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);
499 static unsigned ngg_nogs_vertex_size(struct si_shader
*shader
)
501 unsigned lds_vertex_size
= 0;
503 if (shader
->selector
->so
.num_outputs
)
504 lds_vertex_size
= 4 * shader
->selector
->info
.num_outputs
+ 1;
506 return lds_vertex_size
;
510 * Returns an `[N x i32] addrspace(LDS)*` pointing at contiguous LDS storage
511 * for the vertex outputs.
513 static LLVMValueRef
ngg_nogs_vertex_ptr(struct si_shader_context
*ctx
,
516 /* The extra dword is used to avoid LDS bank conflicts. */
517 unsigned vertex_size
= ngg_nogs_vertex_size(ctx
->shader
);
518 LLVMTypeRef ai32
= LLVMArrayType(ctx
->i32
, vertex_size
);
519 LLVMTypeRef pai32
= LLVMPointerType(ai32
, AC_ADDR_SPACE_LDS
);
520 LLVMValueRef tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->esgs_ring
, pai32
, "");
521 return LLVMBuildGEP(ctx
->ac
.builder
, tmp
, &vtxid
, 1, "");
525 * Emit the epilogue of an API VS or TES shader compiled as ESGS shader.
527 void gfx10_emit_ngg_epilogue(struct ac_shader_abi
*abi
,
528 unsigned max_outputs
,
531 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
532 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
533 struct tgsi_shader_info
*info
= &sel
->info
;
534 struct si_shader_output_values outputs
[PIPE_MAX_SHADER_OUTPUTS
];
535 LLVMBuilderRef builder
= ctx
->ac
.builder
;
536 struct lp_build_if_state if_state
;
537 LLVMValueRef tmp
, tmp2
;
539 assert(!ctx
->shader
->is_gs_copy_shader
);
540 assert(info
->num_outputs
<= max_outputs
);
542 LLVMValueRef vertex_ptr
= NULL
;
544 if (sel
->so
.num_outputs
)
545 vertex_ptr
= ngg_nogs_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
));
547 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
548 outputs
[i
].semantic_name
= info
->output_semantic_name
[i
];
549 outputs
[i
].semantic_index
= info
->output_semantic_index
[i
];
551 for (unsigned j
= 0; j
< 4; j
++) {
552 outputs
[i
].vertex_stream
[j
] =
553 (info
->output_streams
[i
] >> (2 * j
)) & 3;
555 /* TODO: we may store more outputs than streamout needs,
556 * but streamout performance isn't that important.
558 if (sel
->so
.num_outputs
) {
559 tmp
= ac_build_gep0(&ctx
->ac
, vertex_ptr
,
560 LLVMConstInt(ctx
->i32
, 4 * i
+ j
, false));
561 tmp2
= LLVMBuildLoad(builder
, addrs
[4 * i
+ j
], "");
562 tmp2
= ac_to_integer(&ctx
->ac
, tmp2
);
563 LLVMBuildStore(builder
, tmp2
, tmp
);
568 lp_build_endif(&ctx
->merged_wrap_if_state
);
570 LLVMValueRef prims_in_wave
= si_unpack_param(ctx
, ctx
->param_merged_wave_info
, 8, 8);
571 LLVMValueRef vtx_in_wave
= si_unpack_param(ctx
, ctx
->param_merged_wave_info
, 0, 8);
572 LLVMValueRef is_gs_thread
= LLVMBuildICmp(builder
, LLVMIntULT
,
573 ac_get_thread_id(&ctx
->ac
), prims_in_wave
, "");
574 LLVMValueRef is_es_thread
= LLVMBuildICmp(builder
, LLVMIntULT
,
575 ac_get_thread_id(&ctx
->ac
), vtx_in_wave
, "");
576 LLVMValueRef vtxindex
[] = {
577 si_unpack_param(ctx
, ctx
->param_gs_vtx01_offset
, 0, 16),
578 si_unpack_param(ctx
, ctx
->param_gs_vtx01_offset
, 16, 16),
579 si_unpack_param(ctx
, ctx
->param_gs_vtx23_offset
, 0, 16),
582 /* Determine the number of vertices per primitive. */
583 unsigned num_vertices
;
584 LLVMValueRef num_vertices_val
;
586 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
587 if (info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS
]) {
588 /* Blits always use axis-aligned rectangles with 3 vertices. */
590 num_vertices_val
= LLVMConstInt(ctx
->i32
, 3, 0);
592 /* Extract OUTPRIM field. */
593 tmp
= si_unpack_param(ctx
, ctx
->param_vs_state_bits
, 2, 2);
594 num_vertices_val
= LLVMBuildAdd(builder
, tmp
, ctx
->i32_1
, "");
595 num_vertices
= 3; /* TODO: optimize for points & lines */
598 assert(ctx
->type
== PIPE_SHADER_TESS_EVAL
);
600 if (info
->properties
[TGSI_PROPERTY_TES_POINT_MODE
])
602 else if (info
->properties
[TGSI_PROPERTY_TES_PRIM_MODE
] == PIPE_PRIM_LINES
)
607 num_vertices_val
= LLVMConstInt(ctx
->i32
, num_vertices
, false);
611 LLVMValueRef emitted_prims
= NULL
;
613 if (sel
->so
.num_outputs
) {
614 struct ngg_streamout nggso
= {};
616 nggso
.num_vertices
= num_vertices_val
;
617 nggso
.prim_enable
[0] = is_gs_thread
;
619 for (unsigned i
= 0; i
< num_vertices
; ++i
)
620 nggso
.vertices
[i
] = ngg_nogs_vertex_ptr(ctx
, vtxindex
[i
]);
622 build_streamout(ctx
, &nggso
);
623 emitted_prims
= nggso
.emit
[0];
626 /* Copy Primitive IDs from GS threads to the LDS address corresponding
627 * to the ES thread of the provoking vertex.
