2 * Copyright 2017 Advanced Micro Devices, Inc.
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:
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
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.
25 #include "si_shader_internal.h"
29 #include "util/u_memory.h"
30 #include "util/u_prim.h"
32 static LLVMValueRef
get_wave_id_in_tg(struct si_shader_context
*ctx
)
34 return si_unpack_param(ctx
, ctx
->param_merged_wave_info
, 24, 4);
37 static LLVMValueRef
get_tgsize(struct si_shader_context
*ctx
)
39 return si_unpack_param(ctx
, ctx
->param_merged_wave_info
, 28, 4);
42 static LLVMValueRef
get_thread_id_in_tg(struct si_shader_context
*ctx
)
44 LLVMBuilderRef builder
= ctx
->ac
.builder
;
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
), "");
51 static LLVMValueRef
ngg_get_vtx_cnt(struct si_shader_context
*ctx
)
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),
59 static LLVMValueRef
ngg_get_prim_cnt(struct si_shader_context
*ctx
)
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),
67 static LLVMValueRef
ngg_get_ordered_id(struct si_shader_context
*ctx
)
69 return ac_build_bfe(&ctx
->ac
, ctx
->gs_tg_info
,
71 LLVMConstInt(ctx
->ac
.i32
, 11, false),
75 static LLVMValueRef
ngg_get_query_buf(struct si_shader_context
*ctx
)
77 LLVMValueRef buf_ptr
= LLVMGetParam(ctx
->main_fn
,
78 ctx
->param_rw_buffers
);
80 return ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
,
81 LLVMConstInt(ctx
->i32
, GFX10_GS_QUERY_BUF
, false));
84 /* Send GS Alloc Req message from the first wave of the group to SPI.
86 * - bits 0..10: vertices in group
87 * - bits 12..22: primitives in group
89 static void build_sendmsg_gs_alloc_req(struct si_shader_context
*ctx
,
91 LLVMValueRef prim_cnt
)
93 LLVMBuilderRef builder
= ctx
->ac
.builder
;
96 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, get_wave_id_in_tg(ctx
), ctx
->ac
.i32_0
, "");
97 ac_build_ifcc(&ctx
->ac
, tmp
, 5020);
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
);
103 ac_build_endif(&ctx
->ac
, 5020);
107 unsigned num_vertices
;
109 LLVMValueRef index
[3];
110 LLVMValueRef edgeflag
[3];
113 static void build_export_prim(struct si_shader_context
*ctx
,
114 const struct ngg_prim
*prim
)
116 LLVMBuilderRef builder
= ctx
->ac
.builder
;
117 struct ac_export_args args
;
120 tmp
= LLVMBuildZExt(builder
, prim
->isnull
, ctx
->ac
.i32
, "");
121 args
.out
[0] = LLVMBuildShl(builder
, tmp
, LLVMConstInt(ctx
->ac
.i32
, 31, false), "");
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
, "");
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
);
138 args
.target
= V_008DFC_SQ_EXP_PRIM
;
139 args
.enabled_channels
= 1;
141 args
.valid_mask
= false;
144 ac_build_export(&ctx
->ac
, &args
);
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
)
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] = {};
158 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
159 if (!wg_offset_dw
[buffer
])
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), "");
168 for (unsigned i
= 0; i
< so
->num_outputs
; ++i
) {
169 if (so
->output
[i
].stream
!= stream
)
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
];
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;
185 si_emit_streamout_output(ctx
, so_buffer
, offset
, &so
->output
[i
], &out
);
189 struct ngg_streamout
{
190 LLVMValueRef num_vertices
;
192 /* per-thread data */
193 LLVMValueRef prim_enable
[4]; /* i1 per stream */
194 LLVMValueRef vertices
[3]; /* [N x i32] addrspace(LDS)* */
197 LLVMValueRef emit
[4]; /* per-stream emitted primitives (only valid for used streams) */
201 * Build streamout logic.
205 * Writes number of emitted primitives to gs_ngg_scratch[4:8].
207 * Clobbers gs_ngg_scratch[8:].
209 static void build_streamout(struct si_shader_context
*ctx
,
210 struct ngg_streamout
*nggso
)
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
;
234 ac_llvm_add_target_dep_function_attr(ctx
->main_fn
, "amdgpu-gds-size", 256);
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
;
245 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
246 if (stream_for_buffer
[buffer
] == -1)
249 assert(so
->stride
[buffer
]);
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));
257 so_buffer
[buffer
] = ac_build_load_to_sgpr(
259 LLVMConstInt(ctx
->i32
, SI_VS_STREAMOUT_BUF0
+ buffer
, false));
262 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, get_wave_id_in_tg(ctx
), ctx
->i32_0
, "");
263 ac_build_ifcc(&ctx
->ac
, tmp
, 5200);
265 LLVMTypeRef gdsptr
= LLVMPointerType(ctx
->i32
, AC_ADDR_SPACE_GDS
);
266 LLVMValueRef gdsbase
= LLVMBuildIntToPtr(builder
, ctx
->i32_0
, gdsptr
, "");
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
, "");
272 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, ac_get_thread_id(&ctx
->ac
), i32_4
, "");
273 ac_build_ifcc(&ctx
->ac
, tmp
, 5210);
276 tmp
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, tid
);
277 tmp
= LLVMBuildLoad(builder
, tmp
, "");
279 tmp
= ac_build_writelane(&ctx
->ac
, ctx
->i32_0
,
280 ngg_get_prim_cnt(ctx
), ctx
->i32_0
);
282 LLVMBuildStore(builder
, tmp
, generated_by_stream_vgpr
);
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
;
292 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
293 if (stream_for_buffer
[buffer
] >= 0) {
294 swizzle
[buffer
] = stream_for_buffer
[buffer
];
296 assert(unused_stream
>= 0);
297 swizzle
[buffer
] = unused_stream
;
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
, "");
305 LLVMValueRef args
[] = {
306 LLVMBuildIntToPtr(builder
, ngg_get_ordered_id(ctx
), gdsptr
, ""),
308 ctx
->i32_0
, // ordering
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
315 tmp
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.ds.ordered.add",
316 ctx
->i32
, args
, ARRAY_SIZE(args
), 0);
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
);
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
);
328 ac_build_endif(&ctx
->ac
, 5210);
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.
