1 /*
2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
4 *
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
7 *
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
17 * Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 * IN THE SOFTWARE.
26 */
27
28 #include "radv_private.h"
29 #include "radv_shader.h"
30 #include "radv_shader_helper.h"
31 #include "radv_shader_args.h"
32 #include "radv_debug.h"
33 #include "nir/nir.h"
34
35 #include "sid.h"
36 #include "ac_binary.h"
37 #include "ac_llvm_util.h"
38 #include "ac_llvm_build.h"
39 #include "ac_shader_abi.h"
40 #include "ac_shader_util.h"
41 #include "ac_exp_param.h"
42
43 #define RADEON_LLVM_MAX_INPUTS (VARYING_SLOT_VAR31 + 1)
44
45 struct radv_shader_context {
46 struct ac_llvm_context ac;
47 const struct nir_shader *shader;
48 struct ac_shader_abi abi;
49 const struct radv_shader_args *args;
50
51 gl_shader_stage stage;
52
53 unsigned max_workgroup_size;
54 LLVMContextRef context;
55 LLVMValueRef main_function;
56
57 LLVMValueRef descriptor_sets[MAX_SETS];
58
59 LLVMValueRef ring_offsets;
60
61 LLVMValueRef rel_auto_id;
62
63 LLVMValueRef gs_wave_id;
64 LLVMValueRef gs_vtx_offset[6];
65
66 LLVMValueRef esgs_ring;
67 LLVMValueRef gsvs_ring[4];
68 LLVMValueRef hs_ring_tess_offchip;
69 LLVMValueRef hs_ring_tess_factor;
70
71 LLVMValueRef inputs[RADEON_LLVM_MAX_INPUTS * 4];
72
73 uint64_t output_mask;
74
75 LLVMValueRef gs_next_vertex[4];
76 LLVMValueRef gs_curprim_verts[4];
77 LLVMValueRef gs_generated_prims[4];
78 LLVMValueRef gs_ngg_emit;
79 LLVMValueRef gs_ngg_scratch;
80
81 uint32_t tcs_num_inputs;
82 uint32_t tcs_num_patches;
83 uint32_t tcs_tess_lvl_inner;
84 uint32_t tcs_tess_lvl_outer;
85
86 LLVMValueRef vertexptr; /* GFX10 only */
87 };
88
89 struct radv_shader_output_values {
90 LLVMValueRef values[4];
91 unsigned slot_name;
92 unsigned slot_index;
93 unsigned usage_mask;
94 };
95
96 static inline struct radv_shader_context *
radv_shader_context_from_abi(struct ac_shader_abi * abi)97 radv_shader_context_from_abi(struct ac_shader_abi *abi)
98 {
99 struct radv_shader_context *ctx = NULL;
100 return container_of(abi, ctx, abi);
101 }
102
get_rel_patch_id(struct radv_shader_context * ctx)103 static LLVMValueRef get_rel_patch_id(struct radv_shader_context *ctx)
104 {
105 switch (ctx->stage) {
106 case MESA_SHADER_TESS_CTRL:
107 return ac_unpack_param(&ctx->ac,
108 ac_get_arg(&ctx->ac, ctx->args->ac.tcs_rel_ids),
109 0, 8);
110 case MESA_SHADER_TESS_EVAL:
111 return ac_get_arg(&ctx->ac, ctx->args->tes_rel_patch_id);
112 break;
113 default:
114 unreachable("Illegal stage");
115 }
116 }
117
118 /* Tessellation shaders pass outputs to the next shader using LDS.
119 *
120 * LS outputs = TCS inputs
121 * TCS outputs = TES inputs
122 *
123 * The LDS layout is:
124 * - TCS inputs for patch 0
125 * - TCS inputs for patch 1
126 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
127 * - ...
128 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
129 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
130 * - TCS outputs for patch 1
131 * - Per-patch TCS outputs for patch 1
132 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
133 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
134 * - ...
135 *
136 * All three shaders VS(LS), TCS, TES share the same LDS space.
137 */
138 static LLVMValueRef
get_tcs_in_patch_stride(struct radv_shader_context * ctx)139 get_tcs_in_patch_stride(struct radv_shader_context *ctx)
140 {
141 assert(ctx->stage == MESA_SHADER_TESS_CTRL);
142 uint32_t input_vertex_size = ctx->tcs_num_inputs * 16;
143 uint32_t input_patch_size = ctx->args->options->key.tcs.input_vertices * input_vertex_size;
144
145 input_patch_size /= 4;
146 return LLVMConstInt(ctx->ac.i32, input_patch_size, false);
147 }
148
149 static LLVMValueRef
get_tcs_out_patch_stride(struct radv_shader_context * ctx)150 get_tcs_out_patch_stride(struct radv_shader_context *ctx)
151 {
152 uint32_t num_tcs_outputs = ctx->args->shader_info->tcs.num_linked_outputs;
153 uint32_t num_tcs_patch_outputs = ctx->args->shader_info->tcs.num_linked_patch_outputs;
154 uint32_t output_vertex_size = num_tcs_outputs * 16;
155 uint32_t pervertex_output_patch_size = ctx->shader->info.tess.tcs_vertices_out * output_vertex_size;
156 uint32_t output_patch_size = pervertex_output_patch_size + num_tcs_patch_outputs * 16;
157 output_patch_size /= 4;
158 return LLVMConstInt(ctx->ac.i32, output_patch_size, false);
159 }
160
161 static LLVMValueRef
get_tcs_out_vertex_stride(struct radv_shader_context * ctx)162 get_tcs_out_vertex_stride(struct radv_shader_context *ctx)
163 {
164 uint32_t num_tcs_outputs = ctx->args->shader_info->tcs.num_linked_outputs;
165 uint32_t output_vertex_size = num_tcs_outputs * 16;
166 output_vertex_size /= 4;
167 return LLVMConstInt(ctx->ac.i32, output_vertex_size, false);
168 }
169
170 static LLVMValueRef
get_tcs_out_patch0_offset(struct radv_shader_context * ctx)171 get_tcs_out_patch0_offset(struct radv_shader_context *ctx)
172 {
173 assert (ctx->stage == MESA_SHADER_TESS_CTRL);
174 uint32_t input_vertex_size = ctx->tcs_num_inputs * 16;
175 uint32_t input_patch_size = ctx->args->options->key.tcs.input_vertices * input_vertex_size;
176 uint32_t output_patch0_offset = input_patch_size;
177 unsigned num_patches = ctx->tcs_num_patches;
178
179 output_patch0_offset *= num_patches;
180 output_patch0_offset /= 4;
181 return LLVMConstInt(ctx->ac.i32, output_patch0_offset, false);
182 }
183
184 static LLVMValueRef
get_tcs_out_patch0_patch_data_offset(struct radv_shader_context * ctx)185 get_tcs_out_patch0_patch_data_offset(struct radv_shader_context *ctx)
186 {
187 assert (ctx->stage == MESA_SHADER_TESS_CTRL);
188 uint32_t input_vertex_size = ctx->tcs_num_inputs * 16;
189 uint32_t input_patch_size = ctx->args->options->key.tcs.input_vertices * input_vertex_size;
190 uint32_t output_patch0_offset = input_patch_size;
191
192 uint32_t num_tcs_outputs = ctx->args->shader_info->tcs.num_linked_outputs;
193 uint32_t output_vertex_size = num_tcs_outputs * 16;
194 uint32_t pervertex_output_patch_size = ctx->shader->info.tess.tcs_vertices_out * output_vertex_size;
195 unsigned num_patches = ctx->tcs_num_patches;
196
197 output_patch0_offset *= num_patches;
198 output_patch0_offset += pervertex_output_patch_size;
199 output_patch0_offset /= 4;
200 return LLVMConstInt(ctx->ac.i32, output_patch0_offset, false);
201 }
202
203 static LLVMValueRef
get_tcs_in_current_patch_offset(struct radv_shader_context * ctx)204 get_tcs_in_current_patch_offset(struct radv_shader_context *ctx)
205 {
206 LLVMValueRef patch_stride = get_tcs_in_patch_stride(ctx);
207 LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
208
209 return LLVMBuildMul(ctx->ac.builder, patch_stride, rel_patch_id, "");
210 }
211
212 static LLVMValueRef
get_tcs_out_current_patch_offset(struct radv_shader_context * ctx)213 get_tcs_out_current_patch_offset(struct radv_shader_context *ctx)
214 {
215 LLVMValueRef patch0_offset = get_tcs_out_patch0_offset(ctx);
216 LLVMValueRef patch_stride = get_tcs_out_patch_stride(ctx);
217 LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
218
219 return ac_build_imad(&ctx->ac, patch_stride, rel_patch_id,
220 patch0_offset);
221 }
222
223 static LLVMValueRef
get_tcs_out_current_patch_data_offset(struct radv_shader_context * ctx)224 get_tcs_out_current_patch_data_offset(struct radv_shader_context *ctx)
225 {
226 LLVMValueRef patch0_patch_data_offset =
227 get_tcs_out_patch0_patch_data_offset(ctx);
228 LLVMValueRef patch_stride = get_tcs_out_patch_stride(ctx);
229 LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
230
231 return ac_build_imad(&ctx->ac, patch_stride, rel_patch_id,
232 patch0_patch_data_offset);
233 }
234
235 static LLVMValueRef
create_llvm_function(struct ac_llvm_context * ctx,LLVMModuleRef module,LLVMBuilderRef builder,const struct ac_shader_args * args,enum ac_llvm_calling_convention convention,unsigned max_workgroup_size,const struct radv_nir_compiler_options * options)236 create_llvm_function(struct ac_llvm_context *ctx, LLVMModuleRef module,
237 LLVMBuilderRef builder,
238 const struct ac_shader_args *args,
239 enum ac_llvm_calling_convention convention,
240 unsigned max_workgroup_size,
241 const struct radv_nir_compiler_options *options)
242 {
243 LLVMValueRef main_function =
244 ac_build_main(args, ctx, convention, "main", ctx->voidt, module);
245
246 if (options->address32_hi) {
247 ac_llvm_add_target_dep_function_attr(main_function,
248 "amdgpu-32bit-address-high-bits",
249 options->address32_hi);
250 }
251
252 ac_llvm_set_workgroup_size(main_function, max_workgroup_size);
253
254 return main_function;
255 }
256
257 static void
load_descriptor_sets(struct radv_shader_context * ctx)258 load_descriptor_sets(struct radv_shader_context *ctx)
259 {
260 uint32_t mask = ctx->args->shader_info->desc_set_used_mask;
261 if (ctx->args->shader_info->need_indirect_descriptor_sets) {
262 LLVMValueRef desc_sets =
263 ac_get_arg(&ctx->ac, ctx->args->descriptor_sets[0]);
264 while (mask) {
265 int i = u_bit_scan(&mask);
266
267 ctx->descriptor_sets[i] =
268 ac_build_load_to_sgpr(&ctx->ac, desc_sets,
269 LLVMConstInt(ctx->ac.i32, i, false));
270
271 }
272 } else {
273 while (mask) {
274 int i = u_bit_scan(&mask);
275
276 ctx->descriptor_sets[i] =
277 ac_get_arg(&ctx->ac, ctx->args->descriptor_sets[i]);
278 }
279 }
280 }
281
282 static enum ac_llvm_calling_convention
get_llvm_calling_convention(LLVMValueRef func,gl_shader_stage stage)283 get_llvm_calling_convention(LLVMValueRef func, gl_shader_stage stage)
284 {
285 switch (stage) {
286 case MESA_SHADER_VERTEX:
287 case MESA_SHADER_TESS_EVAL:
288 return AC_LLVM_AMDGPU_VS;
289 break;
290 case MESA_SHADER_GEOMETRY:
291 return AC_LLVM_AMDGPU_GS;
292 break;
293 case MESA_SHADER_TESS_CTRL:
294 return AC_LLVM_AMDGPU_HS;
295 break;
296 case MESA_SHADER_FRAGMENT:
297 return AC_LLVM_AMDGPU_PS;
298 break;
299 case MESA_SHADER_COMPUTE:
300 return AC_LLVM_AMDGPU_CS;
301 break;
302 default:
303 unreachable("Unhandle shader type");
304 }
305 }
306
307 /* Returns whether the stage is a stage that can be directly before the GS */
is_pre_gs_stage(gl_shader_stage stage)308 static bool is_pre_gs_stage(gl_shader_stage stage)
309 {
310 return stage == MESA_SHADER_VERTEX || stage == MESA_SHADER_TESS_EVAL;
311 }
312
create_function(struct radv_shader_context * ctx,gl_shader_stage stage,bool has_previous_stage)313 static void create_function(struct radv_shader_context *ctx,
314 gl_shader_stage stage,
315 bool has_previous_stage)
316 {
317 if (ctx->ac.chip_class >= GFX10) {
318 if (is_pre_gs_stage(stage) && ctx->args->options->key.vs_common_out.as_ngg) {
319 /* On GFX10, VS is merged into GS for NGG. */
320 stage = MESA_SHADER_GEOMETRY;
321 has_previous_stage = true;
322 }
323 }
324
325 ctx->main_function = create_llvm_function(
326 &ctx->ac, ctx->ac.module, ctx->ac.builder, &ctx->args->ac,
327 get_llvm_calling_convention(ctx->main_function, stage),
328 ctx->max_workgroup_size,
329 ctx->args->options);
330
331 ctx->ring_offsets = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.implicit.buffer.ptr",
332 LLVMPointerType(ctx->ac.i8, AC_ADDR_SPACE_CONST),
333 NULL, 0, AC_FUNC_ATTR_READNONE);
334 ctx->ring_offsets = LLVMBuildBitCast(ctx->ac.builder, ctx->ring_offsets,
335 ac_array_in_const_addr_space(ctx->ac.v4i32), "");
336
337 load_descriptor_sets(ctx);
338
339 if (stage == MESA_SHADER_TESS_CTRL ||
340 (stage == MESA_SHADER_VERTEX && ctx->args->options->key.vs_common_out.as_ls) ||
341 /* GFX9 has the ESGS ring buffer in LDS. */
342 (stage == MESA_SHADER_GEOMETRY && has_previous_stage)) {
343 ac_declare_lds_as_pointer(&ctx->ac);
344 }
345
346 }
347
348
349 static LLVMValueRef
radv_load_resource(struct ac_shader_abi * abi,LLVMValueRef index,unsigned desc_set,unsigned binding)350 radv_load_resource(struct ac_shader_abi *abi, LLVMValueRef index,
351 unsigned desc_set, unsigned binding)
352 {
353 struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
354 LLVMValueRef desc_ptr = ctx->descriptor_sets[desc_set];
355 struct radv_pipeline_layout *pipeline_layout = ctx->args->options->layout;
356 struct radv_descriptor_set_layout *layout = pipeline_layout->set[desc_set].layout;
357 unsigned base_offset = layout->binding[binding].offset;
358 LLVMValueRef offset, stride;
359
360 if (layout->binding[binding].type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC ||
361 layout->binding[binding].type == VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC) {
362 unsigned idx = pipeline_layout->set[desc_set].dynamic_offset_start +
363 layout->binding[binding].dynamic_offset_offset;
364 desc_ptr = ac_get_arg(&ctx->ac, ctx->args->ac.push_constants);
365 base_offset = pipeline_layout->push_constant_size + 16 * idx;
366 stride = LLVMConstInt(ctx->ac.i32, 16, false);
367 } else
368 stride = LLVMConstInt(ctx->ac.i32, layout->binding[binding].size, false);
369
370 offset = LLVMConstInt(ctx->ac.i32, base_offset, false);
371
372 if (layout->binding[binding].type != VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) {
373 offset = ac_build_imad(&ctx->ac, index, stride, offset);
374 }
375
376 desc_ptr = LLVMBuildGEP(ctx->ac.builder, desc_ptr, &offset, 1, "");
377 desc_ptr = ac_cast_ptr(&ctx->ac, desc_ptr, ctx->ac.v4i32);
378 LLVMSetMetadata(desc_ptr, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
379
380 return desc_ptr;
381 }
382
383
384 /* The offchip buffer layout for TCS->TES is
385 *
386 * - attribute 0 of patch 0 vertex 0
387 * - attribute 0 of patch 0 vertex 1
388 * - attribute 0 of patch 0 vertex 2
389 * ...
390 * - attribute 0 of patch 1 vertex 0
391 * - attribute 0 of patch 1 vertex 1
392 * ...
393 * - attribute 1 of patch 0 vertex 0
394 * - attribute 1 of patch 0 vertex 1
395 * ...
396 * - per patch attribute 0 of patch 0
397 * - per patch attribute 0 of patch 1
398 * ...
399 *
400 * Note that every attribute has 4 components.
401 */
get_non_vertex_index_offset(struct radv_shader_context * ctx)402 static LLVMValueRef get_non_vertex_index_offset(struct radv_shader_context *ctx)
403 {
404 uint32_t num_patches = ctx->tcs_num_patches;
405 uint32_t num_tcs_outputs;
406 if (ctx->stage == MESA_SHADER_TESS_CTRL)
407 num_tcs_outputs = ctx->args->shader_info->tcs.num_linked_outputs;
408 else
409 num_tcs_outputs = ctx->args->shader_info->tes.num_linked_inputs;
410
411 uint32_t output_vertex_size = num_tcs_outputs * 16;
412 uint32_t pervertex_output_patch_size = ctx->shader->info.tess.tcs_vertices_out * output_vertex_size;
413
414 return LLVMConstInt(ctx->ac.i32, pervertex_output_patch_size * num_patches, false);
415 }
416
calc_param_stride(struct radv_shader_context * ctx,LLVMValueRef vertex_index)417 static LLVMValueRef calc_param_stride(struct radv_shader_context *ctx,
418 LLVMValueRef vertex_index)
419 {
420 LLVMValueRef param_stride;
421 if (vertex_index)
422 param_stride = LLVMConstInt(ctx->ac.i32, ctx->shader->info.tess.tcs_vertices_out * ctx->tcs_num_patches, false);
423 else
424 param_stride = LLVMConstInt(ctx->ac.i32, ctx->tcs_num_patches, false);
425 return param_stride;
426 }
427
get_tcs_tes_buffer_address(struct radv_shader_context * ctx,LLVMValueRef vertex_index,LLVMValueRef param_index)428 static LLVMValueRef get_tcs_tes_buffer_address(struct radv_shader_context *ctx,
429 LLVMValueRef vertex_index,
430 LLVMValueRef param_index)
431 {
432 LLVMValueRef base_addr;
433 LLVMValueRef param_stride, constant16;
434 LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
435 LLVMValueRef vertices_per_patch = LLVMConstInt(ctx->ac.i32, ctx->shader->info.tess.tcs_vertices_out, false);
436 constant16 = LLVMConstInt(ctx->ac.i32, 16, false);
437 param_stride = calc_param_stride(ctx, vertex_index);
438 if (vertex_index) {
439 base_addr = ac_build_imad(&ctx->ac, rel_patch_id,
440 vertices_per_patch, vertex_index);
441 } else {
442 base_addr = rel_patch_id;
443 }
444
445 base_addr = LLVMBuildAdd(ctx->ac.builder, base_addr,
446 LLVMBuildMul(ctx->ac.builder, param_index,
447 param_stride, ""), "");
448
449 base_addr = LLVMBuildMul(ctx->ac.builder, base_addr, constant16, "");
450
451 if (!vertex_index) {
452 LLVMValueRef patch_data_offset = get_non_vertex_index_offset(ctx);
453
454 base_addr = LLVMBuildAdd(ctx->ac.builder, base_addr,
455 patch_data_offset, "");
456 }
457 return base_addr;
458 }
459
get_tcs_tes_buffer_address_params(struct radv_shader_context * ctx,unsigned param,LLVMValueRef vertex_index,LLVMValueRef indir_index)460 static LLVMValueRef get_tcs_tes_buffer_address_params(struct radv_shader_context *ctx,
461 unsigned param,
462 LLVMValueRef vertex_index,
463 LLVMValueRef indir_index)
464 {
465 LLVMValueRef param_index;
466
467 if (indir_index)
468 param_index = LLVMBuildAdd(ctx->ac.builder, LLVMConstInt(ctx->ac.i32, param, false),
469 indir_index, "");
470 else {
471 param_index = LLVMConstInt(ctx->ac.i32, param, false);
472 }
473 return get_tcs_tes_buffer_address(ctx, vertex_index, param_index);
474 }
475
476 static LLVMValueRef
get_dw_address(struct radv_shader_context * ctx,LLVMValueRef dw_addr,unsigned param,LLVMValueRef vertex_index,LLVMValueRef stride,LLVMValueRef indir_index)477 get_dw_address(struct radv_shader_context *ctx,
478 LLVMValueRef dw_addr,
479 unsigned param,
480 LLVMValueRef vertex_index,
481 LLVMValueRef stride,
482 LLVMValueRef indir_index)
483
484 {
485
486 if (vertex_index) {
487 dw_addr = LLVMBuildAdd(ctx->ac.builder, dw_addr,
488 LLVMBuildMul(ctx->ac.builder,
489 vertex_index,
490 stride, ""), "");
491 }
492
493 if (indir_index)
494 dw_addr = LLVMBuildAdd(ctx->ac.builder, dw_addr,
495 LLVMBuildMul(ctx->ac.builder, indir_index,
496 LLVMConstInt(ctx->ac.i32, 4, false), ""), "");
497
498 dw_addr = LLVMBuildAdd(ctx->ac.builder, dw_addr,
499 LLVMConstInt(ctx->ac.i32, param * 4, false), "");
500
501 return dw_addr;
502 }
503
504 static LLVMValueRef
load_tcs_varyings(struct ac_shader_abi * abi,LLVMTypeRef type,LLVMValueRef vertex_index,LLVMValueRef indir_index,unsigned driver_location,unsigned component,unsigned num_components,bool load_input)505 load_tcs_varyings(struct ac_shader_abi *abi,
506 LLVMTypeRef type,
507 LLVMValueRef vertex_index,
508 LLVMValueRef indir_index,
509 unsigned driver_location,
510 unsigned component,
511 unsigned num_components,
512 bool load_input)
513 {
514 struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
515 LLVMValueRef dw_addr, stride;
516 LLVMValueRef value[4], result;
517 unsigned param = driver_location;
518
519 bool is_patch = vertex_index == NULL;
520
521 if (load_input) {
522 uint32_t input_vertex_size = (ctx->tcs_num_inputs * 16) / 4;
523 stride = LLVMConstInt(ctx->ac.i32, input_vertex_size, false);
524 dw_addr = get_tcs_in_current_patch_offset(ctx);
525 } else {
526 if (!is_patch) {
527 stride = get_tcs_out_vertex_stride(ctx);
528 dw_addr = get_tcs_out_current_patch_offset(ctx);
529 } else {
530 dw_addr = get_tcs_out_current_patch_data_offset(ctx);
531 stride = NULL;
532 }
533 }
534
535 dw_addr = get_dw_address(ctx, dw_addr, param, vertex_index, stride, indir_index);
536
537 for (unsigned i = 0; i < num_components + component; i++) {
538 value[i] = ac_lds_load(&ctx->ac, dw_addr);
539 dw_addr = LLVMBuildAdd(ctx->ac.builder, dw_addr,
540 ctx->ac.i32_1, "");
541 }
542 result = ac_build_varying_gather_values(&ctx->ac, value, num_components, component);
543 return result;
544 }
545
546 static void
store_tcs_output(struct ac_shader_abi * abi,LLVMValueRef vertex_index,LLVMValueRef param_index,LLVMValueRef src,unsigned writemask,unsigned component,unsigned location,unsigned driver_location)547 store_tcs_output(struct ac_shader_abi *abi,
548 LLVMValueRef vertex_index,
549 LLVMValueRef param_index,
550 LLVMValueRef src,
551 unsigned writemask,
552 unsigned component,
553 unsigned location,
554 unsigned driver_location)
555 {
556 struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
557 const bool is_patch = vertex_index == NULL;
558 LLVMValueRef dw_addr;
559 LLVMValueRef stride = NULL;
560 LLVMValueRef buf_addr = NULL;
561 LLVMValueRef oc_lds = ac_get_arg(&ctx->ac, ctx->args->oc_lds);
562 unsigned param = driver_location;
563 bool store_lds = true;
564
565 if (is_patch) {
566 if (!(ctx->shader->info.patch_outputs_read & (1U << (location - VARYING_SLOT_PATCH0))))
567 store_lds = false;
568 } else {
569 if (!(ctx->shader->info.outputs_read & (1ULL << location)))
570 store_lds = false;
571 }
572
573 if (!is_patch) {
574 stride = get_tcs_out_vertex_stride(ctx);
575 dw_addr = get_tcs_out_current_patch_offset(ctx);
576 } else {
577 dw_addr = get_tcs_out_current_patch_data_offset(ctx);
578 }
579
580 dw_addr = get_dw_address(ctx, dw_addr, param, vertex_index, stride, param_index);
581 buf_addr = get_tcs_tes_buffer_address_params(ctx, param, vertex_index, param_index);
582
583 bool is_tess_factor = false;
584 if (location == VARYING_SLOT_TESS_LEVEL_INNER ||
585 location == VARYING_SLOT_TESS_LEVEL_OUTER)
586 is_tess_factor = true;
587
588 for (unsigned chan = 0; chan < 8; chan++) {
589 if (!(writemask & (1 << chan)))
590 continue;
591 LLVMValueRef value = ac_llvm_extract_elem(&ctx->ac, src, chan - component);
592 value = ac_to_integer(&ctx->ac, value);
