projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radeonsi/gfx10: don't use MALLOC for outputs
[mesa.git] / src / gallium / drivers / radeonsi / gfx10_shader_ngg.c
1 /*
2 * Copyright 2017 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 */
23
24 #include "si_pipe.h"
25 #include "si_shader_internal.h"
26
27 #include "sid.h"
28
29 #include "util/u_memory.h"
30 #include "util/u_prim.h"
31
32 static LLVMValueRef get_wave_id_in_tg(struct si_shader_context *ctx)
33 {
34 return si_unpack_param(ctx, ctx->param_merged_wave_info, 24, 4);
35 }
36
37 static LLVMValueRef get_tgsize(struct si_shader_context *ctx)
38 {
39 return si_unpack_param(ctx, ctx->param_merged_wave_info, 28, 4);
40 }
41
42 static LLVMValueRef get_thread_id_in_tg(struct si_shader_context *ctx)
43 {
44 LLVMBuilderRef builder = ctx->ac.builder;
45 LLVMValueRef tmp;
46 tmp = LLVMBuildMul(builder, get_wave_id_in_tg(ctx),
47 LLVMConstInt(ctx->ac.i32, 64, false), "");
48 return LLVMBuildAdd(builder, tmp, ac_get_thread_id(&ctx->ac), "");
49 }
50
51 static LLVMValueRef ngg_get_vtx_cnt(struct si_shader_context *ctx)
52 {
53 return ac_build_bfe(&ctx->ac, ctx->gs_tg_info,
54 LLVMConstInt(ctx->ac.i32, 12, false),
55 LLVMConstInt(ctx->ac.i32, 9, false),
56 false);
57 }
58
59 static LLVMValueRef ngg_get_prim_cnt(struct si_shader_context *ctx)
60 {
61 return ac_build_bfe(&ctx->ac, ctx->gs_tg_info,
62 LLVMConstInt(ctx->ac.i32, 22, false),
63 LLVMConstInt(ctx->ac.i32, 9, false),
64 false);
65 }
66
67 static LLVMValueRef ngg_get_ordered_id(struct si_shader_context *ctx)
68 {
69 return ac_build_bfe(&ctx->ac, ctx->gs_tg_info,
70 ctx->i32_0,
71 LLVMConstInt(ctx->ac.i32, 11, false),
72 false);
73 }
74
75 static LLVMValueRef ngg_get_query_buf(struct si_shader_context *ctx)
76 {
77 LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn,
78 ctx->param_rw_buffers);
79
80 return ac_build_load_to_sgpr(&ctx->ac, buf_ptr,
81 LLVMConstInt(ctx->i32, GFX10_GS_QUERY_BUF, false));
82 }
83
84 /* Send GS Alloc Req message from the first wave of the group to SPI.
85 * Message payload is:
86 * - bits 0..10: vertices in group
87 * - bits 12..22: primitives in group
88 */
89 static void build_sendmsg_gs_alloc_req(struct si_shader_context *ctx,
90 LLVMValueRef vtx_cnt,
91 LLVMValueRef prim_cnt)
92 {
93 LLVMBuilderRef builder = ctx->ac.builder;
94 LLVMValueRef tmp;
95
96 tmp = LLVMBuildICmp(builder, LLVMIntEQ, get_wave_id_in_tg(ctx), ctx->ac.i32_0, "");
97 ac_build_ifcc(&ctx->ac, tmp, 5020);
98
99 tmp = LLVMBuildShl(builder, prim_cnt, LLVMConstInt(ctx->ac.i32, 12, false),"");
100 tmp = LLVMBuildOr(builder, tmp, vtx_cnt, "");
101 ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_ALLOC_REQ, tmp);
102
103 ac_build_endif(&ctx->ac, 5020);
104 }
105
106 struct ngg_prim {
107 unsigned num_vertices;
108 LLVMValueRef isnull;
109 LLVMValueRef index[3];
110 LLVMValueRef edgeflag[3];
111 };
112
113 static void build_export_prim(struct si_shader_context *ctx,
114 const struct ngg_prim *prim)
115 {
116 LLVMBuilderRef builder = ctx->ac.builder;
117 struct ac_export_args args;
118 LLVMValueRef tmp;
119
120 tmp = LLVMBuildZExt(builder, prim->isnull, ctx->ac.i32, "");
121 args.out[0] = LLVMBuildShl(builder, tmp, LLVMConstInt(ctx->ac.i32, 31, false), "");
122
123 for (unsigned i = 0; i < prim->num_vertices; ++i) {
124 tmp = LLVMBuildShl(builder, prim->index[i],
125 LLVMConstInt(ctx->ac.i32, 10 * i, false), "");
126 args.out[0] = LLVMBuildOr(builder, args.out[0], tmp, "");
127 tmp = LLVMBuildZExt(builder, prim->edgeflag[i], ctx->ac.i32, "");
128 tmp = LLVMBuildShl(builder, tmp,
129 LLVMConstInt(ctx->ac.i32, 10 * i + 9, false), "");
130 args.out[0] = LLVMBuildOr(builder, args.out[0], tmp, "");
131 }
132
133 args.out[0] = LLVMBuildBitCast(builder, args.out[0], ctx->ac.f32, "");
134 args.out[1] = LLVMGetUndef(ctx->ac.f32);
135 args.out[2] = LLVMGetUndef(ctx->ac.f32);
136 args.out[3] = LLVMGetUndef(ctx->ac.f32);
137
138 args.target = V_008DFC_SQ_EXP_PRIM;
139 args.enabled_channels = 1;
140 args.done = true;
141 args.valid_mask = false;
142 args.compr = false;
143
144 ac_build_export(&ctx->ac, &args);
145 }
146
147 static void build_streamout_vertex(struct si_shader_context *ctx,
148 LLVMValueRef *so_buffer, LLVMValueRef *wg_offset_dw,
149 unsigned stream, LLVMValueRef offset_vtx,
150 LLVMValueRef vertexptr)
151 {
152 struct tgsi_shader_info *info = &ctx->shader->selector->info;
153 struct pipe_stream_output_info *so = &ctx->shader->selector->so;
154 LLVMBuilderRef builder = ctx->ac.builder;
155 LLVMValueRef offset[4] = {};
156 LLVMValueRef tmp;
157
158 for (unsigned buffer = 0; buffer < 4; ++buffer) {
159 if (!wg_offset_dw[buffer])
160 continue;
161
162 tmp = LLVMBuildMul(builder, offset_vtx,
163 LLVMConstInt(ctx->i32, so->stride[buffer], false), "");
164 tmp = LLVMBuildAdd(builder, wg_offset_dw[buffer], tmp, "");
165 offset[buffer] = LLVMBuildShl(builder, tmp, LLVMConstInt(ctx->i32, 2, false), "");
166 }
167
168 for (unsigned i = 0; i < so->num_outputs; ++i) {
169 if (so->output[i].stream != stream)
170 continue;
171
172 unsigned reg = so->output[i].register_index;
173 struct si_shader_output_values out;
174 out.semantic_name = info->output_semantic_name[reg];
175 out.semantic_index = info->output_semantic_index[reg];
176
177 for (unsigned comp = 0; comp < 4; comp++) {
178 tmp = ac_build_gep0(&ctx->ac, vertexptr,
179 LLVMConstInt(ctx->i32, 4 * reg + comp, false));
180 out.values[comp] = LLVMBuildLoad(builder, tmp, "");
181 out.vertex_stream[comp] =
182 (info->output_streams[reg] >> (2 * comp)) & 3;
183 }
184
185 si_emit_streamout_output(ctx, so_buffer, offset, &so->output[i], &out);
186 }
187 }
188
189 struct ngg_streamout {
190 LLVMValueRef num_vertices;
191
192 /* per-thread data */
193 LLVMValueRef prim_enable[4]; /* i1 per stream */
194 LLVMValueRef vertices[3]; /* [N x i32] addrspace(LDS)* */
195
196 /* Output */
197 LLVMValueRef emit[4]; /* per-stream emitted primitives (only valid for used streams) */
198 };
199
200 /**
201 * Build streamout logic.
202 *
203 * Implies a barrier.
204 *
205 * Writes number of emitted primitives to gs_ngg_scratch[4:8].
206 *
207 * Clobbers gs_ngg_scratch[8:].