629 if (ctx
->type
== PIPE_SHADER_VERTEX
&&
630 ctx
->shader
->key
.mono
.u
.vs_export_prim_id
) {
631 /* Streamout uses LDS. We need to wait for it before we can reuse it. */
632 if (sel
->so
.num_outputs
)
633 ac_build_s_barrier(&ctx
->ac
);
635 ac_build_ifcc(&ctx
->ac
, is_gs_thread
, 5400);
636 /* Extract the PROVOKING_VTX_INDEX field. */
637 LLVMValueRef provoking_vtx_in_prim
=
638 si_unpack_param(ctx
, ctx
->param_vs_state_bits
, 4, 2);
640 /* provoking_vtx_index = vtxindex[provoking_vtx_in_prim]; */
641 LLVMValueRef indices
= ac_build_gather_values(&ctx
->ac
, vtxindex
, 3);
642 LLVMValueRef provoking_vtx_index
=
643 LLVMBuildExtractElement(builder
, indices
, provoking_vtx_in_prim
, "");
645 LLVMBuildStore(builder
, ctx
->abi
.gs_prim_id
,
646 ac_build_gep0(&ctx
->ac
, ctx
->esgs_ring
, provoking_vtx_index
));
647 ac_build_endif(&ctx
->ac
, 5400);
650 /* TODO: primitive culling */
652 build_sendmsg_gs_alloc_req(ctx
, ngg_get_vtx_cnt(ctx
), ngg_get_prim_cnt(ctx
));
654 /* Update query buffer */
655 /* TODO: this won't catch 96-bit clear_buffer via transform feedback. */
656 if (!info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS
]) {
657 tmp
= si_unpack_param(ctx
, ctx
->param_vs_state_bits
, 6, 1);
658 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->i1
, "");
659 ac_build_ifcc(&ctx
->ac
, tmp
, 5029); /* if (STREAMOUT_QUERY_ENABLED) */
660 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, get_wave_id_in_tg(ctx
), ctx
->ac
.i32_0
, "");
661 ac_build_ifcc(&ctx
->ac
, tmp
, 5030);
662 tmp
= LLVMBuildICmp(builder
, LLVMIntULE
, ac_get_thread_id(&ctx
->ac
),
663 sel
->so
.num_outputs
? ctx
->ac
.i32_1
: ctx
->ac
.i32_0
, "");
664 ac_build_ifcc(&ctx
->ac
, tmp
, 5031);
666 LLVMValueRef args
[] = {
667 ngg_get_prim_cnt(ctx
),
668 ngg_get_query_buf(ctx
),
669 LLVMConstInt(ctx
->i32
, 16, false), /* offset of stream[0].generated_primitives */
670 ctx
->i32_0
, /* soffset */
671 ctx
->i32_0
, /* cachepolicy */
674 if (sel
->so
.num_outputs
) {
675 args
[0] = ac_build_writelane(&ctx
->ac
, args
[0], emitted_prims
, ctx
->i32_1
);
676 args
[2] = ac_build_writelane(&ctx
->ac
, args
[2],
677 LLVMConstInt(ctx
->i32
, 24, false), ctx
->i32_1
);
680 /* TODO: should this be 64-bit atomics? */
681 ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.raw.buffer.atomic.add.i32",
682 ctx
->i32
, args
, 5, 0);
684 ac_build_endif(&ctx
->ac
, 5031);
685 ac_build_endif(&ctx
->ac
, 5030);
686 ac_build_endif(&ctx
->ac
, 5029);
689 /* Export primitive data to the index buffer. Format is:
690 * - bits 0..8: index 0
691 * - bit 9: edge flag 0
692 * - bits 10..18: index 1
693 * - bit 19: edge flag 1
694 * - bits 20..28: index 2
695 * - bit 29: edge flag 2
696 * - bit 31: null primitive (skip)
698 * For the first version, we will always build up all three indices
699 * independent of the primitive type. The additional garbage data
702 * TODO: culling depends on the primitive type, so can have some
705 lp_build_if(&if_state
, &ctx
->gallivm
, is_gs_thread
);
707 struct ngg_prim prim
= {};
709 prim
.num_vertices
= num_vertices
;
710 prim
.isnull
= ctx
->ac
.i1false
;
711 memcpy(prim
.index
, vtxindex
, sizeof(vtxindex
[0]) * 3);
713 for (unsigned i
= 0; i
< num_vertices
; ++i
) {
714 tmp
= LLVMBuildLShr(builder
, ctx
->abi
.gs_invocation_id
,
715 LLVMConstInt(ctx
->ac
.i32
, 8 + i
, false), "");
716 prim
.edgeflag
[i
] = LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
719 build_export_prim(ctx
, &prim
);
721 lp_build_endif(&if_state
);
723 /* Export per-vertex data (positions and parameters). */
724 lp_build_if(&if_state
, &ctx
->gallivm
, is_es_thread
);
728 /* Unconditionally (re-)load the values for proper SSA form. */
729 for (i
= 0; i
< info
->num_outputs
; i
++) {
730 for (unsigned j
= 0; j
< 4; j
++) {
731 outputs
[i
].values
[j
] =
732 LLVMBuildLoad(builder
,
738 if (ctx
->shader
->key
.mono
.u
.vs_export_prim_id
) {
739 outputs
[i
].semantic_name
= TGSI_SEMANTIC_PRIMID
;
740 outputs
[i
].semantic_index
= 0;
742 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
743 /* Wait for GS stores to finish. */
744 ac_build_s_barrier(&ctx
->ac
);
746 tmp
= ac_build_gep0(&ctx
->ac
, ctx
->esgs_ring
,
747 get_thread_id_in_tg(ctx
));
748 outputs
