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.
24 #include "ac_llvm_cull.h"
26 #include "si_shader_internal.h"
28 #include "util/u_memory.h"
29 #include "util/u_prim.h"
31 static LLVMValueRef
get_wave_id_in_tg(struct si_shader_context
*ctx
)
33 return si_unpack_param(ctx
, ctx
->merged_wave_info
, 24, 4);
36 static LLVMValueRef
get_tgsize(struct si_shader_context
*ctx
)
38 return si_unpack_param(ctx
, ctx
->merged_wave_info
, 28, 4);
41 static LLVMValueRef
get_thread_id_in_tg(struct si_shader_context
*ctx
)
43 LLVMBuilderRef builder
= ctx
->ac
.builder
;
45 tmp
= LLVMBuildMul(builder
, get_wave_id_in_tg(ctx
),
46 LLVMConstInt(ctx
->ac
.i32
, ctx
->ac
.wave_size
, false), "");
47 return LLVMBuildAdd(builder
, tmp
, ac_get_thread_id(&ctx
->ac
), "");
50 static LLVMValueRef
ngg_get_vtx_cnt(struct si_shader_context
*ctx
)
52 return si_unpack_param(ctx
, ctx
->gs_tg_info
, 12, 9);
55 static LLVMValueRef
ngg_get_prim_cnt(struct si_shader_context
*ctx
)
57 return si_unpack_param(ctx
, ctx
->gs_tg_info
, 22, 9);
60 static LLVMValueRef
ngg_get_ordered_id(struct si_shader_context
*ctx
)
62 return si_unpack_param(ctx
, ctx
->gs_tg_info
, 0, 12);
65 static LLVMValueRef
ngg_get_query_buf(struct si_shader_context
*ctx
)
67 LLVMValueRef buf_ptr
= ac_get_arg(&ctx
->ac
, ctx
->rw_buffers
);
69 return ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
,
70 LLVMConstInt(ctx
->ac
.i32
, GFX10_GS_QUERY_BUF
, false));
73 static LLVMValueRef
ngg_get_initial_edgeflag(struct si_shader_context
*ctx
, unsigned index
)
75 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
77 tmp
= LLVMBuildLShr(ctx
->ac
.builder
, ac_get_arg(&ctx
->ac
, ctx
->args
.gs_invocation_id
),
78 LLVMConstInt(ctx
->ac
.i32
, 8 + index
, false), "");
79 return LLVMBuildTrunc(ctx
->ac
.builder
, tmp
, ctx
->ac
.i1
, "");
81 return ctx
->ac
.i1false
;
85 * Return the number of vertices as a constant in \p num_vertices,
86 * and return a more precise value as LLVMValueRef from the function.
88 static LLVMValueRef
ngg_get_vertices_per_prim(struct si_shader_context
*ctx
, unsigned *num_vertices
)
90 const struct si_shader_info
*info
= &ctx
->shader
->selector
->info
;
92 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
93 if (info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
]) {
94 /* Blits always use axis-aligned rectangles with 3 vertices. */
96 return LLVMConstInt(ctx
->ac
.i32
, 3, 0);
98 /* We always build up all three indices for the prim export
99 * independent of the primitive type. The additional garbage
100 * data shouldn't hurt. This number doesn't matter with
105 /* Extract OUTPRIM field. */
106 LLVMValueRef num
= si_unpack_param(ctx
, ctx
->vs_state_bits
, 2, 2);
107 return LLVMBuildAdd(ctx
->ac
.builder
, num
, ctx
->ac
.i32_1
, "");
110 assert(ctx
->type
== PIPE_SHADER_TESS_EVAL
);
112 if (info
->properties
[TGSI_PROPERTY_TES_POINT_MODE
])
114 else if (info
->properties
[TGSI_PROPERTY_TES_PRIM_MODE
] == PIPE_PRIM_LINES
)
119 return LLVMConstInt(ctx
->ac
.i32
, *num_vertices
, false);
123 bool gfx10_ngg_export_prim_early(struct si_shader
*shader
)
125 struct si_shader_selector
*sel
= shader
->selector
;
127 assert(shader
->key
.as_ngg
&& !shader
->key
.as_es
);
129 return sel
->type
!= PIPE_SHADER_GEOMETRY
&& !sel
->info
.writes_edgeflag
;
132 void gfx10_ngg_build_sendmsg_gs_alloc_req(struct si_shader_context
*ctx
)
134 ac_build_sendmsg_gs_alloc_req(&ctx
->ac
, get_wave_id_in_tg(ctx
), ngg_get_vtx_cnt(ctx
),
135 ngg_get_prim_cnt(ctx
));
138 void gfx10_ngg_build_export_prim(struct si_shader_context
*ctx
, LLVMValueRef user_edgeflags
[3],
139 LLVMValueRef prim_passthrough
)
141 LLVMBuilderRef builder
= ctx
->ac
.builder
;
143 if (gfx10_is_ngg_passthrough(ctx
->shader
) || ctx
->shader
->key
.opt
.ngg_culling
) {
144 ac_build_ifcc(&ctx
->ac
, si_is_gs_thread(ctx
), 6001);
146 struct ac_ngg_prim prim
= {};
148 if (prim_passthrough
)
149 prim
.passthrough
= prim_passthrough
;
151 prim
.passthrough
= ac_get_arg(&ctx
->ac
, ctx
->gs_vtx01_offset
);
153 /* This is only used with NGG culling, which returns the NGG
154 * passthrough prim export encoding.
156 if (ctx
->shader
->selector
->info
.writes_edgeflag
) {
157 unsigned all_bits_no_edgeflags
= ~SI_NGG_PRIM_EDGE_FLAG_BITS
;
158 LLVMValueRef edgeflags
= LLVMConstInt(ctx
->ac
.i32
, all_bits_no_edgeflags
, 0);
160 unsigned num_vertices
;
161 ngg_get_vertices_per_prim(ctx
, &num_vertices
);
163 for (unsigned i
= 0; i
< num_vertices
; i
++) {
164 unsigned shift
= 9 + i
* 10;
167 edge
= LLVMBuildLoad(builder
, user_edgeflags
[i
], "");
168 edge
= LLVMBuildZExt(builder
, edge
, ctx
->ac
.i32
, "");
169 edge
= LLVMBuildShl(builder
, edge
, LLVMConstInt(ctx
->ac
.i32
, shift
, 0), "");
170 edgeflags
= LLVMBuildOr(builder
, edgeflags
, edge
, "");
172 prim
.passthrough
= LLVMBuildAnd(builder
, prim
.passthrough
, edgeflags
, "");
175 ac_build_export_prim(&ctx
->ac
, &prim
);
177 ac_build_endif(&ctx
->ac
, 6001);
181 ac_build_ifcc(&ctx
->ac
, si_is_gs_thread(ctx
), 6001);
183 struct ac_ngg_prim prim
= {};
185 ngg_get_vertices_per_prim(ctx
, &prim
.num_vertices
);
187 prim
.isnull
= ctx
->ac
.i1false
;
188 prim
.index
[0] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 0, 16);
189 prim
.index
[1] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 16, 16);
190 prim
.index
[2] = si_unpack_param(ctx
, ctx
->gs_vtx23_offset
, 0, 16);
192 for (unsigned i
= 0; i
< prim
.num_vertices
; ++i
) {
193 prim
.edgeflag
[i
] = ngg_get_initial_edgeflag(ctx
, i
);
195 if (ctx
->shader
->selector
->info
.writes_edgeflag
) {
198 edge
= LLVMBuildLoad(ctx
->ac
.builder
, user_edgeflags
[i
], "");
199 edge
= LLVMBuildAnd(ctx
->ac
.builder
, prim
.edgeflag
[i
], edge
, "");
200 prim
.edgeflag
[i
] = edge
;
204 ac_build_export_prim(&ctx
->ac
, &prim
);
206 ac_build_endif(&ctx
->ac
, 6001);
209 static void build_streamout_vertex(struct si_shader_context
*ctx
, LLVMValueRef
*so_buffer
,
210 LLVMValueRef
*wg_offset_dw
, unsigned stream
,
211 LLVMValueRef offset_vtx
, LLVMValueRef vertexptr
)
213 struct si_shader_info
*info
= &ctx
->shader
->selector
->info
;
214 struct pipe_stream_output_info
*so
= &ctx
->shader
->selector
->so
;
215 LLVMBuilderRef builder
= ctx
->ac
.builder
;
216 LLVMValueRef offset
[4] = {};
219 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
220 if (!wg_offset_dw
[buffer
])
223 tmp
= LLVMBuildMul(builder
, offset_vtx
, LLVMConstInt(ctx
->ac
.i32
, so
->stride
[buffer
], false),
225 tmp
= LLVMBuildAdd(builder
, wg_offset_dw
[buffer
], tmp
, "");
226 offset
[buffer
] = LLVMBuildShl(builder
, tmp
, LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
229 for (unsigned i
= 0; i
< so
->num_outputs
; ++i
) {
230 if (so
->output
[i
].stream
!= stream
)
233 unsigned reg
= so
->output
[i
].register_index
;
234 struct si_shader_output_values out
;
235 out
.semantic_name
= info
->output_semantic_name
[reg
];
236 out
.semantic_index
= info
->output_semantic_index
[reg
];
238 for (unsigned comp
= 0; comp
< 4; comp
++) {
239 tmp
= ac_build_gep0(&ctx
->ac
, vertexptr
, LLVMConstInt(ctx
->ac
.i32
, 4 * reg
+ comp
, false));
240 out
.values
[comp
] = LLVMBuildLoad(builder
, tmp
, "");
241 out
.vertex_stream
[comp
] = (info
->output_streams
[reg
] >> (2 * comp
)) & 3;
244 si_llvm_streamout_store_output(ctx
, so_buffer
, offset
, &so
->output
[i
], &out
);
248 struct ngg_streamout
{
249 LLVMValueRef num_vertices
;
251 /* per-thread data */
252 LLVMValueRef prim_enable
[4]; /* i1 per stream */
253 LLVMValueRef vertices
[3]; /* [N x i32] addrspace(LDS)* */
256 LLVMValueRef emit
[4]; /* per-stream emitted primitives (only valid for used streams) */
260 * Build streamout logic.
264 * Writes number of emitted primitives to gs_ngg_scratch[4:8].
266 * Clobbers gs_ngg_scratch[8:].