334 LLVMValueRef max_emit
[4] = {};
335 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
336 if (stream_for_buffer
[buffer
] == -1)
339 LLVMValueRef bufsize_dw
=
340 LLVMBuildLShr(builder
,
341 LLVMBuildExtractElement(builder
, so_buffer
[buffer
], i32_2
, ""),
344 tmp
= LLVMBuildLoad(builder
, offsets_vgpr
, "");
345 LLVMValueRef offset_dw
=
346 ac_build_readlane(&ctx
->ac
, tmp
,
347 LLVMConstInt(ctx
->i32
, buffer
, false));
349 tmp
= LLVMBuildSub(builder
, bufsize_dw
, offset_dw
, "");
350 tmp
= LLVMBuildUDiv(builder
, tmp
, prim_stride_dw
[buffer
], "");
352 tmp2
= LLVMBuildICmp(builder
, LLVMIntULT
, bufsize_dw
, offset_dw
, "");
353 max_emit
[buffer
] = LLVMBuildSelect(builder
, tmp2
, ctx
->i32_0
, tmp
, "");
356 /* Determine the number of emitted primitives per stream and fixup the
357 * GDS counter if necessary.
359 * This is complicated by the fact that a single stream can emit to
360 * multiple buffers (but luckily not vice versa).
362 LLVMValueRef emit_vgpr
= ctx
->i32_0
;
364 for (unsigned stream
= 0; stream
< 4; ++stream
) {
365 if (!info
->num_stream_output_components
[stream
])
368 tmp
= LLVMBuildLoad(builder
, generated_by_stream_vgpr
, "");
369 LLVMValueRef generated
=
370 ac_build_readlane(&ctx
->ac
, tmp
,
371 LLVMConstInt(ctx
->i32
, stream
, false));
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
]);
379 emit_vgpr
= ac_build_writelane(&ctx
->ac
, emit_vgpr
, emit
,
380 LLVMConstInt(ctx
->i32
, stream
, false));
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);
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);
397 ac_build_endif(&ctx
->ac
, 5222);
398 ac_build_endif(&ctx
->ac
, 5221);
401 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, ac_get_thread_id(&ctx
->ac
), i32_4
, "");
402 ac_build_ifcc(&ctx
->ac
, tmp
, 5225);
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
);
409 ac_build_endif(&ctx
->ac
, 5225);
411 ac_build_endif(&ctx
->ac
, 5200);
413 /* Determine the workgroup-relative per-thread / primitive offset into
414 * the streamout buffers */
415 struct ac_wg_scan primemit_scan
[4] = {};
418 for (unsigned stream
= 0; stream
< 4; ++stream
) {
419 if (!info
->num_stream_output_components
[stream
])
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
]);
435 ac_build_s_barrier(&ctx
->ac
);
437 /* Fetch the per-buffer offsets and per-stream emit counts in all waves. */
438 LLVMValueRef wgoffset_dw
[4] = {};
441 LLVMValueRef scratch_vgpr
;
443 tmp
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, ac_get_thread_id(&ctx
->ac
));
444 scratch_vgpr
= LLVMBuildLoad(builder
, tmp
, "");
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));
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));
463 /* Write out primitive data */
464 for (unsigned stream
= 0; stream
< 4; ++stream
) {
465 if (!info
->num_stream_output_components
[stream
])
469 ac_build_wg_scan_bottom(&ctx
->ac
, &primemit_scan
[stream
]);
471 primemit_scan
[stream
].result_exclusive
= tid
;
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);
480 LLVMValueRef offset_vtx
=
481 LLVMBuildMul(builder
, primemit_scan
[stream
].result_exclusive
,
482 nggso
->num_vertices
, "");
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
, "");
495 ac_build_endif(&ctx
->ac
, 5240);
500 * Returns an `[N x i32] addrspace(LDS)*` pointing at contiguous LDS storage
501 * for the vertex outputs.
503 static LLVMValueRef
ngg_nogs_vertex_ptr(struct si_shader_context
*ctx
,
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, "");
515 * Emit the epilogue of an API VS or TES shader compiled as ESGS shader.
517 void gfx10_emit_ngg_epilogue(struct ac_shader_abi
*abi
,
518 unsigned max_outputs
,
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
;
529 assert(!ctx
->shader
->is_gs_copy_shader
);
530 assert(info
->num_outputs
<= max_outputs
);
532 outputs
= MALLOC((info
->num_outputs
+ 1) * sizeof(outputs
[0]));
534 LLVMValueRef vertex_ptr
= NULL
;
536 if (sel
->so
.num_outputs
)
537 vertex_ptr
= ngg_nogs_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
));
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
];
543 /* This is used only by streamout. */
544 for (unsigned j
= 0; j
< 4; j
++) {
545 outputs
[i
].values
[j
] =
546 LLVMBuildLoad(builder
,
549 outputs
[i
].vertex_stream
[j
] =
550 (info
->output_streams
[i
] >> (2 * j
)) & 3;
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
);
561 lp_build_endif(&ctx
->merged_wrap_if_state
);
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),
575 /* Determine the number of vertices per primitive. */
576 unsigned num_vertices
;
577 LLVMValueRef num_vertices_val
;
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. */
583 num_vertices_val
= LLVMConstInt(ctx
->i32
, 3, 0);
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 */
591 assert(ctx
->type
== PIPE_SHADER_TESS_EVAL
);
593 if (info
->properties
[TGSI_PROPERTY_TES_POINT_MODE
])
595 else if (info
->properties
[TGSI_PROPERTY_TES_PRIM_MODE
] == PIPE_PRIM_LINES
)
600 num_vertices_val
= LLVMConstInt(ctx
->i32
, num_vertices
, false);
604 LLVMValueRef emitted_prims
= NULL
;
606 if (sel
->so
.num_outputs
) {
607 struct ngg_streamout nggso
= {};
609 nggso
.num_vertices
= num_vertices_val
;
610 nggso
.prim_enable
[0] = is_gs_thread
;
612 for (unsigned i
= 0; i
< num_vertices
; ++i
)
613 nggso
.vertices
[i
] = ngg_nogs_vertex_ptr(ctx
, vtxindex
[i
]);
615 build_streamout(ctx
, &nggso
);
616 emitted_prims
= nggso
.emit
[0];
619 /* Copy Primitive IDs from GS threads to the LDS address corresponding
620 * to the ES thread of the provoking vertex.
