1 /*
2 * Copyright © Microsoft 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 "nir.h"
25 #include "nir_clc_helpers.h"
26 #include "nir_serialize.h"
27 #include "glsl_types.h"
28 #include "clc_compiler.h"
29 #include "clc_helpers.h"
30 #include "clc_nir.h"
31 #include "../compiler/dxil_nir.h"
32 #include "../compiler/dxil_nir_lower_int_samplers.h"
33 #include "../compiler/nir_to_dxil.h"
34
35 #include "util/u_debug.h"
36 #include <util/u_math.h>
37 #include "spirv/nir_spirv.h"
38 #include "spirv/spirv_info.h"
39 #include "nir_builder.h"
40 #include "nir_builtin_builder.h"
41
42 #include "git_sha1.h"
43
44 struct clc_image_lower_context
45 {
46 struct clc_dxil_metadata *metadata;
47 unsigned *num_srvs;
48 unsigned *num_uavs;
49 nir_deref_instr *deref;
50 unsigned num_buf_ids;
51 int metadata_index;
52 };
53
54 static int
lower_image_deref_impl(nir_builder * b,struct clc_image_lower_context * context,const struct glsl_type * new_var_type,nir_variable_mode var_mode,unsigned * num_bindings)55 lower_image_deref_impl(nir_builder *b, struct clc_image_lower_context *context,
56 const struct glsl_type *new_var_type,
57 nir_variable_mode var_mode,
58 unsigned *num_bindings)
59 {
60 nir_variable *in_var = nir_deref_instr_get_variable(context->deref);
61 nir_foreach_variable_with_modes(var, b->shader, var_mode) {
62 // Check if we've already created a variable for this image
63 if (var->data.driver_location == in_var->data.driver_location &&
64 var->type == new_var_type)
65 return var->data.binding;
66 }
67 nir_variable *image = nir_variable_create(b->shader, var_mode, new_var_type, NULL);
68 image->data.access = in_var->data.access;
69 image->data.binding = in_var->data.binding;
70 image->data.driver_location = in_var->data.driver_location;
71 if (context->num_buf_ids > 0) {
72 // Need to assign a new binding
73 context->metadata->args[context->metadata_index].
74 image.buf_ids[context->num_buf_ids] = image->data.binding = (*num_bindings)++;
75 }
76 context->num_buf_ids++;
77 return image->data.binding;
78 }
79
80 static int
lower_read_only_image_deref(nir_builder * b,struct clc_image_lower_context * context,nir_alu_type image_type)81 lower_read_only_image_deref(nir_builder *b, struct clc_image_lower_context *context,
82 nir_alu_type image_type)
83 {
84 nir_variable *in_var = nir_deref_instr_get_variable(context->deref);
85
86 // Non-writeable images should be converted to samplers,
87 // since they may have texture operations done on them
88 const struct glsl_type *new_var_type =
89 glsl_texture_type(glsl_get_sampler_dim(in_var->type),
90 glsl_sampler_type_is_array(in_var->type),
91 nir_get_glsl_base_type_for_nir_type(image_type | 32));
92 return lower_image_deref_impl(b, context, new_var_type, nir_var_uniform, context->num_srvs);
93 }
94
95 static int
lower_read_write_image_deref(nir_builder * b,struct clc_image_lower_context * context,nir_alu_type image_type)96 lower_read_write_image_deref(nir_builder *b, struct clc_image_lower_context *context,
97 nir_alu_type image_type)
98 {
99 nir_variable *in_var = nir_deref_instr_get_variable(context->deref);
100 const struct glsl_type *new_var_type =
101 glsl_image_type(glsl_get_sampler_dim(in_var->type),
102 glsl_sampler_type_is_array(in_var->type),
103 nir_get_glsl_base_type_for_nir_type(image_type | 32));
104 return lower_image_deref_impl(b, context, new_var_type, nir_var_image, context->num_uavs);
105 }
106
107 static void
clc_lower_input_image_deref(nir_builder * b,struct clc_image_lower_context * context)108 clc_lower_input_image_deref(nir_builder *b, struct clc_image_lower_context *context)
109 {
110 // The input variable here isn't actually an image, it's just the
111 // image format data.
112 //
113 // For every use of an image in a different way, we'll add an
114 // appropriate image to match it. That can result in up to
115 // 3 images (float4, int4, uint4) for each image. Only one of these
116 // formats will actually produce correct data, but a single kernel
117 // could use runtime conditionals to potentially access any of them.
118 //
119 // If the image is used in a query that doesn't have a corresponding
120 // DXIL intrinsic (CL image channel order or channel format), then
121 // we'll add a kernel input for that data that'll be lowered by the
122 // explicit IO pass later on.
123 //
124 // After all that, we can remove the image input variable and deref.
125
126 enum image_type {
127 FLOAT4,
128 INT4,
129 UINT4,
130 IMAGE_TYPE_COUNT
131 };
132
133 int image_bindings[IMAGE_TYPE_COUNT] = {-1, -1, -1};
134 nir_def *format_deref_dest = NULL, *order_deref_dest = NULL;
135
136 nir_variable *in_var = nir_deref_instr_get_variable(context->deref);
137
138 context->metadata_index = 0;
139 while (context->metadata->args[context->metadata_index].offset != in_var->data.driver_location)
140 context->metadata_index++;
141
142 context->num_buf_ids = context->metadata->args[context->metadata_index].image.num_buf_ids;
143
144 /* Do this in 2 passes:
145 * 1. When encountering a strongly-typed access (load/store), replace the deref
146 * with one that references an appropriately typed variable. When encountering
147 * an untyped access (size query), if we have a strongly-typed variable already,
148 * replace the deref to point to it.
149 * 2. If there's any references left, they should all be untyped. If we found
150 * a strongly-typed access later in the 1st pass, then just replace the reference.
151 * If we didn't, e.g. the resource is only used for a size query, then pick an
152 * arbitrary type for it.
153 */
154 for (int pass = 0; pass < 2; ++pass) {
155 nir_foreach_use_safe(src, &context->deref->def) {
156 enum image_type type;
157
158 if (nir_src_parent_instr(src)->type == nir_instr_type_intrinsic) {
159 nir_intrinsic_instr *intrinsic = nir_instr_as_intrinsic(nir_src_parent_instr(src));
160 nir_alu_type dest_type;
161
162 b->cursor = nir_before_instr(&intrinsic->instr);
163
164 switch (intrinsic->intrinsic) {
165 case nir_intrinsic_image_deref_load:
166 case nir_intrinsic_image_deref_store: {
167 dest_type = intrinsic->intrinsic == nir_intrinsic_image_deref_load ?