268 static void build_streamout(struct si_shader_context
*ctx
, struct ngg_streamout
*nggso
)
270 struct si_shader_info
*info
= &ctx
->shader
->selector
->info
;
271 struct pipe_stream_output_info
*so
= &ctx
->shader
->selector
->so
;
272 LLVMBuilderRef builder
= ctx
->ac
.builder
;
273 LLVMValueRef buf_ptr
= ac_get_arg(&ctx
->ac
, ctx
->rw_buffers
);
274 LLVMValueRef tid
= get_thread_id_in_tg(ctx
);
275 LLVMValueRef tmp
, tmp2
;
276 LLVMValueRef i32_2
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
277 LLVMValueRef i32_4
= LLVMConstInt(ctx
->ac
.i32
, 4, false);
278 LLVMValueRef i32_8
= LLVMConstInt(ctx
->ac
.i32
, 8, false);
279 LLVMValueRef so_buffer
[4] = {};
280 unsigned max_num_vertices
= 1 + (nggso
->vertices
[1] ? 1 : 0) + (nggso
->vertices
[2] ? 1 : 0);
281 LLVMValueRef prim_stride_dw
[4] = {};
282 LLVMValueRef prim_stride_dw_vgpr
= LLVMGetUndef(ctx
->ac
.i32
);
283 int stream_for_buffer
[4] = {-1, -1, -1, -1};
284 unsigned bufmask_for_stream
[4] = {};
285 bool isgs
= ctx
->type
== PIPE_SHADER_GEOMETRY
;
286 unsigned scratch_emit_base
= isgs
? 4 : 0;
287 LLVMValueRef scratch_emit_basev
= isgs
? i32_4
: ctx
->ac
.i32_0
;
288 unsigned scratch_offset_base
= isgs
? 8 : 4;
289 LLVMValueRef scratch_offset_basev
= isgs
? i32_8
: i32_4
;
291 ac_llvm_add_target_dep_function_attr(ctx
->main_fn
, "amdgpu-gds-size", 256);
293 /* Determine the mapping of streamout buffers to vertex streams. */
294 for (unsigned i
= 0; i
< so
->num_outputs
; ++i
) {
295 unsigned buf
= so
->output
[i
].output_buffer
;
296 unsigned stream
= so
->output
[i
].stream
;
297 assert(stream_for_buffer
[buf
] < 0 || stream_for_buffer
[buf
] == stream
);
298 stream_for_buffer
[buf
] = stream
;
299 bufmask_for_stream
[stream
] |= 1 << buf
;
302 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
303 if (stream_for_buffer
[buffer
] == -1)
306 assert(so
->stride
[buffer
]);
308 tmp
= LLVMConstInt(ctx
->ac
.i32
, so
->stride
[buffer
], false);
309 prim_stride_dw
[buffer
] = LLVMBuildMul(builder
, tmp
, nggso
->num_vertices
, "");
310 prim_stride_dw_vgpr
=
311 ac_build_writelane(&ctx
->ac
, prim_stride_dw_vgpr
, prim_stride_dw
[buffer
],
312 LLVMConstInt(ctx
->ac
.i32
, buffer
, false));
314 so_buffer
[buffer
] = ac_build_load_to_sgpr(
315 &ctx
->ac
, buf_ptr
, LLVMConstInt(ctx
->ac
.i32
, SI_VS_STREAMOUT_BUF0
+ buffer
, false));
318 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, get_wave_id_in_tg(ctx
), ctx
->ac
.i32_0
, "");
319 ac_build_ifcc(&ctx
->ac
, tmp
, 5200);
321 LLVMTypeRef gdsptr
= LLVMPointerType(ctx
->ac
.i32
, AC_ADDR_SPACE_GDS
);
322 LLVMValueRef gdsbase
= LLVMBuildIntToPtr(builder
, ctx
->ac
.i32_0
, gdsptr
, "");
324 /* Advance the streamout offsets in GDS. */
325 LLVMValueRef offsets_vgpr
= ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i32
, "");
326 LLVMValueRef generated_by_stream_vgpr
= ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i32
, "");
328 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, ac_get_thread_id(&ctx
->ac
), i32_4
, "");
329 ac_build_ifcc(&ctx
->ac
, tmp
, 5210);
332 tmp
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, tid
);
333 tmp
= LLVMBuildLoad(builder
, tmp
, "");
335 tmp
= ac_build_writelane(&ctx
->ac
, ctx
->ac
.i32_0
, ngg_get_prim_cnt(ctx
), ctx
->ac
.i32_0
);
337 LLVMBuildStore(builder
, tmp
, generated_by_stream_vgpr
);
340 int unused_stream
= -1;
341 for (unsigned stream
= 0; stream
< 4; ++stream
) {
342 if (!info
->num_stream_output_components
[stream
]) {
343 unused_stream
= stream
;
347 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
348 if (stream_for_buffer
[buffer
] >= 0) {
349 swizzle
[buffer
] = stream_for_buffer
[buffer
];
351 assert(unused_stream
>= 0);
352 swizzle
[buffer
] = unused_stream
;
356 tmp
= ac_build_quad_swizzle(&ctx
->ac
, tmp
, swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
357 tmp
= LLVMBuildMul(builder
, tmp
, prim_stride_dw_vgpr
, "");
359 LLVMValueRef args
[] = {
360 LLVMBuildIntToPtr(builder
, ngg_get_ordered_id(ctx
), gdsptr
, ""),
362 ctx
->ac
.i32_0
, // ordering
363 ctx
->ac
.i32_0
, // scope
364 ctx
->ac
.i1false
, // isVolatile
365 LLVMConstInt(ctx
->ac
.i32
, 4 << 24, false), // OA index
366 ctx
->ac
.i1true
, // wave release
367 ctx
->ac
.i1true
, // wave done
369 tmp
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.ds.ordered.add", ctx
->ac
.i32
, args
,
370 ARRAY_SIZE(args
), 0);
372 /* Keep offsets in a VGPR for quick retrieval via readlane by
373 * the first wave for bounds checking, and also store in LDS
374 * for retrieval by all waves later. */
375 LLVMBuildStore(builder
, tmp
, offsets_vgpr
);
377 tmp2
= LLVMBuildAdd(builder
, ac_get_thread_id(&ctx
->ac
), scratch_offset_basev
, "");
378 tmp2
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, tmp2
);
379 LLVMBuildStore(builder
, tmp
, tmp2
);
381 ac_build_endif(&ctx
->ac
, 5210);
383 /* Determine the max emit per buffer. This is done via the SALU, in part
384 * because LLVM can't generate divide-by-multiply if we try to do this
385 * via VALU with one lane per buffer.
387 LLVMValueRef max_emit
[4] = {};
388 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
389 if (stream_for_buffer
[buffer
] == -1)
392 LLVMValueRef bufsize_dw
= LLVMBuildLShr(
393 builder
, LLVMBuildExtractElement(builder
, so_buffer
[buffer
], i32_2
, ""), i32_2
, "");
395 tmp
= LLVMBuildLoad(builder
, offsets_vgpr
, "");
396 LLVMValueRef offset_dw
=
397 ac_build_readlane(&ctx
->ac
, tmp
, LLVMConstInt(ctx
->ac
.i32
, buffer
, false));
399 tmp
= LLVMBuildSub(builder
, bufsize_dw
, offset_dw
, "");
400 tmp
= LLVMBuildUDiv(builder
, tmp
, prim_stride_dw
[buffer
], "");
402 tmp2
= LLVMBuildICmp(builder
, LLVMIntULT
, bufsize_dw
, offset_dw
, "");
403 max_emit
[buffer
] = LLVMBuildSelect(builder
, tmp2
, ctx
->ac
.i32_0
, tmp
, "");
406 /* Determine the number of emitted primitives per stream and fixup the
407 * GDS counter if necessary.
409 * This is complicated by the fact that a single stream can emit to
410 * multiple buffers (but luckily not vice versa).
412 LLVMValueRef emit_vgpr
= ctx
->ac
.i32_0
;
414 for (unsigned stream
= 0; stream
< 4; ++stream
) {
415 if (!info
->num_stream_output_components
[stream
])
418 tmp
= LLVMBuildLoad(builder
, generated_by_stream_vgpr
, "");
419 LLVMValueRef generated
=
420 ac_build_readlane(&ctx
->ac
, tmp
, LLVMConstInt(ctx
->ac
.i32
, stream
, false));
422 LLVMValueRef emit
= generated
;
423 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
424 if (stream_for_buffer
[buffer
] == stream
)
425 emit
= ac_build_umin(&ctx
->ac
, emit
, max_emit
[buffer
]);
429 ac_build_writelane(&ctx
->ac
, emit_vgpr
, emit
, LLVMConstInt(ctx
->ac
.i32
, stream
, false));
431 /* Fixup the offset using a plain GDS atomic if we overflowed. */
432 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, emit
, generated
, "");
433 ac_build_ifcc(&ctx
->ac
, tmp
, 5221); /* scalar branch */
434 tmp
= LLVMBuildLShr(builder
, LLVMConstInt(ctx
->ac
.i32
, bufmask_for_stream
[stream
], false),
435 ac_get_thread_id(&ctx
->ac
), "");
436 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
437 ac_build_ifcc(&ctx
->ac
, tmp
, 5222);
439 tmp
= LLVMBuildSub(builder
, generated
, emit
, "");
440 tmp
= LLVMBuildMul(builder
, tmp
, prim_stride_dw_vgpr
, "");
441 tmp2
= LLVMBuildGEP(builder
, gdsbase
, &tid
, 1, "");
442 LLVMBuildAtomicRMW(builder
, LLVMAtomicRMWBinOpSub
, tmp2
, tmp
,
443 LLVMAtomicOrderingMonotonic
, false);
445 ac_build_endif(&ctx
->ac
, 5222);
446 ac_build_endif(&ctx
->ac
, 5221);
449 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, ac_get_thread_id(&ctx
->ac
), i32_4
, "");
450 ac_build_ifcc(&ctx
->ac
, tmp
, 5225);
452 tmp
= LLVMBuildAdd(builder
, ac_get_thread_id(&ctx
->ac
), scratch_emit_basev
, "");
453 tmp
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, tmp
);
454 LLVMBuildStore(builder
, emit_vgpr
, tmp
);
456 ac_build_endif(&ctx
->ac
, 5225);
458 ac_build_endif(&ctx
->ac
, 5200);
460 /* Determine the workgroup-relative per-thread / primitive offset into
461 * the streamout buffers */
462 struct ac_wg_scan primemit_scan
[4] = {};
465 for (unsigned stream
= 0; stream
< 4; ++stream
) {
466 if (!info
->num_stream_output_components
[stream
])
469 primemit_scan
[stream
].enable_exclusive
= true;
470 primemit_scan
[stream
].op
= nir_op_iadd
;
471 primemit_scan
[stream
].src
= nggso
->prim_enable
[stream
];
472 primemit_scan
[stream
].scratch
= ac_build_gep0(
473 &ctx
->ac
, ctx
->gs_ngg_scratch
, LLVMConstInt(ctx
->ac
.i32
, 12 + 8 * stream
, false));
474 primemit_scan
[stream
].waveidx
= get_wave_id_in_tg(ctx
);
475 primemit_scan
[stream
].numwaves
= get_tgsize(ctx
);
476 primemit_scan
[stream
].maxwaves
= 8;
477 ac_build_wg_scan_top(&ctx
->ac
, &primemit_scan
[stream
]);
481 ac_build_s_barrier(&ctx
->ac
);
483 /* Fetch the per-buffer offsets and per-stream emit counts in all waves. */
484 LLVMValueRef wgoffset_dw
[4] = {};
487 LLVMValueRef scratch_vgpr
;
489 tmp
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, ac_get_thread_id(&ctx
->ac
));
490 scratch_vgpr
= LLVMBuildLoad(builder
, tmp
, "");
492 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
493 if (stream_for_buffer
[buffer
] >= 0) {
494 wgoffset_dw
[buffer
] =
495 ac_build_readlane(&ctx
->ac
, scratch_vgpr
,
496 LLVMConstInt(ctx
->ac
.i32
, scratch_offset_base
+ buffer
, false));
500 for (unsigned stream
= 0; stream
< 4; ++stream
) {
501 if (info
->num_stream_output_components
[stream
]) {
502 nggso
->emit
[stream
] =
503 ac_build_readlane(&ctx
->ac
, scratch_vgpr
,
504 LLVMConstInt(ctx
->ac
.i32
, scratch_emit_base
+ stream
, false));
509 /* Write out primitive data */
510 for (unsigned stream
= 0; stream
< 4; ++stream
) {
511 if (!info
->num_stream_output_components
[stream
])
515 ac_build_wg_scan_bottom(&ctx
->ac
, &primemit_scan
[stream
]);
517 primemit_scan
[stream
].result_exclusive
= tid
;
520 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, primemit_scan
[stream
].result_exclusive
,
521 nggso
->emit
[stream
], "");
522 tmp
= LLVMBuildAnd(builder
, tmp
, nggso
->prim_enable
[stream
], "");
523 ac_build_ifcc(&ctx
->ac
, tmp
, 5240);
525 LLVMValueRef offset_vtx
=
526 LLVMBuildMul(builder
, primemit_scan
[stream
].result_exclusive
, nggso
->num_vertices
, "");
528 for (unsigned i
= 0; i
< max_num_vertices
; ++i
) {
529 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, LLVMConstInt(ctx
->ac
.i32
, i
, false),
530 nggso
->num_vertices
, "");
531 ac_build_ifcc(&ctx
->ac
, tmp
, 5241);
532 build_streamout_vertex(ctx
, so_buffer
, wgoffset_dw
, stream
, offset_vtx
,
534 ac_build_endif(&ctx
->ac
, 5241);
535 offset_vtx
= LLVMBuildAdd(builder
, offset_vtx
, ctx
->ac
.i32_1
, "");
538 ac_build_endif(&ctx
->ac
, 5240);
542 /* LDS layout of ES vertex data for NGG culling. */
545 /* Byte 0: Boolean ES thread accepted (unculled) flag, and later the old
546 * ES thread ID. After vertex compaction, compacted ES threads
547 * store the old thread ID here to copy input VGPRs from uncompacted
549 * Byte 1: New ES thread ID, loaded by GS to prepare the prim export value.
550 * Byte 2: TES rel patch ID
553 lds_byte0_accept_flag
= 0,
554 lds_byte0_old_thread_id
= 0,
555 lds_byte1_new_thread_id
,
556 lds_byte2_tes_rel_patch_id
,
559 lds_packed_data
= 0, /* lds_byteN_... */
569 lds_instance_id
, /* optional */
571 lds_tes_u
= lds_vertex_id
,
572 lds_tes_v
= lds_instance_id
,
573 lds_tes_patch_id
, /* optional */
576 static LLVMValueRef
si_build_gep_i8(struct si_shader_context
*ctx
, LLVMValueRef ptr
,
579 assert(byte_index
< 4);
580 LLVMTypeRef pi8
= LLVMPointerType(ctx
->ac
.i8
, AC_ADDR_SPACE_LDS
);
581 LLVMValueRef index
= LLVMConstInt(ctx
->ac
.i32
, byte_index
, 0);
583 return LLVMBuildGEP(ctx
->ac
.builder
, LLVMBuildPointerCast(ctx
->ac
.builder
, ptr
, pi8
, ""), &index
,
587 static unsigned ngg_nogs_vertex_size(struct si_shader
*shader
)
589 unsigned lds_vertex_size
= 0;
591 /* The edgeflag is always stored in the last element that's also
592 * used for padding to reduce LDS bank conflicts. */
593 if (shader
->selector
->so
.num_outputs
)
594 lds_vertex_size
= 4 * shader
->selector
->info
.num_outputs
+ 1;
595 if (shader
->selector
->info
.writes_edgeflag
)
596 lds_vertex_size
= MAX2(lds_vertex_size
, 1);
598 /* LDS size for passing data from GS to ES.