622 if (ctx
->type
== PIPE_SHADER_VERTEX
&&
623 ctx
->shader
->key
.mono
.u
.vs_export_prim_id
) {
624 /* Streamout uses LDS. We need to wait for it before we can reuse it. */
625 if (sel
->so
.num_outputs
)
626 ac_build_s_barrier(&ctx
->ac
);
628 ac_build_ifcc(&ctx
->ac
, is_gs_thread
, 5400);
629 /* Extract the PROVOKING_VTX_INDEX field. */
630 LLVMValueRef provoking_vtx_in_prim
=
631 si_unpack_param(ctx
, ctx
->param_vs_state_bits
, 4, 2);
633 /* provoking_vtx_index = vtxindex[provoking_vtx_in_prim]; */
634 LLVMValueRef indices
= ac_build_gather_values(&ctx
->ac
, vtxindex
, 3);
635 LLVMValueRef provoking_vtx_index
=
636 LLVMBuildExtractElement(builder
, indices
, provoking_vtx_in_prim
, "");
638 LLVMBuildStore(builder
, ctx
->abi
.gs_prim_id
,
639 ac_build_gep0(&ctx
->ac
, ctx
->esgs_ring
, provoking_vtx_index
));
640 ac_build_endif(&ctx
->ac
, 5400);
643 /* TODO: primitive culling */
645 build_sendmsg_gs_alloc_req(ctx
, ngg_get_vtx_cnt(ctx
), ngg_get_prim_cnt(ctx
));
647 /* Update query buffer */
648 /* TODO: this won't catch 96-bit clear_buffer via transform feedback. */
649 if (!info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS
]) {
650 tmp
= si_unpack_param(ctx
, ctx
->param_vs_state_bits
, 6, 1);
651 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->i1
, "");
652 ac_build_ifcc(&ctx
->ac
, tmp
, 5029); /* if (STREAMOUT_QUERY_ENABLED) */
653 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, get_wave_id_in_tg(ctx
), ctx
->ac
.i32_0
, "");
654 ac_build_ifcc(&ctx
->ac
, tmp
, 5030);
655 tmp
= LLVMBuildICmp(builder
, LLVMIntULE
, ac_get_thread_id(&ctx
->ac
),
656 sel
->so
.num_outputs
? ctx
->ac
.i32_1
: ctx
->ac
.i32_0
, "");
657 ac_build_ifcc(&ctx
->ac
, tmp
, 5031);
659 LLVMValueRef args
[] = {
660 ngg_get_prim_cnt(ctx
),
661 ngg_get_query_buf(ctx
),
662 LLVMConstInt(ctx
->i32
, 16, false), /* offset of stream[0].generated_primitives */
663 ctx
->i32_0
, /* soffset */
664 ctx
->i32_0
, /* cachepolicy */
667 if (sel
->so
.num_outputs
) {
668 args
[0] = ac_build_writelane(&ctx
->ac
, args
[0], emitted_prims
, ctx
->i32_1
);
669 args
[2] = ac_build_writelane(&ctx
->ac
, args
[2],
670 LLVMConstInt(ctx
->i32
, 24, false), ctx
->i32_1
);
673 /* TODO: should this be 64-bit atomics? */
674 ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.raw.buffer.atomic.add.i32",
675 ctx
->i32
, args
, 5, 0);
677 ac_build_endif(&ctx
->ac
, 5031);
678 ac_build_endif(&ctx
->ac
, 5030);
679 ac_build_endif(&ctx
->ac
, 5029);
682 /* Export primitive data to the index buffer. Format is:
683 * - bits 0..8: index 0
684 * - bit 9: edge flag 0
685 * - bits 10..18: index 1
686 * - bit 19: edge flag 1
687 * - bits 20..28: index 2
688 * - bit 29: edge flag 2
689 * - bit 31: null primitive (skip)
691 * For the first version, we will always build up all three indices
692 * independent of the primitive type. The additional garbage data
695 * TODO: culling depends on the primitive type, so can have some
698 lp_build_if(&if_state
, &ctx
->gallivm
, is_gs_thread
);
700 struct ngg_prim prim
= {};
702 prim
.num_vertices
= num_vertices
;
703 prim
.isnull
= ctx
->ac
.i1false
;
704 memcpy(prim
.index
, vtxindex
, sizeof(vtxindex
[0]) * 3);
706 for (unsigned i
= 0; i
< num_vertices
; ++i
) {
707 tmp
= LLVMBuildLShr(builder
, ctx
->abi
.gs_invocation_id
,
708 LLVMConstInt(ctx
->ac
.i32
, 8 + i
, false), "");
709 prim
.edgeflag
[i
] = LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
712 build_export_prim(ctx
, &prim
);
714 lp_build_endif(&if_state
);
716 /* Export per-vertex data (positions and parameters). */
717 lp_build_if(&if_state
, &ctx
->gallivm
, is_es_thread
);
721 /* Unconditionally (re-)load the values for proper SSA form. */
722 for (i
= 0; i
< info
->num_outputs
; i
++) {
723 for (unsigned j
= 0; j
< 4; j
++) {
724 outputs
[i
].values
[j
] =
725 LLVMBuildLoad(builder
,
731 if (ctx
->shader
->key
.mono
.u
.vs_export_prim_id
) {
732 outputs
[i
].semantic_name
= TGSI_SEMANTIC_PRIMID
;
733 outputs
[i
].semantic_index
= 0;
735 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
736 /* Wait for GS stores to finish. */
737 ac_build_s_barrier(&ctx
->ac
);
739 tmp
= ac_build_gep0(&ctx
->ac
, ctx
->esgs_ring
,
740 get_thread_id_in_tg(ctx
));
741 outputs
[i
].values
[0] = LLVMBuildLoad(builder
, tmp
, "");
743 assert(ctx
->type
== PIPE_SHADER_TESS_EVAL
);
744 outputs
[i
].values
[0] = si_get_primitive_id(ctx
, 0);
747 outputs
[i
].values
[0] = ac_to_float(&ctx
->ac
, outputs
[i
].values
[0]);
748 for (unsigned j
= 1; j
< 4; j
++)
749 outputs
[i
].values
[j
] = LLVMGetUndef(ctx
->f32
);
751 memset(outputs
[i
].vertex_stream
, 0,
752 sizeof(outputs
[i
].vertex_stream
));
756 si_llvm_export_vs(ctx
, outputs
, i
);
758 lp_build_endif(&if_state
);
764 ngg_gs_get_vertex_storage(struct si_shader_context
*ctx
)
766 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
767 const struct tgsi_shader_info
*info
= &sel
->info
;
769 LLVMTypeRef elements
[2] = {
770 LLVMArrayType(ctx
->ac
.i32
, 4 * info
->num_outputs
),
771 LLVMArrayType(ctx
->ac
.i8
, 4),
773 LLVMTypeRef type
= LLVMStructTypeInContext(ctx
->ac
.context
, elements
, 2, false);
774 type
= LLVMPointerType(LLVMArrayType(type
, 0), AC_ADDR_SPACE_LDS
);
775 return LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->gs_ngg_emit
, type
, "");
779 * Return a pointer to the LDS storage reserved for the N'th vertex, where N
780 * is in emit order; that is:
781 * - during the epilogue, N is the threadidx (relative to the entire threadgroup)
782 * - during vertex emit, i.e. while the API GS shader invocation is running,
783 * N = threadidx * gs_max_out_vertices + emitidx
785 * Goals of the LDS memory layout:
786 * 1. Eliminate bank conflicts on write for geometry shaders that have all emits
787 * in uniform control flow
788 * 2. Eliminate bank conflicts on read for export if, additionally, there is no
790 * 3. Agnostic to the number of waves (since we don't know it before compiling)
791 * 4. Allow coalescing of LDS instructions (ds_write_b128 etc.)