168 nir_intrinsic_dest_type(intrinsic) : nir_intrinsic_src_type(intrinsic);
169
170 switch (nir_alu_type_get_base_type(dest_type)) {
171 case nir_type_float: type = FLOAT4; break;
172 case nir_type_int: type = INT4; break;
173 case nir_type_uint: type = UINT4; break;
174 default: unreachable("Unsupported image type for load.");
175 }
176
177 int image_binding = image_bindings[type];
178 if (image_binding < 0) {
179 image_binding = image_bindings[type] =
180 lower_read_write_image_deref(b, context, dest_type);
181 }
182
183 assert((in_var->data.access & ACCESS_NON_WRITEABLE) == 0);
184 nir_rewrite_image_intrinsic(intrinsic, nir_imm_int(b, image_binding), false);
185 break;
186 }
187
188 case nir_intrinsic_image_deref_size: {
189 int image_binding = -1;
190 for (unsigned i = 0; i < IMAGE_TYPE_COUNT; ++i) {
191 if (image_bindings[i] >= 0) {
192 image_binding = image_bindings[i];
193 break;
194 }
195 }
196 if (image_binding < 0) {
197 // Skip for now and come back to it
198 if (pass == 0)
199 break;
200
201 type = FLOAT4;
202 image_binding = image_bindings[type] =
203 lower_read_write_image_deref(b, context, nir_type_float32);
204 }
205
206 assert((in_var->data.access & ACCESS_NON_WRITEABLE) == 0);
207 nir_rewrite_image_intrinsic(intrinsic, nir_imm_int(b, image_binding), false);
208 break;
209 }
210
211 case nir_intrinsic_image_deref_format:
212 case nir_intrinsic_image_deref_order: {
213 nir_def **cached_deref = intrinsic->intrinsic == nir_intrinsic_image_deref_format ?
214 &format_deref_dest : &order_deref_dest;
215 if (!*cached_deref) {
216 unsigned driver_location = in_var->data.driver_location;
217 if (intrinsic->intrinsic == nir_intrinsic_image_deref_format) {
218 /* Match cl_image_format { image_channel_order, image_channel_data_type }; */
219 driver_location += 4;
220 }
221
222 nir_variable *new_input = NULL;
223 nir_foreach_variable_with_modes(var, b->shader, nir_var_uniform) {
224 if (var->data.driver_location == driver_location &&
225 var->type == glsl_uint_type()) {
226 new_input = var;
227 break;
228 }
229 }
230 if (!new_input) {
231 new_input = nir_variable_create(b->shader, nir_var_uniform, glsl_uint_type(), NULL);
232 new_input->data.driver_location = driver_location;
233 }
234
235 b->cursor = nir_after_instr(&context->deref->instr);
236 *cached_deref = nir_load_var(b, new_input);
237 }
238
239 /* No actual intrinsic needed here, just reference the loaded variable */
240 nir_def_replace(&intrinsic->def, *cached_deref);
241 break;
242 }
243
244 default:
245 unreachable("Unsupported image intrinsic");
246 }
247 } else if (nir_src_parent_instr(src)->type == nir_instr_type_tex) {
248 assert(in_var->data.access & ACCESS_NON_WRITEABLE);
249 nir_tex_instr *tex = nir_instr_as_tex(nir_src_parent_instr(src));
250
251 switch (nir_alu_type_get_base_type(tex->dest_type)) {
252 case nir_type_float: type = FLOAT4; break;
253 case nir_type_int: type = INT4; break;
254 case nir_type_uint: type = UINT4; break;
255 default: unreachable("Unsupported image format for sample.");
256 }
257
258 int image_binding = image_bindings[type];
259 if (image_binding < 0) {
260 image_binding = image_bindings[type] =
261 lower_read_only_image_deref(b, context, tex->dest_type);
262 }
263
264 nir_tex_instr_remove_src(tex, nir_tex_instr_src_index(tex, nir_tex_src_texture_deref));
265 tex->texture_index = image_binding;
266 }
267 }
268 }
269
270 context->metadata->args[context->metadata_index].image.num_buf_ids = context->num_buf_ids;
271
272 nir_instr_remove(&context->deref->instr);
273 }
274
275 static void
clc_lower_images(nir_shader * nir,struct clc_image_lower_context * context)276 clc_lower_images(nir_shader *nir, struct clc_image_lower_context *context)
277 {
278 nir_foreach_function(func, nir) {
279 if (!func->is_entrypoint)
280 continue;
281 assert(func->impl);
282
283 nir_builder b = nir_builder_create(func->impl);
284
285 nir_foreach_block(block, func->impl) {
286 nir_foreach_instr_safe(instr, block) {
287 if (instr->type == nir_instr_type_deref) {
288 context->deref = nir_instr_as_deref(instr);
289
290 if (glsl_type_is_image(context->deref->type)) {
291 assert(context->deref->deref_type == nir_deref_type_var);
292 clc_lower_input_image_deref(&b, context);
293 }
294 }
295 }
296 }
297 }
298
299 nir_foreach_variable_with_modes_safe(var, nir, nir_var_image) {
300 if (glsl_type_is_image(var->type) && glsl_get_sampler_result_type(var->type) == GLSL_TYPE_VOID)
301 exec_node_remove(&var->node);
302 }
303 }
304
305 static void
clc_lower_64bit_semantics(nir_shader * nir)306 clc_lower_64bit_semantics(nir_shader *nir)
307 {
308 nir_foreach_function_impl(impl, nir) {
309 nir_builder b = nir_builder_create(impl);
310
311 nir_foreach_block(block, impl) {
312 nir_foreach_instr_safe(instr, block) {
313 if (instr->type == nir_instr_type_intrinsic) {
314 nir_intrinsic_instr *intrinsic = nir_instr_as_intrinsic(instr);
315 switch (intrinsic->intrinsic) {
316 case nir_intrinsic_load_global_invocation_id:
317 case nir_intrinsic_load_base_global_invocation_id:
318 case nir_intrinsic_load_local_invocation_id:
319 case nir_intrinsic_load_workgroup_id:
320 case nir_intrinsic_load_base_workgroup_id:
321 case nir_intrinsic_load_num_workgroups:
322 break;
323 default:
324 continue;
325 }
326
327 if (nir_instr_def(instr)->bit_size != 64)
328 continue;
329
330 intrinsic->def.bit_size = 32;
331 b.cursor = nir_after_instr(instr);
332
333 nir_def *i64 = nir_u2u64(&b, &intrinsic->def);
334 nir_def_rewrite_uses_after(
335 &intrinsic->def,
336 i64,
337 i64->parent_instr);
338 }
339 }
340 }
341 }
342 }
343
344 static void
clc_lower_nonnormalized_samplers(nir_shader * nir,const dxil_wrap_sampler_state * states)345 clc_lower_nonnormalized_samplers(nir_shader *nir,
346 const dxil_wrap_sampler_state *states)
347 {
348 nir_foreach_function(func, nir) {
349 if (!func->is_entrypoint)
350 continue;
351 assert(func->impl);
352
353 nir_builder b = nir_builder_create(func->impl);
354
355 nir_foreach_block(block, func->impl) {
356 nir_foreach_instr_safe(instr, block) {
357 if (instr->type != nir_instr_type_tex)
358 continue;
359 nir_tex_instr *tex = nir_instr_as_tex(instr);
360
361 int sampler_src_idx = nir_tex_instr_src_index(tex, nir_tex_src_sampler_deref);
362 if (sampler_src_idx == -1)
363 continue;
364
365 nir_src *sampler_src = &tex->src[sampler_src_idx].src;
366 assert(sampler_src->ssa->parent_instr->type == nir_instr_type_deref);
367 nir_variable *sampler = nir_deref_instr_get_variable(
368 nir_instr_as_deref(sampler_src->ssa->parent_instr));
369
370 // If the sampler returns ints, we'll handle this in the int lowering pass
371 if (nir_alu_type_get_base_type(tex->dest_type) != nir_type_float)
372 continue;
373
374 // If sampler uses normalized coords, nothing to do
375 if (!states[sampler->data.binding].is_nonnormalized_coords)
376 continue;
377
378 b.cursor = nir_before_instr(&tex->instr);
379
380 int coords_idx = nir_tex_instr_src_index(tex, nir_tex_src_coord);
381 assert(coords_idx != -1);
382 nir_def *coords =
383 tex->src[coords_idx].src.ssa;
384
385 nir_def *txs = nir_i2f32(&b, nir_get_texture_size(&b, tex));
386
387 // Normalize coords for tex
388 nir_def *scale = nir_frcp(&b, txs);
389 nir_def *comps[4];
390 for (unsigned i = 0; i < coords->num_components; ++i) {
391 comps[i] = nir_channel(&b, coords, i);
392 if (tex->is_array && i == coords->num_components - 1) {
393 // Don't scale the array index, but do clamp it
394 comps[i] = nir_fround_even(&b, comps[i]);
395 comps[i] = nir_fmax(&b, comps[i], nir_imm_float(&b, 0.0f));
396 comps[i] = nir_fmin(&b, comps[i], nir_fadd_imm(&b, nir_channel(&b, txs, i), -1.0f));
397 break;
398 }
399
400 // The CTS is pretty clear that this value has to be floored for nearest sampling
401 // but must not be for linear sampling.
402 if (!states[sampler->data.binding].is_linear_filtering)
403 comps[i] = nir_fadd_imm(&b, nir_ffloor(&b, comps[i]), 0.5f);
404 comps[i] = nir_fmul(&b, comps[i], nir_channel(&b, scale, i));
405 }
406 nir_def *normalized_coords = nir_vec(&b, comps, coords->num_components);
407 nir_src_rewrite(&tex->src[coords_idx].src, normalized_coords);
408 }
409 }
410 }
411 }
412
413 static nir_variable *
add_kernel_inputs_var(struct clc_dxil_object * dxil,nir_shader * nir,unsigned * cbv_id)414 add_kernel_inputs_var(struct clc_dxil_object *dxil, nir_shader *nir,
415 unsigned *cbv_id)
416 {
417 if (!dxil->kernel->num_args)
418 return NULL;
419
420 unsigned size = 0;
421
422 nir_foreach_variable_with_modes(var, nir, nir_var_uniform)
423 size = MAX2(size,
424 var->data.driver_location +
425 glsl_get_cl_size(var->type));
426
427 size = align(size, 4);
428
429 const struct glsl_type *array_type = glsl_array_type(glsl_uint_type(), size / 4, 4);
430 const struct glsl_struct_field field = { array_type, "arr" };
431 nir_variable *var =
432 nir_variable_create(nir, nir_var_mem_ubo,
433 glsl_struct_type(&field, 1, "kernel_inputs", false),
434 "kernel_inputs");
435 var->data.binding = (*cbv_id)++;
436 var->data.how_declared = nir_var_hidden;
437 return var;
438 }
439
440 static nir_variable *
add_work_properties_var(struct clc_dxil_object * dxil,struct nir_shader * nir,unsigned * cbv_id)441 add_work_properties_var(struct clc_dxil_object *dxil,
442 struct nir_shader *nir, unsigned *cbv_id)
443 {
444 const struct glsl_type *array_type =
445 glsl_array_type(glsl_uint_type(),
446 sizeof(struct clc_work_properties_data) / sizeof(unsigned),
447 sizeof(unsigned));
448 const struct glsl_struct_field field = { array_type, "arr" };
449 nir_variable *var =
450 nir_variable_create(nir, nir_var_mem_ubo,
451 glsl_struct_type(&field, 1, "kernel_work_properties", false),
452 "kernel_work_properies");
453 var->data.binding = (*cbv_id)++;
454 var->data.how_declared = nir_var_hidden;
455 return var;
456 }
457
458 static void
clc_lower_constant_to_ssbo(nir_shader * nir,const struct clc_kernel_info * kerninfo,unsigned * uav_id)459 clc_lower_constant_to_ssbo(nir_shader *nir,
460 const struct clc_kernel_info *kerninfo, unsigned *uav_id)
461 {
462 /* Update UBO vars and assign them a binding. */
463 nir_foreach_variable_with_modes(var, nir, nir_var_mem_constant) {
464 var->data.mode = nir_var_mem_ssbo;
465 var->data.binding = (*uav_id)++;
466 }
467
468 /* And finally patch all the derefs referincing the constant
469 * variables/pointers.