599 * GS stores Primitive IDs into LDS at the address corresponding
600 * to the ES thread of the provoking vertex. All ES threads
601 * load and export PrimitiveID for their thread.
603 if (shader
->selector
->type
== PIPE_SHADER_VERTEX
&& shader
->key
.mono
.u
.vs_export_prim_id
)
604 lds_vertex_size
= MAX2(lds_vertex_size
, 1);
606 if (shader
->key
.opt
.ngg_culling
) {
607 if (shader
->selector
->type
== PIPE_SHADER_VERTEX
) {
608 STATIC_ASSERT(lds_instance_id
+ 1 == 9);
609 lds_vertex_size
= MAX2(lds_vertex_size
, 9);
611 assert(shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
);
613 if (shader
->selector
->info
.uses_primid
|| shader
->key
.mono
.u
.vs_export_prim_id
) {
614 STATIC_ASSERT(lds_tes_patch_id
+ 2 == 11);
615 lds_vertex_size
= MAX2(lds_vertex_size
, 11);
617 STATIC_ASSERT(lds_tes_v
+ 1 == 9);
618 lds_vertex_size
= MAX2(lds_vertex_size
, 9);
623 return lds_vertex_size
;
627 * Returns an `[N x i32] addrspace(LDS)*` pointing at contiguous LDS storage
628 * for the vertex outputs.
630 static LLVMValueRef
ngg_nogs_vertex_ptr(struct si_shader_context
*ctx
, LLVMValueRef vtxid
)
632 /* The extra dword is used to avoid LDS bank conflicts. */
633 unsigned vertex_size
= ngg_nogs_vertex_size(ctx
->shader
);
634 LLVMTypeRef ai32
= LLVMArrayType(ctx
->ac
.i32
, vertex_size
);
635 LLVMTypeRef pai32
= LLVMPointerType(ai32
, AC_ADDR_SPACE_LDS
);
636 LLVMValueRef tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->esgs_ring
, pai32
, "");
637 return LLVMBuildGEP(ctx
->ac
.builder
, tmp
, &vtxid
, 1, "");
640 static LLVMValueRef
si_insert_input_v4i32(struct si_shader_context
*ctx
, LLVMValueRef ret
,
641 struct ac_arg param
, unsigned return_index
)
643 LLVMValueRef v
= ac_get_arg(&ctx
->ac
, param
);
645 for (unsigned i
= 0; i
< 4; i
++) {
646 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, ac_llvm_extract_elem(&ctx
->ac
, v
, i
),
647 return_index
+ i
, "");
652 static void load_bitmasks_2x64(struct si_shader_context
*ctx
, LLVMValueRef lds_ptr
,
653 unsigned dw_offset
, LLVMValueRef mask
[2],
654 LLVMValueRef
*total_bitcount
)
656 LLVMBuilderRef builder
= ctx
->ac
.builder
;
657 LLVMValueRef ptr64
= LLVMBuildPointerCast(
658 builder
, lds_ptr
, LLVMPointerType(LLVMArrayType(ctx
->ac
.i64
, 2), AC_ADDR_SPACE_LDS
), "");
659 for (unsigned i
= 0; i
< 2; i
++) {
660 LLVMValueRef index
= LLVMConstInt(ctx
->ac
.i32
, dw_offset
/ 2 + i
, 0);
661 mask
[i
] = LLVMBuildLoad(builder
, ac_build_gep0(&ctx
->ac
, ptr64
, index
), "");
664 /* We get better code if we don't use the 128-bit bitcount. */
665 *total_bitcount
= LLVMBuildAdd(builder
, ac_build_bit_count(&ctx
->ac
, mask
[0]),
666 ac_build_bit_count(&ctx
->ac
, mask
[1]), "");
670 * Given a total thread count, update total and per-wave thread counts in input SGPRs
671 * and return the per-wave thread count.
673 * \param new_num_threads Total thread count on the input, per-wave thread count on the output.
674 * \param tg_info tg_info SGPR value
675 * \param tg_info_num_bits the bit size of thread count field in tg_info
676 * \param tg_info_shift the bit offset of the thread count field in tg_info
677 * \param wave_info merged_wave_info SGPR value
678 * \param wave_info_num_bits the bit size of thread count field in merged_wave_info
679 * \param wave_info_shift the bit offset of the thread count field in merged_wave_info
681 static void update_thread_counts(struct si_shader_context
*ctx
, LLVMValueRef
*new_num_threads
,
682 LLVMValueRef
*tg_info
, unsigned tg_info_num_bits
,
683 unsigned tg_info_shift
, LLVMValueRef
*wave_info
,
684 unsigned wave_info_num_bits
, unsigned wave_info_shift
)
686 LLVMBuilderRef builder
= ctx
->ac
.builder
;
688 /* Update the total thread count. */
689 unsigned tg_info_mask
= ~(u_bit_consecutive(0, tg_info_num_bits
) << tg_info_shift
);
690 *tg_info
= LLVMBuildAnd(builder
, *tg_info
, LLVMConstInt(ctx
->ac
.i32
, tg_info_mask
, 0), "");
691 *tg_info
= LLVMBuildOr(
693 LLVMBuildShl(builder
, *new_num_threads
, LLVMConstInt(ctx
->ac
.i32
, tg_info_shift
, 0), ""), "");
695 /* Update the per-wave thread count. */
696 LLVMValueRef prev_threads
= LLVMBuildMul(builder
, get_wave_id_in_tg(ctx
),
697 LLVMConstInt(ctx
->ac
.i32
, ctx
->ac
.wave_size
, 0), "");
698 *new_num_threads
= LLVMBuildSub(builder
, *new_num_threads
, prev_threads
, "");
699 *new_num_threads
= ac_build_imax(&ctx
->ac
, *new_num_threads
, ctx
->ac
.i32_0
);
701 ac_build_imin(&ctx
->ac
, *new_num_threads
, LLVMConstInt(ctx
->ac
.i32
, ctx
->ac
.wave_size
, 0));
702 unsigned wave_info_mask
= ~(u_bit_consecutive(0, wave_info_num_bits
) << wave_info_shift
);
703 *wave_info
= LLVMBuildAnd(builder
, *wave_info
, LLVMConstInt(ctx
->ac
.i32
, wave_info_mask
, 0), "");
704 *wave_info
= LLVMBuildOr(
706 LLVMBuildShl(builder
, *new_num_threads
, LLVMConstInt(ctx
->ac
.i32
, wave_info_shift
, 0), ""),
711 * Cull primitives for NGG VS or TES, then compact vertices, which happens
712 * before the VS or TES main function. Return values for the main function.
713 * Also return the position, which is passed to the shader as an input,
714 * so that we don't compute it twice.
716 void gfx10_emit_ngg_culling_epilogue_4x_wave32(struct ac_shader_abi
*abi
, unsigned max_outputs
,
719 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
720 struct si_shader
*shader
= ctx
->shader
;
721 struct si_shader_selector
*sel
= shader
->selector
;
722 struct si_shader_info
*info
= &sel
->info
;
723 LLVMBuilderRef builder
= ctx
->ac
.builder
;
725 assert(shader
->key
.opt
.ngg_culling
);
726 assert(shader
->key
.as_ngg
);
727 assert(sel
->type
== PIPE_SHADER_VERTEX
||
728 (sel
->type
== PIPE_SHADER_TESS_EVAL
&& !shader
->key
.as_es
));
730 LLVMValueRef position
[4] = {};
731 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
732 switch (info
->output_semantic_name
[i
]) {
733 case TGSI_SEMANTIC_POSITION
:
734 for (unsigned j
= 0; j
< 4; j
++) {
735 position
[j
] = LLVMBuildLoad(ctx
->ac
.builder
, addrs
[4 * i
+ j
], "");
742 /* Store Position.XYZW into LDS. */
743 LLVMValueRef es_vtxptr
= ngg_nogs_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
));
744 for (unsigned chan
= 0; chan
< 4; chan
++) {
746 builder
, ac_to_integer(&ctx
->ac
, position
[chan
]),
747 ac_build_gep0(&ctx
->ac
, es_vtxptr
, LLVMConstInt(ctx
->ac
.i32
, lds_pos_x
+ chan
, 0)));
749 /* Store Position.XY / W into LDS. */
750 for (unsigned chan
= 0; chan
< 2; chan
++) {
751 LLVMValueRef val
= ac_build_fdiv(&ctx
->ac
, position
[chan
], position
[3]);
753 builder
, ac_to_integer(&ctx
->ac
, val
),
754 ac_build_gep0(&ctx
->ac
, es_vtxptr
, LLVMConstInt(ctx
->ac
.i32
, lds_pos_x_div_w
+ chan
, 0)));
757 /* Store VertexID and InstanceID. ES threads will have to load them
758 * from LDS after vertex compaction and use them instead of their own
761 bool uses_instance_id
= false;
762 bool uses_tes_prim_id
= false;
763 LLVMValueRef packed_data
= ctx
->ac
.i32_0
;
765 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
766 uses_instance_id
= sel
->info
.uses_instanceid
||
767 shader
->key
.part
.vs
.prolog
.instance_divisor_is_one
||
768 shader
->key
.part
.vs
.prolog
.instance_divisor_is_fetched
;
771 builder
, ctx
->abi
.vertex_id
,
772 ac_build_gep0(&ctx
->ac
, es_vtxptr
, LLVMConstInt(ctx
->ac
.i32
, lds_vertex_id
, 0)));
773 if (uses_instance_id
) {
775 builder
, ctx
->abi
.instance_id
,
776 ac_build_gep0(&ctx
->ac
, es_vtxptr
, LLVMConstInt(ctx
->ac
.i32
, lds_instance_id
, 0)));
779 uses_tes_prim_id
= sel
->info
.uses_primid
|| shader
->key
.mono
.u
.vs_export_prim_id
;
781 assert(ctx
->type
== PIPE_SHADER_TESS_EVAL
);
782 LLVMBuildStore(builder
, ac_to_integer(&ctx
->ac
, ac_get_arg(&ctx
->ac
, ctx
->tes_u
)),
783 ac_build_gep0(&ctx
->ac
, es_vtxptr
, LLVMConstInt(ctx
->ac
.i32
, lds_tes_u
, 0)));
784 LLVMBuildStore(builder
, ac_to_integer(&ctx
->ac
, ac_get_arg(&ctx
->ac
, ctx
->tes_v
)),
785 ac_build_gep0(&ctx
->ac
, es_vtxptr
, LLVMConstInt(ctx
->ac
.i32
, lds_tes_v
, 0)));
786 packed_data
= LLVMBuildShl(builder
, ac_get_arg(&ctx
->ac
, ctx
->tes_rel_patch_id
),
787 LLVMConstInt(ctx
->ac
.i32
, lds_byte2_tes_rel_patch_id
* 8, 0), "");
788 if (uses_tes_prim_id
) {
790 builder
, ac_get_arg(&ctx
->ac
, ctx
->args
.tes_patch_id
),
791 ac_build_gep0(&ctx
->ac
, es_vtxptr
, LLVMConstInt(ctx
->ac
.i32
, lds_tes_patch_id
, 0)));
794 /* Initialize the packed data. */
796 builder
, packed_data
,
797 ac_build_gep0(&ctx
->ac
, es_vtxptr
, LLVMConstInt(ctx
->ac
.i32
, lds_packed_data
, 0)));
798 ac_build_endif(&ctx
->ac
, ctx
->merged_wrap_if_label
);
800 LLVMValueRef tid
= ac_get_thread_id(&ctx
->ac
);
802 /* Initialize the last 3 gs_ngg_scratch dwords to 0, because we may have less
803 * than 4 waves, but we always read all 4 values. This is where the thread
804 * bitmasks of unculled threads will be stored.
806 * gs_ngg_scratch layout: esmask[0..3]
808 ac_build_ifcc(&ctx
->ac
,
809 LLVMBuildICmp(builder
, LLVMIntULT
, get_thread_id_in_tg(ctx
),
810 LLVMConstInt(ctx
->ac
.i32
, 3, 0), ""),
813 LLVMValueRef index
= LLVMBuildAdd(builder
, tid
, ctx
->ac
.i32_1
, "");
814 LLVMBuildStore(builder
, ctx
->ac
.i32_0
, ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, index
));
816 ac_build_endif(&ctx
->ac
, 16101);
817 ac_build_s_barrier(&ctx
->ac
);
819 /* The hardware requires that there are no holes between unculled vertices,
820 * which means we have to pack ES threads, i.e. reduce the ES thread count
821 * and move ES input VGPRs to lower threads. The upside is that varyings
822 * are only fetched and computed for unculled vertices.
824 * Vertex compaction in GS threads:
826 * Part 1: Compute the surviving vertex mask in GS threads:
827 * - Compute 4 32-bit surviving vertex masks in LDS. (max 4 waves)
828 * - In GS, notify ES threads whether the vertex survived.
830 * - ES threads will create the mask and store it in LDS.
832 * - Each GS thread loads the vertex masks from LDS.
834 * Part 2: Compact ES threads in GS threads:
835 * - Compute the prefix sum for all 3 vertices from the masks. These are the new
836 * thread IDs for each vertex within the primitive.
837 * - Write the value of the old thread ID into the LDS address of the new thread ID.
