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
->stage
== MESA_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
->stage
== MESA_SHADER_VERTEX
) {
93 if (info
->base
.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
->stage
== MESA_SHADER_TESS_EVAL
);
112 if (info
->base
.tess
.point_mode
)
114 else if (info
->base
.tess
.primitive_mode
== GL_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
->info
.stage
!= MESA_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
= info
->output_semantic
[reg
];
237 for (unsigned comp
= 0; comp
< 4; comp
++) {
238 tmp
= ac_build_gep0(&ctx
->ac
, vertexptr
, LLVMConstInt(ctx
->ac
.i32
, 4 * reg
+ comp
, false));
239 out
.values
[comp
] = LLVMBuildLoad(builder
, tmp
, "");
240 out
.vertex_stream
[comp
] = (info
->output_streams
[reg
] >> (2 * comp
)) & 3;
243 si_llvm_streamout_store_output(ctx
, so_buffer
, offset
, &so
->output
[i
], &out
);
247 struct ngg_streamout
{
248 LLVMValueRef num_vertices
;
250 /* per-thread data */
251 LLVMValueRef prim_enable
[4]; /* i1 per stream */
252 LLVMValueRef vertices
[3]; /* [N x i32] addrspace(LDS)* */
255 LLVMValueRef emit
[4]; /* per-stream emitted primitives (only valid for used streams) */
259 * Build streamout logic.
263 * Writes number of emitted primitives to gs_ngg_scratch[4:8].
265 * Clobbers gs_ngg_scratch[8:].
267 static void build_streamout(struct si_shader_context
*ctx
, struct ngg_streamout
*nggso
)
269 struct si_shader_info
*info
= &ctx
->shader
->selector
->info
;
270 struct pipe_stream_output_info
*so
= &ctx
->shader
->selector
->so
;
271 LLVMBuilderRef builder
= ctx
->ac
.builder
;
272 LLVMValueRef buf_ptr
= ac_get_arg(&ctx
->ac
, ctx
->rw_buffers
);
273 LLVMValueRef tid
= get_thread_id_in_tg(ctx
);
274 LLVMValueRef tmp
, tmp2
;
275 LLVMValueRef i32_2
= LLVMConstInt(ctx
->ac
.i32
, 2, false);
276 LLVMValueRef i32_4
= LLVMConstInt(ctx
->ac
.i32
, 4, false);
277 LLVMValueRef i32_8
= LLVMConstInt(ctx
->ac
.i32
, 8, false);
278 LLVMValueRef so_buffer
[4] = {};
279 unsigned max_num_vertices
= 1 + (nggso
->vertices
[1] ? 1 : 0) + (nggso
->vertices
[2] ? 1 : 0);
280 LLVMValueRef prim_stride_dw
[4] = {};
281 LLVMValueRef prim_stride_dw_vgpr
= LLVMGetUndef(ctx
->ac
.i32
);
282 int stream_for_buffer
[4] = {-1, -1, -1, -1};
283 unsigned bufmask_for_stream
[4] = {};
284 bool isgs
= ctx
->stage
== MESA_SHADER_GEOMETRY
;
285 unsigned scratch_emit_base
= isgs
? 4 : 0;
286 LLVMValueRef scratch_emit_basev
= isgs
? i32_4
: ctx
->ac
.i32_0
;
287 unsigned scratch_offset_base
= isgs
? 8 : 4;
288 LLVMValueRef scratch_offset_basev
= isgs
? i32_8
: i32_4
;
290 ac_llvm_add_target_dep_function_attr(ctx
->main_fn
, "amdgpu-gds-size", 256);
292 /* Determine the mapping of streamout buffers to vertex streams. */
293 for (unsigned i
= 0; i
< so
->num_outputs
; ++i
) {
294 unsigned buf
= so
->output
[i
].output_buffer
;
295 unsigned stream
= so
->output
[i
].stream
;
296 assert(stream_for_buffer
[buf
] < 0 || stream_for_buffer
[buf
] == stream
);
297 stream_for_buffer
[buf
] = stream
;
298 bufmask_for_stream
[stream
] |= 1 << buf
;
301 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
302 if (stream_for_buffer
[buffer
] == -1)
305 assert(so
->stride
[buffer
]);
307 tmp
= LLVMConstInt(ctx
->ac
.i32
, so
->stride
[buffer
], false);
308 prim_stride_dw
[buffer
] = LLVMBuildMul(builder
, tmp
, nggso
->num_vertices
, "");
309 prim_stride_dw_vgpr
=
310 ac_build_writelane(&ctx
->ac
, prim_stride_dw_vgpr
, prim_stride_dw
[buffer
],
311 LLVMConstInt(ctx
->ac
.i32
, buffer
, false));
313 so_buffer
[buffer
] = ac_build_load_to_sgpr(
314 &ctx
->ac
, buf_ptr
, LLVMConstInt(ctx
->ac
.i32
, SI_VS_STREAMOUT_BUF0
+ buffer
, false));
317 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, get_wave_id_in_tg(ctx
), ctx
->ac
.i32_0
, "");
318 ac_build_ifcc(&ctx
->ac
, tmp
, 5200);
320 LLVMTypeRef gdsptr
= LLVMPointerType(ctx
->ac
.i32
, AC_ADDR_SPACE_GDS
);
321 LLVMValueRef gdsbase
= LLVMBuildIntToPtr(builder
, ctx
->ac
.i32_0
, gdsptr
, "");
323 /* Advance the streamout offsets in GDS. */
324 LLVMValueRef offsets_vgpr
= ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i32
, "");
325 LLVMValueRef generated_by_stream_vgpr
= ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i32
, "");
327 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, ac_get_thread_id(&ctx
->ac
), i32_4
, "");
328 ac_build_ifcc(&ctx
->ac
, tmp
, 5210);
331 tmp
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, tid
);
332 tmp
= LLVMBuildLoad(builder
, tmp
, "");
334 tmp
= ac_build_writelane(&ctx
->ac
, ctx
->ac
.i32_0
, ngg_get_prim_cnt(ctx
), ctx
->ac
.i32_0
);
336 LLVMBuildStore(builder
, tmp
, generated_by_stream_vgpr
);
339 int unused_stream
= -1;
340 for (unsigned stream
= 0; stream
< 4; ++stream
) {
341 if (!info
->num_stream_output_components
[stream
]) {
342 unused_stream
= stream
;
346 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
347 if (stream_for_buffer
[buffer
] >= 0) {
348 swizzle
[buffer
] = stream_for_buffer
[buffer
];
350 assert(unused_stream
>= 0);
351 swizzle
[buffer
] = unused_stream
;
355 tmp
= ac_build_quad_swizzle(&ctx
->ac
, tmp
, swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
356 tmp
= LLVMBuildMul(builder
, tmp
, prim_stride_dw_vgpr
, "");
358 LLVMValueRef args
[] = {
359 LLVMBuildIntToPtr(builder
, ngg_get_ordered_id(ctx
), gdsptr
, ""),
361 ctx
->ac
.i32_0
, // ordering
362 ctx
->ac
.i32_0
, // scope
363 ctx
->ac
.i1false
, // isVolatile
364 LLVMConstInt(ctx
->ac
.i32
, 4 << 24, false), // OA index
365 ctx
->ac
.i1true
, // wave release
366 ctx
->ac
.i1true
, // wave done
368 tmp
= ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.ds.ordered.add", ctx
->ac
.i32
, args
,
369 ARRAY_SIZE(args
), 0);
371 /* Keep offsets in a VGPR for quick retrieval via readlane by
372 * the first wave for bounds checking, and also store in LDS
373 * for retrieval by all waves later. */
374 LLVMBuildStore(builder
, tmp
, offsets_vgpr
);
376 tmp2
= LLVMBuildAdd(builder
, ac_get_thread_id(&ctx
->ac
), scratch_offset_basev
, "");
377 tmp2
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, tmp2
);
378 LLVMBuildStore(builder
, tmp
, tmp2
);
380 ac_build_endif(&ctx
->ac
, 5210);
382 /* Determine the max emit per buffer. This is done via the SALU, in part
383 * because LLVM can't generate divide-by-multiply if we try to do this
384 * via VALU with one lane per buffer.
386 LLVMValueRef max_emit
[4] = {};
387 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
388 if (stream_for_buffer
[buffer
] == -1)
391 LLVMValueRef bufsize_dw
= LLVMBuildLShr(
392 builder
, LLVMBuildExtractElement(builder
, so_buffer
[buffer
], i32_2
, ""), i32_2
, "");
394 tmp
= LLVMBuildLoad(builder
, offsets_vgpr
, "");
395 LLVMValueRef offset_dw
=
396 ac_build_readlane(&ctx
->ac
, tmp
, LLVMConstInt(ctx
->ac
.i32
, buffer
, false));
398 tmp
= LLVMBuildSub(builder
, bufsize_dw
, offset_dw
, "");
399 tmp
= LLVMBuildUDiv(builder
, tmp
, prim_stride_dw
[buffer
], "");
401 tmp2
= LLVMBuildICmp(builder
, LLVMIntULT
, bufsize_dw
, offset_dw
, "");
402 max_emit
[buffer
] = LLVMBuildSelect(builder
, tmp2
, ctx
->ac
.i32_0
, tmp
, "");
405 /* Determine the number of emitted primitives per stream and fixup the
406 * GDS counter if necessary.
408 * This is complicated by the fact that a single stream can emit to
409 * multiple buffers (but luckily not vice versa).