208 */
209 static void build_streamout(struct si_shader_context *ctx,
210 struct ngg_streamout *nggso)
211 {
212 struct tgsi_shader_info *info = &ctx->shader->selector->info;
213 struct pipe_stream_output_info *so = &ctx->shader->selector->so;
214 LLVMBuilderRef builder = ctx->ac.builder;
215 LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn, ctx->param_rw_buffers);
216 LLVMValueRef tid = get_thread_id_in_tg(ctx);
217 LLVMValueRef tmp, tmp2;
218 LLVMValueRef i32_2 = LLVMConstInt(ctx->i32, 2, false);
219 LLVMValueRef i32_4 = LLVMConstInt(ctx->i32, 4, false);
220 LLVMValueRef i32_8 = LLVMConstInt(ctx->i32, 8, false);
221 LLVMValueRef so_buffer[4] = {};
222 unsigned max_num_vertices = 1 + (nggso->vertices[1] ? 1 : 0) +
223 (nggso->vertices[2] ? 1 : 0);
224 LLVMValueRef prim_stride_dw[4] = {};
225 LLVMValueRef prim_stride_dw_vgpr = LLVMGetUndef(ctx->i32);
226 int stream_for_buffer[4] = { -1, -1, -1, -1 };
227 unsigned bufmask_for_stream[4] = {};
228 bool isgs = ctx->type == PIPE_SHADER_GEOMETRY;
229 unsigned scratch_emit_base = isgs ? 4 : 0;
230 LLVMValueRef scratch_emit_basev = isgs ? i32_4 : ctx->i32_0;
231 unsigned scratch_offset_base = isgs ? 8 : 4;
232 LLVMValueRef scratch_offset_basev = isgs ? i32_8 : i32_4;
233
234 ac_llvm_add_target_dep_function_attr(ctx->main_fn, "amdgpu-gds-size", 256);
235
236 /* Determine the mapping of streamout buffers to vertex streams. */
237 for (unsigned i = 0; i < so->num_outputs; ++i) {
238 unsigned buf = so->output[i].output_buffer;
239 unsigned stream = so->output[i].stream;
240 assert(stream_for_buffer[buf] < 0 || stream_for_buffer[buf] == stream);
241 stream_for_buffer[buf] = stream;
242 bufmask_for_stream[stream] |= 1 << buf;
243 }
244
245 for (unsigned buffer = 0; buffer < 4; ++buffer) {
246 if (stream_for_buffer[buffer] == -1)
247 continue;
248
249 assert(so->stride[buffer]);
250
251 tmp = LLVMConstInt(ctx->i32, so->stride[buffer], false);
252 prim_stride_dw[buffer] = LLVMBuildMul(builder, tmp, nggso->num_vertices, "");
253 prim_stride_dw_vgpr = ac_build_writelane(
254 &ctx->ac, prim_stride_dw_vgpr, prim_stride_dw[buffer],
255 LLVMConstInt(ctx->i32, buffer, false));
256
257 so_buffer[buffer] = ac_build_load_to_sgpr(
258 &ctx->ac, buf_ptr,
259 LLVMConstInt(ctx->i32, SI_VS_STREAMOUT_BUF0 + buffer, false));
260 }
261
262 tmp = LLVMBuildICmp(builder, LLVMIntEQ, get_wave_id_in_tg(ctx), ctx->i32_0, "");
263 ac_build_ifcc(&ctx->ac, tmp, 5200);
264 {
265 LLVMTypeRef gdsptr = LLVMPointerType(ctx->i32, AC_ADDR_SPACE_GDS);
266 LLVMValueRef gdsbase = LLVMBuildIntToPtr(builder, ctx->i32_0, gdsptr, "");
267
268 /* Advance the streamout offsets in GDS. */
269 LLVMValueRef offsets_vgpr = ac_build_alloca_undef(&ctx->ac, ctx->i32, "");
270 LLVMValueRef generated_by_stream_vgpr = ac_build_alloca_undef(&ctx->ac, ctx->i32, "");
271
272 tmp = LLVMBuildICmp(builder, LLVMIntULT, ac_get_thread_id(&ctx->ac), i32_4, "");
273 ac_build_ifcc(&ctx->ac, tmp, 5210);
274 {
275 if (isgs) {
276 tmp = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tid);
277 tmp = LLVMBuildLoad(builder, tmp, "");
278 } else {
279 tmp = ac_build_writelane(&ctx->ac, ctx->i32_0,
280 ngg_get_prim_cnt(ctx), ctx->i32_0);
281 }
282 LLVMBuildStore(builder, tmp, generated_by_stream_vgpr);
283
284 unsigned swizzle[4];
285 int unused_stream = -1;
286 for (unsigned stream = 0; stream < 4; ++stream) {
287 if (!info->num_stream_output_components[stream]) {
288 unused_stream = stream;
289 break;
290 }
291 }
292 for (unsigned buffer = 0; buffer < 4; ++buffer) {
293 if (stream_for_buffer[buffer] >= 0) {
294 swizzle[buffer] = stream_for_buffer[buffer];
295 } else {
296 assert(unused_stream >= 0);
297 swizzle[buffer] = unused_stream;
298 }
299 }
300
301 tmp = ac_build_quad_swizzle(&ctx->ac, tmp,
302 swizzle[0], swizzle[1], swizzle[2], swizzle[3]);
303 tmp = LLVMBuildMul(builder, tmp, prim_stride_dw_vgpr, "");
304
305 LLVMValueRef args[] = {
306 LLVMBuildIntToPtr(builder, ngg_get_ordered_id(ctx), gdsptr, ""),
307 tmp,
308 ctx->i32_0, // ordering
309 ctx->i32_0, // scope
310 ctx->ac.i1false, // isVolatile
311 LLVMConstInt(ctx->i32, 4 << 24, false), // OA index
312 ctx->ac.i1true, // wave release
313 ctx->ac.i1true, // wave done
314 };
315 tmp = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ds.ordered.add",
316 ctx->i32, args, ARRAY_SIZE(args), 0);
317
318 /* Keep offsets in a VGPR for quick retrieval via readlane by
319 * the first wave for bounds checking, and also store in LDS
320 * for retrieval by all waves later. */
321 LLVMBuildStore(builder, tmp, offsets_vgpr);
322
323 tmp2 = LLVMBuildAdd(builder, ac_get_thread_id(&ctx->ac),
324 scratch_offset_basev, "");
325 tmp2 = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tmp2);
326 LLVMBuildStore(builder, tmp, tmp2);
327 }
328 ac_build_endif(&ctx->ac, 5210);
329
330 /* Determine the max emit per buffer. This is done via the SALU, in part
331 * because LLVM can't generate divide-by-multiply if we try to do this
332 * via VALU with one lane per buffer.
333 */
334 LLVMValueRef max_emit[4] = {};
335 for (unsigned buffer = 0; buffer < 4; ++buffer) {
336 if (stream_for_buffer[buffer] == -1)
337 continue;
338
339 LLVMValueRef bufsize_dw =
340 LLVMBuildLShr(builder,
341 LLVMBuildExtractElement(builder, so_buffer[buffer], i32_2, ""),
342 i32_2, "");
343
344 tmp = LLVMBuildLoad(builder, offsets_vgpr, "");
345 LLVMValueRef offset_dw =
346 ac_build_readlane(&ctx->ac, tmp,
347 LLVMConstInt(ctx->i32, buffer, false));
348
349 tmp = LLVMBuildSub(builder, bufsize_dw, offset_dw, "");
350 tmp = LLVMBuildUDiv(builder, tmp, prim_stride_dw[buffer], "");
351
352 tmp2 = LLVMBuildICmp(builder, LLVMIntULT, bufsize_dw, offset_dw, "");
353 max_emit[buffer] = LLVMBuildSelect(builder, tmp2, ctx->i32_0, tmp, "");
354 }
355
356 /* Determine the number of emitted primitives per stream and fixup the
357 * GDS counter if necessary.
358 *
359 * This is complicated by the fact that a single stream can emit to
360 * multiple buffers (but luckily not vice versa).