838 * The ES thread will load the old thread ID and use it to load the position, VertexID,
840 * - Update vertex indices and null flag in the GS input VGPRs.
843 * Part 3: Update inputs GPRs
844 * - For all waves, update per-wave thread counts in input SGPRs.
845 * - In ES threads, update the ES input VGPRs (VertexID, InstanceID, TES inputs).
848 LLVMValueRef vtxindex
[3];
849 if (shader
->key
.opt
.ngg_culling
& SI_NGG_CULL_GS_FAST_LAUNCH_ALL
) {
850 /* For the GS fast launch, the VS prologs simply puts the Vertex IDs
853 vtxindex
[0] = ac_get_arg(&ctx
->ac
, ctx
->gs_vtx01_offset
);
854 vtxindex
[1] = ac_get_arg(&ctx
->ac
, ctx
->gs_vtx23_offset
);
855 vtxindex
[2] = ac_get_arg(&ctx
->ac
, ctx
->gs_vtx45_offset
);
857 vtxindex
[0] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 0, 16);
858 vtxindex
[1] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 16, 16);
859 vtxindex
[2] = si_unpack_param(ctx
, ctx
->gs_vtx23_offset
, 0, 16);
861 LLVMValueRef gs_vtxptr
[] = {
862 ngg_nogs_vertex_ptr(ctx
, vtxindex
[0]),
863 ngg_nogs_vertex_ptr(ctx
, vtxindex
[1]),
864 ngg_nogs_vertex_ptr(ctx
, vtxindex
[2]),
866 es_vtxptr
= ngg_nogs_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
));
868 LLVMValueRef gs_accepted
= ac_build_alloca(&ctx
->ac
, ctx
->ac
.i32
, "");
870 /* Do culling in GS threads. */
871 ac_build_ifcc(&ctx
->ac
, si_is_gs_thread(ctx
), 16002);
873 /* Load positions. */
874 LLVMValueRef pos
[3][4] = {};
875 for (unsigned vtx
= 0; vtx
< 3; vtx
++) {
876 for (unsigned chan
= 0; chan
< 4; chan
++) {
878 if (chan
== 0 || chan
== 1)
879 index
= lds_pos_x_div_w
+ chan
;
886 ac_build_gep0(&ctx
->ac
, gs_vtxptr
[vtx
], LLVMConstInt(ctx
->ac
.i32
, index
, 0));
887 pos
[vtx
][chan
] = LLVMBuildLoad(builder
, addr
, "");
888 pos
[vtx
][chan
] = ac_to_float(&ctx
->ac
, pos
[vtx
][chan
]);
892 /* Load the viewport state for small prim culling. */
893 LLVMValueRef vp
= ac_build_load_invariant(
894 &ctx
->ac
, ac_get_arg(&ctx
->ac
, ctx
->small_prim_cull_info
), ctx
->ac
.i32_0
);
895 vp
= LLVMBuildBitCast(builder
, vp
, ctx
->ac
.v4f32
, "");
896 LLVMValueRef vp_scale
[2], vp_translate
[2];
897 vp_scale
[0] = ac_llvm_extract_elem(&ctx
->ac
, vp
, 0);
898 vp_scale
[1] = ac_llvm_extract_elem(&ctx
->ac
, vp
, 1);
899 vp_translate
[0] = ac_llvm_extract_elem(&ctx
->ac
, vp
, 2);
900 vp_translate
[1] = ac_llvm_extract_elem(&ctx
->ac
, vp
, 3);
902 /* Get the small prim filter precision. */
903 LLVMValueRef small_prim_precision
= si_unpack_param(ctx
, ctx
->vs_state_bits
, 7, 4);
904 small_prim_precision
=
905 LLVMBuildOr(builder
, small_prim_precision
, LLVMConstInt(ctx
->ac
.i32
, 0x70, 0), "");
906 small_prim_precision
=
907 LLVMBuildShl(builder
, small_prim_precision
, LLVMConstInt(ctx
->ac
.i32
, 23, 0), "");
908 small_prim_precision
= LLVMBuildBitCast(builder
, small_prim_precision
, ctx
->ac
.f32
, "");
910 /* Execute culling code. */
911 struct ac_cull_options options
= {};
912 options
.cull_front
= shader
->key
.opt
.ngg_culling
& SI_NGG_CULL_FRONT_FACE
;
913 options
.cull_back
= shader
->key
.opt
.ngg_culling
& SI_NGG_CULL_BACK_FACE
;
914 options
.cull_view_xy
= shader
->key
.opt
.ngg_culling
& SI_NGG_CULL_VIEW_SMALLPRIMS
;
915 options
.cull_small_prims
= options
.cull_view_xy
;
916 options
.cull_zero_area
= options
.cull_front
|| options
.cull_back
;
917 options
.cull_w
= true;
919 /* Tell ES threads whether their vertex survived. */
920 ac_build_ifcc(&ctx
->ac
,
921 ac_cull_triangle(&ctx
->ac
, pos
, ctx
->ac
.i1true
, vp_scale
, vp_translate
,
922 small_prim_precision
, &options
),
925 LLVMBuildStore(builder
, ctx
->ac
.i32_1
, gs_accepted
);
926 for (unsigned vtx
= 0; vtx
< 3; vtx
++) {
927 LLVMBuildStore(builder
, ctx
->ac
.i8_1
,
928 si_build_gep_i8(ctx
, gs_vtxptr
[vtx
], lds_byte0_accept_flag
));
931 ac_build_endif(&ctx
->ac
, 16003);
933 ac_build_endif(&ctx
->ac
, 16002);
934 ac_build_s_barrier(&ctx
->ac
);
936 gs_accepted
= LLVMBuildLoad(builder
, gs_accepted
, "");
938 LLVMValueRef es_accepted
= ac_build_alloca(&ctx
->ac
, ctx
->ac
.i1
, "");
940 /* Convert the per-vertex flag to a thread bitmask in ES threads and store it in LDS. */
941 ac_build_ifcc(&ctx
->ac
, si_is_es_thread(ctx
), 16007);
943 LLVMValueRef es_accepted_flag
=
944 LLVMBuildLoad(builder
, si_build_gep_i8(ctx
, es_vtxptr
, lds_byte0_accept_flag
), "");
946 LLVMValueRef es_accepted_bool
=
947 LLVMBuildICmp(builder
, LLVMIntNE
, es_accepted_flag
, ctx
->ac
.i8_0
, "");
948 LLVMValueRef es_mask
= ac_get_i1_sgpr_mask(&ctx
->ac
, es_accepted_bool
);
950 LLVMBuildStore(builder
, es_accepted_bool
, es_accepted
);
952 ac_build_ifcc(&ctx
->ac
, LLVMBuildICmp(builder
, LLVMIntEQ
, tid
, ctx
->ac
.i32_0
, ""), 16008);
954 LLVMBuildStore(builder
, es_mask
,
955 ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, get_wave_id_in_tg(ctx
)));
957 ac_build_endif(&ctx
->ac
, 16008);
959 ac_build_endif(&ctx
->ac
, 16007);
960 ac_build_s_barrier(&ctx
->ac
);
962 /* Load the vertex masks and compute the new ES thread count. */
963 LLVMValueRef es_mask
[2], new_num_es_threads
, kill_wave
;
964 load_bitmasks_2x64(ctx
, ctx
->gs_ngg_scratch
, 0, es_mask
, &new_num_es_threads
);
965 new_num_es_threads
= ac_build_readlane_no_opt_barrier(&ctx
->ac
, new_num_es_threads
, NULL
);
967 /* ES threads compute their prefix sum, which is the new ES thread ID.
968 * Then they write the value of the old thread ID into the LDS address
969 * of the new thread ID. It will be used it to load input VGPRs from
970 * the old thread's LDS location.
972 ac_build_ifcc(&ctx
->ac
, LLVMBuildLoad(builder
, es_accepted
, ""), 16009);
974 LLVMValueRef old_id
= get_thread_id_in_tg(ctx
);
975 LLVMValueRef new_id
= ac_prefix_bitcount_2x64(&ctx
->ac
, es_mask
, old_id
);
978 builder
, LLVMBuildTrunc(builder
, old_id
, ctx
->ac
.i8
, ""),
979 si_build_gep_i8(ctx
, ngg_nogs_vertex_ptr(ctx
, new_id
), lds_byte0_old_thread_id
));
980 LLVMBuildStore(builder
, LLVMBuildTrunc(builder
, new_id
, ctx
->ac
.i8
, ""),
981 si_build_gep_i8(ctx
, es_vtxptr
, lds_byte1_new_thread_id
));
983 ac_build_endif(&ctx
->ac
, 16009);
985 /* Kill waves that have inactive threads. */
986 kill_wave
= LLVMBuildICmp(builder
, LLVMIntULE
,
987 ac_build_imax(&ctx
->ac
, new_num_es_threads
, ngg_get_prim_cnt(ctx
)),
988 LLVMBuildMul(builder
, get_wave_id_in_tg(ctx
),
989 LLVMConstInt(ctx
->ac
.i32
, ctx
->ac
.wave_size
, 0), ""),
991 ac_build_ifcc(&ctx
->ac
, kill_wave
, 19202);
993 /* If we are killing wave 0, send that there are no primitives
994 * in this threadgroup.
996 ac_build_sendmsg_gs_alloc_req(&ctx
->ac
, get_wave_id_in_tg(ctx
), ctx
->ac
.i32_0
, ctx
->ac
.i32_0
);
997 ac_build_s_endpgm(&ctx
->ac
);
999 ac_build_endif(&ctx
->ac
, 19202);
1000 ac_build_s_barrier(&ctx
->ac
);
1002 /* Send the final vertex and primitive counts. */
1003 ac_build_sendmsg_gs_alloc_req(&ctx
->ac
, get_wave_id_in_tg(ctx
), new_num_es_threads
,
1004 ngg_get_prim_cnt(ctx
));
1006 /* Update thread counts in SGPRs. */
1007 LLVMValueRef new_gs_tg_info
= ac_get_arg(&ctx
->ac
, ctx
->gs_tg_info
);
1008 LLVMValueRef new_merged_wave_info
= ac_get_arg(&ctx
->ac
, ctx
->merged_wave_info
);
1010 /* This also converts the thread count from the total count to the per-wave count. */
1011 update_thread_counts(ctx
, &new_num_es_threads
, &new_gs_tg_info
, 9, 12, &new_merged_wave_info
, 8,
1014 /* Update vertex indices in VGPR0 (same format as NGG passthrough). */
1015 LLVMValueRef new_vgpr0
= ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i32
, "");
1017 /* Set the null flag at the beginning (culled), and then
1018 * overwrite it for accepted primitives.