411 LLVMValueRef emit_vgpr
= ctx
->ac
.i32_0
;
413 for (unsigned stream
= 0; stream
< 4; ++stream
) {
414 if (!info
->num_stream_output_components
[stream
])
417 tmp
= LLVMBuildLoad(builder
, generated_by_stream_vgpr
, "");
418 LLVMValueRef generated
=
419 ac_build_readlane(&ctx
->ac
, tmp
, LLVMConstInt(ctx
->ac
.i32
, stream
, false));
421 LLVMValueRef emit
= generated
;
422 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
423 if (stream_for_buffer
[buffer
] == stream
)
424 emit
= ac_build_umin(&ctx
->ac
, emit
, max_emit
[buffer
]);
428 ac_build_writelane(&ctx
->ac
, emit_vgpr
, emit
, LLVMConstInt(ctx
->ac
.i32
, stream
, false));
430 /* Fixup the offset using a plain GDS atomic if we overflowed. */
431 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, emit
, generated
, "");
432 ac_build_ifcc(&ctx
->ac
, tmp
, 5221); /* scalar branch */
433 tmp
= LLVMBuildLShr(builder
, LLVMConstInt(ctx
->ac
.i32
, bufmask_for_stream
[stream
], false),
434 ac_get_thread_id(&ctx
->ac
), "");
435 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
436 ac_build_ifcc(&ctx
->ac
, tmp
, 5222);
438 tmp
= LLVMBuildSub(builder
, generated
, emit
, "");
439 tmp
= LLVMBuildMul(builder
, tmp
, prim_stride_dw_vgpr
, "");
440 tmp2
= LLVMBuildGEP(builder
, gdsbase
, &tid
, 1, "");
441 LLVMBuildAtomicRMW(builder
, LLVMAtomicRMWBinOpSub
, tmp2
, tmp
,
442 LLVMAtomicOrderingMonotonic
, false);
444 ac_build_endif(&ctx
->ac
, 5222);
445 ac_build_endif(&ctx
->ac
, 5221);
448 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, ac_get_thread_id(&ctx
->ac
), i32_4
, "");
449 ac_build_ifcc(&ctx
->ac
, tmp
, 5225);
451 tmp
= LLVMBuildAdd(builder
, ac_get_thread_id(&ctx
->ac
), scratch_emit_basev
, "");
452 tmp
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, tmp
);
453 LLVMBuildStore(builder
, emit_vgpr
, tmp
);
455 ac_build_endif(&ctx
->ac
, 5225);
457 ac_build_endif(&ctx
->ac
, 5200);
459 /* Determine the workgroup-relative per-thread / primitive offset into
460 * the streamout buffers */
461 struct ac_wg_scan primemit_scan
[4] = {};
464 for (unsigned stream
= 0; stream
< 4; ++stream
) {
465 if (!info
->num_stream_output_components
[stream
])
468 primemit_scan
[stream
].enable_exclusive
= true;
469 primemit_scan
[stream
].op
= nir_op_iadd
;
470 primemit_scan
[stream
].src
= nggso
->prim_enable
[stream
];
471 primemit_scan
[stream
].scratch
= ac_build_gep0(
472 &ctx
->ac
, ctx
->gs_ngg_scratch
, LLVMConstInt(ctx
->ac
.i32
, 12 + 8 * stream
, false));
473 primemit_scan
[stream
].waveidx
= get_wave_id_in_tg(ctx
);
474 primemit_scan
[stream
].numwaves
= get_tgsize(ctx
);
475 primemit_scan
[stream
].maxwaves
= 8;
476 ac_build_wg_scan_top(&ctx
->ac
, &primemit_scan
[stream
]);
480 ac_build_s_barrier(&ctx
->ac
);
482 /* Fetch the per-buffer offsets and per-stream emit counts in all waves. */
483 LLVMValueRef wgoffset_dw
[4] = {};
486 LLVMValueRef scratch_vgpr
;
488 tmp
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, ac_get_thread_id(&ctx
->ac
));
489 scratch_vgpr
= LLVMBuildLoad(builder
, tmp
, "");
491 for (unsigned buffer
= 0; buffer
< 4; ++buffer
) {
492 if (stream_for_buffer
[buffer
] >= 0) {
493 wgoffset_dw
[buffer
] =
494 ac_build_readlane(&ctx
->ac
, scratch_vgpr
,
495 LLVMConstInt(ctx
->ac
.i32
, scratch_offset_base
+ buffer
, false));
499 for (unsigned stream
= 0; stream
< 4; ++stream
) {
500 if (info
->num_stream_output_components
[stream
]) {
501 nggso
->emit
[stream
] =
502 ac_build_readlane(&ctx
->ac
, scratch_vgpr
,
503 LLVMConstInt(ctx
->ac
.i32
, scratch_emit_base
+ stream
, false));
508 /* Write out primitive data */
509 for (unsigned stream
= 0; stream
< 4; ++stream
) {
510 if (!info
->num_stream_output_components
[stream
])
514 ac_build_wg_scan_bottom(&ctx
->ac
, &primemit_scan
[stream
]);
516 primemit_scan
[stream
].result_exclusive
= tid
;
519 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, primemit_scan
[stream
].result_exclusive
,
520 nggso
->emit
[stream
], "");
521 tmp
= LLVMBuildAnd(builder
, tmp
, nggso
->prim_enable
[stream
], "");
522 ac_build_ifcc(&ctx
->ac
, tmp
, 5240);
524 LLVMValueRef offset_vtx
=
525 LLVMBuildMul(builder
, primemit_scan
[stream
].result_exclusive
, nggso
->num_vertices
, "");
527 for (unsigned i
= 0; i
< max_num_vertices
; ++i
) {
528 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, LLVMConstInt(ctx
->ac
.i32
, i
, false),
529 nggso
->num_vertices
, "");
530 ac_build_ifcc(&ctx
->ac
, tmp
, 5241);
531 build_streamout_vertex(ctx
, so_buffer
, wgoffset_dw
, stream
, offset_vtx
,
533 ac_build_endif(&ctx
->ac
, 5241);
534 offset_vtx
= LLVMBuildAdd(builder
, offset_vtx
, ctx
->ac
.i32_1
, "");
537 ac_build_endif(&ctx
->ac
, 5240);
541 /* LDS layout of ES vertex data for NGG culling. */
544 /* Byte 0: Boolean ES thread accepted (unculled) flag, and later the old
545 * ES thread ID. After vertex compaction, compacted ES threads
546 * store the old thread ID here to copy input VGPRs from uncompacted
548 * Byte 1: New ES thread ID, loaded by GS to prepare the prim export value.
549 * Byte 2: TES rel patch ID
552 lds_byte0_accept_flag
= 0,
553 lds_byte0_old_thread_id
= 0,
554 lds_byte1_new_thread_id
,
555 lds_byte2_tes_rel_patch_id
,
558 lds_packed_data
= 0, /* lds_byteN_... */
568 lds_instance_id
, /* optional */
570 lds_tes_u
= lds_vertex_id
,
571 lds_tes_v
= lds_instance_id
,
572 lds_tes_patch_id
, /* optional */
575 static LLVMValueRef
si_build_gep_i8(struct si_shader_context
*ctx
, LLVMValueRef ptr
,
578 assert(byte_index
< 4);
579 LLVMTypeRef pi8
= LLVMPointerType(ctx
->ac
.i8
, AC_ADDR_SPACE_LDS
);
580 LLVMValueRef index
= LLVMConstInt(ctx
->ac
.i32
, byte_index
, 0);
582 return LLVMBuildGEP(ctx
->ac
.builder
, LLVMBuildPointerCast(ctx
->ac
.builder
, ptr
, pi8
, ""), &index
,
586 static unsigned ngg_nogs_vertex_size(struct si_shader
*shader
)
588 unsigned lds_vertex_size
= 0;
590 /* The edgeflag is always stored in the last element that's also
591 * used for padding to reduce LDS bank conflicts. */
592 if (shader
->selector
->so
.num_outputs
)
593 lds_vertex_size
= 4 * shader
->selector
->info
.num_outputs
+ 1;
594 if (shader
->selector
->info
.writes_edgeflag
)
595 lds_vertex_size
= MAX2(lds_vertex_size
, 1);
597 /* LDS size for passing data from GS to ES.
598 * GS stores Primitive IDs into LDS at the address corresponding
599 * to the ES thread of the provoking vertex. All ES threads
600 * load and export PrimitiveID for their thread.
602 if (shader
->selector
->info
.stage
== MESA_SHADER_VERTEX
&& shader
->key
.mono
.u
.vs_export_prim_id
)
603 lds_vertex_size
= MAX2(lds_vertex_size
, 1);
605 if (shader
->key
.opt
.ngg_culling
) {
606 if (shader
->selector
->info
.stage
== MESA_SHADER_VERTEX
) {
607 STATIC_ASSERT(lds_instance_id
+ 1 == 9);
608 lds_vertex_size
= MAX2(lds_vertex_size
, 9);
610 assert(shader
->selector
->info
.stage
== MESA_SHADER_TESS_EVAL
);
612 if (shader
->selector
->info
.uses_primid
|| shader
->key
.mono
.u
.vs_export_prim_id
) {
613 STATIC_ASSERT(lds_tes_patch_id
+ 2 == 11);
614 lds_vertex_size
= MAX2(lds_vertex_size
, 11);
616 STATIC_ASSERT(lds_tes_v
+ 1 == 9);
617 lds_vertex_size
= MAX2(lds_vertex_size
, 9);
622 return lds_vertex_size
;
626 * Returns an `[N x i32] addrspace(LDS)*` pointing at contiguous LDS storage
627 * for the vertex outputs.
629 static LLVMValueRef
ngg_nogs_vertex_ptr(struct si_shader_context
*ctx
, LLVMValueRef vtxid
)
631 /* The extra dword is used to avoid LDS bank conflicts. */
632 unsigned vertex_size
= ngg_nogs_vertex_size(ctx
->shader
);
633 LLVMTypeRef ai32
= LLVMArrayType(ctx
->ac
.i32
, vertex_size
);
634 LLVMTypeRef pai32
= LLVMPointerType(ai32
, AC_ADDR_SPACE_LDS
);
635 LLVMValueRef tmp
= LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->esgs_ring
, pai32
, "");
636 return LLVMBuildGEP(ctx
->ac
.builder
, tmp
, &vtxid
, 1, "");
639 static LLVMValueRef
si_insert_input_v4i32(struct si_shader_context
*ctx
, LLVMValueRef ret
,
640 struct ac_arg param
, unsigned return_index
)
642 LLVMValueRef v
= ac_get_arg(&ctx
->ac
, param
);
644 for (unsigned i
= 0; i
< 4; i
++) {
645 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, ac_llvm_extract_elem(&ctx
->ac
, v
, i
),
646 return_index
+ i
, "");
651 static void load_bitmasks_2x64(struct si_shader_context
*ctx
, LLVMValueRef lds_ptr
,
652 unsigned dw_offset
, LLVMValueRef mask
[2],
653 LLVMValueRef
*total_bitcount
)
655 LLVMBuilderRef builder
= ctx
->ac
.builder
;
656 LLVMValueRef ptr64
= LLVMBuildPointerCast(
657 builder
, lds_ptr
, LLVMPointerType(LLVMArrayType(ctx
->ac
.i64
, 2), AC_ADDR_SPACE_LDS
), "");
658 for (unsigned i
= 0; i
< 2; i
++) {
659 LLVMValueRef index
= LLVMConstInt(ctx
->ac
.i32
, dw_offset
/ 2 + i
, 0);
660 mask
[i
] = LLVMBuildLoad(builder
, ac_build_gep0(&ctx
->ac
, ptr64
, index
), "");
663 /* We get better code if we don't use the 128-bit bitcount. */
664 *total_bitcount
= LLVMBuildAdd(builder
, ac_build_bit_count(&ctx
->ac
, mask
[0]),
665 ac_build_bit_count(&ctx
->ac
, mask
[1]), "");
669 * Given a total thread count, update total and per-wave thread counts in input SGPRs
670 * and return the per-wave thread count.
672 * \param new_num_threads Total thread count on the input, per-wave thread count on the output.
673 * \param tg_info tg_info SGPR value
674 * \param tg_info_num_bits the bit size of thread count field in tg_info
675 * \param tg_info_shift the bit offset of the thread count field in tg_info
676 * \param wave_info merged_wave_info SGPR value
677 * \param wave_info_num_bits the bit size of thread count field in merged_wave_info
678 * \param wave_info_shift the bit offset of the thread count field in merged_wave_info
680 static void update_thread_counts(struct si_shader_context
*ctx
, LLVMValueRef
*new_num_threads
,
681 LLVMValueRef
*tg_info
, unsigned tg_info_num_bits
,
682 unsigned tg_info_shift
, LLVMValueRef
*wave_info
,
683 unsigned wave_info_num_bits
, unsigned wave_info_shift
)
685 LLVMBuilderRef builder
= ctx
->ac
.builder
;
687 /* Update the total thread count. */
688 unsigned tg_info_mask
= ~(u_bit_consecutive(0, tg_info_num_bits
) << tg_info_shift
);
689 *tg_info
= LLVMBuildAnd(builder
, *tg_info
, LLVMConstInt(ctx
->ac
.i32
, tg_info_mask
, 0), "");
690 *tg_info
= LLVMBuildOr(
692 LLVMBuildShl(builder
, *new_num_threads
, LLVMConstInt(ctx
->ac
.i32
, tg_info_shift
, 0), ""), "");
694 /* Update the per-wave thread count. */
695 LLVMValueRef prev_threads
= LLVMBuildMul(builder
, get_wave_id_in_tg(ctx
),
696 LLVMConstInt(ctx
->ac
.i32
, ctx
->ac
.wave_size
, 0), "");
697 *new_num_threads
= LLVMBuildSub(builder
, *new_num_threads
, prev_threads
, "");
698 *new_num_threads
= ac_build_imax(&ctx
->ac
, *new_num_threads
, ctx
->ac
.i32_0
);
700 ac_build_imin(&ctx
->ac
, *new_num_threads
, LLVMConstInt(ctx
->ac
.i32
, ctx
->ac
.wave_size
, 0));
701 unsigned wave_info_mask
= ~(u_bit_consecutive(0, wave_info_num_bits
) << wave_info_shift
);
702 *wave_info
= LLVMBuildAnd(builder
, *wave_info
, LLVMConstInt(ctx
->ac
.i32
, wave_info_mask
, 0), "");
703 *wave_info
= LLVMBuildOr(
705 LLVMBuildShl(builder
, *new_num_threads
, LLVMConstInt(ctx
->ac
.i32
, wave_info_shift
, 0), ""),
710 * Cull primitives for NGG VS or TES, then compact vertices, which happens
711 * before the VS or TES main function. Return values for the main function.
712 * Also return the position, which is passed to the shader as an input,
713 * so that we don't compute it twice.
