1 /*
2 * Copyright © 2017 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include "vk_nir_convert_ycbcr.h"
25
26 #include "vk_format.h"
27 #include "vk_ycbcr_conversion.h"
28
29 #include <math.h>
30
31 static nir_def *
y_range(nir_builder * b,nir_def * y_channel,int bpc,VkSamplerYcbcrRange range)32 y_range(nir_builder *b,
33 nir_def *y_channel,
34 int bpc,
35 VkSamplerYcbcrRange range)
36 {
37 switch (range) {
38 case VK_SAMPLER_YCBCR_RANGE_ITU_FULL:
39 return y_channel;
40 case VK_SAMPLER_YCBCR_RANGE_ITU_NARROW:
41 return nir_fmul_imm(b,
42 nir_fadd_imm(b,
43 nir_fmul_imm(b, y_channel,
44 pow(2, bpc) - 1),
45 -16.0f * pow(2, bpc - 8)),
46 1.0f / (219.0f * pow(2, bpc - 8)));
47
48 default:
49 unreachable("missing Ycbcr range");
50 return NULL;
51 }
52 }
53
54 static nir_def *
chroma_range(nir_builder * b,nir_def * chroma_channel,int bpc,VkSamplerYcbcrRange range)55 chroma_range(nir_builder *b,
56 nir_def *chroma_channel,
57 int bpc,
58 VkSamplerYcbcrRange range)
59 {
60 switch (range) {
61 case VK_SAMPLER_YCBCR_RANGE_ITU_FULL:
62 return nir_fadd(b, chroma_channel,
63 nir_imm_float(b, -pow(2, bpc - 1) / (pow(2, bpc) - 1.0f)));
64 case VK_SAMPLER_YCBCR_RANGE_ITU_NARROW:
65 return nir_fmul_imm(b,
66 nir_fadd_imm(b,
67 nir_fmul_imm(b, chroma_channel,
68 pow(2, bpc) - 1),
69 -128.0f * pow(2, bpc - 8)),
70 1.0f / (224.0f * pow(2, bpc - 8)));
71 default:
72 unreachable("missing Ycbcr range");
73 return NULL;
74 }
75 }
76
77 typedef struct nir_const_value_3_4 {
78 nir_const_value v[3][4];
79 } nir_const_value_3_4;
80
81 static const nir_const_value_3_4 *
ycbcr_model_to_rgb_matrix(VkSamplerYcbcrModelConversion model)82 ycbcr_model_to_rgb_matrix(VkSamplerYcbcrModelConversion model)
83 {
84 switch (model) {
85 case VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_601: {
86 static const nir_const_value_3_4 bt601 = { {
87 { { .f32 = 1.402f }, { .f32 = 1.0f }, { .f32 = 0.0f }, { .f32 = 0.0f } },
88 { { .f32 = -0.714136286201022f }, { .f32 = 1.0f }, { .f32 = -0.344136286201022f }, { .f32 = 0.0f } },
89 { { .f32 = 0.0f }, { .f32 = 1.0f }, { .f32 = 1.772f }, { .f32 = 0.0f } },
90 } };
91
92 return &bt601;
93 }
94 case VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_709: {
95 static const nir_const_value_3_4 bt709 = { {
96 { { .f32 = 1.5748031496063f }, { .f32 = 1.0f }, { .f32 = 0.0f }, { .f32 = 0.0f } },
97 { { .f32 = -0.468125209181067f }, { .f32 = 1.0f }, { .f32 = -0.187327487470334f }, { .f32 = 0.0f } },
98 { { .f32 = 0.0f }, { .f32 = 1.0f }, { .f32 = 1.85563184264242f }, { .f32 = 0.0f } },
99 } };
100
101 return &bt709;
102 }
103 case VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_2020: {
104 static const nir_const_value_3_4 bt2020 = { {
105 { { .f32 = 1.4746f }, { .f32 = 1.0f }, { .f32 = 0.0f }, { .f32 = 0.0f } },
106 { { .f32 = -0.571353126843658f }, { .f32 = 1.0f }, { .f32 = -0.164553126843658f }, { .f32 = 0.0f } },
107 { { .f32 = 0.0f }, { .f32 = 1.0f }, { .f32 = 1.8814f }, { .f32 = 0.0f } },
108 } };
109
110 return &bt2020;
111 }
112 default:
113 unreachable("missing Ycbcr model");
114 return NULL;
115 }
116 }
117
118 nir_def *
nir_convert_ycbcr_to_rgb(nir_builder * b,VkSamplerYcbcrModelConversion model,VkSamplerYcbcrRange range,nir_def * raw_channels,uint32_t * bpcs)119 nir_convert_ycbcr_to_rgb(nir_builder *b,
