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"
31 #include "ac_llvm_cull.h"
33 static LLVMValueRef
get_wave_id_in_tg(struct si_shader_context
*ctx
)
35 return si_unpack_param(ctx
, ctx
->merged_wave_info
, 24, 4);
38 static LLVMValueRef
get_tgsize(struct si_shader_context
*ctx
)
40 return si_unpack_param(ctx
, ctx
->merged_wave_info
, 28, 4);
43 static LLVMValueRef
get_thread_id_in_tg(struct si_shader_context
*ctx
)
45 LLVMBuilderRef builder
= ctx
->ac
.builder
;
47 tmp
= LLVMBuildMul(builder
, get_wave_id_in_tg(ctx
),
48 LLVMConstInt(ctx
->ac
.i32
, ctx
->ac
.wave_size
, false), "");
49 return LLVMBuildAdd(builder
, tmp
, ac_get_thread_id(&ctx
->ac
), "");
52 static LLVMValueRef
ngg_get_vtx_cnt(struct si_shader_context
*ctx
)
54 return si_unpack_param(ctx
, ctx
->gs_tg_info
, 12, 9);
57 static LLVMValueRef
ngg_get_prim_cnt(struct si_shader_context
*ctx
)
59 return si_unpack_param(ctx
, ctx
->gs_tg_info
, 22, 9);
62 static LLVMValueRef
ngg_get_ordered_id(struct si_shader_context
*ctx
)
64 return si_unpack_param(ctx
, ctx
->gs_tg_info
, 0, 12);
67 static LLVMValueRef
ngg_get_query_buf(struct si_shader_context
*ctx
)
69 LLVMValueRef buf_ptr
= ac_get_arg(&ctx
->ac
, ctx
->rw_buffers
);
71 return ac_build_load_to_sgpr(&ctx
->ac
, buf_ptr
,
72 LLVMConstInt(ctx
->ac
.i32
, GFX10_GS_QUERY_BUF
, false));
75 static LLVMValueRef
ngg_get_initial_edgeflag(struct si_shader_context
*ctx
, unsigned index
)
77 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
79 tmp
= LLVMBuildLShr(ctx
->ac
.builder
,
80 ac_get_arg(&ctx
->ac
, ctx
->args
.gs_invocation_id
),
81 LLVMConstInt(ctx
->ac
.i32
, 8 + index
, false), "");
82 return LLVMBuildTrunc(ctx
->ac
.builder
, tmp
, ctx
->ac
.i1
, "");
84 return ctx
->ac
.i1false
;
88 * Return the number of vertices as a constant in \p num_vertices,
89 * and return a more precise value as LLVMValueRef from the function.
91 static LLVMValueRef
ngg_get_vertices_per_prim(struct si_shader_context
*ctx
,
92 unsigned *num_vertices
)
94 const struct si_shader_info
*info
= &ctx
->shader
->selector
->info
;
96 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
97 if (info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
]) {
98 /* Blits always use axis-aligned rectangles with 3 vertices. */
100 return LLVMConstInt(ctx
->ac
.i32
, 3, 0);
102 /* We always build up all three indices for the prim export
103 * independent of the primitive type. The additional garbage
104 * data shouldn't hurt. This number doesn't matter with
109 /* Extract OUTPRIM field. */
110 LLVMValueRef num
= si_unpack_param(ctx
, ctx
->vs_state_bits
, 2, 2);
111 return LLVMBuildAdd(ctx
->ac
.builder
, num
, ctx
->ac
.i32_1
, "");
114 assert(ctx
->type
== PIPE_SHADER_TESS_EVAL
);
116 if (info
->properties
[TGSI_PROPERTY_TES_POINT_MODE
])
118 else if (info
->properties
[TGSI_PROPERTY_TES_PRIM_MODE
] == PIPE_PRIM_LINES
)
123 return LLVMConstInt(ctx
->ac
.i32
, *num_vertices
, false);
127 bool gfx10_ngg_export_prim_early(struct si_shader
*shader
)
129 struct si_shader_selector
*sel
= shader
->selector
;
131 assert(shader
->key
.as_ngg
&& !shader
->key
.as_es
);
133 return sel
->type
!= PIPE_SHADER_GEOMETRY
&&
134 !sel
->info
.writes_edgeflag
;
137 void gfx10_ngg_build_sendmsg_gs_alloc_req(struct si_shader_context
*ctx
)
139 ac_build_sendmsg_gs_alloc_req(&ctx
->ac
, get_wave_id_in_tg(ctx
),
140 ngg_get_vtx_cnt(ctx
),
141 ngg_get_prim_cnt(ctx
));
144 void gfx10_ngg_build_export_prim(struct si_shader_context
*ctx
,
145 LLVMValueRef user_edgeflags
[3],
146 LLVMValueRef prim_passthrough
)
148 LLVMBuilderRef builder
= ctx
->ac
.builder
;
150 if (gfx10_is_ngg_passthrough(ctx
->shader
) ||
151 ctx
->shader
->key
.opt
.ngg_culling
) {
152 ac_build_ifcc(&ctx
->ac
, si_is_gs_thread(ctx
), 6001);
154 struct ac_ngg_prim prim
= {};
156 if (prim_passthrough
)
157 prim
.passthrough
= prim_passthrough
;
159 prim
.passthrough
= ac_get_arg(&ctx
->ac
, ctx
->gs_vtx01_offset
);
161 /* This is only used with NGG culling, which returns the NGG
162 * passthrough prim export encoding.
164 if (ctx
->shader
->selector
->info
.writes_edgeflag
) {
165 unsigned all_bits_no_edgeflags
= ~SI_NGG_PRIM_EDGE_FLAG_BITS
;
166 LLVMValueRef edgeflags
= LLVMConstInt(ctx
->ac
.i32
, all_bits_no_edgeflags
, 0);
168 unsigned num_vertices
;
169 ngg_get_vertices_per_prim(ctx
, &num_vertices
);
171 for (unsigned i
= 0; i
< num_vertices
; i
++) {
172 unsigned shift
= 9 + i
*10;
175 edge
= LLVMBuildLoad(builder
, user_edgeflags
[i
], "");
176 edge
= LLVMBuildZExt(builder
, edge
, ctx
->ac
.i32
, "");
177 edge
= LLVMBuildShl(builder
, edge
, LLVMConstInt(ctx
->ac
.i32
, shift
, 0), "");
178 edgeflags
= LLVMBuildOr(builder
, edgeflags
, edge
, "");
180 prim
.passthrough
= LLVMBuildAnd(builder
, prim
.passthrough
, edgeflags
, "");
183 ac_build_export_prim(&ctx
->ac
, &prim
);
185 ac_build_endif(&ctx
->ac
, 6001);
189 ac_build_ifcc(&ctx
->ac
, si_is_gs_thread(ctx
), 6001);
191 struct ac_ngg_prim prim
= {};
193 ngg_get_vertices_per_prim(ctx
, &prim
.num_vertices
);
195 prim
.isnull
= ctx
->ac
.i1false
;
196 prim
.index
[0] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 0, 16);
197 prim
.index
[1] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 16, 16);
198 prim
.index
[2] = si_unpack_param(ctx
, ctx
->gs_vtx23_offset
, 0, 16);
200 for (unsigned i
= 0; i
< prim
.num_vertices
; ++i
) {
201 prim
.edgeflag
[i
] = ngg_get_initial_edgeflag(ctx
, i
);
203 if (ctx
->shader
->selector
->info
.writes_edgeflag
) {
206 edge
= LLVMBuildLoad(ctx
->ac
.builder
, user_edgeflags
[i
], "");
207 edge
= LLVMBuildAnd(ctx
->ac
.builder
, prim
.edgeflag
[i
], edge
, "");
208 prim
.edgeflag
[i
] = edge
;
212 ac_build_export_prim(&ctx
->ac
, &prim
);
214 ac_build_endif(&ctx
->ac
, 6001);
217 static void build_streamout_vertex(struct si_shader_context
*ctx
,
218 LLVMValueRef
*so_buffer
, LLVMValueRef
*wg_offset_dw
,
219 unsigned stream
, LLVMValueRef offset_vtx
,
220 LLVMValueRef vertexptr
)
222 struct si_shader_info
*info
= &ctx
->shader
->selector
->info
;
223 struct pipe_stream_output_info
*so
= &ctx
->shader
->selector
->so
;
224 LLVMBuilderRef builder
= ctx
->ac
.builder
;
225 LLVMValueRef offset
[4] = {};
228 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
229 if (!wg_offset_dw
[buffer
])
232 tmp
= LLVMBuildMul(builder
, offset_vtx
,
233 LLVMConstInt(ctx
->ac
.i32
, so
->stride
[buffer
], false), "");
234 tmp
= LLVMBuildAdd(builder
, wg_offset_dw
[buffer
], tmp
, "");
235 offset
[buffer
] = LLVMBuildShl(builder
, tmp
, LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
238 for (unsigned i
= 0; i
< so
->num_outputs
; ++i
) {
239 if (so
->output
[i
].stream
!= stream
)
242 unsigned reg
= so
->output
[i
].register_index
;
243 struct si_shader_output_values out
;
244 out
.semantic_name
= info
->output_semantic_name
[reg
];
245 out
.semantic_index
= info
->output_semantic_index
[reg
];
247 for (unsigned comp
= 0; comp
< 4; comp
++) {
248 tmp
= ac_build_gep0(&ctx
->ac
, vertexptr
,
249 LLVMConstInt(ctx
->ac
.i32
, 4 * reg
+ comp
, false));
250 out
.values
[comp
] = LLVMBuildLoad(builder
, tmp
, "");
251 out
.vertex_stream
[comp
] =
252 (info
->output_streams
[reg
] >> (2 * comp
)) & 3;
255 si_llvm_streamout_store_output(ctx
, so_buffer
, offset
, &so
->output
[i
], &out
);
259 struct ngg_streamout
{
260 LLVMValueRef num_vertices
;
262 /* per-thread data */
263 LLVMValueRef prim_enable
[4]; /* i1 per stream */
264 LLVMValueRef vertices
[3]; /* [N x i32] addrspace(LDS)* */
267 LLVMValueRef emit
[4]; /* per-stream emitted primitives (only valid for used streams) */
271 * Build streamout logic.
275 * Writes number of emitted primitives to gs_ngg_scratch[4:8].
277 * Clobbers gs_ngg_scratch[8:].
