projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radeonsi/gfx10: don't do the query buffer atomic for blit shaders
[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 = NULL;
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 outputs = MALLOC((info->num_outputs + 1) * sizeof(outputs[0]));
533
534 LLVMValueRef vertex_ptr = NULL;
535
536 if (sel->so.num_outputs)
537 vertex_ptr = ngg_nogs_vertex_ptr(ctx, get_thread_id_in_tg(ctx));
538
539 for (unsigned i = 0; i < info->num_outputs; i++) {
540 outputs[i].semantic_name = info->output_semantic_name[i];
541 outputs[i].semantic_index = info->output_semantic_index[i];
542
543 /* This is used only by streamout. */
544 for (unsigned j = 0; j < 4; j++) {
545 outputs[i].values[j] =
546 LLVMBuildLoad(builder,
547 addrs[4 * i + j],
548 "");
549 outputs[i].vertex_stream[j] =
550 (info->output_streams[i] >> (2 * j)) & 3;
551
552 if (vertex_ptr) {
553 tmp = ac_build_gep0(&ctx->ac, vertex_ptr,
554 LLVMConstInt(ctx->i32, 4 * i + j, false));
555 tmp2 = ac_to_integer(&ctx->ac, outputs[i].values[j]);
556 LLVMBuildStore(builder, tmp2, tmp);
557 }
558 }
559 }
560
561 lp_build_endif(&ctx->merged_wrap_if_state);
562
563 LLVMValueRef prims_in_wave = si_unpack_param(ctx, ctx->param_merged_wave_info, 8, 8);
564 LLVMValueRef vtx_in_wave = si_unpack_param(ctx, ctx->param_merged_wave_info, 0, 8);
565 LLVMValueRef is_gs_thread = LLVMBuildICmp(builder, LLVMIntULT,
566 ac_get_thread_id(&ctx->ac), prims_in_wave, "");
567 LLVMValueRef is_es_thread = LLVMBuildICmp(builder, LLVMIntULT,
568 ac_get_thread_id(&ctx->ac), vtx_in_wave, "");
569 LLVMValueRef vtxindex[] = {
570 si_unpack_param(ctx, ctx->param_gs_vtx01_offset, 0, 16),
571 si_unpack_param(ctx, ctx->param_gs_vtx01_offset, 16, 16),
572 si_unpack_param(ctx, ctx->param_gs_vtx23_offset, 0, 16),
573 };
574
575 /* Determine the number of vertices per primitive. */
576 unsigned num_vertices;
577 LLVMValueRef num_vertices_val;
578
579 if (ctx->type == PIPE_SHADER_VERTEX) {
580 if (info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS]) {
581 /* Blits always use axis-aligned rectangles with 3 vertices. */
582 num_vertices = 3;
583 num_vertices_val = LLVMConstInt(ctx->i32, 3, 0);
584 } else {
585 /* Extract OUTPRIM field. */
586 tmp = si_unpack_param(ctx, ctx->param_vs_state_bits, 2, 2);
587 num_vertices_val = LLVMBuildAdd(builder, tmp, ctx->i32_1, "");
588 num_vertices = 3; /* TODO: optimize for points & lines */
589 }
590 } else {
591 assert(ctx->type == PIPE_SHADER_TESS_EVAL);
592
593 if (info->properties[TGSI_PROPERTY_TES_POINT_MODE])
594 num_vertices = 1;
595 else if (info->properties[TGSI_PROPERTY_TES_PRIM_MODE] == PIPE_PRIM_LINES)
596 num_vertices = 2;
597 else
598 num_vertices = 3;
599
600 num_vertices_val = LLVMConstInt(ctx->i32, num_vertices, false);
601 }
602
603 /* Streamout */
604 LLVMValueRef emitted_prims = NULL;
605
606 if (sel->so.num_outputs) {
607 struct ngg_streamout nggso = {};
608
609 nggso.num_vertices = num_vertices_val;
610 nggso.prim_enable[0] = is_gs_thread;
611
612 for (unsigned i = 0; i < num_vertices; ++i)
613 nggso.vertices[i] = ngg_nogs_vertex_ptr(ctx, vtxindex[i]);
614
615 build_streamout(ctx, &nggso);
616 emitted_prims = nggso.emit[0];
617 }
618
619 /* TODO: primitive culling */
620
621 build_sendmsg_gs_alloc_req(ctx, ngg_get_vtx_cnt(ctx), ngg_get_prim_cnt(ctx));
622
623 /* Update query buffer */
624 /* TODO: this won't catch 96-bit clear_buffer via transform feedback. */
625 if (!info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS]) {
626 tmp = LLVMBuildICmp(builder, LLVMIntEQ, get_wave_id_in_tg(ctx), ctx->ac.i32_0, "");
627 ac_build_ifcc(&ctx->ac, tmp, 5030);
628 tmp = LLVMBuildICmp(builder, LLVMIntULE, ac_get_thread_id(&ctx->ac),
629 sel->so.num_outputs ? ctx->ac.i32_1 : ctx->ac.i32_0, "");
630 ac_build_ifcc(&ctx->ac, tmp, 5031);
631 {
632 LLVMValueRef args[] = {
633 ngg_get_prim_cnt(ctx),
634 ngg_get_query_buf(ctx),
635 LLVMConstInt(ctx->i32, 16, false), /* offset of stream[0].generated_primitives */
636 ctx->i32_0, /* soffset */
637 ctx->i32_0, /* cachepolicy */
638 };
639
640 if (sel->so.num_outputs) {
641 args[0] = ac_build_writelane(&ctx->ac, args[0], emitted_prims, ctx->i32_1);
642 args[2] = ac_build_writelane(&ctx->ac, args[2],
643 LLVMConstInt(ctx->i32, 24, false), ctx->i32_1);
644 }
645
646 /* TODO: should this be 64-bit atomics? */
647 ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.raw.buffer.atomic.add.i32",
648 ctx->i32, args, 5, 0);
649 }
650 ac_build_endif(&ctx->ac, 5031);
651 ac_build_endif(&ctx->ac, 5030);
652 }
653
654 /* Export primitive data to the index buffer. Format is:
655 * - bits 0..8: index 0
656 * - bit 9: edge flag 0
657 * - bits 10..18: index 1
658 * - bit 19: edge flag 1
659 * - bits 20..28: index 2
660 * - bit 29: edge flag 2
661 * - bit 31: null primitive (skip)
662 *
663 * For the first version, we will always build up all three indices
664 * independent of the primitive type. The additional garbage data
665 * shouldn't hurt.