361 */
362 LLVMValueRef emit_vgpr = ctx->i32_0;
363
364 for (unsigned stream = 0; stream < 4; ++stream) {
365 if (!info->num_stream_output_components[stream])
366 continue;
367
368 tmp = LLVMBuildLoad(builder, generated_by_stream_vgpr, "");
369 LLVMValueRef generated =
370 ac_build_readlane(&ctx->ac, tmp,
371 LLVMConstInt(ctx->i32, stream, false));
372
373 LLVMValueRef emit = generated;
374 for (unsigned buffer = 0; buffer < 4; ++buffer) {
375 if (stream_for_buffer[buffer] == stream)
376 emit = ac_build_umin(&ctx->ac, emit, max_emit[buffer]);
377 }
378
379 emit_vgpr = ac_build_writelane(&ctx->ac, emit_vgpr, emit,
380 LLVMConstInt(ctx->i32, stream, false));
381
382 /* Fixup the offset using a plain GDS atomic if we overflowed. */
383 tmp = LLVMBuildICmp(builder, LLVMIntULT, emit, generated, "");
384 ac_build_ifcc(&ctx->ac, tmp, 5221); /* scalar branch */
385 tmp = LLVMBuildLShr(builder,
386 LLVMConstInt(ctx->i32, bufmask_for_stream[stream], false),
387 ac_get_thread_id(&ctx->ac), "");
388 tmp = LLVMBuildTrunc(builder, tmp, ctx->i1, "");
389 ac_build_ifcc(&ctx->ac, tmp, 5222);
390 {
391 tmp = LLVMBuildSub(builder, generated, emit, "");
392 tmp = LLVMBuildMul(builder, tmp, prim_stride_dw_vgpr, "");
393 tmp2 = LLVMBuildGEP(builder, gdsbase, &tid, 1, "");
394 LLVMBuildAtomicRMW(builder, LLVMAtomicRMWBinOpSub, tmp2, tmp,
395 LLVMAtomicOrderingMonotonic, false);
396 }
397 ac_build_endif(&ctx->ac, 5222);
398 ac_build_endif(&ctx->ac, 5221);
399 }
400
401 tmp = LLVMBuildICmp(builder, LLVMIntULT, ac_get_thread_id(&ctx->ac), i32_4, "");
402 ac_build_ifcc(&ctx->ac, tmp, 5225);
403 {
404 tmp = LLVMBuildAdd(builder, ac_get_thread_id(&ctx->ac),
405 scratch_emit_basev, "");
406 tmp = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tmp);
407 LLVMBuildStore(builder, emit_vgpr, tmp);
408 }
409 ac_build_endif(&ctx->ac, 5225);
410 }
411 ac_build_endif(&ctx->ac, 5200);
412
413 /* Determine the workgroup-relative per-thread / primitive offset into
414 * the streamout buffers */
415 struct ac_wg_scan primemit_scan[4] = {};
416
417 if (isgs) {
418 for (unsigned stream = 0; stream < 4; ++stream) {
419 if (!info->num_stream_output_components[stream])
420 continue;
421
422 primemit_scan[stream].enable_exclusive = true;
423 primemit_scan[stream].op = nir_op_iadd;
424 primemit_scan[stream].src = nggso->prim_enable[stream];
425 primemit_scan[stream].scratch =
426 ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch,
427 LLVMConstInt(ctx->i32, 12 + 8 * stream, false));
428 primemit_scan[stream].waveidx = get_wave_id_in_tg(ctx);
429 primemit_scan[stream].numwaves = get_tgsize(ctx);
430 primemit_scan[stream].maxwaves = 8;
431 ac_build_wg_scan_top(&ctx->ac, &primemit_scan[stream]);
432 }
433 }
434
435 ac_build_s_barrier(&ctx->ac);
436
437 /* Fetch the per-buffer offsets and per-stream emit counts in all waves. */
438 LLVMValueRef wgoffset_dw[4] = {};
439
440 {
441 LLVMValueRef scratch_vgpr;
442
443 tmp = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, ac_get_thread_id(&ctx->ac));
444 scratch_vgpr = LLVMBuildLoad(builder, tmp, "");
445
446 for (unsigned buffer = 0; buffer < 4; ++buffer) {
447 if (stream_for_buffer[buffer] >= 0) {
448 wgoffset_dw[buffer] = ac_build_readlane(
449 &ctx->ac, scratch_vgpr,
450 LLVMConstInt(ctx->i32, scratch_offset_base + buffer, false));
451 }
452 }
453
454 for (unsigned stream = 0; stream < 4; ++stream) {
455 if (info->num_stream_output_components[stream]) {
456 nggso->emit[stream] = ac_build_readlane(
457 &ctx->ac, scratch_vgpr,
458 LLVMConstInt(ctx->i32, scratch_emit_base + stream, false));
459 }
460 }
461 }
462
463 /* Write out primitive data */
464 for (unsigned stream = 0; stream < 4; ++stream) {
465 if (!info->num_stream_output_components[stream])
466 continue;
467
468 if (isgs) {
469 ac_build_wg_scan_bottom(&ctx->ac, &primemit_scan[stream]);
470 } else {
471 primemit_scan[stream].result_exclusive = tid;
472 }
473
474 tmp = LLVMBuildICmp(builder, LLVMIntULT,
475 primemit_scan[stream].result_exclusive,
476 nggso->emit[stream], "");
477 tmp = LLVMBuildAnd(builder, tmp, nggso->prim_enable[stream], "");
478 ac_build_ifcc(&ctx->ac, tmp, 5240);
479 {
480 LLVMValueRef offset_vtx =
481 LLVMBuildMul(builder, primemit_scan[stream].result_exclusive,
482 nggso->num_vertices, "");
483
484 for (unsigned i = 0; i < max_num_vertices; ++i) {
485 tmp = LLVMBuildICmp(builder, LLVMIntULT,
486 LLVMConstInt(ctx->i32, i, false),
487 nggso->num_vertices, "");
488 ac_build_ifcc(&ctx->ac, tmp, 5241);
489 build_streamout_vertex(ctx, so_buffer, wgoffset_dw,
490 stream, offset_vtx, nggso->vertices[i]);
491 ac_build_endif(&ctx->ac, 5241);
492 offset_vtx = LLVMBuildAdd(builder, offset_vtx, ctx->i32_1, "");
493 }
494 }
495 ac_build_endif(&ctx->ac, 5240);
496 }
497 }
498
499 /**
500 * Returns an `[N x i32] addrspace(LDS)*` pointing at contiguous LDS storage
501 * for the vertex outputs.
502 */
503 static LLVMValueRef ngg_nogs_vertex_ptr(struct si_shader_context *ctx,
504 LLVMValueRef vtxid)
505 {
506 /* The extra dword is used to avoid LDS bank conflicts. */
507 unsigned vertex_size = 4 * ctx->shader->selector->info.num_outputs + 1;
508 LLVMTypeRef ai32 = LLVMArrayType(ctx->i32, vertex_size);
509 LLVMTypeRef pai32 = LLVMPointerType(ai32, AC_ADDR_SPACE_LDS);
510 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, ctx->esgs_ring, pai32, "");
511 return LLVMBuildGEP(ctx->ac.builder, tmp, &vtxid, 1, "");
512 }
513
514 /**
515 * Emit the epilogue of an API VS or TES shader compiled as ESGS shader.