792 * 5. Avoid wasting memory.
794 * We use an AoS layout due to point 4 (this also helps point 3). In an AoS
795 * layout, elimination of bank conflicts requires that each vertex occupy an
796 * odd number of dwords. We use the additional dword to store the output stream
797 * index as well as a flag to indicate whether this vertex ends a primitive
800 * Swizzling is required to satisfy points 1 and 2 simultaneously.
802 * Vertices are stored in export order (gsthread * gs_max_out_vertices + emitidx).
803 * Indices are swizzled in groups of 32, which ensures point 1 without
804 * disturbing point 2.
806 * \return an LDS pointer to type {[N x i32], [4 x i8]}
809 ngg_gs_vertex_ptr(struct si_shader_context
*ctx
, LLVMValueRef vertexidx
)
811 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
812 LLVMBuilderRef builder
= ctx
->ac
.builder
;
813 LLVMValueRef storage
= ngg_gs_get_vertex_storage(ctx
);
815 /* gs_max_out_vertices = 2^(write_stride_2exp) * some odd number */
816 unsigned write_stride_2exp
= ffs(sel
->gs_max_out_vertices
) - 1;
817 if (write_stride_2exp
) {
819 LLVMBuildLShr(builder
, vertexidx
,
820 LLVMConstInt(ctx
->ac
.i32
, 5, false), "");
821 LLVMValueRef swizzle
=
822 LLVMBuildAnd(builder
, row
,
823 LLVMConstInt(ctx
->ac
.i32
, (1u << write_stride_2exp
) - 1,
825 vertexidx
= LLVMBuildXor(builder
, vertexidx
, swizzle
, "");
828 return ac_build_gep0(&ctx
->ac
, storage
, vertexidx
);
832 ngg_gs_emit_vertex_ptr(struct si_shader_context
*ctx
, LLVMValueRef gsthread
,
833 LLVMValueRef emitidx
)
835 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
836 LLVMBuilderRef builder
= ctx
->ac
.builder
;
839 tmp
= LLVMConstInt(ctx
->ac
.i32
, sel
->gs_max_out_vertices
, false);
840 tmp
= LLVMBuildMul(builder
, tmp
, gsthread
, "");
841 const LLVMValueRef vertexidx
= LLVMBuildAdd(builder
, tmp
, emitidx
, "");
842 return ngg_gs_vertex_ptr(ctx
, vertexidx
);
845 void gfx10_ngg_gs_emit_vertex(struct si_shader_context
*ctx
,
849 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
850 const struct tgsi_shader_info
*info
= &sel
->info
;
851 LLVMBuilderRef builder
= ctx
->ac
.builder
;
852 struct lp_build_if_state if_state
;
854 const LLVMValueRef vertexidx
=
855 LLVMBuildLoad(builder
, ctx
->gs_next_vertex
[stream
], "");
857 /* If this thread has already emitted the declared maximum number of
858 * vertices, skip the write: excessive vertex emissions are not
859 * supposed to have any effect.
861 const LLVMValueRef can_emit
=
862 LLVMBuildICmp(builder
, LLVMIntULT
, vertexidx
,
863 LLVMConstInt(ctx
->i32
, sel
->gs_max_out_vertices
, false), "");
865 tmp
= LLVMBuildAdd(builder
, vertexidx
, ctx
->ac
.i32_1
, "");
866 tmp
= LLVMBuildSelect(builder
, can_emit
, tmp
, vertexidx
, "");
867 LLVMBuildStore(builder
, tmp
, ctx
->gs_next_vertex
[stream
]);
869 lp_build_if(&if_state
, &ctx
->gallivm
, can_emit
);
871 const LLVMValueRef vertexptr
=
872 ngg_gs_emit_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
), vertexidx
);
873 unsigned out_idx
= 0;
874 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
875 for (unsigned chan
= 0; chan
< 4; chan
++, out_idx
++) {
876 if (!(info
->output_usagemask
[i
] & (1 << chan
)) ||
877 ((info
->output_streams
[i
] >> (2 * chan
)) & 3) != stream
)
880 LLVMValueRef out_val
= LLVMBuildLoad(builder
, addrs
[4 * i
+ chan
], "");
881 LLVMValueRef gep_idx
[3] = {
882 ctx
->ac
.i32_0
, /* implied C-style array */
883 ctx
->ac
.i32_0
, /* first entry of struct */
884 LLVMConstInt(ctx
->ac
.i32
, out_idx
, false),
886 LLVMValueRef ptr
= LLVMBuildGEP(builder
, vertexptr
, gep_idx
, 3, "");
888 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
889 LLVMBuildStore(builder
, out_val
, ptr
);
892 assert(out_idx
* 4 == sel
->gsvs_vertex_size
);
894 /* Determine and store whether this vertex completed a primitive. */
895 const LLVMValueRef curverts
= LLVMBuildLoad(builder
, ctx
->gs_curprim_verts
[stream
], "");
897 tmp
= LLVMConstInt(ctx
->ac
.i32
, u_vertices_per_prim(sel
->gs_output_prim
) - 1, false);
898 const LLVMValueRef iscompleteprim
=
899 LLVMBuildICmp(builder
, LLVMIntUGE
, curverts
, tmp
, "");
901 tmp
= LLVMBuildAdd(builder
, curverts
, ctx
->ac
.i32_1
, "");
902 LLVMBuildStore(builder
, tmp
, ctx
->gs_curprim_verts
[stream
]);
904 LLVMValueRef gep_idx
[3] = {
905 ctx
->ac
.i32_0
, /* implied C-style array */
906 ctx
->ac
.i32_1
, /* second struct entry */
907 LLVMConstInt(ctx
->ac
.i32
, stream
, false),
909 const LLVMValueRef primflagptr
=
910 LLVMBuildGEP(builder
, vertexptr
, gep_idx
, 3, "");
912 tmp
= LLVMBuildZExt(builder
, iscompleteprim
, ctx
->ac
.i8
, "");
913 LLVMBuildStore(builder
, tmp
, primflagptr
);
915 tmp
= LLVMBuildLoad(builder
, ctx
->gs_generated_prims
[stream
], "");
916 tmp
= LLVMBuildAdd(builder
, tmp
, LLVMBuildZExt(builder
, iscompleteprim
, ctx
->ac
.i32
, ""), "");
917 LLVMBuildStore(builder
, tmp
, ctx
->gs_generated_prims
[stream
]);
919 lp_build_endif(&if_state
);
922 void gfx10_ngg_gs_emit_prologue(struct si_shader_context
*ctx
)
924 /* Zero out the part of LDS scratch that is used to accumulate the
925 * per-stream generated primitive count.