666 *
667 * TODO: culling depends on the primitive type, so can have some
668 * interaction here.
669 */
670 lp_build_if(&if_state, &ctx->gallivm, is_gs_thread);
671 {
672 struct ngg_prim prim = {};
673
674 prim.num_vertices = num_vertices;
675 prim.isnull = ctx->ac.i1false;
676 memcpy(prim.index, vtxindex, sizeof(vtxindex[0]) * 3);
677
678 for (unsigned i = 0; i < num_vertices; ++i) {
679 tmp = LLVMBuildLShr(builder, ctx->abi.gs_invocation_id,
680 LLVMConstInt(ctx->ac.i32, 8 + i, false), "");
681 prim.edgeflag[i] = LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
682 }
683
684 build_export_prim(ctx, &prim);
685 }
686 lp_build_endif(&if_state);
687
688 /* Export per-vertex data (positions and parameters). */
689 lp_build_if(&if_state, &ctx->gallivm, is_es_thread);
690 {
691 unsigned i;
692
693 /* Unconditionally (re-)load the values for proper SSA form. */
694 for (i = 0; i < info->num_outputs; i++) {
695 for (unsigned j = 0; j < 4; j++) {
696 outputs[i].values[j] =
697 LLVMBuildLoad(builder,
698 addrs[4 * i + j],
699 "");
700 }
701 }
702
703 /* TODO: Vertex shaders have to get PrimitiveID from GS VGPRs. */
704 if (ctx->type == PIPE_SHADER_TESS_EVAL &&
705 ctx->shader->key.mono.u.vs_export_prim_id) {
706 outputs[i].semantic_name = TGSI_SEMANTIC_PRIMID;
707 outputs[i].semantic_index = 0;
708 outputs[i].values[0] = ac_to_float(&ctx->ac, si_get_primitive_id(ctx, 0));
709 for (unsigned j = 1; j < 4; j++)
710 outputs[i].values[j] = LLVMGetUndef(ctx->f32);
711
712 memset(outputs[i].vertex_stream, 0,
713 sizeof(outputs[i].vertex_stream));
714 i++;
715 }
716
717 si_llvm_export_vs(ctx, outputs, i);
718 }
719 lp_build_endif(&if_state);
720
721 FREE(outputs);
722 }
723
724 static LLVMValueRef
725 ngg_gs_get_vertex_storage(struct si_shader_context *ctx)
726 {
727 const struct si_shader_selector *sel = ctx->shader->selector;
728 const struct tgsi_shader_info *info = &sel->info;
729
730 LLVMTypeRef elements[2] = {
731 LLVMArrayType(ctx->ac.i32, 4 * info->num_outputs),
732 LLVMArrayType(ctx->ac.i8, 4),
733 };
734 LLVMTypeRef type = LLVMStructTypeInContext(ctx->ac.context, elements, 2, false);
735 type = LLVMPointerType(LLVMArrayType(type, 0), AC_ADDR_SPACE_LDS);
736 return LLVMBuildBitCast(ctx->ac.builder, ctx->gs_ngg_emit, type, "");
737 }
738
739 /**
740 * Return a pointer to the LDS storage reserved for the N'th vertex, where N
741 * is in emit order; that is:
742 * - during the epilogue, N is the threadidx (relative to the entire threadgroup)
743 * - during vertex emit, i.e. while the API GS shader invocation is running,
744 * N = threadidx * gs_max_out_vertices + emitidx
745 *
746 * Goals of the LDS memory layout:
747 * 1. Eliminate bank conflicts on write for geometry shaders that have all emits
748 * in uniform control flow
749 * 2. Eliminate bank conflicts on read for export if, additionally, there is no
750 * culling
751 * 3. Agnostic to the number of waves (since we don't know it before compiling)
752 * 4. Allow coalescing of LDS instructions (ds_write_b128 etc.)
753 * 5. Avoid wasting memory.
