1/*
2 * Copyright 2020 Advanced Micro Devices, Inc.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
23 */
24
25#include "si_pipe.h"
26#include "si_shader_internal.h"
27#include "sid.h"
28#include "util/u_memory.h"
29
30LLVMValueRef si_is_es_thread(struct si_shader_context *ctx)
31{
32   /* Return true if the current thread should execute an ES thread. */
33   return LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, ac_get_thread_id(&ctx->ac),
34                        si_unpack_param(ctx, ctx->args.merged_wave_info, 0, 8), "");
35}
36
37LLVMValueRef si_is_gs_thread(struct si_shader_context *ctx)
38{
39   /* Return true if the current thread should execute a GS thread. */
40   return LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, ac_get_thread_id(&ctx->ac),
41                        si_unpack_param(ctx, ctx->args.merged_wave_info, 8, 8), "");
42}
43
44static LLVMValueRef si_llvm_load_input_gs(struct ac_shader_abi *abi, unsigned input_index,
45                                          unsigned vtx_offset_param, LLVMTypeRef type,
46                                          unsigned swizzle)
47{
48   struct si_shader_context *ctx = si_shader_context_from_abi(abi);
49   struct si_shader *shader = ctx->shader;
50   LLVMValueRef vtx_offset, soffset;
51   struct si_shader_info *info = &shader->selector->info;
52   unsigned param;
53   LLVMValueRef value;
54
55   param = si_shader_io_get_unique_index(info->input[input_index].semantic, false);
56
57   /* GFX9 has the ESGS ring in LDS. */
58   if (ctx->screen->info.chip_class >= GFX9) {
59      unsigned index = vtx_offset_param;
60      vtx_offset =
61         si_unpack_param(ctx, ctx->args.gs_vtx_offset[index / 2], (index & 1) * 16, 16);
62
63      unsigned offset = param * 4 + swizzle;
64      vtx_offset =
65         LLVMBuildAdd(ctx->ac.builder, vtx_offset, LLVMConstInt(ctx->ac.i32, offset, false), "");
66
67      LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->esgs_ring, vtx_offset);
68      LLVMValueRef value = LLVMBuildLoad(ctx->ac.builder, ptr, "");
69      return LLVMBuildBitCast(ctx->ac.builder, value, type, "");
70   }
71
72   /* GFX6: input load from the ESGS ring in memory. */
73   /* Get the vertex offset parameter on GFX6. */
74   LLVMValueRef gs_vtx_offset = ac_get_arg(&ctx->ac, ctx->args.gs_vtx_offset[vtx_offset_param]);
75
76   vtx_offset = LLVMBuildMul(ctx->ac.builder, gs_vtx_offset, LLVMConstInt(ctx->ac.i32, 4, 0), "");
77
78   soffset = LLVMConstInt(ctx->ac.i32, (param * 4 + swizzle) * 256, 0);
79
80   value = ac_build_buffer_load(&ctx->ac, ctx->esgs_ring, 1, ctx->ac.i32_0, vtx_offset, soffset, 0,
81                                ctx->ac.f32, ac_glc, true, false);
82   return LLVMBuildBitCast(ctx->ac.builder, value, type, "");
83}
84
85static LLVMValueRef si_nir_load_input_gs(struct ac_shader_abi *abi,
86                                         unsigned driver_location, unsigned component,
87                                         unsigned num_components, unsigned vertex_index,
88                                         LLVMTypeRef type)
89{
90   struct si_shader_context *ctx = si_shader_context_from_abi(abi);
91
92   LLVMValueRef value[4];
93   for (unsigned i = component; i < component + num_components; i++) {
94      value[i] = si_llvm_load_input_gs(&ctx->abi, driver_location,