470 */
471 nir_foreach_function(func, nir) {
472 if (!func->is_entrypoint)
473 continue;
474
475 assert(func->impl);
476
477 nir_foreach_block(block, func->impl) {
478 nir_foreach_instr(instr, block) {
479 if (instr->type != nir_instr_type_deref)
480 continue;
481
482 nir_deref_instr *deref = nir_instr_as_deref(instr);
483
484 if (deref->modes != nir_var_mem_constant)
485 continue;
486
487 deref->modes = nir_var_mem_ssbo;
488 }
489 }
490 }
491 }
492
493 static void
clc_change_variable_mode(nir_shader * nir,nir_variable_mode from,nir_variable_mode to)494 clc_change_variable_mode(nir_shader *nir, nir_variable_mode from, nir_variable_mode to)
495 {
496 nir_foreach_variable_with_modes(var, nir, from)
497 var->data.mode = to;
498
499 nir_foreach_function(func, nir) {
500 if (!func->is_entrypoint)
501 continue;
502
503 assert(func->impl);
504
505 nir_foreach_block(block, func->impl) {
506 nir_foreach_instr(instr, block) {
507 if (instr->type != nir_instr_type_deref)
508 continue;
509
510 nir_deref_instr *deref = nir_instr_as_deref(instr);
511
512 if (deref->modes != from)
513 continue;
514
515 deref->modes = to;
516 }
517 }
518 }
519 }
520
521 static void
copy_const_initializer(const nir_constant * constant,const struct glsl_type * type,uint8_t * data)522 copy_const_initializer(const nir_constant *constant, const struct glsl_type *type,
523 uint8_t *data)
524 {
525 if (glsl_type_is_array(type)) {
526 const struct glsl_type *elm_type = glsl_get_array_element(type);
527 unsigned step_size = glsl_get_explicit_stride(type);
528
529 for (unsigned i = 0; i < constant->num_elements; i++) {
530 copy_const_initializer(constant->elements[i], elm_type,
531 data + (i * step_size));
532 }
533 } else if (glsl_type_is_struct(type)) {
534 for (unsigned i = 0; i < constant->num_elements; i++) {
535 const struct glsl_type *elm_type = glsl_get_struct_field(type, i);
536 int offset = glsl_get_struct_field_offset(type, i);
537 copy_const_initializer(constant->elements[i], elm_type, data + offset);
538 }
539 } else {
540 assert(glsl_type_is_vector_or_scalar(type));
541
542 for (unsigned i = 0; i < glsl_get_components(type); i++) {
543 switch (glsl_get_bit_size(type)) {
544 case 64:
545 *((uint64_t *)data) = constant->values[i].u64;
546 break;
547 case 32:
548 *((uint32_t *)data) = constant->values[i].u32;
549 break;
550 case 16:
551 *((uint16_t *)data) = constant->values[i].u16;
552 break;
553 case 8:
554 *((uint8_t *)data) = constant->values[i].u8;
555 break;
556 default:
557 unreachable("Invalid base type");
558 }
559
560 data += glsl_get_bit_size(type) / 8;
561 }
562 }
563 }
564
565 static enum dxil_tex_wrap
wrap_from_cl_addressing(unsigned addressing_mode)566 wrap_from_cl_addressing(unsigned addressing_mode)
567 {
568 switch (addressing_mode)
569 {
570 default:
571 case SAMPLER_ADDRESSING_MODE_NONE:
572 case SAMPLER_ADDRESSING_MODE_CLAMP:
573 // Since OpenCL's only border color is 0's and D3D specs out-of-bounds loads to return 0, don't apply any wrap mode
574 return (enum dxil_tex_wrap)-1;
575 case SAMPLER_ADDRESSING_MODE_CLAMP_TO_EDGE: return DXIL_TEX_WRAP_CLAMP_TO_EDGE;
576 case SAMPLER_ADDRESSING_MODE_REPEAT: return DXIL_TEX_WRAP_REPEAT;
577 case SAMPLER_ADDRESSING_MODE_REPEAT_MIRRORED: return DXIL_TEX_WRAP_MIRROR_REPEAT;
578 }
579 }
580
shader_has_double(nir_shader * nir)581 static bool shader_has_double(nir_shader *nir)
582 {
583 foreach_list_typed(nir_function, func, node, &nir->functions) {
584 if (!func->is_entrypoint)
585 continue;
586
587 assert(func->impl);
588
589 nir_foreach_block(block, func->impl) {
590 nir_foreach_instr_safe(instr, block) {
591 if (instr->type != nir_instr_type_alu)
592 continue;
593
594 nir_alu_instr *alu = nir_instr_as_alu(instr);
595 const nir_op_info *info = &nir_op_infos[alu->op];
596
597 if (info->output_type & nir_type_float &&
598 alu->def.bit_size == 64)
599 return true;
600 }
601 }
602 }
603
604 return false;
605 }
606
607 struct clc_libclc {
608 const nir_shader *libclc_nir;
609 };
610
611 struct clc_libclc *
clc_libclc_new(const struct clc_logger * logger,const struct clc_libclc_options * options)612 clc_libclc_new(const struct clc_logger *logger, const struct clc_libclc_options *options)
613 {
614 struct clc_libclc *ctx = rzalloc(NULL, struct clc_libclc);
615 if (!ctx) {
616 clc_error(logger, "D3D12: failed to allocate a clc_libclc");
617 return NULL;
618 }
619
620 const struct spirv_capabilities libclc_spirv_caps = {
621 .Addresses = true,
622 .Float64 = true,
623 .Int8 = true,
624 .Int16 = true,
625 .Int64 = true,
626 .Kernel = true,
627 .Linkage = true,
628 };
629 const struct spirv_to_nir_options libclc_spirv_options = {
630 .environment = NIR_SPIRV_OPENCL,
631 .create_library = true,
632 .constant_addr_format = nir_address_format_32bit_index_offset_pack64,
633 .global_addr_format = nir_address_format_32bit_index_offset_pack64,
634 .shared_addr_format = nir_address_format_32bit_offset_as_64bit,
635 .temp_addr_format = nir_address_format_32bit_offset_as_64bit,
636 .float_controls_execution_mode = FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32,
637 .capabilities = &libclc_spirv_caps,
638 };
639
640 glsl_type_singleton_init_or_ref();
641 bool optimize = options && options->optimize;
642 nir_shader *s =
643 nir_load_libclc_shader(64, NULL, &libclc_spirv_options, options->nir_options, optimize);
644 if (!s) {
645 clc_error(logger, "D3D12: spirv_to_nir failed on libclc blob");
646 ralloc_free(ctx);
647 return NULL;
648 }
649
650 ralloc_steal(ctx, s);
651 ctx->libclc_nir = s;
652
653 return ctx;
654 }
655
clc_free_libclc(struct clc_libclc * ctx)656 void clc_free_libclc(struct clc_libclc *ctx)
657 {
658 ralloc_free(ctx);
659 glsl_type_singleton_decref();
660 }
661
clc_libclc_get_clc_shader(struct clc_libclc * ctx)662 const nir_shader *clc_libclc_get_clc_shader(struct clc_libclc *ctx)
663 {
664 return ctx->libclc_nir;
665 }