593 value = LLVMBuildZExtOrBitCast(ctx->ac.builder, value, ctx->ac.i32, "");
594
595 if (store_lds || is_tess_factor) {
596 LLVMValueRef dw_addr_chan =
597 LLVMBuildAdd(ctx->ac.builder, dw_addr,
598 LLVMConstInt(ctx->ac.i32, chan, false), "");
599 ac_lds_store(&ctx->ac, dw_addr_chan, value);
600 }
601
602 if (!is_tess_factor && writemask != 0xF)
603 ac_build_buffer_store_dword(&ctx->ac, ctx->hs_ring_tess_offchip, value, 1,
604 buf_addr, oc_lds,
605 4 * chan, ac_glc);
606 }
607
608 if (writemask == 0xF) {
609 ac_build_buffer_store_dword(&ctx->ac, ctx->hs_ring_tess_offchip, src, 4,
610 buf_addr, oc_lds, 0, ac_glc);
611 }
612 }
613
614 static LLVMValueRef
load_tes_input(struct ac_shader_abi * abi,LLVMTypeRef type,LLVMValueRef vertex_index,LLVMValueRef param_index,unsigned driver_location,unsigned component,unsigned num_components,bool load_input)615 load_tes_input(struct ac_shader_abi *abi,
616 LLVMTypeRef type,
617 LLVMValueRef vertex_index,
618 LLVMValueRef param_index,
619 unsigned driver_location,
620 unsigned component,
621 unsigned num_components,
622 bool load_input)
623 {
624 struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
625 LLVMValueRef buf_addr;
626 LLVMValueRef result;
627 LLVMValueRef oc_lds = ac_get_arg(&ctx->ac, ctx->args->oc_lds);
628 unsigned param = driver_location;
629
630 buf_addr = get_tcs_tes_buffer_address_params(ctx, param, vertex_index, param_index);
631
632 LLVMValueRef comp_offset = LLVMConstInt(ctx->ac.i32, component * 4, false);
633 buf_addr = LLVMBuildAdd(ctx->ac.builder, buf_addr, comp_offset, "");
634
635 result = ac_build_buffer_load(&ctx->ac, ctx->hs_ring_tess_offchip, num_components, NULL,
636 buf_addr, oc_lds, 0, ac_glc, true, false);
637 result = ac_trim_vector(&ctx->ac, result, num_components);
638 return result;
639 }
640
641 static LLVMValueRef
load_gs_input(struct ac_shader_abi * abi,unsigned driver_location,unsigned component,unsigned num_components,unsigned vertex_index,LLVMTypeRef type)642 load_gs_input(struct ac_shader_abi *abi,
643 unsigned driver_location,
644 unsigned component,
645 unsigned num_components,
646 unsigned vertex_index,
647 LLVMTypeRef type)
648 {
649 struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
650 LLVMValueRef vtx_offset;
651 unsigned param = driver_location;
652 unsigned vtx_offset_param;
653 LLVMValueRef value[4], result;
654
655 vtx_offset_param = vertex_index;
656 assert(vtx_offset_param < 6);
657 vtx_offset = LLVMBuildMul(ctx->ac.builder, ctx->gs_vtx_offset[vtx_offset_param],
658 LLVMConstInt(ctx->ac.i32, 4, false), "");
659
660 for (unsigned i = component; i < num_components + component; i++) {
661 if (ctx->ac.chip_class >= GFX9) {
662 LLVMValueRef dw_addr = ctx->gs_vtx_offset[vtx_offset_param];
663 dw_addr = LLVMBuildAdd(ctx->ac.builder, dw_addr,
664 LLVMConstInt(ctx->ac.i32, param * 4 + i, 0), "");
665 value[i] = ac_lds_load(&ctx->ac, dw_addr);
666 } else {
667 LLVMValueRef soffset =
668 LLVMConstInt(ctx->ac.i32,
669 (param * 4 + i) * 256,
670 false);
671
672 value[i] = ac_build_buffer_load(&ctx->ac,
673 ctx->esgs_ring, 1,
674 ctx->ac.i32_0,
675 vtx_offset, soffset,
676 0, ac_glc, true, false);
677 }
678
679 if (ac_get_type_size(type) == 2) {
680 value[i] = LLVMBuildBitCast(ctx->ac.builder, value[i], ctx->ac.i32, "");
681 value[i] = LLVMBuildTrunc(ctx->ac.builder, value[i], ctx->ac.i16, "");
682 }
683 value[i] = LLVMBuildBitCast(ctx->ac.builder, value[i], type, "");
684 }
685 result = ac_build_varying_gather_values(&ctx->ac, value, num_components, component);
686 result = ac_to_integer(&ctx->ac, result);
687 return result;
688 }
689
690 static uint32_t
radv_get_sample_pos_offset(uint32_t num_samples)691 radv_get_sample_pos_offset(uint32_t num_samples)
692 {
693 uint32_t sample_pos_offset = 0;
694
695 switch (num_samples) {
696 case 2:
697 sample_pos_offset = 1;
698 break;
699 case 4:
700 sample_pos_offset = 3;
701 break;
702 case 8:
703 sample_pos_offset = 7;
704 break;
705 default:
706 break;
707 }
708 return sample_pos_offset;
709 }
710
load_sample_position(struct ac_shader_abi * abi,LLVMValueRef sample_id)711 static LLVMValueRef load_sample_position(struct ac_shader_abi *abi,
712 LLVMValueRef sample_id)
713 {
714 struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
715
716 LLVMValueRef result;
717 LLVMValueRef index = LLVMConstInt(ctx->ac.i32, RING_PS_SAMPLE_POSITIONS, false);
718 LLVMValueRef ptr = LLVMBuildGEP(ctx->ac.builder, ctx->ring_offsets, &index, 1, "");
719
720 ptr = LLVMBuildBitCast(ctx->ac.builder, ptr,
721 ac_array_in_const_addr_space(ctx->ac.v2f32), "");
722
723 uint32_t sample_pos_offset =
724 radv_get_sample_pos_offset(ctx->args->options->key.fs.num_samples);
725
726 sample_id =
727 LLVMBuildAdd(ctx->ac.builder, sample_id,
728 LLVMConstInt(ctx->ac.i32, sample_pos_offset, false), "");
729 result = ac_build_load_invariant(&ctx->ac, ptr, sample_id);
730
731 return result;
732 }
733
734
load_sample_mask_in(struct ac_shader_abi * abi)735 static LLVMValueRef load_sample_mask_in(struct ac_shader_abi *abi)
736 {
737 struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
738 uint8_t log2_ps_iter_samples;
739
740 if (ctx->args->shader_info->ps.force_persample) {
741 log2_ps_iter_samples =
742 util_logbase2(ctx->args->options->key.fs.num_samples);
743 } else {
744 log2_ps_iter_samples = ctx->args->options->key.fs.log2_ps_iter_samples;
745 }
746
747 LLVMValueRef result, sample_id;
748 if (log2_ps_iter_samples) {
749 /* gl_SampleMaskIn[0] = (SampleCoverage & (1 << gl_SampleID)). */
750 sample_id = ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->ac.ancillary), 8, 4);
751 sample_id = LLVMBuildShl(ctx->ac.builder, LLVMConstInt(ctx->ac.i32, 1, false), sample_id, "");
752 result = LLVMBuildAnd(ctx->ac.builder, sample_id,
753 ac_get_arg(&ctx->ac, ctx->args->ac.sample_coverage), "");
754 } else {
755 result = ac_get_arg(&ctx->ac, ctx->args->ac.sample_coverage);
756 }
757
758 return result;
759 }
760
761
762 static void gfx10_ngg_gs_emit_vertex(struct radv_shader_context *ctx,
763 unsigned stream,
764 LLVMValueRef vertexidx,
765 LLVMValueRef *addrs);
766
767 static void
visit_emit_vertex_with_counter(struct ac_shader_abi * abi,unsigned stream,LLVMValueRef vertexidx,LLVMValueRef * addrs)768 visit_emit_vertex_with_counter(struct ac_shader_abi *abi, unsigned stream,
769 LLVMValueRef vertexidx, LLVMValueRef *addrs)
770 {
771 unsigned offset = 0;
772 struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
773
774 if (ctx->args->options->key.vs_common_out.as_ngg) {
775 gfx10_ngg_gs_emit_vertex(ctx, stream, vertexidx, addrs);
776 return;
777 }
778
779 for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
780 unsigned output_usage_mask =
781 ctx->args->shader_info->gs.output_usage_mask[i];
782 uint8_t output_stream =
783 ctx->args->shader_info->gs.output_streams[i];
784 LLVMValueRef *out_ptr = &addrs[i * 4];
785 int length = util_last_bit(output_usage_mask);
786
787 if (!(ctx->output_mask & (1ull << i)) ||
788 output_stream != stream)
789 continue;
790
791 for (unsigned j = 0; j < length; j++) {
792 if (!(output_usage_mask & (1 << j)))
793 continue;
794
795 LLVMValueRef out_val = LLVMBuildLoad(ctx->ac.builder,
796 out_ptr[j], "");
797 LLVMValueRef voffset =
798 LLVMConstInt(ctx->ac.i32, offset *
799 ctx->shader->info.gs.vertices_out, false);
800
801 offset++;
802
803 voffset = LLVMBuildAdd(ctx->ac.builder, voffset, vertexidx, "");
804 voffset = LLVMBuildMul(ctx->ac.builder, voffset, LLVMConstInt(ctx->ac.i32, 4, false), "");
805
806 out_val = ac_to_integer(&ctx->ac, out_val);
807 out_val = LLVMBuildZExtOrBitCast(ctx->ac.builder, out_val, ctx->ac.i32, "");
808
809 ac_build_buffer_store_dword(&ctx->ac,
810 ctx->gsvs_ring[stream],
811 out_val, 1,
812 voffset,
813 ac_get_arg(&ctx->ac,
814 ctx->args->gs2vs_offset),
815 0, ac_glc | ac_slc | ac_swizzled);
816 }
817 }
818
819 ac_build_sendmsg(&ctx->ac,
820 AC_SENDMSG_GS_OP_EMIT | AC_SENDMSG_GS | (stream << 8),
821 ctx->gs_wave_id);
822 }
823
824 static void
visit_end_primitive(struct ac_shader_abi * abi,unsigned stream)825 visit_end_primitive(struct ac_shader_abi *abi, unsigned stream)
826 {
827 struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
828
829 if (ctx->args->options->key.vs_common_out.as_ngg) {
830 LLVMBuildStore(ctx->ac.builder, ctx->ac.i32_0, ctx->gs_curprim_verts[stream]);
831 return;
832 }
833
834 ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_CUT | AC_SENDMSG_GS | (stream << 8), ctx->gs_wave_id);
835 }
836
837 static LLVMValueRef
load_tess_coord(struct ac_shader_abi * abi)838 load_tess_coord(struct ac_shader_abi *abi)
839 {
840 struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
841
842 LLVMValueRef coord[4] = {
843 ac_get_arg(&ctx->ac, ctx->args->tes_u),
844 ac_get_arg(&ctx->ac, ctx->args->tes_v),
845 ctx->ac.f32_0,
846 ctx->ac.f32_0,
847 };
848
849 if (ctx->shader->info.tess.primitive_mode == GL_TRIANGLES)
850 coord[2] = LLVMBuildFSub(ctx->ac.builder, ctx->ac.f32_1,
851 LLVMBuildFAdd(ctx->ac.builder, coord[0], coord[1], ""), "");
852
853 return ac_build_gather_values(&ctx->ac, coord, 3);
854 }
855
856 static LLVMValueRef
load_patch_vertices_in(struct ac_shader_abi * abi)857 load_patch_vertices_in(struct ac_shader_abi *abi)
858 {
859 struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
860 return LLVMConstInt(ctx->ac.i32, ctx->args->options->key.tcs.input_vertices, false);
861 }
862
863
radv_load_base_vertex(struct ac_shader_abi * abi)864 static LLVMValueRef radv_load_base_vertex(struct ac_shader_abi *abi)
865 {
866 struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
867 return ac_get_arg(&ctx->ac, ctx->args->ac.base_vertex);
868 }
869
radv_load_ssbo(struct ac_shader_abi * abi,LLVMValueRef buffer_ptr,bool write)870 static LLVMValueRef radv_load_ssbo(struct ac_shader_abi *abi,
871 LLVMValueRef buffer_ptr, bool write)
872 {
873 struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
874 LLVMValueRef result;
875
876 LLVMSetMetadata(buffer_ptr, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
877
878 result = LLVMBuildLoad(ctx->ac.builder, buffer_ptr, "");
879 LLVMSetMetadata(result, ctx->ac.invariant_load_md_kind, ctx->ac.empty_md);
880
881 return result;
882 }
883
radv_load_ubo(struct ac_shader_abi * abi,unsigned desc_set,unsigned binding,bool valid_binding,LLVMValueRef buffer_ptr)884 static LLVMValueRef radv_load_ubo(struct ac_shader_abi *abi,
885 unsigned desc_set, unsigned binding,
886 bool valid_binding, LLVMValueRef buffer_ptr)
887 {
888 struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
889 LLVMValueRef result;
890
891 if (valid_binding) {
892 struct radv_pipeline_layout *pipeline_layout = ctx->args->options->layout;
893 struct radv_descriptor_set_layout *layout = pipeline_layout->set[desc_set].layout;
894
895 if (layout->binding[binding].type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) {
896 uint32_t desc_type = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
897 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
898 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
899 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
900
901 if (ctx->ac.chip_class >= GFX10) {
902 desc_type |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) |
903 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW) |
904 S_008F0C_RESOURCE_LEVEL(1);
905 } else {
906 desc_type |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
907 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
908 }
909
910 LLVMValueRef desc_components[4] = {
911 LLVMBuildPtrToInt(ctx->ac.builder, buffer_ptr, ctx->ac.intptr, ""),
912 LLVMConstInt(ctx->ac.i32, S_008F04_BASE_ADDRESS_HI(ctx->args->options->address32_hi), false),
913 LLVMConstInt(ctx->ac.i32, 0xffffffff, false),
914 LLVMConstInt(ctx->ac.i32, desc_type, false),
915 };
916
917 return ac_build_gather_values(&ctx->ac, desc_components, 4);
918 }
919 }
920
921 LLVMSetMetadata(buffer_ptr, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
922
923 result = LLVMBuildLoad(ctx->ac.builder, buffer_ptr, "");
924 LLVMSetMetadata(result, ctx->ac.invariant_load_md_kind, ctx->ac.empty_md);
925
926 return result;
927 }
928
radv_get_sampler_desc(struct ac_shader_abi * abi,unsigned descriptor_set,unsigned base_index,unsigned constant_index,LLVMValueRef index,enum ac_descriptor_type desc_type,bool image,bool write,bool bindless)929 static LLVMValueRef radv_get_sampler_desc(struct ac_shader_abi *abi,
930 unsigned descriptor_set,
931 unsigned base_index,
932 unsigned constant_index,
933 LLVMValueRef index,
934 enum ac_descriptor_type desc_type,
935 bool image, bool write,
936 bool bindless)
937 {
938 struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
939 LLVMValueRef list = ctx->descriptor_sets[descriptor_set];
940 struct radv_descriptor_set_layout *layout = ctx->args->options->layout->set[descriptor_set].layout;
941 struct radv_descriptor_set_binding_layout *binding = layout->binding + base_index;
942 unsigned offset = binding->offset;
943 unsigned stride = binding->size;
944 unsigned type_size;
945 LLVMBuilderRef builder = ctx->ac.builder;
946 LLVMTypeRef type;
947
948 assert(base_index < layout->binding_count);
949
950 switch (desc_type) {
951 case AC_DESC_IMAGE:
952 type = ctx->ac.v8i32;
953 type_size = 32;
954 break;
955 case AC_DESC_FMASK:
956 type = ctx->ac.v8i32;
957 offset += 32;
958 type_size = 32;
959 break;
960 case AC_DESC_SAMPLER:
961 type = ctx->ac.v4i32;
962 if (binding->type == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER) {
963 offset += radv_combined_image_descriptor_sampler_offset(binding);
964 }
965
966 type_size = 16;
967 break;
968 case AC_DESC_BUFFER:
969 type = ctx->ac.v4i32;
970 type_size = 16;
971 break;
972 case AC_DESC_PLANE_0:
973 case AC_DESC_PLANE_1:
974 case AC_DESC_PLANE_2:
975 type = ctx->ac.v8i32;
976 type_size = 32;
977 offset += 32 * (desc_type - AC_DESC_PLANE_0);
978 break;
979 default:
980 unreachable("invalid desc_type\n");
981 }
982
983 offset += constant_index * stride;
984
985 if (desc_type == AC_DESC_SAMPLER && binding->immutable_samplers_offset &&
986 (!index || binding->immutable_samplers_equal)) {
987 if (binding->immutable_samplers_equal)
988 constant_index = 0;
989
990 const uint32_t *samplers = radv_immutable_samplers(layout, binding);
991
992 LLVMValueRef constants[] = {
993 LLVMConstInt(ctx->ac.i32, samplers[constant_index * 4 + 0], 0),
994 LLVMConstInt(ctx->ac.i32, samplers[constant_index * 4 + 1], 0),
995 LLVMConstInt(ctx->ac.i32, samplers[constant_index * 4 + 2], 0),
996 LLVMConstInt(ctx->ac.i32, samplers[constant_index * 4 + 3], 0),
997 };
998 return ac_build_gather_values(&ctx->ac, constants, 4);
999 }
1000
1001 assert(stride % type_size == 0);
1002
1003 LLVMValueRef adjusted_index = index;
1004 if (!adjusted_index)
1005 adjusted_index = ctx->ac.i32_0;
1006
1007 adjusted_index = LLVMBuildMul(builder, adjusted_index, LLVMConstInt(ctx->ac.i32, stride / type_size, 0), "");
1008
1009 LLVMValueRef val_offset = LLVMConstInt(ctx->ac.i32, offset, 0);
1010 list = LLVMBuildGEP(builder, list, &val_offset, 1, "");
1011 list = LLVMBuildPointerCast(builder, list,
1012 ac_array_in_const32_addr_space(type), "");
1013
1014 LLVMValueRef descriptor = ac_build_load_to_sgpr(&ctx->ac, list, adjusted_index);
1015
1016 /* 3 plane formats always have same size and format for plane 1 & 2, so
1017 * use the tail from plane 1 so that we can store only the first 16 bytes
1018 * of the last plane. */
1019 if (desc_type == AC_DESC_PLANE_2) {
1020 LLVMValueRef descriptor2 = radv_get_sampler_desc(abi, descriptor_set, base_index, constant_index, index, AC_DESC_PLANE_1,image, write, bindless);
1021
1022 LLVMValueRef components[8];
1023 for (unsigned i = 0; i < 4; ++i)
1024 components[i] = ac_llvm_extract_elem(&ctx->ac, descriptor, i);
1025
1026 for (unsigned i = 4; i < 8; ++i)
1027 components[i] = ac_llvm_extract_elem(&ctx->ac, descriptor2, i);
1028 descriptor = ac_build_gather_values(&ctx->ac, components, 8);
1029 }
1030
1031 return descriptor;
1032 }
1033
1034 /* For 2_10_10_10 formats the alpha is handled as unsigned by pre-vega HW.
1035 * so we may need to fix it up. */
1036 static LLVMValueRef
adjust_vertex_fetch_alpha(struct radv_shader_context * ctx,unsigned adjustment,LLVMValueRef alpha)1037 adjust_vertex_fetch_alpha(struct radv_shader_context *ctx,
1038 unsigned adjustment,
1039 LLVMValueRef alpha)
1040 {
1041 if (adjustment == AC_FETCH_FORMAT_NONE)
1042 return alpha;
1043
1044 LLVMValueRef c30 = LLVMConstInt(ctx->ac.i32, 30, 0);
1045
1046 alpha = LLVMBuildBitCast(ctx->ac.builder, alpha, ctx->ac.f32, "");
1047
1048 if (adjustment == AC_FETCH_FORMAT_SSCALED)
1049 alpha = LLVMBuildFPToUI(ctx->ac.builder, alpha, ctx->ac.i32, "");
1050 else
1051 alpha = ac_to_integer(&ctx->ac, alpha);
1052
1053 /* For the integer-like cases, do a natural sign extension.
1054 *
1055 * For the SNORM case, the values are 0.0, 0.333, 0.666, 1.0
1056 * and happen to contain 0, 1, 2, 3 as the two LSBs of the
1057 * exponent.
1058 */
1059 alpha = LLVMBuildShl(ctx->ac.builder, alpha,
1060 adjustment == AC_FETCH_FORMAT_SNORM ?
1061 LLVMConstInt(ctx->ac.i32, 7, 0) : c30, "");
1062 alpha = LLVMBuildAShr(ctx->ac.builder, alpha, c30, "");
1063
1064 /* Convert back to the right type. */
1065 if (adjustment == AC_FETCH_FORMAT_SNORM) {
1066 LLVMValueRef clamp;
1067 LLVMValueRef neg_one = LLVMConstReal(ctx->ac.f32, -1.0);
1068 alpha = LLVMBuildSIToFP(ctx->ac.builder, alpha, ctx->ac.f32, "");
1069 clamp = LLVMBuildFCmp(ctx->ac.builder, LLVMRealULT, alpha, neg_one, "");
1070 alpha = LLVMBuildSelect(ctx->ac.builder, clamp, neg_one, alpha, "");
1071 } else if (adjustment == AC_FETCH_FORMAT_SSCALED) {
1072 alpha = LLVMBuildSIToFP(ctx->ac.builder, alpha, ctx->ac.f32, "");
1073 }
1074
1075 return LLVMBuildBitCast(ctx->ac.builder, alpha, ctx->ac.i32, "");
1076 }
1077
1078 static LLVMValueRef
radv_fixup_vertex_input_fetches(struct radv_shader_context * ctx,LLVMValueRef value,unsigned num_channels,bool is_float)1079 radv_fixup_vertex_input_fetches(struct radv_shader_context *ctx,
1080 LLVMValueRef value,
1081 unsigned num_channels,
1082 bool is_float)
1083 {
1084 LLVMValueRef zero = is_float ? ctx->ac.f32_0 : ctx->ac.i32_0;
1085 LLVMValueRef one = is_float ? ctx->ac.f32_1 : ctx->ac.i32_1;
1086 LLVMValueRef chan[4];
1087
1088 if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMVectorTypeKind) {
1089 unsigned vec_size = LLVMGetVectorSize(LLVMTypeOf(value));
1090
1091 if (num_channels == 4 && num_channels == vec_size)
1092 return value;
1093
1094 num_channels = MIN2(num_channels, vec_size);
1095
1096 for (unsigned i = 0; i < num_channels; i++)
1097 chan[i] = ac_llvm_extract_elem(&ctx->ac, value, i);
1098 } else {
1099 assert(num_channels == 1);
1100 chan[0] = value;
1101 }
1102
1103 for (unsigned i = num_channels; i < 4; i++) {
1104 chan[i] = i == 3 ? one : zero;
1105 chan[i] = ac_to_integer(&ctx->ac, chan[i]);
1106 }
1107
1108 return ac_build_gather_values(&ctx->ac, chan, 4);
1109 }
1110
1111 static void
handle_vs_input_decl(struct radv_shader_context * ctx,struct nir_variable * variable)1112 handle_vs_input_decl(struct radv_shader_context *ctx,
1113 struct nir_variable *variable)
1114 {
1115 LLVMValueRef t_list_ptr = ac_get_arg(&ctx->ac, ctx->args->vertex_buffers);
1116 LLVMValueRef t_offset;
1117 LLVMValueRef t_list;
1118 LLVMValueRef input;
1119 LLVMValueRef buffer_index;
1120 unsigned attrib_count = glsl_count_attribute_slots(variable->type, true);
1121
1122
1123 enum glsl_base_type type = glsl_get_base_type(variable->type);
1124 for (unsigned i = 0; i < attrib_count; ++i) {
1125 LLVMValueRef output[4];
1126 unsigned attrib_index = variable->data.location + i - VERT_ATTRIB_GENERIC0;
1127 unsigned attrib_format = ctx->args->options->key.vs.vertex_attribute_formats[attrib_index];
1128 unsigned data_format = attrib_format & 0x0f;
1129 unsigned num_format = (attrib_format >> 4) & 0x07;
1130 bool is_float = num_format != V_008F0C_BUF_NUM_FORMAT_UINT &&
1131 num_format != V_008F0C_BUF_NUM_FORMAT_SINT;
1132 uint8_t input_usage_mask =
1133 ctx->args->shader_info->vs.input_usage_mask[variable->data.location + i];
1134 unsigned num_input_channels = util_last_bit(input_usage_mask);
1135
1136 if (num_input_channels == 0)
1137 continue;
1138
1139 if (ctx->args->options->key.vs.instance_rate_inputs & (1u << attrib_index)) {
1140 uint32_t divisor = ctx->args->options->key.vs.instance_rate_divisors[attrib_index];
1141
1142 if (divisor) {
1143 buffer_index = ctx->abi.instance_id;
1144
1145 if (divisor != 1) {
1146 buffer_index = LLVMBuildUDiv(ctx->ac.builder, buffer_index,
1147 LLVMConstInt(ctx->ac.i32, divisor, 0), "");
1148 }
1149 } else {
1150 buffer_index = ctx->ac.i32_0;
1151 }
1152
1153 buffer_index = LLVMBuildAdd(ctx->ac.builder,
1154 ac_get_arg(&ctx->ac,
1155 ctx->args->ac.start_instance),\
1156 buffer_index, "");
1157 } else {
1158 buffer_index = LLVMBuildAdd(ctx->ac.builder,
1159 ctx->abi.vertex_id,
1160 ac_get_arg(&ctx->ac,
1161 ctx->args->ac.base_vertex), "");
1162 }
1163
1164 const struct ac_data_format_info *vtx_info = ac_get_data_format_info(data_format);
1165
1166 /* Adjust the number of channels to load based on the vertex
1167 * attribute format.