[i
].values
[0] = LLVMBuildLoad(builder
, tmp
, "");
750 assert(ctx
->type
== PIPE_SHADER_TESS_EVAL
);
751 outputs
[i
].values
[0] = si_get_primitive_id(ctx
, 0);
754 outputs
[i
].values
[0] = ac_to_float(&ctx
->ac
, outputs
[i
].values
[0]);
755 for (unsigned j
= 1; j
< 4; j
++)
756 outputs
[i
].values
[j
] = LLVMGetUndef(ctx
->f32
);
758 memset(outputs
[i
].vertex_stream
, 0,
759 sizeof(outputs
[i
].vertex_stream
));
763 si_llvm_export_vs(ctx
, outputs
, i
);
765 lp_build_endif(&if_state
);
769 ngg_gs_get_vertex_storage(struct si_shader_context
*ctx
)
771 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
772 const struct tgsi_shader_info
*info
= &sel
->info
;
774 LLVMTypeRef elements
[2] = {
775 LLVMArrayType(ctx
->ac
.i32
, 4 * info
->num_outputs
),
776 LLVMArrayType(ctx
->ac
.i8
, 4),
778 LLVMTypeRef type
= LLVMStructTypeInContext(ctx
->ac
.context
, elements
, 2, false);
779 type
= LLVMPointerType(LLVMArrayType(type
, 0), AC_ADDR_SPACE_LDS
);
780 return LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->gs_ngg_emit
, type
, "");
784 * Return a pointer to the LDS storage reserved for the N'th vertex, where N
785 * is in emit order; that is:
786 * - during the epilogue, N is the threadidx (relative to the entire threadgroup)
787 * - during vertex emit, i.e. while the API GS shader invocation is running,
788 * N = threadidx * gs_max_out_vertices + emitidx
790 * Goals of the LDS memory layout:
791 * 1. Eliminate bank conflicts on write for geometry shaders that have all emits
792 * in uniform control flow
793 * 2. Eliminate bank conflicts on read for export if, additionally, there is no
795 * 3. Agnostic to the number of waves (since we don't know it before compiling)
796 * 4. Allow coalescing of LDS instructions (ds_write_b128 etc.)
797 * 5. Avoid wasting memory.
799 * We use an AoS layout due to point 4 (this also helps point 3). In an AoS
800 * layout, elimination of bank conflicts requires that each vertex occupy an
801 * odd number of dwords. We use the additional dword to store the output stream
802 * index as well as a flag to indicate whether this vertex ends a primitive
805 * Swizzling is required to satisfy points 1 and 2 simultaneously.
807 * Vertices are stored in export order (gsthread * gs_max_out_vertices + emitidx).
808 * Indices are swizzled in groups of 32, which ensures point 1 without
809 * disturbing point 2.
811 * \return an LDS pointer to type {[N x i32], [4 x i8]}
814 ngg_gs_vertex_ptr(struct si_shader_context
*ctx
, LLVMValueRef vertexidx
)
816 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
817 LLVMBuilderRef builder
= ctx
->ac
.builder
;
818 LLVMValueRef storage
= ngg_gs_get_vertex_storage(ctx
);
820 /* gs_max_out_vertices = 2^(write_stride_2exp) * some odd number */
821 unsigned write_stride_2exp
= ffs(sel
->gs_max_out_vertices
) - 1;
822 if (write_stride_2exp
) {
824 LLVMBuildLShr(builder
, vertexidx
,
825 LLVMConstInt(ctx
->ac
.i32
, 5, false), "");
826 LLVMValueRef swizzle
=
827 LLVMBuildAnd(builder
, row
,
828 LLVMConstInt(ctx
->ac
.i32
, (1u << write_stride_2exp
) - 1,
830 vertexidx
= LLVMBuildXor(builder
, vertexidx
, swizzle
, "");
833 return ac_build_gep0(&ctx
->ac
, storage
, vertexidx
);
837 ngg_gs_emit_vertex_ptr(struct si_shader_context
*ctx
, LLVMValueRef gsthread
,
838 LLVMValueRef emitidx
)
840 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
841 LLVMBuilderRef builder
= ctx
->ac
.builder
;
844 tmp
= LLVMConstInt(ctx
->ac
.i32
, sel
->gs_max_out_vertices
, false);
845 tmp
= LLVMBuildMul(builder
, tmp
, gsthread
, "");
846 const LLVMValueRef vertexidx
= LLVMBuildAdd(builder
, tmp
, emitidx
, "");
847 return ngg_gs_vertex_ptr(ctx
, vertexidx
);
850 void gfx10_ngg_gs_emit_vertex(struct si_shader_context
*ctx
,
854 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
855 const struct tgsi_shader_info
*info
= &sel
->info
;
856 LLVMBuilderRef builder
= ctx
->ac
.builder
;
857 struct lp_build_if_state if_state
;
859 const LLVMValueRef vertexidx
=
860 LLVMBuildLoad(builder
, ctx
->gs_next_vertex
[stream
], "");
862 /* If this thread has already emitted the declared maximum number of
863 * vertices, skip the write: excessive vertex emissions are not
864 * supposed to have any effect.