754 *
755 * We use an AoS layout due to point 4 (this also helps point 3). In an AoS
756 * layout, elimination of bank conflicts requires that each vertex occupy an
757 * odd number of dwords. We use the additional dword to store the output stream
758 * index as well as a flag to indicate whether this vertex ends a primitive
759 * for rasterization.
760 *
761 * Swizzling is required to satisfy points 1 and 2 simultaneously.
762 *
763 * Vertices are stored in export order (gsthread * gs_max_out_vertices + emitidx).
764 * Indices are swizzled in groups of 32, which ensures point 1 without
765 * disturbing point 2.
766 *
767 * \return an LDS pointer to type {[N x i32], [4 x i8]}
768 */
769 static LLVMValueRef
770 ngg_gs_vertex_ptr(struct si_shader_context *ctx, LLVMValueRef vertexidx)
771 {
772 struct si_shader_selector *sel = ctx->shader->selector;
773 LLVMBuilderRef builder = ctx->ac.builder;
774 LLVMValueRef storage = ngg_gs_get_vertex_storage(ctx);
775
776 /* gs_max_out_vertices = 2^(write_stride_2exp) * some odd number */
777 unsigned write_stride_2exp = ffs(sel->gs_max_out_vertices) - 1;
778 if (write_stride_2exp) {
779 LLVMValueRef row =
780 LLVMBuildLShr(builder, vertexidx,
781 LLVMConstInt(ctx->ac.i32, 5, false), "");
782 LLVMValueRef swizzle =
783 LLVMBuildAnd(builder, row,
784 LLVMConstInt(ctx->ac.i32, (1u << write_stride_2exp) - 1,
785 false), "");
786 vertexidx = LLVMBuildXor(builder, vertexidx, swizzle, "");
787 }
788
789 return ac_build_gep0(&ctx->ac, storage, vertexidx);
790 }
791
792 static LLVMValueRef
793 ngg_gs_emit_vertex_ptr(struct si_shader_context *ctx, LLVMValueRef gsthread,
794 LLVMValueRef emitidx)
795 {
796 struct si_shader_selector *sel = ctx->shader->selector;
797 LLVMBuilderRef builder = ctx->ac.builder;
798 LLVMValueRef tmp;
799
800 tmp = LLVMConstInt(ctx->ac.i32, sel->gs_max_out_vertices, false);
801 tmp = LLVMBuildMul(builder, tmp, gsthread, "");
802 const LLVMValueRef vertexidx = LLVMBuildAdd(builder, tmp, emitidx, "");
803 return ngg_gs_vertex_ptr(ctx, vertexidx);
804 }
805
806 void gfx10_ngg_gs_emit_vertex(struct si_shader_context *ctx,
807 unsigned stream,
808 LLVMValueRef *addrs)
809 {
810 const struct si_shader_selector *sel = ctx->shader->selector;
811 const struct tgsi_shader_info *info = &sel->info;
812 LLVMBuilderRef builder = ctx->ac.builder;
813 struct lp_build_if_state if_state;
814 LLVMValueRef tmp;
815 const LLVMValueRef vertexidx =
816 LLVMBuildLoad(builder, ctx->gs_next_vertex[stream], "");
817
818 /* If this thread has already emitted the declared maximum number of
819 * vertices, skip the write: excessive vertex emissions are not
820 * supposed to have any effect.
821 */
822 const LLVMValueRef can_emit =
823 LLVMBuildICmp(builder, LLVMIntULT, vertexidx,
824 LLVMConstInt(ctx->i32, sel->gs_max_out_vertices, false), "");
825
826 tmp = LLVMBuildAdd(builder, vertexidx, ctx->ac.i32_1, "");
827 tmp = LLVMBuildSelect(builder, can_emit, tmp, vertexidx, "");
828 LLVMBuildStore(builder, tmp, ctx->gs_next_vertex[stream]);
829
830 lp_build_if(&if_state, &ctx->gallivm, can_emit);
831
832 const LLVMValueRef vertexptr =
833 ngg_gs_emit_vertex_ptr(ctx, get_thread_id_in_tg(ctx), vertexidx);
834 unsigned out_idx = 0;
835 for (unsigned i = 0; i < info->num_outputs; i++) {
836 for (unsigned chan = 0; chan < 4; chan++, out_idx++) {
837 if (!(info->output_usagemask[i] & (1 << chan)) ||
838 ((info->output_streams[i] >> (2 * chan)) & 3) != stream)
839 continue;
840
841 LLVMValueRef out_val = LLVMBuildLoad(builder, addrs[4 * i + chan], "");
842 LLVMValueRef gep_idx[3] = {
843 ctx->ac.i32_0, /* implied C-style array */
844 ctx->ac.i32_0, /* first entry of struct */
845 LLVMConstInt(ctx->ac.i32, out_idx, false),
846 };
847 LLVMValueRef ptr = LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
848
849 out_val = ac_to_integer(&ctx->ac, out_val);
850 LLVMBuildStore(builder, out_val, ptr);
851 }
852 }
853 assert(out_idx * 4 == sel->gsvs_vertex_size);
854
855 /* Determine and store whether this vertex completed a primitive. */
856 const LLVMValueRef curverts = LLVMBuildLoad(builder, ctx->gs_curprim_verts[stream], "");
857
858 tmp = LLVMConstInt(ctx->ac.i32, u_vertices_per_prim(sel->gs_output_prim) - 1, false);
859 const LLVMValueRef iscompleteprim =
860 LLVMBuildICmp(builder, LLVMIntUGE, curverts, tmp, "");
861
862 tmp = LLVMBuildAdd(builder, curverts, ctx->ac.i32_1, "");
863 LLVMBuildStore(builder, tmp, ctx->gs_curprim_verts[stream]);
864
865 LLVMValueRef gep_idx[3] = {
866 ctx->ac.i32_0, /* implied C-style array */
867 ctx->ac.i32_1, /* second struct entry */
868 LLVMConstInt(ctx->ac.i32, stream, false),
869 };
870 const LLVMValueRef primflagptr =
871 LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
872
873 tmp = LLVMBuildZExt(builder, iscompleteprim, ctx->ac.i8, "");
874 LLVMBuildStore(builder, tmp, primflagptr);
875
876 tmp = LLVMBuildLoad(builder, ctx->gs_generated_prims[stream], "");
877 tmp = LLVMBuildAdd(builder, tmp, LLVMBuildZExt(builder, iscompleteprim, ctx->ac.i32, ""), "");
878 LLVMBuildStore(builder, tmp, ctx->gs_generated_prims[stream]);
879
880 lp_build_endif(&if_state);
881 }
882
883 void gfx10_ngg_gs_emit_prologue(struct si_shader_context *ctx)
884 {
885 /* Zero out the part of LDS scratch that is used to accumulate the
886 * per-stream generated primitive count.