1168 */
1169 unsigned num_channels = MIN2(num_input_channels, vtx_info->num_channels);
1170 unsigned attrib_binding = ctx->args->options->key.vs.vertex_attribute_bindings[attrib_index];
1171 unsigned attrib_offset = ctx->args->options->key.vs.vertex_attribute_offsets[attrib_index];
1172 unsigned attrib_stride = ctx->args->options->key.vs.vertex_attribute_strides[attrib_index];
1173 unsigned alpha_adjust = ctx->args->options->key.vs.alpha_adjust[attrib_index];
1174
1175 if (ctx->args->options->key.vs.post_shuffle & (1 << attrib_index)) {
1176 /* Always load, at least, 3 channels for formats that
1177 * need to be shuffled because X<->Z.
1178 */
1179 num_channels = MAX2(num_channels, 3);
1180 }
1181
1182 t_offset = LLVMConstInt(ctx->ac.i32, attrib_binding, false);
1183 t_list = ac_build_load_to_sgpr(&ctx->ac, t_list_ptr, t_offset);
1184
1185 /* Always split typed vertex buffer loads on GFX6 and GFX10+
1186 * to avoid any alignment issues that triggers memory
1187 * violations and eventually a GPU hang. This can happen if
1188 * the stride (static or dynamic) is unaligned and also if the
1189 * VBO offset is aligned to a scalar (eg. stride is 8 and VBO
1190 * offset is 2 for R16G16B16A16_SNORM).
1191 */
1192 if (ctx->ac.chip_class == GFX6 ||
1193 ctx->ac.chip_class >= GFX10) {
1194 unsigned chan_format = vtx_info->chan_format;
1195 LLVMValueRef values[4];
1196
1197 assert(ctx->ac.chip_class == GFX6 ||
1198 ctx->ac.chip_class >= GFX10);
1199
1200 for (unsigned chan = 0; chan < num_channels; chan++) {
1201 unsigned chan_offset = attrib_offset + chan * vtx_info->chan_byte_size;
1202 LLVMValueRef chan_index = buffer_index;
1203
1204 if (attrib_stride != 0 && chan_offset > attrib_stride) {
1205 LLVMValueRef buffer_offset =
1206 LLVMConstInt(ctx->ac.i32,
1207 chan_offset / attrib_stride, false);
1208
1209 chan_index = LLVMBuildAdd(ctx->ac.builder,
1210 buffer_index,
1211 buffer_offset, "");
1212
1213 chan_offset = chan_offset % attrib_stride;
1214 }
1215
1216 values[chan] = ac_build_struct_tbuffer_load(&ctx->ac, t_list,
1217 chan_index,
1218 LLVMConstInt(ctx->ac.i32, chan_offset, false),
1219 ctx->ac.i32_0, ctx->ac.i32_0, 1,
1220 chan_format, num_format, 0, true);
1221 }
1222
1223 input = ac_build_gather_values(&ctx->ac, values, num_channels);
1224 } else {
1225 if (attrib_stride != 0 && attrib_offset > attrib_stride) {
1226 LLVMValueRef buffer_offset =
1227 LLVMConstInt(ctx->ac.i32,
1228 attrib_offset / attrib_stride, false);
1229
1230 buffer_index = LLVMBuildAdd(ctx->ac.builder,
1231 buffer_index,
1232 buffer_offset, "");
1233
1234 attrib_offset = attrib_offset % attrib_stride;
1235 }
1236
1237 input = ac_build_struct_tbuffer_load(&ctx->ac, t_list,
1238 buffer_index,
1239 LLVMConstInt(ctx->ac.i32, attrib_offset, false),
1240 ctx->ac.i32_0, ctx->ac.i32_0,
1241 num_channels,
1242 data_format, num_format, 0, true);
1243 }
1244
1245 if (ctx->args->options->key.vs.post_shuffle & (1 << attrib_index)) {
1246 LLVMValueRef c[4];
1247 c[0] = ac_llvm_extract_elem(&ctx->ac, input, 2);
1248 c[1] = ac_llvm_extract_elem(&ctx->ac, input, 1);
1249 c[2] = ac_llvm_extract_elem(&ctx->ac, input, 0);
1250 c[3] = ac_llvm_extract_elem(&ctx->ac, input, 3);
1251
1252 input = ac_build_gather_values(&ctx->ac, c, 4);
1253 }
1254
1255 input = radv_fixup_vertex_input_fetches(ctx, input, num_channels,
1256 is_float);
1257
1258 for (unsigned chan = 0; chan < 4; chan++) {
1259 LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, chan, false);
1260 output[chan] = LLVMBuildExtractElement(ctx->ac.builder, input, llvm_chan, "");
1261 if (type == GLSL_TYPE_FLOAT16) {
1262 output[chan] = LLVMBuildBitCast(ctx->ac.builder, output[chan], ctx->ac.f32, "");
1263 output[chan] = LLVMBuildFPTrunc(ctx->ac.builder, output[chan], ctx->ac.f16, "");
1264 }
1265 }
1266
1267 output[3] = adjust_vertex_fetch_alpha(ctx, alpha_adjust, output[3]);
1268
1269 for (unsigned chan = 0; chan < 4; chan++) {
1270 output[chan] = ac_to_integer(&ctx->ac, output[chan]);
1271 if (type == GLSL_TYPE_UINT16 || type == GLSL_TYPE_INT16)
1272 output[chan] = LLVMBuildTrunc(ctx->ac.builder, output[chan], ctx->ac.i16, "");
1273
1274 ctx->inputs[ac_llvm_reg_index_soa(variable->data.location + i, chan)] = output[chan];
1275 }
1276 }
1277 }
1278
1279 static void
handle_vs_inputs(struct radv_shader_context * ctx,struct nir_shader * nir)1280 handle_vs_inputs(struct radv_shader_context *ctx,
1281 struct nir_shader *nir) {
1282 nir_foreach_shader_in_variable(variable, nir)
1283 handle_vs_input_decl(ctx, variable);
1284 }
1285
1286 static void
prepare_interp_optimize(struct radv_shader_context * ctx,struct nir_shader * nir)1287 prepare_interp_optimize(struct radv_shader_context *ctx,
1288 struct nir_shader *nir)
1289 {
1290 bool uses_center = false;
1291 bool uses_centroid = false;
1292 nir_foreach_shader_in_variable(variable, nir) {
1293 if (glsl_get_base_type(glsl_without_array(variable->type)) != GLSL_TYPE_FLOAT ||
1294 variable->data.sample)
1295 continue;
1296
1297 if (variable->data.centroid)
1298 uses_centroid = true;
1299 else
1300 uses_center = true;
1301 }
1302
1303 ctx->abi.persp_centroid = ac_get_arg(&ctx->ac, ctx->args->ac.persp_centroid);
1304 ctx->abi.linear_centroid = ac_get_arg(&ctx->ac, ctx->args->ac.linear_centroid);
1305
1306 if (uses_center && uses_centroid) {
1307 LLVMValueRef sel = LLVMBuildICmp(ctx->ac.builder, LLVMIntSLT,
1308 ac_get_arg(&ctx->ac, ctx->args->ac.prim_mask),
1309 ctx->ac.i32_0, "");
1310 ctx->abi.persp_centroid =
1311 LLVMBuildSelect(ctx->ac.builder, sel,
1312 ac_get_arg(&ctx->ac, ctx->args->ac.persp_center),
1313 ctx->abi.persp_centroid, "");
1314 ctx->abi.linear_centroid =
1315 LLVMBuildSelect(ctx->ac.builder, sel,
1316 ac_get_arg(&ctx->ac, ctx->args->ac.linear_center),
1317 ctx->abi.linear_centroid, "");
1318 }
1319 }
1320
1321 static void
scan_shader_output_decl(struct radv_shader_context * ctx,struct nir_variable * variable,struct nir_shader * shader,gl_shader_stage stage)1322 scan_shader_output_decl(struct radv_shader_context *ctx,
1323 struct nir_variable *variable,
1324 struct nir_shader *shader,
1325 gl_shader_stage stage)
1326 {
1327 int idx = variable->data.driver_location;
1328 unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
1329 uint64_t mask_attribs;
1330
1331 /* tess ctrl has it's own load/store paths for outputs */
1332 if (stage == MESA_SHADER_TESS_CTRL) {
1333 /* Remember driver location of tess factors, so we can read
1334 * them later, in write_tess_factors.
1335 */
1336 if (variable->data.location == VARYING_SLOT_TESS_LEVEL_INNER) {
1337 ctx->tcs_tess_lvl_inner = idx;
1338 } else if (variable->data.location == VARYING_SLOT_TESS_LEVEL_OUTER) {
1339 ctx->tcs_tess_lvl_outer = idx;
1340 }
1341 return;
1342 }
1343
1344 if (variable->data.compact) {
1345 unsigned component_count = variable->data.location_frac +
1346 glsl_get_length(variable->type);
1347 attrib_count = (component_count + 3) / 4;
1348 }
1349
1350 mask_attribs = ((1ull << attrib_count) - 1) << idx;
1351
1352 ctx->output_mask |= mask_attribs;
1353 }
1354
1355
1356 /* Initialize arguments for the shader export intrinsic */
1357 static void
si_llvm_init_export_args(struct radv_shader_context * ctx,LLVMValueRef * values,unsigned enabled_channels,unsigned target,struct ac_export_args * args)1358 si_llvm_init_export_args(struct radv_shader_context *ctx,
1359 LLVMValueRef *values,
1360 unsigned enabled_channels,
1361 unsigned target,
1362 struct ac_export_args *args)
1363 {
1364 /* Specify the channels that are enabled. */
1365 args->enabled_channels = enabled_channels;
1366
1367 /* Specify whether the EXEC mask represents the valid mask */
1368 args->valid_mask = 0;
1369
1370 /* Specify whether this is the last export */
1371 args->done = 0;
1372
1373 /* Specify the target we are exporting */
1374 args->target = target;
1375
1376 args->compr = false;
1377 args->out[0] = LLVMGetUndef(ctx->ac.f32);
1378 args->out[1] = LLVMGetUndef(ctx->ac.f32);
1379 args->out[2] = LLVMGetUndef(ctx->ac.f32);
1380 args->out[3] = LLVMGetUndef(ctx->ac.f32);
1381
1382 if (!values)
1383 return;
1384
1385 bool is_16bit = ac_get_type_size(LLVMTypeOf(values[0])) == 2;
1386 if (ctx->stage == MESA_SHADER_FRAGMENT) {
1387 unsigned index = target - V_008DFC_SQ_EXP_MRT;
1388 unsigned col_format = (ctx->args->options->key.fs.col_format >> (4 * index)) & 0xf;
1389 bool is_int8 = (ctx->args->options->key.fs.is_int8 >> index) & 1;
1390 bool is_int10 = (ctx->args->options->key.fs.is_int10 >> index) & 1;
1391 unsigned chan;
1392
1393 LLVMValueRef (*packf)(struct ac_llvm_context *ctx, LLVMValueRef args[2]) = NULL;
1394 LLVMValueRef (*packi)(struct ac_llvm_context *ctx, LLVMValueRef args[2],
1395 unsigned bits, bool hi) = NULL;
1396
1397 switch(col_format) {
1398 case V_028714_SPI_SHADER_ZERO:
1399 args->enabled_channels = 0; /* writemask */
1400 args->target = V_008DFC_SQ_EXP_NULL;
1401 break;
1402
1403 case V_028714_SPI_SHADER_32_R:
1404 args->enabled_channels = 1;
1405 args->out[0] = values[0];
1406 break;
1407
1408 case V_028714_SPI_SHADER_32_GR:
1409 args->enabled_channels = 0x3;
1410 args->out[0] = values[0];
1411 args->out[1] = values[1];
1412 break;
1413
1414 case V_028714_SPI_SHADER_32_AR:
1415 if (ctx->ac.chip_class >= GFX10) {
1416 args->enabled_channels = 0x3;
1417 args->out[0] = values[0];
1418 args->out[1] = values[3];
1419 } else {
1420 args->enabled_channels = 0x9;
1421 args->out[0] = values[0];
1422 args->out[3] = values[3];
1423 }
1424 break;
1425
1426 case V_028714_SPI_SHADER_FP16_ABGR:
1427 args->enabled_channels = 0x5;
1428 packf = ac_build_cvt_pkrtz_f16;
1429 if (is_16bit) {
1430 for (unsigned chan = 0; chan < 4; chan++)
1431 values[chan] = LLVMBuildFPExt(ctx->ac.builder,
1432 values[chan],
1433 ctx->ac.f32, "");
1434 }
1435 break;
1436
1437 case V_028714_SPI_SHADER_UNORM16_ABGR:
1438 args->enabled_channels = 0x5;
1439 packf = ac_build_cvt_pknorm_u16;
1440 break;
1441
1442 case V_028714_SPI_SHADER_SNORM16_ABGR:
1443 args->enabled_channels = 0x5;
1444 packf = ac_build_cvt_pknorm_i16;
1445 break;
1446
1447 case V_028714_SPI_SHADER_UINT16_ABGR:
1448 args->enabled_channels = 0x5;
1449 packi = ac_build_cvt_pk_u16;
1450 if (is_16bit) {
1451 for (unsigned chan = 0; chan < 4; chan++)
1452 values[chan] = LLVMBuildZExt(ctx->ac.builder,
1453 ac_to_integer(&ctx->ac, values[chan]),
1454 ctx->ac.i32, "");
1455 }
1456 break;
1457
1458 case V_028714_SPI_SHADER_SINT16_ABGR:
1459 args->enabled_channels = 0x5;
1460 packi = ac_build_cvt_pk_i16;
1461 if (is_16bit) {
1462 for (unsigned chan = 0; chan < 4; chan++)
1463 values[chan] = LLVMBuildSExt(ctx->ac.builder,
1464 ac_to_integer(&ctx->ac, values[chan]),
1465 ctx->ac.i32, "");
1466 }
1467 break;
1468
1469 default:
1470 case V_028714_SPI_SHADER_32_ABGR:
1471 memcpy(&args->out[0], values, sizeof(values[0]) * 4);
1472 break;
1473 }
1474
1475 /* Replace NaN by zero (only 32-bit) to fix game bugs if
1476 * requested.
1477 */
1478 if (ctx->args->options->enable_mrt_output_nan_fixup &&
1479 !is_16bit &&
1480 (col_format == V_028714_SPI_SHADER_32_R ||
1481 col_format == V_028714_SPI_SHADER_32_GR ||
1482 col_format == V_028714_SPI_SHADER_32_AR ||
1483 col_format == V_028714_SPI_SHADER_32_ABGR ||
1484 col_format == V_028714_SPI_SHADER_FP16_ABGR)) {
1485 for (unsigned i = 0; i < 4; i++) {
1486 LLVMValueRef args[2] = {
1487 values[i],
1488 LLVMConstInt(ctx->ac.i32, S_NAN | Q_NAN, false)
1489 };
1490 LLVMValueRef isnan =
1491 ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.class.f32", ctx->ac.i1,
1492 args, 2, AC_FUNC_ATTR_READNONE);
1493 values[i] = LLVMBuildSelect(ctx->ac.builder, isnan,
1494 ctx->ac.f32_0,
1495 values[i], "");
1496 }
1497 }
1498
1499 /* Pack f16 or norm_i16/u16. */
1500 if (packf) {
1501 for (chan = 0; chan < 2; chan++) {
1502 LLVMValueRef pack_args[2] = {
1503 values[2 * chan],
1504 values[2 * chan + 1]
1505 };
1506 LLVMValueRef packed;
1507
1508 packed = packf(&ctx->ac, pack_args);
1509 args->out[chan] = ac_to_float(&ctx->ac, packed);
1510 }
1511 args->compr = 1; /* COMPR flag */
1512 }
1513
1514 /* Pack i16/u16. */
1515 if (packi) {
1516 for (chan = 0; chan < 2; chan++) {
1517 LLVMValueRef pack_args[2] = {
1518 ac_to_integer(&ctx->ac, values[2 * chan]),
1519 ac_to_integer(&ctx->ac, values[2 * chan + 1])
1520 };
1521 LLVMValueRef packed;
1522
1523 packed = packi(&ctx->ac, pack_args,
1524 is_int8 ? 8 : is_int10 ? 10 : 16,
1525 chan == 1);
1526 args->out[chan] = ac_to_float(&ctx->ac, packed);
1527 }
1528 args->compr = 1; /* COMPR flag */
1529 }
1530 return;
1531 }
1532
1533 if (is_16bit) {
1534 for (unsigned chan = 0; chan < 4; chan++) {
1535 values[chan] = LLVMBuildBitCast(ctx->ac.builder, values[chan], ctx->ac.i16, "");
1536 args->out[chan] = LLVMBuildZExt(ctx->ac.builder, values[chan], ctx->ac.i32, "");
1537 }
1538 } else
1539 memcpy(&args->out[0], values, sizeof(values[0]) * 4);
1540
1541 for (unsigned i = 0; i < 4; ++i)
1542 args->out[i] = ac_to_float(&ctx->ac, args->out[i]);
1543 }
1544
1545 static void
radv_export_param(struct radv_shader_context * ctx,unsigned index,LLVMValueRef * values,unsigned enabled_channels)1546 radv_export_param(struct radv_shader_context *ctx, unsigned index,
1547 LLVMValueRef *values, unsigned enabled_channels)
1548 {
1549 struct ac_export_args args;
1550
1551 si_llvm_init_export_args(ctx, values, enabled_channels,
1552 V_008DFC_SQ_EXP_PARAM + index, &args);
1553 ac_build_export(&ctx->ac, &args);
1554 }
1555
1556 static LLVMValueRef
radv_load_output(struct radv_shader_context * ctx,unsigned index,unsigned chan)1557 radv_load_output(struct radv_shader_context *ctx, unsigned index, unsigned chan)
1558 {
1559 LLVMValueRef output = ctx->abi.outputs[ac_llvm_reg_index_soa(index, chan)];
1560 return LLVMBuildLoad(ctx->ac.builder, output, "");
1561 }
1562
1563 static void
radv_emit_stream_output(struct radv_shader_context * ctx,LLVMValueRef const * so_buffers,LLVMValueRef const * so_write_offsets,const struct radv_stream_output * output,struct radv_shader_output_values * shader_out)1564 radv_emit_stream_output(struct radv_shader_context *ctx,
1565 LLVMValueRef const *so_buffers,
1566 LLVMValueRef const *so_write_offsets,
1567 const struct radv_stream_output *output,
1568 struct radv_shader_output_values *shader_out)
1569 {
1570 unsigned num_comps = util_bitcount(output->component_mask);
1571 unsigned buf = output->buffer;
1572 unsigned offset = output->offset;
1573 unsigned start;
1574 LLVMValueRef out[4];
1575
1576 assert(num_comps && num_comps <= 4);
1577 if (!num_comps || num_comps > 4)
1578 return;
1579
1580 /* Get the first component. */
1581 start = ffs(output->component_mask) - 1;
1582
1583 /* Load the output as int. */
1584 for (int i = 0; i < num_comps; i++) {
1585 out[i] = ac_to_integer(&ctx->ac, shader_out->values[start + i]);
1586 }
1587
1588 /* Pack the output. */
1589 LLVMValueRef vdata = NULL;
1590
1591 switch (num_comps) {
1592 case 1: /* as i32 */
1593 vdata = out[0];
1594 break;
1595 case 2: /* as v2i32 */
1596 case 3: /* as v4i32 (aligned to 4) */
1597 out[3] = LLVMGetUndef(ctx->ac.i32);
1598 /* fall through */
1599 case 4: /* as v4i32 */
1600 vdata = ac_build_gather_values(&ctx->ac, out,
1601 !ac_has_vec3_support(ctx->ac.chip_class, false) ?
1602 util_next_power_of_two(num_comps) :
1603 num_comps);
1604 break;
1605 }
1606
1607 ac_build_buffer_store_dword(&ctx->ac, so_buffers[buf],
1608 vdata, num_comps, so_write_offsets[buf],
1609 ctx->ac.i32_0, offset,
1610 ac_glc | ac_slc);
1611 }
1612
1613 static void
radv_emit_streamout(struct radv_shader_context * ctx,unsigned stream)1614 radv_emit_streamout(struct radv_shader_context *ctx, unsigned stream)
1615 {
1616 int i;
1617
1618 /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
1619 assert(ctx->args->streamout_config.used);
1620 LLVMValueRef so_vtx_count =
1621 ac_build_bfe(&ctx->ac,
1622 ac_get_arg(&ctx->ac, ctx->args->streamout_config),
1623 LLVMConstInt(ctx->ac.i32, 16, false),
1624 LLVMConstInt(ctx->ac.i32, 7, false), false);
1625
1626 LLVMValueRef tid = ac_get_thread_id(&ctx->ac);
1627
1628 /* can_emit = tid < so_vtx_count; */
1629 LLVMValueRef can_emit = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT,
1630 tid, so_vtx_count, "");
1631
1632 /* Emit the streamout code conditionally. This actually avoids
1633 * out-of-bounds buffer access. The hw tells us via the SGPR
1634 * (so_vtx_count) which threads are allowed to emit streamout data.
1635 */
1636 ac_build_ifcc(&ctx->ac, can_emit, 6501);
1637 {
1638 /* The buffer offset is computed as follows:
1639 * ByteOffset = streamout_offset[buffer_id]*4 +
1640 * (streamout_write_index + thread_id)*stride[buffer_id] +
1641 * attrib_offset
1642 */
1643 LLVMValueRef so_write_index =
1644 ac_get_arg(&ctx->ac, ctx->args->streamout_write_idx);
1645
1646 /* Compute (streamout_write_index + thread_id). */
1647 so_write_index =
1648 LLVMBuildAdd(ctx->ac.builder, so_write_index, tid, "");
1649
1650 /* Load the descriptor and compute the write offset for each
1651 * enabled buffer.
1652 */
1653 LLVMValueRef so_write_offset[4] = {0};
1654 LLVMValueRef so_buffers[4] = {0};
1655 LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->args->streamout_buffers);
1656
1657 for (i = 0; i < 4; i++) {
1658 uint16_t stride = ctx->args->shader_info->so.strides[i];
1659
1660 if (!stride)
1661 continue;
1662
1663 LLVMValueRef offset =
1664 LLVMConstInt(ctx->ac.i32, i, false);
1665
1666 so_buffers[i] = ac_build_load_to_sgpr(&ctx->ac,
1667 buf_ptr, offset);
1668
1669 LLVMValueRef so_offset =
1670 ac_get_arg(&ctx->ac, ctx->args->streamout_offset[i]);
1671
1672 so_offset = LLVMBuildMul(ctx->ac.builder, so_offset,
1673 LLVMConstInt(ctx->ac.i32, 4, false), "");
1674
1675 so_write_offset[i] =
1676 ac_build_imad(&ctx->ac, so_write_index,
1677 LLVMConstInt(ctx->ac.i32,
1678 stride * 4, false),
1679 so_offset);
1680 }
1681
1682 /* Write streamout data. */
1683 for (i = 0; i < ctx->args->shader_info->so.num_outputs; i++) {
1684 struct radv_shader_output_values shader_out = {0};
1685 struct radv_stream_output *output =
1686 &ctx->args->shader_info->so.outputs[i];
1687
1688 if (stream != output->stream)
1689 continue;
1690
1691 for (int j = 0; j < 4; j++) {
1692 shader_out.values[j] =
1693 radv_load_output(ctx, output->location, j);
1694 }
1695
1696 radv_emit_stream_output(ctx, so_buffers,so_write_offset,
1697 output, &shader_out);
1698 }
1699 }
1700 ac_build_endif(&ctx->ac, 6501);
1701 }
1702
1703 static void
radv_build_param_exports(struct radv_shader_context * ctx,struct radv_shader_output_values * outputs,unsigned noutput,struct radv_vs_output_info * outinfo,bool export_clip_dists)1704 radv_build_param_exports(struct radv_shader_context *ctx,
1705 struct radv_shader_output_values *outputs,
1706 unsigned noutput,
1707 struct radv_vs_output_info *outinfo,
1708 bool export_clip_dists)
1709 {
1710 unsigned param_count = 0;
1711
1712 for (unsigned i = 0; i < noutput; i++) {
1713 unsigned slot_name = outputs[i].slot_name;
1714 unsigned usage_mask = outputs[i].usage_mask;
1715
1716 if (slot_name != VARYING_SLOT_LAYER &&
1717 slot_name != VARYING_SLOT_PRIMITIVE_ID &&
1718 slot_name != VARYING_SLOT_VIEWPORT &&
1719 slot_name != VARYING_SLOT_CLIP_DIST0 &&
1720 slot_name != VARYING_SLOT_CLIP_DIST1 &&
1721 slot_name < VARYING_SLOT_VAR0)
1722 continue;
1723
1724 if ((slot_name == VARYING_SLOT_CLIP_DIST0 ||
1725 slot_name == VARYING_SLOT_CLIP_DIST1) && !export_clip_dists)
1726 continue;
1727
1728 radv_export_param(ctx, param_count, outputs[i].values, usage_mask);
1729
1730 assert(i < ARRAY_SIZE(outinfo->vs_output_param_offset));
1731 outinfo->vs_output_param_offset[slot_name] = param_count++;
1732 }
1733
1734 outinfo->param_exports = param_count;
1735 }
1736
1737 /* Generate export instructions for hardware VS shader stage or NGG GS stage
1738 * (position and parameter data only).
