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 "anv_nir.h"
25 #include "anv_private.h"
26 #include "nir/nir.h"
27 #include "nir/nir_builder.h"
28
29 struct ycbcr_state {
30 nir_builder *builder;
31 nir_ssa_def *image_size;
32 nir_tex_instr *origin_tex;
33 struct anv_ycbcr_conversion *conversion;
34 };
35
36 static nir_ssa_def *
y_range(nir_builder * b,nir_ssa_def * y_channel,int bpc,VkSamplerYcbcrRangeKHR range)37 y_range(nir_builder *b,
38 nir_ssa_def *y_channel,
39 int bpc,
40 VkSamplerYcbcrRangeKHR range)
41 {
42 switch (range) {
43 case VK_SAMPLER_YCBCR_RANGE_ITU_FULL_KHR:
44 return y_channel;
45 case VK_SAMPLER_YCBCR_RANGE_ITU_NARROW_KHR:
46 return nir_fmul(b,
47 nir_fadd(b,
48 nir_fmul(b, y_channel,
49 nir_imm_float(b, pow(2, bpc) - 1)),
50 nir_imm_float(b, -16.0f * pow(2, bpc - 8))),
51 nir_frcp(b, nir_imm_float(b, 219.0f * pow(2, bpc - 8))));
52 default:
53 unreachable("missing Ycbcr range");
54 return NULL;
55 }
56 }
57
58 static nir_ssa_def *
chroma_range(nir_builder * b,nir_ssa_def * chroma_channel,int bpc,VkSamplerYcbcrRangeKHR range)59 chroma_range(nir_builder *b,
60 nir_ssa_def *chroma_channel,
61 int bpc,
62 VkSamplerYcbcrRangeKHR range)
63 {
64 switch (range) {
65 case VK_SAMPLER_YCBCR_RANGE_ITU_FULL_KHR:
66 return nir_fadd(b, chroma_channel,
67 nir_imm_float(b, -pow(2, bpc - 1) / (pow(2, bpc) - 1.0f)));
68 case VK_SAMPLER_YCBCR_RANGE_ITU_NARROW_KHR:
69 return nir_fmul(b,
70 nir_fadd(b,
71 nir_fmul(b, chroma_channel,
72 nir_imm_float(b, pow(2, bpc) - 1)),
73 nir_imm_float(b, -128.0f * pow(2, bpc - 8))),
74 nir_frcp(b, nir_imm_float(b, 224.0f * pow(2, bpc - 8))));
75 default:
76 unreachable("missing Ycbcr range");
77 return NULL;
78 }
79 }
80
81 static const nir_const_value *
ycbcr_model_to_rgb_matrix(VkSamplerYcbcrModelConversionKHR model)82 ycbcr_model_to_rgb_matrix(VkSamplerYcbcrModelConversionKHR model)
83 {
84 switch (model) {
85 case VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_601_KHR: {
86 static const nir_const_value bt601[3] = {
87 { .f32 = { 1.402f, 1.0f, 0.0f, 0.0f } },
88 { .f32 = { -0.714136286201022f, 1.0f, -0.344136286201022f, 0.0f } },
89 { .f32 = { 0.0f, 1.0f, 1.772f, 0.0f } }
90 };
91
92 return bt601;
93 }
94 case VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_709_KHR: {
95 static const nir_const_value bt709[3] = {
96 { .f32 = { 1.5748031496063f, 1.0f, 0.0, 0.0f } },
97 { .f32 = { -0.468125209181067f, 1.0f, -0.187327487470334f, 0.0f } },
98 { .f32 = { 0.0f, 1.0f, 1.85563184264242f, 0.0f } }
99 };
100
101 return bt709;
102 }
103 case VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_2020_KHR: {
104 static const nir_const_value bt2020[3] = {
105 { .f32 = { 1.4746f, 1.0f, 0.0f, 0.0f } },
106 { .f32 = { -0.571353126843658f, 1.0f, -0.164553126843658f, 0.0f } },
107 { .f32 = { 0.0f, 1.0f, 1.8814f, 0.0f } }
108 };
109
110 return bt2020;
111 }
112 default:
113 unreachable("missing Ycbcr model");
114 return NULL;
115 }
116 }
117
118 static nir_ssa_def *
convert_ycbcr(struct ycbcr_state * state,nir_ssa_def * raw_channels,uint32_t * bpcs)119 convert_ycbcr(struct ycbcr_state *state,
120 nir_ssa_def *raw_channels,
121 uint32_t *bpcs)
122 {
123 nir_builder *b = state->builder;
124 struct anv_ycbcr_conversion *conversion = state->conversion;
125
126 nir_ssa_def *expanded_channels =
127 nir_vec4(b,
128 chroma_range(b, nir_channel(b, raw_channels, 0),
129 bpcs[0], conversion->ycbcr_range),
130 y_range(b, nir_channel(b, raw_channels, 1),