887 */
888 LLVMBuilderRef builder = ctx->ac.builder;
889 LLVMValueRef scratchptr = ctx->gs_ngg_scratch;
890 LLVMValueRef tid = get_thread_id_in_tg(ctx);
891 LLVMValueRef tmp;
892
893 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, LLVMConstInt(ctx->i32, 4, false), "");
894 ac_build_ifcc(&ctx->ac, tmp, 5090);
895 {
896 LLVMValueRef ptr = ac_build_gep0(&ctx->ac, scratchptr, tid);
897 LLVMBuildStore(builder, ctx->i32_0, ptr);
898 }
899 ac_build_endif(&ctx->ac, 5090);
900
901 ac_build_s_barrier(&ctx->ac);
902 }
903
904 void gfx10_ngg_gs_emit_epilogue(struct si_shader_context *ctx)
905 {
906 const struct si_shader_selector *sel = ctx->shader->selector;
907 const struct tgsi_shader_info *info = &sel->info;
908 const unsigned verts_per_prim = u_vertices_per_prim(sel->gs_output_prim);
909 LLVMBuilderRef builder = ctx->ac.builder;
910 LLVMValueRef i8_0 = LLVMConstInt(ctx->ac.i8, 0, false);
911 LLVMValueRef tmp, tmp2;
912
913 /* Zero out remaining (non-emitted) primitive flags.
914 *
915 * Note: Alternatively, we could pass the relevant gs_next_vertex to
916 * the emit threads via LDS. This is likely worse in the expected
917 * typical case where each GS thread emits the full set of
918 * vertices.
919 */
920 for (unsigned stream = 0; stream < 4; ++stream) {
921 if (!info->num_stream_output_components[stream])
922 continue;
923
924 const LLVMValueRef gsthread = get_thread_id_in_tg(ctx);
925
926 ac_build_bgnloop(&ctx->ac, 5100);
927
928 const LLVMValueRef vertexidx =
929 LLVMBuildLoad(builder, ctx->gs_next_vertex[stream], "");
930 tmp = LLVMBuildICmp(builder, LLVMIntUGE, vertexidx,
931 LLVMConstInt(ctx->ac.i32, sel->gs_max_out_vertices, false), "");
932 ac_build_ifcc(&ctx->ac, tmp, 5101);
933 ac_build_break(&ctx->ac);
934 ac_build_endif(&ctx->ac, 5101);
935
936 tmp = LLVMBuildAdd(builder, vertexidx, ctx->ac.i32_1, "");
937 LLVMBuildStore(builder, tmp, ctx->gs_next_vertex[stream]);
938
939 tmp = ngg_gs_emit_vertex_ptr(ctx, gsthread, vertexidx);
940 LLVMValueRef gep_idx[3] = {
941 ctx->ac.i32_0, /* implied C-style array */
942 ctx->ac.i32_1, /* second entry of struct */
943 LLVMConstInt(ctx->ac.i32, stream, false),
944 };
945 tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
946 LLVMBuildStore(builder, i8_0, tmp);
947
948 ac_build_endloop(&ctx->ac, 5100);
949 }
950
951 /* Accumulate generated primitives counts across the entire threadgroup. */
952 for (unsigned stream = 0; stream < 4; ++stream) {
953 if (!info->num_stream_output_components[stream])
954 continue;
955
956 LLVMValueRef numprims =
957 LLVMBuildLoad(builder, ctx->gs_generated_prims[stream], "");
958 numprims = ac_build_reduce(&ctx->ac, numprims, nir_op_iadd, 64);
959
960 tmp = LLVMBuildICmp(builder, LLVMIntEQ, ac_get_thread_id(&ctx->ac), ctx->i32_0, "");
961 ac_build_ifcc(&ctx->ac, tmp, 5105);
962 {
963 LLVMBuildAtomicRMW(builder, LLVMAtomicRMWBinOpAdd,
964 ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch,
965 LLVMConstInt(ctx->i32, stream, false)),
966 numprims, LLVMAtomicOrderingMonotonic, false);
967 }
968 ac_build_endif(&ctx->ac, 5105);
969 }
970
971 lp_build_endif(&ctx->merged_wrap_if_state);
972
973 ac_build_s_barrier(&ctx->ac);
974
975 const LLVMValueRef tid = get_thread_id_in_tg(ctx);
976 LLVMValueRef num_emit_threads = ngg_get_prim_cnt(ctx);
977
978 /* Streamout */
979 if (sel->so.num_outputs) {
980 struct ngg_streamout nggso = {};
981
982 nggso.num_vertices = LLVMConstInt(ctx->i32, verts_per_prim, false);
983
984 LLVMValueRef vertexptr = ngg_gs_vertex_ptr(ctx, tid);
985 for (unsigned stream = 0; stream < 4; ++stream) {
986 if (!info->num_stream_output_components[stream])
987 continue;
988
989 LLVMValueRef gep_idx[3] = {
990 ctx->i32_0, /* implicit C-style array */
991 ctx->i32_1, /* second value of struct */
992 LLVMConstInt(ctx->i32, stream, false),
993 };
994 tmp = LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