279 static void build_streamout(struct si_shader_context
*ctx
,
280 struct ngg_streamout
*nggso
)
282 struct si_shader_info
*info
= &ctx
->shader
->selector
->info
;
283 struct pipe_stream_output_info
*so
= &ctx
->shader
->selector
->so
;
284 LLVMBuilderRef builder
= ctx
->ac
.builder
;
285 LLVMValueRef buf_ptr
= ac_get_arg(&ctx
->ac
, ctx
->rw_buffers
);
286 LLVMValueRef tid
= get_thread_id_in_tg(ctx
);
287 LLVMValueRef tmp
, tmp2
;
288 LLVMValueRef i32_2
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
289 LLVMValueRef i32_4
= LLVMConstInt(ctx
->ac
.i32
, 4, false);
290 LLVMValueRef i32_8
= LLVMConstInt(ctx
->ac
.i32
, 8, false);
291 LLVMValueRef so_buffer
[4] = {};
292 unsigned max_num_vertices
= 1 + (nggso
->vertices
[1] ? 1 : 0) +
293 (nggso
->vertices
[2] ? 1 : 0);
294 LLVMValueRef prim_stride_dw
[4] = {};
295 LLVMValueRef prim_stride_dw_vgpr
= LLVMGetUndef(ctx
->ac
.i32
);
296 int stream_for_buffer
[4] = { -1, -1, -1, -1 };
297 unsigned bufmask_for_stream
[4] = {};
298 bool isgs
= ctx
->type
== PIPE_SHADER_GEOMETRY
;
299 unsigned scratch_emit_base
= isgs
? 4 : 0;
300 LLVMValueRef scratch_emit_basev
= isgs
? i32_4
: ctx
->ac
.i32_0
;
301 unsigned scratch_offset_base
= isgs
? 8 : 4;
302 LLVMValueRef scratch_offset_basev
= isgs
? i32_8
: i32_4
;
304 ac_llvm_add_target_dep_function_attr(ctx
->main_fn
, "amdgpu-gds-size", 256);
306 /* Determine the mapping of streamout buffers to vertex streams. */
307 for (unsigned i
= 0; i
< so
->num_outputs
; ++i
) {
308 unsigned buf
= so
->output
[i
].output_buffer
;
309 unsigned stream
= so
->output
[i
].stream
;
310 assert(stream_for_buffer
[buf
] < 0 || stream_for_buffer
[buf
] == stream
);
311 stream_for_buffer
[buf
] = stream
;
312 bufmask_for_stream
[stream
] |= 1 << buf
;
315 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
316 if (stream_for_buffer
[buffer
] == -1)
319 assert(so
->stride
[buffer
]);
321 tmp
= LLVMConstInt(ctx
->ac
.i32
, so
->stride
[buffer
], false);
322 prim_stride_dw
[buffer
] = LLVMBuildMul(builder
, tmp
, nggso
->num_vertices
, "");
323 prim_stride_dw_vgpr
= ac_build_writelane(
324 &ctx
->ac
, prim_stride_dw_vgpr
, prim_stride_dw
[buffer
],
325 LLVMConstInt(ctx
->ac
.i32
, buffer
, false));
327 so_buffer
[buffer
] = ac_build_load_to_sgpr(
329 LLVMConstInt(ctx
->ac
.i32
, SI_VS_STREAMOUT_BUF0
+ buffer
, false));
332 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, get_wave_id_in_tg(ctx
), ctx
->ac
.i32_0
, "");
333 ac_build_ifcc(&ctx
->ac
, tmp
, 5200);
335 LLVMTypeRef gdsptr
= LLVMPointerType(ctx
->ac
.i32
, AC_ADDR_SPACE_GDS
);
336 LLVMValueRef gdsbase
= LLVMBuildIntToPtr(builder
, ctx
->ac
.i32_0
, gdsptr
, "");
338 /* Advance the streamout offsets in GDS. */
339 LLVMValueRef offsets_vgpr
= ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i32
, "");
340 LLVMValueRef generated_by_stream_vgpr
= ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i32
, "");
342 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, ac_get_thread_id(&ctx
->ac
), i32_4
, "");
343 ac_build_ifcc(&ctx
->ac
, tmp
, 5210);
346 tmp
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, tid
);
347 tmp
= LLVMBuildLoad(builder
, tmp
, "");
349 tmp
= ac_build_writelane(&ctx
->ac
, ctx
->ac
.i32_0
,
350 ngg_get_prim_cnt(ctx
), ctx
->ac
.i32_0
);
352 LLVMBuildStore(builder
, tmp
, generated_by_stream_vgpr
);
355 int unused_stream
= -1;
356 for (unsigned stream
= 0; stream
< 4; ++stream
) {
357 if (!info
->num_stream_output_components
[stream
]) {
358 unused_stream
= stream
;
362 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
363 if (stream_for_buffer
[buffer
] >= 0) {
364 swizzle
[buffer
] = stream_for_buffer
[buffer
];
366 assert(unused_stream
>= 0);
367 swizzle
[buffer
] = unused_stream
;
371 tmp
= ac_build_quad_swizzle(&ctx
->ac
, tmp
,
372 swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
373 tmp
= LLVMBuildMul(builder
, tmp
, prim_stride_dw_vgpr
, "");
375 LLVMValueRef args
[] = {
376 LLVMBuildIntToPtr(builder
, ngg_get_ordered_id(ctx
), gdsptr
, ""),
378 ctx
->ac
.i32_0
, // ordering
379 ctx
->ac
.i32_0
, // scope
380 ctx
->ac
.i1false
, // isVolatile
381 LLVMConstInt(ctx
->ac
.i32
, 4 << 24, false), // OA index
382 ctx
->ac
.i1true
, // wave release
383 ctx
->ac
.i1true
, // wave done
385 tmp
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.ds.ordered.add",
386 ctx
->ac
.i32
, args
, ARRAY_SIZE(args
), 0);
388 /* Keep offsets in a VGPR for quick retrieval via readlane by
389 * the first wave for bounds checking, and also store in LDS
390 * for retrieval by all waves later. */
391 LLVMBuildStore(builder
, tmp
, offsets_vgpr
);
393 tmp2
= LLVMBuildAdd(builder
, ac_get_thread_id(&ctx
->ac
),
394 scratch_offset_basev
, "");
395 tmp2
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, tmp2
);
396 LLVMBuildStore(builder
, tmp
, tmp2
);
398 ac_build_endif(&ctx
->ac
, 5210);
400 /* Determine the max emit per buffer. This is done via the SALU, in part
401 * because LLVM can't generate divide-by-multiply if we try to do this
402 * via VALU with one lane per buffer.
404 LLVMValueRef max_emit
[4] = {};
405 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
406 if (stream_for_buffer
[buffer
] == -1)
409 LLVMValueRef bufsize_dw
=
410 LLVMBuildLShr(builder
,
411 LLVMBuildExtractElement(builder
, so_buffer
[buffer
], i32_2
, ""),
414 tmp
= LLVMBuildLoad(builder
, offsets_vgpr
, "");
415 LLVMValueRef offset_dw
=
416 ac_build_readlane(&ctx
->ac
, tmp
,
417 LLVMConstInt(ctx
->ac
.i32
, buffer
, false));
419 tmp
= LLVMBuildSub(builder
, bufsize_dw
, offset_dw
, "");
420 tmp
= LLVMBuildUDiv(builder
, tmp
, prim_stride_dw
[buffer
], "");
422 tmp2
= LLVMBuildICmp(builder
, LLVMIntULT
, bufsize_dw
, offset_dw
, "");
423 max_emit
[buffer
] = LLVMBuildSelect(builder
, tmp2
, ctx
->ac
.i32_0
, tmp
, "");
426 /* Determine the number of emitted primitives per stream and fixup the
427 * GDS counter if necessary.
429 * This is complicated by the fact that a single stream can emit to
430 * multiple buffers (but luckily not vice versa).
432 LLVMValueRef emit_vgpr
= ctx
->ac
.i32_0
;
434 for (unsigned stream
= 0; stream
< 4; ++stream
) {
435 if (!info
->num_stream_output_components
[stream
])
438 tmp
= LLVMBuildLoad(builder
, generated_by_stream_vgpr
, "");
439 LLVMValueRef generated
=
440 ac_build_readlane(&ctx
->ac
, tmp
,
441 LLVMConstInt(ctx
->ac
.i32
, stream
, false));
443 LLVMValueRef emit
= generated
;
444 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
445 if (stream_for_buffer
[buffer
] == stream
)
446 emit
= ac_build_umin(&ctx
->ac
, emit
, max_emit
[buffer
]);
449 emit_vgpr
= ac_build_writelane(&ctx
->ac
, emit_vgpr
, emit
,
450 LLVMConstInt(ctx
->ac
.i32
, stream
, false));
452 /* Fixup the offset using a plain GDS atomic if we overflowed. */
453 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, emit
, generated
, "");
454 ac_build_ifcc(&ctx
->ac
, tmp
, 5221); /* scalar branch */
455 tmp
= LLVMBuildLShr(builder
,
456 LLVMConstInt(ctx
->ac
.i32
, bufmask_for_stream
[stream
], false),
457 ac_get_thread_id(&ctx
->ac
), "");
458 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
459 ac_build_ifcc(&ctx
->ac
, tmp
, 5222);
461 tmp
= LLVMBuildSub(builder
, generated
, emit
, "");
462 tmp
= LLVMBuildMul(builder
, tmp
, prim_stride_dw_vgpr
, "");
463 tmp2
= LLVMBuildGEP(builder
, gdsbase
, &tid
, 1, "");
464 LLVMBuildAtomicRMW(builder
, LLVMAtomicRMWBinOpSub
, tmp2
, tmp
,
465 LLVMAtomicOrderingMonotonic
, false);
467 ac_build_endif(&ctx
->ac
, 5222);
468 ac_build_endif(&ctx
->ac
, 5221);
471 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, ac_get_thread_id(&ctx
->ac
), i32_4
, "");
472 ac_build_ifcc(&ctx
->ac
, tmp
, 5225);
474 tmp
= LLVMBuildAdd(builder
, ac_get_thread_id(&ctx
->ac
),
475 scratch_emit_basev
, "");
476 tmp
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, tmp
);
477 LLVMBuildStore(builder
, emit_vgpr
, tmp
);
479 ac_build_endif(&ctx
->ac
, 5225);
481 ac_build_endif(&ctx
->ac
, 5200);
483 /* Determine the workgroup-relative per-thread / primitive offset into
484 * the streamout buffers */
485 struct ac_wg_scan primemit_scan
[4] = {};
488 for (unsigned stream
= 0; stream
< 4; ++stream
) {
489 if (!info
->num_stream_output_components
[stream
])
492 primemit_scan
[stream
].enable_exclusive
= true;
493 primemit_scan
[stream
].op
= nir_op_iadd
;
494 primemit_scan
[stream
].src
= nggso
->prim_enable
[stream
];
495 primemit_scan
[stream
].scratch
=
496 ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
,
497 LLVMConstInt(ctx
->ac
.i32
, 12 + 8 * stream
, false));
498 primemit_scan
[stream
].waveidx
= get_wave_id_in_tg(ctx
);
499 primemit_scan
[stream
].numwaves
= get_tgsize(ctx
);
500 primemit_scan
[stream
].maxwaves
= 8;
501 ac_build_wg_scan_top(&ctx
->ac
, &primemit_scan
[stream
]);
505 ac_build_s_barrier(&ctx
->ac
);
507 /* Fetch the per-buffer offsets and per-stream emit counts in all waves. */
508 LLVMValueRef wgoffset_dw
[4] = {};
511 LLVMValueRef scratch_vgpr
;
513 tmp
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, ac_get_thread_id(&ctx
->ac
));
514 scratch_vgpr
= LLVMBuildLoad(builder
, tmp
, "");
516 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
517 if (stream_for_buffer
[buffer
] >= 0) {
518 wgoffset_dw
[buffer
] = ac_build_readlane(
519 &ctx
->ac
, scratch_vgpr
,
520 LLVMConstInt(ctx
->ac
.i32
, scratch_offset_base
+ buffer
, false));
524 for (unsigned stream
= 0; stream
< 4; ++stream
) {
525 if (info
->num_stream_output_components
[stream
]) {
526 nggso
->emit
[stream
] = ac_build_readlane(
527 &ctx
->ac
, scratch_vgpr
,
528 LLVMConstInt(ctx
->ac
.i32
, scratch_emit_base
+ stream
, false));
533 /* Write out primitive data */
534 for (unsigned stream
= 0; stream
< 4; ++stream
) {
535 if (!info
->num_stream_output_components
[stream
])
539 ac_build_wg_scan_bottom(&ctx
->ac
, &primemit_scan
[stream
]);
541 primemit_scan
[stream
].result_exclusive
= tid
;
544 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
,
545 primemit_scan
[stream
].result_exclusive
,
546 nggso
->emit
[stream
], "");
547 tmp
= LLVMBuildAnd(builder
, tmp
, nggso
->prim_enable
[stream
], "");
548 ac_build_ifcc(&ctx
->ac
, tmp
, 5240);
550 LLVMValueRef offset_vtx
=
551 LLVMBuildMul(builder
, primemit_scan
[stream
].result_exclusive
,
552 nggso
->num_vertices
, "");
554 for (unsigned i
= 0; i
< max_num_vertices
; ++i
) {
555 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
,
556 LLVMConstInt(ctx
->ac
.i32
, i
, false),
557 nggso
->num_vertices
, "");
558 ac_build_ifcc(&ctx
->ac
, tmp
, 5241);
559 build_streamout_vertex(ctx
, so_buffer
, wgoffset_dw
,
560 stream
, offset_vtx
, nggso
->vertices
[i
]);
561 ac_build_endif(&ctx
->ac
, 5241);
562 offset_vtx
= LLVMBuildAdd(builder
, offset_vtx
, ctx
->ac
.i32_1
, "");
565 ac_build_endif(&ctx
->ac
, 5240);
569 /* LDS layout of ES vertex data for NGG culling. */
571 /* Byte 0: Boolean ES thread accepted (unculled) flag, and later the old
572 * ES thread ID. After vertex compaction, compacted ES threads
573 * store the old thread ID here to copy input VGPRs from uncompacted
575 * Byte 1: New ES thread ID, loaded by GS to prepare the prim export value.
576 * Byte 2: TES rel patch ID
579 lds_byte0_accept_flag
= 0,
580 lds_byte0_old_thread_id
= 0,
581 lds_byte1_new_thread_id
,
582 lds_byte2_tes_rel_patch_id
,
585 lds_packed_data
= 0, /* lds_byteN_... */
595 lds_instance_id
, /* optional */
597 lds_tes_u
= lds_vertex_id
,
598 lds_tes_v
= lds_instance_id
,
599 lds_tes_patch_id
, /* optional */
602 static LLVMValueRef
si_build_gep_i8(struct si_shader_context
*ctx
,
603 LLVMValueRef ptr
, unsigned byte_index
)
605 assert(byte_index
< 4);
606 LLVMTypeRef pi8
= LLVMPointerType(ctx
->ac
.i8
, AC_ADDR_SPACE_LDS
);
607 LLVMValueRef index
= LLVMConstInt(ctx
->ac
.i32
, byte_index
, 0);
609 return LLVMBuildGEP(ctx
->ac
.builder
,
610 LLVMBuildPointerCast(ctx
->ac
.builder
, ptr
, pi8
, ""),
614 static unsigned ngg_nogs_vertex_size(struct si_shader
*shader
)
616 unsigned lds_vertex_size
= 0;
618 /* The edgeflag is always stored in the last element that's also
619 * used for padding to reduce LDS bank conflicts. */
620 if (shader
->selector
->so
.num_outputs
)
621 lds_vertex_size
= 4 * shader
->selector
->info
.num_outputs
+ 1;
622 if (shader
->selector
->info
.writes_edgeflag
)
623 lds_vertex_size
= MAX2(lds_vertex_size
, 1);
625 /* LDS size for passing data from GS to ES.
626 * GS stores Primitive IDs into LDS at the address corresponding
627 * to the ES thread of the provoking vertex. All ES threads
628 * load and export PrimitiveID for their thread.