1020 LLVMBuildStore(builder
, LLVMConstInt(ctx
->ac
.i32
, 1u << 31, 0), new_vgpr0
);
1022 /* Get vertex indices after vertex compaction. */
1023 ac_build_ifcc(&ctx
->ac
, LLVMBuildTrunc(builder
, gs_accepted
, ctx
->ac
.i1
, ""), 16011);
1025 struct ac_ngg_prim prim
= {};
1026 prim
.num_vertices
= 3;
1027 prim
.isnull
= ctx
->ac
.i1false
;
1029 for (unsigned vtx
= 0; vtx
< 3; vtx
++) {
1030 prim
.index
[vtx
] = LLVMBuildLoad(
1031 builder
, si_build_gep_i8(ctx
, gs_vtxptr
[vtx
], lds_byte1_new_thread_id
), "");
1032 prim
.index
[vtx
] = LLVMBuildZExt(builder
, prim
.index
[vtx
], ctx
->ac
.i32
, "");
1033 prim
.edgeflag
[vtx
] = ngg_get_initial_edgeflag(ctx
, vtx
);
1036 /* Set the new GS input VGPR. */
1037 LLVMBuildStore(builder
, ac_pack_prim_export(&ctx
->ac
, &prim
), new_vgpr0
);
1039 ac_build_endif(&ctx
->ac
, 16011);
1041 if (gfx10_ngg_export_prim_early(shader
))
1042 gfx10_ngg_build_export_prim(ctx
, NULL
, LLVMBuildLoad(builder
, new_vgpr0
, ""));
1044 /* Set the new ES input VGPRs. */
1045 LLVMValueRef es_data
[4];
1046 LLVMValueRef old_thread_id
= ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i32
, "");
1048 for (unsigned i
= 0; i
< 4; i
++)
1049 es_data
[i
] = ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i32
, "");
1051 ac_build_ifcc(&ctx
->ac
, LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
, tid
, new_num_es_threads
, ""),
1054 LLVMValueRef old_id
, old_es_vtxptr
, tmp
;
1056 /* Load ES input VGPRs from the ES thread before compaction. */
1057 old_id
= LLVMBuildLoad(builder
, si_build_gep_i8(ctx
, es_vtxptr
, lds_byte0_old_thread_id
), "");
1058 old_id
= LLVMBuildZExt(builder
, old_id
, ctx
->ac
.i32
, "");
1060 LLVMBuildStore(builder
, old_id
, old_thread_id
);
1061 old_es_vtxptr
= ngg_nogs_vertex_ptr(ctx
, old_id
);
1063 for (unsigned i
= 0; i
< 2; i
++) {
1064 tmp
= LLVMBuildLoad(
1066 ac_build_gep0(&ctx
->ac
, old_es_vtxptr
, LLVMConstInt(ctx
->ac
.i32
, lds_vertex_id
+ i
, 0)),
1068 LLVMBuildStore(builder
, tmp
, es_data
[i
]);
1071 if (ctx
->type
== PIPE_SHADER_TESS_EVAL
) {
1072 tmp
= LLVMBuildLoad(builder
,
1073 si_build_gep_i8(ctx
, old_es_vtxptr
, lds_byte2_tes_rel_patch_id
), "");
1074 tmp
= LLVMBuildZExt(builder
, tmp
, ctx
->ac
.i32
, "");
1075 LLVMBuildStore(builder
, tmp
, es_data
[2]);
1077 if (uses_tes_prim_id
) {
1078 tmp
= LLVMBuildLoad(builder
,
1079 ac_build_gep0(&ctx
->ac
, old_es_vtxptr
,
1080 LLVMConstInt(ctx
->ac
.i32
, lds_tes_patch_id
, 0)),
1082 LLVMBuildStore(builder
, tmp
, es_data
[3]);
1086 ac_build_endif(&ctx
->ac
, 16012);
1088 /* Return values for the main function. */
1089 LLVMValueRef ret
= ctx
->return_value
;
1092 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, new_gs_tg_info
, 2, "");
1093 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, new_merged_wave_info
, 3, "");
1094 if (ctx
->type
== PIPE_SHADER_TESS_EVAL
)
1095 ret
= si_insert_input_ret(ctx
, ret
, ctx
->tcs_offchip_offset
, 4);
1097 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->rw_buffers
, 8 + SI_SGPR_RW_BUFFERS
);
1098 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->bindless_samplers_and_images
,
1099 8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES
);
1100 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->const_and_shader_buffers
,
1101 8 + SI_SGPR_CONST_AND_SHADER_BUFFERS
);
1102 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->samplers_and_images
, 8 + SI_SGPR_SAMPLERS_AND_IMAGES
);
1103 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->vs_state_bits
, 8 + SI_SGPR_VS_STATE_BITS
);
1105 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
1106 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->args
.base_vertex
, 8 + SI_SGPR_BASE_VERTEX
);
1107 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->args
.start_instance
, 8 + SI_SGPR_START_INSTANCE
);
1108 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->args
.draw_id
, 8 + SI_SGPR_DRAWID
);
1109 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->vertex_buffers
, 8 + SI_VS_NUM_USER_SGPR
);
1111 for (unsigned i
= 0; i
< shader
->selector
->num_vbos_in_user_sgprs
; i
++) {
1112 ret
= si_insert_input_v4i32(ctx
, ret
, ctx
->vb_descriptors
[i
],
1113 8 + SI_SGPR_VS_VB_DESCRIPTOR_FIRST
+ i
* 4);
1116 assert(ctx
->type
== PIPE_SHADER_TESS_EVAL
);
1117 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->tcs_offchip_layout
, 8 + SI_SGPR_TES_OFFCHIP_LAYOUT
);
1118 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->tes_offchip_addr
, 8 + SI_SGPR_TES_OFFCHIP_ADDR
);
1122 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
1123 if (shader
->selector
->num_vbos_in_user_sgprs
) {
1124 vgpr
= 8 + SI_SGPR_VS_VB_DESCRIPTOR_FIRST
+ shader
->selector
->num_vbos_in_user_sgprs
* 4;
1126 vgpr
= 8 + GFX9_VSGS_NUM_USER_SGPR
+ 1;
1129 vgpr
= 8 + GFX9_TESGS_NUM_USER_SGPR
;
1132 val
= LLVMBuildLoad(builder
, new_vgpr0
, "");
1133 ret
= LLVMBuildInsertValue(builder
, ret
, ac_to_float(&ctx
->ac
, val
), vgpr
++, "");
1134 vgpr
++; /* gs_vtx23_offset */
1136 ret
= si_insert_input_ret_float(ctx
, ret
, ctx
->args
.gs_prim_id
, vgpr
++);
1137 ret
= si_insert_input_ret_float(ctx
, ret
, ctx
->args
.gs_invocation_id
, vgpr
++);
1138 vgpr
++; /* gs_vtx45_offset */
1140 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
1141 val
= LLVMBuildLoad(builder
, es_data
[0], "");
1142 ret
= LLVMBuildInsertValue(builder
, ret
, ac_to_float(&ctx
->ac
, val
), vgpr
++,
1143 ""); /* VGPR5 - VertexID */
1145 if (uses_instance_id
) {
1146 val
= LLVMBuildLoad(builder
, es_data
[1], "");
1147 ret
= LLVMBuildInsertValue(builder
, ret
, ac_to_float(&ctx
->ac
, val
), vgpr
++,
1148 ""); /* VGPR8 - InstanceID */
1153 assert(ctx
->type
== PIPE_SHADER_TESS_EVAL
);
1154 unsigned num_vgprs
= uses_tes_prim_id
? 4 : 3;
1155 for (unsigned i
= 0; i
< num_vgprs
; i
++) {
1156 val
= LLVMBuildLoad(builder
, es_data
[i
], "");
1157 ret
= LLVMBuildInsertValue(builder
, ret
, ac_to_float(&ctx
->ac
, val
), vgpr
++, "");
1162 /* Return the old thread ID. */
1163 val
= LLVMBuildLoad(builder
, old_thread_id
, "");
1164 ret
= LLVMBuildInsertValue(builder
, ret
, ac_to_float(&ctx
->ac
, val
), vgpr
++, "");
1166 /* These two also use LDS. */
1167 if (sel
->info
.writes_edgeflag
||
1168 (ctx
->type
== PIPE_SHADER_VERTEX
&& shader
->key
.mono
.u
.vs_export_prim_id
))
1169 ac_build_s_barrier(&ctx
->ac
);
1171 ctx
->return_value
= ret
;
1175 * Emit the epilogue of an API VS or TES shader compiled as ESGS shader.
1177 void gfx10_emit_ngg_epilogue(struct ac_shader_abi
*abi
, unsigned max_outputs
, LLVMValueRef
*addrs
)
1179 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1180 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1181 struct si_shader_info
*info
= &sel
->info
;
1182 struct si_shader_output_values outputs
[PIPE_MAX_SHADER_OUTPUTS
];
1183 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1184 LLVMValueRef tmp
, tmp2
;
1186 assert(!ctx
->shader
->is_gs_copy_shader
);
1187 assert(info
->num_outputs
<= max_outputs
);
1189 LLVMValueRef vertex_ptr
= NULL
;
1191 if (sel
->so
.num_outputs
|| sel
->info
.writes_edgeflag
)
1192 vertex_ptr
= ngg_nogs_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
));
1194 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
1195 outputs
[i
].semantic_name
= info
->output_semantic_name
[i
];
1196 outputs
[i
].semantic_index
= info
->output_semantic_index
[i
];
1198 for (unsigned j
= 0; j
< 4; j
++) {
1199 outputs
[i
].vertex_stream
[j
] = (info
->output_streams
[i
] >> (2 * j
)) & 3;
1201 /* TODO: we may store more outputs than streamout needs,
1202 * but streamout performance isn't that important.
1204 if (sel
->so
.num_outputs
) {
1205 tmp
= ac_build_gep0(&ctx
->ac
, vertex_ptr
, LLVMConstInt(ctx
->ac
.i32
, 4 * i
+ j
, false));
1206 tmp2
= LLVMBuildLoad(builder
, addrs
[4 * i
+ j
], "");
1207 tmp2
= ac_to_integer(&ctx
->ac
, tmp2
);
1208 LLVMBuildStore(builder
, tmp2
, tmp
);
1212 /* Store the edgeflag at the end (if streamout is enabled) */
1213 if (info
->output_semantic_name
[i
] == TGSI_SEMANTIC_EDGEFLAG
&& sel
->info
.writes_edgeflag
) {
1214 LLVMValueRef edgeflag
= LLVMBuildLoad(builder
, addrs
[4 * i
], "");
1215 /* The output is a float, but the hw expects a 1-bit integer. */
1216 edgeflag
= LLVMBuildFPToUI(ctx
->ac
.builder
, edgeflag
, ctx
->ac
.i32
, "");
1217 edgeflag
= ac_build_umin(&ctx
->ac
, edgeflag
, ctx
->ac
.i32_1
);
1219 tmp
= LLVMConstInt(ctx
->ac
.i32
, ngg_nogs_vertex_size(ctx
->shader
) - 1, 0);
1220 tmp
= ac_build_gep0(&ctx
->ac
, vertex_ptr
, tmp
);
1221 LLVMBuildStore(builder
, edgeflag
, tmp
);
1225 bool unterminated_es_if_block
=
1226 !sel
->so
.num_outputs
&& !sel
->info
.writes_edgeflag
&&
1227 !ctx
->screen
->use_ngg_streamout
&& /* no query buffer */
1228 (ctx
->type
!= PIPE_SHADER_VERTEX
|| !ctx
->shader
->key
.mono
.u
.vs_export_prim_id
);
1230 if (!unterminated_es_if_block
)
1231 ac_build_endif(&ctx
->ac
, ctx
->merged_wrap_if_label
);
1233 LLVMValueRef is_gs_thread
= si_is_gs_thread(ctx
);
1234 LLVMValueRef is_es_thread
= si_is_es_thread(ctx
);
1235 LLVMValueRef vtxindex
[3];
1237 if (ctx
->shader
->key
.opt
.ngg_culling
) {
1238 vtxindex
[0] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 0, 9);
1239 vtxindex
[1] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 10, 9);
1240 vtxindex
[2] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 20, 9);
1242 vtxindex
[0] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 0, 16);
1243 vtxindex
[1] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 16, 16);
1244 vtxindex
[2] = si_unpack_param(ctx
, ctx
->gs_vtx23_offset
, 0, 16);
1247 /* Determine the number of vertices per primitive. */
1248 unsigned num_vertices
;
1249 LLVMValueRef num_vertices_val
= ngg_get_vertices_per_prim(ctx
, &num_vertices
);
1252 LLVMValueRef emitted_prims
= NULL
;
1254 if (sel
->so
.num_outputs
) {
1255 assert(!unterminated_es_if_block
);
1257 struct ngg_streamout nggso
= {};
1258 nggso
.num_vertices
= num_vertices_val
;
1259 nggso
.prim_enable
[0] = is_gs_thread
;
1261 for (unsigned i
= 0; i
< num_vertices
; ++i
)
1262 nggso
.vertices
[i
] = ngg_nogs_vertex_ptr(ctx
, vtxindex
[i
]);
1264 build_streamout(ctx
, &nggso
);
1265 emitted_prims
= nggso
.emit
[0];
1268 LLVMValueRef user_edgeflags
[3] = {};
1270 if (sel
->info
.writes_edgeflag
) {
1271 assert(!unterminated_es_if_block
);
1273 /* Streamout already inserted the barrier, so don't insert it again. */
1274 if (!sel
->so
.num_outputs
)
1275 ac_build_s_barrier(&ctx
->ac
);
1277 ac_build_ifcc(&ctx
->ac
, is_gs_thread
, 5400);
1278 /* Load edge flags from ES threads and store them into VGPRs in GS threads. */
1279 for (unsigned i
= 0; i
< num_vertices
; i
++) {
1280 tmp
= ngg_nogs_vertex_ptr(ctx
, vtxindex
[i
]);
1281 tmp2
= LLVMConstInt(ctx
->ac
.i32
, ngg_nogs_vertex_size(ctx
->shader
) - 1, 0);
1282 tmp
= ac_build_gep0(&ctx
->ac
, tmp
, tmp2
);
1283 tmp
= LLVMBuildLoad(builder
, tmp
, "");
1284 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
1286 user_edgeflags
[i
] = ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i1
, "");
1287 LLVMBuildStore(builder
, tmp
, user_edgeflags
[i
]);
1289 ac_build_endif(&ctx
->ac
, 5400);
1292 /* Copy Primitive IDs from GS threads to the LDS address corresponding
1293 * to the ES thread of the provoking vertex.