120 VkSamplerYcbcrModelConversion model,
121 VkSamplerYcbcrRange range,
122 nir_def *raw_channels,
123 uint32_t *bpcs)
124 {
125 nir_def *expanded_channels =
126 nir_vec4(b,
127 chroma_range(b, nir_channel(b, raw_channels, 0), bpcs[0], range),
128 y_range(b, nir_channel(b, raw_channels, 1), bpcs[1], range),
129 chroma_range(b, nir_channel(b, raw_channels, 2), bpcs[2], range),
130 nir_channel(b, raw_channels, 3));
131
132 if (model == VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_IDENTITY)
133 return expanded_channels;
134
135 const nir_const_value_3_4 *conversion_matrix =
136 ycbcr_model_to_rgb_matrix(model);
137
138 nir_def *converted_channels[] = {
139 nir_fdot(b, expanded_channels, nir_build_imm(b, 4, 32, conversion_matrix->v[0])),
140 nir_fdot(b, expanded_channels, nir_build_imm(b, 4, 32, conversion_matrix->v[1])),
141 nir_fdot(b, expanded_channels, nir_build_imm(b, 4, 32, conversion_matrix->v[2]))
142 };
143
144 return nir_vec4(b,
145 converted_channels[0], converted_channels[1],
146 converted_channels[2], nir_channel(b, raw_channels, 3));
147 }
148
149 struct ycbcr_state {
150 nir_builder *builder;
151 nir_def *image_size;
152 nir_tex_instr *origin_tex;
153 nir_deref_instr *tex_deref;
154 const struct vk_ycbcr_conversion_state *conversion;
155 const struct vk_format_ycbcr_info *format_ycbcr_info;
156 };
157
158 /* TODO: we should probably replace this with a push constant/uniform. */
159 static nir_def *
get_texture_size(struct ycbcr_state * state,nir_deref_instr * texture)160 get_texture_size(struct ycbcr_state *state, nir_deref_instr *texture)
161 {
162 if (state->image_size)
163 return state->image_size;
164
165 nir_builder *b = state->builder;
166 const struct glsl_type *type = texture->type;
167 nir_tex_instr *tex = nir_tex_instr_create(b->shader, 1);
168
169 tex->op = nir_texop_txs;
170 tex->sampler_dim = glsl_get_sampler_dim(type);
171 tex->is_array = glsl_sampler_type_is_array(type);
172 tex->is_shadow = glsl_sampler_type_is_shadow(type);
173 tex->dest_type = nir_type_int32;
174
175 tex->src[0] = nir_tex_src_for_ssa(nir_tex_src_texture_deref,
176 &texture->def);
177
178 nir_def_init(&tex->instr, &tex->def, nir_tex_instr_dest_size(tex), 32);
179 nir_builder_instr_insert(b, &tex->instr);
180
181 state->image_size = nir_i2f32(b, &tex->def);
182
183 return state->image_size;
184 }
185
186 static nir_def *
implicit_downsampled_coord(nir_builder * b,nir_def * value,nir_def * max_value,int div_scale)187 implicit_downsampled_coord(nir_builder *b,
188 nir_def *value,
189 nir_def *max_value,
190 int div_scale)
191 {
192 return nir_fadd(b,
193 value,
194 nir_frcp(b,
195 nir_fmul(b,
196 nir_imm_float(b, div_scale),
197 max_value)));
198 }
199
200 static nir_def *
implicit_downsampled_coords(struct ycbcr_state * state,nir_def * old_coords,const struct vk_format_ycbcr_plane * format_plane)201 implicit_downsampled_coords(struct ycbcr_state *state,
202 nir_def *old_coords,
203 const struct vk_format_ycbcr_plane *format_plane)
204 {
205 nir_builder *b = state->builder;
206 const struct vk_ycbcr_conversion_state *conversion = state->conversion;
207 nir_def *image_size = get_texture_size(state, state->tex_deref);
208 nir_def *comp[4] = { NULL, };
209 int c;
210
211 for (c = 0; c < ARRAY_SIZE(conversion->chroma_offsets); c++) {