866 const LLVMValueRef can_emit
=
867 LLVMBuildICmp(builder
, LLVMIntULT
, vertexidx
,
868 LLVMConstInt(ctx
->i32
, sel
->gs_max_out_vertices
, false), "");
870 tmp
= LLVMBuildAdd(builder
, vertexidx
, ctx
->ac
.i32_1
, "");
871 tmp
= LLVMBuildSelect(builder
, can_emit
, tmp
, vertexidx
, "");
872 LLVMBuildStore(builder
, tmp
, ctx
->gs_next_vertex
[stream
]);
874 lp_build_if(&if_state
, &ctx
->gallivm
, can_emit
);
876 const LLVMValueRef vertexptr
=
877 ngg_gs_emit_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
), vertexidx
);
878 unsigned out_idx
= 0;
879 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
880 for (unsigned chan
= 0; chan
< 4; chan
++, out_idx
++) {
881 if (!(info
->output_usagemask
[i
] & (1 << chan
)) ||
882 ((info
->output_streams
[i
] >> (2 * chan
)) & 3) != stream
)
885 LLVMValueRef out_val
= LLVMBuildLoad(builder
, addrs
[4 * i
+ chan
], "");
886 LLVMValueRef gep_idx
[3] = {
887 ctx
->ac
.i32_0
, /* implied C-style array */
888 ctx
->ac
.i32_0
, /* first entry of struct */
889 LLVMConstInt(ctx
->ac
.i32
, out_idx
, false),
891 LLVMValueRef ptr
= LLVMBuildGEP(builder
, vertexptr
, gep_idx
, 3, "");
893 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
894 LLVMBuildStore(builder
, out_val
, ptr
);
897 assert(out_idx
* 4 == sel
->gsvs_vertex_size
);
899 /* Determine and store whether this vertex completed a primitive. */
900 const LLVMValueRef curverts
= LLVMBuildLoad(builder
, ctx
->gs_curprim_verts
[stream
], "");
902 tmp
= LLVMConstInt(ctx
->ac
.i32
, u_vertices_per_prim(sel
->gs_output_prim
) - 1, false);
903 const LLVMValueRef iscompleteprim
=
904 LLVMBuildICmp(builder
, LLVMIntUGE
, curverts
, tmp
, "");
906 tmp
= LLVMBuildAdd(builder
, curverts
, ctx
->ac
.i32_1
, "");
907 LLVMBuildStore(builder
, tmp
, ctx
->gs_curprim_verts
[stream
]);
909 LLVMValueRef gep_idx
[3] = {
910 ctx
->ac
.i32_0
, /* implied C-style array */
911 ctx
->ac
.i32_1
, /* second struct entry */
912 LLVMConstInt(ctx
->ac
.i32
, stream
, false),
914 const LLVMValueRef primflagptr
=
915 LLVMBuildGEP(builder
, vertexptr
, gep_idx
, 3, "");
917 tmp
= LLVMBuildZExt(builder
, iscompleteprim
, ctx
->ac
.i8
, "");
918 LLVMBuildStore(builder
, tmp
, primflagptr
);
920 tmp
= LLVMBuildLoad(builder
, ctx
->gs_generated_prims
[stream
], "");
921 tmp
= LLVMBuildAdd(builder
, tmp
, LLVMBuildZExt(builder
, iscompleteprim
, ctx
->ac
.i32
, ""), "");
922 LLVMBuildStore(builder
, tmp
, ctx
->gs_generated_prims
[stream
]);
924 lp_build_endif(&if_state
);
927 void gfx10_ngg_gs_emit_prologue(struct si_shader_context
*ctx
)
929 /* Zero out the part of LDS scratch that is used to accumulate the
930 * per-stream generated primitive count.
932 LLVMBuilderRef builder
= ctx
->ac
.builder
;
933 LLVMValueRef scratchptr
= ctx
->gs_ngg_scratch
;
934 LLVMValueRef tid
= get_thread_id_in_tg(ctx
);
937 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, LLVMConstInt(ctx
->i32
, 4, false), "");
938 ac_build_ifcc(&ctx
->ac
, tmp
, 5090);
940 LLVMValueRef ptr
= ac_build_gep0(&ctx
->ac
, scratchptr
, tid
);
941 LLVMBuildStore(builder
, ctx
->i32_0
, ptr
);
943 ac_build_endif(&ctx
->ac
, 5090);
945 ac_build_s_barrier(&ctx
->ac
);
948 void gfx10_ngg_gs_emit_epilogue(struct si_shader_context
*ctx
)
950 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
951 const struct tgsi_shader_info
*info
= &sel
->info
;
952 const unsigned verts_per_prim
= u_vertices_per_prim(sel
->gs_output_prim
);
953 LLVMBuilderRef builder
= ctx
->ac
.builder
;
954 LLVMValueRef i8_0
= LLVMConstInt(ctx
->ac
.i8
, 0, false);
955 LLVMValueRef tmp
, tmp2
;
957 /* Zero out remaining (non-emitted) primitive flags.