666
clc_libclc_serialize(struct clc_libclc * context,void ** serialized,size_t * serialized_size)667 void clc_libclc_serialize(struct clc_libclc *context,
668 void **serialized,
669 size_t *serialized_size)
670 {
671 struct blob tmp;
672 blob_init(&tmp);
673 nir_serialize(&tmp, context->libclc_nir, true);
674
675 blob_finish_get_buffer(&tmp, serialized, serialized_size);
676 }
677
clc_libclc_free_serialized(void * serialized)678 void clc_libclc_free_serialized(void *serialized)
679 {
680 free(serialized);
681 }
682
683 struct clc_libclc *
clc_libclc_deserialize(const void * serialized,size_t serialized_size)684 clc_libclc_deserialize(const void *serialized, size_t serialized_size)
685 {
686 struct clc_libclc *ctx = rzalloc(NULL, struct clc_libclc);
687 if (!ctx) {
688 return NULL;
689 }
690
691 glsl_type_singleton_init_or_ref();
692
693 struct blob_reader tmp;
694 blob_reader_init(&tmp, serialized, serialized_size);
695
696 nir_shader *s = nir_deserialize(NULL, NULL, &tmp);
697 if (!s) {
698 ralloc_free(ctx);
699 return NULL;
700 }
701
702 ralloc_steal(ctx, s);
703 ctx->libclc_nir = s;
704
705 return ctx;
706 }
707
708 struct clc_libclc *
clc_libclc_new_dxil(const struct clc_logger * logger,const struct clc_libclc_dxil_options * options)709 clc_libclc_new_dxil(const struct clc_logger *logger,
710 const struct clc_libclc_dxil_options *options)
711 {
712 struct clc_libclc_options clc_options = {
713 .optimize = options->optimize,
714 .nir_options = dxil_get_base_nir_compiler_options(),
715 };
716
717 return clc_libclc_new(logger, &clc_options);
718 }
719
720 bool
clc_spirv_to_dxil(struct clc_libclc * lib,const struct clc_binary * linked_spirv,const struct clc_parsed_spirv * parsed_data,const char * entrypoint,const struct clc_runtime_kernel_conf * conf,const struct clc_spirv_specialization_consts * consts,const struct clc_logger * logger,struct clc_dxil_object * out_dxil)721 clc_spirv_to_dxil(struct clc_libclc *lib,
722 const struct clc_binary *linked_spirv,
723 const struct clc_parsed_spirv *parsed_data,
724 const char *entrypoint,
725 const struct clc_runtime_kernel_conf *conf,
726 const struct clc_spirv_specialization_consts *consts,
727 const struct clc_logger *logger,
728 struct clc_dxil_object *out_dxil)
729 {
730 struct nir_shader *nir;
731
732 for (unsigned i = 0; i < parsed_data->num_kernels; i++) {
733 if (!strcmp(parsed_data->kernels[i].name, entrypoint)) {
734 out_dxil->kernel = &parsed_data->kernels[i];
735 break;
736 }
737 }
738
739 if (!out_dxil->kernel) {
740 clc_error(logger, "no '%s' kernel found", entrypoint);
741 return false;
742 }
743
744 const struct spirv_capabilities libclc_spirv_caps = {
745 .Addresses = true,
746 .Float64 = true,
747 .Int8 = true,
748 .Int16 = true,
749 .Int64 = true,
750 .Kernel = true,
751 .ImageBasic = true,
752 .ImageReadWrite = true,
753 .LiteralSampler = true,
754
755 // These aren't fully supported, but silence warnings about them from
756 // code that doesn't really use them.
757 .Linkage = true,
758 .GenericPointer = true,
759 };
760 const struct spirv_to_nir_options spirv_options = {
761 .environment = NIR_SPIRV_OPENCL,
762 .clc_shader = clc_libclc_get_clc_shader(lib),
763 .constant_addr_format = nir_address_format_32bit_index_offset_pack64,
764 .global_addr_format = nir_address_format_32bit_index_offset_pack64,
765 .shared_addr_format = nir_address_format_32bit_offset_as_64bit,
766 .temp_addr_format = nir_address_format_32bit_offset_as_64bit,
767 .float_controls_execution_mode = FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32,
768 .printf = true,
769 .capabilities = &libclc_spirv_caps,
770 };
771 unsigned supported_int_sizes = (16 | 32 | 64);
772 unsigned supported_float_sizes = (16 | 32);
773 if (conf) {
774 supported_int_sizes &= ~conf->lower_bit_size;
775 supported_float_sizes &= ~conf->lower_bit_size;
776 }
777 nir_shader_compiler_options nir_options;
778 dxil_get_nir_compiler_options(&nir_options,
779 conf ? conf->max_shader_model : SHADER_MODEL_6_2,
780 supported_int_sizes,
781 supported_float_sizes);
782
783 glsl_type_singleton_init_or_ref();
784
785 nir = spirv_to_nir(linked_spirv->data, linked_spirv->size / 4,
786 consts ? (struct nir_spirv_specialization *)consts->specializations : NULL,
787 consts ? consts->num_specializations : 0,
788 MESA_SHADER_KERNEL, entrypoint,
789 &spirv_options,
790 &nir_options);
791 if (!nir) {
792 clc_error(logger, "spirv_to_nir() failed");
793 goto err_free_dxil;
794 }
795 nir->info.workgroup_size_variable = true;
796
797 NIR_PASS_V(nir, nir_lower_goto_ifs);
798 NIR_PASS_V(nir, nir_opt_dead_cf);
799
800 struct clc_dxil_metadata *metadata = &out_dxil->metadata;
801
802 metadata->args = calloc(out_dxil->kernel->num_args,
803 sizeof(*metadata->args));
804 if (!metadata->args) {
805 clc_error(logger, "failed to allocate arg positions");
806 goto err_free_dxil;
807 }
808
809 {
810 bool progress;
811 do
812 {
813 progress = false;
814 NIR_PASS(progress, nir, nir_copy_prop);
815 NIR_PASS(progress, nir, nir_opt_copy_prop_vars);
816 NIR_PASS(progress, nir, nir_opt_deref);
817 NIR_PASS(progress, nir, nir_opt_dce);
818 NIR_PASS(progress, nir, nir_opt_undef);
819 NIR_PASS(progress, nir, nir_opt_constant_folding);
820 NIR_PASS(progress, nir, nir_opt_cse);
821 NIR_PASS(progress, nir, nir_split_var_copies);
822 NIR_PASS(progress, nir, nir_lower_var_copies);
823 NIR_PASS(progress, nir, nir_lower_vars_to_ssa);
824 NIR_PASS(progress, nir, nir_opt_algebraic);
825 } while (progress);
826 }
827
828 // Inline all functions first.
829 // according to the comment on nir_inline_functions
830 NIR_PASS_V(nir, nir_lower_variable_initializers, nir_var_function_temp);
831 NIR_PASS_V(nir, nir_lower_returns);
832 NIR_PASS_V(nir, nir_link_shader_functions, clc_libclc_get_clc_shader(lib));
833 NIR_PASS_V(nir, nir_inline_functions);
834
835 // Pick off the single entrypoint that we want.