212 if (format_plane->denominator_scales[c] > 1 &&
213 conversion->chroma_offsets[c] == VK_CHROMA_LOCATION_COSITED_EVEN) {
214 comp[c] = implicit_downsampled_coord(b,
215 nir_channel(b, old_coords, c),
216 nir_channel(b, image_size, c),
217 format_plane->denominator_scales[c]);
218 } else {
219 comp[c] = nir_channel(b, old_coords, c);
220 }
221 }
222
223 /* Leave other coordinates untouched */
224 for (; c < old_coords->num_components; c++)
225 comp[c] = nir_channel(b, old_coords, c);
226
227 return nir_vec(b, comp, old_coords->num_components);
228 }
229
230 static nir_def *
create_plane_tex_instr_implicit(struct ycbcr_state * state,uint32_t plane)231 create_plane_tex_instr_implicit(struct ycbcr_state *state,
232 uint32_t plane)
233 {
234 nir_builder *b = state->builder;
235 const struct vk_ycbcr_conversion_state *conversion = state->conversion;
236 const struct vk_format_ycbcr_plane *format_plane =
237 &state->format_ycbcr_info->planes[plane];
238 nir_tex_instr *old_tex = state->origin_tex;
239 nir_tex_instr *tex = nir_tex_instr_create(b->shader, old_tex->num_srcs + 1);
240
241 for (uint32_t i = 0; i < old_tex->num_srcs; i++) {
242 tex->src[i].src_type = old_tex->src[i].src_type;
243
244 switch (old_tex->src[i].src_type) {
245 case nir_tex_src_coord:
246 if (format_plane->has_chroma && conversion->chroma_reconstruction) {
247 tex->src[i].src =
248 nir_src_for_ssa(implicit_downsampled_coords(state,
249 old_tex->src[i].src.ssa,
250 format_plane));
251 break;
252 }
253 FALLTHROUGH;
254 default:
255 tex->src[i].src = nir_src_for_ssa(old_tex->src[i].src.ssa);
256 break;
257 }
258 }
259 tex->src[tex->num_srcs - 1] = nir_tex_src_for_ssa(nir_tex_src_plane,
260 nir_imm_int(b, plane));
261 tex->sampler_dim = old_tex->sampler_dim;
262 tex->dest_type = old_tex->dest_type;
263
264 tex->op = old_tex->op;
265 tex->coord_components = old_tex->coord_components;
266 tex->is_new_style_shadow = old_tex->is_new_style_shadow;
267 tex->component = old_tex->component;
268
269 tex->texture_index = old_tex->texture_index;
270 tex->sampler_index = old_tex->sampler_index;
271 tex->is_array = old_tex->is_array;
272
273 nir_def_init(&tex->instr, &tex->def, old_tex->def.num_components,
274 old_tex->def.bit_size);
275 nir_builder_instr_insert(b, &tex->instr);
276
277 return &tex->def;
278 }
279
280 static unsigned
swizzle_to_component(VkComponentSwizzle swizzle)281 swizzle_to_component(VkComponentSwizzle swizzle)
282 {
283 switch (swizzle) {
284 case VK_COMPONENT_SWIZZLE_R:
285 return 0;
286 case VK_COMPONENT_SWIZZLE_G:
287 return 1;
288 case VK_COMPONENT_SWIZZLE_B:
289 return 2;
290 case VK_COMPONENT_SWIZZLE_A:
291 return 3;
292 default:
293 unreachable("invalid channel");
294 return 0;
295 }
296 }
297
298 struct lower_ycbcr_tex_state {
299 nir_vk_ycbcr_conversion_lookup_cb cb;
300 const void *cb_data;
301 };
302
303 static bool
lower_ycbcr_tex_instr(nir_builder * b,nir_instr * instr,void * _state)304 lower_ycbcr_tex_instr(nir_builder *b, nir_instr *instr, void *_state)
305 {
306 const struct lower_ycbcr_tex_state *state = _state;
307
308 if (instr->type != nir_instr_type_tex)
309 return false;
310
311 nir_tex_instr *tex = nir_instr_as_tex(instr);
312
313 /* For the following instructions, we don't apply any change and let the
314 * instruction apply to the first plane.