95                                       vertex_index, type, i);
96   }
97
98   return ac_build_varying_gather_values(&ctx->ac, value, num_components, component);
99}
100
101/* Pass GS inputs from ES to GS on GFX9. */
102static void si_set_es_return_value_for_gs(struct si_shader_context *ctx)
103{
104   if (!ctx->shader->is_monolithic)
105      ac_build_endif(&ctx->ac, ctx->merged_wrap_if_label);
106
107   LLVMValueRef ret = ctx->return_value;
108
109   ret = si_insert_input_ptr(ctx, ret, ctx->other_const_and_shader_buffers, 0);
110   ret = si_insert_input_ptr(ctx, ret, ctx->other_samplers_and_images, 1);
111   if (ctx->shader->key.as_ngg)
112      ret = si_insert_input_ptr(ctx, ret, ctx->args.gs_tg_info, 2);
113   else
114      ret = si_insert_input_ret(ctx, ret, ctx->args.gs2vs_offset, 2);
115   ret = si_insert_input_ret(ctx, ret, ctx->args.merged_wave_info, 3);
116   ret = si_insert_input_ret(ctx, ret, ctx->args.scratch_offset, 5);
117
118   ret = si_insert_input_ptr(ctx, ret, ctx->internal_bindings, 8 + SI_SGPR_INTERNAL_BINDINGS);
119   ret = si_insert_input_ptr(ctx, ret, ctx->bindless_samplers_and_images,
120                             8 + SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES);
121   if (ctx->screen->use_ngg) {
122      ret = si_insert_input_ptr(ctx, ret, ctx->vs_state_bits, 8 + SI_SGPR_VS_STATE_BITS);
123   }
124
125   unsigned vgpr = 8 + SI_NUM_VS_STATE_RESOURCE_SGPRS;
126
127   ret = si_insert_input_ret_float(ctx, ret, ctx->args.gs_vtx_offset[0], vgpr++);
128   ret = si_insert_input_ret_float(ctx, ret, ctx->args.gs_vtx_offset[1], vgpr++);
129   ret = si_insert_input_ret_float(ctx, ret, ctx->args.gs_prim_id, vgpr++);
130   ret = si_insert_input_ret_float(ctx, ret, ctx->args.gs_invocation_id, vgpr++);
131   ret = si_insert_input_ret_float(ctx, ret, ctx->args.gs_vtx_offset[2], vgpr++);
132   ctx->return_value = ret;
133}
134
135void si_llvm_emit_es_epilogue(struct ac_shader_abi *abi)
136{
137   struct si_shader_context *ctx = si_shader_context_from_abi(abi);
138   struct si_shader *es = ctx->shader;
139   struct si_shader_info *info = &es->selector->info;
140   LLVMValueRef *addrs = abi->outputs;
141   LLVMValueRef lds_base = NULL;
142   unsigned chan;
143   int i;
144
145   if (ctx->screen->info.chip_class >= GFX9 && info->num_outputs) {
146      unsigned itemsize_dw = es->selector->esgs_itemsize / 4;
147      LLVMValueRef vertex_idx = ac_get_thread_id(&ctx->ac);
148      LLVMValueRef wave_idx = si_unpack_param(ctx, ctx->args.merged_wave_info, 24, 4);
149      vertex_idx =
150         LLVMBuildOr(ctx->ac.builder, vertex_idx,
151                     LLVMBuildMul(ctx->ac.builder, wave_idx,
152                                  LLVMConstInt(ctx->ac.i32, ctx->ac.wave_size, false), ""),
153                     "");
154      lds_base =
155         LLVMBuildMul(ctx->ac.builder, vertex_idx, LLVMConstInt(ctx->ac.i32, itemsize_dw, 0), "");
156   }
157
158   for (i = 0; i < info->num_outputs; i++) {
159      int param;
160
161      if (info->output_semantic[i] == VARYING_SLOT_VIEWPORT ||
162          info->output_semantic[i] == VARYING_SLOT_LAYER)
163         continue;
164
165      param = si_shader_io_get_unique_index(info->output_semantic[i], false);
166
167      for (chan = 0; chan < 4; chan++) {