836 nir_remove_non_entrypoints(nir);
837
838 {
839 bool progress;
840 do
841 {
842 progress = false;
843 NIR_PASS(progress, nir, nir_copy_prop);
844 NIR_PASS(progress, nir, nir_opt_copy_prop_vars);
845 NIR_PASS(progress, nir, nir_opt_deref);
846 NIR_PASS(progress, nir, nir_opt_dce);
847 NIR_PASS(progress, nir, nir_opt_undef);
848 NIR_PASS(progress, nir, nir_opt_constant_folding);
849 NIR_PASS(progress, nir, nir_opt_cse);
850 NIR_PASS(progress, nir, nir_split_var_copies);
851 NIR_PASS(progress, nir, nir_lower_var_copies);
852 NIR_PASS(progress, nir, nir_lower_vars_to_ssa);
853 NIR_PASS(progress, nir, nir_opt_algebraic);
854 NIR_PASS(progress, nir, nir_opt_if, nir_opt_if_optimize_phi_true_false);
855 NIR_PASS(progress, nir, nir_opt_dead_cf);
856 NIR_PASS(progress, nir, nir_opt_remove_phis);
857 NIR_PASS(progress, nir, nir_opt_peephole_select, 8, true, true);
858 NIR_PASS(progress, nir, nir_lower_vec3_to_vec4, nir_var_mem_generic | nir_var_uniform);
859 NIR_PASS(progress, nir, nir_opt_memcpy);
860 } while (progress);
861 }
862
863 NIR_PASS_V(nir, nir_scale_fdiv);
864
865 /* 128 is the minimum value for CL_DEVICE_MAX_READ_IMAGE_ARGS and used by CLOn12 */
866 dxil_wrap_sampler_state int_sampler_states[128] = { {{0}} };
867 unsigned sampler_id = 0;
868
869 NIR_PASS_V(nir, nir_lower_variable_initializers, ~(nir_var_function_temp | nir_var_shader_temp));
870
871 // Ensure the printf struct has explicit types, but we'll throw away the scratch size, because we haven't
872 // necessarily removed all temp variables (e.g. the printf struct itself) at this point, so we'll rerun this later
873 assert(nir->scratch_size == 0);
874 NIR_PASS_V(nir, nir_lower_vars_to_explicit_types, nir_var_function_temp, glsl_get_cl_type_size_align);
875
876 nir_lower_printf_options printf_options = {
877 .max_buffer_size = 1024 * 1024
878 };
879 NIR_PASS_V(nir, nir_lower_printf, &printf_options);
880
881 metadata->printf.info_count = nir->printf_info_count;
882 metadata->printf.infos = calloc(nir->printf_info_count, sizeof(struct clc_printf_info));
883 for (unsigned i = 0; i < nir->printf_info_count; i++) {
884 metadata->printf.infos[i].str = malloc(nir->printf_info[i].string_size);
885 memcpy(metadata->printf.infos[i].str, nir->printf_info[i].strings, nir->printf_info[i].string_size);
886 metadata->printf.infos[i].num_args = nir->printf_info[i].num_args;
887 metadata->printf.infos[i].arg_sizes = malloc(nir->printf_info[i].num_args * sizeof(unsigned));
888 memcpy(metadata->printf.infos[i].arg_sizes, nir->printf_info[i].arg_sizes, nir->printf_info[i].num_args * sizeof(unsigned));
889 }
890
891 // For uniforms (kernel inputs, minus images), run this before adjusting variable list via image/sampler lowering
892 NIR_PASS_V(nir, nir_lower_vars_to_explicit_types, nir_var_uniform, glsl_get_cl_type_size_align);
893
894 // Calculate input offsets/metadata.
895 unsigned uav_id = 0;
896 nir_foreach_variable_with_modes(var, nir, nir_var_uniform) {
897 int i = var->data.location;
898 if (i < 0)
899 continue;
900
901 unsigned size = glsl_get_cl_size(var->type);
902
903 metadata->args[i].offset = var->data.driver_location;
904 metadata->args[i].size = size;
905 metadata->kernel_inputs_buf_size = MAX2(metadata->kernel_inputs_buf_size,
906 var->data.driver_location + size);
907 if (out_dxil->kernel->args[i].address_qualifier == CLC_KERNEL_ARG_ADDRESS_GLOBAL ||
908 out_dxil->kernel->args[i].address_qualifier == CLC_KERNEL_ARG_ADDRESS_CONSTANT) {
909 metadata->args[i].globconstptr.buf_id = var->data.binding = uav_id++;
910 } else if (glsl_type_is_sampler(var->type)) {
911 unsigned address_mode = conf ? conf->args[i].sampler.addressing_mode : 0u;
912 int_sampler_states[sampler_id].wrap[0] =
913 int_sampler_states[sampler_id].wrap[1] =
914 int_sampler_states[sampler_id].wrap[2] = wrap_from_cl_addressing(address_mode);
915 int_sampler_states[sampler_id].is_nonnormalized_coords =
916 conf ? !conf->args[i].sampler.normalized_coords : 0;
917 int_sampler_states[sampler_id].is_linear_filtering =
918 conf ? conf->args[i].sampler.linear_filtering : 0;
919 metadata->args[i].sampler.sampler_id = var->data.binding = sampler_id++;
920 }
921 }
922
923 unsigned num_global_inputs = uav_id;
924
925 // Before removing dead uniforms, dedupe inline samplers to make more dead uniforms
926 NIR_PASS_V(nir, nir_dedup_inline_samplers);
927 NIR_PASS_V(nir, nir_remove_dead_variables, nir_var_uniform | nir_var_mem_ubo |
928 nir_var_mem_constant | nir_var_function_temp | nir_var_image, NULL);
929
930 nir->scratch_size = 0;
931 NIR_PASS_V(nir, nir_lower_vars_to_explicit_types,
932 nir_var_mem_shared | nir_var_function_temp | nir_var_mem_global | nir_var_mem_constant,
933 glsl_get_cl_type_size_align);
934
935 // Lower memcpy - needs to wait until types are sized
936 {
937 bool progress;
938 do {
939 progress = false;
940 NIR_PASS(progress, nir, nir_opt_memcpy);
941 NIR_PASS(progress, nir, nir_copy_prop);
942 NIR_PASS(progress, nir, nir_opt_copy_prop_vars);
943 NIR_PASS(progress, nir, nir_opt_deref);
944 NIR_PASS(progress, nir, nir_opt_dce);
945 NIR_PASS(progress, nir, nir_split_var_copies);
946 NIR_PASS(progress, nir, nir_lower_var_copies);
947 NIR_PASS(progress, nir, nir_lower_vars_to_ssa);
948 NIR_PASS(progress, nir, nir_opt_constant_folding);
949 NIR_PASS(progress, nir, nir_opt_cse);
950 } while (progress);
951 }
952 NIR_PASS_V(nir, nir_lower_memcpy);
953
954 NIR_PASS_V(nir, clc_nir_lower_global_pointers_to_constants);
955
956 // Attempt to preserve derefs to constants by moving them to shader_temp
957 NIR_PASS_V(nir, dxil_nir_lower_constant_to_temp);
958 // While inserting new var derefs for our "logical" addressing mode, temporarily
959 // switch the pointer size to 32-bit.
