;
396 		ret = -EINVAL;
397 	}
398 
399 	return ret;
400 }
401 
402 int vkms_verify_crc_source(struct drm_crtc *crtc, const char *src_name,
403 			   size_t *values_cnt)
404 {
405 	bool enabled;
406 
407 	if (vkms_crc_parse_source(src_name, &enabled) < 0) {
408 		DRM_DEBUG_DRIVER("unknown source %s\n", src_name);
409 		return -EINVAL;
410 	}
411 
412 	*values_cnt = 1;
413 
414 	return 0;
415 }
416 
417 void vkms_set_composer(struct vkms_output *out, bool enabled)
418 {
419 	bool old_enabled;
420 
421 	if (enabled)
422 		drm_crtc_vblank_get(&out->crtc);
423 
424 	spin_lock_irq(&out->lock);
425 	old_enabled = out->composer_enabled;
426 	out->composer_enabled = enabled;
427 	spin_unlock_irq(&out->lock);
428 
429 	if (old_enabled)
430 		drm_crtc_vblank_put(&out->crtc);
431 }
432 
433 int vkms_set_crc_source(struct drm_crtc *crtc, const char *src_name)
434 {
435 	struct vkms_output *out = drm_crtc_to_vkms_output(crtc);
436 	bool enabled = false;
437 	int ret = 0;
438 
439 	ret = vkms_crc_parse_source(src_name, &enabled);
440 
441 	vkms_set_composer(out, enabled);
442 
443 	return ret;
444 }
445