995 tmp = LLVMBuildLoad(builder, tmp, "");
996 tmp = LLVMBuildTrunc(builder, tmp, ctx->i1, "");
997 tmp2 = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
998 nggso.prim_enable[stream] = LLVMBuildAnd(builder, tmp, tmp2, "");
999 }
1000
1001 for (unsigned i = 0; i < verts_per_prim; ++i) {
1002 tmp = LLVMBuildSub(builder, tid,
1003 LLVMConstInt(ctx->i32, verts_per_prim - i - 1, false), "");
1004 tmp = ngg_gs_vertex_ptr(ctx, tmp);
1005 nggso.vertices[i] = ac_build_gep0(&ctx->ac, tmp, ctx->i32_0);
1006 }
1007
1008 build_streamout(ctx, &nggso);
1009 }
1010
1011 /* Write shader query data. */
1012 unsigned num_query_comps = sel->so.num_outputs ? 8 : 4;
1013 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid,
1014 LLVMConstInt(ctx->i32, num_query_comps, false), "");
1015 ac_build_ifcc(&ctx->ac, tmp, 5110);
1016 {
1017 LLVMValueRef offset;
1018 tmp = tid;
1019 if (sel->so.num_outputs)
1020 tmp = LLVMBuildAnd(builder, tmp, LLVMConstInt(ctx->i32, 3, false), "");
1021 offset = LLVMBuildNUWMul(builder, tmp, LLVMConstInt(ctx->i32, 32, false), "");
1022 if (sel->so.num_outputs) {
1023 tmp = LLVMBuildLShr(builder, tid, LLVMConstInt(ctx->i32, 2, false), "");
1024 tmp = LLVMBuildNUWMul(builder, tmp, LLVMConstInt(ctx->i32, 8, false), "");
1025 offset = LLVMBuildAdd(builder, offset, tmp, "");
1026 }
1027
1028 tmp = LLVMBuildLoad(builder, ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tid), "");
1029 LLVMValueRef args[] = {
1030 tmp,
1031 ngg_get_query_buf(ctx),
1032 offset,
1033 LLVMConstInt(ctx->i32, 16, false), /* soffset */
1034 ctx->i32_0, /* cachepolicy */
1035 };
1036 ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.raw.buffer.atomic.add.i32",
1037 ctx->i32, args, 5, 0);
1038 }
1039 ac_build_endif(&ctx->ac, 5110);
1040
1041 /* TODO: culling */
1042
1043 /* Determine vertex liveness. */
1044 LLVMValueRef vertliveptr = lp_build_alloca(&ctx->gallivm, ctx->ac.i1, "vertexlive");
1045
1046 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
1047 ac_build_ifcc(&ctx->ac, tmp, 5120);
1048 {
1049 for (unsigned i = 0; i < verts_per_prim; ++i) {
1050 const LLVMValueRef primidx =
1051 LLVMBuildAdd(builder, tid,
1052 LLVMConstInt(ctx->ac.i32, i, false), "");
1053
1054 if (i > 0) {
1055 tmp = LLVMBuildICmp(builder, LLVMIntULT, primidx, num_emit_threads, "");
1056 ac_build_ifcc(&ctx->ac, tmp, 5121 + i);
1057 }
1058
1059 /* Load primitive liveness */
1060 tmp = ngg_gs_vertex_ptr(ctx, primidx);
1061 LLVMValueRef gep_idx[3] = {
1062 ctx->ac.i32_0, /* implicit C-style array */
1063 ctx->ac.i32_1, /* second value of struct */
1064 ctx->ac.i32_0, /* stream 0 */
1065 };
1066 tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
1067 tmp = LLVMBuildLoad(builder, tmp, "");
1068 const LLVMValueRef primlive =
1069 LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
1070
1071 tmp = LLVMBuildLoad(builder, vertliveptr, "");
1072 tmp = LLVMBuildOr(builder, tmp, primlive, ""),
1073 LLVMBuildStore(builder, tmp, vertliveptr);
1074
1075 if (i > 0)
1076 ac_build_endif(&ctx->ac, 5121 + i);
1077 }
1078 }
1079 ac_build_endif(&ctx->ac, 5120);
1080
1081 /* Inclusive scan addition across the current wave. */
1082 LLVMValueRef vertlive = LLVMBuildLoad(builder, vertliveptr, "");
1083 struct ac_wg_scan vertlive_scan = {};
1084 vertlive_scan.op = nir_op_iadd;
1085 vertlive_scan.enable_reduce = true;
1086 vertlive_scan.enable_exclusive = true;
1087 vertlive_scan.src = vertlive;
1088 vertlive_scan.scratch = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, ctx->i32_0);
1089 vertlive_scan.waveidx = get_wave_id_in_tg(ctx);
1090 vertlive_scan.numwaves = get_tgsize(ctx);
1091 vertlive_scan.maxwaves = 8;
1092
1093 ac_build_wg_scan(&ctx->ac, &vertlive_scan);
1094
1095 /* Skip all exports (including index exports) when possible. At least on
1096 * early gfx10 revisions this is also to avoid hangs.