131 bpcs[1], conversion->ycbcr_range),
132 chroma_range(b, nir_channel(b, raw_channels, 2),
133 bpcs[2], conversion->ycbcr_range),
134 nir_imm_float(b, 1.0f));
135
136 if (conversion->ycbcr_model == VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_IDENTITY_KHR)
137 return expanded_channels;
138
139 const nir_const_value *conversion_matrix =
140 ycbcr_model_to_rgb_matrix(conversion->ycbcr_model);
141
142 nir_ssa_def *converted_channels[] = {
143 nir_fdot4(b, expanded_channels, nir_build_imm(b, 4, 32, conversion_matrix[0])),
144 nir_fdot4(b, expanded_channels, nir_build_imm(b, 4, 32, conversion_matrix[1])),
145 nir_fdot4(b, expanded_channels, nir_build_imm(b, 4, 32, conversion_matrix[2]))
146 };
147
148 return nir_vec4(b,
149 converted_channels[0], converted_channels[1],
150 converted_channels[2], nir_imm_float(b, 1.0f));
151 }
152
153 /* TODO: we should probably replace this with a push constant/uniform. */
154 static nir_ssa_def *
get_texture_size(struct ycbcr_state * state,nir_deref_var * texture)155 get_texture_size(struct ycbcr_state *state, nir_deref_var *texture)
156 {
157 if (state->image_size)
158 return state->image_size;
159
160 nir_builder *b = state->builder;
161 const struct glsl_type *type = nir_deref_tail(&texture->deref)->type;
162 nir_tex_instr *tex = nir_tex_instr_create(b->shader, 0);
163
164 tex->op = nir_texop_txs;
165 tex->sampler_dim = glsl_get_sampler_dim(type);
166 tex->is_array = glsl_sampler_type_is_array(type);
167 tex->is_shadow = glsl_sampler_type_is_shadow(type);
168 tex->texture = nir_deref_var_clone(texture, tex);
169 tex->dest_type = nir_type_int;
170
171 nir_ssa_dest_init(&tex->instr, &tex->dest,
172 nir_tex_instr_dest_size(tex), 32, NULL);
173 nir_builder_instr_insert(b, &tex->instr);
174
175 state->image_size = nir_i2f32(b, &tex->dest.ssa);
176
177 return state->image_size;
178 }
179
180 static nir_ssa_def *
implicit_downsampled_coord(nir_builder * b,nir_ssa_def * value,nir_ssa_def * max_value,int div_scale)181 implicit_downsampled_coord(nir_builder *b,
182 nir_ssa_def *value,
183 nir_ssa_def *max_value,
184 int div_scale)
185 {
186 return nir_fadd(b,
187 value,
188 nir_fdiv(b,
189 nir_imm_float(b, 1.0f),
190 nir_fmul(b,
191 nir_imm_float(b, div_scale),
192 max_value)));
193 }
194
195 static nir_ssa_def *
implicit_downsampled_coords(struct ycbcr_state * state,nir_ssa_def * old_coords,const struct anv_format_plane * plane_format)196 implicit_downsampled_coords(struct ycbcr_state *state,
197 nir_ssa_def *old_coords,
198 const struct anv_format_plane *plane_format)
199 {
200 nir_builder *b = state->builder;
201 struct anv_ycbcr_conversion *conversion = state->conversion;
202 nir_ssa_def *image_size = get_texture_size(state,
203 state->origin_tex->texture);
204 nir_ssa_def *comp[4] = { NULL, };
205 int c;
206
207 for (c = 0; c < ARRAY_SIZE(conversion->chroma_offsets); c++) {
208 if (plane_format->denominator_scales[c] > 1 &&
209 conversion->chroma_offsets[c] == VK_CHROMA_LOCATION_COSITED_EVEN_KHR) {
210 comp[c] = implicit_downsampled_coord(b,
211 nir_channel(b, old_coords, c),
212 nir_channel(b, image_size, c),
213 plane_format->denominator_scales[c]);
214 } else {
215 comp[c] = nir_channel(b, old_coords, c);
216 }
217 }
218
219 /* Leave other coordinates untouched */
220 for (; c < old_coords->num_components; c++)
221 comp[c] = nir_channel(b, old_coords, c);
222
223 return nir_vec(b, comp, old_coords->num_components);
224 }
225