715 void gfx10_emit_ngg_culling_epilogue(struct ac_shader_abi
*abi
, unsigned max_outputs
,
718 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
719 struct si_shader
*shader
= ctx
->shader
;
720 struct si_shader_selector
*sel
= shader
->selector
;
721 struct si_shader_info
*info
= &sel
->info
;
722 LLVMBuilderRef builder
= ctx
->ac
.builder
;
723 unsigned max_waves
= ctx
->ac
.wave_size
== 64 ? 2 : 4;
724 LLVMValueRef ngg_scratch
= ctx
->gs_ngg_scratch
;
726 if (ctx
->ac
.wave_size
== 64) {
727 ngg_scratch
= LLVMBuildPointerCast(builder
, ngg_scratch
,
728 LLVMPointerType(LLVMArrayType(ctx
->ac
.i64
, max_waves
),
729 AC_ADDR_SPACE_LDS
), "");
732 assert(shader
->key
.opt
.ngg_culling
);
733 assert(shader
->key
.as_ngg
);
734 assert(sel
->info
.stage
== MESA_SHADER_VERTEX
||
735 (sel
->info
.stage
== MESA_SHADER_TESS_EVAL
&& !shader
->key
.as_es
));
737 LLVMValueRef position
[4] = {};
738 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
739 switch (info
->output_semantic
[i
]) {
740 case VARYING_SLOT_POS
:
741 for (unsigned j
= 0; j
< 4; j
++) {
742 position
[j
] = LLVMBuildLoad(ctx
->ac
.builder
, addrs
[4 * i
+ j
], "");
749 /* Store Position.XYZW into LDS. */
750 LLVMValueRef es_vtxptr
= ngg_nogs_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
));
751 for (unsigned chan
= 0; chan
< 4; chan
++) {
753 builder
, ac_to_integer(&ctx
->ac
, position
[chan
]),
754 ac_build_gep0(&ctx
->ac
, es_vtxptr
, LLVMConstInt(ctx
->ac
.i32
, lds_pos_x
+ chan
, 0)));
756 /* Store Position.XY / W into LDS. */
757 for (unsigned chan
= 0; chan
< 2; chan
++) {
758 LLVMValueRef val
= ac_build_fdiv(&ctx
->ac
, position
[chan
], position
[3]);
760 builder
, ac_to_integer(&ctx
->ac
, val
),
761 ac_build_gep0(&ctx
->ac
, es_vtxptr
, LLVMConstInt(ctx
->ac
.i32
, lds_pos_x_div_w
+ chan
, 0)));
764 /* Store VertexID and InstanceID. ES threads will have to load them
765 * from LDS after vertex compaction and use them instead of their own
768 bool uses_instance_id
= false;
769 bool uses_tes_prim_id
= false;
770 LLVMValueRef packed_data
= ctx
->ac
.i32_0
;
772 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
773 uses_instance_id
= sel
->info
.uses_instanceid
||
774 shader
->key
.part
.vs
.prolog
.instance_divisor_is_one
||
775 shader
->key
.part
.vs
.prolog
.instance_divisor_is_fetched
;
778 builder
, ctx
->abi
.vertex_id
,
779 ac_build_gep0(&ctx
->ac
, es_vtxptr
, LLVMConstInt(ctx
->ac
.i32
, lds_vertex_id
, 0)));
780 if (uses_instance_id
) {
782 builder
, ctx
->abi
.instance_id
,
783 ac_build_gep0(&ctx
->ac
, es_vtxptr
, LLVMConstInt(ctx
->ac
.i32
, lds_instance_id
, 0)));
786 uses_tes_prim_id
= sel
->info
.uses_primid
|| shader
->key
.mono
.u
.vs_export_prim_id
;
788 assert(ctx
->stage
== MESA_SHADER_TESS_EVAL
);
789 LLVMBuildStore(builder
, ac_to_integer(&ctx
->ac
, ac_get_arg(&ctx
->ac
, ctx
->tes_u
)),
790 ac_build_gep0(&ctx
->ac
, es_vtxptr
, LLVMConstInt(ctx
->ac
.i32
, lds_tes_u
, 0)));
791 LLVMBuildStore(builder
, ac_to_integer(&ctx
->ac
, ac_get_arg(&ctx
->ac
, ctx
->tes_v
)),
792 ac_build_gep0(&ctx
->ac
, es_vtxptr
, LLVMConstInt(ctx
->ac
.i32
, lds_tes_v
, 0)));
793 packed_data
= LLVMBuildShl(builder
, ac_get_arg(&ctx
->ac
, ctx
->tes_rel_patch_id
),
794 LLVMConstInt(ctx
->ac
.i32
, lds_byte2_tes_rel_patch_id
* 8, 0), "");
795 if (uses_tes_prim_id
) {
797 builder
, ac_get_arg(&ctx
->ac
, ctx
->args
.tes_patch_id
),
798 ac_build_gep0(&ctx
->ac
, es_vtxptr
, LLVMConstInt(ctx
->ac
.i32
, lds_tes_patch_id
, 0)));
801 /* Initialize the packed data. */
803 builder
, packed_data
,
804 ac_build_gep0(&ctx
->ac
, es_vtxptr
, LLVMConstInt(ctx
->ac
.i32
, lds_packed_data
, 0)));
805 ac_build_endif(&ctx
->ac
, ctx
->merged_wrap_if_label
);
807 LLVMValueRef tid
= ac_get_thread_id(&ctx
->ac
);
809 /* Initialize all but the first element of ngg_scratch to 0, because we may have less
810 * than the maximum number of waves, but we always read all values. This is where
811 * the thread bitmasks of unculled threads will be stored.
813 * ngg_scratch layout: iN_wavemask esmask[0..n]
815 ac_build_ifcc(&ctx
->ac
,
816 LLVMBuildICmp(builder
, LLVMIntULT
, get_thread_id_in_tg(ctx
),
817 LLVMConstInt(ctx
->ac
.i32
, max_waves
- 1, 0), ""),
820 LLVMValueRef index
= LLVMBuildAdd(builder
, tid
, ctx
->ac
.i32_1
, "");
821 LLVMBuildStore(builder
, LLVMConstInt(ctx
->ac
.iN_wavemask
, 0, 0),
822 ac_build_gep0(&ctx
->ac
, ngg_scratch
, index
));
824 ac_build_endif(&ctx
->ac
, 16101);
825 ac_build_s_barrier(&ctx
->ac
);
827 /* The hardware requires that there are no holes between unculled vertices,
828 * which means we have to pack ES threads, i.e. reduce the ES thread count
829 * and move ES input VGPRs to lower threads. The upside is that varyings
830 * are only fetched and computed for unculled vertices.
832 * Vertex compaction in GS threads:
834 * Part 1: Compute the surviving vertex mask in GS threads:
835 * - Compute 4 32-bit surviving vertex masks in LDS. (max 4 waves)
836 * - In GS, notify ES threads whether the vertex survived.
838 * - ES threads will create the mask and store it in LDS.
840 * - Each GS thread loads the vertex masks from LDS.
842 * Part 2: Compact ES threads in GS threads:
843 * - Compute the prefix sum for all 3 vertices from the masks. These are the new
844 * thread IDs for each vertex within the primitive.
845 * - Write the value of the old thread ID into the LDS address of the new thread ID.
846 * The ES thread will load the old thread ID and use it to load the position, VertexID,
848 * - Update vertex indices and null flag in the GS input VGPRs.
851 * Part 3: Update inputs GPRs
852 * - For all waves, update per-wave thread counts in input SGPRs.
853 * - In ES threads, update the ES input VGPRs (VertexID, InstanceID, TES inputs).
856 LLVMValueRef vtxindex
[3];
857 if (shader
->key
.opt
.ngg_culling
& SI_NGG_CULL_GS_FAST_LAUNCH_ALL
) {
858 /* For the GS fast launch, the VS prologs simply puts the Vertex IDs
861 vtxindex
[0] = ac_get_arg(&ctx
->ac
, ctx
->gs_vtx01_offset
);
862 vtxindex
[1] = ac_get_arg(&ctx
->ac
, ctx
->gs_vtx23_offset
);
863 vtxindex
[2] = ac_get_arg(&ctx
->ac
, ctx
->gs_vtx45_offset
);
865 vtxindex
[0] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 0, 16);
866 vtxindex
[1] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 16, 16);
867 vtxindex
[2] = si_unpack_param(ctx
, ctx
->gs_vtx23_offset
, 0, 16);
869 LLVMValueRef gs_vtxptr
[] = {
870 ngg_nogs_vertex_ptr(ctx
, vtxindex
[0]),
871 ngg_nogs_vertex_ptr(ctx
, vtxindex
[1]),
872 ngg_nogs_vertex_ptr(ctx
, vtxindex
[2]),
874 es_vtxptr
= ngg_nogs_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
));
876 LLVMValueRef gs_accepted
= ac_build_alloca(&ctx
->ac
, ctx
->ac
.i32
, "");
878 /* Do culling in GS threads. */
879 ac_build_ifcc(&ctx
->ac
, si_is_gs_thread(ctx
), 16002);
881 /* Load positions. */
882 LLVMValueRef pos
[3][4] = {};
883 for (unsigned vtx
= 0; vtx
< 3; vtx
++) {
884 for (unsigned chan
= 0; chan
< 4; chan
++) {
886 if (chan
== 0 || chan
== 1)
887 index
= lds_pos_x_div_w
+ chan
;
894 ac_build_gep0(&ctx
->ac
, gs_vtxptr
[vtx
], LLVMConstInt(ctx
->ac
.i32
, index
, 0));
895 pos
[vtx
][chan
] = LLVMBuildLoad(builder
, addr
, "");
896 pos
[vtx
][chan
] = ac_to_float(&ctx
->ac
, pos
[vtx
][chan
]);
900 /* Load the viewport state for small prim culling. */
901 LLVMValueRef vp
= ac_build_load_invariant(
902 &ctx
->ac
, ac_get_arg(&ctx
->ac
, ctx
->small_prim_cull_info
), ctx
->ac
.i32_0
);
903 vp
= LLVMBuildBitCast(builder
, vp
, ctx
->ac
.v4f32
, "");
904 LLVMValueRef vp_scale
[2], vp_translate
[2];
905 vp_scale
[0] = ac_llvm_extract_elem(&ctx
->ac
, vp
, 0);
906 vp_scale
[1] = ac_llvm_extract_elem(&ctx
->ac
, vp
, 1);
907 vp_translate
[0] = ac_llvm_extract_elem(&ctx
->ac
, vp
, 2);
908 vp_translate
[1] = ac_llvm_extract_elem(&ctx
->ac
, vp
, 3);
910 /* Get the small prim filter precision. */
911 LLVMValueRef small_prim_precision
= si_unpack_param(ctx
, ctx
->vs_state_bits
, 7, 4);
912 small_prim_precision
=
913 LLVMBuildOr(builder
, small_prim_precision
, LLVMConstInt(ctx
->ac
.i32
, 0x70, 0), "");
914 small_prim_precision
=
915 LLVMBuildShl(builder
, small_prim_precision
, LLVMConstInt(ctx
->ac
.i32
, 23, 0), "");
916 small_prim_precision
= LLVMBuildBitCast(builder
, small_prim_precision
, ctx
->ac
.f32
, "");
918 /* Execute culling code. */
919 struct ac_cull_options options
= {};
920 options
.cull_front
= shader
->key
.opt
.ngg_culling
& SI_NGG_CULL_FRONT_FACE
;
921 options
.cull_back
= shader
->key
.opt
.ngg_culling
& SI_NGG_CULL_BACK_FACE
;
922 options
.cull_view_xy
= shader
->key
.opt
.ngg_culling
& SI_NGG_CULL_VIEW_SMALLPRIMS
;
923 options
.cull_small_prims
= options
.cull_view_xy
;
924 options
.cull_zero_area
= options
.cull_front
|| options
.cull_back
;
925 options
.cull_w
= true;
927 /* Tell ES threads whether their vertex survived. */
928 ac_build_ifcc(&ctx
->ac
,
929 ac_cull_triangle(&ctx
->ac
, pos
, ctx
->ac
.i1true
, vp_scale
, vp_translate
,
930 small_prim_precision
, &options
),
933 LLVMBuildStore(builder
, ctx
->ac
.i32_1
, gs_accepted
);
934 for (unsigned vtx
= 0; vtx
< 3; vtx
++) {
935 LLVMBuildStore(builder
, ctx
->ac
.i8_1
,
936 si_build_gep_i8(ctx
, gs_vtxptr
[vtx
], lds_byte0_accept_flag
));
939 ac_build_endif(&ctx
->ac
, 16003);
941 ac_build_endif(&ctx
->ac
, 16002);
942 ac_build_s_barrier(&ctx
->ac
);
944 gs_accepted
= LLVMBuildLoad(builder
, gs_accepted
, "");
946 LLVMValueRef es_accepted
= ac_build_alloca(&ctx
->ac
, ctx
->ac
.i1
, "");
948 /* Convert the per-vertex flag to a thread bitmask in ES threads and store it in LDS. */
949 ac_build_ifcc(&ctx
->ac
, si_is_es_thread(ctx
), 16007);
951 LLVMValueRef es_accepted_flag
=
952 LLVMBuildLoad(builder
, si_build_gep_i8(ctx
, es_vtxptr
, lds_byte0_accept_flag
), "");
954 LLVMValueRef es_accepted_bool
=
955 LLVMBuildICmp(builder
, LLVMIntNE
, es_accepted_flag
, ctx
->ac
.i8_0
, "");
956 LLVMValueRef es_mask
= ac_get_i1_sgpr_mask(&ctx
->ac
, es_accepted_bool
);
958 LLVMBuildStore(builder
, es_accepted_bool
, es_accepted
);
960 ac_build_ifcc(&ctx
->ac
, LLVMBuildICmp(builder
, LLVMIntEQ
, tid
, ctx
->ac
.i32_0
, ""), 16008);
962 LLVMBuildStore(builder
, es_mask
,
963 ac_build_gep0(&ctx
->ac
, ngg_scratch
, get_wave_id_in_tg(ctx
)));
965 ac_build_endif(&ctx
->ac
, 16008);
967 ac_build_endif(&ctx
->ac
, 16007);
968 ac_build_s_barrier(&ctx
->ac
);
970 /* Load the vertex masks and compute the new ES thread count. */
971 LLVMValueRef es_mask
[2], new_num_es_threads
, kill_wave
;
972 load_bitmasks_2x64(ctx
, ngg_scratch
, 0, es_mask
, &new_num_es_threads
);
973 new_num_es_threads
= ac_build_readlane_no_opt_barrier(&ctx
->ac
, new_num_es_threads
, NULL
);
975 /* ES threads compute their prefix sum, which is the new ES thread ID.