1739 */
1740 static void
radv_llvm_export_vs(struct radv_shader_context * ctx,struct radv_shader_output_values * outputs,unsigned noutput,struct radv_vs_output_info * outinfo,bool export_clip_dists)1741 radv_llvm_export_vs(struct radv_shader_context *ctx,
1742 struct radv_shader_output_values *outputs,
1743 unsigned noutput,
1744 struct radv_vs_output_info *outinfo,
1745 bool export_clip_dists)
1746 {
1747 LLVMValueRef psize_value = NULL, layer_value = NULL, viewport_value = NULL;
1748 struct ac_export_args pos_args[4] = {0};
1749 unsigned pos_idx, index;
1750 int i;
1751
1752 /* Build position exports */
1753 for (i = 0; i < noutput; i++) {
1754 switch (outputs[i].slot_name) {
1755 case VARYING_SLOT_POS:
1756 si_llvm_init_export_args(ctx, outputs[i].values, 0xf,
1757 V_008DFC_SQ_EXP_POS, &pos_args[0]);
1758 break;
1759 case VARYING_SLOT_PSIZ:
1760 psize_value = outputs[i].values[0];
1761 break;
1762 case VARYING_SLOT_LAYER:
1763 layer_value = outputs[i].values[0];
1764 break;
1765 case VARYING_SLOT_VIEWPORT:
1766 viewport_value = outputs[i].values[0];
1767 break;
1768 case VARYING_SLOT_CLIP_DIST0:
1769 case VARYING_SLOT_CLIP_DIST1:
1770 index = 2 + outputs[i].slot_index;
1771 si_llvm_init_export_args(ctx, outputs[i].values, 0xf,
1772 V_008DFC_SQ_EXP_POS + index,
1773 &pos_args[index]);
1774 break;
1775 default:
1776 break;
1777 }
1778 }
1779
1780 /* We need to add the position output manually if it's missing. */
1781 if (!pos_args[0].out[0]) {
1782 pos_args[0].enabled_channels = 0xf; /* writemask */
1783 pos_args[0].valid_mask = 0; /* EXEC mask */
1784 pos_args[0].done = 0; /* last export? */
1785 pos_args[0].target = V_008DFC_SQ_EXP_POS;
1786 pos_args[0].compr = 0; /* COMPR flag */
1787 pos_args[0].out[0] = ctx->ac.f32_0; /* X */
1788 pos_args[0].out[1] = ctx->ac.f32_0; /* Y */
1789 pos_args[0].out[2] = ctx->ac.f32_0; /* Z */
1790 pos_args[0].out[3] = ctx->ac.f32_1; /* W */
1791 }
1792
1793 if (outinfo->writes_pointsize ||
1794 outinfo->writes_layer ||
1795 outinfo->writes_viewport_index) {
1796 pos_args[1].enabled_channels = ((outinfo->writes_pointsize == true ? 1 : 0) |
1797 (outinfo->writes_layer == true ? 4 : 0));
1798 pos_args[1].valid_mask = 0;
1799 pos_args[1].done = 0;
1800 pos_args[1].target = V_008DFC_SQ_EXP_POS + 1;
1801 pos_args[1].compr = 0;
1802 pos_args[1].out[0] = ctx->ac.f32_0; /* X */
1803 pos_args[1].out[1] = ctx->ac.f32_0; /* Y */
1804 pos_args[1].out[2] = ctx->ac.f32_0; /* Z */
1805 pos_args[1].out[3] = ctx->ac.f32_0; /* W */
1806
1807 if (outinfo->writes_pointsize == true)
1808 pos_args[1].out[0] = psize_value;
1809 if (outinfo->writes_layer == true)
1810 pos_args[1].out[2] = layer_value;
1811 if (outinfo->writes_viewport_index == true) {
1812 if (ctx->args->options->chip_class >= GFX9) {
1813 /* GFX9 has the layer in out.z[10:0] and the viewport
1814 * index in out.z[19:16].
1815 */
1816 LLVMValueRef v = viewport_value;
1817 v = ac_to_integer(&ctx->ac, v);
1818 v = LLVMBuildShl(ctx->ac.builder, v,
1819 LLVMConstInt(ctx->ac.i32, 16, false),
1820 "");
1821 v = LLVMBuildOr(ctx->ac.builder, v,
1822 ac_to_integer(&ctx->ac, pos_args[1].out[2]), "");
1823
1824 pos_args[1].out[2] = ac_to_float(&ctx->ac, v);
1825 pos_args[1].enabled_channels |= 1 << 2;
1826 } else {
1827 pos_args[1].out[3] = viewport_value;
1828 pos_args[1].enabled_channels |= 1 << 3;
1829 }
1830 }
1831 }
1832
1833 for (i = 0; i < 4; i++) {
1834 if (pos_args[i].out[0])
1835 outinfo->pos_exports++;
1836 }
1837
1838 /* GFX10 skip POS0 exports if EXEC=0 and DONE=0, causing a hang.
1839 * Setting valid_mask=1 prevents it and has no other effect.
1840 */
1841 if (ctx->ac.chip_class == GFX10)
1842 pos_args[0].valid_mask = 1;
1843
1844 pos_idx = 0;
1845 for (i = 0; i < 4; i++) {
1846 if (!pos_args[i].out[0])
1847 continue;
1848
1849 /* Specify the target we are exporting */
1850 pos_args[i].target = V_008DFC_SQ_EXP_POS + pos_idx++;
1851
1852 if (pos_idx == outinfo->pos_exports)
1853 /* Specify that this is the last export */
1854 pos_args[i].done = 1;
1855
1856 ac_build_export(&ctx->ac, &pos_args[i]);
1857 }
1858
1859 /* Build parameter exports */
1860 radv_build_param_exports(ctx, outputs, noutput, outinfo, export_clip_dists);
1861 }
1862
1863 static void
handle_vs_outputs_post(struct radv_shader_context * ctx,bool export_prim_id,bool export_clip_dists,struct radv_vs_output_info * outinfo)1864 handle_vs_outputs_post(struct radv_shader_context *ctx,
1865 bool export_prim_id,
1866 bool export_clip_dists,
1867 struct radv_vs_output_info *outinfo)
1868 {
1869 struct radv_shader_output_values *outputs;
1870 unsigned noutput = 0;
1871
1872 if (ctx->args->options->key.has_multiview_view_index) {
1873 LLVMValueRef* tmp_out = &ctx->abi.outputs[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER, 0)];
1874 if(!*tmp_out) {
1875 for(unsigned i = 0; i < 4; ++i)
1876 ctx->abi.outputs[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER, i)] =
1877 ac_build_alloca_undef(&ctx->ac, ctx->ac.f32, "");
1878 }
1879
1880 LLVMValueRef view_index = ac_get_arg(&ctx->ac, ctx->args->ac.view_index);
1881 LLVMBuildStore(ctx->ac.builder, ac_to_float(&ctx->ac, view_index), *tmp_out);
1882 ctx->output_mask |= 1ull << VARYING_SLOT_LAYER;
1883 }
1884
1885 memset(outinfo->vs_output_param_offset, AC_EXP_PARAM_UNDEFINED,
1886 sizeof(outinfo->vs_output_param_offset));
1887 outinfo->pos_exports = 0;
1888
1889 if (!ctx->args->options->use_ngg_streamout &&
1890 ctx->args->shader_info->so.num_outputs &&
1891 !ctx->args->is_gs_copy_shader) {
1892 /* The GS copy shader emission already emits streamout. */
1893 radv_emit_streamout(ctx, 0);
1894 }
1895
1896 /* Allocate a temporary array for the output values. */
1897 unsigned num_outputs = util_bitcount64(ctx->output_mask) + export_prim_id;
1898 outputs = malloc(num_outputs * sizeof(outputs[0]));
1899
1900 for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
1901 if (!(ctx->output_mask & (1ull << i)))
1902 continue;
1903
1904 outputs[noutput].slot_name = i;
1905 outputs[noutput].slot_index = i == VARYING_SLOT_CLIP_DIST1;
1906
1907 if (ctx->stage == MESA_SHADER_VERTEX &&
1908 !ctx->args->is_gs_copy_shader) {
1909 outputs[noutput].usage_mask =
1910 ctx->args->shader_info->vs.output_usage_mask[i];
1911 } else if (ctx->stage == MESA_SHADER_TESS_EVAL) {
1912 outputs[noutput].usage_mask =
1913 ctx->args->shader_info->tes.output_usage_mask[i];
1914 } else {
1915 assert(ctx->args->is_gs_copy_shader);
1916 outputs[noutput].usage_mask =
1917 ctx->args->shader_info->gs.output_usage_mask[i];
1918 }
1919
1920 for (unsigned j = 0; j < 4; j++) {
1921 outputs[noutput].values[j] =
1922 ac_to_float(&ctx->ac, radv_load_output(ctx, i, j));
1923 }
1924
1925 noutput++;
1926 }
1927
1928 /* Export PrimitiveID. */
1929 if (export_prim_id) {
1930 outputs[noutput].slot_name = VARYING_SLOT_PRIMITIVE_ID;
1931 outputs[noutput].slot_index = 0;
1932 outputs[noutput].usage_mask = 0x1;
1933 if (ctx->stage == MESA_SHADER_TESS_EVAL)
1934 outputs[noutput].values[0] =
1935 ac_get_arg(&ctx->ac, ctx->args->ac.tes_patch_id);
1936 else
1937 outputs[noutput].values[0] =
1938 ac_get_arg(&ctx->ac, ctx->args->vs_prim_id);
1939 for (unsigned j = 1; j < 4; j++)
1940 outputs[noutput].values[j] = ctx->ac.f32_0;
1941 noutput++;
1942 }
1943
1944 radv_llvm_export_vs(ctx, outputs, noutput, outinfo, export_clip_dists);
1945
1946 free(outputs);
1947 }
1948
1949 static void
handle_es_outputs_post(struct radv_shader_context * ctx,struct radv_es_output_info * outinfo)1950 handle_es_outputs_post(struct radv_shader_context *ctx,
1951 struct radv_es_output_info *outinfo)
1952 {
1953 int j;
1954 LLVMValueRef lds_base = NULL;
1955
1956 if (ctx->ac.chip_class >= GFX9) {
1957 unsigned itemsize_dw = outinfo->esgs_itemsize / 4;
1958 LLVMValueRef vertex_idx = ac_get_thread_id(&ctx->ac);
1959 LLVMValueRef wave_idx =
1960 ac_unpack_param(&ctx->ac,
1961 ac_get_arg(&ctx->ac, ctx->args->merged_wave_info), 24, 4);
1962 vertex_idx = LLVMBuildOr(ctx->ac.builder, vertex_idx,
1963 LLVMBuildMul(ctx->ac.builder, wave_idx,
1964 LLVMConstInt(ctx->ac.i32,
1965 ctx->ac.wave_size, false), ""), "");
1966 lds_base = LLVMBuildMul(ctx->ac.builder, vertex_idx,
1967 LLVMConstInt(ctx->ac.i32, itemsize_dw, 0), "");
1968 }
1969
1970 for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
1971 LLVMValueRef dw_addr = NULL;
1972 LLVMValueRef *out_ptr = &ctx->abi.outputs[i * 4];
1973 unsigned output_usage_mask;
1974
1975 if (!(ctx->output_mask & (1ull << i)))
1976 continue;
1977
1978 if (ctx->stage == MESA_SHADER_VERTEX) {
1979 output_usage_mask =
1980 ctx->args->shader_info->vs.output_usage_mask[i];
1981 } else {
1982 assert(ctx->stage == MESA_SHADER_TESS_EVAL);
1983 output_usage_mask =
1984 ctx->args->shader_info->tes.output_usage_mask[i];
1985 }
1986
1987 if (lds_base) {
1988 dw_addr = LLVMBuildAdd(ctx->ac.builder, lds_base,
1989 LLVMConstInt(ctx->ac.i32, i * 4, false),
1990 "");
1991 }
1992
1993 for (j = 0; j < 4; j++) {
1994 if (!(output_usage_mask & (1 << j)))
1995 continue;
1996
1997 LLVMValueRef out_val = LLVMBuildLoad(ctx->ac.builder, out_ptr[j], "");
1998 out_val = ac_to_integer(&ctx->ac, out_val);
1999 out_val = LLVMBuildZExtOrBitCast(ctx->ac.builder, out_val, ctx->ac.i32, "");
2000
2001 if (ctx->ac.chip_class >= GFX9) {
2002 LLVMValueRef dw_addr_offset =
2003 LLVMBuildAdd(ctx->ac.builder, dw_addr,
2004 LLVMConstInt(ctx->ac.i32,
2005 j, false), "");
2006
2007 ac_lds_store(&ctx->ac, dw_addr_offset, out_val);
2008 } else {
2009 ac_build_buffer_store_dword(&ctx->ac,
2010 ctx->esgs_ring,
2011 out_val, 1,
2012 NULL,
2013 ac_get_arg(&ctx->ac, ctx->args->es2gs_offset),
2014 (4 * i + j) * 4,
2015 ac_glc | ac_slc | ac_swizzled);
2016 }
2017 }
2018 }
2019 }
2020
2021 static void
handle_ls_outputs_post(struct radv_shader_context * ctx)2022 handle_ls_outputs_post(struct radv_shader_context *ctx)
2023 {
2024 LLVMValueRef vertex_id = ctx->rel_auto_id;
2025 uint32_t num_tcs_inputs = ctx->args->shader_info->vs.num_linked_outputs;
2026 LLVMValueRef vertex_dw_stride = LLVMConstInt(ctx->ac.i32, num_tcs_inputs * 4, false);
2027 LLVMValueRef base_dw_addr = LLVMBuildMul(ctx->ac.builder, vertex_id,
2028 vertex_dw_stride, "");
2029
2030 for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
2031 LLVMValueRef *out_ptr = &ctx->abi.outputs[i * 4];
2032
2033 if (!(ctx->output_mask & (1ull << i)))
2034 continue;
2035
2036 LLVMValueRef dw_addr = LLVMBuildAdd(ctx->ac.builder, base_dw_addr,
2037 LLVMConstInt(ctx->ac.i32, i * 4, false),
2038 "");
2039 for (unsigned j = 0; j < 4; j++) {
2040 LLVMValueRef value = LLVMBuildLoad(ctx->ac.builder, out_ptr[j], "");
2041 value = ac_to_integer(&ctx->ac, value);
2042 value = LLVMBuildZExtOrBitCast(ctx->ac.builder, value, ctx->ac.i32, "");
2043 ac_lds_store(&ctx->ac, dw_addr, value);
2044 dw_addr = LLVMBuildAdd(ctx->ac.builder, dw_addr, ctx->ac.i32_1, "");
2045 }
2046 }
2047 }
2048
get_wave_id_in_tg(struct radv_shader_context * ctx)2049 static LLVMValueRef get_wave_id_in_tg(struct radv_shader_context *ctx)
2050 {
2051 return ac_unpack_param(&ctx->ac,
2052 ac_get_arg(&ctx->ac, ctx->args->merged_wave_info), 24, 4);
2053 }
2054
get_tgsize(struct radv_shader_context * ctx)2055 static LLVMValueRef get_tgsize(struct radv_shader_context *ctx)
2056 {
2057 return ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->merged_wave_info), 28, 4);
2058 }
2059
get_thread_id_in_tg(struct radv_shader_context * ctx)2060 static LLVMValueRef get_thread_id_in_tg(struct radv_shader_context *ctx)
2061 {
2062 LLVMBuilderRef builder = ctx->ac.builder;
2063 LLVMValueRef tmp;
2064 tmp = LLVMBuildMul(builder, get_wave_id_in_tg(ctx),
2065 LLVMConstInt(ctx->ac.i32, ctx->ac.wave_size, false), "");
2066 return LLVMBuildAdd(builder, tmp, ac_get_thread_id(&ctx->ac), "");
2067 }
2068
ngg_get_vtx_cnt(struct radv_shader_context * ctx)2069 static LLVMValueRef ngg_get_vtx_cnt(struct radv_shader_context *ctx)
2070 {
2071 return ac_build_bfe(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->gs_tg_info),
2072 LLVMConstInt(ctx->ac.i32, 12, false),
2073 LLVMConstInt(ctx->ac.i32, 9, false),
2074 false);
2075 }
2076
ngg_get_prim_cnt(struct radv_shader_context * ctx)2077 static LLVMValueRef ngg_get_prim_cnt(struct radv_shader_context *ctx)
2078 {
2079 return ac_build_bfe(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->gs_tg_info),
2080 LLVMConstInt(ctx->ac.i32, 22, false),
2081 LLVMConstInt(ctx->ac.i32, 9, false),
2082 false);
2083 }
2084
ngg_get_ordered_id(struct radv_shader_context * ctx)2085 static LLVMValueRef ngg_get_ordered_id(struct radv_shader_context *ctx)
2086 {
2087 return ac_build_bfe(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->gs_tg_info),
2088 ctx->ac.i32_0,
2089 LLVMConstInt(ctx->ac.i32, 12, false),
2090 false);
2091 }
2092
2093 static LLVMValueRef
ngg_gs_get_vertex_storage(struct radv_shader_context * ctx)2094 ngg_gs_get_vertex_storage(struct radv_shader_context *ctx)
2095 {
2096 unsigned num_outputs = util_bitcount64(ctx->output_mask);
2097
2098 if (ctx->args->options->key.has_multiview_view_index)
2099 num_outputs++;
2100
2101 LLVMTypeRef elements[2] = {
2102 LLVMArrayType(ctx->ac.i32, 4 * num_outputs),
2103 LLVMArrayType(ctx->ac.i8, 4),
2104 };
2105 LLVMTypeRef type = LLVMStructTypeInContext(ctx->ac.context, elements, 2, false);
2106 type = LLVMPointerType(LLVMArrayType(type, 0), AC_ADDR_SPACE_LDS);
2107 return LLVMBuildBitCast(ctx->ac.builder, ctx->gs_ngg_emit, type, "");
2108 }
2109
2110 /**
2111 * Return a pointer to the LDS storage reserved for the N'th vertex, where N
2112 * is in emit order; that is:
2113 * - during the epilogue, N is the threadidx (relative to the entire threadgroup)
2114 * - during vertex emit, i.e. while the API GS shader invocation is running,
2115 * N = threadidx * gs_max_out_vertices + emitidx
2116 *
2117 * Goals of the LDS memory layout:
2118 * 1. Eliminate bank conflicts on write for geometry shaders that have all emits
2119 * in uniform control flow
2120 * 2. Eliminate bank conflicts on read for export if, additionally, there is no
2121 * culling
2122 * 3. Agnostic to the number of waves (since we don't know it before compiling)
2123 * 4. Allow coalescing of LDS instructions (ds_write_b128 etc.)
2124 * 5. Avoid wasting memory.
2125 *
2126 * We use an AoS layout due to point 4 (this also helps point 3). In an AoS
2127 * layout, elimination of bank conflicts requires that each vertex occupy an
2128 * odd number of dwords. We use the additional dword to store the output stream
2129 * index as well as a flag to indicate whether this vertex ends a primitive
2130 * for rasterization.
2131 *
2132 * Swizzling is required to satisfy points 1 and 2 simultaneously.
2133 *
2134 * Vertices are stored in export order (gsthread * gs_max_out_vertices + emitidx).
2135 * Indices are swizzled in groups of 32, which ensures point 1 without
2136 * disturbing point 2.
2137 *
2138 * \return an LDS pointer to type {[N x i32], [4 x i8]}
2139 */
2140 static LLVMValueRef
ngg_gs_vertex_ptr(struct radv_shader_context * ctx,LLVMValueRef vertexidx)2141 ngg_gs_vertex_ptr(struct radv_shader_context *ctx, LLVMValueRef vertexidx)
2142 {
2143 LLVMBuilderRef builder = ctx->ac.builder;
2144 LLVMValueRef storage = ngg_gs_get_vertex_storage(ctx);
2145
2146 /* gs_max_out_vertices = 2^(write_stride_2exp) * some odd number */
2147 unsigned write_stride_2exp = ffs(ctx->shader->info.gs.vertices_out) - 1;
2148 if (write_stride_2exp) {
2149 LLVMValueRef row =
2150 LLVMBuildLShr(builder, vertexidx,
2151 LLVMConstInt(ctx->ac.i32, 5, false), "");
2152 LLVMValueRef swizzle =
2153 LLVMBuildAnd(builder, row,
2154 LLVMConstInt(ctx->ac.i32, (1u << write_stride_2exp) - 1,
2155 false), "");
2156 vertexidx = LLVMBuildXor(builder, vertexidx, swizzle, "");
2157 }
2158
2159 return ac_build_gep0(&ctx->ac, storage, vertexidx);
2160 }
2161
2162 static LLVMValueRef
ngg_gs_emit_vertex_ptr(struct radv_shader_context * ctx,LLVMValueRef gsthread,LLVMValueRef emitidx)2163 ngg_gs_emit_vertex_ptr(struct radv_shader_context *ctx, LLVMValueRef gsthread,
2164 LLVMValueRef emitidx)
2165 {
2166 LLVMBuilderRef builder = ctx->ac.builder;
2167 LLVMValueRef tmp;
2168
2169 tmp = LLVMConstInt(ctx->ac.i32, ctx->shader->info.gs.vertices_out, false);
2170 tmp = LLVMBuildMul(builder, tmp, gsthread, "");
2171 const LLVMValueRef vertexidx = LLVMBuildAdd(builder, tmp, emitidx, "");
2172 return ngg_gs_vertex_ptr(ctx, vertexidx);
2173 }
2174
2175 static LLVMValueRef
ngg_gs_get_emit_output_ptr(struct radv_shader_context * ctx,LLVMValueRef vertexptr,unsigned out_idx)2176 ngg_gs_get_emit_output_ptr(struct radv_shader_context *ctx, LLVMValueRef vertexptr,
2177 unsigned out_idx)
2178 {
2179 LLVMValueRef gep_idx[3] = {
2180 ctx->ac.i32_0, /* implied C-style array */
2181 ctx->ac.i32_0, /* first struct entry */
2182 LLVMConstInt(ctx->ac.i32, out_idx, false),
2183 };
2184 return LLVMBuildGEP(ctx->ac.builder, vertexptr, gep_idx, 3, "");
2185 }
2186
2187 static LLVMValueRef
ngg_gs_get_emit_primflag_ptr(struct radv_shader_context * ctx,LLVMValueRef vertexptr,unsigned stream)2188 ngg_gs_get_emit_primflag_ptr(struct radv_shader_context *ctx, LLVMValueRef vertexptr,
2189 unsigned stream)
2190 {
2191 LLVMValueRef gep_idx[3] = {
2192 ctx->ac.i32_0, /* implied C-style array */
2193 ctx->ac.i32_1, /* second struct entry */
2194 LLVMConstInt(ctx->ac.i32, stream, false),
2195 };
2196 return LLVMBuildGEP(ctx->ac.builder, vertexptr, gep_idx, 3, "");
2197 }
2198
2199 static struct radv_stream_output *
radv_get_stream_output_by_loc(struct radv_streamout_info * so,unsigned location)2200 radv_get_stream_output_by_loc(struct radv_streamout_info *so, unsigned location)
2201 {
2202 for (unsigned i = 0; i < so->num_outputs; ++i) {
2203 if (so->outputs[i].location == location)
2204 return &so->outputs[i];
2205 }
2206
2207 return NULL;
2208 }
2209
build_streamout_vertex(struct radv_shader_context * ctx,LLVMValueRef * so_buffer,LLVMValueRef * wg_offset_dw,unsigned stream,LLVMValueRef offset_vtx,LLVMValueRef vertexptr)2210 static void build_streamout_vertex(struct radv_shader_context *ctx,
2211 LLVMValueRef *so_buffer, LLVMValueRef *wg_offset_dw,
2212 unsigned stream, LLVMValueRef offset_vtx,
2213 LLVMValueRef vertexptr)
2214 {
2215 struct radv_streamout_info *so = &ctx->args->shader_info->so;
2216 LLVMBuilderRef builder = ctx->ac.builder;
2217 LLVMValueRef offset[4] = {0};
2218 LLVMValueRef tmp;
2219
2220 for (unsigned buffer = 0; buffer < 4; ++buffer) {
2221 if (!wg_offset_dw[buffer])
2222 continue;
2223
2224 tmp = LLVMBuildMul(builder, offset_vtx,
2225 LLVMConstInt(ctx->ac.i32, so->strides[buffer], false), "");
2226 tmp = LLVMBuildAdd(builder, wg_offset_dw[buffer], tmp, "");
2227 offset[buffer] = LLVMBuildShl(builder, tmp, LLVMConstInt(ctx->ac.i32, 2, false), "");
2228 }
2229
2230 if (ctx->stage == MESA_SHADER_GEOMETRY) {
2231 struct radv_shader_output_values outputs[AC_LLVM_MAX_OUTPUTS];
2232 unsigned noutput = 0;
2233 unsigned out_idx = 0;
2234
2235 for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
2236 unsigned output_usage_mask =
2237 ctx->args->shader_info->gs.output_usage_mask[i];
2238 uint8_t output_stream = ctx->args->shader_info->gs.output_streams[i];
2239
2240 if (!(ctx->output_mask & (1ull << i)) ||
2241 output_stream != stream)
2242 continue;
2243
2244 outputs[noutput].slot_name = i;
2245 outputs[noutput].slot_index = i == VARYING_SLOT_CLIP_DIST1;
2246 outputs[noutput].usage_mask = output_usage_mask;
2247
2248 int length = util_last_bit(output_usage_mask);
2249
2250 for (unsigned j = 0; j < length; j++, out_idx++) {
2251 if (!(output_usage_mask & (1 << j)))
2252 continue;
2253
2254 tmp = ac_build_gep0(&ctx->ac, vertexptr,
2255 LLVMConstInt(ctx->ac.i32, out_idx, false));
2256 outputs[noutput].values[j] = LLVMBuildLoad(builder, tmp, "");
2257 }
2258
2259 for (unsigned j = length; j < 4; j++)
2260 outputs[noutput].values[j] = LLVMGetUndef(ctx->ac.f32);
2261
2262 noutput++;
2263 }
2264
2265 for (unsigned i = 0; i < noutput; i++) {
2266 struct radv_stream_output *output =
2267 radv_get_stream_output_by_loc(so, outputs[i].slot_name);
2268
2269 if (!output ||
2270 output->stream != stream)
2271 continue;
2272
2273 struct radv_shader_output_values out = {0};
2274
2275 for (unsigned j = 0; j < 4; j++) {
2276 out.values[j] = outputs[i].values[j];
2277 }
2278
2279 radv_emit_stream_output(ctx, so_buffer, offset, output, &out);
2280 }
2281 } else {
2282 for (unsigned i = 0; i < so->num_outputs; ++i) {
2283 struct radv_stream_output *output =
2284 &ctx->args->shader_info->so.outputs[i];
2285
2286 if (stream != output->stream)
2287 continue;
2288
2289 struct radv_shader_output_values out = {0};
2290
2291 for (unsigned comp = 0; comp < 4; comp++) {
2292 if (!(output->component_mask & (1 << comp)))
2293 continue;
2294
2295 tmp = ac_build_gep0(&ctx->ac, vertexptr,
2296 LLVMConstInt(ctx->ac.i32, 4 * i + comp, false));
2297 out.values[comp] = LLVMBuildLoad(builder, tmp, "");
2298 }
2299
2300 radv_emit_stream_output(ctx, so_buffer, offset, output, &out);
2301 }
2302 }
2303 }
2304
2305 struct ngg_streamout {
2306 LLVMValueRef num_vertices;
2307
2308 /* per-thread data */
2309 LLVMValueRef prim_enable[4]; /* i1 per stream */
2310 LLVMValueRef vertices[3]; /* [N x i32] addrspace(LDS)* */
2311
2312 /* Output */
2313 LLVMValueRef emit[4]; /* per-stream emitted primitives (only valid for used streams) */
2314 };
2315
2316 /**
2317 * Build streamout logic.