226 static nir_ssa_def *
create_plane_tex_instr_implicit(struct ycbcr_state * state,uint32_t plane)227 create_plane_tex_instr_implicit(struct ycbcr_state *state,
228 uint32_t plane)
229 {
230 nir_builder *b = state->builder;
231 struct anv_ycbcr_conversion *conversion = state->conversion;
232 const struct anv_format_plane *plane_format =
233 &conversion->format->planes[plane];
234 nir_tex_instr *old_tex = state->origin_tex;
235 nir_tex_instr *tex = nir_tex_instr_create(b->shader, old_tex->num_srcs + 1);
236
237 for (uint32_t i = 0; i < old_tex->num_srcs; i++) {
238 tex->src[i].src_type = old_tex->src[i].src_type;
239
240 switch (old_tex->src[i].src_type) {
241 case nir_tex_src_coord:
242 if (plane_format->has_chroma && conversion->chroma_reconstruction) {
243 assert(old_tex->src[i].src.is_ssa);
244 tex->src[i].src =
245 nir_src_for_ssa(implicit_downsampled_coords(state,
246 old_tex->src[i].src.ssa,
247 plane_format));
248 break;
249 }
250 /* fall through */
251 default:
252 nir_src_copy(&tex->src[i].src, &old_tex->src[i].src, tex);
253 break;
254 }
255 }
256 tex->src[tex->num_srcs - 1].src = nir_src_for_ssa(nir_imm_int(b, plane));
257 tex->src[tex->num_srcs - 1].src_type = nir_tex_src_plane;
258
259 tex->sampler_dim = old_tex->sampler_dim;
260 tex->dest_type = old_tex->dest_type;
261
262 tex->op = old_tex->op;
263 tex->coord_components = old_tex->coord_components;
264 tex->is_new_style_shadow = old_tex->is_new_style_shadow;
265 tex->component = old_tex->component;
266
267 tex->texture_index = old_tex->texture_index;
268 tex->texture_array_size = old_tex->texture_array_size;
269 tex->texture = nir_deref_var_clone(old_tex->texture, tex);
270
271 tex->sampler_index = old_tex->sampler_index;
272 tex->sampler = nir_deref_var_clone(old_tex->sampler, tex);
273
274 nir_ssa_dest_init(&tex->instr, &tex->dest,
275 old_tex->dest.ssa.num_components,
276 nir_dest_bit_size(old_tex->dest), NULL);
277 nir_builder_instr_insert(b, &tex->instr);
278
279 return &tex->dest.ssa;
280 }
281
282 static unsigned
channel_to_component(enum isl_channel_select channel)283 channel_to_component(enum isl_channel_select channel)
284 {
285 switch (channel) {
286 case ISL_CHANNEL_SELECT_RED:
287 return 0;
288 case ISL_CHANNEL_SELECT_GREEN:
289 return 1;
290 case ISL_CHANNEL_SELECT_BLUE:
291 return 2;
292 case ISL_CHANNEL_SELECT_ALPHA:
293 return 3;
294 default:
295 unreachable("invalid channel");
296 return 0;
297 }
298 }
299
300 static enum isl_channel_select
swizzle_channel(struct isl_swizzle swizzle,unsigned channel)301 swizzle_channel(struct isl_swizzle swizzle, unsigned channel)
302 {
303 switch (channel) {
304 case 0:
305 return swizzle.r;
306 case 1:
307 return swizzle.g;
308 case 2:
309 return swizzle.b;
310 case 3:
311 return swizzle.a;
312 default:
313 unreachable("invalid channel");
314 return 0;
315 }
316 }
317
318 static bool
try_lower_tex_ycbcr(struct anv_pipeline * pipeline,nir_builder * builder,nir_tex_instr * tex)319 try_lower_tex_ycbcr(struct anv_pipeline *pipeline,
320 nir_builder *builder,
321 nir_tex_instr *tex)
322 {
323 nir_variable *var = tex->texture->var;
324 const struct anv_descriptor_set_layout *set_layout =
325 pipeline->layout->set[var->data.descriptor_set].layout;
326 const struct anv_descriptor_set_binding_layout *binding =
327 &set_layout->binding[var->data.binding];
328
329 /* For the following instructions, we don't apply any change and let the
330 * instruction apply to the first plane.