927 LLVMBuilderRef builder
= ctx
->ac
.builder
;
928 LLVMValueRef scratchptr
= ctx
->gs_ngg_scratch
;
929 LLVMValueRef tid
= get_thread_id_in_tg(ctx
);
932 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, LLVMConstInt(ctx
->i32
, 4, false), "");
933 ac_build_ifcc(&ctx
->ac
, tmp
, 5090);
935 LLVMValueRef ptr
= ac_build_gep0(&ctx
->ac
, scratchptr
, tid
);
936 LLVMBuildStore(builder
, ctx
->i32_0
, ptr
);
938 ac_build_endif(&ctx
->ac
, 5090);
940 ac_build_s_barrier(&ctx
->ac
);
943 void gfx10_ngg_gs_emit_epilogue(struct si_shader_context
*ctx
)
945 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
946 const struct tgsi_shader_info
*info
= &sel
->info
;
947 const unsigned verts_per_prim
= u_vertices_per_prim(sel
->gs_output_prim
);
948 LLVMBuilderRef builder
= ctx
->ac
.builder
;
949 LLVMValueRef i8_0
= LLVMConstInt(ctx
->ac
.i8
, 0, false);
950 LLVMValueRef tmp
, tmp2
;
952 /* Zero out remaining (non-emitted) primitive flags.
954 * Note: Alternatively, we could pass the relevant gs_next_vertex to
955 * the emit threads via LDS. This is likely worse in the expected
956 * typical case where each GS thread emits the full set of
959 for (unsigned stream
= 0; stream
< 4; ++stream
) {
960 if (!info
->num_stream_output_components
[stream
])
963 const LLVMValueRef gsthread
= get_thread_id_in_tg(ctx
);
965 ac_build_bgnloop(&ctx
->ac
, 5100);
967 const LLVMValueRef vertexidx
=
968 LLVMBuildLoad(builder
, ctx
->gs_next_vertex
[stream
], "");
969 tmp
= LLVMBuildICmp(builder
, LLVMIntUGE
, vertexidx
,
970 LLVMConstInt(ctx
->ac
.i32
, sel
->gs_max_out_vertices
, false), "");
971 ac_build_ifcc(&ctx
->ac
, tmp
, 5101);
972 ac_build_break(&ctx
->ac
);
973 ac_build_endif(&ctx
->ac
, 5101);
975 tmp
= LLVMBuildAdd(builder
, vertexidx
, ctx
->ac
.i32_1
, "");
976 LLVMBuildStore(builder
, tmp
, ctx
->gs_next_vertex
[stream
]);
978 tmp
= ngg_gs_emit_vertex_ptr(ctx
, gsthread
, vertexidx
);
979 LLVMValueRef gep_idx
[3] = {
980 ctx
->ac
.i32_0
, /* implied C-style array */
981 ctx
->ac
.i32_1
, /* second entry of struct */
982 LLVMConstInt(ctx
->ac
.i32
, stream
, false),
984 tmp
= LLVMBuildGEP(builder
, tmp
, gep_idx
, 3, "");
985 LLVMBuildStore(builder
, i8_0
, tmp
);
987 ac_build_endloop(&ctx
->ac
, 5100);
990 /* Accumulate generated primitives counts across the entire threadgroup. */
991 for (unsigned stream
= 0; stream
< 4; ++stream
) {
992 if (!info
->num_stream_output_components
[stream
])
995 LLVMValueRef numprims
=
996 LLVMBuildLoad(builder
, ctx
->gs_generated_prims
[stream
], "");
997 numprims
= ac_build_reduce(&ctx
->ac
, numprims
, nir_op_iadd
, 64);
999 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, ac_get_thread_id(&ctx
->ac
), ctx
->i32_0
, "");
1000 ac_build_ifcc(&ctx
->ac
, tmp
, 5105);
1002 LLVMBuildAtomicRMW(builder
, LLVMAtomicRMWBinOpAdd
,
1003 ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
,
1004 LLVMConstInt(ctx
->i32
, stream
, false)),
1005 numprims
, LLVMAtomicOrderingMonotonic
, false);
1007 ac_build_endif(&ctx
->ac
, 5105);
1010 lp_build_endif(&ctx
->merged_wrap_if_state
);
1012 ac_build_s_barrier(&ctx
->ac
);
1014 const LLVMValueRef tid
= get_thread_id_in_tg(ctx
);
1015 LLVMValueRef num_emit_threads
= ngg_get_prim_cnt(ctx
);
1018 if (sel
->so
.num_outputs
) {
1019 struct ngg_streamout nggso
= {};
1021 nggso
.num_vertices
= LLVMConstInt(ctx
->i32
, verts_per_prim
, false);
1023 LLVMValueRef vertexptr
= ngg_gs_vertex_ptr(ctx
, tid
);
1024 for (unsigned stream
= 0; stream
< 4; ++stream
) {
1025 if (!info
->num_stream_output_components
[stream
])
1028 LLVMValueRef gep_idx
[3] = {
1029 ctx
->i32_0
, /* implicit C-style array */
1030 ctx
->i32_1
, /* second value of struct */
1031 LLVMConstInt(ctx
->i32
, stream
, false),
1033 tmp
= LLVMBuildGEP(builder
, vertexptr
, gep_idx
, 3, "");
1034 tmp
= LLVMBuildLoad(builder
, tmp
, "");
1035 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->i1
, "");
1036 tmp2
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, num_emit_threads
, "");
1037 nggso
.prim_enable
[stream
] = LLVMBuildAnd(builder
, tmp
, tmp2
, "");
1040 for (unsigned i
= 0; i
< verts_per_prim
; ++i
) {
1041 tmp
= LLVMBuildSub(builder
, tid
,
1042 LLVMConstInt(ctx
->i32
, verts_per_prim
- i
- 1, false), "");
1043 tmp
= ngg_gs_vertex_ptr(ctx
, tmp
);
1044 nggso
.vertices
[i