516 */
517 void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi,
518 unsigned max_outputs,
519 LLVMValueRef *addrs)
520 {
521 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
522 struct si_shader_selector *sel = ctx->shader->selector;
523 struct tgsi_shader_info *info = &sel->info;
524 struct si_shader_output_values outputs[PIPE_MAX_SHADER_OUTPUTS];
525 LLVMBuilderRef builder = ctx->ac.builder;
526 struct lp_build_if_state if_state;
527 LLVMValueRef tmp, tmp2;
528
529 assert(!ctx->shader->is_gs_copy_shader);
530 assert(info->num_outputs <= max_outputs);
531
532 LLVMValueRef vertex_ptr = NULL;
533
534 if (sel->so.num_outputs)
535 vertex_ptr = ngg_nogs_vertex_ptr(ctx, get_thread_id_in_tg(ctx));
536
537 for (unsigned i = 0; i < info->num_outputs; i++) {
538 outputs[i].semantic_name = info->output_semantic_name[i];
539 outputs[i].semantic_index = info->output_semantic_index[i];
540
541 /* This is used only by streamout. */
542 for (unsigned j = 0; j < 4; j++) {
543 outputs[i].values[j] =
544 LLVMBuildLoad(builder,
545 addrs[4 * i + j],
546 "");
547 outputs[i].vertex_stream[j] =
548 (info->output_streams[i] >> (2 * j)) & 3;
549
550 if (vertex_ptr) {
551 tmp = ac_build_gep0(&ctx->ac, vertex_ptr,
552 LLVMConstInt(ctx->i32, 4 * i + j, false));
553 tmp2 = ac_to_integer(&ctx->ac, outputs[i].values[j]);
554 LLVMBuildStore(builder, tmp2, tmp);
555 }
556 }
557 }
558
559 lp_build_endif(&ctx->merged_wrap_if_state);
560
561 LLVMValueRef prims_in_wave = si_unpack_param(ctx, ctx->param_merged_wave_info, 8, 8);
562 LLVMValueRef vtx_in_wave = si_unpack_param(ctx, ctx->param_merged_wave_info, 0, 8);
563 LLVMValueRef is_gs_thread = LLVMBuildICmp(builder, LLVMIntULT,
564 ac_get_thread_id(&ctx->ac), prims_in_wave, "");
565 LLVMValueRef is_es_thread = LLVMBuildICmp(builder, LLVMIntULT,
566 ac_get_thread_id(&ctx->ac), vtx_in_wave, "");
567 LLVMValueRef vtxindex[] = {
568 si_unpack_param(ctx, ctx->param_gs_vtx01_offset, 0, 16),
569 si_unpack_param(ctx, ctx->param_gs_vtx01_offset, 16, 16),
570 si_unpack_param(ctx, ctx->param_gs_vtx23_offset, 0, 16),
571 };
572
573 /* Determine the number of vertices per primitive. */
574 unsigned num_vertices;
575 LLVMValueRef num_vertices_val;
576
577 if (ctx->type == PIPE_SHADER_VERTEX) {
578 if (info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS]) {
579 /* Blits always use axis-aligned rectangles with 3 vertices. */
580 num_vertices = 3;
581 num_vertices_val = LLVMConstInt(ctx->i32, 3, 0);
582 } else {
583 /* Extract OUTPRIM field. */
584 tmp = si_unpack_param(ctx, ctx->param_vs_state_bits, 2, 2);
585 num_vertices_val = LLVMBuildAdd(builder, tmp, ctx->i32_1, "");
586 num_vertices = 3; /* TODO: optimize for points & lines */
587 }
588 } else {
589 assert(ctx->type == PIPE_SHADER_TESS_EVAL);
590
591 if (info->properties[TGSI_PROPERTY_TES_POINT_MODE])
592 num_vertices = 1;
593 else if (info->properties[TGSI_PROPERTY_TES_PRIM_MODE] == PIPE_PRIM_LINES)
594 num_vertices = 2;
595 else
596 num_vertices = 3;
597
598 num_vertices_val = LLVMConstInt(ctx->i32, num_vertices, false);
599 }
600
601 /* Streamout */
602 LLVMValueRef emitted_prims = NULL;
603
604 if (sel->so.num_outputs) {
605 struct ngg_streamout nggso = {};
606
607 nggso.num_vertices = num_vertices_val;
608 nggso.prim_enable[0] = is_gs_thread;
609
610 for (unsigned i = 0; i < num_vertices; ++i)
611 nggso.vertices[i] = ngg_nogs_vertex_ptr(ctx, vtxindex[i]);
612
613 build_streamout(ctx, &nggso);
614 emitted_prims = nggso.emit[0];
615 }
616
617 /* Copy Primitive IDs from GS threads to the LDS address corresponding
618 * to the ES thread of the provoking vertex.
619 */
620 if (ctx->type == PIPE_SHADER_VERTEX &&
621 ctx->shader->key.mono.u.vs_export_prim_id) {
622 /* Streamout uses LDS. We need to wait for it before we can reuse it. */
623 if (sel->so.num_outputs)
624 ac_build_s_barrier(&ctx->ac);
625
626 ac_build_ifcc(&ctx->ac, is_gs_thread, 5400);
627 /* Extract the PROVOKING_VTX_INDEX field. */
628 LLVMValueRef provoking_vtx_in_prim =
629 si_unpack_param(ctx, ctx->param_vs_state_bits, 4, 2);
630
631 /* provoking_vtx_index = vtxindex[provoking_vtx_in_prim]; */
632 LLVMValueRef indices = ac_build_gather_values(&ctx->ac, vtxindex, 3);
633 LLVMValueRef provoking_vtx_index =
634 LLVMBuildExtractElement(builder, indices, provoking_vtx_in_prim, "");
635
636 LLVMBuildStore(builder, ctx->abi.gs_prim_id,
637 ac_build_gep0(&ctx->ac, ctx->esgs_ring, provoking_vtx_index));
638 ac_build_endif(&ctx->ac, 5400);
639 }
640
641 /* TODO: primitive culling */
642
643 build_sendmsg_gs_alloc_req(ctx, ngg_get_vtx_cnt(ctx), ngg_get_prim_cnt(ctx));
644
645 /* Update query buffer */
646 /* TODO: this won't catch 96-bit clear_buffer via transform feedback. */
647 if (!info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS]) {
648 tmp = si_unpack_param(ctx, ctx->param_vs_state_bits, 6, 1);
649 tmp = LLVMBuildTrunc(builder, tmp, ctx->i1, "");
650 ac_build_ifcc(&ctx->ac, tmp, 5029); /* if (STREAMOUT_QUERY_ENABLED) */
651 tmp = LLVMBuildICmp(builder, LLVMIntEQ, get_wave_id_in_tg(ctx), ctx->ac.i32_0, "");
652 ac_build_ifcc(&ctx->ac, tmp, 5030);
653 tmp = LLVMBuildICmp(builder, LLVMIntULE, ac_get_thread_id(&ctx->ac),
654 sel->so.num_outputs ? ctx->ac.i32_1 : ctx->ac.i32_0, "");
655 ac_build_ifcc(&ctx->ac, tmp, 5031);
656 {
657 LLVMValueRef args[] = {
658 ngg_get_prim_cnt(ctx),
659 ngg_get_query_buf(ctx),
660 LLVMConstInt(ctx->i32, 16, false), /* offset of stream[0].generated_primitives */
661 ctx->i32_0, /* soffset */
662 ctx->i32_0, /* cachepolicy */
663 };
664
665 if (sel->so.num_outputs) {
666 args[0] = ac_build_writelane(&ctx->ac, args[0], emitted_prims, ctx->i32_1);
667 args[2] = ac_build_writelane(&ctx->ac, args[2],
668 LLVMConstInt(ctx->i32, 24, false), ctx->i32_1);
669 }
670
671 /* TODO: should this be 64-bit atomics? */
672 ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.raw.buffer.atomic.add.i32",
673 ctx->i32, args, 5, 0);
674 }
675 ac_build_endif(&ctx->ac, 5031);
676 ac_build_endif(&ctx->ac, 5030);
677 ac_build_endif(&ctx->ac, 5029);
678 }
679
680 /* Export primitive data to the index buffer. Format is:
681 * - bits 0..8: index 0
682 * - bit 9: edge flag 0
683 * - bits 10..18: index 1
684 * - bit 19: edge flag 1
685 * - bits 20..28: index 2
686 * - bit 29: edge flag 2
687 * - bit 31: null primitive (skip)
688 *
689 * For the first version, we will always build up all three indices
690 * independent of the primitive type. The additional garbage data
691 * shouldn't hurt.
692 *
693 * TODO: culling depends on the primitive type, so can have some
694 * interaction here.