315 */
316 if (tex->op == nir_texop_txs ||
317 tex->op == nir_texop_query_levels ||
318 tex->op == nir_texop_lod)
319 return false;
320
321 int deref_src_idx = nir_tex_instr_src_index(tex, nir_tex_src_texture_deref);
322 assert(deref_src_idx >= 0);
323 nir_deref_instr *deref = nir_src_as_deref(tex->src[deref_src_idx].src);
324
325 nir_variable *var = nir_deref_instr_get_variable(deref);
326 uint32_t set = var->data.descriptor_set;
327 uint32_t binding = var->data.binding;
328
329 assert(tex->texture_index == 0);
330 unsigned array_index = 0;
331 if (deref->deref_type != nir_deref_type_var) {
332 assert(deref->deref_type == nir_deref_type_array);
333 if (!nir_src_is_const(deref->arr.index))
334 return false;
335 array_index = nir_src_as_uint(deref->arr.index);
336 }
337
338 const struct vk_ycbcr_conversion_state *conversion =
339 state->cb(state->cb_data, set, binding, array_index);
340 if (conversion == NULL)
341 return false;
342
343 const struct vk_format_ycbcr_info *format_ycbcr_info =
344 vk_format_get_ycbcr_info(conversion->format);
345
346 /* This can happen if the driver hasn't done a good job of filtering on
347 * sampler creation and lets through a VkYcbcrConversion object which isn't
348 * actually YCbCr. We're supposed to ignore those.
349 */
350 if (format_ycbcr_info == NULL)
351 return false;
352
353 b->cursor = nir_before_instr(&tex->instr);
354
355 VkFormat y_format = VK_FORMAT_UNDEFINED;
356 for (uint32_t p = 0; p < format_ycbcr_info->n_planes; p++) {
357 if (!format_ycbcr_info->planes[p].has_chroma)
358 y_format = format_ycbcr_info->planes[p].format;
359 }
360 assert(y_format != VK_FORMAT_UNDEFINED);
361 const struct util_format_description *y_format_desc =
362 util_format_description(vk_format_to_pipe_format(y_format));
363 uint8_t y_bpc = y_format_desc->channel[0].size;
364
365 /* |ycbcr_comp| holds components in the order : Cr-Y-Cb */
366 nir_def *zero = nir_imm_float(b, 0.0f);
367 nir_def *one = nir_imm_float(b, 1.0f);
368 /* Use extra 2 channels for following swizzle */
369 nir_def *ycbcr_comp[5] = { zero, zero, zero, one, zero };
370
371 uint8_t ycbcr_bpcs[5];
372 memset(ycbcr_bpcs, y_bpc, sizeof(ycbcr_bpcs));
373
374 /* Go through all the planes and gather the samples into a |ycbcr_comp|
375 * while applying a swizzle required by the spec:
376 *
377 * R, G, B should respectively map to Cr, Y, Cb
378 */
379 for (uint32_t p = 0; p < format_ycbcr_info->n_planes; p++) {
380 const struct vk_format_ycbcr_plane *format_plane =
381 &format_ycbcr_info->planes[p];
382
383 struct ycbcr_state tex_state = {