630 if (shader
->selector
->type
== PIPE_SHADER_VERTEX
&&
631 shader
->key
.mono
.u
.vs_export_prim_id
)
632 lds_vertex_size
= MAX2(lds_vertex_size
, 1);
634 if (shader
->key
.opt
.ngg_culling
) {
635 if (shader
->selector
->type
== PIPE_SHADER_VERTEX
) {
636 STATIC_ASSERT(lds_instance_id
+ 1 == 9);
637 lds_vertex_size
= MAX2(lds_vertex_size
, 9);
639 assert(shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
);
641 if (shader
->selector
->info
.uses_primid
||
642 shader
->key
.mono
.u
.vs_export_prim_id
) {
643 STATIC_ASSERT(lds_tes_patch_id
+ 2 == 11);
644 lds_vertex_size
= MAX2(lds_vertex_size
, 11);
646 STATIC_ASSERT(lds_tes_v
+ 1 == 9);
647 lds_vertex_size
= MAX2(lds_vertex_size
, 9);
652 return lds_vertex_size
;
656 * Returns an `[N x i32] addrspace(LDS)*` pointing at contiguous LDS storage
657 * for the vertex outputs.
659 static LLVMValueRef
ngg_nogs_vertex_ptr(struct si_shader_context
*ctx
,
662 /* The extra dword is used to avoid LDS bank conflicts. */
663 unsigned vertex_size
= ngg_nogs_vertex_size(ctx
->shader
);
664 LLVMTypeRef ai32
= LLVMArrayType(ctx
->ac
.i32
, vertex_size
);
665 LLVMTypeRef pai32
= LLVMPointerType(ai32
, AC_ADDR_SPACE_LDS
);
666 LLVMValueRef tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->esgs_ring
, pai32
, "");
667 return LLVMBuildGEP(ctx
->ac
.builder
, tmp
, &vtxid
, 1, "");
670 static LLVMValueRef
si_insert_input_v4i32(struct si_shader_context
*ctx
,
671 LLVMValueRef ret
, struct ac_arg param
,
672 unsigned return_index
)
674 LLVMValueRef v
= ac_get_arg(&ctx
->ac
, param
);
676 for (unsigned i
= 0; i
< 4; i
++) {
677 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
,
678 ac_llvm_extract_elem(&ctx
->ac
, v
, i
),
679 return_index
+ i
, "");
684 static void load_bitmasks_2x64(struct si_shader_context
*ctx
,
685 LLVMValueRef lds_ptr
, unsigned dw_offset
,
686 LLVMValueRef mask
[2], LLVMValueRef
*total_bitcount
)
688 LLVMBuilderRef builder
= ctx
->ac
.builder
;
689 LLVMValueRef ptr64
= LLVMBuildPointerCast(builder
, lds_ptr
,
690 LLVMPointerType(LLVMArrayType(ctx
->ac
.i64
, 2),
691 AC_ADDR_SPACE_LDS
), "");
692 for (unsigned i
= 0; i
< 2; i
++) {
693 LLVMValueRef index
= LLVMConstInt(ctx
->ac
.i32
, dw_offset
/ 2 + i
, 0);
694 mask
[i
] = LLVMBuildLoad(builder
, ac_build_gep0(&ctx
->ac
, ptr64
, index
), "");
697 /* We get better code if we don't use the 128-bit bitcount. */
698 *total_bitcount
= LLVMBuildAdd(builder
, ac_build_bit_count(&ctx
->ac
, mask
[0]),
699 ac_build_bit_count(&ctx
->ac
, mask
[1]), "");
703 * Given a total thread count, update total and per-wave thread counts in input SGPRs
704 * and return the per-wave thread count.
706 * \param new_num_threads Total thread count on the input, per-wave thread count on the output.
707 * \param tg_info tg_info SGPR value
708 * \param tg_info_num_bits the bit size of thread count field in tg_info
709 * \param tg_info_shift the bit offset of the thread count field in tg_info
710 * \param wave_info merged_wave_info SGPR value
711 * \param wave_info_num_bits the bit size of thread count field in merged_wave_info
712 * \param wave_info_shift the bit offset of the thread count field in merged_wave_info
714 static void update_thread_counts(struct si_shader_context
*ctx
,
715 LLVMValueRef
*new_num_threads
,
716 LLVMValueRef
*tg_info
,
717 unsigned tg_info_num_bits
,
718 unsigned tg_info_shift
,
719 LLVMValueRef
*wave_info
,
720 unsigned wave_info_num_bits
,
721 unsigned wave_info_shift
)
723 LLVMBuilderRef builder
= ctx
->ac
.builder
;
725 /* Update the total thread count. */
726 unsigned tg_info_mask
= ~(u_bit_consecutive(0, tg_info_num_bits
) << tg_info_shift
);
727 *tg_info
= LLVMBuildAnd(builder
, *tg_info
,
728 LLVMConstInt(ctx
->ac
.i32
, tg_info_mask
, 0), "");
729 *tg_info
= LLVMBuildOr(builder
, *tg_info
,
730 LLVMBuildShl(builder
, *new_num_threads
,
731 LLVMConstInt(ctx
->ac
.i32
, tg_info_shift
, 0), ""), "");
733 /* Update the per-wave thread count. */
734 LLVMValueRef prev_threads
= LLVMBuildMul(builder
, get_wave_id_in_tg(ctx
),
735 LLVMConstInt(ctx
->ac
.i32
, ctx
->ac
.wave_size
, 0), "");
736 *new_num_threads
= LLVMBuildSub(builder
, *new_num_threads
, prev_threads
, "");
737 *new_num_threads
= ac_build_imax(&ctx
->ac
, *new_num_threads
, ctx
->ac
.i32_0
);
738 *new_num_threads
= ac_build_imin(&ctx
->ac
, *new_num_threads
,
739 LLVMConstInt(ctx
->ac
.i32
, ctx
->ac
.wave_size
, 0));
740 unsigned wave_info_mask
= ~(u_bit_consecutive(0, wave_info_num_bits
) << wave_info_shift
);
741 *wave_info
= LLVMBuildAnd(builder
, *wave_info
,
742 LLVMConstInt(ctx
->ac
.i32
, wave_info_mask
, 0), "");
743 *wave_info
= LLVMBuildOr(builder
, *wave_info
,
744 LLVMBuildShl(builder
, *new_num_threads
,
745 LLVMConstInt(ctx
->ac
.i32
, wave_info_shift
, 0), ""), "");
749 * Cull primitives for NGG VS or TES, then compact vertices, which happens
750 * before the VS or TES main function. Return values for the main function.
751 * Also return the position, which is passed to the shader as an input,
752 * so that we don't compute it twice.
754 void gfx10_emit_ngg_culling_epilogue_4x_wave32(struct ac_shader_abi
*abi
,
755 unsigned max_outputs
,
758 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
759 struct si_shader
*shader
= ctx
->shader
;
760 struct si_shader_selector
*sel
= shader
->selector
;
761 struct si_shader_info
*info
= &sel
->info
;
762 LLVMBuilderRef builder
= ctx
->ac
.builder
;
764 assert(shader
->key
.opt
.ngg_culling
);
765 assert(shader
->key
.as_ngg
);
766 assert(sel
->type
== PIPE_SHADER_VERTEX
||
767 (sel
->type
== PIPE_SHADER_TESS_EVAL
&& !shader
->key
.as_es
));
769 LLVMValueRef position
[4] = {};
770 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
771 switch (info
->output_semantic_name
[i
]) {
772 case TGSI_SEMANTIC_POSITION
:
773 for (unsigned j
= 0; j
< 4; j
++) {
774 position
[j
] = LLVMBuildLoad(ctx
->ac
.builder
,
775 addrs
[4 * i
+ j
], "");
782 /* Store Position.XYZW into LDS. */
783 LLVMValueRef es_vtxptr
= ngg_nogs_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
));
784 for (unsigned chan
= 0; chan
< 4; chan
++) {
785 LLVMBuildStore(builder
, ac_to_integer(&ctx
->ac
, position
[chan
]),
786 ac_build_gep0(&ctx
->ac
, es_vtxptr
,
787 LLVMConstInt(ctx
->ac
.i32
, lds_pos_x
+ chan
, 0)));
789 /* Store Position.XY / W into LDS. */
790 for (unsigned chan
= 0; chan
< 2; chan
++) {
791 LLVMValueRef val
= ac_build_fdiv(&ctx
->ac
, position
[chan
], position
[3]);
792 LLVMBuildStore(builder
, ac_to_integer(&ctx
->ac
, val
),
793 ac_build_gep0(&ctx
->ac
, es_vtxptr
,
794 LLVMConstInt(ctx
->ac
.i32
, lds_pos_x_div_w
+ chan
, 0)));
797 /* Store VertexID and InstanceID. ES threads will have to load them
798 * from LDS after vertex compaction and use them instead of their own
801 bool uses_instance_id
= false;
802 bool uses_tes_prim_id
= false;
803 LLVMValueRef packed_data
= ctx
->ac
.i32_0
;
805 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
806 uses_instance_id
= sel
->info
.uses_instanceid
||
807 shader
->key
.part
.vs
.prolog
.instance_divisor_is_one
||
808 shader
->key
.part
.vs
.prolog
.instance_divisor_is_fetched
;
810 LLVMBuildStore(builder
, ctx
->abi
.vertex_id
,
811 ac_build_gep0(&ctx
->ac
, es_vtxptr
,
812 LLVMConstInt(ctx
->ac
.i32
, lds_vertex_id
, 0)));
813 if (uses_instance_id
) {
814 LLVMBuildStore(builder
, ctx
->abi
.instance_id
,
815 ac_build_gep0(&ctx
->ac
, es_vtxptr
,
816 LLVMConstInt(ctx
->ac
.i32
, lds_instance_id
, 0)));
819 uses_tes_prim_id
= sel
->info
.uses_primid
||
820 shader
->key
.mono
.u
.vs_export_prim_id
;
822 assert(ctx
->type
== PIPE_SHADER_TESS_EVAL
);
823 LLVMBuildStore(builder
, ac_to_integer(&ctx
->ac
, ac_get_arg(&ctx
->ac
, ctx
->tes_u
)),
824 ac_build_gep0(&ctx
->ac
, es_vtxptr
,
825 LLVMConstInt(ctx
->ac
.i32
, lds_tes_u
, 0)));
826 LLVMBuildStore(builder
, ac_to_integer(&ctx
->ac
, ac_get_arg(&ctx
->ac
, ctx
->tes_v
)),
827 ac_build_gep0(&ctx
->ac
, es_vtxptr
,
828 LLVMConstInt(ctx
->ac
.i32
, lds_tes_v
, 0)));
829 packed_data
= LLVMBuildShl(builder
, ac_get_arg(&ctx
->ac
, ctx
->tes_rel_patch_id
),
830 LLVMConstInt(ctx
->ac
.i32
, lds_byte2_tes_rel_patch_id
* 8, 0), "");
831 if (uses_tes_prim_id
) {
832 LLVMBuildStore(builder
, ac_get_arg(&ctx
->ac
, ctx
->args
.tes_patch_id
),
833 ac_build_gep0(&ctx
->ac
, es_vtxptr
,
834 LLVMConstInt(ctx
->ac
.i32
, lds_tes_patch_id
, 0)));
837 /* Initialize the packed data. */
838 LLVMBuildStore(builder
, packed_data
,
839 ac_build_gep0(&ctx
->ac
, es_vtxptr
,
840 LLVMConstInt(ctx
->ac
.i32
, lds_packed_data
, 0)));
841 ac_build_endif(&ctx
->ac
, ctx
->merged_wrap_if_label
);
843 LLVMValueRef tid
= ac_get_thread_id(&ctx
->ac
);
845 /* Initialize the last 3 gs_ngg_scratch dwords to 0, because we may have less
846 * than 4 waves, but we always read all 4 values. This is where the thread
847 * bitmasks of unculled threads will be stored.
849 * gs_ngg_scratch layout: esmask[0..3]
851 ac_build_ifcc(&ctx
->ac
,
852 LLVMBuildICmp(builder
, LLVMIntULT
, get_thread_id_in_tg(ctx
),
853 LLVMConstInt(ctx
->ac
.i32
, 3, 0), ""), 16101);
855 LLVMValueRef index
= LLVMBuildAdd(builder
, tid
, ctx
->ac
.i32_1
, "");
856 LLVMBuildStore(builder
, ctx
->ac
.i32_0
,
857 ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, index
));
859 ac_build_endif(&ctx
->ac
, 16101);
860 ac_build_s_barrier(&ctx
->ac
);
862 /* The hardware requires that there are no holes between unculled vertices,
863 * which means we have to pack ES threads, i.e. reduce the ES thread count
864 * and move ES input VGPRs to lower threads. The upside is that varyings
865 * are only fetched and computed for unculled vertices.
867 * Vertex compaction in GS threads:
869 * Part 1: Compute the surviving vertex mask in GS threads:
870 * - Compute 4 32-bit surviving vertex masks in LDS. (max 4 waves)
871 * - In GS, notify ES threads whether the vertex survived.
873 * - ES threads will create the mask and store it in LDS.
875 * - Each GS thread loads the vertex masks from LDS.