959 * Note: Alternatively, we could pass the relevant gs_next_vertex to
960 * the emit threads via LDS. This is likely worse in the expected
961 * typical case where each GS thread emits the full set of
964 for (unsigned stream
= 0; stream
< 4; ++stream
) {
965 if (!info
->num_stream_output_components
[stream
])
968 const LLVMValueRef gsthread
= get_thread_id_in_tg(ctx
);
970 ac_build_bgnloop(&ctx
->ac
, 5100);
972 const LLVMValueRef vertexidx
=
973 LLVMBuildLoad(builder
, ctx
->gs_next_vertex
[stream
], "");
974 tmp
= LLVMBuildICmp(builder
, LLVMIntUGE
, vertexidx
,
975 LLVMConstInt(ctx
->ac
.i32
, sel
->gs_max_out_vertices
, false), "");
976 ac_build_ifcc(&ctx
->ac
, tmp
, 5101);
977 ac_build_break(&ctx
->ac
);
978 ac_build_endif(&ctx
->ac
, 5101);
980 tmp
= LLVMBuildAdd(builder
, vertexidx
, ctx
->ac
.i32_1
, "");
981 LLVMBuildStore(builder
, tmp
, ctx
->gs_next_vertex
[stream
]);
983 tmp
= ngg_gs_emit_vertex_ptr(ctx
, gsthread
, vertexidx
);
984 LLVMValueRef gep_idx
[3] = {
985 ctx
->ac
.i32_0
, /* implied C-style array */
986 ctx
->ac
.i32_1
, /* second entry of struct */
987 LLVMConstInt(ctx
->ac
.i32
, stream
, false),
989 tmp
= LLVMBuildGEP(builder
, tmp
, gep_idx
, 3, "");
990 LLVMBuildStore(builder
, i8_0
, tmp
);
992 ac_build_endloop(&ctx
->ac
, 5100);
995 /* Accumulate generated primitives counts across the entire threadgroup. */
996 for (unsigned stream
= 0; stream
< 4; ++stream
) {
997 if (!info
->num_stream_output_components
[stream
])
1000 LLVMValueRef numprims
=
1001 LLVMBuildLoad(builder
, ctx
->gs_generated_prims
[stream
], "");
1002 numprims
= ac_build_reduce(&ctx
->ac
, numprims
, nir_op_iadd
, 64);
1004 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, ac_get_thread_id(&ctx
->ac
), ctx
->i32_0
, "");
1005 ac_build_ifcc(&ctx
->ac
, tmp
, 5105);
1007 LLVMBuildAtomicRMW(builder
, LLVMAtomicRMWBinOpAdd
,
1008 ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
,
1009 LLVMConstInt(ctx
->i32
, stream
, false)),
1010 numprims
, LLVMAtomicOrderingMonotonic
, false);
1012 ac_build_endif(&ctx
->ac
, 5105);
1015 lp_build_endif(&ctx
->merged_wrap_if_state
);
1017 ac_build_s_barrier(&ctx
->ac
);
1019 const LLVMValueRef tid
= get_thread_id_in_tg(ctx
);
1020 LLVMValueRef num_emit_threads
= ngg_get_prim_cnt(ctx
);
1023 if (sel
->so
.num_outputs
) {
1024 struct ngg_streamout nggso
= {};
1026 nggso
.num_vertices
= LLVMConstInt(ctx
->i32
, verts_per_prim
, false);
1028 LLVMValueRef vertexptr
= ngg_gs_vertex_ptr(ctx
, tid
);
1029 for (unsigned stream
= 0; stream
< 4; ++stream
) {
1030 if (!info
->num_stream_output_components
[stream
])
1033 LLVMValueRef gep_idx
[3] = {
1034 ctx
->i32_0
, /* implicit C-style array */
1035 ctx
->i32_1
, /* second value of struct */
1036 LLVMConstInt(ctx
->i32
, stream
, false),
1038 tmp
= LLVMBuildGEP(builder
, vertexptr
, gep_idx
, 3, "");
1039 tmp
= LLVMBuildLoad(builder
, tmp
, "");
1040 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->i1
, "");
1041 tmp2
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, num_emit_threads
, "");
1042 nggso
.prim_enable
[stream
] = LLVMBuildAnd(builder
, tmp
, tmp2
, "");
1045 for (unsigned i
= 0; i
< verts_per_prim
; ++i
) {
1046 tmp
= LLVMBuildSub(builder
, tid
,
1047 LLVMConstInt(ctx
->i32
, verts_per_prim
- i
- 1, false), "");
1048 tmp
= ngg_gs_vertex_ptr(ctx
, tmp
);
1049 nggso
.vertices
[i
] = ac_build_gep0(&ctx
->ac
, tmp
, ctx
->i32_0
);
1052 build_streamout(ctx
, &nggso
);
1055 /* Write shader query data. */
1056 tmp
= si_unpack_param(ctx
, ctx
->param_vs_state_bits
, 6, 1);
1057 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->i1
, "");
1058 ac_build_ifcc(&ctx
->ac
, tmp
, 5109); /* if (STREAMOUT_QUERY_ENABLED) */
1059 unsigned num_query_comps
= sel
->so
.num_outputs
? 8 : 4;
1060 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
,
1061 LLVMConstInt(ctx
->i32
, num_query_comps
, false), "");
1062 ac_build_ifcc(&ctx
->ac
, tmp
, 5110);
1064 LLVMValueRef offset
;
1066 if (sel
->so
.num_outputs
)
1067 tmp
= LLVMBuildAnd(builder
, tmp
, LLVMConstInt(ctx
->i32
, 3, false), "");
1068 offset
= LLVMBuildNUWMul(builder
, tmp
, LLVMConstInt(ctx
->i32
, 32, false), "");
1069 if (sel
->so
.num_outputs
) {
1070 tmp
= LLVMBuildLShr(builder
, tid
, LLVMConstInt(ctx
->i32
, 2, false), "");
1071 tmp
= LLVMBuildNUWMul(builder
, tmp
, LLVMConstInt(ctx
->i32
, 8, false), "");
1072 offset
= LLVMBuildAdd(builder
, offset
, tmp
, "");
1075 tmp
= LLVMBuildLoad(builder
, ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, tid
), "");
1076 LLVMValueRef args
[] = {
1078 ngg_get_query_buf(ctx
),
1080 LLVMConstInt(ctx
->i32
, 16, false), /* soffset */
1081 ctx
->i32_0
, /* cachepolicy */
1083 ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.raw.buffer.atomic.add.i32",
1084 ctx
->i32
, args
, 5, 0);
1086 ac_build_endif(&ctx