331 */
332 if (tex->op == nir_texop_txs ||
333 tex->op == nir_texop_query_levels ||
334 tex->op == nir_texop_lod)
335 return false;
336
337 if (binding->immutable_samplers == NULL)
338 return false;
339
340 unsigned texture_index = tex->texture_index;
341 if (tex->texture->deref.child) {
342 assert(tex->texture->deref.child->deref_type == nir_deref_type_array);
343 nir_deref_array *deref_array = nir_deref_as_array(tex->texture->deref.child);
344 if (deref_array->deref_array_type != nir_deref_array_type_direct)
345 return false;
346 size_t hw_binding_size =
347 anv_descriptor_set_binding_layout_get_hw_size(binding);
348 texture_index += MIN2(deref_array->base_offset, hw_binding_size - 1);
349 }
350 const struct anv_sampler *sampler =
351 binding->immutable_samplers[texture_index];
352
353 if (sampler->conversion == NULL)
354 return false;
355
356 struct ycbcr_state state = {
357 .builder = builder,
358 .origin_tex = tex,
359 .conversion = sampler->conversion,
360 };
361
362 builder->cursor = nir_before_instr(&tex->instr);
363
364 const struct anv_format *format = state.conversion->format;
365 const struct isl_format_layout *y_isl_layout = NULL;
366 for (uint32_t p = 0; p < format->n_planes; p++) {
367 if (!format->planes[p].has_chroma)
368 y_isl_layout = isl_format_get_layout(format->planes[p].isl_format);
369 }
370 assert(y_isl_layout != NULL);
371 uint8_t y_bpc = y_isl_layout->channels_array[0].bits;
372
373 /* |ycbcr_comp| holds components in the order : Cr-Y-Cb */
374 nir_ssa_def *ycbcr_comp[5] = { NULL, NULL, NULL,
375 /* Use extra 2 channels for following swizzle */
376 nir_imm_float(builder, 1.0f),
377 nir_imm_float(builder, 0.0f),
378 };
379 uint8_t ycbcr_bpcs[5];
380 memset(ycbcr_bpcs, y_bpc, sizeof(ycbcr_bpcs));
381
382 /* Go through all the planes and gather the samples into a |ycbcr_comp|
383 * while applying a swizzle required by the spec:
384 *
385 * R, G, B should respectively map to Cr, Y, Cb
386 */
387 for (uint32_t p = 0; p < format->n_planes; p++) {
388 const struct anv_format_plane *plane_format = &format->planes[p];
389 nir_ssa_def *plane_sample = create_plane_tex_instr_implicit(&state, p);
390
391 for (uint32_t pc = 0; pc < 4; pc++) {
392 enum isl_channel_select ycbcr_swizzle =
393 swizzle_channel(plane_format->ycbcr_swizzle, pc);
394 if (ycbcr_swizzle == ISL_CHANNEL_SELECT_ZERO)
395 continue;
396
397 unsigned ycbcr_component = channel_to_component(ycbcr_swizzle);
398 ycbcr_comp[ycbcr_component] = nir_channel(builder, plane_sample, pc);
399
400 /* Also compute the number of bits for each component. */
401 const struct isl_format_layout *isl_layout =
402 isl_format_get_layout(plane_format->isl_format);
403 ycbcr_bpcs[ycbcr_component] = isl_layout->channels_array[pc].bits;
404 }
405 }
406
407 /* Now remaps components to the order specified by the conversion. */
408 nir_ssa_def *swizzled_comp[4] = { NULL, };
409 uint32_t swizzled_bpcs[4] = { 0, };
410
411 for (uint32_t i = 0; i < ARRAY_SIZE(state.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 = state.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_ssa_def *result = nir_vec(builder, swizzled_comp, 4);
433 if (state.conversion->ycbcr_model != VK_SAMPLER_YCBCR_MODEL_CONVERSION_RGB_IDENTITY_KHR)
434 result = convert_ycbcr(&state, result, swizzled_bpcs);
435
436 nir_ssa_def_rewrite_uses(&tex->dest.ssa, nir_src_for_ssa(result));
437 nir_instr_remove(&tex->instr);
438
439 return true;
440 }
441
442 bool
anv_nir_lower_ycbcr_textures(nir_shader * shader,struct anv_pipeline * pipeline)443 anv_nir_lower_ycbcr_textures(nir_shader *shader, struct anv_pipeline *pipeline)
444 {
445 bool progress = false;
446
447 nir_foreach_function(function, shader) {
448 if (!function->impl)
449 continue;
450
451 bool function_progress = false;
452 nir_builder builder;
453 nir_builder_init(&builder, function->impl);
454
455 nir_foreach_block(block, function->impl) {
456 nir_foreach_instr_safe(instr, block) {
457 if (instr->type != nir_instr_type_tex)
458 continue;
459
460 nir_tex_instr *tex = nir_instr_as_tex(instr);
461 function_progress |= try_lower_tex_ycbcr(pipeline, &builder, tex);
462 }
463 }
464
465 if (function_progress) {
466 nir_metadata_preserve(function->impl,
467 nir_metadata_block_index |
468 nir_metadata_dominance);
469 }
470
471 progress |= function_progress;
472 }
473
474 return progress;
475 }
476