695 */
696 lp_build_if(&if_state, &ctx->gallivm, is_gs_thread);
697 {
698 struct ngg_prim prim = {};
699
700 prim.num_vertices = num_vertices;
701 prim.isnull = ctx->ac.i1false;
702 memcpy(prim.index, vtxindex, sizeof(vtxindex[0]) * 3);
703
704 for (unsigned i = 0; i < num_vertices; ++i) {
705 tmp = LLVMBuildLShr(builder, ctx->abi.gs_invocation_id,
706 LLVMConstInt(ctx->ac.i32, 8 + i, false), "");
707 prim.edgeflag[i] = LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
708 }
709
710 build_export_prim(ctx, &prim);
711 }
712 lp_build_endif(&if_state);
713
714 /* Export per-vertex data (positions and parameters). */
715 lp_build_if(&if_state, &ctx->gallivm, is_es_thread);
716 {
717 unsigned i;
718
719 /* Unconditionally (re-)load the values for proper SSA form. */
720 for (i = 0; i < info->num_outputs; i++) {
721 for (unsigned j = 0; j < 4; j++) {
722 outputs[i].values[j] =
723 LLVMBuildLoad(builder,
724 addrs[4 * i + j],
725 "");
726 }
727 }
728
729 if (ctx->shader->key.mono.u.vs_export_prim_id) {
730 outputs[i].semantic_name = TGSI_SEMANTIC_PRIMID;
731 outputs[i].semantic_index = 0;
732
733 if (ctx->type == PIPE_SHADER_VERTEX) {
734 /* Wait for GS stores to finish. */
735 ac_build_s_barrier(&ctx->ac);
736
737 tmp = ac_build_gep0(&ctx->ac, ctx->esgs_ring,
738 get_thread_id_in_tg(ctx));
739 outputs[i].values[0] = LLVMBuildLoad(builder, tmp, "");
740 } else {
741 assert(ctx->type == PIPE_SHADER_TESS_EVAL);
742 outputs[i].values[0] = si_get_primitive_id(ctx, 0);
743 }
744
745 outputs[i].values[0] = ac_to_float(&ctx->ac, outputs[i].values[0]);
746 for (unsigned j = 1; j < 4; j++)
747 outputs[i].values[j] = LLVMGetUndef(ctx->f32);
748
749 memset(outputs[i].vertex_stream, 0,
750 sizeof(outputs[i].vertex_stream));
751 i++;
752 }
753
754 si_llvm_export_vs(ctx, outputs, i);
755 }
756 lp_build_endif(&if_state);
757 }
758
759 static LLVMValueRef
760 ngg_gs_get_vertex_storage(struct si_shader_context *ctx)
761 {
762 const struct si_shader_selector *sel = ctx->shader->selector;
763 const struct tgsi_shader_info *info = &sel->info;
764
765 LLVMTypeRef elements[2] = {
766 LLVMArrayType(ctx->ac.i32, 4 * info->num_outputs),
767 LLVMArrayType(ctx->ac.i8, 4),
768 };
769 LLVMTypeRef type = LLVMStructTypeInContext(ctx->ac.context, elements, 2, false);
770 type = LLVMPointerType(LLVMArrayType(type, 0), AC_ADDR_SPACE_LDS);
771 return LLVMBuildBitCast(ctx->ac.builder, ctx->gs_ngg_emit, type, "");
772 }
773
774 /**
775 * Return a pointer to the LDS storage reserved for the N'th vertex, where N
776 * is in emit order; that is:
777 * - during the epilogue, N is the threadidx (relative to the entire threadgroup)
778 * - during vertex emit, i.e. while the API GS shader invocation is running,
779 * N = threadidx * gs_max_out_vertices + emitidx
780 *
781 * Goals of the LDS memory layout:
782 * 1. Eliminate bank conflicts on write for geometry shaders that have all emits
783 * in uniform control flow
784 * 2. Eliminate bank conflicts on read for export if, additionally, there is no
785 * culling
786 * 3. Agnostic to the number of waves (since we don't know it before compiling)
787 * 4. Allow coalescing of LDS instructions (ds_write_b128 etc.)
788 * 5. Avoid wasting memory.
789 *
790 * We use an AoS layout due to point 4 (this also helps point 3). In an AoS
791 * layout, elimination of bank conflicts requires that each vertex occupy an
792 * odd number of dwords. We use the additional dword to store the output stream
793 * index as well as a flag to indicate whether this vertex ends a primitive
794 * for rasterization.
795 *
796 * Swizzling is required to satisfy points 1 and 2 simultaneously.
797 *
798 * Vertices are stored in export order (gsthread * gs_max_out_vertices + emitidx).
799 * Indices are swizzled in groups of 32, which ensures point 1 without
800 * disturbing point 2.
801 *
802 * \return an LDS pointer to type {[N x i32], [4 x i8]}
803 */
804 static LLVMValueRef
805 ngg_gs_vertex_ptr(struct si_shader_context *ctx, LLVMValueRef vertexidx)
806 {
807 struct si_shader_selector *sel = ctx->shader->selector;
808 LLVMBuilderRef builder = ctx->ac.builder;
809 LLVMValueRef storage = ngg_gs_get_vertex_storage(ctx);
810
811 /* gs_max_out_vertices = 2^(write_stride_2exp) * some odd number */
812 unsigned write_stride_2exp = ffs(sel->gs_max_out_vertices) - 1;
813 if (write_stride_2exp) {
814 LLVMValueRef row =
815 LLVMBuildLShr(builder, vertexidx,
816 LLVMConstInt(ctx->ac.i32, 5, false), "");
817 LLVMValueRef swizzle =
818 LLVMBuildAnd(builder, row,
819 LLVMConstInt(ctx->ac.i32, (1u << write_stride_2exp) - 1,
820 false), "");
821 vertexidx = LLVMBuildXor(builder, vertexidx, swizzle, "");
822 }
823
824 return ac_build_gep0(&ctx->ac, storage, vertexidx);
825 }
826
827 static LLVMValueRef
828 ngg_gs_emit_vertex_ptr(struct si_shader_context *ctx, LLVMValueRef gsthread,
829 LLVMValueRef emitidx)
830 {
831 struct si_shader_selector *sel = ctx->shader->selector;
832 LLVMBuilderRef builder = ctx->ac.builder;
833 LLVMValueRef tmp;
834
835 tmp = LLVMConstInt(ctx->ac.i32, sel->gs_max_out_vertices, false);
836 tmp = LLVMBuildMul(builder, tmp, gsthread, "");
837 const LLVMValueRef vertexidx = LLVMBuildAdd(builder, tmp, emitidx, "");
838 return ngg_gs_vertex_ptr(ctx, vertexidx);
839 }
840
841 void gfx10_ngg_gs_emit_vertex(struct si_shader_context *ctx,
842 unsigned stream,
843 LLVMValueRef *addrs)
844 {
845 const struct si_shader_selector *sel = ctx->shader->selector;
846 const struct tgsi_shader_info *info = &sel->info;
847 LLVMBuilderRef builder = ctx->ac.builder;
848 struct lp_build_if_state if_state;
849 LLVMValueRef tmp;
850 const LLVMValueRef vertexidx =
851 LLVMBuildLoad(builder, ctx->gs_next_vertex[stream], "");
852
853 /* If this thread has already emitted the declared maximum number of
854 * vertices, skip the write: excessive vertex emissions are not
855 * supposed to have any effect.
856 */
857 const LLVMValueRef can_emit =
858 LLVMBuildICmp(builder, LLVMIntULT, vertexidx,
859 LLVMConstInt(ctx->i32, sel->gs_max_out_vertices, false), "");
860
861 tmp = LLVMBuildAdd(builder, vertexidx, ctx->ac.i32_1, "");
862 tmp = LLVMBuildSelect(builder, can_emit, tmp, vertexidx, "");
863 LLVMBuildStore(builder, tmp, ctx->gs_next_vertex[stream]);
864
865 lp_build_if(&if_state, &ctx->gallivm, can_emit);
866
867 const LLVMValueRef vertexptr =
868 ngg_gs_emit_vertex_ptr(ctx, get_thread_id_in_tg(ctx), vertexidx);
869 unsigned out_idx = 0;
870 for (unsigned i = 0; i < info->num_outputs; i++) {
871 for (unsigned chan = 0; chan < 4; chan++, out_idx++) {
872 if (!(info->output_usagemask[i] & (1 << chan)) ||
873 ((info->output_streams[i] >> (2 * chan)) & 3) != stream)
874 continue;
875
876 LLVMValueRef out_val = LLVMBuildLoad(builder, addrs[4 * i + chan], "");
877 LLVMValueRef gep_idx[3] = {
878 ctx->ac.i32_0, /* implied C-style array */
879 ctx->ac.i32_0, /* first entry of struct */
880 LLVMConstInt(ctx->ac.i32, out_idx, false),
881 };
882 LLVMValueRef ptr = LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
883
884 out_val = ac_to_integer(&ctx->ac, out_val);
885 LLVMBuildStore(builder, out_val, ptr);
886 }
887 }
888 assert(out_idx * 4 == sel->gsvs_vertex_size);
889
890 /* Determine and store whether this vertex completed a primitive. */
891 const LLVMValueRef curverts = LLVMBuildLoad(builder, ctx->gs_curprim_verts[stream], "");
892
893 tmp = LLVMConstInt(ctx->ac.i32, u_vertices_per_prim(sel->gs_output_prim) - 1, false);
894 const LLVMValueRef iscompleteprim =
895 LLVMBuildICmp(builder, LLVMIntUGE, curverts, tmp, "");
896
897 tmp = LLVMBuildAdd(builder, curverts, ctx->ac.i32_1, "");
898 LLVMBuildStore(builder, tmp, ctx->gs_curprim_verts[stream]);
899
900 LLVMValueRef gep_idx[3] = {
901 ctx->ac.i32_0, /* implied C-style array */
902 ctx->ac.i32_1, /* second struct entry */
903 LLVMConstInt(ctx->ac.i32, stream, false),
904 };
905 const LLVMValueRef primflagptr =
906 LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
907
908 tmp = LLVMBuildZExt(builder, iscompleteprim, ctx->ac.i8, "");
909 LLVMBuildStore(builder, tmp, primflagptr);
910
911 tmp = LLVMBuildLoad(builder, ctx->gs_generated_prims[stream], "");
912 tmp = LLVMBuildAdd(builder, tmp, LLVMBuildZExt(builder, iscompleteprim, ctx->ac.i32, ""), "");
913 LLVMBuildStore(builder, tmp, ctx->gs_generated_prims[stream]);
914
915 lp_build_endif(&if_state);
916 }
917
918 void gfx10_ngg_gs_emit_prologue(struct si_shader_context *ctx)
919 {
920 /* Zero out the part of LDS scratch that is used to accumulate the
921 * per-stream generated primitive count.