1295 if (ctx
->type
== PIPE_SHADER_VERTEX
&& ctx
->shader
->key
.mono
.u
.vs_export_prim_id
) {
1296 assert(!unterminated_es_if_block
);
1298 /* Streamout and edge flags use LDS. Make it idle, so that we can reuse it. */
1299 if (sel
->so
.num_outputs
|| sel
->info
.writes_edgeflag
)
1300 ac_build_s_barrier(&ctx
->ac
);
1302 ac_build_ifcc(&ctx
->ac
, is_gs_thread
, 5400);
1303 /* Extract the PROVOKING_VTX_INDEX field. */
1304 LLVMValueRef provoking_vtx_in_prim
= si_unpack_param(ctx
, ctx
->vs_state_bits
, 4, 2);
1306 /* provoking_vtx_index = vtxindex[provoking_vtx_in_prim]; */
1307 LLVMValueRef indices
= ac_build_gather_values(&ctx
->ac
, vtxindex
, 3);
1308 LLVMValueRef provoking_vtx_index
=
1309 LLVMBuildExtractElement(builder
, indices
, provoking_vtx_in_prim
, "");
1310 LLVMValueRef vertex_ptr
= ngg_nogs_vertex_ptr(ctx
, provoking_vtx_index
);
1312 LLVMBuildStore(builder
, ac_get_arg(&ctx
->ac
, ctx
->args
.gs_prim_id
),
1313 ac_build_gep0(&ctx
->ac
, vertex_ptr
, ctx
->ac
.i32_0
));
1314 ac_build_endif(&ctx
->ac
, 5400);
1317 /* Update query buffer */
1318 if (ctx
->screen
->use_ngg_streamout
&& !info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
]) {
1319 assert(!unterminated_es_if_block
);
1321 tmp
= si_unpack_param(ctx
, ctx
->vs_state_bits
, 6, 1);
1322 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
1323 ac_build_ifcc(&ctx
->ac
, tmp
, 5029); /* if (STREAMOUT_QUERY_ENABLED) */
1324 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, get_wave_id_in_tg(ctx
), ctx
->ac
.i32_0
, "");
1325 ac_build_ifcc(&ctx
->ac
, tmp
, 5030);
1326 tmp
= LLVMBuildICmp(builder
, LLVMIntULE
, ac_get_thread_id(&ctx
->ac
),
1327 sel
->so
.num_outputs
? ctx
->ac
.i32_1
: ctx
->ac
.i32_0
, "");
1328 ac_build_ifcc(&ctx
->ac
, tmp
, 5031);
1330 LLVMValueRef args
[] = {
1331 ngg_get_prim_cnt(ctx
),
1332 ngg_get_query_buf(ctx
),
1333 LLVMConstInt(ctx
->ac
.i32
, 16, false), /* offset of stream[0].generated_primitives */
1334 ctx
->ac
.i32_0
, /* soffset */
1335 ctx
->ac
.i32_0
, /* cachepolicy */
1338 if (sel
->so
.num_outputs
) {
1339 args
[0] = ac_build_writelane(&ctx
->ac
, args
[0], emitted_prims
, ctx
->ac
.i32_1
);
1340 args
[2] = ac_build_writelane(&ctx
->ac
, args
[2], LLVMConstInt(ctx
->ac
.i32
, 24, false),
1344 /* TODO: should this be 64-bit atomics? */
1345 ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.raw.buffer.atomic.add.i32", ctx
->ac
.i32
, args
, 5,
1348 ac_build_endif(&ctx
->ac
, 5031);
1349 ac_build_endif(&ctx
->ac
, 5030);
1350 ac_build_endif(&ctx
->ac
, 5029);
1353 /* Build the primitive export. */
1354 if (!gfx10_ngg_export_prim_early(ctx
->shader
)) {
1355 assert(!unterminated_es_if_block
);
1356 gfx10_ngg_build_export_prim(ctx
, user_edgeflags
, NULL
);
1359 /* Export per-vertex data (positions and parameters). */
1360 if (!unterminated_es_if_block
)
1361 ac_build_ifcc(&ctx
->ac
, is_es_thread
, 6002);
1365 /* Unconditionally (re-)load the values for proper SSA form. */
1366 for (i
= 0; i
< info
->num_outputs
; i
++) {
1367 /* If the NGG cull shader part computed the position, don't
1368 * use the position from the current shader part. Instead,
1371 if (info
->output_semantic_name
[i
] == TGSI_SEMANTIC_POSITION
&&
1372 ctx
->shader
->key
.opt
.ngg_culling
) {
1373 vertex_ptr
= ngg_nogs_vertex_ptr(ctx
, ac_get_arg(&ctx
->ac
, ctx
->ngg_old_thread_id
));
1375 for (unsigned j
= 0; j
< 4; j
++) {
1376 tmp
= LLVMConstInt(ctx
->ac
.i32
, lds_pos_x
+ j
, 0);
1377 tmp
= ac_build_gep0(&ctx
->ac
, vertex_ptr
, tmp
);
1378 tmp
= LLVMBuildLoad(builder
, tmp
, "");
1379 outputs
[i
].values
[j
] = ac_to_float(&ctx
->ac
, tmp
);
1382 for (unsigned j
= 0; j
< 4; j
++) {
1383 outputs
[i
].values
[j
] = LLVMBuildLoad(builder
, addrs
[4 * i
+ j
], "");
1388 if (ctx
->shader
->key
.mono
.u
.vs_export_prim_id
) {
1389 outputs
[i
].semantic_name
= TGSI_SEMANTIC_PRIMID
;
1390 outputs
[i
].semantic_index
= 0;
1392 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
1393 /* Wait for GS stores to finish. */
1394 ac_build_s_barrier(&ctx
->ac
);
1396 tmp
= ngg_nogs_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
));
1397 tmp
= ac_build_gep0(&ctx
->ac
, tmp
, ctx
->ac
.i32_0
);
1398 outputs
[i
].values
[0] = LLVMBuildLoad(builder
, tmp
, "");
1400 assert(ctx
->type
== PIPE_SHADER_TESS_EVAL
);
1401 outputs
[i
].values
[0] = si_get_primitive_id(ctx
, 0);
1404 outputs
[i
].values
[0] = ac_to_float(&ctx
->ac
, outputs
[i
].values
[0]);
1405 for (unsigned j
= 1; j
< 4; j
++)
1406 outputs
[i
].values
[j
] = LLVMGetUndef(ctx
->ac
.f32
);
1408 memset(outputs
[i
].vertex_stream
, 0, sizeof(outputs
[i
].vertex_stream
));
1412 si_llvm_build_vs_exports(ctx
, outputs
, i
);
1414 ac_build_endif(&ctx
->ac
, 6002);
1417 static LLVMValueRef
ngg_gs_get_vertex_storage(struct si_shader_context
*ctx
)
1419 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1420 const struct si_shader_info
*info
= &sel
->info
;
1422 LLVMTypeRef elements
[2] = {
1423 LLVMArrayType(ctx
->ac
.i32
, 4 * info
->num_outputs
),
1424 LLVMArrayType(ctx
->ac
.i8
, 4),
1426 LLVMTypeRef type
= LLVMStructTypeInContext(ctx
->ac
.context
, elements
, 2, false);
1427 type
= LLVMPointerType(LLVMArrayType(type
, 0), AC_ADDR_SPACE_LDS
);
1428 return LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->gs_ngg_emit
, type
, "");
1432 * Return a pointer to the LDS storage reserved for the N'th vertex, where N
1433 * is in emit order; that is:
1434 * - during the epilogue, N is the threadidx (relative to the entire threadgroup)
1435 * - during vertex emit, i.e. while the API GS shader invocation is running,
1436 * N = threadidx * gs_max_out_vertices + emitidx
1438 * Goals of the LDS memory layout:
1439 * 1. Eliminate bank conflicts on write for geometry shaders that have all emits
1440 * in uniform control flow
1441 * 2. Eliminate bank conflicts on read for export if, additionally, there is no
1443 * 3. Agnostic to the number of waves (since we don't know it before compiling)
1444 * 4. Allow coalescing of LDS instructions (ds_write_b128 etc.)
1445 * 5. Avoid wasting memory.
1447 * We use an AoS layout due to point 4 (this also helps point 3). In an AoS
1448 * layout, elimination of bank conflicts requires that each vertex occupy an
1449 * odd number of dwords. We use the additional dword to store the output stream
1450 * index as well as a flag to indicate whether this vertex ends a primitive
1451 * for rasterization.
1453 * Swizzling is required to satisfy points 1 and 2 simultaneously.
1455 * Vertices are stored in export order (gsthread * gs_max_out_vertices + emitidx).
1456 * Indices are swizzled in groups of 32, which ensures point 1 without
1457 * disturbing point 2.
1459 * \return an LDS pointer to type {[N x i32], [4 x i8]}
1461 static LLVMValueRef
ngg_gs_vertex_ptr(struct si_shader_context
*ctx
, LLVMValueRef vertexidx
)
1463 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1464 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1465 LLVMValueRef storage
= ngg_gs_get_vertex_storage(ctx
);
1467 /* gs_max_out_vertices = 2^(write_stride_2exp) * some odd number */
1468 unsigned write_stride_2exp
= ffs(sel
->gs_max_out_vertices
) - 1;
1469 if (write_stride_2exp
) {
1470 LLVMValueRef row
= LLVMBuildLShr(builder
, vertexidx
, LLVMConstInt(ctx
->ac
.i32
, 5, false), "");
1471 LLVMValueRef swizzle
= LLVMBuildAnd(
1472 builder
, row
, LLVMConstInt(ctx
->ac
.i32
, (1u << write_stride_2exp
) - 1, false), "");
1473 vertexidx
= LLVMBuildXor(builder
, vertexidx
, swizzle
, "");
1476 return ac_build_gep0(&ctx
->ac
, storage
, vertexidx
);
1479 static LLVMValueRef
ngg_gs_emit_vertex_ptr(struct si_shader_context
*ctx
, LLVMValueRef gsthread
,
1480 LLVMValueRef emitidx
)
1482 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1483 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1486 tmp
= LLVMConstInt(ctx
->ac
.i32
, sel
->gs_max_out_vertices
, false);
1487 tmp
= LLVMBuildMul(builder
, tmp
, gsthread
, "");
1488 const LLVMValueRef vertexidx
= LLVMBuildAdd(builder
, tmp
, emitidx
, "");
1489 return ngg_gs_vertex_ptr(ctx
, vertexidx
);
1492 static LLVMValueRef
ngg_gs_get_emit_output_ptr(struct si_shader_context
*ctx
,
1493 LLVMValueRef vertexptr
, unsigned out_idx
)
1495 LLVMValueRef gep_idx
[3] = {
1496 ctx
->ac
.i32_0
, /* implied C-style array */
1497 ctx
->ac
.i32_0
, /* first struct entry */
1498 LLVMConstInt(ctx
->ac
.i32
, out_idx
, false),
1500 return LLVMBuildGEP(ctx
->ac
.builder
, vertexptr
, gep_idx
, 3, "");
1503 static LLVMValueRef
ngg_gs_get_emit_primflag_ptr(struct si_shader_context
*ctx
,
1504 LLVMValueRef vertexptr
, unsigned stream
)
1506 LLVMValueRef gep_idx
[3] = {
1507 ctx
->ac
.i32_0
, /* implied C-style array */
1508 ctx
->ac
.i32_1
, /* second struct entry */
1509 LLVMConstInt(ctx
->ac
.i32
, stream
, false),
1511 return LLVMBuildGEP(ctx
->ac
.builder
, vertexptr
, gep_idx
, 3, "");
1514 void gfx10_ngg_gs_emit_vertex(struct si_shader_context
*ctx
, unsigned stream
, LLVMValueRef
*addrs
)
1516 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1517 const struct si_shader_info
*info
= &sel
->info
;
1518 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1520 const LLVMValueRef vertexidx
= LLVMBuildLoad(builder
, ctx
->gs_next_vertex
[stream
], "");
1522 /* If this thread has already emitted the declared maximum number of
1523 * vertices, skip the write: excessive vertex emissions are not
1524 * supposed to have any effect.
1526 const LLVMValueRef can_emit
=
1527 LLVMBuildICmp(builder
, LLVMIntULT
, vertexidx
,
1528 LLVMConstInt(ctx
->ac
.i32
, sel
->gs_max_out_vertices
, false), "");
1530 tmp
= LLVMBuildAdd(builder
, vertexidx
, ctx
->ac
.i32_1
, "");
1531 tmp
= LLVMBuildSelect(builder
, can_emit
, tmp
, vertexidx
, "");
1532 LLVMBuildStore(builder
, tmp
, ctx
->gs_next_vertex
[stream
]);
1534 ac_build_ifcc(&ctx
->ac
, can_emit
, 9001);
1536 const LLVMValueRef vertexptr
= ngg_gs_emit_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
), vertexidx
);
1537 unsigned out_idx
= 0;
1538 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
1539 for (unsigned chan
= 0; chan
< 4; chan
++, out_idx
++) {
1540 if (!(info
->output_usagemask
[i
] & (1 << chan
)) ||
1541 ((info
->output_streams
[i
] >> (2 * chan
)) & 3) != stream
)
1544 LLVMValueRef out_val
= LLVMBuildLoad(builder
, addrs
[4 * i
+ chan
], "");
1545 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
1546 LLVMBuildStore(builder
, out_val
, ngg_gs_get_emit_output_ptr(ctx
, vertexptr
, out_idx
));
1549 assert(out_idx
* 4 == sel
->gsvs_vertex_size
);
1551 /* Determine and store whether this vertex completed a primitive. */
1552 const LLVMValueRef curverts
= LLVMBuildLoad(builder
, ctx
->gs_curprim_verts
[stream
], "");
1554 tmp
= LLVMConstInt(ctx
->ac
.i32
, u_vertices_per_prim(sel
->gs_output_prim
) - 1, false);
1555 const LLVMValueRef iscompleteprim
= LLVMBuildICmp(builder
, LLVMIntUGE
, curverts
, tmp
, "");
1557 /* Since the geometry shader emits triangle strips, we need to
1558 * track which primitive is odd and swap vertex indices to get
1559 * the correct vertex order.