877 * Part 2: Compact ES threads in GS threads:
878 * - Compute the prefix sum for all 3 vertices from the masks. These are the new
879 * thread IDs for each vertex within the primitive.
880 * - Write the value of the old thread ID into the LDS address of the new thread ID.
881 * The ES thread will load the old thread ID and use it to load the position, VertexID,
883 * - Update vertex indices and null flag in the GS input VGPRs.
886 * Part 3: Update inputs GPRs
887 * - For all waves, update per-wave thread counts in input SGPRs.
888 * - In ES threads, update the ES input VGPRs (VertexID, InstanceID, TES inputs).
891 LLVMValueRef vtxindex
[3];
892 if (shader
->key
.opt
.ngg_culling
& SI_NGG_CULL_GS_FAST_LAUNCH_ALL
) {
893 /* For the GS fast launch, the VS prologs simply puts the Vertex IDs
896 vtxindex
[0] = ac_get_arg(&ctx
->ac
, ctx
->gs_vtx01_offset
);
897 vtxindex
[1] = ac_get_arg(&ctx
->ac
, ctx
->gs_vtx23_offset
);
898 vtxindex
[2] = ac_get_arg(&ctx
->ac
, ctx
->gs_vtx45_offset
);
900 vtxindex
[0] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 0, 16);
901 vtxindex
[1] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 16, 16);
902 vtxindex
[2] = si_unpack_param(ctx
, ctx
->gs_vtx23_offset
, 0, 16);
904 LLVMValueRef gs_vtxptr
[] = {
905 ngg_nogs_vertex_ptr(ctx
, vtxindex
[0]),
906 ngg_nogs_vertex_ptr(ctx
, vtxindex
[1]),
907 ngg_nogs_vertex_ptr(ctx
, vtxindex
[2]),
909 es_vtxptr
= ngg_nogs_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
));
911 LLVMValueRef gs_accepted
= ac_build_alloca(&ctx
->ac
, ctx
->ac
.i32
, "");
913 /* Do culling in GS threads. */
914 ac_build_ifcc(&ctx
->ac
, si_is_gs_thread(ctx
), 16002);
916 /* Load positions. */
917 LLVMValueRef pos
[3][4] = {};
918 for (unsigned vtx
= 0; vtx
< 3; vtx
++) {
919 for (unsigned chan
= 0; chan
< 4; chan
++) {
921 if (chan
== 0 || chan
== 1)
922 index
= lds_pos_x_div_w
+ chan
;
928 LLVMValueRef addr
= ac_build_gep0(&ctx
->ac
, gs_vtxptr
[vtx
],
929 LLVMConstInt(ctx
->ac
.i32
, index
, 0));
930 pos
[vtx
][chan
] = LLVMBuildLoad(builder
, addr
, "");
931 pos
[vtx
][chan
] = ac_to_float(&ctx
->ac
, pos
[vtx
][chan
]);
935 /* Load the viewport state for small prim culling. */
936 LLVMValueRef vp
= ac_build_load_invariant(&ctx
->ac
,
937 ac_get_arg(&ctx
->ac
, ctx
->small_prim_cull_info
),
939 vp
= LLVMBuildBitCast(builder
, vp
, ctx
->ac
.v4f32
, "");
940 LLVMValueRef vp_scale
[2], vp_translate
[2];
941 vp_scale
[0] = ac_llvm_extract_elem(&ctx
->ac
, vp
, 0);
942 vp_scale
[1] = ac_llvm_extract_elem(&ctx
->ac
, vp
, 1);
943 vp_translate
[0] = ac_llvm_extract_elem(&ctx
->ac
, vp
, 2);
944 vp_translate
[1] = ac_llvm_extract_elem(&ctx
->ac
, vp
, 3);
946 /* Get the small prim filter precision. */
947 LLVMValueRef small_prim_precision
= si_unpack_param(ctx
, ctx
->vs_state_bits
, 7, 4);
948 small_prim_precision
= LLVMBuildOr(builder
, small_prim_precision
,
949 LLVMConstInt(ctx
->ac
.i32
, 0x70, 0), "");
950 small_prim_precision
= LLVMBuildShl(builder
, small_prim_precision
,
951 LLVMConstInt(ctx
->ac
.i32
, 23, 0), "");
952 small_prim_precision
= LLVMBuildBitCast(builder
, small_prim_precision
, ctx
->ac
.f32
, "");
954 /* Execute culling code. */
955 struct ac_cull_options options
= {};
956 options
.cull_front
= shader
->key
.opt
.ngg_culling
& SI_NGG_CULL_FRONT_FACE
;
957 options
.cull_back
= shader
->key
.opt
.ngg_culling
& SI_NGG_CULL_BACK_FACE
;
958 options
.cull_view_xy
= shader
->key
.opt
.ngg_culling
& SI_NGG_CULL_VIEW_SMALLPRIMS
;
959 options
.cull_small_prims
= options
.cull_view_xy
;
960 options
.cull_zero_area
= options
.cull_front
|| options
.cull_back
;
961 options
.cull_w
= true;
963 /* Tell ES threads whether their vertex survived. */
964 ac_build_ifcc(&ctx
->ac
, ac_cull_triangle(&ctx
->ac
, pos
, ctx
->ac
.i1true
,
965 vp_scale
, vp_translate
,
966 small_prim_precision
, &options
), 16003);
968 LLVMBuildStore(builder
, ctx
->ac
.i32_1
, gs_accepted
);
969 for (unsigned vtx
= 0; vtx
< 3; vtx
++) {
970 LLVMBuildStore(builder
, ctx
->ac
.i8_1
,
971 si_build_gep_i8(ctx
, gs_vtxptr
[vtx
], lds_byte0_accept_flag
));
974 ac_build_endif(&ctx
->ac
, 16003);
976 ac_build_endif(&ctx
->ac
, 16002);
977 ac_build_s_barrier(&ctx
->ac
);
979 gs_accepted
= LLVMBuildLoad(builder
, gs_accepted
, "");
981 LLVMValueRef es_accepted
= ac_build_alloca(&ctx
->ac
, ctx
->ac
.i1
, "");
983 /* Convert the per-vertex flag to a thread bitmask in ES threads and store it in LDS. */
984 ac_build_ifcc(&ctx
->ac
, si_is_es_thread(ctx
), 16007);
986 LLVMValueRef es_accepted_flag
=
987 LLVMBuildLoad(builder
,
988 si_build_gep_i8(ctx
, es_vtxptr
, lds_byte0_accept_flag
), "");
990 LLVMValueRef es_accepted_bool
= LLVMBuildICmp(builder
, LLVMIntNE
,
991 es_accepted_flag
, ctx
->ac
.i8_0
, "");
992 LLVMValueRef es_mask
= ac_get_i1_sgpr_mask(&ctx
->ac
, es_accepted_bool
);
994 LLVMBuildStore(builder
, es_accepted_bool
, es_accepted
);
996 ac_build_ifcc(&ctx
->ac
, LLVMBuildICmp(builder
, LLVMIntEQ
,
997 tid
, ctx
->ac
.i32_0
, ""), 16008);
999 LLVMBuildStore(builder
, es_mask
,
1000 ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
,
1001 get_wave_id_in_tg(ctx
)));
1003 ac_build_endif(&ctx
->ac
, 16008);
1005 ac_build_endif(&ctx
->ac
, 16007);
1006 ac_build_s_barrier(&ctx
->ac
);
1008 /* Load the vertex masks and compute the new ES thread count. */
1009 LLVMValueRef es_mask
[2], new_num_es_threads
, kill_wave
;
1010 load_bitmasks_2x64(ctx
, ctx
->gs_ngg_scratch
, 0, es_mask
, &new_num_es_threads
);
1011 new_num_es_threads
= ac_build_readlane_no_opt_barrier(&ctx
->ac
, new_num_es_threads
, NULL
);
1013 /* ES threads compute their prefix sum, which is the new ES thread ID.
1014 * Then they write the value of the old thread ID into the LDS address
1015 * of the new thread ID. It will be used it to load input VGPRs from
1016 * the old thread's LDS location.
1018 ac_build_ifcc(&ctx
->ac
, LLVMBuildLoad(builder
, es_accepted
, ""), 16009);
1020 LLVMValueRef old_id
= get_thread_id_in_tg(ctx
);
1021 LLVMValueRef new_id
= ac_prefix_bitcount_2x64(&ctx
->ac
, es_mask
, old_id
);
1023 LLVMBuildStore(builder
, LLVMBuildTrunc(builder
, old_id
, ctx
->ac
.i8
, ""),
1024 si_build_gep_i8(ctx
, ngg_nogs_vertex_ptr(ctx
, new_id
),
1025 lds_byte0_old_thread_id
));
1026 LLVMBuildStore(builder
, LLVMBuildTrunc(builder
, new_id
, ctx
->ac
.i8
, ""),
1027 si_build_gep_i8(ctx
, es_vtxptr
, lds_byte1_new_thread_id
));
1029 ac_build_endif(&ctx
->ac
, 16009);
1031 /* Kill waves that have inactive threads. */
1032 kill_wave
= LLVMBuildICmp(builder
, LLVMIntULE
,
1033 ac_build_imax(&ctx
->ac
, new_num_es_threads
, ngg_get_prim_cnt(ctx
)),
1034 LLVMBuildMul(builder
, get_wave_id_in_tg(ctx
),
1035 LLVMConstInt(ctx
->ac
.i32
, ctx
->ac
.wave_size
, 0), ""), "");
1036 ac_build_ifcc(&ctx
->ac
, kill_wave
, 19202);
1038 /* If we are killing wave 0, send that there are no primitives
1039 * in this threadgroup.
1041 ac_build_sendmsg_gs_alloc_req(&ctx
->ac
, get_wave_id_in_tg(ctx
),
1042 ctx
->ac
.i32_0
, ctx
->ac
.i32_0
);
1043 ac_build_s_endpgm(&ctx
->ac
);
1045 ac_build_endif(&ctx
->ac
, 19202);
1046 ac_build_s_barrier(&ctx
->ac
);
1048 /* Send the final vertex and primitive counts. */
1049 ac_build_sendmsg_gs_alloc_req(&ctx
->ac
, get_wave_id_in_tg(ctx
),
1050 new_num_es_threads
, ngg_get_prim_cnt(ctx
));
1052 /* Update thread counts in SGPRs. */
1053 LLVMValueRef new_gs_tg_info
= ac_get_arg(&ctx
->ac
, ctx
->gs_tg_info
);
1054 LLVMValueRef new_merged_wave_info
= ac_get_arg(&ctx
->ac
, ctx
->merged_wave_info
);
1056 /* This also converts the thread count from the total count to the per-wave count. */
1057 update_thread_counts(ctx
, &new_num_es_threads
, &new_gs_tg_info
, 9, 12,
1058 &new_merged_wave_info
, 8, 0);
1060 /* Update vertex indices in VGPR0 (same format as NGG passthrough). */
1061 LLVMValueRef new_vgpr0
= ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i32
, "");
1063 /* Set the null flag at the beginning (culled), and then
1064 * overwrite it for accepted primitives.