168         if (!(info->output_usagemask[i] & (1 << chan)))
169            continue;
170
171         LLVMValueRef out_val = LLVMBuildLoad(ctx->ac.builder, addrs[4 * i + chan], "");
172         out_val = ac_to_integer(&ctx->ac, out_val);
173
174         /* GFX9 has the ESGS ring in LDS. */
175         if (ctx->screen->info.chip_class >= GFX9) {
176            LLVMValueRef idx = LLVMConstInt(ctx->ac.i32, param * 4 + chan, false);
177            idx = LLVMBuildAdd(ctx->ac.builder, lds_base, idx, "");
178            ac_build_indexed_store(&ctx->ac, ctx->esgs_ring, idx, out_val);
179            continue;
180         }
181
182         ac_build_buffer_store_dword(&ctx->ac, ctx->esgs_ring, out_val, 1, NULL,
183                                     ac_get_arg(&ctx->ac, ctx->args.es2gs_offset),
184                                     (4 * param + chan) * 4, ac_glc | ac_slc | ac_swizzled);
185      }
186   }
187
188   if (ctx->screen->info.chip_class >= GFX9)
189      si_set_es_return_value_for_gs(ctx);
190}
191
192static LLVMValueRef si_get_gs_wave_id(struct si_shader_context *ctx)
193{
194   if (ctx->screen->info.chip_class >= GFX9)
195      return si_unpack_param(ctx, ctx->args.merged_wave_info, 16, 8);
196   else
197      return ac_get_arg(&ctx->ac, ctx->args.gs_wave_id);
198}
199
200static void emit_gs_epilogue(struct si_shader_context *ctx)
201{
202   if (ctx->shader->key.as_ngg) {
203      gfx10_ngg_gs_emit_epilogue(ctx);
204      return;
205   }
206
207   if (ctx->screen->info.chip_class >= GFX10)
208      LLVMBuildFence(ctx->ac.builder, LLVMAtomicOrderingRelease, false, "");
209
210   ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_NOP | AC_SENDMSG_GS_DONE, si_get_gs_wave_id(ctx));
211
212   if (ctx->screen->info.chip_class >= GFX9)
213      ac_build_endif(&ctx->ac, ctx->merged_wrap_if_label);
214}
215
216static void si_llvm_emit_gs_epilogue(struct ac_shader_abi *abi)
217{
218   struct si_shader_context *ctx = si_shader_context_from_abi(abi);
219   struct si_shader_info UNUSED *info = &ctx->shader->selector->info;
220
221   assert(info->num_outputs <= AC_LLVM_MAX_OUTPUTS);
222
223   emit_gs_epilogue(ctx);
224}
225
226/* Emit one vertex from the geometry shader */
227static void si_llvm_emit_vertex(struct ac_shader_abi *abi, unsigned stream, LLVMValueRef *addrs)
228{
229   struct si_shader_context *ctx = si_shader_context_from_abi(abi);
230
231   if (ctx->shader->key.as_ngg) {
232      gfx10_ngg_gs_emit_vertex(ctx, stream, addrs);
233      return;
234   }
235
236   struct si_shader_info *info = &ctx->shader->selector->info;
237   struct si_shader *shader = ctx->shader;
238   LLVMValueRef soffset = ac_get_arg(&ctx->ac, ctx->args.gs2vs_offset);
239   LLVMValueRef gs_next_vertex;
240   LLVMValueRef can_emit;
241   unsigned chan, offset;
242   int i;
243
244   /* Write vertex attribute values to GSVS ring */
245   gs_next_vertex = LLVMBuildLoad(ctx->ac.builder, ctx->gs_next_vertex[stream], "");
246
247   /* If this thread has already emitted the declared maximum number of
248    * vertices, skip the write: excessive vertex emissions are not
249    * supposed to have any effect.
250    *
251    * If the shader has no writes to memory, kill it instead. This skips
252    * further memory loads and may allow LLVM to skip to the end
253    * altogether.