2318 *
2319 * Implies a barrier.
2320 *
2321 * Writes number of emitted primitives to gs_ngg_scratch[4:7].
2322 *
2323 * Clobbers gs_ngg_scratch[8:].
2324 */
build_streamout(struct radv_shader_context * ctx,struct ngg_streamout * nggso)2325 static void build_streamout(struct radv_shader_context *ctx,
2326 struct ngg_streamout *nggso)
2327 {
2328 struct radv_streamout_info *so = &ctx->args->shader_info->so;
2329 LLVMBuilderRef builder = ctx->ac.builder;
2330 LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->args->streamout_buffers);
2331 LLVMValueRef tid = get_thread_id_in_tg(ctx);
2332 LLVMValueRef cond, tmp, tmp2;
2333 LLVMValueRef i32_2 = LLVMConstInt(ctx->ac.i32, 2, false);
2334 LLVMValueRef i32_4 = LLVMConstInt(ctx->ac.i32, 4, false);
2335 LLVMValueRef i32_8 = LLVMConstInt(ctx->ac.i32, 8, false);
2336 LLVMValueRef so_buffer[4] = {0};
2337 unsigned max_num_vertices = 1 + (nggso->vertices[1] ? 1 : 0) +
2338 (nggso->vertices[2] ? 1 : 0);
2339 LLVMValueRef prim_stride_dw[4] = {0};
2340 LLVMValueRef prim_stride_dw_vgpr = LLVMGetUndef(ctx->ac.i32);
2341 int stream_for_buffer[4] = { -1, -1, -1, -1 };
2342 unsigned bufmask_for_stream[4] = {0};
2343 bool isgs = ctx->stage == MESA_SHADER_GEOMETRY;
2344 unsigned scratch_emit_base = isgs ? 4 : 0;
2345 LLVMValueRef scratch_emit_basev = isgs ? i32_4 : ctx->ac.i32_0;
2346 unsigned scratch_offset_base = isgs ? 8 : 4;
2347 LLVMValueRef scratch_offset_basev = isgs ? i32_8 : i32_4;
2348
2349 ac_llvm_add_target_dep_function_attr(ctx->main_function,
2350 "amdgpu-gds-size", 256);
2351
2352 /* Determine the mapping of streamout buffers to vertex streams. */
2353 for (unsigned i = 0; i < so->num_outputs; ++i) {
2354 unsigned buf = so->outputs[i].buffer;
2355 unsigned stream = so->outputs[i].stream;
2356 assert(stream_for_buffer[buf] < 0 || stream_for_buffer[buf] == stream);
2357 stream_for_buffer[buf] = stream;
2358 bufmask_for_stream[stream] |= 1 << buf;
2359 }
2360
2361 for (unsigned buffer = 0; buffer < 4; ++buffer) {
2362 if (stream_for_buffer[buffer] == -1)
2363 continue;
2364
2365 assert(so->strides[buffer]);
2366
2367 LLVMValueRef stride_for_buffer =
2368 LLVMConstInt(ctx->ac.i32, so->strides[buffer], false);
2369 prim_stride_dw[buffer] =
2370 LLVMBuildMul(builder, stride_for_buffer,
2371 nggso->num_vertices, "");
2372 prim_stride_dw_vgpr = ac_build_writelane(
2373 &ctx->ac, prim_stride_dw_vgpr, prim_stride_dw[buffer],
2374 LLVMConstInt(ctx->ac.i32, buffer, false));
2375
2376 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, buffer, false);
2377 so_buffer[buffer] = ac_build_load_to_sgpr(&ctx->ac, buf_ptr,
2378 offset);
2379 }
2380
2381 cond = LLVMBuildICmp(builder, LLVMIntEQ, get_wave_id_in_tg(ctx), ctx->ac.i32_0, "");
2382 ac_build_ifcc(&ctx->ac, cond, 5200);
2383 {
2384 LLVMTypeRef gdsptr = LLVMPointerType(ctx->ac.i32, AC_ADDR_SPACE_GDS);
2385 LLVMValueRef gdsbase = LLVMBuildIntToPtr(builder, ctx->ac.i32_0, gdsptr, "");
2386
2387 /* Advance the streamout offsets in GDS. */
2388 LLVMValueRef offsets_vgpr = ac_build_alloca_undef(&ctx->ac, ctx->ac.i32, "");
2389 LLVMValueRef generated_by_stream_vgpr = ac_build_alloca_undef(&ctx->ac, ctx->ac.i32, "");
2390
2391 cond = LLVMBuildICmp(builder, LLVMIntULT, ac_get_thread_id(&ctx->ac), i32_4, "");
2392 ac_build_ifcc(&ctx->ac, cond, 5210);
2393 {
2394 /* Fetch the number of generated primitives and store
2395 * it in GDS for later use.
2396 */
2397 if (isgs) {
2398 tmp = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tid);
2399 tmp = LLVMBuildLoad(builder, tmp, "");
2400 } else {
2401 tmp = ac_build_writelane(&ctx->ac, ctx->ac.i32_0,
2402 ngg_get_prim_cnt(ctx), ctx->ac.i32_0);
2403 }
2404 LLVMBuildStore(builder, tmp, generated_by_stream_vgpr);
2405
2406 unsigned swizzle[4];
2407 int unused_stream = -1;
2408 for (unsigned stream = 0; stream < 4; ++stream) {
2409 if (!ctx->args->shader_info->gs.num_stream_output_components[stream]) {
2410 unused_stream = stream;
2411 break;
2412 }
2413 }
2414 for (unsigned buffer = 0; buffer < 4; ++buffer) {
2415 if (stream_for_buffer[buffer] >= 0) {
2416 swizzle[buffer] = stream_for_buffer[buffer];
2417 } else {
2418 assert(unused_stream >= 0);
2419 swizzle[buffer] = unused_stream;
2420 }
2421 }
2422
2423 tmp = ac_build_quad_swizzle(&ctx->ac, tmp,
2424 swizzle[0], swizzle[1], swizzle[2], swizzle[3]);
2425 tmp = LLVMBuildMul(builder, tmp, prim_stride_dw_vgpr, "");
2426
2427 LLVMValueRef args[] = {
2428 LLVMBuildIntToPtr(builder, ngg_get_ordered_id(ctx), gdsptr, ""),
2429 tmp,
2430 ctx->ac.i32_0, // ordering
2431 ctx->ac.i32_0, // scope
2432 ctx->ac.i1false, // isVolatile
2433 LLVMConstInt(ctx->ac.i32, 4 << 24, false), // OA index
2434 ctx->ac.i1true, // wave release
2435 ctx->ac.i1true, // wave done
2436 };
2437
2438 tmp = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ds.ordered.add",
2439 ctx->ac.i32, args, ARRAY_SIZE(args), 0);
2440
2441 /* Keep offsets in a VGPR for quick retrieval via readlane by
2442 * the first wave for bounds checking, and also store in LDS
2443 * for retrieval by all waves later. */
2444 LLVMBuildStore(builder, tmp, offsets_vgpr);
2445
2446 tmp2 = LLVMBuildAdd(builder, ac_get_thread_id(&ctx->ac),
2447 scratch_offset_basev, "");
2448 tmp2 = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tmp2);
2449 LLVMBuildStore(builder, tmp, tmp2);
2450 }
2451 ac_build_endif(&ctx->ac, 5210);
2452
2453 /* Determine the max emit per buffer. This is done via the SALU, in part
2454 * because LLVM can't generate divide-by-multiply if we try to do this
2455 * via VALU with one lane per buffer.
2456 */
2457 LLVMValueRef max_emit[4] = {0};
2458 for (unsigned buffer = 0; buffer < 4; ++buffer) {
2459 if (stream_for_buffer[buffer] == -1)
2460 continue;
2461
2462 /* Compute the streamout buffer size in DWORD. */
2463 LLVMValueRef bufsize_dw =
2464 LLVMBuildLShr(builder,
2465 LLVMBuildExtractElement(builder, so_buffer[buffer], i32_2, ""),
2466 i32_2, "");
2467
2468 /* Load the streamout buffer offset from GDS. */
2469 tmp = LLVMBuildLoad(builder, offsets_vgpr, "");
2470 LLVMValueRef offset_dw =
2471 ac_build_readlane(&ctx->ac, tmp,
2472 LLVMConstInt(ctx->ac.i32, buffer, false));
2473
2474 /* Compute the remaining size to emit. */
2475 LLVMValueRef remaining_dw =
2476 LLVMBuildSub(builder, bufsize_dw, offset_dw, "");
2477 tmp = LLVMBuildUDiv(builder, remaining_dw,
2478 prim_stride_dw[buffer], "");
2479
2480 cond = LLVMBuildICmp(builder, LLVMIntULT,
2481 bufsize_dw, offset_dw, "");
2482 max_emit[buffer] = LLVMBuildSelect(builder, cond,
2483 ctx->ac.i32_0, tmp, "");
2484 }
2485
2486 /* Determine the number of emitted primitives per stream and fixup the
2487 * GDS counter if necessary.
2488 *
2489 * This is complicated by the fact that a single stream can emit to
2490 * multiple buffers (but luckily not vice versa).
2491 */
2492 LLVMValueRef emit_vgpr = ctx->ac.i32_0;
2493
2494 for (unsigned stream = 0; stream < 4; ++stream) {
2495 if (!ctx->args->shader_info->gs.num_stream_output_components[stream])
2496 continue;
2497
2498 /* Load the number of generated primitives from GDS and
2499 * determine that number for the given stream.
2500 */
2501 tmp = LLVMBuildLoad(builder, generated_by_stream_vgpr, "");
2502 LLVMValueRef generated =
2503 ac_build_readlane(&ctx->ac, tmp,
2504 LLVMConstInt(ctx->ac.i32, stream, false));
2505
2506
2507 /* Compute the number of emitted primitives. */
2508 LLVMValueRef emit = generated;
2509 for (unsigned buffer = 0; buffer < 4; ++buffer) {
2510 if (stream_for_buffer[buffer] == stream)
2511 emit = ac_build_umin(&ctx->ac, emit, max_emit[buffer]);
2512 }
2513
2514 /* Store the number of emitted primitives for that
2515 * stream.
2516 */
2517 emit_vgpr = ac_build_writelane(&ctx->ac, emit_vgpr, emit,
2518 LLVMConstInt(ctx->ac.i32, stream, false));
2519
2520 /* Fixup the offset using a plain GDS atomic if we overflowed. */
2521 cond = LLVMBuildICmp(builder, LLVMIntULT, emit, generated, "");
2522 ac_build_ifcc(&ctx->ac, cond, 5221); /* scalar branch */
2523 tmp = LLVMBuildLShr(builder,
2524 LLVMConstInt(ctx->ac.i32, bufmask_for_stream[stream], false),
2525 ac_get_thread_id(&ctx->ac), "");
2526 tmp = LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
2527 ac_build_ifcc(&ctx->ac, tmp, 5222);
2528 {
2529 tmp = LLVMBuildSub(builder, generated, emit, "");
2530 tmp = LLVMBuildMul(builder, tmp, prim_stride_dw_vgpr, "");
2531 tmp2 = LLVMBuildGEP(builder, gdsbase, &tid, 1, "");
2532 LLVMBuildAtomicRMW(builder, LLVMAtomicRMWBinOpSub, tmp2, tmp,
2533 LLVMAtomicOrderingMonotonic, false);
2534 }
2535 ac_build_endif(&ctx->ac, 5222);
2536 ac_build_endif(&ctx->ac, 5221);
2537 }
2538
2539 /* Store the number of emitted primitives to LDS for later use. */
2540 cond = LLVMBuildICmp(builder, LLVMIntULT, ac_get_thread_id(&ctx->ac), i32_4, "");
2541 ac_build_ifcc(&ctx->ac, cond, 5225);
2542 {
2543 tmp = LLVMBuildAdd(builder, ac_get_thread_id(&ctx->ac),
2544 scratch_emit_basev, "");
2545 tmp = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tmp);
2546 LLVMBuildStore(builder, emit_vgpr, tmp);
2547 }
2548 ac_build_endif(&ctx->ac, 5225);
2549 }
2550 ac_build_endif(&ctx->ac, 5200);
2551
2552 /* Determine the workgroup-relative per-thread / primitive offset into
2553 * the streamout buffers */
2554 struct ac_wg_scan primemit_scan[4] = {0};
2555
2556 if (isgs) {
2557 for (unsigned stream = 0; stream < 4; ++stream) {
2558 if (!ctx->args->shader_info->gs.num_stream_output_components[stream])
2559 continue;
2560
2561 primemit_scan[stream].enable_exclusive = true;
2562 primemit_scan[stream].op = nir_op_iadd;
2563 primemit_scan[stream].src = nggso->prim_enable[stream];
2564 primemit_scan[stream].scratch =
2565 ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch,
2566 LLVMConstInt(ctx->ac.i32, 12 + 8 * stream, false));
2567 primemit_scan[stream].waveidx = get_wave_id_in_tg(ctx);
2568 primemit_scan[stream].numwaves = get_tgsize(ctx);
2569 primemit_scan[stream].maxwaves = 8;
2570 ac_build_wg_scan_top(&ctx->ac, &primemit_scan[stream]);
2571 }
2572 }
2573
2574 ac_build_s_barrier(&ctx->ac);
2575
2576 /* Fetch the per-buffer offsets and per-stream emit counts in all waves. */
2577 LLVMValueRef wgoffset_dw[4] = {0};
2578
2579 {
2580 LLVMValueRef scratch_vgpr;
2581
2582 tmp = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, ac_get_thread_id(&ctx->ac));
2583 scratch_vgpr = LLVMBuildLoad(builder, tmp, "");
2584
2585 for (unsigned buffer = 0; buffer < 4; ++buffer) {
2586 if (stream_for_buffer[buffer] >= 0) {
2587 wgoffset_dw[buffer] = ac_build_readlane(
2588 &ctx->ac, scratch_vgpr,
2589 LLVMConstInt(ctx->ac.i32, scratch_offset_base + buffer, false));
2590 }
2591 }
2592
2593 for (unsigned stream = 0; stream < 4; ++stream) {
2594 if (ctx->args->shader_info->gs.num_stream_output_components[stream]) {
2595 nggso->emit[stream] = ac_build_readlane(
2596 &ctx->ac, scratch_vgpr,
2597 LLVMConstInt(ctx->ac.i32, scratch_emit_base + stream, false));
2598 }
2599 }
2600 }
2601
2602 /* Write out primitive data */
2603 for (unsigned stream = 0; stream < 4; ++stream) {
2604 if (!ctx->args->shader_info->gs.num_stream_output_components[stream])
2605 continue;
2606
2607 if (isgs) {
2608 ac_build_wg_scan_bottom(&ctx->ac, &primemit_scan[stream]);
2609 } else {
2610 primemit_scan[stream].result_exclusive = tid;
2611 }
2612
2613 cond = LLVMBuildICmp(builder, LLVMIntULT,
2614 primemit_scan[stream].result_exclusive,
2615 nggso->emit[stream], "");
2616 cond = LLVMBuildAnd(builder, cond, nggso->prim_enable[stream], "");
2617 ac_build_ifcc(&ctx->ac, cond, 5240);
2618 {
2619 LLVMValueRef offset_vtx =
2620 LLVMBuildMul(builder, primemit_scan[stream].result_exclusive,
2621 nggso->num_vertices, "");
2622
2623 for (unsigned i = 0; i < max_num_vertices; ++i) {
2624 cond = LLVMBuildICmp(builder, LLVMIntULT,
2625 LLVMConstInt(ctx->ac.i32, i, false),
2626 nggso->num_vertices, "");
2627 ac_build_ifcc(&ctx->ac, cond, 5241);
2628 build_streamout_vertex(ctx, so_buffer, wgoffset_dw,
2629 stream, offset_vtx, nggso->vertices[i]);
2630 ac_build_endif(&ctx->ac, 5241);
2631 offset_vtx = LLVMBuildAdd(builder, offset_vtx, ctx->ac.i32_1, "");
2632 }
2633 }
2634 ac_build_endif(&ctx->ac, 5240);
2635 }
2636 }
2637
ngg_nogs_vertex_size(struct radv_shader_context * ctx)2638 static unsigned ngg_nogs_vertex_size(struct radv_shader_context *ctx)
2639 {
2640 unsigned lds_vertex_size = 0;
2641
2642 if (ctx->args->shader_info->so.num_outputs)
2643 lds_vertex_size = 4 * ctx->args->shader_info->so.num_outputs + 1;
2644
2645 return lds_vertex_size;
2646 }
2647
2648 /**
2649 * Returns an `[N x i32] addrspace(LDS)*` pointing at contiguous LDS storage
2650 * for the vertex outputs.
2651 */
ngg_nogs_vertex_ptr(struct radv_shader_context * ctx,LLVMValueRef vtxid)2652 static LLVMValueRef ngg_nogs_vertex_ptr(struct radv_shader_context *ctx,
2653 LLVMValueRef vtxid)
2654 {
2655 /* The extra dword is used to avoid LDS bank conflicts. */
2656 unsigned vertex_size = ngg_nogs_vertex_size(ctx);
2657 LLVMTypeRef ai32 = LLVMArrayType(ctx->ac.i32, vertex_size);
2658 LLVMTypeRef pai32 = LLVMPointerType(ai32, AC_ADDR_SPACE_LDS);
2659 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, ctx->esgs_ring, pai32, "");
2660 return LLVMBuildGEP(ctx->ac.builder, tmp, &vtxid, 1, "");
2661 }
2662
2663 static void
handle_ngg_outputs_post_1(struct radv_shader_context * ctx)2664 handle_ngg_outputs_post_1(struct radv_shader_context *ctx)
2665 {
2666 struct radv_streamout_info *so = &ctx->args->shader_info->so;
2667 LLVMBuilderRef builder = ctx->ac.builder;
2668 LLVMValueRef vertex_ptr = NULL;
2669 LLVMValueRef tmp, tmp2;
2670
2671 assert((ctx->stage == MESA_SHADER_VERTEX ||
2672 ctx->stage == MESA_SHADER_TESS_EVAL) && !ctx->args->is_gs_copy_shader);
2673
2674 if (!ctx->args->shader_info->so.num_outputs)
2675 return;
2676
2677 vertex_ptr = ngg_nogs_vertex_ptr(ctx, get_thread_id_in_tg(ctx));
2678
2679 for (unsigned i = 0; i < so->num_outputs; ++i) {
2680 struct radv_stream_output *output =
2681 &ctx->args->shader_info->so.outputs[i];
2682
2683 unsigned loc = output->location;
2684
2685 for (unsigned comp = 0; comp < 4; comp++) {
2686 if (!(output->component_mask & (1 << comp)))
2687 continue;
2688
2689 tmp = ac_build_gep0(&ctx->ac, vertex_ptr,
2690 LLVMConstInt(ctx->ac.i32, 4 * i + comp, false));
2691 tmp2 = LLVMBuildLoad(builder,
2692 ctx->abi.outputs[4 * loc + comp], "");
2693 tmp2 = ac_to_integer(&ctx->ac, tmp2);
2694 LLVMBuildStore(builder, tmp2, tmp);
2695 }
2696 }
2697 }
2698
2699 static void
handle_ngg_outputs_post_2(struct radv_shader_context * ctx)2700 handle_ngg_outputs_post_2(struct radv_shader_context *ctx)
2701 {
2702 LLVMBuilderRef builder = ctx->ac.builder;
2703 LLVMValueRef tmp;
2704
2705 assert((ctx->stage == MESA_SHADER_VERTEX ||
2706 ctx->stage == MESA_SHADER_TESS_EVAL) && !ctx->args->is_gs_copy_shader);
2707
2708 LLVMValueRef prims_in_wave = ac_unpack_param(&ctx->ac,
2709 ac_get_arg(&ctx->ac, ctx->args->merged_wave_info), 8, 8);
2710 LLVMValueRef vtx_in_wave = ac_unpack_param(&ctx->ac,
2711 ac_get_arg(&ctx->ac, ctx->args->merged_wave_info), 0, 8);
2712 LLVMValueRef is_gs_thread = LLVMBuildICmp(builder, LLVMIntULT,
2713 ac_get_thread_id(&ctx->ac), prims_in_wave, "");
2714 LLVMValueRef is_es_thread = LLVMBuildICmp(builder, LLVMIntULT,
2715 ac_get_thread_id(&ctx->ac), vtx_in_wave, "");
2716 LLVMValueRef vtxindex[] = {
2717 ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->gs_vtx_offset[0]), 0, 16),
2718 ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->gs_vtx_offset[0]), 16, 16),
2719 ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->gs_vtx_offset[2]), 0, 16),
2720 };
2721
2722 /* Determine the number of vertices per primitive. */
2723 unsigned num_vertices;
2724 LLVMValueRef num_vertices_val;
2725
2726 if (ctx->stage == MESA_SHADER_VERTEX) {
2727 LLVMValueRef outprim_val =
2728 LLVMConstInt(ctx->ac.i32,
2729 ctx->args->options->key.vs.outprim, false);
2730 num_vertices_val = LLVMBuildAdd(builder, outprim_val,
2731 ctx->ac.i32_1, "");
2732 num_vertices = 3; /* TODO: optimize for points & lines */
2733 } else {
2734 assert(ctx->stage == MESA_SHADER_TESS_EVAL);
2735
2736 if (ctx->shader->info.tess.point_mode)
2737 num_vertices = 1;
2738 else if (ctx->shader->info.tess.primitive_mode == GL_ISOLINES)
2739 num_vertices = 2;
2740 else
2741 num_vertices = 3;
2742
2743 num_vertices_val = LLVMConstInt(ctx->ac.i32, num_vertices, false);
2744 }
2745
2746 /* Streamout */
2747 if (ctx->args->shader_info->so.num_outputs) {
2748 struct ngg_streamout nggso = {0};
2749
2750 nggso.num_vertices = num_vertices_val;
2751 nggso.prim_enable[0] = is_gs_thread;
2752
2753 for (unsigned i = 0; i < num_vertices; ++i)
2754 nggso.vertices[i] = ngg_nogs_vertex_ptr(ctx, vtxindex[i]);
2755
2756 build_streamout(ctx, &nggso);
2757 }
2758
2759 /* Copy Primitive IDs from GS threads to the LDS address corresponding
2760 * to the ES thread of the provoking vertex.
2761 */
2762 if (ctx->stage == MESA_SHADER_VERTEX &&
2763 ctx->args->options->key.vs_common_out.export_prim_id) {
2764 if (ctx->args->shader_info->so.num_outputs)
2765 ac_build_s_barrier(&ctx->ac);
2766
2767 ac_build_ifcc(&ctx->ac, is_gs_thread, 5400);
2768 /* Extract the PROVOKING_VTX_INDEX field. */
2769 LLVMValueRef provoking_vtx_in_prim =
2770 LLVMConstInt(ctx->ac.i32, 0, false);
2771
2772 /* provoking_vtx_index = vtxindex[provoking_vtx_in_prim]; */
2773 LLVMValueRef indices = ac_build_gather_values(&ctx->ac, vtxindex, 3);
2774 LLVMValueRef provoking_vtx_index =
2775 LLVMBuildExtractElement(builder, indices, provoking_vtx_in_prim, "");
2776
2777 LLVMBuildStore(builder, ac_get_arg(&ctx->ac, ctx->args->ac.gs_prim_id),
2778 ac_build_gep0(&ctx->ac, ctx->esgs_ring, provoking_vtx_index));
2779 ac_build_endif(&ctx->ac, 5400);
2780 }
2781
2782 /* TODO: primitive culling */
2783
2784 ac_build_sendmsg_gs_alloc_req(&ctx->ac, get_wave_id_in_tg(ctx),
2785 ngg_get_vtx_cnt(ctx), ngg_get_prim_cnt(ctx));
2786
2787 /* TODO: streamout queries */
2788 /* Export primitive data to the index buffer.