->ac
, 5110);
1087 ac_build_endif(&ctx
->ac
, 5109);
1091 /* Determine vertex liveness. */
1092 LLVMValueRef vertliveptr
= lp_build_alloca(&ctx
->gallivm
, ctx
->ac
.i1
, "vertexlive");
1094 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, num_emit_threads
, "");
1095 ac_build_ifcc(&ctx
->ac
, tmp
, 5120);
1097 for (unsigned i
= 0; i
< verts_per_prim
; ++i
) {
1098 const LLVMValueRef primidx
=
1099 LLVMBuildAdd(builder
, tid
,
1100 LLVMConstInt(ctx
->ac
.i32
, i
, false), "");
1103 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, primidx
, num_emit_threads
, "");
1104 ac_build_ifcc(&ctx
->ac
, tmp
, 5121 + i
);
1107 /* Load primitive liveness */
1108 tmp
= ngg_gs_vertex_ptr(ctx
, primidx
);
1109 LLVMValueRef gep_idx
[3] = {
1110 ctx
->ac
.i32_0
, /* implicit C-style array */
1111 ctx
->ac
.i32_1
, /* second value of struct */
1112 ctx
->ac
.i32_0
, /* stream 0 */
1114 tmp
= LLVMBuildGEP(builder
, tmp
, gep_idx
, 3, "");
1115 tmp
= LLVMBuildLoad(builder
, tmp
, "");
1116 const LLVMValueRef primlive
=
1117 LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
1119 tmp
= LLVMBuildLoad(builder
, vertliveptr
, "");
1120 tmp
= LLVMBuildOr(builder
, tmp
, primlive
, ""),
1121 LLVMBuildStore(builder
, tmp
, vertliveptr
);
1124 ac_build_endif(&ctx
->ac
, 5121 + i
);
1127 ac_build_endif(&ctx
->ac
, 5120);
1129 /* Inclusive scan addition across the current wave. */
1130 LLVMValueRef vertlive
= LLVMBuildLoad(builder
, vertliveptr
, "");
1131 struct ac_wg_scan vertlive_scan
= {};
1132 vertlive_scan
.op
= nir_op_iadd
;
1133 vertlive_scan
.enable_reduce
= true;
1134 vertlive_scan
.enable_exclusive
= true;
1135 vertlive_scan
.src
= vertlive
;
1136 vertlive_scan
.scratch
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, ctx
->i32_0
);
1137 vertlive_scan
.waveidx
= get_wave_id_in_tg(ctx
);
1138 vertlive_scan
.numwaves
= get_tgsize(ctx
);
1139 vertlive_scan
.maxwaves
= 8;
1141 ac_build_wg_scan(&ctx
->ac
, &vertlive_scan
);
1143 /* Skip all exports (including index exports) when possible. At least on
1144 * early gfx10 revisions this is also to avoid hangs.
1146 LLVMValueRef have_exports
=
1147 LLVMBuildICmp(builder
, LLVMIntNE
, vertlive_scan
.result_reduce
, ctx
->ac
.i32_0
, "");
1149 LLVMBuildSelect(builder
, have_exports
, num_emit_threads
, ctx
->ac
.i32_0
, "");
1151 /* Allocate export space. Send this message as early as possible, to
1152 * hide the latency of the SQ <-> SPI roundtrip.
1154 * Note: We could consider compacting primitives for export as well.
1155 * PA processes 1 non-null prim / clock, but it fetches 4 DW of
1156 * prim data per clock and skips null primitives at no additional
1157 * cost. So compacting primitives can only be beneficial when
1158 * there are 4 or more contiguous null primitives in the export
1159 * (in the common case of single-dword prim exports).
1161 build_sendmsg_gs_alloc_req(ctx
, vertlive_scan
.result_reduce
, num_emit_threads
);
1163 /* Setup the reverse vertex compaction permutation. We re-use stream 1
1164 * of the primitive liveness flags, relying on the fact that each
1165 * threadgroup can have at most 256 threads. */
1166 ac_build_ifcc(&ctx
->ac
, vertlive
, 5130);
1168 tmp
= ngg_gs_vertex_ptr(ctx
, vertlive_scan
.result_exclusive
);
1169 LLVMValueRef gep_idx
[3] = {
1170 ctx
->ac
.i32_0
, /* implicit C-style array */
1171 ctx
->ac
.i32_1
, /* second value of struct */
1172 ctx
->ac
.i32_1
, /* stream 1 */
1174 tmp
= LLVMBuildGEP(builder
, tmp
, gep_idx
, 3, "");
1175 tmp2
= LLVMBuildTrunc(builder
, tid
, ctx
->ac
.i8
, "");
1176 LLVMBuildStore(builder
, tmp2
, tmp
);
1178 ac_build_endif(&ctx
->ac
, 5130);
1180 ac_build_s_barrier(&ctx
->ac
);
1182 /* Export primitive data */
1183 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, num_emit_threads
, "");
1184 ac_build_ifcc(&ctx
->ac
, tmp
, 5140);
1186 struct ngg_prim prim
= {};
1187 prim
.num_vertices
= verts_per_prim
;
1189 tmp
= ngg_gs_vertex_ptr(ctx
, tid
);
1190 LLVMValueRef gep_idx
[3] = {
1191 ctx
->ac
.i32_0
, /* implicit C-style array */
1192 ctx
->ac
.i32_1
, /* second value of struct */
1193 ctx
->ac
.i32_0
, /* primflag */
1195 tmp
= LLVMBuildGEP(builder
, tmp
, gep_idx
, 3, "");
1196 tmp
= LLVMBuildLoad(builder
, tmp
, "");
1197 prim
.isnull
= LLVMBuildICmp(builder
, LLVMIntEQ
, tmp
,
1198 LLVMConstInt(ctx
->ac
.i8
, 0, false), "");
1200 for (unsigned i
= 0; i
< verts_per_prim
; ++i
) {
1201 prim
.index
[i
] = LLVMBuildSub(builder
, vertlive_scan
.result_exclusive
,
1202 LLVMConstInt(ctx
->ac
.i32
, verts_per_prim
- i
- 1, false), "");
1203 prim
.edgeflag
[i
] = ctx
->ac
.i1false
;
1206 build_export_prim(ctx
, &prim
);
1208 ac_build_endif(&ctx
->ac
, 5140);
1210 /* Export position and parameter data */
1211 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, vertlive_scan
.result_reduce