254    */
255   can_emit =
256      LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, gs_next_vertex,
257                    LLVMConstInt(ctx->ac.i32, shader->selector->info.base.gs.vertices_out, 0), "");
258
259   bool use_kill = !info->base.writes_memory;
260   if (use_kill) {
261      ac_build_kill_if_false(&ctx->ac, can_emit);
262   } else {
263      ac_build_ifcc(&ctx->ac, can_emit, 6505);
264   }
265
266   offset = 0;
267   for (i = 0; i < info->num_outputs; i++) {
268      for (chan = 0; chan < 4; chan++) {
269         if (!(info->output_usagemask[i] & (1 << chan)) ||
270             ((info->output_streams[i] >> (2 * chan)) & 3) != stream)
271            continue;
272
273         LLVMValueRef out_val = LLVMBuildLoad(ctx->ac.builder, addrs[4 * i + chan], "");
274         LLVMValueRef voffset =
275            LLVMConstInt(ctx->ac.i32, offset * shader->selector->info.base.gs.vertices_out, 0);
276         offset++;
277
278         voffset = LLVMBuildAdd(ctx->ac.builder, voffset, gs_next_vertex, "");
279         voffset = LLVMBuildMul(ctx->ac.builder, voffset, LLVMConstInt(ctx->ac.i32, 4, 0), "");
280
281         out_val = ac_to_integer(&ctx->ac, out_val);
282
283         ac_build_buffer_store_dword(&ctx->ac, ctx->gsvs_ring[stream], out_val, 1, voffset, soffset,
284                                     0, ac_glc | ac_slc | ac_swizzled);
285      }
286   }
287
288   gs_next_vertex = LLVMBuildAdd(ctx->ac.builder, gs_next_vertex, ctx->ac.i32_1, "");
289   LLVMBuildStore(ctx->ac.builder, gs_next_vertex, ctx->gs_next_vertex[stream]);
290
291   /* Signal vertex emission if vertex data was written. */
292   if (offset) {
293      ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_EMIT | AC_SENDMSG_GS | (stream << 8),
294                       si_get_gs_wave_id(ctx));
295   }
296
297   if (!use_kill)
298      ac_build_endif(&ctx->ac, 6505);
299}
300
301/* Cut one primitive from the geometry shader */
302static void si_llvm_emit_primitive(struct ac_shader_abi *abi, unsigned stream)
303{
304   struct si_shader_context *ctx = si_shader_context_from_abi(abi);
305
306   if (ctx->shader->key.as_ngg) {
307      LLVMBuildStore(ctx->ac.builder, ctx->ac.i32_0, ctx->gs_curprim_verts[stream]);
308      return;
309   }
310
311   /* Signal primitive cut */
312   ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_CUT | AC_SENDMSG_GS | (stream << 8),
313                    si_get_gs_wave_id(ctx));
314}
315
316void si_preload_esgs_ring(struct si_shader_context *ctx)
317{
318   if (ctx->screen->info.chip_class <= GFX8) {
319      unsigned ring = ctx->stage == MESA_SHADER_GEOMETRY ? SI_GS_RING_ESGS : SI_ES_RING_ESGS;
320      LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, ring, 0);
321      LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->internal_bindings);
322
323      ctx->esgs_ring = ac_build_load_to_sgpr(&ctx->ac, buf_ptr, offset);
324   } else {
325      if (USE_LDS_SYMBOLS) {
326         /* Declare the ESGS ring as an explicit LDS symbol. */
327         si_llvm_declare_esgs_ring(ctx);
328      } else {
329         ac_declare_lds_as_pointer(&ctx->ac);
330         ctx->esgs_ring = ctx->ac.lds;
331      }
332   }
333}
334
335void si_preload_gs_rings(struct si_shader_context *ctx)
336{
337   const struct si_shader_selector *sel = ctx->shader->selector;
338   LLVMBuilderRef builder = ctx->ac.builder;
339   LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, SI_RING_GSVS, 0);
340   LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->internal_bindings);
341   LLVMValueRef base_ring = ac_build_load_to_sgpr(&ctx->ac, buf_ptr, offset);
342
343   /* The conceptual layout of the GSVS ring is
344    *   v0c0 .. vLv0 v0c1 .. vLc1 ..
345    * but the real memory layout is swizzled across
346    * threads:
347    *   t0v0c0 .. t15v0c0 t0v1c0 .. t15v1c0 ... t15vLcL
348    *   t16v0c0 ..
349    * Override the buffer descriptor accordingly.