1066 LLVMBuildStore(builder
, LLVMConstInt(ctx
->ac
.i32
, 1u << 31, 0), new_vgpr0
);
1068 /* Get vertex indices after vertex compaction. */
1069 ac_build_ifcc(&ctx
->ac
, LLVMBuildTrunc(builder
, gs_accepted
, ctx
->ac
.i1
, ""), 16011);
1071 struct ac_ngg_prim prim
= {};
1072 prim
.num_vertices
= 3;
1073 prim
.isnull
= ctx
->ac
.i1false
;
1075 for (unsigned vtx
= 0; vtx
< 3; vtx
++) {
1077 LLVMBuildLoad(builder
,
1078 si_build_gep_i8(ctx
, gs_vtxptr
[vtx
],
1079 lds_byte1_new_thread_id
), "");
1080 prim
.index
[vtx
] = LLVMBuildZExt(builder
, prim
.index
[vtx
], ctx
->ac
.i32
, "");
1081 prim
.edgeflag
[vtx
] = ngg_get_initial_edgeflag(ctx
, vtx
);
1084 /* Set the new GS input VGPR. */
1085 LLVMBuildStore(builder
, ac_pack_prim_export(&ctx
->ac
, &prim
), new_vgpr0
);
1087 ac_build_endif(&ctx
->ac
, 16011);
1089 if (gfx10_ngg_export_prim_early(shader
))
1090 gfx10_ngg_build_export_prim(ctx
, NULL
, LLVMBuildLoad(builder
, new_vgpr0
, ""));
1092 /* Set the new ES input VGPRs. */
1093 LLVMValueRef es_data
[4];
1094 LLVMValueRef old_thread_id
= ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i32
, "");
1096 for (unsigned i
= 0; i
< 4; i
++)
1097 es_data
[i
] = ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i32
, "");
1099 ac_build_ifcc(&ctx
->ac
, LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
, tid
,
1100 new_num_es_threads
, ""), 16012);
1102 LLVMValueRef old_id
, old_es_vtxptr
, tmp
;
1104 /* Load ES input VGPRs from the ES thread before compaction. */
1105 old_id
= LLVMBuildLoad(builder
,
1106 si_build_gep_i8(ctx
, es_vtxptr
, lds_byte0_old_thread_id
), "");
1107 old_id
= LLVMBuildZExt(builder
, old_id
, ctx
->ac
.i32
, "");
1109 LLVMBuildStore(builder
, old_id
, old_thread_id
);
1110 old_es_vtxptr
= ngg_nogs_vertex_ptr(ctx
, old_id
);
1112 for (unsigned i
= 0; i
< 2; i
++) {
1113 tmp
= LLVMBuildLoad(builder
,
1114 ac_build_gep0(&ctx
->ac
, old_es_vtxptr
,
1115 LLVMConstInt(ctx
->ac
.i32
, lds_vertex_id
+ i
, 0)), "");
1116 LLVMBuildStore(builder
, tmp
, es_data
[i
]);
1119 if (ctx
->type
== PIPE_SHADER_TESS_EVAL
) {
1120 tmp
= LLVMBuildLoad(builder
,
1121 si_build_gep_i8(ctx
, old_es_vtxptr
,
1122 lds_byte2_tes_rel_patch_id
), "");
1123 tmp
= LLVMBuildZExt(builder
, tmp
, ctx
->ac
.i32
, "");
1124 LLVMBuildStore(builder
, tmp
, es_data
[2]);
1126 if (uses_tes_prim_id
) {
1127 tmp
= LLVMBuildLoad(builder
,
1128 ac_build_gep0(&ctx
->ac
, old_es_vtxptr
,
1129 LLVMConstInt(ctx
->ac
.i32
, lds_tes_patch_id
, 0)), "");
1130 LLVMBuildStore(builder
, tmp
, es_data
[3]);
1134 ac_build_endif(&ctx
->ac
, 16012);
1136 /* Return values for the main function. */
1137 LLVMValueRef ret
= ctx
->return_value
;
1140 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, new_gs_tg_info
, 2, "");
1141 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, new_merged_wave_info
, 3, "");
1142 if (ctx
->type
== PIPE_SHADER_TESS_EVAL
)
1143 ret
= si_insert_input_ret(ctx
, ret
, ctx
->tcs_offchip_offset
, 4);
1145 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->rw_buffers
,
1146 8 + SI_SGPR_RW_BUFFERS
);
1147 ret
= si_insert_input_ptr(ctx
, ret
,
1148 ctx
->bindless_samplers_and_images
,
1149 8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES
);
1150 ret
= si_insert_input_ptr(ctx
, ret
,
1151 ctx
->const_and_shader_buffers
,
1152 8 + SI_SGPR_CONST_AND_SHADER_BUFFERS
);
1153 ret
= si_insert_input_ptr(ctx
, ret
,
1154 ctx
->samplers_and_images
,
1155 8 + SI_SGPR_SAMPLERS_AND_IMAGES
);
1156 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->vs_state_bits
,
1157 8 + SI_SGPR_VS_STATE_BITS
);
1159 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
1160 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->args
.base_vertex
,
1161 8 + SI_SGPR_BASE_VERTEX
);
1162 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->args
.start_instance
,
1163 8 + SI_SGPR_START_INSTANCE
);
1164 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->args
.draw_id
,
1165 8 + SI_SGPR_DRAWID
);
1166 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->vertex_buffers
,
1167 8 + SI_VS_NUM_USER_SGPR
);
1169 for (unsigned i
= 0; i
< shader
->selector
->num_vbos_in_user_sgprs
; i
++) {
1170 ret
= si_insert_input_v4i32(ctx
, ret
, ctx
->vb_descriptors
[i
],
1171 8 + SI_SGPR_VS_VB_DESCRIPTOR_FIRST
+ i
* 4);
1174 assert(ctx
->type
== PIPE_SHADER_TESS_EVAL
);
1175 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->tcs_offchip_layout
,
1176 8 + SI_SGPR_TES_OFFCHIP_LAYOUT
);
1177 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->tes_offchip_addr
,
1178 8 + SI_SGPR_TES_OFFCHIP_ADDR
);
1182 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
1183 if (shader
->selector
->num_vbos_in_user_sgprs
) {
1184 vgpr
= 8 + SI_SGPR_VS_VB_DESCRIPTOR_FIRST
+
1185 shader
->selector
->num_vbos_in_user_sgprs
* 4;
1187 vgpr
= 8 + GFX9_VSGS_NUM_USER_SGPR
+ 1;
1190 vgpr
= 8 + GFX9_TESGS_NUM_USER_SGPR
;
1193 val
= LLVMBuildLoad(builder
, new_vgpr0
, "");
1194 ret
= LLVMBuildInsertValue(builder
, ret
, ac_to_float(&ctx
->ac
, val
),
1196 vgpr
++; /* gs_vtx23_offset */
1198 ret
= si_insert_input_ret_float(ctx
, ret
, ctx
->args
.gs_prim_id
, vgpr
++);
1199 ret
= si_insert_input_ret_float(ctx
, ret
, ctx
->args
.gs_invocation_id
, vgpr
++);
1200 vgpr
++; /* gs_vtx45_offset */
1202 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
1203 val
= LLVMBuildLoad(builder
, es_data
[0], "");
1204 ret
= LLVMBuildInsertValue(builder
, ret
, ac_to_float(&ctx
->ac
, val
),
1205 vgpr
++, ""); /* VGPR5 - VertexID */
1207 if (uses_instance_id
) {
1208 val
= LLVMBuildLoad(builder
, es_data
[1], "");
1209 ret
= LLVMBuildInsertValue(builder
, ret
, ac_to_float(&ctx
->ac
, val
),
1210 vgpr
++, ""); /* VGPR8 - InstanceID */
1215 assert(ctx
->type
== PIPE_SHADER_TESS_EVAL
);
1216 unsigned num_vgprs
= uses_tes_prim_id
? 4 : 3;
1217 for (unsigned i
= 0; i
< num_vgprs
; i
++) {
1218 val
= LLVMBuildLoad(builder
, es_data
[i
], "");
1219 ret
= LLVMBuildInsertValue(builder
, ret
, ac_to_float(&ctx
->ac
, val
),
1225 /* Return the old thread ID. */
1226 val
= LLVMBuildLoad(builder
, old_thread_id
, "");
1227 ret
= LLVMBuildInsertValue(builder
, ret
, ac_to_float(&ctx
->ac
, val
), vgpr
++, "");
1229 /* These two also use LDS. */
1230 if (sel
->info
.writes_edgeflag
||
1231 (ctx
->type
== PIPE_SHADER_VERTEX
&& shader
->key
.mono
.u
.vs_export_prim_id
))
1232 ac_build_s_barrier(&ctx
->ac
);
1234 ctx
->return_value
= ret
;
1238 * Emit the epilogue of an API VS or TES shader compiled as ESGS shader.
1240 void gfx10_emit_ngg_epilogue(struct ac_shader_abi
*abi
,
1241 unsigned max_outputs
,
1242 LLVMValueRef
*addrs
)
1244 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1245 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1246 struct si_shader_info
*info
= &sel
->info
;
1247 struct si_shader_output_values outputs
[PIPE_MAX_SHADER_OUTPUTS
];
1248 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1249 LLVMValueRef tmp
, tmp2
;
1251 assert(!ctx
->shader
->is_gs_copy_shader
);
1252 assert(info
->num_outputs
<= max_outputs
);
1254 LLVMValueRef vertex_ptr
= NULL
;
1256 if (sel
->so
.num_outputs
|| sel
->info
.writes_edgeflag
)
1257 vertex_ptr
= ngg_nogs_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
));
1259 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
1260 outputs
[i
].semantic_name
= info
->output_semantic_name
[i
];
1261 outputs
[i
].semantic_index
= info
->output_semantic_index
[i
];
1263 for (unsigned j
= 0; j
< 4; j
++) {
1264 outputs
[i
].vertex_stream
[j
] =
1265 (info
->output_streams
[i
] >> (2 * j
)) & 3;
1267 /* TODO: we may store more outputs than streamout needs,
1268 * but streamout performance isn't that important.
1270 if (sel
->so
.num_outputs
) {
1271 tmp
= ac_build_gep0(&ctx
->ac
, vertex_ptr
,
1272 LLVMConstInt(ctx
->ac
.i32
, 4 * i
+ j
, false));
1273 tmp2
= LLVMBuildLoad(builder
, addrs
[4 * i
+ j
], "");
1274 tmp2
= ac_to_integer(&ctx
->ac
, tmp2
);
1275 LLVMBuildStore(builder
, tmp2
, tmp
);
1279 /* Store the edgeflag at the end (if streamout is enabled) */
1280 if (info
->output_semantic_name
[i
] == TGSI_SEMANTIC_EDGEFLAG
&&
1281 sel
->info
.writes_edgeflag
) {
1282 LLVMValueRef edgeflag
= LLVMBuildLoad(builder
, addrs
[4 * i
], "");
1283 /* The output is a float, but the hw expects a 1-bit integer. */
1284 edgeflag
= LLVMBuildFPToUI(ctx
->ac
.builder
, edgeflag
, ctx
->ac
.i32
, "");
1285 edgeflag
= ac_build_umin(&ctx
->ac
, edgeflag
, ctx
->ac
.i32_1
);
1287 tmp
= LLVMConstInt(ctx
->ac
.i32
, ngg_nogs_vertex_size(ctx
->shader
) - 1, 0);
1288 tmp
= ac_build_gep0(&ctx
->ac
, vertex_ptr
, tmp
);
1289 LLVMBuildStore(builder
, edgeflag
, tmp
);
1293 bool unterminated_es_if_block
=
1294 !sel
->so
.num_outputs
&&
1295 !sel
->info
.writes_edgeflag
&&
1296 !ctx
->screen
->use_ngg_streamout
&& /* no query buffer */
1297 (ctx
->type
!= PIPE_SHADER_VERTEX
||
1298 !ctx
->shader
->key
.mono
.u
.vs_export_prim_id
);
1300 if (!unterminated_es_if_block
)
1301 ac_build_endif(&ctx
->ac
, ctx
->merged_wrap_if_label
);
1303 LLVMValueRef is_gs_thread
= si_is_gs_thread(ctx
);
1304 LLVMValueRef is_es_thread
= si_is_es_thread(ctx
);
1305 LLVMValueRef vtxindex
[3];
1307 if (ctx
->shader
->key
.opt
.ngg_culling
) {
1308 vtxindex
[0] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 0, 9);
1309 vtxindex
[1] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 10, 9);
1310 vtxindex
[2] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 20, 9);
1312 vtxindex
[0] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 0, 16);
1313 vtxindex
[1] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 16, 16);
1314 vtxindex
[2] = si_unpack_param(ctx
, ctx
->gs_vtx23_offset
, 0, 16);
1317 /* Determine the number of vertices per primitive. */
1318 unsigned num_vertices
;
1319 LLVMValueRef num_vertices_val
= ngg_get_vertices_per_prim(ctx
, &num_vertices
);
1322 LLVMValueRef emitted_prims
= NULL
;
1324 if (sel
->so
.num_outputs
) {
1325 assert(!unterminated_es_if_block
);
1327 struct ngg_streamout nggso
= {};
1328 nggso
.num_vertices
= num_vertices_val
;
1329 nggso
.prim_enable
[0] = is_gs_thread
;
1331 for (unsigned i
= 0; i
< num_vertices
; ++i
)
1332 nggso
.vertices
[i
] = ngg_nogs_vertex_ptr(ctx
, vtxindex
[i
]);
1334 build_streamout(ctx
, &nggso
);
1335 emitted_prims
= nggso
.emit
[0];
1338 LLVMValueRef user_edgeflags
[3] = {};
1340 if (sel
->info
.writes_edgeflag
) {
1341 assert(!unterminated_es_if_block
);
1343 /* Streamout already inserted the barrier, so don't insert it again. */
1344 if (!sel
->so
.num_outputs
)
1345 ac_build_s_barrier(&ctx
->ac
);
1347 ac_build_ifcc(&ctx
->ac
, is_gs_thread
, 5400);
1348 /* Load edge flags from ES threads and store them into VGPRs in GS threads. */
1349 for (unsigned i
= 0; i
< num_vertices
; i
++) {
1350 tmp
= ngg_nogs_vertex_ptr(ctx
, vtxindex
[i
]);
1351 tmp2
= LLVMConstInt(ctx
->ac
.i32
, ngg_nogs_vertex_size(ctx
->shader
) - 1, 0);
1352 tmp
= ac_build_gep0(&ctx
->ac
, tmp
, tmp2
);
1353 tmp
= LLVMBuildLoad(builder
, tmp
, "");
1354 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
1356 user_edgeflags
[i
] = ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i1
, "");
1357 LLVMBuildStore(builder
, tmp
, user_edgeflags
[i
]);
1359 ac_build_endif(&ctx
->ac
, 5400);
1362 /* Copy Primitive IDs from GS threads to the LDS address corresponding
1363 * to the ES thread of the provoking vertex.