2789 *
2790 * For the first version, we will always build up all three indices
2791 * independent of the primitive type. The additional garbage data
2792 * shouldn't hurt.
2793 *
2794 * TODO: culling depends on the primitive type, so can have some
2795 * interaction here.
2796 */
2797 ac_build_ifcc(&ctx->ac, is_gs_thread, 6001);
2798 {
2799 struct ac_ngg_prim prim = {0};
2800
2801 if (ctx->args->options->key.vs_common_out.as_ngg_passthrough) {
2802 prim.passthrough = ac_get_arg(&ctx->ac, ctx->args->gs_vtx_offset[0]);
2803 } else {
2804 prim.num_vertices = num_vertices;
2805 prim.isnull = ctx->ac.i1false;
2806 memcpy(prim.index, vtxindex, sizeof(vtxindex[0]) * 3);
2807
2808 for (unsigned i = 0; i < num_vertices; ++i) {
2809 tmp = LLVMBuildLShr(builder,
2810 ac_get_arg(&ctx->ac, ctx->args->ac.gs_invocation_id),
2811 LLVMConstInt(ctx->ac.i32, 8 + i, false), "");
2812 prim.edgeflag[i] = LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
2813 }
2814 }
2815
2816 ac_build_export_prim(&ctx->ac, &prim);
2817 }
2818 ac_build_endif(&ctx->ac, 6001);
2819
2820 /* Export per-vertex data (positions and parameters). */
2821 ac_build_ifcc(&ctx->ac, is_es_thread, 6002);
2822 {
2823 struct radv_vs_output_info *outinfo =
2824 ctx->stage == MESA_SHADER_TESS_EVAL ?
2825 &ctx->args->shader_info->tes.outinfo : &ctx->args->shader_info->vs.outinfo;
2826
2827 /* Exporting the primitive ID is handled below. */
2828 /* TODO: use the new VS export path */
2829 handle_vs_outputs_post(ctx, false,
2830 ctx->args->options->key.vs_common_out.export_clip_dists,
2831 outinfo);
2832
2833 if (ctx->args->options->key.vs_common_out.export_prim_id) {
2834 unsigned param_count = outinfo->param_exports;
2835 LLVMValueRef values[4];
2836
2837 if (ctx->stage == MESA_SHADER_VERTEX) {
2838 /* Wait for GS stores to finish. */
2839 ac_build_s_barrier(&ctx->ac);
2840
2841 tmp = ac_build_gep0(&ctx->ac, ctx->esgs_ring,
2842 get_thread_id_in_tg(ctx));
2843 values[0] = LLVMBuildLoad(builder, tmp, "");
2844 } else {
2845 assert(ctx->stage == MESA_SHADER_TESS_EVAL);
2846 values[0] = ac_get_arg(&ctx->ac, ctx->args->ac.tes_patch_id);
2847 }
2848
2849 values[0] = ac_to_float(&ctx->ac, values[0]);
2850 for (unsigned j = 1; j < 4; j++)
2851 values[j] = ctx->ac.f32_0;
2852
2853 radv_export_param(ctx, param_count, values, 0x1);
2854
2855 outinfo->vs_output_param_offset[VARYING_SLOT_PRIMITIVE_ID] = param_count++;
2856 outinfo->param_exports = param_count;
2857 }
2858 }
2859 ac_build_endif(&ctx->ac, 6002);
2860 }
2861
gfx10_ngg_gs_emit_prologue(struct radv_shader_context * ctx)2862 static void gfx10_ngg_gs_emit_prologue(struct radv_shader_context *ctx)
2863 {
2864 /* Zero out the part of LDS scratch that is used to accumulate the
2865 * per-stream generated primitive count.
2866 */
2867 LLVMBuilderRef builder = ctx->ac.builder;
2868 LLVMValueRef scratchptr = ctx->gs_ngg_scratch;
2869 LLVMValueRef tid = get_thread_id_in_tg(ctx);
2870 LLVMBasicBlockRef merge_block;
2871 LLVMValueRef cond;
2872
2873 LLVMValueRef fn = LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx->ac.builder));
2874 LLVMBasicBlockRef then_block = LLVMAppendBasicBlockInContext(ctx->ac.context, fn, "");
2875 merge_block = LLVMAppendBasicBlockInContext(ctx->ac.context, fn, "");
2876
2877 cond = LLVMBuildICmp(builder, LLVMIntULT, tid, LLVMConstInt(ctx->ac.i32, 4, false), "");
2878 LLVMBuildCondBr(ctx->ac.builder, cond, then_block, merge_block);
2879 LLVMPositionBuilderAtEnd(ctx->ac.builder, then_block);
2880
2881 LLVMValueRef ptr = ac_build_gep0(&ctx->ac, scratchptr, tid);
2882 LLVMBuildStore(builder, ctx->ac.i32_0, ptr);
2883
2884 LLVMBuildBr(ctx->ac.builder, merge_block);
2885 LLVMPositionBuilderAtEnd(ctx->ac.builder, merge_block);
2886
2887 ac_build_s_barrier(&ctx->ac);
2888 }
2889
gfx10_ngg_gs_emit_epilogue_1(struct radv_shader_context * ctx)2890 static void gfx10_ngg_gs_emit_epilogue_1(struct radv_shader_context *ctx)
2891 {
2892 LLVMBuilderRef builder = ctx->ac.builder;
2893 LLVMValueRef i8_0 = LLVMConstInt(ctx->ac.i8, 0, false);
2894 LLVMValueRef tmp;
2895
2896 /* Zero out remaining (non-emitted) primitive flags.
2897 *
2898 * Note: Alternatively, we could pass the relevant gs_next_vertex to
2899 * the emit threads via LDS. This is likely worse in the expected
2900 * typical case where each GS thread emits the full set of
2901 * vertices.
2902 */
2903 for (unsigned stream = 0; stream < 4; ++stream) {
2904 unsigned num_components;
2905
2906 num_components =
2907 ctx->args->shader_info->gs.num_stream_output_components[stream];
2908 if (!num_components)
2909 continue;
2910
2911 const LLVMValueRef gsthread = get_thread_id_in_tg(ctx);
2912
2913 ac_build_bgnloop(&ctx->ac, 5100);
2914
2915 const LLVMValueRef vertexidx =
2916 LLVMBuildLoad(builder, ctx->gs_next_vertex[stream], "");
2917 tmp = LLVMBuildICmp(builder, LLVMIntUGE, vertexidx,
2918 LLVMConstInt(ctx->ac.i32, ctx->shader->info.gs.vertices_out, false), "");
2919 ac_build_ifcc(&ctx->ac, tmp, 5101);
2920 ac_build_break(&ctx->ac);
2921 ac_build_endif(&ctx->ac, 5101);
2922
2923 tmp = LLVMBuildAdd(builder, vertexidx, ctx->ac.i32_1, "");
2924 LLVMBuildStore(builder, tmp, ctx->gs_next_vertex[stream]);
2925
2926 tmp = ngg_gs_emit_vertex_ptr(ctx, gsthread, vertexidx);
2927 LLVMBuildStore(builder, i8_0,
2928 ngg_gs_get_emit_primflag_ptr(ctx, tmp, stream));
2929
2930 ac_build_endloop(&ctx->ac, 5100);
2931 }
2932
2933 /* Accumulate generated primitives counts across the entire threadgroup. */
2934 for (unsigned stream = 0; stream < 4; ++stream) {
2935 unsigned num_components;
2936
2937 num_components =
2938 ctx->args->shader_info->gs.num_stream_output_components[stream];
2939 if (!num_components)
2940 continue;
2941
2942 LLVMValueRef numprims =
2943 LLVMBuildLoad(builder, ctx->gs_generated_prims[stream], "");
2944 numprims = ac_build_reduce(&ctx->ac, numprims, nir_op_iadd, ctx->ac.wave_size);
2945
2946 tmp = LLVMBuildICmp(builder, LLVMIntEQ, ac_get_thread_id(&ctx->ac), ctx->ac.i32_0, "");
2947 ac_build_ifcc(&ctx->ac, tmp, 5105);
2948 {
2949 LLVMBuildAtomicRMW(builder, LLVMAtomicRMWBinOpAdd,
2950 ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch,
2951 LLVMConstInt(ctx->ac.i32, stream, false)),
2952 numprims, LLVMAtomicOrderingMonotonic, false);
2953 }
2954 ac_build_endif(&ctx->ac, 5105);
2955 }
2956 }
2957
gfx10_ngg_gs_emit_epilogue_2(struct radv_shader_context * ctx)2958 static void gfx10_ngg_gs_emit_epilogue_2(struct radv_shader_context *ctx)
2959 {
2960 const unsigned verts_per_prim = si_conv_gl_prim_to_vertices(ctx->shader->info.gs.output_primitive);
2961 LLVMBuilderRef builder = ctx->ac.builder;
2962 LLVMValueRef tmp, tmp2;
2963
2964 ac_build_s_barrier(&ctx->ac);
2965
2966 const LLVMValueRef tid = get_thread_id_in_tg(ctx);
2967 LLVMValueRef num_emit_threads = ngg_get_prim_cnt(ctx);
2968
2969 /* Streamout */
2970 if (ctx->args->shader_info->so.num_outputs) {
2971 struct ngg_streamout nggso = {0};
2972
2973 nggso.num_vertices = LLVMConstInt(ctx->ac.i32, verts_per_prim, false);
2974
2975 LLVMValueRef vertexptr = ngg_gs_vertex_ptr(ctx, tid);
2976 for (unsigned stream = 0; stream < 4; ++stream) {
2977 if (!ctx->args->shader_info->gs.num_stream_output_components[stream])
2978 continue;
2979
2980 tmp = LLVMBuildLoad(builder,
2981 ngg_gs_get_emit_primflag_ptr(ctx, vertexptr, stream), "");
2982 tmp = LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
2983 tmp2 = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
2984 nggso.prim_enable[stream] = LLVMBuildAnd(builder, tmp, tmp2, "");
2985 }
2986
2987 for (unsigned i = 0; i < verts_per_prim; ++i) {
2988 tmp = LLVMBuildSub(builder, tid,
2989 LLVMConstInt(ctx->ac.i32, verts_per_prim - i - 1, false), "");
2990 tmp = ngg_gs_vertex_ptr(ctx, tmp);
2991 nggso.vertices[i] = ac_build_gep0(&ctx->ac, tmp, ctx->ac.i32_0);
2992 }
2993
2994 build_streamout(ctx, &nggso);
2995 }
2996
2997 /* Write shader query data. */
2998 tmp = ac_get_arg(&ctx->ac, ctx->args->ngg_gs_state);
2999 tmp = LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
3000 ac_build_ifcc(&ctx->ac, tmp, 5109);
3001 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid,
3002 LLVMConstInt(ctx->ac.i32, 4, false), "");
3003 ac_build_ifcc(&ctx->ac, tmp, 5110);
3004 {
3005 tmp = LLVMBuildLoad(builder, ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tid), "");
3006
3007 ac_llvm_add_target_dep_function_attr(ctx->main_function,
3008 "amdgpu-gds-size", 256);
3009
3010 LLVMTypeRef gdsptr = LLVMPointerType(ctx->ac.i32, AC_ADDR_SPACE_GDS);
3011 LLVMValueRef gdsbase = LLVMBuildIntToPtr(builder, ctx->ac.i32_0, gdsptr, "");
3012
3013 const char *sync_scope = LLVM_VERSION_MAJOR >= 9 ? "workgroup-one-as" : "workgroup";
3014
3015 /* Use a plain GDS atomic to accumulate the number of generated
3016 * primitives.
3017 */
3018 ac_build_atomic_rmw(&ctx->ac, LLVMAtomicRMWBinOpAdd, gdsbase,
3019 tmp, sync_scope);
3020 }
3021 ac_build_endif(&ctx->ac, 5110);
3022 ac_build_endif(&ctx->ac, 5109);
3023
3024 /* TODO: culling */
3025
3026 /* Determine vertex liveness. */
3027 LLVMValueRef vertliveptr = ac_build_alloca(&ctx->ac, ctx->ac.i1, "vertexlive");
3028
3029 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
3030 ac_build_ifcc(&ctx->ac, tmp, 5120);
3031 {
3032 for (unsigned i = 0; i < verts_per_prim; ++i) {
3033 const LLVMValueRef primidx =
3034 LLVMBuildAdd(builder, tid,
3035 LLVMConstInt(ctx->ac.i32, i, false), "");
3036
3037 if (i > 0) {
3038 tmp = LLVMBuildICmp(builder, LLVMIntULT, primidx, num_emit_threads, "");
3039 ac_build_ifcc(&ctx->ac, tmp, 5121 + i);
3040 }
3041
3042 /* Load primitive liveness */
3043 tmp = ngg_gs_vertex_ptr(ctx, primidx);
3044 tmp = LLVMBuildLoad(builder,
3045 ngg_gs_get_emit_primflag_ptr(ctx, tmp, 0), "");
3046 const LLVMValueRef primlive =
3047 LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
3048
3049 tmp = LLVMBuildLoad(builder, vertliveptr, "");
3050 tmp = LLVMBuildOr(builder, tmp, primlive, ""),
3051 LLVMBuildStore(builder, tmp, vertliveptr);
3052
3053 if (i > 0)
3054 ac_build_endif(&ctx->ac, 5121 + i);
3055 }
3056 }
3057 ac_build_endif(&ctx->ac, 5120);
3058
3059 /* Inclusive scan addition across the current wave. */
3060 LLVMValueRef vertlive = LLVMBuildLoad(builder, vertliveptr, "");
3061 struct ac_wg_scan vertlive_scan = {0};
3062 vertlive_scan.op = nir_op_iadd;
3063 vertlive_scan.enable_reduce = true;
3064 vertlive_scan.enable_exclusive = true;
3065 vertlive_scan.src = vertlive;
3066 vertlive_scan.scratch = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, ctx->ac.i32_0);
3067 vertlive_scan.waveidx = get_wave_id_in_tg(ctx);
3068 vertlive_scan.numwaves = get_tgsize(ctx);
3069 vertlive_scan.maxwaves = 8;
3070
3071 ac_build_wg_scan(&ctx->ac, &vertlive_scan);
3072
3073 /* Skip all exports (including index exports) when possible. At least on
3074 * early gfx10 revisions this is also to avoid hangs.
3075 */
3076 LLVMValueRef have_exports =
3077 LLVMBuildICmp(builder, LLVMIntNE, vertlive_scan.result_reduce, ctx->ac.i32_0, "");
3078 num_emit_threads =
3079 LLVMBuildSelect(builder, have_exports, num_emit_threads, ctx->ac.i32_0, "");
3080
3081 /* Allocate export space. Send this message as early as possible, to
3082 * hide the latency of the SQ <-> SPI roundtrip.
3083 *
3084 * Note: We could consider compacting primitives for export as well.
3085 * PA processes 1 non-null prim / clock, but it fetches 4 DW of
3086 * prim data per clock and skips null primitives at no additional
3087 * cost. So compacting primitives can only be beneficial when
3088 * there are 4 or more contiguous null primitives in the export
3089 * (in the common case of single-dword prim exports).
3090 */
3091 ac_build_sendmsg_gs_alloc_req(&ctx->ac, get_wave_id_in_tg(ctx),
3092 vertlive_scan.result_reduce, num_emit_threads);
3093
3094 /* Setup the reverse vertex compaction permutation. We re-use stream 1
3095 * of the primitive liveness flags, relying on the fact that each
3096 * threadgroup can have at most 256 threads. */
3097 ac_build_ifcc(&ctx->ac, vertlive, 5130);
3098 {
3099 tmp = ngg_gs_vertex_ptr(ctx, vertlive_scan.result_exclusive);
3100 tmp2 = LLVMBuildTrunc(builder, tid, ctx->ac.i8, "");
3101 LLVMBuildStore(builder, tmp2,
3102 ngg_gs_get_emit_primflag_ptr(ctx, tmp, 1));
3103 }
3104 ac_build_endif(&ctx->ac, 5130);
3105
3106 ac_build_s_barrier(&ctx->ac);
3107
3108 /* Export primitive data */
3109 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
3110 ac_build_ifcc(&ctx->ac, tmp, 5140);
3111 {
3112 LLVMValueRef flags;
3113 struct ac_ngg_prim prim = {0};
3114 prim.num_vertices = verts_per_prim;
3115
3116 tmp = ngg_gs_vertex_ptr(ctx, tid);
3117 flags = LLVMBuildLoad(builder,
3118 ngg_gs_get_emit_primflag_ptr(ctx, tmp, 0), "");
3119 prim.isnull = LLVMBuildNot(builder, LLVMBuildTrunc(builder, flags, ctx->ac.i1, ""), "");
3120
3121 for (unsigned i = 0; i < verts_per_prim; ++i) {
3122 prim.index[i] = LLVMBuildSub(builder, vertlive_scan.result_exclusive,
3123 LLVMConstInt(ctx->ac.i32, verts_per_prim - i - 1, false), "");
3124 prim.edgeflag[i] = ctx->ac.i1false;
3125 }
3126
3127 /* Geometry shaders output triangle strips, but NGG expects
3128 * triangles. We need to change the vertex order for odd
3129 * triangles to get correct front/back facing by swapping 2
3130 * vertex indices, but we also have to keep the provoking
3131 * vertex in the same place.
3132 */
3133 if (verts_per_prim == 3) {
3134 LLVMValueRef is_odd = LLVMBuildLShr(builder, flags, ctx->ac.i8_1, "");
3135 is_odd = LLVMBuildTrunc(builder, is_odd, ctx->ac.i1, "");
3136
3137 struct ac_ngg_prim in = prim;
3138 prim.index[0] = in.index[0];
3139 prim.index[1] = LLVMBuildSelect(builder, is_odd,
3140 in.index[2], in.index[1], "");
3141 prim.index[2] = LLVMBuildSelect(builder, is_odd,
3142 in.index[1], in.index[2], "");
3143 }
3144
3145 ac_build_export_prim(&ctx->ac, &prim);
3146 }
3147 ac_build_endif(&ctx->ac, 5140);
3148
3149 /* Export position and parameter data */
3150 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, vertlive_scan.result_reduce, "");
3151 ac_build_ifcc(&ctx->ac, tmp, 5145);
3152 {
3153 struct radv_vs_output_info *outinfo = &ctx->args->shader_info->vs.outinfo;
3154 bool export_view_index = ctx->args->options->key.has_multiview_view_index;
3155 struct radv_shader_output_values *outputs;
3156 unsigned noutput = 0;
3157
3158 /* Allocate a temporary array for the output values. */
3159 unsigned num_outputs = util_bitcount64(ctx->output_mask) + export_view_index;
3160 outputs = calloc(num_outputs, sizeof(outputs[0]));
3161
3162 memset(outinfo->vs_output_param_offset, AC_EXP_PARAM_UNDEFINED,
3163 sizeof(outinfo->vs_output_param_offset));
3164 outinfo->pos_exports = 0;
3165
3166 tmp = ngg_gs_vertex_ptr(ctx, tid);
3167 tmp = LLVMBuildLoad(builder,
3168 ngg_gs_get_emit_primflag_ptr(ctx, tmp, 1), "");
3169 tmp = LLVMBuildZExt(builder, tmp, ctx->ac.i32, "");
3170 const LLVMValueRef vertexptr = ngg_gs_vertex_ptr(ctx, tmp);
3171
3172 unsigned out_idx = 0;
3173 for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
3174 unsigned output_usage_mask =
3175 ctx->args->shader_info->gs.output_usage_mask[i];
3176 int length = util_last_bit(output_usage_mask);
3177
3178 if (!(ctx->output_mask & (1ull << i)))
3179 continue;
3180
3181 outputs[noutput].slot_name = i;
3182 outputs[noutput].slot_index = i == VARYING_SLOT_CLIP_DIST1;
3183 outputs[noutput].usage_mask = output_usage_mask;
3184
3185 for (unsigned j = 0; j < length; j++, out_idx++) {
3186 if (!(output_usage_mask & (1 << j)))
3187 continue;
3188
3189 tmp = ngg_gs_get_emit_output_ptr(ctx, vertexptr, out_idx);
3190 tmp = LLVMBuildLoad(builder, tmp, "");
3191
3192 LLVMTypeRef type = LLVMGetAllocatedType(ctx->abi.outputs[ac_llvm_reg_index_soa(i, j)]);
3193 if (ac_get_type_size(type) == 2) {
3194 tmp = ac_to_integer(&ctx->ac, tmp);
3195 tmp = LLVMBuildTrunc(ctx->ac.builder, tmp, ctx->ac.i16, "");
3196 }
3197
3198 outputs[noutput].values[j] = ac_to_float(&ctx->ac, tmp);
3199 }
3200
3201 for (unsigned j = length; j < 4; j++)
3202 outputs[noutput].values[j] = LLVMGetUndef(ctx->ac.f32);
3203
3204 noutput++;
3205 }
3206
3207 /* Export ViewIndex. */
3208 if (export_view_index) {
3209 outputs[noutput].slot_name = VARYING_SLOT_LAYER;
3210 outputs[noutput].slot_index = 0;
3211 outputs[noutput].usage_mask = 0x1;
3212 outputs[noutput].values[0] =
3213 ac_to_float(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->ac.view_index));
3214 for (unsigned j = 1; j < 4; j++)
3215 outputs[noutput].values[j] = ctx->ac.f32_0;
3216 noutput++;
3217 }
3218
3219 radv_llvm_export_vs(ctx, outputs, noutput, outinfo,
3220 ctx->args->options->key.vs_common_out.export_clip_dists);
3221 FREE(outputs);
3222 }
3223 ac_build_endif(&ctx->ac, 5145);
3224 }
3225
gfx10_ngg_gs_emit_vertex(struct radv_shader_context * ctx,unsigned stream,LLVMValueRef vertexidx,LLVMValueRef * addrs)3226 static void gfx10_ngg_gs_emit_vertex(struct radv_shader_context *ctx,
3227 unsigned stream,
3228 LLVMValueRef vertexidx,
3229 LLVMValueRef *addrs)
3230 {
3231 LLVMBuilderRef builder = ctx->ac.builder;
3232 LLVMValueRef tmp;
3233
3234 const LLVMValueRef vertexptr =
3235 ngg_gs_emit_vertex_ptr(ctx, get_thread_id_in_tg(ctx), vertexidx);
3236 unsigned out_idx = 0;
3237 for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
3238 unsigned output_usage_mask =
3239 ctx->args->shader_info->gs.output_usage_mask[i];
3240 uint8_t output_stream =
3241 ctx->args->shader_info->gs.output_streams[i];
3242 LLVMValueRef *out_ptr = &addrs[i * 4];
3243 int length = util_last_bit(output_usage_mask);
3244
3245 if (!(ctx->output_mask & (1ull << i)) ||
3246 output_stream != stream)
3247 continue;
3248
3249 for (unsigned j = 0; j < length; j++, out_idx++) {
3250 if (!(output_usage_mask & (1 << j)))
3251 continue;
3252
3253 LLVMValueRef out_val = LLVMBuildLoad(ctx->ac.builder,
3254 out_ptr[j], "");
3255 out_val = ac_to_integer(&ctx->ac, out_val);
3256 out_val = LLVMBuildZExtOrBitCast(ctx->ac.builder, out_val, ctx->ac.i32, "");
3257
3258 LLVMBuildStore(builder, out_val,
3259 ngg_gs_get_emit_output_ptr(ctx, vertexptr, out_idx));
3260 }
3261 }
3262 assert(out_idx * 4 <= ctx->args->shader_info->gs.gsvs_vertex_size);
3263
3264 /* Store the current number of emitted vertices to zero out remaining
3265 * primitive flags in case the geometry shader doesn't emit the maximum
3266 * number of vertices.
3267 */
3268 tmp = LLVMBuildAdd(builder, vertexidx, ctx->ac.i32_1, "");
3269 LLVMBuildStore(builder, tmp, ctx->gs_next_vertex[stream]);
3270
3271 /* Determine and store whether this vertex completed a primitive. */
3272 const LLVMValueRef curverts = LLVMBuildLoad(builder, ctx->gs_curprim_verts[stream], "");
3273
3274 tmp = LLVMConstInt(ctx->ac.i32, si_conv_gl_prim_to_vertices(ctx->shader->info.gs.output_primitive) - 1, false);
3275 const LLVMValueRef iscompleteprim =
3276 LLVMBuildICmp(builder, LLVMIntUGE, curverts, tmp, "");
3277
3278 /* Since the geometry shader emits triangle strips, we need to
3279 * track which primitive is odd and swap vertex indices to get
3280 * the correct vertex order.