976 * Then they write the value of the old thread ID into the LDS address
977 * of the new thread ID. It will be used it to load input VGPRs from
978 * the old thread's LDS location.
980 ac_build_ifcc(&ctx
->ac
, LLVMBuildLoad(builder
, es_accepted
, ""), 16009);
982 LLVMValueRef old_id
= get_thread_id_in_tg(ctx
);
983 LLVMValueRef new_id
= ac_prefix_bitcount_2x64(&ctx
->ac
, es_mask
, old_id
);
986 builder
, LLVMBuildTrunc(builder
, old_id
, ctx
->ac
.i8
, ""),
987 si_build_gep_i8(ctx
, ngg_nogs_vertex_ptr(ctx
, new_id
), lds_byte0_old_thread_id
));
988 LLVMBuildStore(builder
, LLVMBuildTrunc(builder
, new_id
, ctx
->ac
.i8
, ""),
989 si_build_gep_i8(ctx
, es_vtxptr
, lds_byte1_new_thread_id
));
991 ac_build_endif(&ctx
->ac
, 16009);
993 /* Kill waves that have inactive threads. */
994 kill_wave
= LLVMBuildICmp(builder
, LLVMIntULE
,
995 ac_build_imax(&ctx
->ac
, new_num_es_threads
, ngg_get_prim_cnt(ctx
)),
996 LLVMBuildMul(builder
, get_wave_id_in_tg(ctx
),
997 LLVMConstInt(ctx
->ac
.i32
, ctx
->ac
.wave_size
, 0), ""),
999 ac_build_ifcc(&ctx
->ac
, kill_wave
, 19202);
1001 /* If we are killing wave 0, send that there are no primitives
1002 * in this threadgroup.
1004 ac_build_sendmsg_gs_alloc_req(&ctx
->ac
, get_wave_id_in_tg(ctx
), ctx
->ac
.i32_0
, ctx
->ac
.i32_0
);
1005 ac_build_s_endpgm(&ctx
->ac
);
1007 ac_build_endif(&ctx
->ac
, 19202);
1008 ac_build_s_barrier(&ctx
->ac
);
1010 /* Send the final vertex and primitive counts. */
1011 ac_build_sendmsg_gs_alloc_req(&ctx
->ac
, get_wave_id_in_tg(ctx
), new_num_es_threads
,
1012 ngg_get_prim_cnt(ctx
));
1014 /* Update thread counts in SGPRs. */
1015 LLVMValueRef new_gs_tg_info
= ac_get_arg(&ctx
->ac
, ctx
->gs_tg_info
);
1016 LLVMValueRef new_merged_wave_info
= ac_get_arg(&ctx
->ac
, ctx
->merged_wave_info
);
1018 /* This also converts the thread count from the total count to the per-wave count. */
1019 update_thread_counts(ctx
, &new_num_es_threads
, &new_gs_tg_info
, 9, 12, &new_merged_wave_info
, 8,
1022 /* Update vertex indices in VGPR0 (same format as NGG passthrough). */
1023 LLVMValueRef new_vgpr0
= ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i32
, "");
1025 /* Set the null flag at the beginning (culled), and then
1026 * overwrite it for accepted primitives.
1028 LLVMBuildStore(builder
, LLVMConstInt(ctx
->ac
.i32
, 1u << 31, 0), new_vgpr0
);
1030 /* Get vertex indices after vertex compaction. */
1031 ac_build_ifcc(&ctx
->ac
, LLVMBuildTrunc(builder
, gs_accepted
, ctx
->ac
.i1
, ""), 16011);
1033 struct ac_ngg_prim prim
= {};
1034 prim
.num_vertices
= 3;
1035 prim
.isnull
= ctx
->ac
.i1false
;
1037 for (unsigned vtx
= 0; vtx
< 3; vtx
++) {
1038 prim
.index
[vtx
] = LLVMBuildLoad(
1039 builder
, si_build_gep_i8(ctx
, gs_vtxptr
[vtx
], lds_byte1_new_thread_id
), "");
1040 prim
.index
[vtx
] = LLVMBuildZExt(builder
, prim
.index
[vtx
], ctx
->ac
.i32
, "");
1041 prim
.edgeflag
[vtx
] = ngg_get_initial_edgeflag(ctx
, vtx
);
1044 /* Set the new GS input VGPR. */
1045 LLVMBuildStore(builder
, ac_pack_prim_export(&ctx
->ac
, &prim
), new_vgpr0
);
1047 ac_build_endif(&ctx
->ac
, 16011);
1049 if (gfx10_ngg_export_prim_early(shader
))
1050 gfx10_ngg_build_export_prim(ctx
, NULL
, LLVMBuildLoad(builder
, new_vgpr0
, ""));
1052 /* Set the new ES input VGPRs. */
1053 LLVMValueRef es_data
[4];
1054 LLVMValueRef old_thread_id
= ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i32
, "");
1056 for (unsigned i
= 0; i
< 4; i
++)
1057 es_data
[i
] = ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i32
, "");
1059 ac_build_ifcc(&ctx
->ac
, LLVMBuildICmp(ctx
->ac
.builder
, LLVMIntULT
, tid
, new_num_es_threads
, ""),
1062 LLVMValueRef old_id
, old_es_vtxptr
, tmp
;
1064 /* Load ES input VGPRs from the ES thread before compaction. */
1065 old_id
= LLVMBuildLoad(builder
, si_build_gep_i8(ctx
, es_vtxptr
, lds_byte0_old_thread_id
), "");
1066 old_id
= LLVMBuildZExt(builder
, old_id
, ctx
->ac
.i32
, "");
1068 LLVMBuildStore(builder
, old_id
, old_thread_id
);
1069 old_es_vtxptr
= ngg_nogs_vertex_ptr(ctx
, old_id
);
1071 for (unsigned i
= 0; i
< 2; i
++) {
1072 tmp
= LLVMBuildLoad(
1074 ac_build_gep0(&ctx
->ac
, old_es_vtxptr
, LLVMConstInt(ctx
->ac
.i32
, lds_vertex_id
+ i
, 0)),
1076 LLVMBuildStore(builder
, tmp
, es_data
[i
]);
1079 if (ctx
->stage
== MESA_SHADER_TESS_EVAL
) {
1080 tmp
= LLVMBuildLoad(builder
,
1081 si_build_gep_i8(ctx
, old_es_vtxptr
, lds_byte2_tes_rel_patch_id
), "");
1082 tmp
= LLVMBuildZExt(builder
, tmp
, ctx
->ac
.i32
, "");
1083 LLVMBuildStore(builder
, tmp
, es_data
[2]);
1085 if (uses_tes_prim_id
) {
1086 tmp
= LLVMBuildLoad(builder
,
1087 ac_build_gep0(&ctx
->ac
, old_es_vtxptr
,
1088 LLVMConstInt(ctx
->ac
.i32
, lds_tes_patch_id
, 0)),
1090 LLVMBuildStore(builder
, tmp
, es_data
[3]);
1094 ac_build_endif(&ctx
->ac
, 16012);
1096 /* Return values for the main function. */
1097 LLVMValueRef ret
= ctx
->return_value
;
1100 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, new_gs_tg_info
, 2, "");
1101 ret
= LLVMBuildInsertValue(ctx
->ac
.builder
, ret
, new_merged_wave_info
, 3, "");
1102 if (ctx
->stage
== MESA_SHADER_TESS_EVAL
)
1103 ret
= si_insert_input_ret(ctx
, ret
, ctx
->tcs_offchip_offset
, 4);
1105 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->rw_buffers
, 8 + SI_SGPR_RW_BUFFERS
);
1106 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->bindless_samplers_and_images
,
1107 8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES
);
1108 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->const_and_shader_buffers
,
1109 8 + SI_SGPR_CONST_AND_SHADER_BUFFERS
);
1110 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->samplers_and_images
, 8 + SI_SGPR_SAMPLERS_AND_IMAGES
);
1111 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->vs_state_bits
, 8 + SI_SGPR_VS_STATE_BITS
);
1113 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1114 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->args
.base_vertex
, 8 + SI_SGPR_BASE_VERTEX
);
1115 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->args
.start_instance
, 8 + SI_SGPR_START_INSTANCE
);
1116 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->args
.draw_id
, 8 + SI_SGPR_DRAWID
);
1117 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->vertex_buffers
, 8 + SI_VS_NUM_USER_SGPR
);
1119 for (unsigned i
= 0; i
< shader
->selector
->num_vbos_in_user_sgprs
; i
++) {
1120 ret
= si_insert_input_v4i32(ctx
, ret
, ctx
->vb_descriptors
[i
],
1121 8 + SI_SGPR_VS_VB_DESCRIPTOR_FIRST
+ i
* 4);
1124 assert(ctx
->stage
== MESA_SHADER_TESS_EVAL
);
1125 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->tcs_offchip_layout
, 8 + SI_SGPR_TES_OFFCHIP_LAYOUT
);
1126 ret
= si_insert_input_ptr(ctx
, ret
, ctx
->tes_offchip_addr
, 8 + SI_SGPR_TES_OFFCHIP_ADDR
);
1130 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1131 if (shader
->selector
->num_vbos_in_user_sgprs
) {
1132 vgpr
= 8 + SI_SGPR_VS_VB_DESCRIPTOR_FIRST
+ shader
->selector
->num_vbos_in_user_sgprs
* 4;
1134 vgpr
= 8 + GFX9_VSGS_NUM_USER_SGPR
+ 1;
1137 vgpr
= 8 + GFX9_TESGS_NUM_USER_SGPR
;
1140 val
= LLVMBuildLoad(builder
, new_vgpr0
, "");
1141 ret
= LLVMBuildInsertValue(builder
, ret
, ac_to_float(&ctx
->ac
, val
), vgpr
++, "");
1142 vgpr
++; /* gs_vtx23_offset */
1144 ret
= si_insert_input_ret_float(ctx
, ret
, ctx
->args
.gs_prim_id
, vgpr
++);
1145 ret
= si_insert_input_ret_float(ctx
, ret
, ctx
->args
.gs_invocation_id
, vgpr
++);
1146 vgpr
++; /* gs_vtx45_offset */
1148 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1149 val
= LLVMBuildLoad(builder
, es_data
[0], "");
1150 ret
= LLVMBuildInsertValue(builder
, ret
, ac_to_float(&ctx
->ac
, val
), vgpr
++,
1151 ""); /* VGPR5 - VertexID */
1153 if (uses_instance_id
) {
1154 val
= LLVMBuildLoad(builder
, es_data
[1], "");
1155 ret
= LLVMBuildInsertValue(builder
, ret
, ac_to_float(&ctx
->ac
, val
), vgpr
++,
1156 ""); /* VGPR8 - InstanceID */
1161 assert(ctx
->stage
== MESA_SHADER_TESS_EVAL
);
1162 unsigned num_vgprs
= uses_tes_prim_id
? 4 : 3;
1163 for (unsigned i
= 0; i
< num_vgprs
; i
++) {
1164 val
= LLVMBuildLoad(builder
, es_data
[i
], "");
1165 ret
= LLVMBuildInsertValue(builder
, ret
, ac_to_float(&ctx
->ac
, val
), vgpr
++, "");
1170 /* Return the old thread ID. */
1171 val
= LLVMBuildLoad(builder
, old_thread_id
, "");
1172 ret
= LLVMBuildInsertValue(builder
, ret
, ac_to_float(&ctx
->ac
, val
), vgpr
++, "");
1174 /* These two also use LDS. */
1175 if (sel
->info
.writes_edgeflag
||
1176 (ctx
->stage
== MESA_SHADER_VERTEX
&& shader
->key
.mono
.u
.vs_export_prim_id
))
1177 ac_build_s_barrier(&ctx
->ac
);
1179 ctx
->return_value
= ret
;
1183 * Emit the epilogue of an API VS or TES shader compiled as ESGS shader.