1365 if (ctx
->type
== PIPE_SHADER_VERTEX
&&
1366 ctx
->shader
->key
.mono
.u
.vs_export_prim_id
) {
1367 assert(!unterminated_es_if_block
);
1369 /* Streamout and edge flags use LDS. Make it idle, so that we can reuse it. */
1370 if (sel
->so
.num_outputs
|| sel
->info
.writes_edgeflag
)
1371 ac_build_s_barrier(&ctx
->ac
);
1373 ac_build_ifcc(&ctx
->ac
, is_gs_thread
, 5400);
1374 /* Extract the PROVOKING_VTX_INDEX field. */
1375 LLVMValueRef provoking_vtx_in_prim
=
1376 si_unpack_param(ctx
, ctx
->vs_state_bits
, 4, 2);
1378 /* provoking_vtx_index = vtxindex[provoking_vtx_in_prim]; */
1379 LLVMValueRef indices
= ac_build_gather_values(&ctx
->ac
, vtxindex
, 3);
1380 LLVMValueRef provoking_vtx_index
=
1381 LLVMBuildExtractElement(builder
, indices
, provoking_vtx_in_prim
, "");
1382 LLVMValueRef vertex_ptr
= ngg_nogs_vertex_ptr(ctx
, provoking_vtx_index
);
1384 LLVMBuildStore(builder
, ac_get_arg(&ctx
->ac
, ctx
->args
.gs_prim_id
),
1385 ac_build_gep0(&ctx
->ac
, vertex_ptr
, ctx
->ac
.i32_0
));
1386 ac_build_endif(&ctx
->ac
, 5400);
1389 /* Update query buffer */
1390 if (ctx
->screen
->use_ngg_streamout
&&
1391 !info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
]) {
1392 assert(!unterminated_es_if_block
);
1394 tmp
= si_unpack_param(ctx
, ctx
->vs_state_bits
, 6, 1);
1395 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
1396 ac_build_ifcc(&ctx
->ac
, tmp
, 5029); /* if (STREAMOUT_QUERY_ENABLED) */
1397 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, get_wave_id_in_tg(ctx
), ctx
->ac
.i32_0
, "");
1398 ac_build_ifcc(&ctx
->ac
, tmp
, 5030);
1399 tmp
= LLVMBuildICmp(builder
, LLVMIntULE
, ac_get_thread_id(&ctx
->ac
),
1400 sel
->so
.num_outputs
? ctx
->ac
.i32_1
: ctx
->ac
.i32_0
, "");
1401 ac_build_ifcc(&ctx
->ac
, tmp
, 5031);
1403 LLVMValueRef args
[] = {
1404 ngg_get_prim_cnt(ctx
),
1405 ngg_get_query_buf(ctx
),
1406 LLVMConstInt(ctx
->ac
.i32
, 16, false), /* offset of stream[0].generated_primitives */
1407 ctx
->ac
.i32_0
, /* soffset */
1408 ctx
->ac
.i32_0
, /* cachepolicy */
1411 if (sel
->so
.num_outputs
) {
1412 args
[0] = ac_build_writelane(&ctx
->ac
, args
[0], emitted_prims
, ctx
->ac
.i32_1
);
1413 args
[2] = ac_build_writelane(&ctx
->ac
, args
[2],
1414 LLVMConstInt(ctx
->ac
.i32
, 24, false), ctx
->ac
.i32_1
);
1417 /* TODO: should this be 64-bit atomics? */
1418 ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.raw.buffer.atomic.add.i32",
1419 ctx
->ac
.i32
, args
, 5, 0);
1421 ac_build_endif(&ctx
->ac
, 5031);
1422 ac_build_endif(&ctx
->ac
, 5030);
1423 ac_build_endif(&ctx
->ac
, 5029);
1426 /* Build the primitive export. */
1427 if (!gfx10_ngg_export_prim_early(ctx
->shader
)) {
1428 assert(!unterminated_es_if_block
);
1429 gfx10_ngg_build_export_prim(ctx
, user_edgeflags
, NULL
);
1432 /* Export per-vertex data (positions and parameters). */
1433 if (!unterminated_es_if_block
)
1434 ac_build_ifcc(&ctx
->ac
, is_es_thread
, 6002);
1438 /* Unconditionally (re-)load the values for proper SSA form. */
1439 for (i
= 0; i
< info
->num_outputs
; i
++) {
1440 /* If the NGG cull shader part computed the position, don't
1441 * use the position from the current shader part. Instead,
1444 if (info
->output_semantic_name
[i
] == TGSI_SEMANTIC_POSITION
&&
1445 ctx
->shader
->key
.opt
.ngg_culling
) {
1446 vertex_ptr
= ngg_nogs_vertex_ptr(ctx
,
1447 ac_get_arg(&ctx
->ac
, ctx
->ngg_old_thread_id
));
1449 for (unsigned j
= 0; j
< 4; j
++) {
1450 tmp
= LLVMConstInt(ctx
->ac
.i32
, lds_pos_x
+ j
, 0);
1451 tmp
= ac_build_gep0(&ctx
->ac
, vertex_ptr
, tmp
);
1452 tmp
= LLVMBuildLoad(builder
, tmp
, "");
1453 outputs
[i
].values
[j
] = ac_to_float(&ctx
->ac
, tmp
);
1456 for (unsigned j
= 0; j
< 4; j
++) {
1457 outputs
[i
].values
[j
] =
1458 LLVMBuildLoad(builder
,
1459 addrs
[4 * i
+ j
], "");
1464 if (ctx
->shader
->key
.mono
.u
.vs_export_prim_id
) {
1465 outputs
[i
].semantic_name
= TGSI_SEMANTIC_PRIMID
;
1466 outputs
[i
].semantic_index
= 0;
1468 if (ctx
->type
== PIPE_SHADER_VERTEX
) {
1469 /* Wait for GS stores to finish. */
1470 ac_build_s_barrier(&ctx
->ac
);
1472 tmp
= ngg_nogs_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
));
1473 tmp
= ac_build_gep0(&ctx
->ac
, tmp
, ctx
->ac
.i32_0
);
1474 outputs
[i
].values
[0] = LLVMBuildLoad(builder
, tmp
, "");
1476 assert(ctx
->type
== PIPE_SHADER_TESS_EVAL
);
1477 outputs
[i
].values
[0] = si_get_primitive_id(ctx
, 0);
1480 outputs
[i
].values
[0] = ac_to_float(&ctx
->ac
, outputs
[i
].values
[0]);
1481 for (unsigned j
= 1; j
< 4; j
++)
1482 outputs
[i
].values
[j
] = LLVMGetUndef(ctx
->ac
.f32
);
1484 memset(outputs
[i
].vertex_stream
, 0,
1485 sizeof(outputs
[i
].vertex_stream
));
1489 si_llvm_build_vs_exports(ctx
, outputs
, i
);
1491 ac_build_endif(&ctx
->ac
, 6002);
1495 ngg_gs_get_vertex_storage(struct si_shader_context
*ctx
)
1497 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1498 const struct si_shader_info
*info
= &sel
->info
;
1500 LLVMTypeRef elements
[2] = {
1501 LLVMArrayType(ctx
->ac
.i32
, 4 * info
->num_outputs
),
1502 LLVMArrayType(ctx
->ac
.i8
, 4),
1504 LLVMTypeRef type
= LLVMStructTypeInContext(ctx
->ac
.context
, elements
, 2, false);
1505 type
= LLVMPointerType(LLVMArrayType(type
, 0), AC_ADDR_SPACE_LDS
);
1506 return LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->gs_ngg_emit
, type
, "");
1510 * Return a pointer to the LDS storage reserved for the N'th vertex, where N
1511 * is in emit order; that is:
1512 * - during the epilogue, N is the threadidx (relative to the entire threadgroup)
1513 * - during vertex emit, i.e. while the API GS shader invocation is running,
1514 * N = threadidx * gs_max_out_vertices + emitidx
1516 * Goals of the LDS memory layout:
1517 * 1. Eliminate bank conflicts on write for geometry shaders that have all emits
1518 * in uniform control flow
1519 * 2. Eliminate bank conflicts on read for export if, additionally, there is no
1521 * 3. Agnostic to the number of waves (since we don't know it before compiling)
1522 * 4. Allow coalescing of LDS instructions (ds_write_b128 etc.)
1523 * 5. Avoid wasting memory.
1525 * We use an AoS layout due to point 4 (this also helps point 3). In an AoS
1526 * layout, elimination of bank conflicts requires that each vertex occupy an
1527 * odd number of dwords. We use the additional dword to store the output stream
1528 * index as well as a flag to indicate whether this vertex ends a primitive
1529 * for rasterization.
1531 * Swizzling is required to satisfy points 1 and 2 simultaneously.
1533 * Vertices are stored in export order (gsthread * gs_max_out_vertices + emitidx).
1534 * Indices are swizzled in groups of 32, which ensures point 1 without
1535 * disturbing point 2.
1537 * \return an LDS pointer to type {[N x i32], [4 x i8]}
1540 ngg_gs_vertex_ptr(struct si_shader_context
*ctx
, LLVMValueRef vertexidx
)
1542 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1543 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1544 LLVMValueRef storage
= ngg_gs_get_vertex_storage(ctx
);
1546 /* gs_max_out_vertices = 2^(write_stride_2exp) * some odd number */
1547 unsigned write_stride_2exp
= ffs(sel
->gs_max_out_vertices
) - 1;
1548 if (write_stride_2exp
) {
1550 LLVMBuildLShr(builder
, vertexidx
,
1551 LLVMConstInt(ctx
->ac
.i32
, 5, false), "");
1552 LLVMValueRef swizzle
=
1553 LLVMBuildAnd(builder
, row
,
1554 LLVMConstInt(ctx
->ac
.i32
, (1u << write_stride_2exp
) - 1,
1556 vertexidx
= LLVMBuildXor(builder
, vertexidx
, swizzle
, "");
1559 return ac_build_gep0(&ctx
->ac
, storage
, vertexidx
);
1563 ngg_gs_emit_vertex_ptr(struct si_shader_context
*ctx
, LLVMValueRef gsthread
,
1564 LLVMValueRef emitidx
)
1566 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1567 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1570 tmp
= LLVMConstInt(ctx
->ac
.i32
, sel
->gs_max_out_vertices
, false);
1571 tmp
= LLVMBuildMul(builder
, tmp
, gsthread
, "");
1572 const LLVMValueRef vertexidx
= LLVMBuildAdd(builder
, tmp
, emitidx
, "");
1573 return ngg_gs_vertex_ptr(ctx
, vertexidx
);
1577 ngg_gs_get_emit_output_ptr(struct si_shader_context
*ctx
, LLVMValueRef vertexptr
,
1580 LLVMValueRef gep_idx
[3] = {
1581 ctx
->ac
.i32_0
, /* implied C-style array */
1582 ctx
->ac
.i32_0
, /* first struct entry */
1583 LLVMConstInt(ctx
->ac
.i32
, out_idx
, false),
1585 return LLVMBuildGEP(ctx
->ac
.builder
, vertexptr
, gep_idx
, 3, "");
1589 ngg_gs_get_emit_primflag_ptr(struct si_shader_context
*ctx
, LLVMValueRef vertexptr
,
1592 LLVMValueRef gep_idx
[3] = {
1593 ctx
->ac
.i32_0
, /* implied C-style array */
1594 ctx
->ac
.i32_1
, /* second struct entry */
1595 LLVMConstInt(ctx
->ac
.i32
, stream
, false),
1597 return LLVMBuildGEP(ctx
->ac
.builder
, vertexptr
, gep_idx
, 3, "");
1600 void gfx10_ngg_gs_emit_vertex(struct si_shader_context
*ctx
,
1602 LLVMValueRef
*addrs
)
1604 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1605 const struct si_shader_info
*info
= &sel
->info
;
1606 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1608 const LLVMValueRef vertexidx
=
1609 LLVMBuildLoad(builder
, ctx
->gs_next_vertex
[stream
], "");
1611 /* If this thread has already emitted the declared maximum number of
1612 * vertices, skip the write: excessive vertex emissions are not
1613 * supposed to have any effect.
1615 const LLVMValueRef can_emit
=
1616 LLVMBuildICmp(builder
, LLVMIntULT
, vertexidx
,
1617 LLVMConstInt(ctx
->ac
.i32
, sel
->gs_max_out_vertices
, false), "");
1619 tmp
= LLVMBuildAdd(builder
, vertexidx
, ctx
->ac
.i32_1
, "");
1620 tmp
= LLVMBuildSelect(builder
, can_emit
, tmp
, vertexidx
, "");
1621 LLVMBuildStore(builder
, tmp
, ctx
->gs_next_vertex
[stream
]);
1623 ac_build_ifcc(&ctx
->ac
, can_emit
, 9001);
1625 const LLVMValueRef vertexptr
=
1626 ngg_gs_emit_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
), vertexidx
);
1627 unsigned out_idx
= 0;
1628 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
1629 for (unsigned chan
= 0; chan
< 4; chan
++, out_idx
++) {
1630 if (!(info
->output_usagemask
[i
] & (1 << chan
)) ||
1631 ((info
->output_streams
[i
] >> (2 * chan
)) & 3) != stream
)
1634 LLVMValueRef out_val
= LLVMBuildLoad(builder
, addrs
[4 * i
+ chan
], "");
1635 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
1636 LLVMBuildStore(builder
, out_val
,
1637 ngg_gs_get_emit_output_ptr(ctx
, vertexptr
, out_idx
));
1640 assert(out_idx
* 4 == sel
->gsvs_vertex_size
);
1642 /* Determine and store whether this vertex completed a primitive. */
1643 const LLVMValueRef curverts
= LLVMBuildLoad(builder
, ctx
->gs_curprim_verts
[stream
], "");
1645 tmp
= LLVMConstInt(ctx
->ac
.i32
, u_vertices_per_prim(sel
->gs_output_prim
) - 1, false);
1646 const LLVMValueRef iscompleteprim
=
1647 LLVMBuildICmp(builder
, LLVMIntUGE
, curverts
, tmp
, "");
1649 /* Since the geometry shader emits triangle strips, we need to
1650 * track which primitive is odd and swap vertex indices to get
1651 * the correct vertex order.