1097 */
1098 LLVMValueRef have_exports =
1099 LLVMBuildICmp(builder, LLVMIntNE, vertlive_scan.result_reduce, ctx->ac.i32_0, "");
1100 num_emit_threads =
1101 LLVMBuildSelect(builder, have_exports, num_emit_threads, ctx->ac.i32_0, "");
1102
1103 /* Allocate export space. Send this message as early as possible, to
1104 * hide the latency of the SQ <-> SPI roundtrip.
1105 *
1106 * Note: We could consider compacting primitives for export as well.
1107 * PA processes 1 non-null prim / clock, but it fetches 4 DW of
1108 * prim data per clock and skips null primitives at no additional
1109 * cost. So compacting primitives can only be beneficial when
1110 * there are 4 or more contiguous null primitives in the export
1111 * (in the common case of single-dword prim exports).
1112 */
1113 build_sendmsg_gs_alloc_req(ctx, vertlive_scan.result_reduce, num_emit_threads);
1114
1115 /* Setup the reverse vertex compaction permutation. We re-use stream 1
1116 * of the primitive liveness flags, relying on the fact that each
1117 * threadgroup can have at most 256 threads. */
1118 ac_build_ifcc(&ctx->ac, vertlive, 5130);
1119 {
1120 tmp = ngg_gs_vertex_ptr(ctx, vertlive_scan.result_exclusive);
1121 LLVMValueRef gep_idx[3] = {
1122 ctx->ac.i32_0, /* implicit C-style array */
1123 ctx->ac.i32_1, /* second value of struct */
1124 ctx->ac.i32_1, /* stream 1 */
1125 };
1126 tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
1127 tmp2 = LLVMBuildTrunc(builder, tid, ctx->ac.i8, "");
1128 LLVMBuildStore(builder, tmp2, tmp);
1129 }
1130 ac_build_endif(&ctx->ac, 5130);
1131
1132 ac_build_s_barrier(&ctx->ac);
1133
1134 /* Export primitive data */
1135 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, num_emit_threads, "");
1136 ac_build_ifcc(&ctx->ac, tmp, 5140);
1137 {
1138 struct ngg_prim prim = {};
1139 prim.num_vertices = verts_per_prim;
1140
1141 tmp = ngg_gs_vertex_ptr(ctx, tid);
1142 LLVMValueRef gep_idx[3] = {
1143 ctx->ac.i32_0, /* implicit C-style array */
1144 ctx->ac.i32_1, /* second value of struct */
1145 ctx->ac.i32_0, /* primflag */
1146 };
1147 tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
1148 tmp = LLVMBuildLoad(builder, tmp, "");
1149 prim.isnull = LLVMBuildICmp(builder, LLVMIntEQ, tmp,
1150 LLVMConstInt(ctx->ac.i8, 0, false), "");
1151
1152 for (unsigned i = 0; i < verts_per_prim; ++i) {
1153 prim.index[i] = LLVMBuildSub(builder, vertlive_scan.result_exclusive,
1154 LLVMConstInt(ctx->ac.i32, verts_per_prim - i - 1, false), "");
1155 prim.edgeflag[i] = ctx->ac.i1false;
1156 }
1157
1158 build_export_prim(ctx, &prim);
1159 }
1160 ac_build_endif(&ctx->ac, 5140);
1161
1162 /* Export position and parameter data */
1163 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, vertlive_scan.result_reduce, "");
1164 ac_build_ifcc(&ctx->ac, tmp, 5145);
1165 {
1166 struct si_shader_output_values *outputs = NULL;
1167 outputs = MALLOC(info->num_outputs * sizeof(outputs[0]));
1168
1169 tmp = ngg_gs_vertex_ptr(ctx, tid);
1170 LLVMValueRef gep_idx[3] = {
1171 ctx->ac.i32_0, /* implicit C-style array */
1172 ctx->ac.i32_1, /* second value of struct */
1173 ctx->ac.i32_1, /* stream 1: source data index */
1174 };
1175 tmp = LLVMBuildGEP(builder, tmp, gep_idx, 3, "");
1176 tmp = LLVMBuildLoad(builder, tmp, "");
1177 tmp = LLVMBuildZExt(builder, tmp, ctx->ac.i32, "");
1178 const LLVMValueRef vertexptr = ngg_gs_vertex_ptr(ctx, tmp);
1179
1180 unsigned out_idx = 0;
1181 gep_idx[1] = ctx->ac.i32_0;
1182 for (unsigned i = 0; i < info->num_outputs; i++) {
1183 outputs[i].semantic_name = info->output_semantic_name[i];
1184 outputs[i].semantic_index = info->output_semantic_index[i];
1185
1186 for (unsigned j = 0; j < 4; j++, out_idx++) {
1187 gep_idx[2] = LLVMConstInt(ctx->ac.i32, out_idx, false);
1188 tmp = LLVMBuildGEP(builder, vertexptr, gep_idx, 3, "");
1189 tmp = LLVMBuildLoad(builder, tmp, "");
1190 outputs[i].values[j] = ac_to_float(&ctx->ac, tmp);
1191 outputs[i].vertex_stream[j] =
1192 (info->output_streams[i] >> (2 * j)) & 3;
1193 }
1194 }
1195
1196 si_llvm_export_vs(ctx, outputs, info->num_outputs);
1197
1198 FREE(outputs);
1199 }
1200 ac_build_endif(&ctx->ac, 5145);
1201 }
1202
1203 static void clamp_gsprims_to_esverts(unsigned *max_gsprims, unsigned max_esverts,
1204 unsigned min_verts_per_prim, bool use_adjacency)
1205 {
1206 unsigned max_reuse = max_esverts - min_verts_per_prim;
1207 if (use_adjacency)
1208 max_reuse /= 2;
1209 *max_gsprims = MIN2(*max_gsprims, 1 + max_reuse);
1210 }