] = ac_build_gep0(&ctx
->ac
, tmp
, ctx
->i32_0
);
1047 build_streamout(ctx
, &nggso
);
1050 /* Write shader query data. */
1051 tmp
= si_unpack_param(ctx
, ctx
->param_vs_state_bits
, 6, 1);
1052 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->i1
, "");
1053 ac_build_ifcc(&ctx
->ac
, tmp
, 5109); /* if (STREAMOUT_QUERY_ENABLED) */
1054 unsigned num_query_comps
= sel
->so
.num_outputs
? 8 : 4;
1055 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
,
1056 LLVMConstInt(ctx
->i32
, num_query_comps
, false), "");
1057 ac_build_ifcc(&ctx
->ac
, tmp
, 5110);
1059 LLVMValueRef offset
;
1061 if (sel
->so
.num_outputs
)
1062 tmp
= LLVMBuildAnd(builder
, tmp
, LLVMConstInt(ctx
->i32
, 3, false), "");
1063 offset
= LLVMBuildNUWMul(builder
, tmp
, LLVMConstInt(ctx
->i32
, 32, false), "");
1064 if (sel
->so
.num_outputs
) {
1065 tmp
= LLVMBuildLShr(builder
, tid
, LLVMConstInt(ctx
->i32
, 2, false), "");
1066 tmp
= LLVMBuildNUWMul(builder
, tmp
, LLVMConstInt(ctx
->i32
, 8, false), "");
1067 offset
= LLVMBuildAdd(builder
, offset
, tmp
, "");
1070 tmp
= LLVMBuildLoad(builder
, ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, tid
), "");
1071 LLVMValueRef args
[] = {
1073 ngg_get_query_buf(ctx
),
1075 LLVMConstInt(ctx
->i32
, 16, false), /* soffset */
1076 ctx
->i32_0
, /* cachepolicy */
1078 ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.raw.buffer.atomic.add.i32",
1079 ctx
->i32
, args
, 5, 0);
1081 ac_build_endif(&ctx
->ac
, 5110);
1082 ac_build_endif(&ctx
->ac
, 5109);
1086 /* Determine vertex liveness. */
1087 LLVMValueRef vertliveptr
= lp_build_alloca(&ctx
->gallivm
, ctx
->ac
.i1
, "vertexlive");
1089 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, num_emit_threads
, "");
1090 ac_build_ifcc(&ctx
->ac
, tmp
, 5120);
1092 for (unsigned i
= 0; i
< verts_per_prim
; ++i
) {
1093 const LLVMValueRef primidx
=
1094 LLVMBuildAdd(builder
, tid
,
1095 LLVMConstInt(ctx
->ac
.i32
, i
, false), "");
1098 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, primidx
, num_emit_threads
, "");
1099 ac_build_ifcc(&ctx
->ac
, tmp
, 5121 + i
);
1102 /* Load primitive liveness */
1103 tmp
= ngg_gs_vertex_ptr(ctx
, primidx
);
1104 LLVMValueRef gep_idx
[3] = {
1105 ctx
->ac
.i32_0
, /* implicit C-style array */
1106 ctx
->ac
.i32_1
, /* second value of struct */
1107 ctx
->ac
.i32_0
, /* stream 0 */
1109 tmp
= LLVMBuildGEP(builder
, tmp
, gep_idx
, 3, "");
1110 tmp
= LLVMBuildLoad(builder
, tmp
, "");
1111 const LLVMValueRef primlive
=
1112 LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
1114 tmp
= LLVMBuildLoad(builder
, vertliveptr
, "");
1115 tmp
= LLVMBuildOr(builder
, tmp
, primlive
, ""),
1116 LLVMBuildStore(builder
, tmp
, vertliveptr
);
1119 ac_build_endif(&ctx
->ac
, 5121 + i
);
1122 ac_build_endif(&ctx
->ac
, 5120);
1124 /* Inclusive scan addition across the current wave. */
1125 LLVMValueRef vertlive
= LLVMBuildLoad(builder
, vertliveptr
, "");
1126 struct ac_wg_scan vertlive_scan
= {};
1127 vertlive_scan
.op
= nir_op_iadd
;
1128 vertlive_scan
.enable_reduce
= true;
1129 vertlive_scan
.enable_exclusive
= true;
1130 vertlive_scan
.src
= vertlive
;
1131 vertlive_scan
.scratch
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, ctx
->i32_0
);
1132 vertlive_scan
.waveidx
= get_wave_id_in_tg(ctx
);
1133 vertlive_scan
.numwaves
= get_tgsize(ctx
);
1134 vertlive_scan
.maxwaves
= 8;
1136 ac_build_wg_scan(&ctx
->ac
, &vertlive_scan
);
1138 /* Skip all exports (including index exports) when possible. At least on
1139 * early gfx10 revisions this is also to avoid hangs.
1141 LLVMValueRef have_exports
=
1142 LLVMBuildICmp(builder
, LLVMIntNE
, vertlive_scan
.result_reduce
, ctx
->ac
.i32_0
, "");
1144 LLVMBuildSelect(builder
, have_exports
, num_emit_threads
, ctx
->ac
.i32_0
, "");
1146 /* Allocate export space. Send this message as early as possible, to
1147 * hide the latency of the SQ <-> SPI roundtrip.
1149 * Note: We could consider compacting primitives for export as well.
1150 * PA processes 1 non-null prim / clock, but it fetches 4 DW of
1151 * prim data per clock and skips null primitives at no additional
1152 * cost. So compacting primitives can only be beneficial when
1153 * there are 4 or more contiguous null primitives in the export
1154 * (in the common case of single-dword prim exports).