1561 LLVMValueRef is_odd
= ctx
->ac
.i1false
;
1562 if (stream
== 0 && u_vertices_per_prim(sel
->gs_output_prim
) == 3) {
1563 tmp
= LLVMBuildAnd(builder
, curverts
, ctx
->ac
.i32_1
, "");
1564 is_odd
= LLVMBuildICmp(builder
, LLVMIntEQ
, tmp
, ctx
->ac
.i32_1
, "");
1567 tmp
= LLVMBuildAdd(builder
, curverts
, ctx
->ac
.i32_1
, "");
1568 LLVMBuildStore(builder
, tmp
, ctx
->gs_curprim_verts
[stream
]);
1570 /* The per-vertex primitive flag encoding:
1571 * bit 0: whether this vertex finishes a primitive
1572 * bit 1: whether the primitive is odd (if we are emitting triangle strips)
1574 tmp
= LLVMBuildZExt(builder
, iscompleteprim
, ctx
->ac
.i8
, "");
1577 LLVMBuildShl(builder
, LLVMBuildZExt(builder
, is_odd
, ctx
->ac
.i8
, ""), ctx
->ac
.i8_1
, ""), "");
1578 LLVMBuildStore(builder
, tmp
, ngg_gs_get_emit_primflag_ptr(ctx
, vertexptr
, stream
));
1580 tmp
= LLVMBuildLoad(builder
, ctx
->gs_generated_prims
[stream
], "");
1581 tmp
= LLVMBuildAdd(builder
, tmp
, LLVMBuildZExt(builder
, iscompleteprim
, ctx
->ac
.i32
, ""), "");
1582 LLVMBuildStore(builder
, tmp
, ctx
->gs_generated_prims
[stream
]);
1584 ac_build_endif(&ctx
->ac
, 9001);
1587 void gfx10_ngg_gs_emit_prologue(struct si_shader_context
*ctx
)
1589 /* Zero out the part of LDS scratch that is used to accumulate the
1590 * per-stream generated primitive count.
1592 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1593 LLVMValueRef scratchptr
= ctx
->gs_ngg_scratch
;
1594 LLVMValueRef tid
= get_thread_id_in_tg(ctx
);
1597 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
1598 ac_build_ifcc(&ctx
->ac
, tmp
, 5090);
1600 LLVMValueRef ptr
= ac_build_gep0(&ctx
->ac
, scratchptr
, tid
);
1601 LLVMBuildStore(builder
, ctx
->ac
.i32_0
, ptr
);
1603 ac_build_endif(&ctx
->ac
, 5090);
1605 ac_build_s_barrier(&ctx
->ac
);
1608 void gfx10_ngg_gs_emit_epilogue(struct si_shader_context
*ctx
)
1610 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1611 const struct si_shader_info
*info
= &sel
->info
;
1612 const unsigned verts_per_prim
= u_vertices_per_prim(sel
->gs_output_prim
);
1613 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1614 LLVMValueRef i8_0
= LLVMConstInt(ctx
->ac
.i8
, 0, false);
1615 LLVMValueRef tmp
, tmp2
;
1617 /* Zero out remaining (non-emitted) primitive flags.
1619 * Note: Alternatively, we could pass the relevant gs_next_vertex to
1620 * the emit threads via LDS. This is likely worse in the expected
1621 * typical case where each GS thread emits the full set of
1624 for (unsigned stream
= 0; stream
< 4; ++stream
) {
1625 if (!info
->num_stream_output_components
[stream
])
1628 const LLVMValueRef gsthread
= get_thread_id_in_tg(ctx
);
1630 ac_build_bgnloop(&ctx
->ac
, 5100);
1632 const LLVMValueRef vertexidx
= LLVMBuildLoad(builder
, ctx
->gs_next_vertex
[stream
], "");
1633 tmp
= LLVMBuildICmp(builder
, LLVMIntUGE
, vertexidx
,
1634 LLVMConstInt(ctx
->ac
.i32
, sel
->gs_max_out_vertices
, false), "");
1635 ac_build_ifcc(&ctx
->ac
, tmp
, 5101);
1636 ac_build_break(&ctx
->ac
);
1637 ac_build_endif(&ctx
->ac
, 5101);
1639 tmp
= LLVMBuildAdd(builder
, vertexidx
, ctx
->ac
.i32_1
, "");
1640 LLVMBuildStore(builder
, tmp
, ctx
->gs_next_vertex
[stream
]);
1642 tmp
= ngg_gs_emit_vertex_ptr(ctx
, gsthread
, vertexidx
);
1643 LLVMBuildStore(builder
, i8_0
, ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, stream
));
1645 ac_build_endloop(&ctx
->ac
, 5100);
1648 /* Accumulate generated primitives counts across the entire threadgroup. */
1649 for (unsigned stream
= 0; stream
< 4; ++stream
) {
1650 if (!info
->num_stream_output_components
[stream
])
1653 LLVMValueRef numprims
= LLVMBuildLoad(builder
, ctx
->gs_generated_prims
[stream
], "");
1654 numprims
= ac_build_reduce(&ctx
->ac
, numprims
, nir_op_iadd
, ctx
->ac
.wave_size
);
1656 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, ac_get_thread_id(&ctx
->ac
), ctx
->ac
.i32_0
, "");
1657 ac_build_ifcc(&ctx
->ac
, tmp
, 5105);
1660 builder
, LLVMAtomicRMWBinOpAdd
,
1661 ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, LLVMConstInt(ctx
->ac
.i32
, stream
, false)),
1662 numprims
, LLVMAtomicOrderingMonotonic
, false);
1664 ac_build_endif(&ctx
->ac
, 5105);
1667 ac_build_endif(&ctx
->ac
, ctx
->merged_wrap_if_label
);
1669 ac_build_s_barrier(&ctx
->ac
);
1671 const LLVMValueRef tid
= get_thread_id_in_tg(ctx
);
1672 LLVMValueRef num_emit_threads
= ngg_get_prim_cnt(ctx
);
1675 if (sel
->so
.num_outputs
) {
1676 struct ngg_streamout nggso
= {};
1678 nggso
.num_vertices
= LLVMConstInt(ctx
->ac
.i32
, verts_per_prim
, false);
1680 LLVMValueRef vertexptr
= ngg_gs_vertex_ptr(ctx
, tid
);
1681 for (unsigned stream
= 0; stream
< 4; ++stream
) {
1682 if (!info
->num_stream_output_components
[stream
])
1685 tmp
= LLVMBuildLoad(builder
, ngg_gs_get_emit_primflag_ptr(ctx
, vertexptr
, stream
), "");
1686 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
1687 tmp2
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, num_emit_threads
, "");
1688 nggso
.prim_enable
[stream
] = LLVMBuildAnd(builder
, tmp
, tmp2
, "");
1691 for (unsigned i
= 0; i
< verts_per_prim
; ++i
) {
1692 tmp
= LLVMBuildSub(builder
, tid
, LLVMConstInt(ctx
->ac
.i32
, verts_per_prim
- i
- 1, false),
1694 tmp
= ngg_gs_vertex_ptr(ctx
, tmp
);
1695 nggso
.vertices
[i
] = ac_build_gep0(&ctx
->ac
, tmp
, ctx
->ac
.i32_0
);
1698 build_streamout(ctx
, &nggso
);
1701 /* Write shader query data. */
1702 if (ctx
->screen
->use_ngg_streamout
) {
1703 tmp
= si_unpack_param(ctx
, ctx
->vs_state_bits
, 6, 1);
1704 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
1705 ac_build_ifcc(&ctx
->ac
, tmp
, 5109); /* if (STREAMOUT_QUERY_ENABLED) */
1706 unsigned num_query_comps
= sel
->so
.num_outputs
? 8 : 4;
1707 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
,
1708 LLVMConstInt(ctx
->ac
.i32
, num_query_comps
, false), "");
1709 ac_build_ifcc(&ctx
->ac
, tmp
, 5110);
1711 LLVMValueRef offset
;
1713 if (sel
->so
.num_outputs
)
1714 tmp
= LLVMBuildAnd(builder
, tmp
, LLVMConstInt(ctx
->ac
.i32
, 3, false), "");
1715 offset
= LLVMBuildNUWMul(builder
, tmp
, LLVMConstInt(ctx
->ac
.i32
, 32, false), "");
1716 if (sel
->so
.num_outputs
) {
1717 tmp
= LLVMBuildLShr(builder
, tid
, LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
1718 tmp
= LLVMBuildNUWMul(builder
, tmp
, LLVMConstInt(ctx
->ac
.i32
, 8, false), "");
1719 offset
= LLVMBuildAdd(builder
, offset
, tmp
, "");
1722 tmp
= LLVMBuildLoad(builder
, ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, tid
), "");
1723 LLVMValueRef args
[] = {
1724 tmp
, ngg_get_query_buf(ctx
),
1725 offset
, LLVMConstInt(ctx
->ac
.i32
, 16, false), /* soffset */
1726 ctx
->ac
.i32_0
, /* cachepolicy */
1728 ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.raw.buffer.atomic.add.i32", ctx
->ac
.i32
, args
, 5,
1731 ac_build_endif(&ctx
->ac
, 5110);
1732 ac_build_endif(&ctx
->ac
, 5109);
1735 /* Determine vertex liveness. */
1736 LLVMValueRef vertliveptr
= ac_build_alloca(&ctx
->ac
, ctx
->ac
.i1
, "vertexlive");
1738 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, num_emit_threads
, "");
1739 ac_build_ifcc(&ctx
->ac
, tmp
, 5120);
1741 for (unsigned i
= 0; i
< verts_per_prim
; ++i
) {
1742 const LLVMValueRef primidx
=
1743 LLVMBuildAdd(builder
, tid
, LLVMConstInt(ctx
->ac
.i32
, i
, false), "");
1746 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, primidx
, num_emit_threads
, "");
1747 ac_build_ifcc(&ctx
->ac
, tmp
, 5121 + i
);
1750 /* Load primitive liveness */
1751 tmp
= ngg_gs_vertex_ptr(ctx
, primidx
);
1752 tmp
= LLVMBuildLoad(builder
, ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, 0), "");
1753 const LLVMValueRef primlive
= LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
1755 tmp
= LLVMBuildLoad(builder
, vertliveptr
, "");
1756 tmp
= LLVMBuildOr(builder
, tmp
, primlive
, ""), LLVMBuildStore(builder
, tmp
, vertliveptr
);
1759 ac_build_endif(&ctx
->ac
, 5121 + i
);
1762 ac_build_endif(&ctx
->ac
, 5120);
1764 /* Inclusive scan addition across the current wave. */
1765 LLVMValueRef vertlive
= LLVMBuildLoad(builder
, vertliveptr
, "");
1766 struct ac_wg_scan vertlive_scan
= {};
1767 vertlive_scan
.op
= nir_op_iadd
;
1768 vertlive_scan
.enable_reduce
= true;
1769 vertlive_scan
.enable_exclusive
= true;
1770 vertlive_scan
.src
= vertlive
;
1771 vertlive_scan
.scratch
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, ctx
->ac
.i32_0
);
1772 vertlive_scan
.waveidx
= get_wave_id_in_tg(ctx
);
1773 vertlive_scan
.numwaves
= get_tgsize(ctx
);
1774 vertlive_scan
.maxwaves
= 8;
1776 ac_build_wg_scan(&ctx
->ac
, &vertlive_scan
);
1778 /* Skip all exports (including index exports) when possible. At least on
1779 * early gfx10 revisions this is also to avoid hangs.
1781 LLVMValueRef have_exports
=
1782 LLVMBuildICmp(builder
, LLVMIntNE
, vertlive_scan
.result_reduce
, ctx
->ac
.i32_0
, "");
1783 num_emit_threads
= LLVMBuildSelect(builder
, have_exports
, num_emit_threads
, ctx
->ac
.i32_0
, "");
1785 /* Allocate export space. Send this message as early as possible, to
1786 * hide the latency of the SQ <-> SPI roundtrip.
1788 * Note: We could consider compacting primitives for export as well.
1789 * PA processes 1 non-null prim / clock, but it fetches 4 DW of
1790 * prim data per clock and skips null primitives at no additional
1791 * cost. So compacting primitives can only be beneficial when
1792 * there are 4 or more contiguous null primitives in the export
1793 * (in the common case of single-dword prim exports).