3281 */
3282 LLVMValueRef is_odd = ctx->ac.i1false;
3283 if (stream == 0 &&
3284 si_conv_gl_prim_to_vertices(ctx->shader->info.gs.output_primitive) == 3) {
3285 tmp = LLVMBuildAnd(builder, curverts, ctx->ac.i32_1, "");
3286 is_odd = LLVMBuildICmp(builder, LLVMIntEQ, tmp, ctx->ac.i32_1, "");
3287 }
3288
3289 tmp = LLVMBuildAdd(builder, curverts, ctx->ac.i32_1, "");
3290 LLVMBuildStore(builder, tmp, ctx->gs_curprim_verts[stream]);
3291
3292 /* The per-vertex primitive flag encoding:
3293 * bit 0: whether this vertex finishes a primitive
3294 * bit 1: whether the primitive is odd (if we are emitting triangle strips)
3295 */
3296 tmp = LLVMBuildZExt(builder, iscompleteprim, ctx->ac.i8, "");
3297 tmp = LLVMBuildOr(builder, tmp,
3298 LLVMBuildShl(builder,
3299 LLVMBuildZExt(builder, is_odd, ctx->ac.i8, ""),
3300 ctx->ac.i8_1, ""), "");
3301 LLVMBuildStore(builder, tmp,
3302 ngg_gs_get_emit_primflag_ptr(ctx, vertexptr, stream));
3303
3304 tmp = LLVMBuildLoad(builder, ctx->gs_generated_prims[stream], "");
3305 tmp = LLVMBuildAdd(builder, tmp, LLVMBuildZExt(builder, iscompleteprim, ctx->ac.i32, ""), "");
3306 LLVMBuildStore(builder, tmp, ctx->gs_generated_prims[stream]);
3307 }
3308
3309 static void
write_tess_factors(struct radv_shader_context * ctx)3310 write_tess_factors(struct radv_shader_context *ctx)
3311 {
3312 unsigned stride, outer_comps, inner_comps;
3313 LLVMValueRef tcs_rel_ids = ac_get_arg(&ctx->ac, ctx->args->ac.tcs_rel_ids);
3314 LLVMValueRef invocation_id = ac_unpack_param(&ctx->ac, tcs_rel_ids, 8, 5);
3315 LLVMValueRef rel_patch_id = ac_unpack_param(&ctx->ac, tcs_rel_ids, 0, 8);
3316 LLVMValueRef lds_base, lds_inner = NULL, lds_outer, byteoffset, buffer;
3317 LLVMValueRef out[6], vec0, vec1, tf_base, inner[4], outer[4];
3318 int i;
3319 ac_emit_barrier(&ctx->ac, ctx->stage);
3320
3321 switch (ctx->args->options->key.tcs.primitive_mode) {
3322 case GL_ISOLINES:
3323 stride = 2;
3324 outer_comps = 2;
3325 inner_comps = 0;
3326 break;
3327 case GL_TRIANGLES:
3328 stride = 4;
3329 outer_comps = 3;
3330 inner_comps = 1;
3331 break;
3332 case GL_QUADS:
3333 stride = 6;
3334 outer_comps = 4;
3335 inner_comps = 2;
3336 break;
3337 default:
3338 return;
3339 }
3340
3341 ac_build_ifcc(&ctx->ac,
3342 LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
3343 invocation_id, ctx->ac.i32_0, ""), 6503);
3344
3345 lds_base = get_tcs_out_current_patch_data_offset(ctx);
3346
3347 if (inner_comps) {
3348 lds_inner = LLVMBuildAdd(ctx->ac.builder, lds_base,
3349 LLVMConstInt(ctx->ac.i32, ctx->tcs_tess_lvl_inner * 4, false), "");
3350 }
3351
3352 lds_outer = LLVMBuildAdd(ctx->ac.builder, lds_base,
3353 LLVMConstInt(ctx->ac.i32, ctx->tcs_tess_lvl_outer * 4, false), "");
3354
3355 for (i = 0; i < 4; i++) {
3356 inner[i] = LLVMGetUndef(ctx->ac.i32);
3357 outer[i] = LLVMGetUndef(ctx->ac.i32);
3358 }
3359
3360 // LINES reversal
3361 if (ctx->args->options->key.tcs.primitive_mode == GL_ISOLINES) {
3362 outer[0] = out[1] = ac_lds_load(&ctx->ac, lds_outer);
3363 lds_outer = LLVMBuildAdd(ctx->ac.builder, lds_outer,
3364 ctx->ac.i32_1, "");
3365 outer[1] = out[0] = ac_lds_load(&ctx->ac, lds_outer);
3366 } else {
3367 for (i = 0; i < outer_comps; i++) {
3368 outer[i] = out[i] =
3369 ac_lds_load(&ctx->ac, lds_outer);
3370 lds_outer = LLVMBuildAdd(ctx->ac.builder, lds_outer,
3371 ctx->ac.i32_1, "");
3372 }
3373 for (i = 0; i < inner_comps; i++) {
3374 inner[i] = out[outer_comps+i] =
3375 ac_lds_load(&ctx->ac, lds_inner);
3376 lds_inner = LLVMBuildAdd(ctx->ac.builder, lds_inner,
3377 ctx->ac.i32_1, "");
3378 }
3379 }
3380
3381 /* Convert the outputs to vectors for stores. */
3382 vec0 = ac_build_gather_values(&ctx->ac, out, MIN2(stride, 4));
3383 vec1 = NULL;
3384
3385 if (stride > 4)
3386 vec1 = ac_build_gather_values(&ctx->ac, out + 4, stride - 4);
3387
3388
3389 buffer = ctx->hs_ring_tess_factor;
3390 tf_base = ac_get_arg(&ctx->ac, ctx->args->tess_factor_offset);
3391 byteoffset = LLVMBuildMul(ctx->ac.builder, rel_patch_id,
3392 LLVMConstInt(ctx->ac.i32, 4 * stride, false), "");
3393 unsigned tf_offset = 0;
3394
3395 if (ctx->ac.chip_class <= GFX8) {
3396 ac_build_ifcc(&ctx->ac,
3397 LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
3398 rel_patch_id, ctx->ac.i32_0, ""), 6504);
3399
3400 /* Store the dynamic HS control word. */
3401 ac_build_buffer_store_dword(&ctx->ac, buffer,
3402 LLVMConstInt(ctx->ac.i32, 0x80000000, false),
3403 1, ctx->ac.i32_0, tf_base,
3404 0, ac_glc);
3405 tf_offset += 4;
3406
3407 ac_build_endif(&ctx->ac, 6504);
3408 }
3409
3410 /* Store the tessellation factors. */
3411 ac_build_buffer_store_dword(&ctx->ac, buffer, vec0,
3412 MIN2(stride, 4), byteoffset, tf_base,
3413 tf_offset, ac_glc);
3414 if (vec1)
3415 ac_build_buffer_store_dword(&ctx->ac, buffer, vec1,
3416 stride - 4, byteoffset, tf_base,
3417 16 + tf_offset, ac_glc);
3418
3419 //store to offchip for TES to read - only if TES reads them
3420 if (ctx->args->options->key.tcs.tes_reads_tess_factors) {
3421 LLVMValueRef inner_vec, outer_vec, tf_outer_offset;
3422 LLVMValueRef tf_inner_offset;
3423
3424 tf_outer_offset = get_tcs_tes_buffer_address(ctx, NULL,
3425 LLVMConstInt(ctx->ac.i32, ctx->tcs_tess_lvl_outer, 0));
3426
3427 outer_vec = ac_build_gather_values(&ctx->ac, outer,
3428 util_next_power_of_two(outer_comps));
3429
3430 ac_build_buffer_store_dword(&ctx->ac, ctx->hs_ring_tess_offchip, outer_vec,
3431 outer_comps, tf_outer_offset,
3432 ac_get_arg(&ctx->ac, ctx->args->oc_lds),
3433 0, ac_glc);
3434 if (inner_comps) {
3435 tf_inner_offset = get_tcs_tes_buffer_address(ctx, NULL,
3436 LLVMConstInt(ctx->ac.i32, ctx->tcs_tess_lvl_inner, 0));
3437
3438 inner_vec = inner_comps == 1 ? inner[0] :
3439 ac_build_gather_values(&ctx->ac, inner, inner_comps);
3440 ac_build_buffer_store_dword(&ctx->ac, ctx->hs_ring_tess_offchip, inner_vec,
3441 inner_comps, tf_inner_offset,
3442 ac_get_arg(&ctx->ac, ctx->args->oc_lds),
3443 0, ac_glc);
3444 }
3445 }
3446
3447 ac_build_endif(&ctx->ac, 6503);
3448 }
3449
3450 static void
handle_tcs_outputs_post(struct radv_shader_context * ctx)3451 handle_tcs_outputs_post(struct radv_shader_context *ctx)
3452 {
3453 write_tess_factors(ctx);
3454 }
3455
3456 static bool
si_export_mrt_color(struct radv_shader_context * ctx,LLVMValueRef * color,unsigned index,struct ac_export_args * args)3457 si_export_mrt_color(struct radv_shader_context *ctx,
3458 LLVMValueRef *color, unsigned index,
3459 struct ac_export_args *args)
3460 {
3461 /* Export */
3462 si_llvm_init_export_args(ctx, color, 0xf,
3463 V_008DFC_SQ_EXP_MRT + index, args);
3464 if (!args->enabled_channels)
3465 return false; /* unnecessary NULL export */
3466
3467 return true;
3468 }
3469
3470 static void
radv_export_mrt_z(struct radv_shader_context * ctx,LLVMValueRef depth,LLVMValueRef stencil,LLVMValueRef samplemask)3471 radv_export_mrt_z(struct radv_shader_context *ctx,
3472 LLVMValueRef depth, LLVMValueRef stencil,
3473 LLVMValueRef samplemask)
3474 {
3475 struct ac_export_args args;
3476
3477 ac_export_mrt_z(&ctx->ac, depth, stencil, samplemask, &args);
3478
3479 ac_build_export(&ctx->ac, &args);
3480 }
3481
3482 static void
handle_fs_outputs_post(struct radv_shader_context * ctx)3483 handle_fs_outputs_post(struct radv_shader_context *ctx)
3484 {
3485 unsigned index = 0;
3486 LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL;
3487 struct ac_export_args color_args[8];
3488
3489 for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
3490 LLVMValueRef values[4];
3491
3492 if (!(ctx->output_mask & (1ull << i)))
3493 continue;
3494
3495 if (i < FRAG_RESULT_DATA0)
3496 continue;
3497
3498 for (unsigned j = 0; j < 4; j++)
3499 values[j] = ac_to_float(&ctx->ac,
3500 radv_load_output(ctx, i, j));
3501
3502 bool ret = si_export_mrt_color(ctx, values,
3503 i - FRAG_RESULT_DATA0,
3504 &color_args[index]);
3505 if (ret)
3506 index++;
3507 }
3508
3509 /* Process depth, stencil, samplemask. */
3510 if (ctx->args->shader_info->ps.writes_z) {
3511 depth = ac_to_float(&ctx->ac,
3512 radv_load_output(ctx, FRAG_RESULT_DEPTH, 0));
3513 }
3514 if (ctx->args->shader_info->ps.writes_stencil) {
3515 stencil = ac_to_float(&ctx->ac,
3516 radv_load_output(ctx, FRAG_RESULT_STENCIL, 0));
3517 }
3518 if (ctx->args->shader_info->ps.writes_sample_mask) {
3519 samplemask = ac_to_float(&ctx->ac,
3520 radv_load_output(ctx, FRAG_RESULT_SAMPLE_MASK, 0));
3521 }
3522
3523 /* Set the DONE bit on last non-null color export only if Z isn't
3524 * exported.
3525 */
3526 if (index > 0 &&
3527 !ctx->args->shader_info->ps.writes_z &&
3528 !ctx->args->shader_info->ps.writes_stencil &&
3529 !ctx->args->shader_info->ps.writes_sample_mask) {
3530 unsigned last = index - 1;
3531
3532 color_args[last].valid_mask = 1; /* whether the EXEC mask is valid */
3533 color_args[last].done = 1; /* DONE bit */
3534 }
3535
3536 /* Export PS outputs. */
3537 for (unsigned i = 0; i < index; i++)
3538 ac_build_export(&ctx->ac, &color_args[i]);
3539
3540 if (depth || stencil || samplemask)
3541 radv_export_mrt_z(ctx, depth, stencil, samplemask);
3542 else if (!index)
3543 ac_build_export_null(&ctx->ac);
3544 }
3545
3546 static void
emit_gs_epilogue(struct radv_shader_context * ctx)3547 emit_gs_epilogue(struct radv_shader_context *ctx)
3548 {
3549 if (ctx->args->options->key.vs_common_out.as_ngg) {
3550 gfx10_ngg_gs_emit_epilogue_1(ctx);
3551 return;
3552 }
3553
3554 if (ctx->ac.chip_class >= GFX10)
3555 LLVMBuildFence(ctx->ac.builder, LLVMAtomicOrderingRelease, false, "");
3556
3557 ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_NOP | AC_SENDMSG_GS_DONE, ctx->gs_wave_id);
3558 }
3559
3560 static void
handle_shader_outputs_post(struct ac_shader_abi * abi,unsigned max_outputs,LLVMValueRef * addrs)3561 handle_shader_outputs_post(struct ac_shader_abi *abi, unsigned max_outputs,
3562 LLVMValueRef *addrs)
3563 {
3564 struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
3565
3566 switch (ctx->stage) {
3567 case MESA_SHADER_VERTEX:
3568 if (ctx->args->options->key.vs_common_out.as_ls)
3569 handle_ls_outputs_post(ctx);
3570 else if (ctx->args->options->key.vs_common_out.as_es)
3571 handle_es_outputs_post(ctx, &ctx->args->shader_info->vs.es_info);
3572 else if (ctx->args->options->key.vs_common_out.as_ngg)
3573 handle_ngg_outputs_post_1(ctx);
3574 else
3575 handle_vs_outputs_post(ctx, ctx->args->options->key.vs_common_out.export_prim_id,
3576 ctx->args->options->key.vs_common_out.export_clip_dists,
3577 &ctx->args->shader_info->vs.outinfo);
3578 break;
3579 case MESA_SHADER_FRAGMENT:
3580 handle_fs_outputs_post(ctx);
3581 break;
3582 case MESA_SHADER_GEOMETRY:
3583 emit_gs_epilogue(ctx);
3584 break;
3585 case MESA_SHADER_TESS_CTRL:
3586 handle_tcs_outputs_post(ctx);
3587 break;
3588 case MESA_SHADER_TESS_EVAL:
3589 if (ctx->args->options->key.vs_common_out.as_es)
3590 handle_es_outputs_post(ctx, &ctx->args->shader_info->tes.es_info);
3591 else if (ctx->args->options->key.vs_common_out.as_ngg)
3592 handle_ngg_outputs_post_1(ctx);
3593 else
3594 handle_vs_outputs_post(ctx, ctx->args->options->key.vs_common_out.export_prim_id,
3595 ctx->args->options->key.vs_common_out.export_clip_dists,
3596 &ctx->args->shader_info->tes.outinfo);
3597 break;
3598 default:
3599 break;
3600 }
3601 }
3602
ac_llvm_finalize_module(struct radv_shader_context * ctx,LLVMPassManagerRef passmgr,const struct radv_nir_compiler_options * options)3603 static void ac_llvm_finalize_module(struct radv_shader_context *ctx,
3604 LLVMPassManagerRef passmgr,
3605 const struct radv_nir_compiler_options *options)
3606 {
3607 LLVMRunPassManager(passmgr, ctx->ac.module);
3608 LLVMDisposeBuilder(ctx->ac.builder);
3609
3610 ac_llvm_context_dispose(&ctx->ac);
3611 }
3612
3613 static void
ac_nir_eliminate_const_vs_outputs(struct radv_shader_context * ctx)3614 ac_nir_eliminate_const_vs_outputs(struct radv_shader_context *ctx)
3615 {
3616 struct radv_vs_output_info *outinfo;
3617
3618 switch (ctx->stage) {
3619 case MESA_SHADER_FRAGMENT:
3620 case MESA_SHADER_COMPUTE:
3621 case MESA_SHADER_TESS_CTRL:
3622 case MESA_SHADER_GEOMETRY:
3623 return;
3624 case MESA_SHADER_VERTEX:
3625 if (ctx->args->options->key.vs_common_out.as_ls ||
3626 ctx->args->options->key.vs_common_out.as_es)
3627 return;
3628 outinfo = &ctx->args->shader_info->vs.outinfo;
3629 break;
3630 case MESA_SHADER_TESS_EVAL:
3631 if (ctx->args->options->key.vs_common_out.as_es)
3632 return;
3633 outinfo = &ctx->args->shader_info->tes.outinfo;
3634 break;
3635 default:
3636 unreachable("Unhandled shader type");
3637 }
3638
3639 ac_optimize_vs_outputs(&ctx->ac,
3640 ctx->main_function,
3641 outinfo->vs_output_param_offset,
3642 VARYING_SLOT_MAX, 0,
3643 &outinfo->param_exports);
3644 }
3645
3646 static void
ac_setup_rings(struct radv_shader_context * ctx)3647 ac_setup_rings(struct radv_shader_context *ctx)
3648 {
3649 if (ctx->args->options->chip_class <= GFX8 &&
3650 (ctx->stage == MESA_SHADER_GEOMETRY ||
3651 ctx->args->options->key.vs_common_out.as_es)) {
3652 unsigned ring = ctx->stage == MESA_SHADER_GEOMETRY ? RING_ESGS_GS
3653 : RING_ESGS_VS;
3654 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, ring, false);
3655
3656 ctx->esgs_ring = ac_build_load_to_sgpr(&ctx->ac,
3657 ctx->ring_offsets,
3658 offset);
3659 }
3660
3661 if (ctx->args->is_gs_copy_shader) {
3662 ctx->gsvs_ring[0] =
3663 ac_build_load_to_sgpr(&ctx->ac, ctx->ring_offsets,
3664 LLVMConstInt(ctx->ac.i32,
3665 RING_GSVS_VS, false));
3666 }
3667
3668 if (ctx->stage == MESA_SHADER_GEOMETRY) {
3669 /* The conceptual layout of the GSVS ring is
3670 * v0c0 .. vLv0 v0c1 .. vLc1 ..
3671 * but the real memory layout is swizzled across
3672 * threads:
3673 * t0v0c0 .. t15v0c0 t0v1c0 .. t15v1c0 ... t15vLcL
3674 * t16v0c0 ..
3675 * Override the buffer descriptor accordingly.
3676 */
3677 LLVMTypeRef v2i64 = LLVMVectorType(ctx->ac.i64, 2);
3678 uint64_t stream_offset = 0;
3679 unsigned num_records = ctx->ac.wave_size;
3680 LLVMValueRef base_ring;
3681
3682 base_ring =
3683 ac_build_load_to_sgpr(&ctx->ac, ctx->ring_offsets,
3684 LLVMConstInt(ctx->ac.i32,
3685 RING_GSVS_GS, false));
3686
3687 for (unsigned stream = 0; stream < 4; stream++) {
3688 unsigned num_components, stride;
3689 LLVMValueRef ring, tmp;
3690
3691 num_components =
3692 ctx->args->shader_info->gs.num_stream_output_components[stream];
3693
3694 if (!num_components)
3695 continue;
3696
3697 stride = 4 * num_components * ctx->shader->info.gs.vertices_out;
3698
3699 /* Limit on the stride field for <= GFX7. */
3700 assert(stride < (1 << 14));
3701
3702 ring = LLVMBuildBitCast(ctx->ac.builder,
3703 base_ring, v2i64, "");
3704 tmp = LLVMBuildExtractElement(ctx->ac.builder,
3705 ring, ctx->ac.i32_0, "");
3706 tmp = LLVMBuildAdd(ctx->ac.builder, tmp,
3707 LLVMConstInt(ctx->ac.i64,
3708 stream_offset, 0), "");
3709 ring = LLVMBuildInsertElement(ctx->ac.builder,
3710 ring, tmp, ctx->ac.i32_0, "");
3711
3712 stream_offset += stride * ctx->ac.wave_size;
3713
3714 ring = LLVMBuildBitCast(ctx->ac.builder, ring,
3715 ctx->ac.v4i32, "");
3716
3717 tmp = LLVMBuildExtractElement(ctx->ac.builder, ring,
3718 ctx->ac.i32_1, "");
3719 tmp = LLVMBuildOr(ctx->ac.builder, tmp,
3720 LLVMConstInt(ctx->ac.i32,
3721 S_008F04_STRIDE(stride), false), "");
3722 ring = LLVMBuildInsertElement(ctx->ac.builder, ring, tmp,
3723 ctx->ac.i32_1, "");
3724
3725 ring = LLVMBuildInsertElement(ctx->ac.builder, ring,
3726 LLVMConstInt(ctx->ac.i32,
3727 num_records, false),
3728 LLVMConstInt(ctx->ac.i32, 2, false), "");
3729
3730 ctx->gsvs_ring[stream] = ring;
3731 }
3732 }
3733
3734 if (ctx->stage == MESA_SHADER_TESS_CTRL ||
3735 ctx->stage == MESA_SHADER_TESS_EVAL) {
3736 ctx->hs_ring_tess_offchip = ac_build_load_to_sgpr(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->ac.i32, RING_HS_TESS_OFFCHIP, false));
3737 ctx->hs_ring_tess_factor = ac_build_load_to_sgpr(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->ac.i32, RING_HS_TESS_FACTOR, false));
3738 }
3739 }
3740
3741 unsigned
radv_nir_get_max_workgroup_size(enum chip_class chip_class,gl_shader_stage stage,const struct nir_shader * nir)3742 radv_nir_get_max_workgroup_size(enum chip_class chip_class,
3743 gl_shader_stage stage,
3744 const struct nir_shader *nir)
3745 {
3746 const unsigned backup_sizes[] = {chip_class >= GFX9 ? 128 : 64, 1, 1};
3747 unsigned sizes[3];
3748 for (unsigned i = 0; i < 3; i++)
3749 sizes[i] = nir ? nir->info.cs.local_size[i] : backup_sizes[i];
3750 return radv_get_max_workgroup_size(chip_class, stage, sizes);
3751 }
3752
3753 /* Fixup the HW not emitting the TCS regs if there are no HS threads. */
ac_nir_fixup_ls_hs_input_vgprs(struct radv_shader_context * ctx)3754 static void ac_nir_fixup_ls_hs_input_vgprs(struct radv_shader_context *ctx)
3755 {
3756 LLVMValueRef count =
3757 ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->merged_wave_info), 8, 8);
3758 LLVMValueRef hs_empty = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, count,
3759 ctx->ac.i32_0, "");
3760 ctx->abi.instance_id = LLVMBuildSelect(ctx->ac.builder, hs_empty,
3761 ac_get_arg(&ctx->ac, ctx->args->rel_auto_id),
3762 ctx->abi.instance_id, "");
3763 ctx->rel_auto_id = LLVMBuildSelect(ctx->ac.builder, hs_empty,
3764 ac_get_arg(&ctx->ac, ctx->args->ac.tcs_rel_ids),
3765 ctx->rel_auto_id,
3766 "");
3767 ctx->abi.vertex_id = LLVMBuildSelect(ctx->ac.builder, hs_empty,
3768 ac_get_arg(&ctx->ac, ctx->args->ac.tcs_patch_id),
3769 ctx->abi.vertex_id, "");
3770 }
3771
prepare_gs_input_vgprs(struct radv_shader_context * ctx,bool merged)3772 static void prepare_gs_input_vgprs(struct radv_shader_context *ctx, bool merged)
3773 {
3774 if (merged) {
3775 for(int i = 5; i >= 0; --i) {
3776 ctx->gs_vtx_offset[i] =
3777 ac_unpack_param(&ctx->ac,
3778 ac_get_arg(&ctx->ac, ctx->args->gs_vtx_offset[i & ~1]),
3779 (i & 1) * 16, 16);
3780 }
3781
3782 ctx->gs_wave_id = ac_unpack_param(&ctx->ac,
3783 ac_get_arg(&ctx->ac, ctx->args->merged_wave_info),
3784 16, 8);
3785 } else {
3786 for (int i = 0; i < 6; i++)
3787 ctx->gs_vtx_offset[i] = ac_get_arg(&ctx->ac, ctx->args->gs_vtx_offset[i]);
3788 ctx->gs_wave_id = ac_get_arg(&ctx->ac, ctx->args->gs_wave_id);
3789 }
3790 }
3791
3792 /* Ensure that the esgs ring is declared.
3793 *
3794 * We declare it with 64KB alignment as a hint that the
3795 * pointer value will always be 0.