960 nir->info.cs.ptr_size = 32;
961 NIR_PASS_V(nir, nir_split_struct_vars, nir_var_shader_temp);
962 NIR_PASS_V(nir, dxil_nir_flatten_var_arrays, nir_var_shader_temp);
963 NIR_PASS_V(nir, dxil_nir_lower_var_bit_size, nir_var_shader_temp,
964 (supported_int_sizes & 16) ? 16 : 32, (supported_int_sizes & 64) ? 64 : 32);
965 nir->info.cs.ptr_size = 64;
966
967 NIR_PASS_V(nir, clc_lower_constant_to_ssbo, out_dxil->kernel, &uav_id);
968 NIR_PASS_V(nir, clc_change_variable_mode, nir_var_shader_temp, nir_var_mem_constant);
969 NIR_PASS_V(nir, clc_change_variable_mode, nir_var_mem_global, nir_var_mem_ssbo);
970
971 bool has_printf = false;
972 NIR_PASS(has_printf, nir, clc_lower_printf_base, uav_id);
973 metadata->printf.uav_id = has_printf ? uav_id++ : -1;
974
975 NIR_PASS_V(nir, dxil_nir_lower_deref_ssbo);
976
977 NIR_PASS_V(nir, dxil_nir_split_unaligned_loads_stores, nir_var_mem_shared | nir_var_function_temp);
978
979 // Second pass over inputs to calculate image bindings
980 unsigned srv_id = 0;
981 nir_foreach_image_variable(var, nir) {
982 int i = var->data.location;
983 if (i < 0)
984 continue;
985
986 assert(glsl_type_is_image(var->type));
987
988 if (var->data.access == ACCESS_NON_WRITEABLE) {
989 metadata->args[i].image.buf_ids[0] = srv_id++;
990 } else {
991 // Write or read-write are UAVs
992 metadata->args[i].image.buf_ids[0] = uav_id++;
993 }
994
995 metadata->args[i].image.num_buf_ids = 0;
996 var->data.binding = metadata->args[i].image.buf_ids[0];
997
998 // Assign location that'll be used for uniforms for format/order
999 var->data.driver_location = metadata->kernel_inputs_buf_size;
1000 metadata->args[i].offset = metadata->kernel_inputs_buf_size;
1001 metadata->args[i].size = 8;
1002 metadata->kernel_inputs_buf_size += metadata->args[i].size;
1003 }
1004
1005 // Fill out inline sampler metadata, now that they've been deduped and dead ones removed
1006 nir_foreach_variable_with_modes(var, nir, nir_var_uniform) {
1007 if (glsl_type_is_sampler(var->type) && var->data.sampler.is_inline_sampler) {
1008 int_sampler_states[sampler_id].wrap[0] =
1009 int_sampler_states[sampler_id].wrap[1] =
1010 int_sampler_states[sampler_id].wrap[2] =
1011 wrap_from_cl_addressing(var->data.sampler.addressing_mode);
1012 int_sampler_states[sampler_id].is_nonnormalized_coords =
1013 !var->data.sampler.normalized_coordinates;
1014 int_sampler_states[sampler_id].is_linear_filtering =
1015 var->data.sampler.filter_mode == SAMPLER_FILTER_MODE_LINEAR;
1016 var->data.binding = sampler_id++;
1017
1018 assert(metadata->num_const_samplers < CLC_MAX_SAMPLERS);
1019 metadata->const_samplers[metadata->num_const_samplers].sampler_id = var->data.binding;
1020 metadata->const_samplers[metadata->num_const_samplers].addressing_mode = var->data.sampler.addressing_mode;
1021 metadata->const_samplers[metadata->num_const_samplers].normalized_coords = var->data.sampler.normalized_coordinates;
1022 metadata->const_samplers[metadata->num_const_samplers].filter_mode = var->data.sampler.filter_mode;
1023 metadata->num_const_samplers++;
1024 }
1025 }
1026
1027 // Needs to come before lower_explicit_io
1028 NIR_PASS_V(nir, nir_lower_readonly_images_to_tex, false);
1029 struct clc_image_lower_context image_lower_context = { metadata, &srv_id, &uav_id };
1030 NIR_PASS_V(nir, clc_lower_images, &image_lower_context);
1031 NIR_PASS_V(nir, clc_lower_nonnormalized_samplers, int_sampler_states);
1032 NIR_PASS_V(nir, nir_lower_samplers);
1033 NIR_PASS_V(nir, dxil_lower_sample_to_txf_for_integer_tex,
1034 sampler_id, int_sampler_states, NULL, 14.0f);
1035
1036 assert(nir->info.cs.ptr_size == 64);
1037 NIR_PASS_V(nir, nir_lower_explicit_io, nir_var_mem_ssbo,
1038 nir_address_format_32bit_index_offset_pack64);
1039 NIR_PASS_V(nir, nir_lower_explicit_io,
1040 nir_var_mem_shared | nir_var_function_temp | nir_var_uniform,
1041 nir_address_format_32bit_offset_as_64bit);
1042
1043 NIR_PASS_V(nir, nir_lower_system_values);
1044
1045 nir_lower_compute_system_values_options compute_options = {
1046 .has_base_global_invocation_id = (conf && conf->support_global_work_id_offsets),
1047 .has_base_workgroup_id = (conf && conf->support_workgroup_id_offsets),
1048 };
1049 NIR_PASS_V(nir, nir_lower_compute_system_values, &compute_options);
1050
1051 NIR_PASS_V(nir, clc_lower_64bit_semantics);
1052
1053 NIR_PASS_V(nir, nir_opt_deref);
1054 NIR_PASS_V(nir, nir_lower_vars_to_ssa);
1055
1056 unsigned cbv_id = 0;
1057
1058 nir_variable *inputs_var =
1059 add_kernel_inputs_var(out_dxil, nir, &cbv_id);
1060 nir_variable *work_properties_var =
1061 add_work_properties_var(out_dxil, nir, &cbv_id);
1062
1063 memcpy(metadata->local_size, nir->info.workgroup_size,
1064 sizeof(metadata->local_size));
1065 memcpy(metadata->local_size_hint, nir->info.cs.workgroup_size_hint,
1066 sizeof(metadata->local_size));
1067
1068 // Patch the localsize before calling clc_nir_lower_system_values().