922 */
923 LLVMBuilderRef builder = ctx->ac.builder;
924 LLVMValueRef scratchptr = ctx->gs_ngg_scratch;
925 LLVMValueRef tid = get_thread_id_in_tg(ctx);
926 LLVMValueRef tmp;
927
928 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, LLVMConstInt(ctx->i32, 4, false), "");
929 ac_build_ifcc(&ctx->ac, tmp, 5090);
930 {
931 LLVMValueRef ptr = ac_build_gep0(&ctx->ac, scratchptr, tid);
932 LLVMBuildStore(builder, ctx->i32_0, ptr);
933 }
934 ac_build_endif(&ctx->ac, 5090);
935
936 ac_build_s_barrier(&ctx->ac);
937 }
938
939 void gfx10_ngg_gs_emit_epilogue(struct si_shader_context *ctx)
940 {
941 const struct si_shader_selector *sel = ctx->shader->selector;
942 const struct tgsi_shader_info *info = &sel->info;
943 const unsigned verts_per_prim = u_vertices_per_prim(sel->gs_output_prim);
944 LLVMBuilderRef builder = ctx->ac.builder;
945 LLVMValueRef i8_0 = LLVMConstInt(ctx->ac.i8, 0, false);
946 LLVMValueRef tmp, tmp2;
947
948 /* Zero out remaining (non-emitted) primitive flags.
949 *
950 * Note: Alternatively, we could pass the relevant gs_next_vertex to
951 * the emit threads via LDS. This is likely worse in the expected
952 * typical case where each GS thread emits the full set of
953 * vertices.
954 */
955 for (unsigned stream = 0; stream < 4; ++stream) {
956 if (!info->num_stream_output_components[stream])
957 continue;
958
959 const LLVMValueRef gsthread = get_thread_id_in_tg(ctx);
960
961 ac_build_bgnloop(&ctx->ac, 5100);
962
963 const LLVMValueRef vertexidx =
964 LLVMBuildLoad(builder, ctx->gs_next_vertex[stream], "");
965 tmp = LLVMBuildICmp(builder, LLVMIntUGE, vertexidx,
966 LLVMConstInt(ctx->ac.i32, sel->gs_max_out_vertices, false), "");
967 ac_build_ifcc(&ctx->ac, tmp, 5101);
968 ac_build_break(&ctx->ac);
969 ac_build_endif(&ctx->ac, 5101);
970
971 tmp = LLVMBuildAdd(builder, vertexidx, ctx->ac.i32_1, "");
972 LLVMBuildStore(builder, tmp, ctx->gs_next_vertex[stream]);
973
974 tmp = ngg_gs_emit_vertex_ptr(ctx, gsthread, vertexidx);
975 LLVMValueRef gep_idx[3] = {
976 ctx->ac.i32_0, /* implied C-style array */
977 ctx->ac.i32_1, /* second entry of struct */
978 LLVMConstInt(ctx->ac.i32, stream, false),
979 };
980 tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
981 LLVMBuildStore(builder, i8_0, tmp);
982
983 ac_build_endloop(&ctx->ac, 5100);
984 }
985
986 /* Accumulate generated primitives counts across the entire threadgroup. */
987 for (unsigned stream = 0; stream < 4; ++stream) {
988 if (!info->num_stream_output_components[stream])
989 continue;
990
991 LLVMValueRef numprims =
992 LLVMBuildLoad(builder, ctx->gs_generated_prims[stream], "");
993 numprims = ac_build_reduce(&ctx->ac, numprims, nir_op_iadd, 64);
994
995 tmp = LLVMBuildICmp(builder, LLVMIntEQ, ac_get_thread_id(&ctx->ac), ctx->i32_0, "");
996 ac_build_ifcc(&ctx->ac, tmp, 5105);
997 {
998 LLVMBuildAtomicRMW(builder, LLVMAtomicRMWBinOpAdd,
999 ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch,
1000 LLVMConstInt(ctx->i32, stream, false)),
1001 numprims, LLVMAtomicOrderingMonotonic, false);
1002 }
1003 ac_build_endif(&ctx->ac, 5105);
1004 }
1005
1006 lp_build_endif(&ctx->merged_wrap_if_state);
1007
1008 ac_build_s_barrier(&ctx->ac);
1009
1010 const LLVMValueRef tid = get_thread_id_in_tg(ctx);
1011 LLVMValueRef num_emit_threads = ngg_get_prim_cnt(ctx);
1012
1013 /* Streamout */
1014 if (sel->so.num_outputs) {
1015 struct ngg_streamout nggso = {};
1016
1017 nggso.num_vertices = LLVMConstInt(ctx->i32, verts_per_prim, false);
1018
1019 LLVMValueRef vertexptr = ngg_gs_vertex_ptr(ctx, tid);
1020 for (unsigned stream = 0; stream < 4; ++stream) {
1021 if (!info->num_stream_output_components[stream])
1022 continue;
1023
1024 LLVMValueRef gep_idx[3] = {
1025 ctx->i32_0, /* implicit C-style array */
1026 ctx->i32_1, /* second value of struct */
1027 LLVMConstInt(ctx->i32, stream, false),
1028 };
1029 tmp = LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
1030 tmp = LLVMBuildLoad(builder, tmp, "");
1031 tmp = LLVMBuildTrunc(builder, tmp, ctx->i1, "");
1032 tmp2 = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
1033 nggso.prim_enable[stream] = LLVMBuildAnd(builder, tmp, tmp2, "");
1034 }
1035
1036 for (unsigned i = 0; i < verts_per_prim; ++i) {
1037 tmp = LLVMBuildSub(builder, tid,
1038 LLVMConstInt(ctx->i32, verts_per_prim - i - 1, false), "");
1039 tmp = ngg_gs_vertex_ptr(ctx, tmp);
1040 nggso.vertices[i] = ac_build_gep0(&ctx->ac, tmp, ctx->i32_0);
1041 }
1042
1043 build_streamout(ctx, &nggso);
1044 }
1045
1046 /* Write shader query data. */
1047 tmp = si_unpack_param(ctx, ctx->param_vs_state_bits, 6, 1);
1048 tmp = LLVMBuildTrunc(builder, tmp, ctx->i1, "");
1049 ac_build_ifcc(&ctx->ac, tmp, 5109); /* if (STREAMOUT_QUERY_ENABLED) */
1050 unsigned num_query_comps = sel->so.num_outputs ? 8 : 4;
1051 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid,
1052 LLVMConstInt(ctx->i32, num_query_comps, false), "");
1053 ac_build_ifcc(&ctx->ac, tmp, 5110);
1054 {
1055 LLVMValueRef offset;
1056 tmp = tid;
1057 if (sel->so.num_outputs)
1058 tmp = LLVMBuildAnd(builder, tmp, LLVMConstInt(ctx->i32, 3, false), "");
1059 offset = LLVMBuildNUWMul(builder, tmp, LLVMConstInt(ctx->i32, 32, false), "");
1060 if (sel->so.num_outputs) {
1061 tmp = LLVMBuildLShr(builder, tid, LLVMConstInt(ctx->i32, 2, false), "");
1062 tmp = LLVMBuildNUWMul(builder, tmp, LLVMConstInt(ctx->i32, 8, false), "");
1063 offset = LLVMBuildAdd(builder, offset, tmp, "");
1064 }
1065
1066 tmp = LLVMBuildLoad(builder, ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tid), "");
1067 LLVMValueRef args[] = {
1068 tmp,
1069 ngg_get_query_buf(ctx),
1070 offset,
1071 LLVMConstInt(ctx->i32, 16, false), /* soffset */
1072 ctx->i32_0, /* cachepolicy */
1073 };
1074 ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.raw.buffer.atomic.add.i32",
1075 ctx->i32, args, 5, 0);
1076 }
1077 ac_build_endif(&ctx->ac, 5110);
1078 ac_build_endif(&ctx->ac, 5109);
1079
1080 /* TODO: culling */
1081
1082 /* Determine vertex liveness. */
1083 LLVMValueRef vertliveptr = lp_build_alloca(&ctx->gallivm, ctx->ac.i1, "vertexlive");
1084
1085 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
1086 ac_build_ifcc(&ctx->ac, tmp, 5120);
1087 {
1088 for (unsigned i = 0; i < verts_per_prim; ++i) {
1089 const LLVMValueRef primidx =
1090 LLVMBuildAdd(builder, tid,