1156 build_sendmsg_gs_alloc_req(ctx
, vertlive_scan
.result_reduce
, num_emit_threads
);
1158 /* Setup the reverse vertex compaction permutation. We re-use stream 1
1159 * of the primitive liveness flags, relying on the fact that each
1160 * threadgroup can have at most 256 threads. */
1161 ac_build_ifcc(&ctx
->ac
, vertlive
, 5130);
1163 tmp
= ngg_gs_vertex_ptr(ctx
, vertlive_scan
.result_exclusive
);
1164 LLVMValueRef gep_idx
[3] = {
1165 ctx
->ac
.i32_0
, /* implicit C-style array */
1166 ctx
->ac
.i32_1
, /* second value of struct */
1167 ctx
->ac
.i32_1
, /* stream 1 */
1169 tmp
= LLVMBuildGEP(builder
, tmp
, gep_idx
, 3, "");
1170 tmp2
= LLVMBuildTrunc(builder
, tid
, ctx
->ac
.i8
, "");
1171 LLVMBuildStore(builder
, tmp2
, tmp
);
1173 ac_build_endif(&ctx
->ac
, 5130);
1175 ac_build_s_barrier(&ctx
->ac
);
1177 /* Export primitive data */
1178 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, num_emit_threads
, "");
1179 ac_build_ifcc(&ctx
->ac
, tmp
, 5140);
1181 struct ngg_prim prim
= {};
1182 prim
.num_vertices
= verts_per_prim
;
1184 tmp
= ngg_gs_vertex_ptr(ctx
, tid
);
1185 LLVMValueRef gep_idx
[3] = {
1186 ctx
->ac
.i32_0
, /* implicit C-style array */
1187 ctx
->ac
.i32_1
, /* second value of struct */
1188 ctx
->ac
.i32_0
, /* primflag */
1190 tmp
= LLVMBuildGEP(builder
, tmp
, gep_idx
, 3, "");
1191 tmp
= LLVMBuildLoad(builder
, tmp
, "");
1192 prim
.isnull
= LLVMBuildICmp(builder
, LLVMIntEQ
, tmp
,
1193 LLVMConstInt(ctx
->ac
.i8
, 0, false), "");
1195 for (unsigned i
= 0; i
< verts_per_prim
; ++i
) {
1196 prim
.index
[i
] = LLVMBuildSub(builder
, vertlive_scan
.result_exclusive
,
1197 LLVMConstInt(ctx
->ac
.i32
, verts_per_prim
- i
- 1, false), "");
1198 prim
.edgeflag
[i
] = ctx
->ac
.i1false
;
1201 build_export_prim(ctx
, &prim
);
1203 ac_build_endif(&ctx
->ac
, 5140);
1205 /* Export position and parameter data */
1206 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, vertlive_scan
.result_reduce
, "");
1207 ac_build_ifcc(&ctx
->ac
, tmp
, 5145);
1209 struct si_shader_output_values
*outputs
= NULL
;
1210 outputs
= MALLOC(info
->num_outputs
* sizeof(outputs
[0]));
1212 tmp
= ngg_gs_vertex_ptr(ctx
, tid
);
1213 LLVMValueRef gep_idx
[3] = {
1214 ctx
->ac
.i32_0
, /* implicit C-style array */
1215 ctx
->ac
.i32_1
, /* second value of struct */
1216 ctx
->ac
.i32_1
, /* stream 1: source data index */
1218 tmp
= LLVMBuildGEP(builder
, tmp
, gep_idx
, 3, "");
1219 tmp
= LLVMBuildLoad(builder
, tmp
, "");
1220 tmp
= LLVMBuildZExt(builder
, tmp
, ctx
->ac
.i32
, "");
1221 const LLVMValueRef vertexptr
= ngg_gs_vertex_ptr(ctx
, tmp
);
1223 unsigned out_idx
= 0;
1224 gep_idx
[1] = ctx
->ac
.i32_0
;
1225 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
1226 outputs
[i
].semantic_name
= info
->output_semantic_name
[i
];
1227 outputs
[i
].semantic_index
= info
->output_semantic_index
[i
];
1229 for (unsigned j
= 0; j
< 4; j
++, out_idx
++) {
1230 gep_idx
[2] = LLVMConstInt(ctx
->ac
.i32
, out_idx
, false);
1231 tmp
= LLVMBuildGEP(builder
, vertexptr
, gep_idx
, 3, "");
1232 tmp
= LLVMBuildLoad(builder
, tmp
, "");
1233 outputs
[i
].values
[j
] = ac_to_float(&ctx
->ac
, tmp
);
1234 outputs
[i
].vertex_stream
[j
] =
1235 (info
->output_streams
[i
] >> (2 * j
)) & 3;
1239 si_llvm_export_vs(ctx
, outputs
, info
->num_outputs
);
1243 ac_build_endif(&ctx
->ac
, 5145);
1246 static void clamp_gsprims_to_esverts(unsigned *max_gsprims
, unsigned max_esverts
,
1247 unsigned min_verts_per_prim
, bool use_adjacency
)
1249 unsigned max_reuse
= max_esverts
- min_verts_per_prim
;
1252 *max_gsprims
= MIN2(*max_gsprims
, 1 + max_reuse
);
1256 * Determine subgroup information like maximum number of vertices and prims.
1258 * This happens before the shader is uploaded, since LDS relocations during
1259 * upload depend on the subgroup size.
1261 void gfx10_ngg_calculate_subgroup_info(struct si_shader
*shader
)
1263 const struct si_shader_selector
*gs_sel
= shader
->selector
;
1264 const struct si_shader_selector
*es_sel
=
1265 shader
->previous_stage_sel
? shader
->previous_stage_sel
: gs_sel
;
1266 const enum pipe_shader_type gs_type
= gs_sel
->type
;
1267 const unsigned gs_num_invocations
= MAX2(gs_sel
->gs_num_invocations
, 1);
1268 /* TODO: Specialize for known primitive type without GS. */
1269 const unsigned input_prim
= gs_type
== PIPE_SHADER_GEOMETRY
?
1270 gs_sel
->info
.properties
[TGSI_PROPERTY_GS_INPUT_PRIM
] :
1271 PIPE_PRIM_TRIANGLES
;
1272 const bool use_adjacency
= input_prim
>= PIPE_PRIM_LINES_ADJACENCY
&&
1273 input_prim
<= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
;
1274 const unsigned max_verts_per_prim
= u_vertices_per_prim(input_prim
);
1275 const unsigned min_verts_per_prim
=
1276 gs_type
== PIPE_SHADER_GEOMETRY
? max_verts_per_prim
: 1;
1278 /* All these are in dwords: */
1279 /* We can't allow using the whole LDS, because GS waves compete with
1280 * other shader stages for LDS space.
1282 * Streamout can increase the ESGS buffer size later on, so be more
1283 * conservative with streamout and use 4K dwords. This may be suboptimal.
1285 * Otherwise, use the limit of 7K dwords. The reason is that we need
1286 * to leave some headroom for the max_esverts increase at the end.