3796 */
declare_esgs_ring(struct radv_shader_context * ctx)3797 static void declare_esgs_ring(struct radv_shader_context *ctx)
3798 {
3799 if (ctx->esgs_ring)
3800 return;
3801
3802 assert(!LLVMGetNamedGlobal(ctx->ac.module, "esgs_ring"));
3803
3804 ctx->esgs_ring = LLVMAddGlobalInAddressSpace(
3805 ctx->ac.module, LLVMArrayType(ctx->ac.i32, 0),
3806 "esgs_ring",
3807 AC_ADDR_SPACE_LDS);
3808 LLVMSetLinkage(ctx->esgs_ring, LLVMExternalLinkage);
3809 LLVMSetAlignment(ctx->esgs_ring, 64 * 1024);
3810 }
3811
3812 static
ac_translate_nir_to_llvm(struct ac_llvm_compiler * ac_llvm,struct nir_shader * const * shaders,int shader_count,const struct radv_shader_args * args)3813 LLVMModuleRef ac_translate_nir_to_llvm(struct ac_llvm_compiler *ac_llvm,
3814 struct nir_shader *const *shaders,
3815 int shader_count,
3816 const struct radv_shader_args *args)
3817 {
3818 struct radv_shader_context ctx = {0};
3819 ctx.args = args;
3820
3821 enum ac_float_mode float_mode = AC_FLOAT_MODE_DEFAULT;
3822
3823 if (args->shader_info->float_controls_mode & FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32) {
3824 float_mode = AC_FLOAT_MODE_DENORM_FLUSH_TO_ZERO;
3825 }
3826
3827 ac_llvm_context_init(&ctx.ac, ac_llvm, args->options->chip_class,
3828 args->options->family, float_mode,
3829 args->shader_info->wave_size,
3830 args->shader_info->ballot_bit_size);
3831 ctx.context = ctx.ac.context;
3832
3833 ctx.max_workgroup_size = 0;
3834 for (int i = 0; i < shader_count; ++i) {
3835 ctx.max_workgroup_size = MAX2(ctx.max_workgroup_size,
3836 radv_nir_get_max_workgroup_size(args->options->chip_class,
3837 shaders[i]->info.stage,
3838 shaders[i]));
3839 }
3840
3841 if (ctx.ac.chip_class >= GFX10) {
3842 if (is_pre_gs_stage(shaders[0]->info.stage) &&
3843 args->options->key.vs_common_out.as_ngg) {
3844 ctx.max_workgroup_size = 128;
3845 }
3846 }
3847
3848 create_function(&ctx, shaders[shader_count - 1]->info.stage, shader_count >= 2);
3849
3850 ctx.abi.inputs = &ctx.inputs[0];
3851 ctx.abi.emit_outputs = handle_shader_outputs_post;
3852 ctx.abi.emit_vertex_with_counter = visit_emit_vertex_with_counter;
3853 ctx.abi.load_ubo = radv_load_ubo;
3854 ctx.abi.load_ssbo = radv_load_ssbo;
3855 ctx.abi.load_sampler_desc = radv_get_sampler_desc;
3856 ctx.abi.load_resource = radv_load_resource;
3857 ctx.abi.clamp_shadow_reference = false;
3858 ctx.abi.robust_buffer_access = args->options->robust_buffer_access;
3859
3860 bool is_ngg = is_pre_gs_stage(shaders[0]->info.stage) && args->options->key.vs_common_out.as_ngg;
3861 if (shader_count >= 2 || is_ngg)
3862 ac_init_exec_full_mask(&ctx.ac);
3863
3864 if (args->ac.vertex_id.used)
3865 ctx.abi.vertex_id = ac_get_arg(&ctx.ac, args->ac.vertex_id);
3866 if (args->rel_auto_id.used)
3867 ctx.rel_auto_id = ac_get_arg(&ctx.ac, args->rel_auto_id);
3868 if (args->ac.instance_id.used)
3869 ctx.abi.instance_id = ac_get_arg(&ctx.ac, args->ac.instance_id);
3870
3871 if (args->options->has_ls_vgpr_init_bug &&
3872 shaders[shader_count - 1]->info.stage == MESA_SHADER_TESS_CTRL)
3873 ac_nir_fixup_ls_hs_input_vgprs(&ctx);
3874
3875 if (is_ngg) {
3876 /* Declare scratch space base for streamout and vertex
3877 * compaction. Whether space is actually allocated is
3878 * determined during linking / PM4 creation.
3879 *
3880 * Add an extra dword per vertex to ensure an odd stride, which
3881 * avoids bank conflicts for SoA accesses.
3882 */
3883 if (!args->options->key.vs_common_out.as_ngg_passthrough)
3884 declare_esgs_ring(&ctx);
3885
3886 /* This is really only needed when streamout and / or vertex
3887 * compaction is enabled.
3888 */
3889 if (args->shader_info->so.num_outputs) {
3890 LLVMTypeRef asi32 = LLVMArrayType(ctx.ac.i32, 8);
3891 ctx.gs_ngg_scratch = LLVMAddGlobalInAddressSpace(ctx.ac.module,
3892 asi32, "ngg_scratch", AC_ADDR_SPACE_LDS);
3893 LLVMSetInitializer(ctx.gs_ngg_scratch, LLVMGetUndef(asi32));
3894 LLVMSetAlignment(ctx.gs_ngg_scratch, 4);
3895 }
3896 }
3897
3898 for(int i = 0; i < shader_count; ++i) {
3899 ctx.stage = shaders[i]->info.stage;
3900 ctx.shader = shaders[i];
3901 ctx.output_mask = 0;
3902
3903 if (shaders[i]->info.stage == MESA_SHADER_GEOMETRY) {
3904 for (int i = 0; i < 4; i++) {
3905 ctx.gs_next_vertex[i] =
3906 ac_build_alloca(&ctx.ac, ctx.ac.i32, "");
3907 }
3908 if (args->options->key.vs_common_out.as_ngg) {
3909 for (unsigned i = 0; i < 4; ++i) {
3910 ctx.gs_curprim_verts[i] =
3911 ac_build_alloca(&ctx.ac, ctx.ac.i32, "");
3912 ctx.gs_generated_prims[i] =
3913 ac_build_alloca(&ctx.ac, ctx.ac.i32, "");
3914 }
3915
3916 unsigned scratch_size = 8;
3917 if (args->shader_info->so.num_outputs)
3918 scratch_size = 44;
3919
3920 LLVMTypeRef ai32 = LLVMArrayType(ctx.ac.i32, scratch_size);
3921 ctx.gs_ngg_scratch =
3922 LLVMAddGlobalInAddressSpace(ctx.ac.module,
3923 ai32, "ngg_scratch", AC_ADDR_SPACE_LDS);
3924 LLVMSetInitializer(ctx.gs_ngg_scratch, LLVMGetUndef(ai32));
3925 LLVMSetAlignment(ctx.gs_ngg_scratch, 4);
3926
3927 ctx.gs_ngg_emit = LLVMAddGlobalInAddressSpace(ctx.ac.module,
3928 LLVMArrayType(ctx.ac.i32, 0), "ngg_emit", AC_ADDR_SPACE_LDS);
3929 LLVMSetLinkage(ctx.gs_ngg_emit, LLVMExternalLinkage);
3930 LLVMSetAlignment(ctx.gs_ngg_emit, 4);
3931 }
3932
3933 ctx.abi.load_inputs = load_gs_input;
3934 ctx.abi.emit_primitive = visit_end_primitive;
3935 } else if (shaders[i]->info.stage == MESA_SHADER_TESS_CTRL) {
3936 ctx.abi.load_tess_varyings = load_tcs_varyings;
3937 ctx.abi.load_patch_vertices_in = load_patch_vertices_in;
3938 ctx.abi.store_tcs_outputs = store_tcs_output;
3939 ctx.tcs_num_inputs = ctx.args->shader_info->tcs.num_linked_inputs;
3940 unsigned tcs_num_outputs = ctx.args->shader_info->tcs.num_linked_outputs;
3941 unsigned tcs_num_patch_outputs = ctx.args->shader_info->tcs.num_linked_patch_outputs;
3942 ctx.tcs_num_patches =
3943 get_tcs_num_patches(
3944 ctx.args->options->key.tcs.input_vertices,
3945 ctx.shader->info.tess.tcs_vertices_out,
3946 ctx.tcs_num_inputs,
3947 tcs_num_outputs,
3948 tcs_num_patch_outputs,
3949 ctx.args->options->tess_offchip_block_dw_size,
3950 ctx.args->options->chip_class,
3951 ctx.args->options->family);
3952 } else if (shaders[i]->info.stage == MESA_SHADER_TESS_EVAL) {
3953 ctx.abi.load_tess_varyings = load_tes_input;
3954 ctx.abi.load_tess_coord = load_tess_coord;
3955 ctx.abi.load_patch_vertices_in = load_patch_vertices_in;
3956 ctx.tcs_num_patches = args->options->key.tes.num_patches;
3957 } else if (shaders[i]->info.stage == MESA_SHADER_VERTEX) {
3958 ctx.abi.load_base_vertex = radv_load_base_vertex;
3959 } else if (shaders[i]->info.stage == MESA_SHADER_FRAGMENT) {
3960 ctx.abi.load_sample_position = load_sample_position;
3961 ctx.abi.load_sample_mask_in = load_sample_mask_in;
3962 }
3963
3964 if (shaders[i]->info.stage == MESA_SHADER_VERTEX &&
3965 args->options->key.vs_common_out.as_ngg &&
3966 args->options->key.vs_common_out.export_prim_id) {
3967 declare_esgs_ring(&ctx);
3968 }
3969
3970 bool nested_barrier = false;
3971
3972 if (i) {
3973 if (shaders[i]->info.stage == MESA_SHADER_GEOMETRY &&
3974 args->options->key.vs_common_out.as_ngg) {
3975 gfx10_ngg_gs_emit_prologue(&ctx);
3976 nested_barrier = false;
3977 } else {
3978 nested_barrier = true;
3979 }
3980 }
3981
3982 if (nested_barrier) {
3983 /* Execute a barrier before the second shader in
3984 * a merged shader.
3985 *
3986 * Execute the barrier inside the conditional block,
3987 * so that empty waves can jump directly to s_endpgm,
3988 * which will also signal the barrier.
3989 *
3990 * This is possible in gfx9, because an empty wave
3991 * for the second shader does not participate in
3992 * the epilogue. With NGG, empty waves may still
3993 * be required to export data (e.g. GS output vertices),
3994 * so we cannot let them exit early.
3995 *
3996 * If the shader is TCS and the TCS epilog is present
3997 * and contains a barrier, it will wait there and then
3998 * reach s_endpgm.
3999 */
4000 ac_emit_barrier(&ctx.ac, ctx.stage);
4001 }
4002
4003 nir_foreach_shader_out_variable(variable, shaders[i])
4004 scan_shader_output_decl(&ctx, variable, shaders[i], shaders[i]->info.stage);
4005
4006 ac_setup_rings(&ctx);
4007
4008 LLVMBasicBlockRef merge_block = NULL;
4009 if (shader_count >= 2 || is_ngg) {
4010 LLVMValueRef fn = LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx.ac.builder));
4011 LLVMBasicBlockRef then_block = LLVMAppendBasicBlockInContext(ctx.ac.context, fn, "");
4012 merge_block = LLVMAppendBasicBlockInContext(ctx.ac.context, fn, "");
4013
4014 LLVMValueRef count =
4015 ac_unpack_param(&ctx.ac,
4016 ac_get_arg(&ctx.ac, args->merged_wave_info),
4017 8 * i, 8);
4018 LLVMValueRef thread_id = ac_get_thread_id(&ctx.ac);
4019 LLVMValueRef cond = LLVMBuildICmp(ctx.ac.builder, LLVMIntULT,
4020 thread_id, count, "");
4021 LLVMBuildCondBr(ctx.ac.builder, cond, then_block, merge_block);
4022
4023 LLVMPositionBuilderAtEnd(ctx.ac.builder, then_block);
4024 }
4025
4026 if (shaders[i]->info.stage == MESA_SHADER_FRAGMENT)
4027 prepare_interp_optimize(&ctx, shaders[i]);
4028 else if(shaders[i]->info.stage == MESA_SHADER_VERTEX)
4029 handle_vs_inputs(&ctx, shaders[i]);
4030 else if(shaders[i]->info.stage == MESA_SHADER_GEOMETRY)
4031 prepare_gs_input_vgprs(&ctx, shader_count >= 2);
4032
4033 ac_nir_translate(&ctx.ac, &ctx.abi, &args->ac, shaders[i]);
4034
4035 if (shader_count >= 2 || is_ngg) {
4036 LLVMBuildBr(ctx.ac.builder, merge_block);
4037 LLVMPositionBuilderAtEnd(ctx.ac.builder, merge_block);
4038 }
4039
4040 /* This needs to be outside the if wrapping the shader body, as sometimes
4041 * the HW generates waves with 0 es/vs threads. */
4042 if (is_pre_gs_stage(shaders[i]->info.stage) &&
4043 args->options->key.vs_common_out.as_ngg &&
4044 i == shader_count - 1) {
4045 handle_ngg_outputs_post_2(&ctx);
4046 } else if (shaders[i]->info.stage == MESA_SHADER_GEOMETRY &&
4047 args->options->key.vs_common_out.as_ngg) {
4048 gfx10_ngg_gs_emit_epilogue_2(&ctx);
4049 }
4050
4051 if (shaders[i]->info.stage == MESA_SHADER_TESS_CTRL) {
4052 unsigned tcs_num_outputs = ctx.args->shader_info->tcs.num_linked_outputs;
4053 unsigned tcs_num_patch_outputs = ctx.args->shader_info->tcs.num_linked_patch_outputs;
4054 args->shader_info->tcs.num_patches = ctx.tcs_num_patches;
4055 args->shader_info->tcs.num_lds_blocks =
4056 calculate_tess_lds_size(
4057 ctx.args->options->chip_class,
4058 ctx.args->options->key.tcs.input_vertices,
4059 ctx.shader->info.tess.tcs_vertices_out,
4060 ctx.tcs_num_inputs,
4061 ctx.tcs_num_patches,
4062 tcs_num_outputs,
4063 tcs_num_patch_outputs);
4064 }
4065 }
4066
4067 LLVMBuildRetVoid(ctx.ac.builder);
4068
4069 if (args->options->dump_preoptir) {
4070 fprintf(stderr, "%s LLVM IR:\n\n",
4071 radv_get_shader_name(args->shader_info,
4072 shaders[shader_count - 1]->info.stage));
4073 ac_dump_module(ctx.ac.module);
4074 fprintf(stderr, "\n");
4075 }
4076
4077 ac_llvm_finalize_module(&ctx, ac_llvm->passmgr, args->options);
4078
4079 if (shader_count == 1)
4080 ac_nir_eliminate_const_vs_outputs(&ctx);
4081
4082 if (args->options->dump_shader) {
4083 args->shader_info->private_mem_vgprs =
4084 ac_count_scratch_private_memory(ctx.main_function);
4085 }
4086
4087 return ctx.ac.module;
4088 }
4089
ac_diagnostic_handler(LLVMDiagnosticInfoRef di,void * context)4090 static void ac_diagnostic_handler(LLVMDiagnosticInfoRef di, void *context)
4091 {
4092 unsigned *retval = (unsigned *)context;
4093 LLVMDiagnosticSeverity severity = LLVMGetDiagInfoSeverity(di);
4094 char *description = LLVMGetDiagInfoDescription(di);
4095
4096 if (severity == LLVMDSError) {
4097 *retval = 1;
4098 fprintf(stderr, "LLVM triggered Diagnostic Handler: %s\n",
4099 description);
4100 }
4101
4102 LLVMDisposeMessage(description);
4103 }
4104
radv_llvm_compile(LLVMModuleRef M,char ** pelf_buffer,size_t * pelf_size,struct ac_llvm_compiler * ac_llvm)4105 static unsigned radv_llvm_compile(LLVMModuleRef M,
4106 char **pelf_buffer, size_t *pelf_size,
4107 struct ac_llvm_compiler *ac_llvm)
4108 {
4109 unsigned retval = 0;
4110 LLVMContextRef llvm_ctx;
4111
4112 /* Setup Diagnostic Handler*/
4113 llvm_ctx = LLVMGetModuleContext(M);
4114
4115 LLVMContextSetDiagnosticHandler(llvm_ctx, ac_diagnostic_handler,
4116 &retval);
4117
4118 /* Compile IR*/
4119 if (!radv_compile_to_elf(ac_llvm, M, pelf_buffer, pelf_size))
4120 retval = 1;
4121 return retval;
4122 }
4123
ac_compile_llvm_module(struct ac_llvm_compiler * ac_llvm,LLVMModuleRef llvm_module,struct radv_shader_binary ** rbinary,gl_shader_stage stage,const char * name,const struct radv_nir_compiler_options * options)4124 static void ac_compile_llvm_module(struct ac_llvm_compiler *ac_llvm,
4125 LLVMModuleRef llvm_module,
4126 struct radv_shader_binary **rbinary,
4127 gl_shader_stage stage,
4128 const char *name,
4129 const struct radv_nir_compiler_options *options)
4130 {
4131 char *elf_buffer = NULL;
4132 size_t elf_size = 0;
4133 char *llvm_ir_string = NULL;
4134
4135 if (options->dump_shader) {
4136 fprintf(stderr, "%s LLVM IR:\n\n", name);
4137 ac_dump_module(llvm_module);
4138 fprintf(stderr, "\n");
4139 }
4140
4141 if (options->record_ir) {
4142 char *llvm_ir = LLVMPrintModuleToString(llvm_module);
4143 llvm_ir_string = strdup(llvm_ir);
4144 LLVMDisposeMessage(llvm_ir);
4145 }
4146
4147 int v = radv_llvm_compile(llvm_module, &elf_buffer, &elf_size, ac_llvm);
4148 if (v) {
4149 fprintf(stderr, "compile failed\n");
4150 }
4151
4152 LLVMContextRef ctx = LLVMGetModuleContext(llvm_module);
4153 LLVMDisposeModule(llvm_module);
4154 LLVMContextDispose(ctx);
4155
4156 size_t llvm_ir_size = llvm_ir_string ? strlen(llvm_ir_string) : 0;
4157 size_t alloc_size = sizeof(struct radv_shader_binary_rtld) + elf_size + llvm_ir_size + 1;
4158 struct radv_shader_binary_rtld *rbin = calloc(1, alloc_size);
4159 memcpy(rbin->data, elf_buffer, elf_size);
4160 if (llvm_ir_string)
4161 memcpy(rbin->data + elf_size, llvm_ir_string, llvm_ir_size + 1);
4162
4163 rbin->base.type = RADV_BINARY_TYPE_RTLD;
4164 rbin->base.stage = stage;
4165 rbin->base.total_size = alloc_size;
4166 rbin->elf_size = elf_size;
4167 rbin->llvm_ir_size = llvm_ir_size;
4168 *rbinary = &rbin->base;
4169
4170 free(llvm_ir_string);
4171 free(elf_buffer);
4172 }
4173
4174 static void
radv_compile_nir_shader(struct ac_llvm_compiler * ac_llvm,struct radv_shader_binary ** rbinary,const struct radv_shader_args * args,struct nir_shader * const * nir,int nir_count)4175 radv_compile_nir_shader(struct ac_llvm_compiler *ac_llvm,
4176 struct radv_shader_binary **rbinary,
4177 const struct radv_shader_args *args,
4178 struct nir_shader *const *nir,
4179 int nir_count)
4180 {
4181
4182 LLVMModuleRef llvm_module;
4183
4184 llvm_module = ac_translate_nir_to_llvm(ac_llvm, nir, nir_count, args);
4185
4186 ac_compile_llvm_module(ac_llvm, llvm_module, rbinary,
4187 nir[nir_count - 1]->info.stage,
4188 radv_get_shader_name(args->shader_info,
4189 nir[nir_count - 1]->info.stage),
4190 args->options);
4191
4192 /* Determine the ES type (VS or TES) for the GS on GFX9. */
4193 if (args->options->chip_class >= GFX9) {
4194 if (nir_count == 2 &&
4195 nir[1]->info.stage == MESA_SHADER_GEOMETRY) {
4196 args->shader_info->gs.es_type = nir[0]->info.stage;
4197 }
4198 }
4199 }
4200
4201 static void
ac_gs_copy_shader_emit(struct radv_shader_context * ctx)4202 ac_gs_copy_shader_emit(struct radv_shader_context *ctx)
4203 {
4204 LLVMValueRef vtx_offset =
4205 LLVMBuildMul(ctx->ac.builder, ac_get_arg(&ctx->ac, ctx->args->ac.vertex_id),
4206 LLVMConstInt(ctx->ac.i32, 4, false), "");
4207 LLVMValueRef stream_id;
4208
4209 /* Fetch the vertex stream ID. */
4210 if (!ctx->args->options->use_ngg_streamout &&
4211 ctx->args->shader_info->so.num_outputs) {
4212 stream_id =
4213 ac_unpack_param(&ctx->ac,
4214 ac_get_arg(&ctx->ac,
4215 ctx->args->streamout_config),
4216 24, 2);
4217 } else {
4218 stream_id = ctx->ac.i32_0;
4219 }
4220
4221 LLVMBasicBlockRef end_bb;
4222 LLVMValueRef switch_inst;
4223
4224 end_bb = LLVMAppendBasicBlockInContext(ctx->ac.context,
4225 ctx->main_function, "end");
4226 switch_inst = LLVMBuildSwitch(ctx->ac.builder, stream_id, end_bb, 4);
4227
4228 for (unsigned stream = 0; stream < 4; stream++) {
4229 unsigned num_components =
4230 ctx->args->shader_info->gs.num_stream_output_components[stream];
4231 LLVMBasicBlockRef bb;
4232 unsigned offset;
4233
4234 if (stream > 0 && !num_components)
4235 continue;
4236
4237 if (stream > 0 && !ctx->args->shader_info->so.num_outputs)
4238 continue;
4239
4240 bb = LLVMInsertBasicBlockInContext(ctx->ac.context, end_bb, "out");
4241 LLVMAddCase(switch_inst, LLVMConstInt(ctx->ac.i32, stream, 0), bb);
4242 LLVMPositionBuilderAtEnd(ctx->ac.builder, bb);
4243
4244 offset = 0;
4245 for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
4246 unsigned output_usage_mask =
4247 ctx->args->shader_info->gs.output_usage_mask[i];
4248 unsigned output_stream =
4249 ctx->args->shader_info->gs.output_streams[i];
4250 int length = util_last_bit(output_usage_mask);
4251
4252 if (!(ctx->output_mask & (1ull << i)) ||
4253 output_stream != stream)
4254 continue;
4255
4256 for (unsigned j = 0; j < length; j++) {
4257 LLVMValueRef value, soffset;
4258
4259 if (!(output_usage_mask & (1 << j)))
4260 continue;
4261
4262 soffset = LLVMConstInt(ctx->ac.i32,
4263 offset *
4264 ctx->shader->info.gs.vertices_out * 16 * 4, false);
4265
4266 offset++;
4267
4268 value = ac_build_buffer_load(&ctx->ac,
4269 ctx->gsvs_ring[0],
4270 1, ctx->ac.i32_0,
4271 vtx_offset, soffset,
4272 0, ac_glc | ac_slc, true, false);
4273
4274 LLVMTypeRef type = LLVMGetAllocatedType(ctx->abi.outputs[ac_llvm_reg_index_soa(i, j)]);
4275 if (ac_get_type_size(type) == 2) {
4276 value = LLVMBuildBitCast(ctx->ac.builder, value, ctx->ac.i32, "");
4277 value = LLVMBuildTrunc(ctx->ac.builder, value, ctx->ac.i16, "");
4278 }
4279
4280 LLVMBuildStore(ctx->ac.builder,
4281 ac_to_float(&ctx->ac, value), ctx->abi.outputs[ac_llvm_reg_index_soa(i, j)]);
4282 }
4283 }
4284
4285 if (!ctx->args->options->use_ngg_streamout &&
4286 ctx->args->shader_info->so.num_outputs)
4287 radv_emit_streamout(ctx, stream);
4288
4289 if (stream == 0) {
4290 handle_vs_outputs_post(ctx, false, true,
4291 &ctx->args->shader_info->vs.outinfo);
4292 }
4293
4294 LLVMBuildBr(ctx->ac.builder, end_bb);
4295 }
4296
4297 LLVMPositionBuilderAtEnd(ctx->ac.builder, end_bb);
4298 }
4299
4300 static void
radv_compile_gs_copy_shader(struct ac_llvm_compiler * ac_llvm,struct nir_shader * geom_shader,struct radv_shader_binary ** rbinary,const struct radv_shader_args * args)4301 radv_compile_gs_copy_shader(struct ac_llvm_compiler *ac_llvm,
4302 struct nir_shader *geom_shader,
4303 struct radv_shader_binary **rbinary,
4304 const struct radv_shader_args *args)
4305 {
4306 struct radv_shader_context ctx = {0};
4307 ctx.args = args;
4308
4309 assert(args->is_gs_copy_shader);
4310
4311 ac_llvm_context_init(&ctx.ac, ac_llvm, args->options->chip_class,
4312 args->options->family, AC_FLOAT_MODE_DEFAULT, 64, 64);
4313 ctx.context = ctx.ac.context;
4314
4315 ctx.stage = MESA_SHADER_VERTEX;
4316 ctx.shader = geom_shader;
4317
4318 create_function(&ctx, MESA_SHADER_VERTEX, false);
4319
4320 ac_setup_rings(&ctx);
4321
4322 nir_foreach_shader_out_variable(variable, geom_shader) {
4323 scan_shader_output_decl(&ctx, variable, geom_shader, MESA_SHADER_VERTEX);
4324 ac_handle_shader_output_decl(&ctx.ac, &ctx.abi, geom_shader,
4325 variable, MESA_SHADER_VERTEX);
4326 }
4327
4328 ac_gs_copy_shader_emit(&ctx);
4329
4330 LLVMBuildRetVoid(ctx.ac.builder);
4331
4332 ac_llvm_finalize_module(&ctx, ac_llvm->passmgr, args->options);
4333
4334 ac_compile_llvm_module(ac_llvm, ctx.ac.module, rbinary,
4335 MESA_SHADER_VERTEX, "GS Copy Shader", args->options);
4336 (*rbinary)->is_gs_copy_shader = true;
4337
4338 }
4339
4340 void
llvm_compile_shader(struct radv_device * device,unsigned shader_count,struct nir_shader * const * shaders,struct radv_shader_binary ** binary,struct radv_shader_args * args)4341 llvm_compile_shader(struct radv_device *device,
4342 unsigned shader_count,
4343 struct nir_shader *const *shaders,
4344 struct radv_shader_binary **binary,
4345 struct radv_shader_args *args)
4346 {
4347 enum ac_target_machine_options tm_options = 0;
4348 struct ac_llvm_compiler ac_llvm;
4349 bool thread_compiler;
4350
4351 tm_options |= AC_TM_SUPPORTS_SPILL;
4352 if (args->options->check_ir)
4353 tm_options |= AC_TM_CHECK_IR;
4354
4355 thread_compiler = !(device->instance->debug_flags & RADV_DEBUG_NOTHREADLLVM);
4356
4357 radv_init_llvm_compiler(&ac_llvm, thread_compiler,
4358 args->options->family, tm_options,
4359 args->shader_info->wave_size);
4360
4361 if (args->is_gs_copy_shader) {
4362 radv_compile_gs_copy_shader(&ac_llvm, *shaders, binary, args);
4363 } else {
4364 radv_compile_nir_shader(&ac_llvm, binary, args,
4365 shaders, shader_count);
4366 }
4367
4368 radv_destroy_llvm_compiler(&ac_llvm, thread_compiler);
4369 }
4370