1211
1212 /**
1213 * Determine subgroup information like maximum number of vertices and prims.
1214 *
1215 * This happens before the shader is uploaded, since LDS relocations during
1216 * upload depend on the subgroup size.
1217 */
1218 void gfx10_ngg_calculate_subgroup_info(struct si_shader *shader)
1219 {
1220 const struct si_shader_selector *gs_sel = shader->selector;
1221 const struct si_shader_selector *es_sel =
1222 shader->previous_stage_sel ? shader->previous_stage_sel : gs_sel;
1223 const enum pipe_shader_type gs_type = gs_sel->type;
1224 const unsigned gs_num_invocations = MAX2(gs_sel->gs_num_invocations, 1);
1225 /* TODO: Specialize for known primitive type without GS. */
1226 const unsigned input_prim = gs_type == PIPE_SHADER_GEOMETRY ?
1227 gs_sel->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM] :
1228 PIPE_PRIM_TRIANGLES;
1229 const bool use_adjacency = input_prim >= PIPE_PRIM_LINES_ADJACENCY &&
1230 input_prim <= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY;
1231 const unsigned max_verts_per_prim = u_vertices_per_prim(input_prim);
1232 const unsigned min_verts_per_prim =
1233 gs_type == PIPE_SHADER_GEOMETRY ? max_verts_per_prim : 1;
1234
1235 /* All these are in dwords: */
1236 /* We can't allow using the whole LDS, because GS waves compete with
1237 * other shader stages for LDS space.
1238 *
1239 * Streamout can increase the ESGS buffer size later on, so be more
1240 * conservative with streamout and use 4K dwords. This may be suboptimal.
1241 *
1242 * Otherwise, use the limit of 7K dwords. The reason is that we need
1243 * to leave some headroom for the max_esverts increase at the end.
1244 *
1245 * TODO: We should really take the shader's internal LDS use into
1246 * account. The linker will fail if the size is greater than
1247 * 8K dwords.
1248 */
1249 const unsigned max_lds_size = (gs_sel->so.num_outputs ? 4 : 7) * 1024 - 128;
1250 const unsigned target_lds_size = max_lds_size;
1251 unsigned esvert_lds_size = 0;
1252 unsigned gsprim_lds_size = 0;
1253
1254 /* All these are per subgroup: */
1255 bool max_vert_out_per_gs_instance = false;
1256 unsigned max_esverts_base = 256;
1257 unsigned max_gsprims_base = 128; /* default prim group size clamp */
1258
1259 /* Hardware has the following non-natural restrictions on the value
1260 * of GE_CNTL.VERT_GRP_SIZE based on based on the primitive type of
1261 * the draw:
1262 * - at most 252 for any line input primitive type
1263 * - at most 251 for any quad input primitive type
1264 * - at most 251 for triangle strips with adjacency (this happens to
1265 * be the natural limit for triangle *lists* with adjacency)
1266 */
1267 max_esverts_base = MIN2(max_esverts_base, 251 + max_verts_per_prim - 1);
1268
1269 if (gs_type == PIPE_SHADER_GEOMETRY) {
1270 unsigned max_out_verts_per_gsprim =
1271 gs_sel->gs_max_out_vertices * gs_num_invocations;
1272
1273 if (max_out_verts_per_gsprim <= 256) {
1274 if (max_out_verts_per_gsprim) {
1275 max_gsprims_base = MIN2(max_gsprims_base,
1276 256 / max_out_verts_per_gsprim);
1277 }
1278 } else {
1279 /* Use special multi-cycling mode in which each GS
1280 * instance gets its own subgroup. Does not work with
1281 * tessellation. */
1282 max_vert_out_per_gs_instance = true;
1283 max_gsprims_base = 1;
1284 max_out_verts_per_gsprim = gs_sel->gs_max_out_vertices;
1285 }
1286
1287 esvert_lds_size = es_sel->esgs_itemsize / 4;
1288 gsprim_lds_size = (gs_sel->gsvs_vertex_size / 4 + 1) * max_out_verts_per_gsprim;
1289 } else {
1290 /* TODO: This needs to be adjusted once LDS use for compaction
1291 * after culling is implemented. */
1292 if (es_sel->so.num_outputs)
1293 esvert_lds_size = 4 * es_sel->info.num_outputs + 1;
1294 }
1295
1296 unsigned max_gsprims = max_gsprims_base;
1297 unsigned max_esverts = max_esverts_base;
1298
1299 if (esvert_lds_size)
1300 max_esverts = MIN2(max_esverts, target_lds_size / esvert_lds_size);
1301 if (gsprim_lds_size)
1302 max_gsprims = MIN2(max_gsprims, target_lds_size / gsprim_lds_size);
1303
1304 max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim);
1305 clamp_gsprims_to_esverts(&max_gsprims, max_esverts, min_verts_per_prim, use_adjacency);
1306 assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1);
1307
1308 if (esvert_lds_size || gsprim_lds_size) {
1309 /* Now that we have a rough proportionality between esverts
1310 * and gsprims based on the primitive type, scale both of them
1311 * down simultaneously based on required LDS space.