, "");
1212 ac_build_ifcc(&ctx
->ac
, tmp
, 5145);
1214 struct si_shader_output_values outputs
[PIPE_MAX_SHADER_OUTPUTS
];
1216 tmp
= ngg_gs_vertex_ptr(ctx
, tid
);
1217 LLVMValueRef gep_idx
[3] = {
1218 ctx
->ac
.i32_0
, /* implicit C-style array */
1219 ctx
->ac
.i32_1
, /* second value of struct */
1220 ctx
->ac
.i32_1
, /* stream 1: source data index */
1222 tmp
= LLVMBuildGEP(builder
, tmp
, gep_idx
, 3, "");
1223 tmp
= LLVMBuildLoad(builder
, tmp
, "");
1224 tmp
= LLVMBuildZExt(builder
, tmp
, ctx
->ac
.i32
, "");
1225 const LLVMValueRef vertexptr
= ngg_gs_vertex_ptr(ctx
, tmp
);
1227 unsigned out_idx
= 0;
1228 gep_idx
[1] = ctx
->ac
.i32_0
;
1229 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
1230 outputs
[i
].semantic_name
= info
->output_semantic_name
[i
];
1231 outputs
[i
].semantic_index
= info
->output_semantic_index
[i
];
1233 for (unsigned j
= 0; j
< 4; j
++, out_idx
++) {
1234 gep_idx
[2] = LLVMConstInt(ctx
->ac
.i32
, out_idx
, false);
1235 tmp
= LLVMBuildGEP(builder
, vertexptr
, gep_idx
, 3, "");
1236 tmp
= LLVMBuildLoad(builder
, tmp
, "");
1237 outputs
[i
].values
[j
] = ac_to_float(&ctx
->ac
, tmp
);
1238 outputs
[i
].vertex_stream
[j
] =
1239 (info
->output_streams
[i
] >> (2 * j
)) & 3;
1243 si_llvm_export_vs(ctx
, outputs
, info
->num_outputs
);
1245 ac_build_endif(&ctx
->ac
, 5145);
1248 static void clamp_gsprims_to_esverts(unsigned *max_gsprims
, unsigned max_esverts
,
1249 unsigned min_verts_per_prim
, bool use_adjacency
)
1251 unsigned max_reuse
= max_esverts
- min_verts_per_prim
;
1254 *max_gsprims
= MIN2(*max_gsprims
, 1 + max_reuse
);
1258 * Determine subgroup information like maximum number of vertices and prims.
1260 * This happens before the shader is uploaded, since LDS relocations during
1261 * upload depend on the subgroup size.
1263 void gfx10_ngg_calculate_subgroup_info(struct si_shader
*shader
)
1265 const struct si_shader_selector
*gs_sel
= shader
->selector
;
1266 const struct si_shader_selector
*es_sel
=
1267 shader
->previous_stage_sel
? shader
->previous_stage_sel
: gs_sel
;
1268 const enum pipe_shader_type gs_type
= gs_sel
->type
;
1269 const unsigned gs_num_invocations
= MAX2(gs_sel
->gs_num_invocations
, 1);
1270 /* TODO: Use QUADS as the worst case because of reuse, but triangles
1271 * will always have 1 additional unoccupied vector lane. We could use
1272 * that lane if the worst case was TRIANGLES. */
1273 const unsigned input_prim
= si_get_input_prim(gs_sel
, PIPE_PRIM_QUADS
);
1274 const bool use_adjacency
= input_prim
>= PIPE_PRIM_LINES_ADJACENCY
&&
1275 input_prim
<= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
;
1276 const unsigned max_verts_per_prim
= u_vertices_per_prim(input_prim
);
1277 const unsigned min_verts_per_prim
=
1278 gs_type
== PIPE_SHADER_GEOMETRY
? max_verts_per_prim
: 1;
1280 /* All these are in dwords: */
1281 /* We can't allow using the whole LDS, because GS waves compete with
1282 * other shader stages for LDS space.
1284 * TODO: We should really take the shader's internal LDS use into
1285 * account. The linker will fail if the size is greater than
1288 const unsigned max_lds_size
= 8 * 1024 - 768;
1289 const unsigned target_lds_size
= max_lds_size
;
1290 unsigned esvert_lds_size
= 0;
1291 unsigned gsprim_lds_size
= 0;
1293 /* All these are per subgroup: */
1294 bool max_vert_out_per_gs_instance
= false;
1295 unsigned max_esverts_base
= 128;
1296 unsigned max_gsprims_base
= 128; /* default prim group size clamp */
1298 /* Hardware has the following non-natural restrictions on the value
1299 * of GE_CNTL.VERT_GRP_SIZE based on based on the primitive type of
1301 * - at most 252 for any line input primitive type
1302 * - at most 251 for any quad input primitive type
1303 * - at most 251 for triangle strips with adjacency (this happens to
1304 * be the natural limit for triangle *lists* with adjacency)
1306 max_esverts_base
= MIN2(max_esverts_base
, 251 + max_verts_per_prim
- 1);
1308 if (gs_type
== PIPE_SHADER_GEOMETRY
) {
1309 unsigned max_out_verts_per_gsprim
=
1310 gs_sel
->gs_max_out_vertices
* gs_num_invocations
;
1312 if (max_out_verts_per_gsprim
<= 256) {
1313 if (max_out_verts_per_gsprim
) {
1314 max_gsprims_base
= MIN2(max_gsprims_base
,
1315 256 / max_out_verts_per_gsprim
);
1318 /* Use special multi-cycling mode in which each GS
1319 * instance gets its own subgroup. Does not work with
1321 max_vert_out_per_gs_instance
= true;
1322 max_gsprims_base
= 1;
1323 max_out_verts_per_gsprim
= gs_sel
->gs_max_out_vertices
;
1326 esvert_lds_size
= es_sel
->esgs_itemsize
/ 4;
1327 gsprim_lds_size
= (gs_sel
->gsvs_vertex_size
/ 4 + 1) * max_out_verts_per_gsprim
;
1330 /* LDS size for passing data from ES to GS. */
1331 esvert_lds_size
= ngg_nogs_vertex_size(shader
);
1333 /* LDS size for passing data from GS to ES.