350    */
351   LLVMTypeRef v2i64 = LLVMVectorType(ctx->ac.i64, 2);
352   uint64_t stream_offset = 0;
353
354   for (unsigned stream = 0; stream < 4; ++stream) {
355      unsigned num_components;
356      unsigned stride;
357      unsigned num_records;
358      LLVMValueRef ring, tmp;
359
360      num_components = sel->info.num_stream_output_components[stream];
361      if (!num_components)
362         continue;
363
364      stride = 4 * num_components * sel->info.base.gs.vertices_out;
365
366      /* Limit on the stride field for <= GFX7. */
367      assert(stride < (1 << 14));
368
369      num_records = ctx->ac.wave_size;
370
371      ring = LLVMBuildBitCast(builder, base_ring, v2i64, "");
372      tmp = LLVMBuildExtractElement(builder, ring, ctx->ac.i32_0, "");
373      tmp = LLVMBuildAdd(builder, tmp, LLVMConstInt(ctx->ac.i64, stream_offset, 0), "");
374      stream_offset += stride * ctx->ac.wave_size;
375
376      ring = LLVMBuildInsertElement(builder, ring, tmp, ctx->ac.i32_0, "");
377      ring = LLVMBuildBitCast(builder, ring, ctx->ac.v4i32, "");
378      tmp = LLVMBuildExtractElement(builder, ring, ctx->ac.i32_1, "");
379      tmp = LLVMBuildOr(
380         builder, tmp,
381         LLVMConstInt(ctx->ac.i32, S_008F04_STRIDE(stride) | S_008F04_SWIZZLE_ENABLE(1), 0), "");
382      ring = LLVMBuildInsertElement(builder, ring, tmp, ctx->ac.i32_1, "");
383      ring = LLVMBuildInsertElement(builder, ring, LLVMConstInt(ctx->ac.i32, num_records, 0),
384                                    LLVMConstInt(ctx->ac.i32, 2, 0), "");
385
386      uint32_t rsrc3 =
387         S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
388         S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
389         S_008F0C_INDEX_STRIDE(1) | /* index_stride = 16 (elements) */
390         S_008F0C_ADD_TID_ENABLE(1);
391
392      if (ctx->ac.chip_class >= GFX10) {
393         rsrc3 |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_FLOAT) |
394                  S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) | S_008F0C_RESOURCE_LEVEL(1);
395      } else {
396         rsrc3 |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
397                  S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32) |
398                  S_008F0C_ELEMENT_SIZE(1); /* element_size = 4 (bytes) */
399      }
400
401      ring = LLVMBuildInsertElement(builder, ring, LLVMConstInt(ctx->ac.i32, rsrc3, false),
402                                    LLVMConstInt(ctx->ac.i32, 3, 0), "");
403
404      ctx->gsvs_ring[stream] = ring;
405   }
406}
407
408/* Generate code for the hardware VS shader stage to go with a geometry shader */
409struct si_shader *si_generate_gs_copy_shader(struct si_screen *sscreen,
410                                             struct ac_llvm_compiler *compiler,
411                                             struct si_shader_selector *gs_selector,
412                                             struct pipe_debug_callback *debug)
413{
414   struct si_shader_context ctx;
415   struct si_shader *shader;
416   LLVMBuilderRef builder;
417   struct si_shader_output_values outputs[SI_MAX_VS_OUTPUTS];
418   struct si_shader_info *gsinfo = &gs_selector->info;
419   int i;
420
421   shader = CALLOC_STRUCT(si_shader);
422   if (!shader)
423      return NULL;
424
425   /* We can leave the fence as permanently signaled because the GS copy
426    * shader only becomes visible globally after it has been compiled. */
427   util_queue_fence_init(&shader->ready);
428
429   shader->selector = gs_selector;
430   shader->is_gs_copy_shader = true;
431
432   si_llvm_context_init(&ctx, sscreen, compiler,
433                        si_get_wave_size(sscreen, MESA_SHADER_VERTEX,
434                                         false, false));
435   ctx.shader = shader;
436   ctx.stage = MESA_SHADER_VERTEX;
437
438   builder = ctx.ac.builder;
439
440   si_llvm_create_main_func(&ctx, false);
441
442   LLVMValueRef buf_ptr = ac_get_arg(&ctx.ac, ctx.internal_bindings);
443   ctx.gsvs_ring[0] =