1185 void gfx10_emit_ngg_epilogue(struct ac_shader_abi
*abi
, unsigned max_outputs
, LLVMValueRef
*addrs
)
1187 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1188 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1189 struct si_shader_info
*info
= &sel
->info
;
1190 struct si_shader_output_values outputs
[PIPE_MAX_SHADER_OUTPUTS
];
1191 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1192 LLVMValueRef tmp
, tmp2
;
1194 assert(!ctx
->shader
->is_gs_copy_shader
);
1195 assert(info
->num_outputs
<= max_outputs
);
1197 LLVMValueRef vertex_ptr
= NULL
;
1199 if (sel
->so
.num_outputs
|| sel
->info
.writes_edgeflag
)
1200 vertex_ptr
= ngg_nogs_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
));
1202 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
1203 outputs
[i
].semantic
= info
->output_semantic
[i
];
1205 for (unsigned j
= 0; j
< 4; j
++) {
1206 outputs
[i
].vertex_stream
[j
] = (info
->output_streams
[i
] >> (2 * j
)) & 3;
1208 /* TODO: we may store more outputs than streamout needs,
1209 * but streamout performance isn't that important.
1211 if (sel
->so
.num_outputs
) {
1212 tmp
= ac_build_gep0(&ctx
->ac
, vertex_ptr
, LLVMConstInt(ctx
->ac
.i32
, 4 * i
+ j
, false));
1213 tmp2
= LLVMBuildLoad(builder
, addrs
[4 * i
+ j
], "");
1214 tmp2
= ac_to_integer(&ctx
->ac
, tmp2
);
1215 LLVMBuildStore(builder
, tmp2
, tmp
);
1219 /* Store the edgeflag at the end (if streamout is enabled) */
1220 if (info
->output_semantic
[i
] == VARYING_SLOT_EDGE
&& sel
->info
.writes_edgeflag
) {
1221 LLVMValueRef edgeflag
= LLVMBuildLoad(builder
, addrs
[4 * i
], "");
1222 /* The output is a float, but the hw expects a 1-bit integer. */
1223 edgeflag
= LLVMBuildFPToUI(ctx
->ac
.builder
, edgeflag
, ctx
->ac
.i32
, "");
1224 edgeflag
= ac_build_umin(&ctx
->ac
, edgeflag
, ctx
->ac
.i32_1
);
1226 tmp
= LLVMConstInt(ctx
->ac
.i32
, ngg_nogs_vertex_size(ctx
->shader
) - 1, 0);
1227 tmp
= ac_build_gep0(&ctx
->ac
, vertex_ptr
, tmp
);
1228 LLVMBuildStore(builder
, edgeflag
, tmp
);
1232 bool unterminated_es_if_block
=
1233 !sel
->so
.num_outputs
&& !sel
->info
.writes_edgeflag
&&
1234 !ctx
->screen
->use_ngg_streamout
&& /* no query buffer */
1235 (ctx
->stage
!= MESA_SHADER_VERTEX
|| !ctx
->shader
->key
.mono
.u
.vs_export_prim_id
);
1237 if (!unterminated_es_if_block
)
1238 ac_build_endif(&ctx
->ac
, ctx
->merged_wrap_if_label
);
1240 LLVMValueRef is_gs_thread
= si_is_gs_thread(ctx
);
1241 LLVMValueRef is_es_thread
= si_is_es_thread(ctx
);
1242 LLVMValueRef vtxindex
[3];
1244 if (ctx
->shader
->key
.opt
.ngg_culling
) {
1245 vtxindex
[0] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 0, 9);
1246 vtxindex
[1] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 10, 9);
1247 vtxindex
[2] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 20, 9);
1249 vtxindex
[0] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 0, 16);
1250 vtxindex
[1] = si_unpack_param(ctx
, ctx
->gs_vtx01_offset
, 16, 16);
1251 vtxindex
[2] = si_unpack_param(ctx
, ctx
->gs_vtx23_offset
, 0, 16);
1254 /* Determine the number of vertices per primitive. */
1255 unsigned num_vertices
;
1256 LLVMValueRef num_vertices_val
= ngg_get_vertices_per_prim(ctx
, &num_vertices
);
1259 LLVMValueRef emitted_prims
= NULL
;
1261 if (sel
->so
.num_outputs
) {
1262 assert(!unterminated_es_if_block
);
1264 struct ngg_streamout nggso
= {};
1265 nggso
.num_vertices
= num_vertices_val
;
1266 nggso
.prim_enable
[0] = is_gs_thread
;
1268 for (unsigned i
= 0; i
< num_vertices
; ++i
)
1269 nggso
.vertices
[i
] = ngg_nogs_vertex_ptr(ctx
, vtxindex
[i
]);
1271 build_streamout(ctx
, &nggso
);
1272 emitted_prims
= nggso
.emit
[0];
1275 LLVMValueRef user_edgeflags
[3] = {};
1277 if (sel
->info
.writes_edgeflag
) {
1278 assert(!unterminated_es_if_block
);
1280 /* Streamout already inserted the barrier, so don't insert it again. */
1281 if (!sel
->so
.num_outputs
)
1282 ac_build_s_barrier(&ctx
->ac
);
1284 ac_build_ifcc(&ctx
->ac
, is_gs_thread
, 5400);
1285 /* Load edge flags from ES threads and store them into VGPRs in GS threads. */
1286 for (unsigned i
= 0; i
< num_vertices
; i
++) {
1287 tmp
= ngg_nogs_vertex_ptr(ctx
, vtxindex
[i
]);
1288 tmp2
= LLVMConstInt(ctx
->ac
.i32
, ngg_nogs_vertex_size(ctx
->shader
) - 1, 0);
1289 tmp
= ac_build_gep0(&ctx
->ac
, tmp
, tmp2
);
1290 tmp
= LLVMBuildLoad(builder
, tmp
, "");
1291 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
1293 user_edgeflags
[i
] = ac_build_alloca_undef(&ctx
->ac
, ctx
->ac
.i1
, "");
1294 LLVMBuildStore(builder
, tmp
, user_edgeflags
[i
]);
1296 ac_build_endif(&ctx
->ac
, 5400);
1299 /* Copy Primitive IDs from GS threads to the LDS address corresponding
1300 * to the ES thread of the provoking vertex.
1302 if (ctx
->stage
== MESA_SHADER_VERTEX
&& ctx
->shader
->key
.mono
.u
.vs_export_prim_id
) {
1303 assert(!unterminated_es_if_block
);
1305 /* Streamout and edge flags use LDS. Make it idle, so that we can reuse it. */
1306 if (sel
->so
.num_outputs
|| sel
->info
.writes_edgeflag
)
1307 ac_build_s_barrier(&ctx
->ac
);
1309 ac_build_ifcc(&ctx
->ac
, is_gs_thread
, 5400);
1310 /* Extract the PROVOKING_VTX_INDEX field. */
1311 LLVMValueRef provoking_vtx_in_prim
= si_unpack_param(ctx
, ctx
->vs_state_bits
, 4, 2);
1313 /* provoking_vtx_index = vtxindex[provoking_vtx_in_prim]; */
1314 LLVMValueRef indices
= ac_build_gather_values(&ctx
->ac
, vtxindex
, 3);
1315 LLVMValueRef provoking_vtx_index
=
1316 LLVMBuildExtractElement(builder
, indices
, provoking_vtx_in_prim
, "");
1317 LLVMValueRef vertex_ptr
= ngg_nogs_vertex_ptr(ctx
, provoking_vtx_index
);
1319 LLVMBuildStore(builder
, ac_get_arg(&ctx
->ac
, ctx
->args
.gs_prim_id
),
1320 ac_build_gep0(&ctx
->ac
, vertex_ptr
, ctx
->ac
.i32_0
));
1321 ac_build_endif(&ctx
->ac
, 5400);
1324 /* Update query buffer */
1325 if (ctx
->screen
->use_ngg_streamout
&& !info
->base
.vs
.blit_sgprs_amd
) {
1326 assert(!unterminated_es_if_block
);
1328 tmp
= si_unpack_param(ctx
, ctx
->vs_state_bits
, 6, 1);
1329 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
1330 ac_build_ifcc(&ctx
->ac
, tmp
, 5029); /* if (STREAMOUT_QUERY_ENABLED) */
1331 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, get_wave_id_in_tg(ctx
), ctx
->ac
.i32_0
, "");
1332 ac_build_ifcc(&ctx
->ac
, tmp
, 5030);
1333 tmp
= LLVMBuildICmp(builder
, LLVMIntULE
, ac_get_thread_id(&ctx
->ac
),
1334 sel
->so
.num_outputs
? ctx
->ac
.i32_1
: ctx
->ac
.i32_0
, "");
1335 ac_build_ifcc(&ctx
->ac
, tmp
, 5031);
1337 LLVMValueRef args
[] = {
1338 ngg_get_prim_cnt(ctx
),
1339 ngg_get_query_buf(ctx
),
1340 LLVMConstInt(ctx
->ac
.i32
, 16, false), /* offset of stream[0].generated_primitives */
1341 ctx
->ac
.i32_0
, /* soffset */
1342 ctx
->ac
.i32_0
, /* cachepolicy */
1345 if (sel
->so
.num_outputs
) {
1346 args
[0] = ac_build_writelane(&ctx
->ac
, args
[0], emitted_prims
, ctx
->ac
.i32_1
);
1347 args
[2] = ac_build_writelane(&ctx
->ac
, args
[2], LLVMConstInt(ctx
->ac
.i32
, 24, false),
1351 /* TODO: should this be 64-bit atomics? */
1352 ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.raw.buffer.atomic.add.i32", ctx
->ac
.i32
, args
, 5,
1355 ac_build_endif(&ctx
->ac
, 5031);
1356 ac_build_endif(&ctx
->ac
, 5030);
1357 ac_build_endif(&ctx
->ac
, 5029);
1360 /* Build the primitive export. */
1361 if (!gfx10_ngg_export_prim_early(ctx
->shader
)) {
1362 assert(!unterminated_es_if_block
);
1363 gfx10_ngg_build_export_prim(ctx
, user_edgeflags
, NULL
);
1366 /* Export per-vertex data (positions and parameters). */
1367 if (!unterminated_es_if_block
)
1368 ac_build_ifcc(&ctx
->ac
, is_es_thread
, 6002);
1372 /* Unconditionally (re-)load the values for proper SSA form. */
1373 for (i
= 0; i
< info
->num_outputs
; i
++) {
1374 /* If the NGG cull shader part computed the position, don't
1375 * use the position from the current shader part. Instead,
1378 if (info
->output_semantic
[i
] == VARYING_SLOT_POS
&&
1379 ctx
->shader
->key
.opt
.ngg_culling
) {
1380 vertex_ptr
= ngg_nogs_vertex_ptr(ctx
, ac_get_arg(&ctx
->ac
, ctx
->ngg_old_thread_id
));
1382 for (unsigned j
= 0; j
< 4; j
++) {
1383 tmp
= LLVMConstInt(ctx
->ac
.i32
, lds_pos_x
+ j
, 0);
1384 tmp
= ac_build_gep0(&ctx
->ac
, vertex_ptr
, tmp
);
1385 tmp
= LLVMBuildLoad(builder
, tmp
, "");
1386 outputs
[i
].values
[j
] = ac_to_float(&ctx
->ac
, tmp
);
1389 for (unsigned j
= 0; j
< 4; j
++) {
1390 outputs
[i
].values
[j
] = LLVMBuildLoad(builder
, addrs
[4 * i
+ j
], "");
1395 if (ctx
->shader
->key
.mono
.u
.vs_export_prim_id
) {
1396 outputs
[i
].semantic
= VARYING_SLOT_PRIMITIVE_ID
;
1398 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1399 /* Wait for GS stores to finish. */
1400 ac_build_s_barrier(&ctx
->ac
);
1402 tmp
= ngg_nogs_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
));
1403 tmp
= ac_build_gep0(&ctx
->ac
, tmp
, ctx
->ac
.i32_0
);
1404 outputs
[i
].values
[0] = LLVMBuildLoad(builder
, tmp
, "");
1406 assert(ctx
->stage
== MESA_SHADER_TESS_EVAL
);
1407 outputs
[i
].values
[0] = si_get_primitive_id(ctx
, 0);
1410 outputs
[i
].values
[0] = ac_to_float(&ctx
->ac
, outputs
[i
].values
[0]);
1411 for (unsigned j
= 1; j
< 4; j
++)
1412 outputs
[i
].values
[j
] = LLVMGetUndef(ctx
->ac
.f32
);
1414 memset(outputs
[i
].vertex_stream
, 0, sizeof(outputs
[i
].vertex_stream
));
1418 si_llvm_build_vs_exports(ctx
, outputs
, i
);
1420 ac_build_endif(&ctx
->ac
, 6002);
1423 static LLVMValueRef
ngg_gs_get_vertex_storage(struct si_shader_context
*ctx
)
1425 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1426 const struct si_shader_info
*info
= &sel
->info
;
1428 LLVMTypeRef elements
[2] = {
1429 LLVMArrayType(ctx
->ac
.i32
, 4 * info
->num_outputs
),
1430 LLVMArrayType(ctx
->ac
.i8
, 4),
1432 LLVMTypeRef type
= LLVMStructTypeInContext(ctx
->ac
.context
, elements
, 2, false);
1433 type
= LLVMPointerType(LLVMArrayType(type
, 0), AC_ADDR_SPACE_LDS
);
1434 return LLVMBuildBitCast(ctx
->ac
.builder
, ctx
->gs_ngg_emit
, type
, "");
1438 * Return a pointer to the LDS storage reserved for the N'th vertex, where N
1439 * is in emit order; that is:
1440 * - during the epilogue, N is the threadidx (relative to the entire threadgroup)
1441 * - during vertex emit, i.e. while the API GS shader invocation is running,
1442 * N = threadidx * gs_max_out_vertices + emitidx
1444 * Goals of the LDS memory layout:
1445 * 1. Eliminate bank conflicts on write for geometry shaders that have all emits
1446 * in uniform control flow
1447 * 2. Eliminate bank conflicts on read for export if, additionally, there is no
1449 * 3. Agnostic to the number of waves (since we don't know it before compiling)
1450 * 4. Allow coalescing of LDS instructions (ds_write_b128 etc.)