1312 *
1313 * We could be smarter about this if we knew how much vertex
1314 * reuse to expect.
1315 */
1316 unsigned lds_total = max_esverts * esvert_lds_size +
1317 max_gsprims * gsprim_lds_size;
1318 if (lds_total > target_lds_size) {
1319 max_esverts = max_esverts * target_lds_size / lds_total;
1320 max_gsprims = max_gsprims * target_lds_size / lds_total;
1321
1322 max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim);
1323 clamp_gsprims_to_esverts(&max_gsprims, max_esverts,
1324 min_verts_per_prim, use_adjacency);
1325 assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1);
1326 }
1327 }
1328
1329 /* Round up towards full wave sizes for better ALU utilization. */
1330 if (!max_vert_out_per_gs_instance) {
1331 const unsigned wavesize = 64;
1332 unsigned orig_max_esverts;
1333 unsigned orig_max_gsprims;
1334 do {
1335 orig_max_esverts = max_esverts;
1336 orig_max_gsprims = max_gsprims;
1337
1338 max_esverts = align(max_esverts, wavesize);
1339 max_esverts = MIN2(max_esverts, max_esverts_base);
1340 if (esvert_lds_size)
1341 max_esverts = MIN2(max_esverts,
1342 (max_lds_size - max_gsprims * gsprim_lds_size) /
1343 esvert_lds_size);
1344 max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim);
1345
1346 max_gsprims = align(max_gsprims, wavesize);
1347 max_gsprims = MIN2(max_gsprims, max_gsprims_base);
1348 if (gsprim_lds_size)
1349 max_gsprims = MIN2(max_gsprims,
1350 (max_lds_size - max_esverts * esvert_lds_size) /
1351 gsprim_lds_size);
1352 clamp_gsprims_to_esverts(&max_gsprims, max_esverts,
1353 min_verts_per_prim, use_adjacency);
1354 assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1);
1355 } while (orig_max_esverts != max_esverts || orig_max_gsprims != max_gsprims);
1356 }
1357
1358 /* Hardware restriction: minimum value of max_esverts */
1359 max_esverts = MAX2(max_esverts, 23 + max_verts_per_prim);
1360
1361 unsigned max_out_vertices =
1362 max_vert_out_per_gs_instance ? gs_sel->gs_max_out_vertices :
1363 gs_type == PIPE_SHADER_GEOMETRY ?
1364 max_gsprims * gs_num_invocations * gs_sel->gs_max_out_vertices :
1365 max_esverts;
1366 assert(max_out_vertices <= 256);
1367
1368 unsigned prim_amp_factor = 1;
1369 if (gs_type == PIPE_SHADER_GEOMETRY) {
1370 /* Number of output primitives per GS input primitive after
1371 * GS instancing. */
1372 prim_amp_factor = gs_sel->gs_max_out_vertices;
1373 }
1374
1375 /* The GE only checks against the maximum number of ES verts after
1376 * allocating a full GS primitive. So we need to ensure that whenever
1377 * this check passes, there is enough space for a full primitive without
1378 * vertex reuse.
1379 */
1380 shader->ngg.hw_max_esverts = max_esverts - max_verts_per_prim + 1;
1381 shader->ngg.max_gsprims = max_gsprims;
1382 shader->ngg.max_out_verts = max_out_vertices;
1383 shader->ngg.prim_amp_factor = prim_amp_factor;
1384 shader->ngg.max_vert_out_per_gs_instance = max_vert_out_per_gs_instance;
1385
1386 shader->gs_info.esgs_ring_size = 4 * max_esverts * esvert_lds_size;
1387 shader->ngg.ngg_emit_size = max_gsprims * gsprim_lds_size;
1388
1389 assert(shader->ngg.hw_max_esverts >= 24); /* HW limitation */
1390 }