1334 * GS stores Primitive IDs into LDS at the address corresponding
1335 * to the ES thread of the provoking vertex. All ES threads
1336 * load and export PrimitiveID for their thread.
1338 if (gs_sel
->type
== PIPE_SHADER_VERTEX
&&
1339 shader
->key
.mono
.u
.vs_export_prim_id
)
1340 esvert_lds_size
= MAX2(esvert_lds_size
, 1);
1343 unsigned max_gsprims
= max_gsprims_base
;
1344 unsigned max_esverts
= max_esverts_base
;
1346 if (esvert_lds_size
)
1347 max_esverts
= MIN2(max_esverts
, target_lds_size
/ esvert_lds_size
);
1348 if (gsprim_lds_size
)
1349 max_gsprims
= MIN2(max_gsprims
, target_lds_size
/ gsprim_lds_size
);
1351 max_esverts
= MIN2(max_esverts
, max_gsprims
* max_verts_per_prim
);
1352 clamp_gsprims_to_esverts(&max_gsprims
, max_esverts
, min_verts_per_prim
, use_adjacency
);
1353 assert(max_esverts
>= max_verts_per_prim
&& max_gsprims
>= 1);
1355 if (esvert_lds_size
|| gsprim_lds_size
) {
1356 /* Now that we have a rough proportionality between esverts
1357 * and gsprims based on the primitive type, scale both of them
1358 * down simultaneously based on required LDS space.
1360 * We could be smarter about this if we knew how much vertex
1363 unsigned lds_total
= max_esverts
* esvert_lds_size
+
1364 max_gsprims
* gsprim_lds_size
;
1365 if (lds_total
> target_lds_size
) {
1366 max_esverts
= max_esverts
* target_lds_size
/ lds_total
;
1367 max_gsprims
= max_gsprims
* target_lds_size
/ lds_total
;
1369 max_esverts
= MIN2(max_esverts
, max_gsprims
* max_verts_per_prim
);
1370 clamp_gsprims_to_esverts(&max_gsprims
, max_esverts
,
1371 min_verts_per_prim
, use_adjacency
);
1372 assert(max_esverts
>= max_verts_per_prim
&& max_gsprims
>= 1);
1376 /* Round up towards full wave sizes for better ALU utilization. */
1377 if (!max_vert_out_per_gs_instance
) {
1378 const unsigned wavesize
= 64;
1379 unsigned orig_max_esverts
;
1380 unsigned orig_max_gsprims
;
1382 orig_max_esverts
= max_esverts
;
1383 orig_max_gsprims
= max_gsprims
;
1385 max_esverts
= align(max_esverts
, wavesize
);
1386 max_esverts
= MIN2(max_esverts
, max_esverts_base
);
1387 if (esvert_lds_size
)
1388 max_esverts
= MIN2(max_esverts
,
1389 (max_lds_size
- max_gsprims
* gsprim_lds_size
) /
1391 max_esverts
= MIN2(max_esverts
, max_gsprims
* max_verts_per_prim
);
1393 max_gsprims
= align(max_gsprims
, wavesize
);
1394 max_gsprims
= MIN2(max_gsprims
, max_gsprims_base
);
1395 if (gsprim_lds_size
)
1396 max_gsprims
= MIN2(max_gsprims
,
1397 (max_lds_size
- max_esverts
* esvert_lds_size
) /
1399 clamp_gsprims_to_esverts(&max_gsprims
, max_esverts
,
1400 min_verts_per_prim
, use_adjacency
);
1401 assert(max_esverts
>= max_verts_per_prim
&& max_gsprims
>= 1);
1402 } while (orig_max_esverts
!= max_esverts
|| orig_max_gsprims
!= max_gsprims
);
1405 /* Hardware restriction: minimum value of max_esverts */
1406 max_esverts
= MAX2(max_esverts
, 23 + max_verts_per_prim
);
1408 unsigned max_out_vertices
=
1409 max_vert_out_per_gs_instance
? gs_sel
->gs_max_out_vertices
:
1410 gs_type
== PIPE_SHADER_GEOMETRY
?
1411 max_gsprims
* gs_num_invocations
* gs_sel
->gs_max_out_vertices
:
1413 assert(max_out_vertices
<= 256);
1415 unsigned prim_amp_factor
= 1;
1416 if (gs_type
== PIPE_SHADER_GEOMETRY
) {
1417 /* Number of output primitives per GS input primitive after
1419 prim_amp_factor
= gs_sel
->gs_max_out_vertices
;
1422 /* The GE only checks against the maximum number of ES verts after
1423 * allocating a full GS primitive. So we need to ensure that whenever
1424 * this check passes, there is enough space for a full primitive without
1427 shader
->ngg
.hw_max_esverts
= max_esverts
- max_verts_per_prim
+ 1;
1428 shader
->ngg
.max_gsprims
= max_gsprims
;
1429 shader
->ngg
.max_out_verts
= max_out_vertices
;
1430 shader
->ngg
.prim_amp_factor
= prim_amp_factor
;
1431 shader
->ngg
.max_vert_out_per_gs_instance
= max_vert_out_per_gs_instance
;
1433 shader
->gs_info
.esgs_ring_size
= 4 * max_esverts
* esvert_lds_size
;
1434 shader
->ngg
.ngg_emit_size
= max_gsprims
* gsprim_lds_size
;
1436 assert(shader
->ngg
.hw_max_esverts
>= 24); /* HW limitation */