384 .builder = b,
385 .origin_tex = tex,
386 .tex_deref = deref,
387 .conversion = conversion,
388 .format_ycbcr_info = format_ycbcr_info,
389 };
390 nir_def *plane_sample = create_plane_tex_instr_implicit(&tex_state, p);
391
392 for (uint32_t pc = 0; pc < 4; pc++) {
393 VkComponentSwizzle ycbcr_swizzle = format_plane->ycbcr_swizzle[pc];
394 if (ycbcr_swizzle == VK_COMPONENT_SWIZZLE_ZERO)
395 continue;
396
397 unsigned ycbcr_component = swizzle_to_component(ycbcr_swizzle);
398 ycbcr_comp[ycbcr_component] = nir_channel(b, plane_sample, pc);
399
400 /* Also compute the number of bits for each component. */
401 const struct util_format_description *plane_format_desc =
402 util_format_description(vk_format_to_pipe_format(format_plane->format));
403 ycbcr_bpcs[ycbcr_component] = plane_format_desc->channel[pc].size;
404 }
405 }
406
407 /* Now remaps components to the order specified by the conversion. */
408 nir_def *swizzled_comp[4] = { NULL, };
409 uint32_t swizzled_bpcs[4] = { 0, };
410
411 for (uint32_t i = 0; i < ARRAY_SIZE(conversion->mapping); i++) {
412 /* Maps to components in |ycbcr_comp| */
413 static const uint32_t swizzle_mapping[] = {
414 [VK_COMPONENT_SWIZZLE_ZERO] = 4,
415 [VK_COMPONENT_SWIZZLE_ONE] = 3,
416 [VK_COMPONENT_SWIZZLE_R] = 0,
417 [VK_COMPONENT_SWIZZLE_G] = 1,
418 [VK_COMPONENT_SWIZZLE_B] = 2,
419 [VK_COMPONENT_SWIZZLE_A] = 3,
420 };
421 const VkComponentSwizzle m = conversion->mapping[i];
422
423 if (m == VK_COMPONENT_SWIZZLE_IDENTITY) {
424 swizzled_comp[i] = ycbcr_comp[i];
425 swizzled_bpcs[i] = ycbcr_bpcs[i];
426 } else {
427 swizzled_comp[i] = ycbcr_comp[swizzle_mapping[m]];
428 swizzled_bpcs[i] = ycbcr_bpcs[swizzle_mapping[m]];
429 }
430 }
431
432 nir_def *result = nir_vec(b, swizzled_comp, 4);
433 if (conversion->ycbcr_model != VK_SAMPLER_YCBCR_MODEL_CONVERSION_RGB_IDENTITY) {
434 result = nir_convert_ycbcr_to_rgb(b, conversion->ycbcr_model,
435 conversion->ycbcr_range,
436 result,
437 swizzled_bpcs);
438 }
439
440 nir_def_rewrite_uses(&tex->def, result);
441 nir_instr_remove(&tex->instr);
442
443 return true;
444 }
445
nir_vk_lower_ycbcr_tex(nir_shader * nir,nir_vk_ycbcr_conversion_lookup_cb cb,const void * cb_data)446 bool nir_vk_lower_ycbcr_tex(nir_shader *nir,
447 nir_vk_ycbcr_conversion_lookup_cb cb,
448 const void *cb_data)
449 {
450 struct lower_ycbcr_tex_state state = {
451 .cb = cb,
452 .cb_data = cb_data,
453 };
454
455 return nir_shader_instructions_pass(nir, lower_ycbcr_tex_instr,
456 nir_metadata_block_index |
457 nir_metadata_dominance,
458 &state);
459 }
460