1069 if (conf) {
1070 for (unsigned i = 0; i < ARRAY_SIZE(nir->info.workgroup_size); i++) {
1071 if (!conf->local_size[i] ||
1072 conf->local_size[i] == nir->info.workgroup_size[i])
1073 continue;
1074
1075 if (nir->info.workgroup_size[i] &&
1076 nir->info.workgroup_size[i] != conf->local_size[i]) {
1077 debug_printf("D3D12: runtime local size does not match reqd_work_group_size() values\n");
1078 goto err_free_dxil;
1079 }
1080
1081 nir->info.workgroup_size[i] = conf->local_size[i];
1082 }
1083 memcpy(metadata->local_size, nir->info.workgroup_size,
1084 sizeof(metadata->local_size));
1085 } else {
1086 /* Make sure there's at least one thread that's set to run */
1087 for (unsigned i = 0; i < ARRAY_SIZE(nir->info.workgroup_size); i++) {
1088 if (nir->info.workgroup_size[i] == 0)
1089 nir->info.workgroup_size[i] = 1;
1090 }
1091 }
1092
1093 NIR_PASS_V(nir, clc_nir_lower_kernel_input_loads, inputs_var);
1094 NIR_PASS_V(nir, nir_lower_explicit_io, nir_var_mem_ubo,
1095 nir_address_format_32bit_index_offset);
1096 NIR_PASS_V(nir, clc_nir_lower_system_values, work_properties_var);
1097 const struct dxil_nir_lower_loads_stores_options loads_stores_options = {
1098 .use_16bit_ssbo = false,
1099 };
1100
1101 /* Now that function-declared local vars have been sized, append args */
1102 for (unsigned i = 0; i < out_dxil->kernel->num_args; i++) {
1103 if (out_dxil->kernel->args[i].address_qualifier != CLC_KERNEL_ARG_ADDRESS_LOCAL)
1104 continue;
1105
1106 /* If we don't have the runtime conf yet, we just create a dummy variable.
1107 * This will be adjusted when clc_spirv_to_dxil() is called with a conf
1108 * argument.
1109 */
1110 unsigned size = 4;
1111 if (conf && conf->args)
1112 size = conf->args[i].localptr.size;
1113
1114 /* The alignment required for the pointee type is not easy to get from
1115 * here, so let's base our logic on the size itself. Anything bigger than
1116 * the maximum alignment constraint (which is 128 bytes, since ulong16 or
1117 * doubl16 size are the biggest base types) should be aligned on this
1118 * maximum alignment constraint. For smaller types, we use the size
1119 * itself to calculate the alignment.
1120 */
1121 unsigned alignment = size < 128 ? (1 << (ffs(size) - 1)) : 128;
1122
1123 nir->info.shared_size = align(nir->info.shared_size, alignment);
1124 metadata->args[i].localptr.sharedmem_offset = nir->info.shared_size;
1125 nir->info.shared_size += size;
1126 }
1127
1128 NIR_PASS_V(nir, dxil_nir_lower_loads_stores_to_dxil, &loads_stores_options);
1129 NIR_PASS_V(nir, dxil_nir_opt_alu_deref_srcs);
1130 NIR_PASS_V(nir, nir_lower_fp16_casts, nir_lower_fp16_all);
1131 NIR_PASS_V(nir, nir_lower_convert_alu_types, NULL);
1132
1133 // Convert pack to pack_split
1134 NIR_PASS_V(nir, nir_lower_pack);
1135 // Lower pack_split to bit math
1136 NIR_PASS_V(nir, nir_opt_algebraic);
1137
1138 NIR_PASS_V(nir, nir_opt_dce);
1139
1140 nir_validate_shader(nir, "Validate before feeding NIR to the DXIL compiler");
1141 struct nir_to_dxil_options opts = {
1142 .interpolate_at_vertex = false,
1143 .lower_int16 = (conf && (conf->lower_bit_size & 16) != 0),
1144 .disable_math_refactoring = true,
1145 .num_kernel_globals = num_global_inputs,
1146 .environment = DXIL_ENVIRONMENT_CL,
1147 .shader_model_max = conf && conf->max_shader_model ? conf->max_shader_model : SHADER_MODEL_6_2,
1148 .validator_version_max = conf ? conf->validator_version : DXIL_VALIDATOR_1_4,
1149 };
1150
1151 metadata->local_mem_size = nir->info.shared_size;
1152 metadata->priv_mem_size = nir->scratch_size;
1153
1154 /* DXIL double math is too limited compared to what NIR expects. Let's refuse
1155 * to compile a shader when it contains double operations until we have
1156 * double lowering hooked up.
1157 */
1158 if (shader_has_double(nir)) {
1159 clc_error(logger, "NIR shader contains doubles, which we don't support yet");
1160 goto err_free_dxil;
1161 }
1162
1163 struct dxil_logger dxil_logger = { .priv = logger ? logger->priv : NULL,
1164 .log = logger ? logger->error : NULL};
1165
1166 struct blob tmp;
1167 if (!nir_to_dxil(nir, &opts, logger ? &dxil_logger : NULL, &tmp)) {
1168 debug_printf("D3D12: nir_to_dxil failed\n");
1169 goto err_free_dxil;
1170 }
1171
1172 nir_foreach_variable_with_modes(var, nir, nir_var_mem_ssbo) {
1173 if (var->constant_initializer) {
1174 if (glsl_type_is_array(var->type)) {
1175 int size = align(glsl_get_cl_size(var->type), 4);
1176 uint8_t *data = malloc(size);
1177 if (!data)
1178 goto err_free_dxil;
1179
1180 copy_const_initializer(var->constant_initializer, var->type, data);
1181 metadata->consts[metadata->num_consts].data = data;
1182 metadata->consts[metadata->num_consts].size = size;
1183 metadata->consts[metadata->num_consts].uav_id = var->data.binding;
1184 metadata->num_consts++;
1185 } else
1186 unreachable("unexpected constant initializer");
1187 }
1188 }
1189
1190 metadata->kernel_inputs_cbv_id = inputs_var ? inputs_var->data.binding : 0;
1191 metadata->work_properties_cbv_id = work_properties_var->data.binding;
1192 metadata->num_uavs = uav_id;
1193 metadata->num_srvs = srv_id;
1194 metadata->num_samplers = sampler_id;
1195
1196 ralloc_free(nir);
1197 glsl_type_singleton_decref();
1198
1199 blob_finish_get_buffer(&tmp, &out_dxil->binary.data,
1200 &out_dxil->binary.size);
1201 return true;
1202
1203 err_free_dxil:
1204 clc_free_dxil_object(out_dxil);
1205 return false;
1206 }
1207
clc_free_dxil_object(struct clc_dxil_object * dxil)1208 void clc_free_dxil_object(struct clc_dxil_object *dxil)
1209 {
1210 for (unsigned i = 0; i < dxil->metadata.num_consts; i++)
1211 free(dxil->metadata.consts[i].data);
1212
1213 for (unsigned i = 0; i < dxil->metadata.printf.info_count; i++) {
1214 free(dxil->metadata.printf.infos[i].arg_sizes);
1215 free(dxil->metadata.printf.infos[i].str);
1216 }
1217 free(dxil->metadata.printf.infos);
1218
1219 free(dxil->binary.data);
1220 }
1221
clc_compiler_get_version(void)1222 uint64_t clc_compiler_get_version(void)
1223 {
1224 const char sha1[] = MESA_GIT_SHA1;
1225 const char* dash = strchr(sha1, '-');
1226 if (dash) {
1227 return strtoull(dash + 1, NULL, 16);
1228 }
1229 return 0;
1230 }
1231