1091 LLVMConstInt(ctx->ac.i32, i, false), "");
1092
1093 if (i > 0) {
1094 tmp = LLVMBuildICmp(builder, LLVMIntULT, primidx, num_emit_threads, "");
1095 ac_build_ifcc(&ctx->ac, tmp, 5121 + i);
1096 }
1097
1098 /* Load primitive liveness */
1099 tmp = ngg_gs_vertex_ptr(ctx, primidx);
1100 LLVMValueRef gep_idx[3] = {
1101 ctx->ac.i32_0, /* implicit C-style array */
1102 ctx->ac.i32_1, /* second value of struct */
1103 ctx->ac.i32_0, /* stream 0 */
1104 };
1105 tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
1106 tmp = LLVMBuildLoad(builder, tmp, "");
1107 const LLVMValueRef primlive =
1108 LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
1109
1110 tmp = LLVMBuildLoad(builder, vertliveptr, "");
1111 tmp = LLVMBuildOr(builder, tmp, primlive, ""),
1112 LLVMBuildStore(builder, tmp, vertliveptr);
1113
1114 if (i > 0)
1115 ac_build_endif(&ctx->ac, 5121 + i);
1116 }
1117 }
1118 ac_build_endif(&ctx->ac, 5120);
1119
1120 /* Inclusive scan addition across the current wave. */
1121 LLVMValueRef vertlive = LLVMBuildLoad(builder, vertliveptr, "");
1122 struct ac_wg_scan vertlive_scan = {};
1123 vertlive_scan.op = nir_op_iadd;
1124 vertlive_scan.enable_reduce = true;
1125 vertlive_scan.enable_exclusive = true;
1126 vertlive_scan.src = vertlive;
1127 vertlive_scan.scratch = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, ctx->i32_0);
1128 vertlive_scan.waveidx = get_wave_id_in_tg(ctx);
1129 vertlive_scan.numwaves = get_tgsize(ctx);
1130 vertlive_scan.maxwaves = 8;
1131
1132 ac_build_wg_scan(&ctx->ac, &vertlive_scan);
1133
1134 /* Skip all exports (including index exports) when possible. At least on
1135 * early gfx10 revisions this is also to avoid hangs.
1136 */
1137 LLVMValueRef have_exports =
1138 LLVMBuildICmp(builder, LLVMIntNE, vertlive_scan.result_reduce, ctx->ac.i32_0, "");
1139 num_emit_threads =
1140 LLVMBuildSelect(builder, have_exports, num_emit_threads, ctx->ac.i32_0, "");
1141
1142 /* Allocate export space. Send this message as early as possible, to
1143 * hide the latency of the SQ <-> SPI roundtrip.
1144 *
1145 * Note: We could consider compacting primitives for export as well.
1146 * PA processes 1 non-null prim / clock, but it fetches 4 DW of
1147 * prim data per clock and skips null primitives at no additional
1148 * cost. So compacting primitives can only be beneficial when
1149 * there are 4 or more contiguous null primitives in the export
1150 * (in the common case of single-dword prim exports).
1151 */
1152 build_sendmsg_gs_alloc_req(ctx, vertlive_scan.result_reduce, num_emit_threads);
1153
1154 /* Setup the reverse vertex compaction permutation. We re-use stream 1
1155 * of the primitive liveness flags, relying on the fact that each
1156 * threadgroup can have at most 256 threads. */
1157 ac_build_ifcc(&ctx->ac, vertlive, 5130);
1158 {
1159 tmp = ngg_gs_vertex_ptr(ctx, vertlive_scan.result_exclusive);
1160 LLVMValueRef gep_idx[3] = {
1161 ctx->ac.i32_0, /* implicit C-style array */
1162 ctx->ac.i32_1, /* second value of struct */
1163 ctx->ac.i32_1, /* stream 1 */
1164 };
1165 tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
1166 tmp2 = LLVMBuildTrunc(builder, tid, ctx->ac.i8, "");
1167 LLVMBuildStore(builder, tmp2, tmp);
1168 }
1169 ac_build_endif(&ctx->ac, 5130);
1170
1171 ac_build_s_barrier(&ctx->ac);
1172
1173 /* Export primitive data */
1174 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
1175 ac_build_ifcc(&ctx->ac, tmp, 5140);
1176 {
1177 struct ngg_prim prim = {};
1178 prim.num_vertices = verts_per_prim;
1179
1180 tmp = ngg_gs_vertex_ptr(ctx, tid);
1181 LLVMValueRef gep_idx[3] = {
1182 ctx->ac.i32_0, /* implicit C-style array */
1183 ctx->ac.i32_1, /* second value of struct */
1184 ctx->ac.i32_0, /* primflag */
1185 };
1186 tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
1187 tmp = LLVMBuildLoad(builder, tmp, "");
1188 prim.isnull = LLVMBuildICmp(builder, LLVMIntEQ, tmp,
1189 LLVMConstInt(ctx->ac.i8, 0, false), "");
1190
1191 for (unsigned i = 0; i < verts_per_prim; ++i) {
1192 prim.index[i] = LLVMBuildSub(builder, vertlive_scan.result_exclusive,
1193 LLVMConstInt(ctx->ac.i32, verts_per_prim - i - 1, false), "");
1194 prim.edgeflag[i] = ctx->ac.i1false;
1195 }
1196
1197 build_export_prim(ctx, &prim);
1198 }
1199 ac_build_endif(&ctx->ac, 5140);
1200
1201 /* Export position and parameter data */
1202 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, vertlive_scan.result_reduce, "");
1203 ac_build_ifcc(&ctx->ac, tmp, 5145);
1204 {
1205 struct si_shader_output_values outputs[PIPE_MAX_SHADER_OUTPUTS];
1206
1207 tmp = ngg_gs_vertex_ptr(ctx, tid);
1208 LLVMValueRef gep_idx[3] = {
1209 ctx->ac.i32_0, /* implicit C-style array */
1210 ctx->ac.i32_1, /* second value of struct */
1211 ctx->ac.i32_1, /* stream 1: source data index */
1212 };
1213 tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
1214 tmp = LLVMBuildLoad(builder, tmp, "");
1215 tmp = LLVMBuildZExt(builder, tmp, ctx->ac.i32, "");
1216 const LLVMValueRef vertexptr = ngg_gs_vertex_ptr(ctx, tmp);
1217
1218 unsigned out_idx = 0;
1219 gep_idx[1] = ctx->ac.i32_0;
1220 for (unsigned i = 0; i < info->num_outputs; i++) {
1221 outputs[i].semantic_name = info->output_semantic_name[i];
1222 outputs[i].semantic_index = info->output_semantic_index[i];
1223
1224 for (unsigned j = 0; j < 4; j++, out_idx++) {
1225 gep_idx[2] = LLVMConstInt(ctx->ac.i32, out_idx, false);
1226 tmp = LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
1227 tmp = LLVMBuildLoad(builder, tmp, "");
1228 outputs[i].values[j] = ac_to_float(&ctx->ac, tmp);
1229 outputs[i].vertex_stream[j] =
1230 (info->output_streams[i] >> (2 * j)) & 3;
1231 }
1232 }
1233
1234 si_llvm_export_vs(ctx, outputs, info->num_outputs);
1235 }
1236 ac_build_endif(&ctx->ac, 5145);
1237 }
1238
1239 static void clamp_gsprims_to_esverts(unsigned *max_gsprims, unsigned max_esverts,
1240 unsigned min_verts_per_prim, bool use_adjacency)
1241 {
1242 unsigned max_reuse = max_esverts - min_verts_per_prim;
1243 if (use_adjacency)
1244 max_reuse /= 2;
1245 *max_gsprims = MIN2(*max_gsprims, 1 + max_reuse);
1246 }
1247
1248 /**
1249 * Determine subgroup information like maximum number of vertices and prims.
1250 *
1251 * This happens before the shader is uploaded, since LDS relocations during
1252 * upload depend on the subgroup size.