1653 LLVMValueRef is_odd
= ctx
->ac
.i1false
;
1654 if (stream
== 0 && u_vertices_per_prim(sel
->gs_output_prim
) == 3) {
1655 tmp
= LLVMBuildAnd(builder
, curverts
, ctx
->ac
.i32_1
, "");
1656 is_odd
= LLVMBuildICmp(builder
, LLVMIntEQ
, tmp
, ctx
->ac
.i32_1
, "");
1659 tmp
= LLVMBuildAdd(builder
, curverts
, ctx
->ac
.i32_1
, "");
1660 LLVMBuildStore(builder
, tmp
, ctx
->gs_curprim_verts
[stream
]);
1662 /* The per-vertex primitive flag encoding:
1663 * bit 0: whether this vertex finishes a primitive
1664 * bit 1: whether the primitive is odd (if we are emitting triangle strips)
1666 tmp
= LLVMBuildZExt(builder
, iscompleteprim
, ctx
->ac
.i8
, "");
1667 tmp
= LLVMBuildOr(builder
, tmp
,
1668 LLVMBuildShl(builder
,
1669 LLVMBuildZExt(builder
, is_odd
, ctx
->ac
.i8
, ""),
1670 ctx
->ac
.i8_1
, ""), "");
1671 LLVMBuildStore(builder
, tmp
, ngg_gs_get_emit_primflag_ptr(ctx
, vertexptr
, stream
));
1673 tmp
= LLVMBuildLoad(builder
, ctx
->gs_generated_prims
[stream
], "");
1674 tmp
= LLVMBuildAdd(builder
, tmp
, LLVMBuildZExt(builder
, iscompleteprim
, ctx
->ac
.i32
, ""), "");
1675 LLVMBuildStore(builder
, tmp
, ctx
->gs_generated_prims
[stream
]);
1677 ac_build_endif(&ctx
->ac
, 9001);
1680 void gfx10_ngg_gs_emit_prologue(struct si_shader_context
*ctx
)
1682 /* Zero out the part of LDS scratch that is used to accumulate the
1683 * per-stream generated primitive count.
1685 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1686 LLVMValueRef scratchptr
= ctx
->gs_ngg_scratch
;
1687 LLVMValueRef tid
= get_thread_id_in_tg(ctx
);
1690 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
1691 ac_build_ifcc(&ctx
->ac
, tmp
, 5090);
1693 LLVMValueRef ptr
= ac_build_gep0(&ctx
->ac
, scratchptr
, tid
);
1694 LLVMBuildStore(builder
, ctx
->ac
.i32_0
, ptr
);
1696 ac_build_endif(&ctx
->ac
, 5090);
1698 ac_build_s_barrier(&ctx
->ac
);
1701 void gfx10_ngg_gs_emit_epilogue(struct si_shader_context
*ctx
)
1703 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1704 const struct si_shader_info
*info
= &sel
->info
;
1705 const unsigned verts_per_prim
= u_vertices_per_prim(sel
->gs_output_prim
);
1706 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1707 LLVMValueRef i8_0
= LLVMConstInt(ctx
->ac
.i8
, 0, false);
1708 LLVMValueRef tmp
, tmp2
;
1710 /* Zero out remaining (non-emitted) primitive flags.
1712 * Note: Alternatively, we could pass the relevant gs_next_vertex to
1713 * the emit threads via LDS. This is likely worse in the expected
1714 * typical case where each GS thread emits the full set of
1717 for (unsigned stream
= 0; stream
< 4; ++stream
) {
1718 if (!info
->num_stream_output_components
[stream
])
1721 const LLVMValueRef gsthread
= get_thread_id_in_tg(ctx
);
1723 ac_build_bgnloop(&ctx
->ac
, 5100);
1725 const LLVMValueRef vertexidx
=
1726 LLVMBuildLoad(builder
, ctx
->gs_next_vertex
[stream
], "");
1727 tmp
= LLVMBuildICmp(builder
, LLVMIntUGE
, vertexidx
,
1728 LLVMConstInt(ctx
->ac
.i32
, sel
->gs_max_out_vertices
, false), "");
1729 ac_build_ifcc(&ctx
->ac
, tmp
, 5101);
1730 ac_build_break(&ctx
->ac
);
1731 ac_build_endif(&ctx
->ac
, 5101);
1733 tmp
= LLVMBuildAdd(builder
, vertexidx
, ctx
->ac
.i32_1
, "");
1734 LLVMBuildStore(builder
, tmp
, ctx
->gs_next_vertex
[stream
]);
1736 tmp
= ngg_gs_emit_vertex_ptr(ctx
, gsthread
, vertexidx
);
1737 LLVMBuildStore(builder
, i8_0
, ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, stream
));
1739 ac_build_endloop(&ctx
->ac
, 5100);
1742 /* Accumulate generated primitives counts across the entire threadgroup. */
1743 for (unsigned stream
= 0; stream
< 4; ++stream
) {
1744 if (!info
->num_stream_output_components
[stream
])
1747 LLVMValueRef numprims
=
1748 LLVMBuildLoad(builder
, ctx
->gs_generated_prims
[stream
], "");
1749 numprims
= ac_build_reduce(&ctx
->ac
, numprims
, nir_op_iadd
, ctx
->ac
.wave_size
);
1751 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, ac_get_thread_id(&ctx
->ac
), ctx
->ac
.i32_0
, "");
1752 ac_build_ifcc(&ctx
->ac
, tmp
, 5105);
1754 LLVMBuildAtomicRMW(builder
, LLVMAtomicRMWBinOpAdd
,
1755 ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
,
1756 LLVMConstInt(ctx
->ac
.i32
, stream
, false)),
1757 numprims
, LLVMAtomicOrderingMonotonic
, false);
1759 ac_build_endif(&ctx
->ac
, 5105);
1762 ac_build_endif(&ctx
->ac
, ctx
->merged_wrap_if_label
);
1764 ac_build_s_barrier(&ctx
->ac
);
1766 const LLVMValueRef tid
= get_thread_id_in_tg(ctx
);
1767 LLVMValueRef num_emit_threads
= ngg_get_prim_cnt(ctx
);
1770 if (sel
->so
.num_outputs
) {
1771 struct ngg_streamout nggso
= {};
1773 nggso
.num_vertices
= LLVMConstInt(ctx
->ac
.i32
, verts_per_prim
, false);
1775 LLVMValueRef vertexptr
= ngg_gs_vertex_ptr(ctx
, tid
);
1776 for (unsigned stream
= 0; stream
< 4; ++stream
) {
1777 if (!info
->num_stream_output_components
[stream
])
1780 tmp
= LLVMBuildLoad(builder
, ngg_gs_get_emit_primflag_ptr(ctx
, vertexptr
, stream
), "");
1781 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
1782 tmp2
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, num_emit_threads
, "");
1783 nggso
.prim_enable
[stream
] = LLVMBuildAnd(builder
, tmp
, tmp2
, "");
1786 for (unsigned i
= 0; i
< verts_per_prim
; ++i
) {
1787 tmp
= LLVMBuildSub(builder
, tid
,
1788 LLVMConstInt(ctx
->ac
.i32
, verts_per_prim
- i
- 1, false), "");
1789 tmp
= ngg_gs_vertex_ptr(ctx
, tmp
);
1790 nggso
.vertices
[i
] = ac_build_gep0(&ctx
->ac
, tmp
, ctx
->ac
.i32_0
);
1793 build_streamout(ctx
, &nggso
);
1796 /* Write shader query data. */
1797 if (ctx
->screen
->use_ngg_streamout
) {
1798 tmp
= si_unpack_param(ctx
, ctx
->vs_state_bits
, 6, 1);
1799 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
1800 ac_build_ifcc(&ctx
->ac
, tmp
, 5109); /* if (STREAMOUT_QUERY_ENABLED) */
1801 unsigned num_query_comps
= sel
->so
.num_outputs
? 8 : 4;
1802 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
,
1803 LLVMConstInt(ctx
->ac
.i32
, num_query_comps
, false), "");
1804 ac_build_ifcc(&ctx
->ac
, tmp
, 5110);
1806 LLVMValueRef offset
;
1808 if (sel
->so
.num_outputs
)
1809 tmp
= LLVMBuildAnd(builder
, tmp
, LLVMConstInt(ctx
->ac
.i32
, 3, false), "");
1810 offset
= LLVMBuildNUWMul(builder
, tmp
, LLVMConstInt(ctx
->ac
.i32
, 32, false), "");
1811 if (sel
->so
.num_outputs
) {
1812 tmp
= LLVMBuildLShr(builder
, tid
, LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
1813 tmp
= LLVMBuildNUWMul(builder
, tmp
, LLVMConstInt(ctx
->ac
.i32
, 8, false), "");
1814 offset
= LLVMBuildAdd(builder
, offset
, tmp
, "");
1817 tmp
= LLVMBuildLoad(builder
, ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, tid
), "");
1818 LLVMValueRef args
[] = {
1820 ngg_get_query_buf(ctx
),
1822 LLVMConstInt(ctx
->ac
.i32
, 16, false), /* soffset */
1823 ctx
->ac
.i32_0
, /* cachepolicy */
1825 ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.raw.buffer.atomic.add.i32",
1826 ctx
->ac
.i32
, args
, 5, 0);
1828 ac_build_endif(&ctx
->ac
, 5110);
1829 ac_build_endif(&ctx
->ac
, 5109);
1832 /* Determine vertex liveness. */
1833 LLVMValueRef vertliveptr
= ac_build_alloca(&ctx
->ac
, ctx
->ac
.i1
, "vertexlive");
1835 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, num_emit_threads
, "");
1836 ac_build_ifcc(&ctx
->ac
, tmp
, 5120);
1838 for (unsigned i
= 0; i
< verts_per_prim
; ++i
) {
1839 const LLVMValueRef primidx
=
1840 LLVMBuildAdd(builder
, tid
,
1841 LLVMConstInt(ctx
->ac
.i32
, i
, false), "");
1844 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, primidx
, num_emit_threads
, "");
1845 ac_build_ifcc(&ctx
->ac
, tmp
, 5121 + i
);
1848 /* Load primitive liveness */
1849 tmp
= ngg_gs_vertex_ptr(ctx
, primidx
);
1850 tmp
= LLVMBuildLoad(builder
, ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, 0), "");
1851 const LLVMValueRef primlive
=
1852 LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
1854 tmp
= LLVMBuildLoad(builder
, vertliveptr
, "");
1855 tmp
= LLVMBuildOr(builder
, tmp
, primlive
, ""),
1856 LLVMBuildStore(builder
, tmp
, vertliveptr
);
1859 ac_build_endif(&ctx
->ac
, 5121 + i
);
1862 ac_build_endif(&ctx
->ac
, 5120);
1864 /* Inclusive scan addition across the current wave. */
1865 LLVMValueRef vertlive
= LLVMBuildLoad(builder
, vertliveptr
, "");
1866 struct ac_wg_scan vertlive_scan
= {};
1867 vertlive_scan
.op
= nir_op_iadd
;
1868 vertlive_scan
.enable_reduce
= true;
1869 vertlive_scan
.enable_exclusive
= true;
1870 vertlive_scan
.src
= vertlive
;
1871 vertlive_scan
.scratch
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, ctx
->ac
.i32_0
);
1872 vertlive_scan
.waveidx
= get_wave_id_in_tg(ctx
);
1873 vertlive_scan
.numwaves
= get_tgsize(ctx
);
1874 vertlive_scan
.maxwaves
= 8;
1876 ac_build_wg_scan(&ctx
->ac
, &vertlive_scan
);
1878 /* Skip all exports (including index exports) when possible. At least on
1879 * early gfx10 revisions this is also to avoid hangs.
1881 LLVMValueRef have_exports
=
1882 LLVMBuildICmp(builder
, LLVMIntNE
, vertlive_scan
.result_reduce
, ctx
->ac
.i32_0
, "");
1884 LLVMBuildSelect(builder
, have_exports
, num_emit_threads
, ctx
->ac
.i32_0
, "");
1886 /* Allocate export space. Send this message as early as possible, to
1887 * hide the latency of the SQ <-> SPI roundtrip.
1889 * Note: We could consider compacting primitives for export as well.
1890 * PA processes 1 non-null prim / clock, but it fetches 4 DW of
1891 * prim data per clock and skips null primitives at no additional
1892 * cost. So compacting primitives can only be beneficial when
1893 * there are 4 or more contiguous null primitives in the export
1894 * (in the common case of single-dword prim exports).