1451 * 5. Avoid wasting memory.
1453 * We use an AoS layout due to point 4 (this also helps point 3). In an AoS
1454 * layout, elimination of bank conflicts requires that each vertex occupy an
1455 * odd number of dwords. We use the additional dword to store the output stream
1456 * index as well as a flag to indicate whether this vertex ends a primitive
1457 * for rasterization.
1459 * Swizzling is required to satisfy points 1 and 2 simultaneously.
1461 * Vertices are stored in export order (gsthread * gs_max_out_vertices + emitidx).
1462 * Indices are swizzled in groups of 32, which ensures point 1 without
1463 * disturbing point 2.
1465 * \return an LDS pointer to type {[N x i32], [4 x i8]}
1467 static LLVMValueRef
ngg_gs_vertex_ptr(struct si_shader_context
*ctx
, LLVMValueRef vertexidx
)
1469 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1470 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1471 LLVMValueRef storage
= ngg_gs_get_vertex_storage(ctx
);
1473 /* gs_max_out_vertices = 2^(write_stride_2exp) * some odd number */
1474 unsigned write_stride_2exp
= ffs(sel
->gs_max_out_vertices
) - 1;
1475 if (write_stride_2exp
) {
1476 LLVMValueRef row
= LLVMBuildLShr(builder
, vertexidx
, LLVMConstInt(ctx
->ac
.i32
, 5, false), "");
1477 LLVMValueRef swizzle
= LLVMBuildAnd(
1478 builder
, row
, LLVMConstInt(ctx
->ac
.i32
, (1u << write_stride_2exp
) - 1, false), "");
1479 vertexidx
= LLVMBuildXor(builder
, vertexidx
, swizzle
, "");
1482 return ac_build_gep0(&ctx
->ac
, storage
, vertexidx
);
1485 static LLVMValueRef
ngg_gs_emit_vertex_ptr(struct si_shader_context
*ctx
, LLVMValueRef gsthread
,
1486 LLVMValueRef emitidx
)
1488 struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1489 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1492 tmp
= LLVMConstInt(ctx
->ac
.i32
, sel
->gs_max_out_vertices
, false);
1493 tmp
= LLVMBuildMul(builder
, tmp
, gsthread
, "");
1494 const LLVMValueRef vertexidx
= LLVMBuildAdd(builder
, tmp
, emitidx
, "");
1495 return ngg_gs_vertex_ptr(ctx
, vertexidx
);
1498 static LLVMValueRef
ngg_gs_get_emit_output_ptr(struct si_shader_context
*ctx
,
1499 LLVMValueRef vertexptr
, unsigned out_idx
)
1501 LLVMValueRef gep_idx
[3] = {
1502 ctx
->ac
.i32_0
, /* implied C-style array */
1503 ctx
->ac
.i32_0
, /* first struct entry */
1504 LLVMConstInt(ctx
->ac
.i32
, out_idx
, false),
1506 return LLVMBuildGEP(ctx
->ac
.builder
, vertexptr
, gep_idx
, 3, "");
1509 static LLVMValueRef
ngg_gs_get_emit_primflag_ptr(struct si_shader_context
*ctx
,
1510 LLVMValueRef vertexptr
, unsigned stream
)
1512 LLVMValueRef gep_idx
[3] = {
1513 ctx
->ac
.i32_0
, /* implied C-style array */
1514 ctx
->ac
.i32_1
, /* second struct entry */
1515 LLVMConstInt(ctx
->ac
.i32
, stream
, false),
1517 return LLVMBuildGEP(ctx
->ac
.builder
, vertexptr
, gep_idx
, 3, "");
1520 void gfx10_ngg_gs_emit_vertex(struct si_shader_context
*ctx
, unsigned stream
, LLVMValueRef
*addrs
)
1522 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1523 const struct si_shader_info
*info
= &sel
->info
;
1524 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1526 const LLVMValueRef vertexidx
= LLVMBuildLoad(builder
, ctx
->gs_next_vertex
[stream
], "");
1528 /* If this thread has already emitted the declared maximum number of
1529 * vertices, skip the write: excessive vertex emissions are not
1530 * supposed to have any effect.
1532 const LLVMValueRef can_emit
=
1533 LLVMBuildICmp(builder
, LLVMIntULT
, vertexidx
,
1534 LLVMConstInt(ctx
->ac
.i32
, sel
->gs_max_out_vertices
, false), "");
1536 tmp
= LLVMBuildAdd(builder
, vertexidx
, ctx
->ac
.i32_1
, "");
1537 tmp
= LLVMBuildSelect(builder
, can_emit
, tmp
, vertexidx
, "");
1538 LLVMBuildStore(builder
, tmp
, ctx
->gs_next_vertex
[stream
]);
1540 ac_build_ifcc(&ctx
->ac
, can_emit
, 9001);
1542 const LLVMValueRef vertexptr
= ngg_gs_emit_vertex_ptr(ctx
, get_thread_id_in_tg(ctx
), vertexidx
);
1543 unsigned out_idx
= 0;
1544 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
1545 for (unsigned chan
= 0; chan
< 4; chan
++, out_idx
++) {
1546 if (!(info
->output_usagemask
[i
] & (1 << chan
)) ||
1547 ((info
->output_streams
[i
] >> (2 * chan
)) & 3) != stream
)
1550 LLVMValueRef out_val
= LLVMBuildLoad(builder
, addrs
[4 * i
+ chan
], "");
1551 out_val
= ac_to_integer(&ctx
->ac
, out_val
);
1552 LLVMBuildStore(builder
, out_val
, ngg_gs_get_emit_output_ptr(ctx
, vertexptr
, out_idx
));
1555 assert(out_idx
* 4 == sel
->gsvs_vertex_size
);
1557 /* Determine and store whether this vertex completed a primitive. */
1558 const LLVMValueRef curverts
= LLVMBuildLoad(builder
, ctx
->gs_curprim_verts
[stream
], "");
1560 tmp
= LLVMConstInt(ctx
->ac
.i32
, u_vertices_per_prim(sel
->gs_output_prim
) - 1, false);
1561 const LLVMValueRef iscompleteprim
= LLVMBuildICmp(builder
, LLVMIntUGE
, curverts
, tmp
, "");
1563 /* Since the geometry shader emits triangle strips, we need to
1564 * track which primitive is odd and swap vertex indices to get
1565 * the correct vertex order.
1567 LLVMValueRef is_odd
= ctx
->ac
.i1false
;
1568 if (stream
== 0 && u_vertices_per_prim(sel
->gs_output_prim
) == 3) {
1569 tmp
= LLVMBuildAnd(builder
, curverts
, ctx
->ac
.i32_1
, "");
1570 is_odd
= LLVMBuildICmp(builder
, LLVMIntEQ
, tmp
, ctx
->ac
.i32_1
, "");
1573 tmp
= LLVMBuildAdd(builder
, curverts
, ctx
->ac
.i32_1
, "");
1574 LLVMBuildStore(builder
, tmp
, ctx
->gs_curprim_verts
[stream
]);
1576 /* The per-vertex primitive flag encoding:
1577 * bit 0: whether this vertex finishes a primitive
1578 * bit 1: whether the primitive is odd (if we are emitting triangle strips)
1580 tmp
= LLVMBuildZExt(builder
, iscompleteprim
, ctx
->ac
.i8
, "");
1583 LLVMBuildShl(builder
, LLVMBuildZExt(builder
, is_odd
, ctx
->ac
.i8
, ""), ctx
->ac
.i8_1
, ""), "");
1584 LLVMBuildStore(builder
, tmp
, ngg_gs_get_emit_primflag_ptr(ctx
, vertexptr
, stream
));
1586 tmp
= LLVMBuildLoad(builder
, ctx
->gs_generated_prims
[stream
], "");
1587 tmp
= LLVMBuildAdd(builder
, tmp
, LLVMBuildZExt(builder
, iscompleteprim
, ctx
->ac
.i32
, ""), "");
1588 LLVMBuildStore(builder
, tmp
, ctx
->gs_generated_prims
[stream
]);
1590 ac_build_endif(&ctx
->ac
, 9001);
1593 void gfx10_ngg_gs_emit_prologue(struct si_shader_context
*ctx
)
1595 /* Zero out the part of LDS scratch that is used to accumulate the
1596 * per-stream generated primitive count.
1598 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1599 LLVMValueRef scratchptr
= ctx
->gs_ngg_scratch
;
1600 LLVMValueRef tid
= get_thread_id_in_tg(ctx
);
1603 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, LLVMConstInt(ctx
->ac
.i32
, 4, false), "");
1604 ac_build_ifcc(&ctx
->ac
, tmp
, 5090);
1606 LLVMValueRef ptr
= ac_build_gep0(&ctx
->ac
, scratchptr
, tid
);
1607 LLVMBuildStore(builder
, ctx
->ac
.i32_0
, ptr
);
1609 ac_build_endif(&ctx
->ac
, 5090);
1611 ac_build_s_barrier(&ctx
->ac
);
1614 void gfx10_ngg_gs_emit_epilogue(struct si_shader_context
*ctx
)
1616 const struct si_shader_selector
*sel
= ctx
->shader
->selector
;
1617 const struct si_shader_info
*info
= &sel
->info
;
1618 const unsigned verts_per_prim
= u_vertices_per_prim(sel
->gs_output_prim
);
1619 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1620 LLVMValueRef i8_0
= LLVMConstInt(ctx
->ac
.i8
, 0, false);
1621 LLVMValueRef tmp
, tmp2
;
1623 /* Zero out remaining (non-emitted) primitive flags.