1795 ac_build_sendmsg_gs_alloc_req(&ctx
->ac
, get_wave_id_in_tg(ctx
), vertlive_scan
.result_reduce
,
1798 /* Setup the reverse vertex compaction permutation. We re-use stream 1
1799 * of the primitive liveness flags, relying on the fact that each
1800 * threadgroup can have at most 256 threads. */
1801 ac_build_ifcc(&ctx
->ac
, vertlive
, 5130);
1803 tmp
= ngg_gs_vertex_ptr(ctx
, vertlive_scan
.result_exclusive
);
1804 tmp2
= LLVMBuildTrunc(builder
, tid
, ctx
->ac
.i8
, "");
1805 LLVMBuildStore(builder
, tmp2
, ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, 1));
1807 ac_build_endif(&ctx
->ac
, 5130);
1809 ac_build_s_barrier(&ctx
->ac
);
1811 /* Export primitive data */
1812 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, num_emit_threads
, "");
1813 ac_build_ifcc(&ctx
->ac
, tmp
, 5140);
1816 struct ac_ngg_prim prim
= {};
1817 prim
.num_vertices
= verts_per_prim
;
1819 tmp
= ngg_gs_vertex_ptr(ctx
, tid
);
1820 flags
= LLVMBuildLoad(builder
, ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, 0), "");
1821 prim
.isnull
= LLVMBuildNot(builder
, LLVMBuildTrunc(builder
, flags
, ctx
->ac
.i1
, ""), "");
1823 for (unsigned i
= 0; i
< verts_per_prim
; ++i
) {
1824 prim
.index
[i
] = LLVMBuildSub(builder
, vertlive_scan
.result_exclusive
,
1825 LLVMConstInt(ctx
->ac
.i32
, verts_per_prim
- i
- 1, false), "");
1826 prim
.edgeflag
[i
] = ctx
->ac
.i1false
;
1829 /* Geometry shaders output triangle strips, but NGG expects triangles. */
1830 if (verts_per_prim
== 3) {
1831 LLVMValueRef is_odd
= LLVMBuildLShr(builder
, flags
, ctx
->ac
.i8_1
, "");
1832 is_odd
= LLVMBuildTrunc(builder
, is_odd
, ctx
->ac
.i1
, "");
1833 LLVMValueRef flatshade_first
= LLVMBuildICmp(
1834 builder
, LLVMIntEQ
, si_unpack_param(ctx
, ctx
->vs_state_bits
, 4, 2), ctx
->ac
.i32_0
, "");
1836 ac_build_triangle_strip_indices_to_triangle(&ctx
->ac
, is_odd
, flatshade_first
, prim
.index
);
1839 ac_build_export_prim(&ctx
->ac
, &prim
);
1841 ac_build_endif(&ctx
->ac
, 5140);
1843 /* Export position and parameter data */
1844 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, vertlive_scan
.result_reduce
, "");
1845 ac_build_ifcc(&ctx
->ac
, tmp
, 5145);
1847 struct si_shader_output_values outputs
[PIPE_MAX_SHADER_OUTPUTS
];
1849 tmp
= ngg_gs_vertex_ptr(ctx
, tid
);
1850 tmp
= LLVMBuildLoad(builder
, ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, 1), "");
1851 tmp
= LLVMBuildZExt(builder
, tmp
, ctx
->ac
.i32
, "");
1852 const LLVMValueRef vertexptr
= ngg_gs_vertex_ptr(ctx
, tmp
);
1854 unsigned out_idx
= 0;
1855 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
1856 outputs
[i
].semantic_name
= info
->output_semantic_name
[i
];
1857 outputs
[i
].semantic_index
= info
->output_semantic_index
[i
];
1859 for (unsigned j
= 0; j
< 4; j
++, out_idx
++) {
1860 tmp
= ngg_gs_get_emit_output_ptr(ctx
, vertexptr
, out_idx
);
1861 tmp
= LLVMBuildLoad(builder
, tmp
, "");
1862 outputs
[i
].values
[j
] = ac_to_float(&ctx
->ac
, tmp
);
1863 outputs
[i
].vertex_stream
[j
] = (info
->output_streams
[i
] >> (2 * j
)) & 3;
1867 si_llvm_build_vs_exports(ctx
, outputs
, info
->num_outputs
);
1869 ac_build_endif(&ctx
->ac
, 5145);
1872 static void clamp_gsprims_to_esverts(unsigned *max_gsprims
, unsigned max_esverts
,
1873 unsigned min_verts_per_prim
, bool use_adjacency
)
1875 unsigned max_reuse
= max_esverts
- min_verts_per_prim
;
1878 *max_gsprims
= MIN2(*max_gsprims
, 1 + max_reuse
);
1882 * Determine subgroup information like maximum number of vertices and prims.
1884 * This happens before the shader is uploaded, since LDS relocations during
1885 * upload depend on the subgroup size.
1887 bool gfx10_ngg_calculate_subgroup_info(struct si_shader
*shader
)
1889 const struct si_shader_selector
*gs_sel
= shader
->selector
;
1890 const struct si_shader_selector
*es_sel
=
1891 shader
->previous_stage_sel
? shader
->previous_stage_sel
: gs_sel
;
1892 const enum pipe_shader_type gs_type
= gs_sel
->type
;
1893 const unsigned gs_num_invocations
= MAX2(gs_sel
->gs_num_invocations
, 1);
1894 const unsigned input_prim
= si_get_input_prim(gs_sel
);
1895 const bool use_adjacency
=
1896 input_prim
>= PIPE_PRIM_LINES_ADJACENCY
&& input_prim
<= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
;
1897 const unsigned max_verts_per_prim
= u_vertices_per_prim(input_prim
);
1898 const unsigned min_verts_per_prim
= gs_type
== PIPE_SHADER_GEOMETRY
? max_verts_per_prim
: 1;
1900 /* All these are in dwords: */
1901 /* We can't allow using the whole LDS, because GS waves compete with
1902 * other shader stages for LDS space.
1904 * TODO: We should really take the shader's internal LDS use into
1905 * account. The linker will fail if the size is greater than
1908 const unsigned max_lds_size
= 8 * 1024 - 768;
1909 const unsigned target_lds_size
= max_lds_size
;
1910 unsigned esvert_lds_size
= 0;
1911 unsigned gsprim_lds_size
= 0;
1913 /* All these are per subgroup: */
1914 bool max_vert_out_per_gs_instance
= false;
1915 unsigned max_gsprims_base
= 128; /* default prim group size clamp */
1916 unsigned max_esverts_base
= 128;
1918 if (shader
->key
.opt
.ngg_culling
& SI_NGG_CULL_GS_FAST_LAUNCH_TRI_LIST
) {
1919 max_gsprims_base
= 128 / 3;
1920 max_esverts_base
= max_gsprims_base
* 3;
1921 } else if (shader
->key
.opt
.ngg_culling
& SI_NGG_CULL_GS_FAST_LAUNCH_TRI_STRIP
) {
1922 max_gsprims_base
= 126;
1923 max_esverts_base
= 128;
1926 /* Hardware has the following non-natural restrictions on the value
1927 * of GE_CNTL.VERT_GRP_SIZE based on based on the primitive type of
1929 * - at most 252 for any line input primitive type
1930 * - at most 251 for any quad input primitive type
1931 * - at most 251 for triangle strips with adjacency (this happens to
1932 * be the natural limit for triangle *lists* with adjacency)
1934 max_esverts_base
= MIN2(max_esverts_base
, 251 + max_verts_per_prim
- 1);
1936 if (gs_type
== PIPE_SHADER_GEOMETRY
) {
1937 bool force_multi_cycling
= false;
1938 unsigned max_out_verts_per_gsprim
= gs_sel
->gs_max_out_vertices
* gs_num_invocations
;
1941 if (max_out_verts_per_gsprim
<= 256 && !force_multi_cycling
) {
1942 if (max_out_verts_per_gsprim
) {
1943 max_gsprims_base
= MIN2(max_gsprims_base
, 256 / max_out_verts_per_gsprim
);
1946 /* Use special multi-cycling mode in which each GS
1947 * instance gets its own subgroup. Does not work with
1949 max_vert_out_per_gs_instance
= true;
1950 max_gsprims_base
= 1;
1951 max_out_verts_per_gsprim
= gs_sel
->gs_max_out_vertices
;
1954 esvert_lds_size
= es_sel
->esgs_itemsize
/ 4;
1955 gsprim_lds_size
= (gs_sel
->gsvs_vertex_size
/ 4 + 1) * max_out_verts_per_gsprim
;
1957 if (gsprim_lds_size
> target_lds_size
&& !force_multi_cycling
) {
1958 if (gs_sel
->tess_turns_off_ngg
|| es_sel
->type
!= PIPE_SHADER_TESS_EVAL
) {
1959 force_multi_cycling
= true;
1960 goto retry_select_mode
;
1965 /* LDS size for passing data from ES to GS. */
1966 esvert_lds_size
= ngg_nogs_vertex_size(shader
);
1969 unsigned max_gsprims
= max_gsprims_base
;
1970 unsigned max_esverts
= max_esverts_base
;
1972 if (esvert_lds_size
)
1973 max_esverts
= MIN2(max_esverts
, target_lds_size
/ esvert_lds_size
);
1974 if (gsprim_lds_size
)
1975 max_gsprims
= MIN2(max_gsprims
, target_lds_size
/ gsprim_lds_size
);
1977 max_esverts
= MIN2(max_esverts
, max_gsprims
* max_verts_per_prim
);
1978 clamp_gsprims_to_esverts(&max_gsprims
, max_esverts
, min_verts_per_prim
, use_adjacency
);
1979 assert(max_esverts
>= max_verts_per_prim
&& max_gsprims
>= 1);
1981 if (esvert_lds_size
|| gsprim_lds_size
) {
1982 /* Now that we have a rough proportionality between esverts
1983 * and gsprims based on the primitive type, scale both of them
1984 * down simultaneously based on required LDS space.
1986 * We could be smarter about this if we knew how much vertex
1989 unsigned lds_total
= max_esverts
* esvert_lds_size
+ max_gsprims
* gsprim_lds_size
;
1990 if (lds_total
> target_lds_size
) {
1991 max_esverts
= max_esverts
* target_lds_size
/ lds_total
;
1992 max_gsprims
= max_gsprims
* target_lds_size
/ lds_total
;
1994 max_esverts
= MIN2(max_esverts
, max_gsprims
* max_verts_per_prim
);
1995 clamp_gsprims_to_esverts(&max_gsprims
, max_esverts
, min_verts_per_prim
, use_adjacency
);
1996 assert(max_esverts
>= max_verts_per_prim
&& max_gsprims
>= 1);
2000 /* Round up towards full wave sizes for better ALU utilization. */
2001 if (!max_vert_out_per_gs_instance
) {
2002 const unsigned wavesize
= gs_sel
->screen
->ge_wave_size
;
2003 unsigned orig_max_esverts
;
2004 unsigned orig_max_gsprims
;
2006 orig_max_esverts
= max_esverts
;
2007 orig_max_gsprims
= max_gsprims
;
2009 max_esverts
= align(max_esverts
, wavesize
);
2010 max_esverts
= MIN2(max_esverts
, max_esverts_base
);
2011 if (esvert_lds_size
)
2013 MIN2(max_esverts
, (max_lds_size
- max_gsprims
* gsprim_lds_size
) / esvert_lds_size
);
2014 max_esverts
= MIN2(max_esverts
, max_gsprims
* max_verts_per_prim
);
2016 max_gsprims
= align(max_gsprims
, wavesize
);
2017 max_gsprims
= MIN2(max_gsprims
, max_gsprims_base
);
2018 if (gsprim_lds_size
)
2020 MIN2(max_gsprims
, (max_lds_size
- max_esverts
* esvert_lds_size
) / gsprim_lds_size
);
2021 clamp_gsprims_to_esverts(&max_gsprims
, max_esverts
, min_verts_per_prim
, use_adjacency
);
2022 assert(max_esverts
>= max_verts_per_prim
&& max_gsprims
>= 1);
2023 } while (orig_max_esverts
!= max_esverts
|| orig_max_gsprims
!= max_gsprims
);
2026 /* Hardware restriction: minimum value of max_esverts */
2027 max_esverts
= MAX2(max_esverts
, 23 + max_verts_per_prim
);
2029 unsigned max_out_vertices
=
2030 max_vert_out_per_gs_instance
2031 ? gs_sel
->gs_max_out_vertices
2032 : gs_type
== PIPE_SHADER_GEOMETRY
2033 ? max_gsprims
* gs_num_invocations
* gs_sel
->gs_max_out_vertices
2035 assert(max_out_vertices
<= 256);
2037 unsigned prim_amp_factor
= 1;
2038 if (gs_type
== PIPE_SHADER_GEOMETRY
) {
2039 /* Number of output primitives per GS input primitive after
2041 prim_amp_factor
= gs_sel
->gs_max_out_vertices
;
2044 /* The GE only checks against the maximum number of ES verts after
2045 * allocating a full GS primitive. So we need to ensure that whenever
2046 * this check passes, there is enough space for a full primitive without
2049 shader
->ngg
.hw_max_esverts
= max_esverts
- max_verts_per_prim
+ 1;
2050 shader
->ngg
.max_gsprims
= max_gsprims
;
2051 shader
->ngg
.max_out_verts
= max_out_vertices
;
2052 shader
->ngg
.prim_amp_factor
= prim_amp_factor
;
2053 shader
->ngg
.max_vert_out_per_gs_instance
= max_vert_out_per_gs_instance
;
2055 shader
->gs_info
.esgs_ring_size
= 4 * max_esverts
* esvert_lds_size
;
2056 shader
->ngg
.ngg_emit_size
= max_gsprims
* gsprim_lds_size
;
2058 assert(shader
->ngg
.hw_max_esverts
>= 24); /* HW limitation */
2060 /* If asserts are disabled, we use the same conditions to return false */
2061 return max_esverts
>= max_verts_per_prim
&& max_gsprims
>= 1 &&
2062 max_out_vertices
<= 256 &&
2063 shader
->ngg
.hw_max_esverts
>= 24;