1896 ac_build_sendmsg_gs_alloc_req(&ctx
->ac
, get_wave_id_in_tg(ctx
),
1897 vertlive_scan
.result_reduce
, num_emit_threads
);
1899 /* Setup the reverse vertex compaction permutation. We re-use stream 1
1900 * of the primitive liveness flags, relying on the fact that each
1901 * threadgroup can have at most 256 threads. */
1902 ac_build_ifcc(&ctx
->ac
, vertlive
, 5130);
1904 tmp
= ngg_gs_vertex_ptr(ctx
, vertlive_scan
.result_exclusive
);
1905 tmp2
= LLVMBuildTrunc(builder
, tid
, ctx
->ac
.i8
, "");
1906 LLVMBuildStore(builder
, tmp2
, ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, 1));
1908 ac_build_endif(&ctx
->ac
, 5130);
1910 ac_build_s_barrier(&ctx
->ac
);
1912 /* Export primitive data */
1913 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, num_emit_threads
, "");
1914 ac_build_ifcc(&ctx
->ac
, tmp
, 5140);
1917 struct ac_ngg_prim prim
= {};
1918 prim
.num_vertices
= verts_per_prim
;
1920 tmp
= ngg_gs_vertex_ptr(ctx
, tid
);
1921 flags
= LLVMBuildLoad(builder
, ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, 0), "");
1922 prim
.isnull
= LLVMBuildNot(builder
, LLVMBuildTrunc(builder
, flags
, ctx
->ac
.i1
, ""), "");
1924 for (unsigned i
= 0; i
< verts_per_prim
; ++i
) {
1925 prim
.index
[i
] = LLVMBuildSub(builder
, vertlive_scan
.result_exclusive
,
1926 LLVMConstInt(ctx
->ac
.i32
, verts_per_prim
- i
- 1, false), "");
1927 prim
.edgeflag
[i
] = ctx
->ac
.i1false
;
1930 /* Geometry shaders output triangle strips, but NGG expects triangles. */
1931 if (verts_per_prim
== 3) {
1932 LLVMValueRef is_odd
= LLVMBuildLShr(builder
, flags
, ctx
->ac
.i8_1
, "");
1933 is_odd
= LLVMBuildTrunc(builder
, is_odd
, ctx
->ac
.i1
, "");
1934 LLVMValueRef flatshade_first
=
1935 LLVMBuildICmp(builder
, LLVMIntEQ
,
1936 si_unpack_param(ctx
, ctx
->vs_state_bits
, 4, 2),
1939 ac_build_triangle_strip_indices_to_triangle(&ctx
->ac
, is_odd
,
1944 ac_build_export_prim(&ctx
->ac
, &prim
);
1946 ac_build_endif(&ctx
->ac
, 5140);
1948 /* Export position and parameter data */
1949 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, vertlive_scan
.result_reduce
, "");
1950 ac_build_ifcc(&ctx
->ac
, tmp
, 5145);
1952 struct si_shader_output_values outputs
[PIPE_MAX_SHADER_OUTPUTS
];
1954 tmp
= ngg_gs_vertex_ptr(ctx
, tid
);
1955 tmp
= LLVMBuildLoad(builder
, ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, 1), "");
1956 tmp
= LLVMBuildZExt(builder
, tmp
, ctx
->ac
.i32
, "");
1957 const LLVMValueRef vertexptr
= ngg_gs_vertex_ptr(ctx
, tmp
);
1959 unsigned out_idx
= 0;
1960 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
1961 outputs
[i
].semantic_name
= info
->output_semantic_name
[i
];
1962 outputs
[i
].semantic_index
= info
->output_semantic_index
[i
];
1964 for (unsigned j
= 0; j
< 4; j
++, out_idx
++) {
1965 tmp
= ngg_gs_get_emit_output_ptr(ctx
, vertexptr
, out_idx
);
1966 tmp
= LLVMBuildLoad(builder
, tmp
, "");
1967 outputs
[i
].values
[j
] = ac_to_float(&ctx
->ac
, tmp
);
1968 outputs
[i
].vertex_stream
[j
] =
1969 (info
->output_streams
[i
] >> (2 * j
)) & 3;
1973 si_llvm_build_vs_exports(ctx
, outputs
, info
->num_outputs
);
1975 ac_build_endif(&ctx
->ac
, 5145);
1978 static void clamp_gsprims_to_esverts(unsigned *max_gsprims
, unsigned max_esverts
,
1979 unsigned min_verts_per_prim
, bool use_adjacency
)
1981 unsigned max_reuse
= max_esverts
- min_verts_per_prim
;
1984 *max_gsprims
= MIN2(*max_gsprims
, 1 + max_reuse
);
1988 * Determine subgroup information like maximum number of vertices and prims.
1990 * This happens before the shader is uploaded, since LDS relocations during
1991 * upload depend on the subgroup size.
1993 void gfx10_ngg_calculate_subgroup_info(struct si_shader
*shader
)
1995 const struct si_shader_selector
*gs_sel
= shader
->selector
;
1996 const struct si_shader_selector
*es_sel
=
1997 shader
->previous_stage_sel
? shader
->previous_stage_sel
: gs_sel
;
1998 const enum pipe_shader_type gs_type
= gs_sel
->type
;
1999 const unsigned gs_num_invocations
= MAX2(gs_sel
->gs_num_invocations
, 1);
2000 const unsigned input_prim
= si_get_input_prim(gs_sel
);
2001 const bool use_adjacency
= input_prim
>= PIPE_PRIM_LINES_ADJACENCY
&&
2002 input_prim
<= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
;
2003 const unsigned max_verts_per_prim
= u_vertices_per_prim(input_prim
);
2004 const unsigned min_verts_per_prim
=
2005 gs_type
== PIPE_SHADER_GEOMETRY
? max_verts_per_prim
: 1;
2007 /* All these are in dwords: */
2008 /* We can't allow using the whole LDS, because GS waves compete with
2009 * other shader stages for LDS space.
2011 * TODO: We should really take the shader's internal LDS use into
2012 * account. The linker will fail if the size is greater than
2015 const unsigned max_lds_size
= 8 * 1024 - 768;
2016 const unsigned target_lds_size
= max_lds_size
;
2017 unsigned esvert_lds_size
= 0;
2018 unsigned gsprim_lds_size
= 0;
2020 /* All these are per subgroup: */
2021 bool max_vert_out_per_gs_instance
= false;
2022 unsigned max_gsprims_base
= 128; /* default prim group size clamp */
2023 unsigned max_esverts_base
= 128;
2025 if (shader
->key
.opt
.ngg_culling
& SI_NGG_CULL_GS_FAST_LAUNCH_TRI_LIST
) {
2026 max_gsprims_base
= 128 / 3;
2027 max_esverts_base
= max_gsprims_base
* 3;
2028 } else if (shader
->key
.opt
.ngg_culling
& SI_NGG_CULL_GS_FAST_LAUNCH_TRI_STRIP
) {
2029 max_gsprims_base
= 126;
2030 max_esverts_base
= 128;
2033 /* Hardware has the following non-natural restrictions on the value
2034 * of GE_CNTL.VERT_GRP_SIZE based on based on the primitive type of
2036 * - at most 252 for any line input primitive type
2037 * - at most 251 for any quad input primitive type
2038 * - at most 251 for triangle strips with adjacency (this happens to
2039 * be the natural limit for triangle *lists* with adjacency)
2041 max_esverts_base
= MIN2(max_esverts_base
, 251 + max_verts_per_prim
- 1);
2043 if (gs_type
== PIPE_SHADER_GEOMETRY
) {
2044 unsigned max_out_verts_per_gsprim
=
2045 gs_sel
->gs_max_out_vertices
* gs_num_invocations
;
2047 if (max_out_verts_per_gsprim
<= 256) {
2048 if (max_out_verts_per_gsprim
) {
2049 max_gsprims_base
= MIN2(max_gsprims_base
,
2050 256 / max_out_verts_per_gsprim
);
2053 /* Use special multi-cycling mode in which each GS
2054 * instance gets its own subgroup. Does not work with
2056 max_vert_out_per_gs_instance
= true;
2057 max_gsprims_base
= 1;
2058 max_out_verts_per_gsprim
= gs_sel
->gs_max_out_vertices
;
2061 esvert_lds_size
= es_sel
->esgs_itemsize
/ 4;
2062 gsprim_lds_size
= (gs_sel
->gsvs_vertex_size
/ 4 + 1) * max_out_verts_per_gsprim
;
2065 /* LDS size for passing data from ES to GS. */
2066 esvert_lds_size
= ngg_nogs_vertex_size(shader
);
2069 unsigned max_gsprims
= max_gsprims_base
;
2070 unsigned max_esverts
= max_esverts_base
;
2072 if (esvert_lds_size
)
2073 max_esverts
= MIN2(max_esverts
, target_lds_size
/ esvert_lds_size
);
2074 if (gsprim_lds_size
)
2075 max_gsprims
= MIN2(max_gsprims
, target_lds_size
/ gsprim_lds_size
);
2077 max_esverts
= MIN2(max_esverts
, max_gsprims
* max_verts_per_prim
);
2078 clamp_gsprims_to_esverts(&max_gsprims
, max_esverts
, min_verts_per_prim
, use_adjacency
);
2079 assert(max_esverts
>= max_verts_per_prim
&& max_gsprims
>= 1);
2081 if (esvert_lds_size
|| gsprim_lds_size
) {
2082 /* Now that we have a rough proportionality between esverts
2083 * and gsprims based on the primitive type, scale both of them
2084 * down simultaneously based on required LDS space.
2086 * We could be smarter about this if we knew how much vertex
2089 unsigned lds_total
= max_esverts
* esvert_lds_size
+
2090 max_gsprims
* gsprim_lds_size
;
2091 if (lds_total
> target_lds_size
) {
2092 max_esverts
= max_esverts
* target_lds_size
/ lds_total
;
2093 max_gsprims
= max_gsprims
* target_lds_size
/ lds_total
;
2095 max_esverts
= MIN2(max_esverts
, max_gsprims
* max_verts_per_prim
);
2096 clamp_gsprims_to_esverts(&max_gsprims
, max_esverts
,
2097 min_verts_per_prim
, use_adjacency
);
2098 assert(max_esverts
>= max_verts_per_prim
&& max_gsprims
>= 1);
2102 /* Round up towards full wave sizes for better ALU utilization. */
2103 if (!max_vert_out_per_gs_instance
) {
2104 const unsigned wavesize
= gs_sel
->screen
->ge_wave_size
;
2105 unsigned orig_max_esverts
;
2106 unsigned orig_max_gsprims
;
2108 orig_max_esverts
= max_esverts
;
2109 orig_max_gsprims
= max_gsprims
;
2111 max_esverts
= align(max_esverts
, wavesize
);
2112 max_esverts
= MIN2(max_esverts
, max_esverts_base
);
2113 if (esvert_lds_size
)
2114 max_esverts
= MIN2(max_esverts
,
2115 (max_lds_size
- max_gsprims
* gsprim_lds_size
) /
2117 max_esverts
= MIN2(max_esverts
, max_gsprims
* max_verts_per_prim
);
2119 max_gsprims
= align(max_gsprims
, wavesize
);
2120 max_gsprims
= MIN2(max_gsprims
, max_gsprims_base
);
2121 if (gsprim_lds_size
)
2122 max_gsprims
= MIN2(max_gsprims
,
2123 (max_lds_size
- max_esverts
* esvert_lds_size
) /
2125 clamp_gsprims_to_esverts(&max_gsprims
, max_esverts
,
2126 min_verts_per_prim
, use_adjacency
);
2127 assert(max_esverts
>= max_verts_per_prim
&& max_gsprims
>= 1);
2128 } while (orig_max_esverts
!= max_esverts
|| orig_max_gsprims
!= max_gsprims
);
2131 /* Hardware restriction: minimum value of max_esverts */
2132 max_esverts
= MAX2(max_esverts
, 23 + max_verts_per_prim
);
2134 unsigned max_out_vertices
=
2135 max_vert_out_per_gs_instance
? gs_sel
->gs_max_out_vertices
:
2136 gs_type
== PIPE_SHADER_GEOMETRY
?
2137 max_gsprims
* gs_num_invocations
* gs_sel
->gs_max_out_vertices
:
2139 assert(max_out_vertices
<= 256);
2141 unsigned prim_amp_factor
= 1;
2142 if (gs_type
== PIPE_SHADER_GEOMETRY
) {
2143 /* Number of output primitives per GS input primitive after
2145 prim_amp_factor
= gs_sel
->gs_max_out_vertices
;
2148 /* The GE only checks against the maximum number of ES verts after
2149 * allocating a full GS primitive. So we need to ensure that whenever
2150 * this check passes, there is enough space for a full primitive without
2153 shader
->ngg
.hw_max_esverts
= max_esverts
- max_verts_per_prim
+ 1;
2154 shader
->ngg
.max_gsprims
= max_gsprims
;
2155 shader
->ngg
.max_out_verts
= max_out_vertices
;
2156 shader
->ngg
.prim_amp_factor
= prim_amp_factor
;
2157 shader
->ngg
.max_vert_out_per_gs_instance
= max_vert_out_per_gs_instance
;
2159 shader
->gs_info
.esgs_ring_size
= 4 * max_esverts
* esvert_lds_size
;
2160 shader
->ngg
.ngg_emit_size
= max_gsprims
* gsprim_lds_size
;
2162 assert(shader
->ngg
.hw_max_esverts
>= 24); /* HW limitation */