1625 * Note: Alternatively, we could pass the relevant gs_next_vertex to
1626 * the emit threads via LDS. This is likely worse in the expected
1627 * typical case where each GS thread emits the full set of
1630 for (unsigned stream
= 0; stream
< 4; ++stream
) {
1631 if (!info
->num_stream_output_components
[stream
])
1634 const LLVMValueRef gsthread
= get_thread_id_in_tg(ctx
);
1636 ac_build_bgnloop(&ctx
->ac
, 5100);
1638 const LLVMValueRef vertexidx
= LLVMBuildLoad(builder
, ctx
->gs_next_vertex
[stream
], "");
1639 tmp
= LLVMBuildICmp(builder
, LLVMIntUGE
, vertexidx
,
1640 LLVMConstInt(ctx
->ac
.i32
, sel
->gs_max_out_vertices
, false), "");
1641 ac_build_ifcc(&ctx
->ac
, tmp
, 5101);
1642 ac_build_break(&ctx
->ac
);
1643 ac_build_endif(&ctx
->ac
, 5101);
1645 tmp
= LLVMBuildAdd(builder
, vertexidx
, ctx
->ac
.i32_1
, "");
1646 LLVMBuildStore(builder
, tmp
, ctx
->gs_next_vertex
[stream
]);
1648 tmp
= ngg_gs_emit_vertex_ptr(ctx
, gsthread
, vertexidx
);
1649 LLVMBuildStore(builder
, i8_0
, ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, stream
));
1651 ac_build_endloop(&ctx
->ac
, 5100);
1654 /* Accumulate generated primitives counts across the entire threadgroup. */
1655 for (unsigned stream
= 0; stream
< 4; ++stream
) {
1656 if (!info
->num_stream_output_components
[stream
])
1659 LLVMValueRef numprims
= LLVMBuildLoad(builder
, ctx
->gs_generated_prims
[stream
], "");
1660 numprims
= ac_build_reduce(&ctx
->ac
, numprims
, nir_op_iadd
, ctx
->ac
.wave_size
);
1662 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, ac_get_thread_id(&ctx
->ac
), ctx
->ac
.i32_0
, "");
1663 ac_build_ifcc(&ctx
->ac
, tmp
, 5105);
1666 builder
, LLVMAtomicRMWBinOpAdd
,
1667 ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, LLVMConstInt(ctx
->ac
.i32
, stream
, false)),
1668 numprims
, LLVMAtomicOrderingMonotonic
, false);
1670 ac_build_endif(&ctx
->ac
, 5105);
1673 ac_build_endif(&ctx
->ac
, ctx
->merged_wrap_if_label
);
1675 ac_build_s_barrier(&ctx
->ac
);
1677 const LLVMValueRef tid
= get_thread_id_in_tg(ctx
);
1678 LLVMValueRef num_emit_threads
= ngg_get_prim_cnt(ctx
);
1681 if (sel
->so
.num_outputs
) {
1682 struct ngg_streamout nggso
= {};
1684 nggso
.num_vertices
= LLVMConstInt(ctx
->ac
.i32
, verts_per_prim
, false);
1686 LLVMValueRef vertexptr
= ngg_gs_vertex_ptr(ctx
, tid
);
1687 for (unsigned stream
= 0; stream
< 4; ++stream
) {
1688 if (!info
->num_stream_output_components
[stream
])
1691 tmp
= LLVMBuildLoad(builder
, ngg_gs_get_emit_primflag_ptr(ctx
, vertexptr
, stream
), "");
1692 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
1693 tmp2
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, num_emit_threads
, "");
1694 nggso
.prim_enable
[stream
] = LLVMBuildAnd(builder
, tmp
, tmp2
, "");
1697 for (unsigned i
= 0; i
< verts_per_prim
; ++i
) {
1698 tmp
= LLVMBuildSub(builder
, tid
, LLVMConstInt(ctx
->ac
.i32
, verts_per_prim
- i
- 1, false),
1700 tmp
= ngg_gs_vertex_ptr(ctx
, tmp
);
1701 nggso
.vertices
[i
] = ac_build_gep0(&ctx
->ac
, tmp
, ctx
->ac
.i32_0
);
1704 build_streamout(ctx
, &nggso
);
1707 /* Write shader query data. */
1708 if (ctx
->screen
->use_ngg_streamout
) {
1709 tmp
= si_unpack_param(ctx
, ctx
->vs_state_bits
, 6, 1);
1710 tmp
= LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
1711 ac_build_ifcc(&ctx
->ac
, tmp
, 5109); /* if (STREAMOUT_QUERY_ENABLED) */
1712 unsigned num_query_comps
= sel
->so
.num_outputs
? 8 : 4;
1713 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
,
1714 LLVMConstInt(ctx
->ac
.i32
, num_query_comps
, false), "");
1715 ac_build_ifcc(&ctx
->ac
, tmp
, 5110);
1717 LLVMValueRef offset
;
1719 if (sel
->so
.num_outputs
)
1720 tmp
= LLVMBuildAnd(builder
, tmp
, LLVMConstInt(ctx
->ac
.i32
, 3, false), "");
1721 offset
= LLVMBuildNUWMul(builder
, tmp
, LLVMConstInt(ctx
->ac
.i32
, 32, false), "");
1722 if (sel
->so
.num_outputs
) {
1723 tmp
= LLVMBuildLShr(builder
, tid
, LLVMConstInt(ctx
->ac
.i32
, 2, false), "");
1724 tmp
= LLVMBuildNUWMul(builder
, tmp
, LLVMConstInt(ctx
->ac
.i32
, 8, false), "");
1725 offset
= LLVMBuildAdd(builder
, offset
, tmp
, "");
1728 tmp
= LLVMBuildLoad(builder
, ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, tid
), "");
1729 LLVMValueRef args
[] = {
1730 tmp
, ngg_get_query_buf(ctx
),
1731 offset
, LLVMConstInt(ctx
->ac
.i32
, 16, false), /* soffset */
1732 ctx
->ac
.i32_0
, /* cachepolicy */
1734 ac_build_intrinsic(&ctx
->ac
, "llvm.amdgcn.raw.buffer.atomic.add.i32", ctx
->ac
.i32
, args
, 5,
1737 ac_build_endif(&ctx
->ac
, 5110);
1738 ac_build_endif(&ctx
->ac
, 5109);
1741 /* Determine vertex liveness. */
1742 LLVMValueRef vertliveptr
= ac_build_alloca(&ctx
->ac
, ctx
->ac
.i1
, "vertexlive");
1744 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, num_emit_threads
, "");
1745 ac_build_ifcc(&ctx
->ac
, tmp
, 5120);
1747 for (unsigned i
= 0; i
< verts_per_prim
; ++i
) {
1748 const LLVMValueRef primidx
=
1749 LLVMBuildAdd(builder
, tid
, LLVMConstInt(ctx
->ac
.i32
, i
, false), "");
1752 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, primidx
, num_emit_threads
, "");
1753 ac_build_ifcc(&ctx
->ac
, tmp
, 5121 + i
);
1756 /* Load primitive liveness */
1757 tmp
= ngg_gs_vertex_ptr(ctx
, primidx
);
1758 tmp
= LLVMBuildLoad(builder
, ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, 0), "");
1759 const LLVMValueRef primlive
= LLVMBuildTrunc(builder
, tmp
, ctx
->ac
.i1
, "");
1761 tmp
= LLVMBuildLoad(builder
, vertliveptr
, "");
1762 tmp
= LLVMBuildOr(builder
, tmp
, primlive
, ""), LLVMBuildStore(builder
, tmp
, vertliveptr
);
1765 ac_build_endif(&ctx
->ac
, 5121 + i
);
1768 ac_build_endif(&ctx
->ac
, 5120);
1770 /* Inclusive scan addition across the current wave. */
1771 LLVMValueRef vertlive
= LLVMBuildLoad(builder
, vertliveptr
, "");
1772 struct ac_wg_scan vertlive_scan
= {};
1773 vertlive_scan
.op
= nir_op_iadd
;
1774 vertlive_scan
.enable_reduce
= true;
1775 vertlive_scan
.enable_exclusive
= true;
1776 vertlive_scan
.src
= vertlive
;
1777 vertlive_scan
.scratch
= ac_build_gep0(&ctx
->ac
, ctx
->gs_ngg_scratch
, ctx
->ac
.i32_0
);
1778 vertlive_scan
.waveidx
= get_wave_id_in_tg(ctx
);
1779 vertlive_scan
.numwaves
= get_tgsize(ctx
);
1780 vertlive_scan
.maxwaves
= 8;
1782 ac_build_wg_scan(&ctx
->ac
, &vertlive_scan
);
1784 /* Skip all exports (including index exports) when possible. At least on
1785 * early gfx10 revisions this is also to avoid hangs.
1787 LLVMValueRef have_exports
=
1788 LLVMBuildICmp(builder
, LLVMIntNE
, vertlive_scan
.result_reduce
, ctx
->ac
.i32_0
, "");
1789 num_emit_threads
= LLVMBuildSelect(builder
, have_exports
, num_emit_threads
, ctx
->ac
.i32_0
, "");
1791 /* Allocate export space. Send this message as early as possible, to
1792 * hide the latency of the SQ <-> SPI roundtrip.
1794 * Note: We could consider compacting primitives for export as well.
1795 * PA processes 1 non-null prim / clock, but it fetches 4 DW of
1796 * prim data per clock and skips null primitives at no additional
1797 * cost. So compacting primitives can only be beneficial when
1798 * there are 4 or more contiguous null primitives in the export
1799 * (in the common case of single-dword prim exports).
1801 ac_build_sendmsg_gs_alloc_req(&ctx
->ac
, get_wave_id_in_tg(ctx
), vertlive_scan
.result_reduce
,
1804 /* Setup the reverse vertex compaction permutation. We re-use stream 1
1805 * of the primitive liveness flags, relying on the fact that each
1806 * threadgroup can have at most 256 threads. */
1807 ac_build_ifcc(&ctx
->ac
, vertlive
, 5130);
1809 tmp
= ngg_gs_vertex_ptr(ctx
, vertlive_scan
.result_exclusive
);
1810 tmp2
= LLVMBuildTrunc(builder
, tid
, ctx
->ac
.i8
, "");
1811 LLVMBuildStore(builder
, tmp2
, ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, 1));
1813 ac_build_endif(&ctx
->ac
, 5130);
1815 ac_build_s_barrier(&ctx
->ac
);
1817 /* Export primitive data */
1818 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, num_emit_threads
, "");
1819 ac_build_ifcc(&ctx
->ac
, tmp
, 5140);
1822 struct ac_ngg_prim prim
= {};
1823 prim
.num_vertices
= verts_per_prim
;
1825 tmp
= ngg_gs_vertex_ptr(ctx
, tid
);
1826 flags
= LLVMBuildLoad(builder
, ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, 0), "");
1827 prim
.isnull
= LLVMBuildNot(builder
, LLVMBuildTrunc(builder
, flags
, ctx
->ac
.i1
, ""), "");
1829 for (unsigned i
= 0; i
< verts_per_prim
; ++i
) {
1830 prim
.index
[i
] = LLVMBuildSub(builder
, vertlive_scan
.result_exclusive
,
1831 LLVMConstInt(ctx
->ac
.i32
, verts_per_prim
- i
- 1, false), "");
1832 prim
.edgeflag
[i
] = ctx
->ac
.i1false
;
1835 /* Geometry shaders output triangle strips, but NGG expects triangles. */
1836 if (verts_per_prim
== 3) {
1837 LLVMValueRef is_odd
= LLVMBuildLShr(builder
, flags
, ctx
->ac
.i8_1
, "");
1838 is_odd
= LLVMBuildTrunc(builder
, is_odd
, ctx
->ac
.i1
, "");
1839 LLVMValueRef flatshade_first
= LLVMBuildICmp(
1840 builder
, LLVMIntEQ
, si_unpack_param(ctx
, ctx
->vs_state_bits
, 4, 2), ctx
->ac
.i32_0
, "");
1842 ac_build_triangle_strip_indices_to_triangle(&ctx
->ac
, is_odd
, flatshade_first
, prim
.index
);
1845 ac_build_export_prim(&ctx
->ac
, &prim
);
1847 ac_build_endif(&ctx
->ac
, 5140);
1849 /* Export position and parameter data */
1850 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, vertlive_scan
.result_reduce
, "");
1851 ac_build_ifcc(&ctx
->ac
, tmp
, 5145);
1853 struct si_shader_output_values outputs
[PIPE_MAX_SHADER_OUTPUTS
];
1855 tmp
= ngg_gs_vertex_ptr(ctx
, tid
);
1856 tmp
= LLVMBuildLoad(builder
, ngg_gs_get_emit_primflag_ptr(ctx
, tmp
, 1), "");
1857 tmp
= LLVMBuildZExt(builder
, tmp
, ctx
->ac
.i32
, "");
1858 const LLVMValueRef vertexptr
= ngg_gs_vertex_ptr(ctx
, tmp
);
1860 unsigned out_idx
= 0;
1861 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
1862 outputs
[i
].semantic
= info
->output_semantic
[i
];
1864 for (unsigned j
= 0; j
< 4; j
++, out_idx
++) {
1865 tmp
= ngg_gs_get_emit_output_ptr(ctx
, vertexptr
, out_idx
);
1866 tmp
= LLVMBuildLoad(builder
, tmp
, "");
1867 outputs
[i
].values
[j
] = ac_to_float(&ctx
->ac
, tmp
);
1868 outputs
[i
].vertex_stream
[j
] = (info
->output_streams
[i
] >> (2 * j
)) & 3;
1872 si_llvm_build_vs_exports(ctx
, outputs
, info
->num_outputs
);
1874 ac_build_endif(&ctx
->ac
, 5145);
1877 static void clamp_gsprims_to_esverts(unsigned *max_gsprims
, unsigned max_esverts
,
1878 unsigned min_verts_per_prim
, bool use_adjacency
)
1880 unsigned max_reuse
= max_esverts
- min_verts_per_prim
;
1883 *max_gsprims
= MIN2(*max_gsprims
, 1 + max_reuse
);
1886 unsigned gfx10_ngg_get_scratch_dw_size(struct si_shader
*shader
)
1888 const struct si_shader_selector
*sel
= shader
->selector
;
1890 if (sel
->info
.stage
== MESA_SHADER_GEOMETRY
&& sel
->so
.num_outputs
)
1897 * Determine subgroup information like maximum number of vertices and prims.