1288 * TODO: We should really take the shader's internal LDS use into
1289 * account. The linker will fail if the size is greater than
1292 const unsigned max_lds_size
= (gs_sel
->so
.num_outputs
? 4 : 7) * 1024 - 128;
1293 const unsigned target_lds_size
= max_lds_size
;
1294 unsigned esvert_lds_size
= 0;
1295 unsigned gsprim_lds_size
= 0;
1297 /* All these are per subgroup: */
1298 bool max_vert_out_per_gs_instance
= false;
1299 unsigned max_esverts_base
= 256;
1300 unsigned max_gsprims_base
= 128; /* default prim group size clamp */
1302 /* Hardware has the following non-natural restrictions on the value
1303 * of GE_CNTL.VERT_GRP_SIZE based on based on the primitive type of
1305 * - at most 252 for any line input primitive type
1306 * - at most 251 for any quad input primitive type
1307 * - at most 251 for triangle strips with adjacency (this happens to
1308 * be the natural limit for triangle *lists* with adjacency)
1310 max_esverts_base
= MIN2(max_esverts_base
, 251 + max_verts_per_prim
- 1);
1312 if (gs_type
== PIPE_SHADER_GEOMETRY
) {
1313 unsigned max_out_verts_per_gsprim
=
1314 gs_sel
->gs_max_out_vertices
* gs_num_invocations
;
1316 if (max_out_verts_per_gsprim
<= 256) {
1317 if (max_out_verts_per_gsprim
) {
1318 max_gsprims_base
= MIN2(max_gsprims_base
,
1319 256 / max_out_verts_per_gsprim
);
1322 /* Use special multi-cycling mode in which each GS
1323 * instance gets its own subgroup. Does not work with
1325 max_vert_out_per_gs_instance
= true;
1326 max_gsprims_base
= 1;
1327 max_out_verts_per_gsprim
= gs_sel
->gs_max_out_vertices
;
1330 esvert_lds_size
= es_sel
->esgs_itemsize
/ 4;
1331 gsprim_lds_size
= (gs_sel
->gsvs_vertex_size
/ 4 + 1) * max_out_verts_per_gsprim
;
1333 /* TODO: This needs to be adjusted once LDS use for compaction
1334 * after culling is implemented. */
1335 if (es_sel
->so
.num_outputs
)
1336 esvert_lds_size
= 4 * es_sel
->info
.num_outputs
+ 1;
1339 unsigned max_gsprims
= max_gsprims_base
;
1340 unsigned max_esverts
= max_esverts_base
;
1342 if (esvert_lds_size
)
1343 max_esverts
= MIN2(max_esverts
, target_lds_size
/ esvert_lds_size
);
1344 if (gsprim_lds_size
)
1345 max_gsprims
= MIN2(max_gsprims
, target_lds_size
/ gsprim_lds_size
);
1347 max_esverts
= MIN2(max_esverts
, max_gsprims
* max_verts_per_prim
);
1348 clamp_gsprims_to_esverts(&max_gsprims
, max_esverts
, min_verts_per_prim
, use_adjacency
);
1349 assert(max_esverts
>= max_verts_per_prim
&& max_gsprims
>= 1);
1351 if (esvert_lds_size
|| gsprim_lds_size
) {
1352 /* Now that we have a rough proportionality between esverts
1353 * and gsprims based on the primitive type, scale both of them
1354 * down simultaneously based on required LDS space.
1356 * We could be smarter about this if we knew how much vertex
1359 unsigned lds_total
= max_esverts
* esvert_lds_size
+
1360 max_gsprims
* gsprim_lds_size
;
1361 if (lds_total
> target_lds_size
) {
1362 max_esverts
= max_esverts
* target_lds_size
/ lds_total
;
1363 max_gsprims
= max_gsprims
* target_lds_size
/ lds_total
;
1365 max_esverts
= MIN2(max_esverts
, max_gsprims
* max_verts_per_prim
);
1366 clamp_gsprims_to_esverts(&max_gsprims
, max_esverts
,
1367 min_verts_per_prim
, use_adjacency
);
1368 assert(max_esverts
>= max_verts_per_prim
&& max_gsprims
>= 1);
1372 /* Round up towards full wave sizes for better ALU utilization. */
1373 if (!max_vert_out_per_gs_instance
) {
1374 const unsigned wavesize
= 64;
1375 unsigned orig_max_esverts
;
1376 unsigned orig_max_gsprims
;
1378 orig_max_esverts
= max_esverts
;
1379 orig_max_gsprims
= max_gsprims
;
1381 max_esverts
= align(max_esverts
, wavesize
);
1382 max_esverts
= MIN2(max_esverts
, max_esverts_base
);
1383 if (esvert_lds_size
)
1384 max_esverts
= MIN2(max_esverts
,
1385 (max_lds_size
- max_gsprims
* gsprim_lds_size
) /
1387 max_esverts
= MIN2(max_esverts
, max_gsprims
* max_verts_per_prim
);
1389 max_gsprims
= align(max_gsprims
, wavesize
);
1390 max_gsprims
= MIN2(max_gsprims
, max_gsprims_base
);
1391 if (gsprim_lds_size
)
1392 max_gsprims
= MIN2(max_gsprims
,
1393 (max_lds_size
- max_esverts
* esvert_lds_size
) /
1395 clamp_gsprims_to_esverts(&max_gsprims
, max_esverts
,
1396 min_verts_per_prim
, use_adjacency
);
1397 assert(max_esverts
>= max_verts_per_prim
&& max_gsprims
>= 1);
1398 } while (orig_max_esverts
!= max_esverts
|| orig_max_gsprims
!= max_gsprims
);
1401 /* Hardware restriction: minimum value of max_esverts */
1402 max_esverts
= MAX2(max_esverts
, 23 + max_verts_per_prim
);
1404 unsigned max_out_vertices
=
1405 max_vert_out_per_gs_instance
? gs_sel
->gs_max_out_vertices
:
1406 gs_type
== PIPE_SHADER_GEOMETRY
?
1407 max_gsprims
* gs_num_invocations
* gs_sel
->gs_max_out_vertices
:
1409 assert(max_out_vertices
<= 256);
1411 unsigned prim_amp_factor
= 1;
1412 if (gs_type
== PIPE_SHADER_GEOMETRY
) {
1413 /* Number of output primitives per GS input primitive after
1415 prim_amp_factor
= gs_sel
->gs_max_out_vertices
;
1418 /* The GE only checks against the maximum number of ES verts after
1419 * allocating a full GS primitive. So we need to ensure that whenever
1420 * this check passes, there is enough space for a full primitive without
1423 shader
->ngg
.hw_max_esverts
= max_esverts
- max_verts_per_prim
+ 1;
1424 shader
->ngg
.max_gsprims
= max_gsprims
;
1425 shader
->ngg
.max_out_verts
= max_out_vertices
;
1426 shader
->ngg
.prim_amp_factor
= prim_amp_factor
;
1427 shader
->ngg
.max_vert_out_per_gs_instance
= max_vert_out_per_gs_instance
;
1429 shader
->gs_info
.esgs_ring_size
= 4 * max_esverts
* esvert_lds_size
;
1430 shader
->ngg
.ngg_emit_size
= max_gsprims
* gsprim_lds_size
;
1432 assert(shader
->ngg
.hw_max_esverts
>= 24); /* HW limitation */