1253 */
1254 void gfx10_ngg_calculate_subgroup_info(struct si_shader *shader)
1255 {
1256 const struct si_shader_selector *gs_sel = shader->selector;
1257 const struct si_shader_selector *es_sel =
1258 shader->previous_stage_sel ? shader->previous_stage_sel : gs_sel;
1259 const enum pipe_shader_type gs_type = gs_sel->type;
1260 const unsigned gs_num_invocations = MAX2(gs_sel->gs_num_invocations, 1);
1261 /* TODO: Use QUADS as the worst case because of reuse, but triangles
1262 * will always have 1 additional unoccupied vector lane. We could use
1263 * that lane if the worst case was TRIANGLES. */
1264 const unsigned input_prim = si_get_input_prim(gs_sel, PIPE_PRIM_QUADS);
1265 const bool use_adjacency = input_prim >= PIPE_PRIM_LINES_ADJACENCY &&
1266 input_prim <= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY;
1267 const unsigned max_verts_per_prim = u_vertices_per_prim(input_prim);
1268 const unsigned min_verts_per_prim =
1269 gs_type == PIPE_SHADER_GEOMETRY ? max_verts_per_prim : 1;
1270
1271 /* All these are in dwords: */
1272 /* We can't allow using the whole LDS, because GS waves compete with
1273 * other shader stages for LDS space.
1274 *
1275 * TODO: We should really take the shader's internal LDS use into
1276 * account. The linker will fail if the size is greater than
1277 * 8K dwords.
1278 */
1279 const unsigned max_lds_size = 8 * 1024 - 768;
1280 const unsigned target_lds_size = max_lds_size;
1281 unsigned esvert_lds_size = 0;
1282 unsigned gsprim_lds_size = 0;
1283
1284 /* All these are per subgroup: */
1285 bool max_vert_out_per_gs_instance = false;
1286 unsigned max_esverts_base = 128;
1287 unsigned max_gsprims_base = 128; /* default prim group size clamp */
1288
1289 /* Hardware has the following non-natural restrictions on the value
1290 * of GE_CNTL.VERT_GRP_SIZE based on based on the primitive type of
1291 * the draw:
1292 * - at most 252 for any line input primitive type
1293 * - at most 251 for any quad input primitive type
1294 * - at most 251 for triangle strips with adjacency (this happens to
1295 * be the natural limit for triangle *lists* with adjacency)
1296 */
1297 max_esverts_base = MIN2(max_esverts_base, 251 + max_verts_per_prim - 1);
1298
1299 if (gs_type == PIPE_SHADER_GEOMETRY) {
1300 unsigned max_out_verts_per_gsprim =
1301 gs_sel->gs_max_out_vertices * gs_num_invocations;
1302
1303 if (max_out_verts_per_gsprim <= 256) {
1304 if (max_out_verts_per_gsprim) {
1305 max_gsprims_base = MIN2(max_gsprims_base,
1306 256 / max_out_verts_per_gsprim);
1307 }
1308 } else {
1309 /* Use special multi-cycling mode in which each GS
1310 * instance gets its own subgroup. Does not work with
1311 * tessellation. */
1312 max_vert_out_per_gs_instance = true;
1313 max_gsprims_base = 1;
1314 max_out_verts_per_gsprim = gs_sel->gs_max_out_vertices;
1315 }
1316
1317 esvert_lds_size = es_sel->esgs_itemsize / 4;
1318 gsprim_lds_size = (gs_sel->gsvs_vertex_size / 4 + 1) * max_out_verts_per_gsprim;
1319 } else {
1320 /* TODO: This needs to be adjusted once LDS use for compaction
1321 * after culling is implemented. */
1322 if (es_sel->so.num_outputs)
1323 esvert_lds_size = 4 * es_sel->info.num_outputs + 1;
1324
1325 /* GS stores Primitive IDs into LDS at the address corresponding
1326 * to the ES thread of the provoking vertex. All ES threads
1327 * load and export PrimitiveID for their thread.
1328 */
1329 if (gs_sel->type == PIPE_SHADER_VERTEX &&
1330 shader->key.mono.u.vs_export_prim_id)
1331 esvert_lds_size = MAX2(esvert_lds_size, 1);
1332 }
1333
1334 unsigned max_gsprims = max_gsprims_base;
1335 unsigned max_esverts = max_esverts_base;
1336
1337 if (esvert_lds_size)
1338 max_esverts = MIN2(max_esverts, target_lds_size / esvert_lds_size);
1339 if (gsprim_lds_size)
1340 max_gsprims = MIN2(max_gsprims, target_lds_size / gsprim_lds_size);
1341
1342 max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim);
1343 clamp_gsprims_to_esverts(&max_gsprims, max_esverts, min_verts_per_prim, use_adjacency);
1344 assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1);
1345
1346 if (esvert_lds_size || gsprim_lds_size) {
1347 /* Now that we have a rough proportionality between esverts
1348 * and gsprims based on the primitive type, scale both of them
1349 * down simultaneously based on required LDS space.
1350 *
1351 * We could be smarter about this if we knew how much vertex
1352 * reuse to expect.
1353 */
1354 unsigned lds_total = max_esverts * esvert_lds_size +
1355 max_gsprims * gsprim_lds_size;
1356 if (lds_total > target_lds_size) {
1357 max_esverts = max_esverts * target_lds_size / lds_total;
1358 max_gsprims = max_gsprims * target_lds_size / lds_total;
1359
1360 max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim);
1361 clamp_gsprims_to_esverts(&max_gsprims, max_esverts,
1362 min_verts_per_prim, use_adjacency);
1363 assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1);
1364 }
1365 }
1366
1367 /* Round up towards full wave sizes for better ALU utilization. */
1368 if (!max_vert_out_per_gs_instance) {
1369 const unsigned wavesize = 64;
1370 unsigned orig_max_esverts;
1371 unsigned orig_max_gsprims;
1372 do {
1373 orig_max_esverts = max_esverts;
1374 orig_max_gsprims = max_gsprims;
1375
1376 max_esverts = align(max_esverts, wavesize);
1377 max_esverts = MIN2(max_esverts, max_esverts_base);
1378 if (esvert_lds_size)
1379 max_esverts = MIN2(max_esverts,
1380 (max_lds_size - max_gsprims * gsprim_lds_size) /
1381 esvert_lds_size);
1382 max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim);
1383
1384 max_gsprims = align(max_gsprims, wavesize);
1385 max_gsprims = MIN2(max_gsprims, max_gsprims_base);
1386 if (gsprim_lds_size)
1387 max_gsprims = MIN2(max_gsprims,
1388 (max_lds_size - max_esverts * esvert_lds_size) /
1389 gsprim_lds_size);
1390 clamp_gsprims_to_esverts(&max_gsprims, max_esverts,
1391 min_verts_per_prim, use_adjacency);
1392 assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1);
1393 } while (orig_max_esverts != max_esverts || orig_max_gsprims != max_gsprims);
1394 }
1395
1396 /* Hardware restriction: minimum value of max_esverts */
1397 max_esverts = MAX2(max_esverts, 23 + max_verts_per_prim);
1398
1399 unsigned max_out_vertices =
1400 max_vert_out_per_gs_instance ? gs_sel->gs_max_out_vertices :
1401 gs_type == PIPE_SHADER_GEOMETRY ?
1402 max_gsprims * gs_num_invocations * gs_sel->gs_max_out_vertices :
1403 max_esverts;
1404 assert(max_out_vertices <= 256);
1405
1406 unsigned prim_amp_factor = 1;
1407 if (gs_type == PIPE_SHADER_GEOMETRY) {
1408 /* Number of output primitives per GS input primitive after
1409 * GS instancing. */
1410 prim_amp_factor = gs_sel->gs_max_out_vertices;
1411 }
1412
1413 /* The GE only checks against the maximum number of ES verts after
1414 * allocating a full GS primitive. So we need to ensure that whenever
1415 * this check passes, there is enough space for a full primitive without
1416 * vertex reuse.
1417 */
1418 shader->ngg.hw_max_esverts = max_esverts - max_verts_per_prim + 1;
1419 shader->ngg.max_gsprims = max_gsprims;
1420 shader->ngg.max_out_verts = max_out_vertices;
1421 shader->ngg.prim_amp_factor = prim_amp_factor;
1422 shader->ngg.max_vert_out_per_gs_instance = max_vert_out_per_gs_instance;
1423
1424 shader->gs_info.esgs_ring_size = 4 * max_esverts * esvert_lds_size;
1425 shader->ngg.ngg_emit_size = max_gsprims * gsprim_lds_size;
1426
1427 assert(shader->ngg.hw_max_esverts >= 24); /* HW limitation */
1428 }