1899 * This happens before the shader is uploaded, since LDS relocations during
1900 * upload depend on the subgroup size.
1902 bool gfx10_ngg_calculate_subgroup_info(struct si_shader
*shader
)
1904 const struct si_shader_selector
*gs_sel
= shader
->selector
;
1905 const struct si_shader_selector
*es_sel
=
1906 shader
->previous_stage_sel
? shader
->previous_stage_sel
: gs_sel
;
1907 const gl_shader_stage gs_stage
= gs_sel
->info
.stage
;
1908 const unsigned gs_num_invocations
= MAX2(gs_sel
->gs_num_invocations
, 1);
1909 const unsigned input_prim
= si_get_input_prim(gs_sel
);
1910 const bool use_adjacency
=
1911 input_prim
>= PIPE_PRIM_LINES_ADJACENCY
&& input_prim
<= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
;
1912 const unsigned max_verts_per_prim
= u_vertices_per_prim(input_prim
);
1913 const unsigned min_verts_per_prim
= gs_stage
== MESA_SHADER_GEOMETRY
? max_verts_per_prim
: 1;
1915 /* All these are in dwords: */
1916 /* GE can only use 8K dwords (32KB) of LDS per workgroup.
1918 const unsigned max_lds_size
= 8 * 1024 - gfx10_ngg_get_scratch_dw_size(shader
);
1919 const unsigned target_lds_size
= max_lds_size
;
1920 unsigned esvert_lds_size
= 0;
1921 unsigned gsprim_lds_size
= 0;
1923 /* All these are per subgroup: */
1924 const unsigned min_esverts
= gs_sel
->screen
->info
.chip_class
>= GFX10_3
? 29 : 24;
1925 bool max_vert_out_per_gs_instance
= false;
1926 unsigned max_gsprims_base
= 128; /* default prim group size clamp */
1927 unsigned max_esverts_base
= 128;
1929 if (shader
->key
.opt
.ngg_culling
& SI_NGG_CULL_GS_FAST_LAUNCH_TRI_LIST
) {
1930 max_gsprims_base
= 128 / 3;
1931 max_esverts_base
= max_gsprims_base
* 3;
1932 } else if (shader
->key
.opt
.ngg_culling
& SI_NGG_CULL_GS_FAST_LAUNCH_TRI_STRIP
) {
1933 max_gsprims_base
= 126;
1934 max_esverts_base
= 128;
1937 /* Hardware has the following non-natural restrictions on the value
1938 * of GE_CNTL.VERT_GRP_SIZE based on based on the primitive type of
1940 * - at most 252 for any line input primitive type
1941 * - at most 251 for any quad input primitive type
1942 * - at most 251 for triangle strips with adjacency (this happens to
1943 * be the natural limit for triangle *lists* with adjacency)
1945 max_esverts_base
= MIN2(max_esverts_base
, 251 + max_verts_per_prim
- 1);
1947 if (gs_stage
== MESA_SHADER_GEOMETRY
) {
1948 bool force_multi_cycling
= false;
1949 unsigned max_out_verts_per_gsprim
= gs_sel
->gs_max_out_vertices
* gs_num_invocations
;
1952 if (max_out_verts_per_gsprim
<= 256 && !force_multi_cycling
) {
1953 if (max_out_verts_per_gsprim
) {
1954 max_gsprims_base
= MIN2(max_gsprims_base
, 256 / max_out_verts_per_gsprim
);
1957 /* Use special multi-cycling mode in which each GS
1958 * instance gets its own subgroup. Does not work with
1960 max_vert_out_per_gs_instance
= true;
1961 max_gsprims_base
= 1;
1962 max_out_verts_per_gsprim
= gs_sel
->gs_max_out_vertices
;
1965 esvert_lds_size
= es_sel
->esgs_itemsize
/ 4;
1966 gsprim_lds_size
= (gs_sel
->gsvs_vertex_size
/ 4 + 1) * max_out_verts_per_gsprim
;
1968 if (gsprim_lds_size
> target_lds_size
&& !force_multi_cycling
) {
1969 if (gs_sel
->tess_turns_off_ngg
|| es_sel
->info
.stage
!= MESA_SHADER_TESS_EVAL
) {
1970 force_multi_cycling
= true;
1971 goto retry_select_mode
;
1976 /* LDS size for passing data from ES to GS. */
1977 esvert_lds_size
= ngg_nogs_vertex_size(shader
);
1980 unsigned max_gsprims
= max_gsprims_base
;
1981 unsigned max_esverts
= max_esverts_base
;
1983 if (esvert_lds_size
)
1984 max_esverts
= MIN2(max_esverts
, target_lds_size
/ esvert_lds_size
);
1985 if (gsprim_lds_size
)
1986 max_gsprims
= MIN2(max_gsprims
, target_lds_size
/ gsprim_lds_size
);
1988 max_esverts
= MIN2(max_esverts
, max_gsprims
* max_verts_per_prim
);
1989 clamp_gsprims_to_esverts(&max_gsprims
, max_esverts
, min_verts_per_prim
, use_adjacency
);
1990 assert(max_esverts
>= max_verts_per_prim
&& max_gsprims
>= 1);
1992 if (esvert_lds_size
|| gsprim_lds_size
) {
1993 /* Now that we have a rough proportionality between esverts
1994 * and gsprims based on the primitive type, scale both of them
1995 * down simultaneously based on required LDS space.
1997 * We could be smarter about this if we knew how much vertex
2000 unsigned lds_total
= max_esverts
* esvert_lds_size
+ max_gsprims
* gsprim_lds_size
;
2001 if (lds_total
> target_lds_size
) {
2002 max_esverts
= max_esverts
* target_lds_size
/ lds_total
;
2003 max_gsprims
= max_gsprims
* target_lds_size
/ lds_total
;
2005 max_esverts
= MIN2(max_esverts
, max_gsprims
* max_verts_per_prim
);
2006 clamp_gsprims_to_esverts(&max_gsprims
, max_esverts
, min_verts_per_prim
, use_adjacency
);
2007 assert(max_esverts
>= max_verts_per_prim
&& max_gsprims
>= 1);
2011 /* Round up towards full wave sizes for better ALU utilization. */
2012 if (!max_vert_out_per_gs_instance
) {
2013 const unsigned wavesize
= si_get_shader_wave_size(shader
);
2014 unsigned orig_max_esverts
;
2015 unsigned orig_max_gsprims
;
2017 orig_max_esverts
= max_esverts
;
2018 orig_max_gsprims
= max_gsprims
;
2020 max_esverts
= align(max_esverts
, wavesize
);
2021 max_esverts
= MIN2(max_esverts
, max_esverts_base
);
2022 if (esvert_lds_size
)
2024 MIN2(max_esverts
, (max_lds_size
- max_gsprims
* gsprim_lds_size
) / esvert_lds_size
);
2025 max_esverts
= MIN2(max_esverts
, max_gsprims
* max_verts_per_prim
);
2026 /* Hardware restriction: minimum value of max_esverts */
2027 max_esverts
= MAX2(max_esverts
, min_esverts
- 1 + max_verts_per_prim
);
2029 max_gsprims
= align(max_gsprims
, wavesize
);
2030 max_gsprims
= MIN2(max_gsprims
, max_gsprims_base
);
2031 if (gsprim_lds_size
) {
2032 /* Don't count unusable vertices to the LDS size. Those are vertices above
2033 * the maximum number of vertices that can occur in the workgroup,
2034 * which is e.g. max_gsprims * 3 for triangles.
2036 unsigned usable_esverts
= MIN2(max_esverts
, max_gsprims
* max_verts_per_prim
);
2038 MIN2(max_gsprims
, (max_lds_size
- usable_esverts
* esvert_lds_size
) / gsprim_lds_size
);
2040 clamp_gsprims_to_esverts(&max_gsprims
, max_esverts
, min_verts_per_prim
, use_adjacency
);
2041 assert(max_esverts
>= max_verts_per_prim
&& max_gsprims
>= 1);
2042 } while (orig_max_esverts
!= max_esverts
|| orig_max_gsprims
!= max_gsprims
);
2044 /* Verify the restriction. */
2045 assert(max_esverts
>= min_esverts
- 1 + max_verts_per_prim
);
2047 /* Hardware restriction: minimum value of max_esverts */
2048 max_esverts
= MAX2(max_esverts
, min_esverts
- 1 + max_verts_per_prim
);
2051 unsigned max_out_vertices
=
2052 max_vert_out_per_gs_instance
2053 ? gs_sel
->gs_max_out_vertices
2054 : gs_stage
== MESA_SHADER_GEOMETRY
2055 ? max_gsprims
* gs_num_invocations
* gs_sel
->gs_max_out_vertices
2057 assert(max_out_vertices
<= 256);
2059 unsigned prim_amp_factor
= 1;
2060 if (gs_stage
== MESA_SHADER_GEOMETRY
) {
2061 /* Number of output primitives per GS input primitive after
2063 prim_amp_factor
= gs_sel
->gs_max_out_vertices
;
2066 /* The GE only checks against the maximum number of ES verts after
2067 * allocating a full GS primitive. So we need to ensure that whenever
2068 * this check passes, there is enough space for a full primitive without
2071 shader
->ngg
.hw_max_esverts
= max_esverts
- max_verts_per_prim
+ 1;
2072 shader
->ngg
.max_gsprims
= max_gsprims
;
2073 shader
->ngg
.max_out_verts
= max_out_vertices
;
2074 shader
->ngg
.prim_amp_factor
= prim_amp_factor
;
2075 shader
->ngg
.max_vert_out_per_gs_instance
= max_vert_out_per_gs_instance
;
2077 /* Don't count unusable vertices. */
2078 shader
->gs_info
.esgs_ring_size
= MIN2(max_esverts
, max_gsprims
* max_verts_per_prim
) *
2080 shader
->ngg
.ngg_emit_size
= max_gsprims
* gsprim_lds_size
;
2082 assert(shader
->ngg
.hw_max_esverts
>= min_esverts
); /* HW limitation */
2084 /* If asserts are disabled, we use the same conditions to return false */
2085 return max_esverts
>= max_verts_per_prim
&& max_gsprims
>= 1 &&
2086 max_out_vertices
<= 256 &&
2087 shader
->ngg
.hw_max_esverts
>= min_esverts
;