444      ac_build_load_to_sgpr(&ctx.ac, buf_ptr, LLVMConstInt(ctx.ac.i32, SI_RING_GSVS, 0));
445
446   LLVMValueRef voffset =
447      LLVMBuildMul(ctx.ac.builder, ctx.abi.vertex_id, LLVMConstInt(ctx.ac.i32, 4, 0), "");
448
449   /* Fetch the vertex stream ID.*/
450   LLVMValueRef stream_id;
451
452   if (!sscreen->use_ngg_streamout && gs_selector->so.num_outputs)
453      stream_id = si_unpack_param(&ctx, ctx.args.streamout_config, 24, 2);
454   else
455      stream_id = ctx.ac.i32_0;
456
457   /* Fill in output information. */
458   for (i = 0; i < gsinfo->num_outputs; ++i) {
459      outputs[i].semantic = gsinfo->output_semantic[i];
460
461      for (int chan = 0; chan < 4; chan++) {
462         outputs[i].vertex_stream[chan] = (gsinfo->output_streams[i] >> (2 * chan)) & 3;
463      }
464   }
465
466   LLVMBasicBlockRef end_bb;
467   LLVMValueRef switch_inst;
468
469   end_bb = LLVMAppendBasicBlockInContext(ctx.ac.context, ctx.main_fn, "end");
470   switch_inst = LLVMBuildSwitch(builder, stream_id, end_bb, 4);
471
472   for (int stream = 0; stream < 4; stream++) {
473      LLVMBasicBlockRef bb;
474      unsigned offset;
475
476      if (!gsinfo->num_stream_output_components[stream])
477         continue;
478
479      if (stream > 0 && !gs_selector->so.num_outputs)
480         continue;
481
482      bb = LLVMInsertBasicBlockInContext(ctx.ac.context, end_bb, "out");
483      LLVMAddCase(switch_inst, LLVMConstInt(ctx.ac.i32, stream, 0), bb);
484      LLVMPositionBuilderAtEnd(builder, bb);
485
486      /* Fetch vertex data from GSVS ring */
487      offset = 0;
488      for (i = 0; i < gsinfo->num_outputs; ++i) {
489         for (unsigned chan = 0; chan < 4; chan++) {
490            if (!(gsinfo->output_usagemask[i] & (1 << chan)) ||
491                outputs[i].vertex_stream[chan] != stream) {
492               outputs[i].values[chan] = LLVMGetUndef(ctx.ac.f32);
493               continue;
494            }
495
496            LLVMValueRef soffset =
497               LLVMConstInt(ctx.ac.i32, offset * gs_selector->info.base.gs.vertices_out * 16 * 4, 0);
498            offset++;
499
500            outputs[i].values[chan] =
501               ac_build_buffer_load(&ctx.ac, ctx.gsvs_ring[0], 1, ctx.ac.i32_0, voffset, soffset, 0,
502                                    ctx.ac.f32, ac_glc | ac_slc, true, false);
503         }
504      }
505
506      /* Streamout and exports. */
507      if (!sscreen->use_ngg_streamout && gs_selector->so.num_outputs) {
508         si_llvm_emit_streamout(&ctx, outputs, gsinfo->num_outputs, stream);
509      }
510
511      if (stream == 0)
512         si_llvm_build_vs_exports(&ctx, outputs, gsinfo->num_outputs);
513
514      LLVMBuildBr(builder, end_bb);
515   }
516
517   LLVMPositionBuilderAtEnd(builder, end_bb);
518
519   LLVMBuildRetVoid(ctx.ac.builder);
520
521   ctx.stage = MESA_SHADER_GEOMETRY; /* override for shader dumping */
522   si_llvm_optimize_module(&ctx);
523
524   bool ok = false;
525   if (si_compile_llvm(sscreen, &ctx.shader->binary, &ctx.shader->config, ctx.compiler, &ctx.ac,
526                       debug, MESA_SHADER_GEOMETRY, "GS Copy Shader", false)) {
527      if (si_can_dump_shader(sscreen, MESA_SHADER_GEOMETRY))
528         fprintf(stderr, "GS Copy Shader:\n");
529      si_shader_dump(sscreen, ctx.shader, debug, stderr, true);
530
531      if (!ctx.shader->config.scratch_bytes_per_wave)
532         ok = si_shader_binary_upload(sscreen, ctx.shader, 0);
533      else
534         ok = true;
535   }
536
537   si_llvm_dispose(&ctx);
538
539   if (!ok) {
540      FREE(shader);
541      shader = NULL;
542   } else {
543      si_fix_resource_usage(sscreen, shader);
544   }
545   return shader;
546}
547
548/**
549 * Build the GS prolog function. Rotate the input vertices for triangle strips
550 * with adjacency.
551 */
552void si_llvm_build_gs_prolog(struct si_shader_context *ctx, union si_shader_part_key *key)
553{
554   unsigned num_sgprs, num_vgprs;
555   LLVMBuilderRef builder = ctx->ac.builder;
556   LLVMTypeRef returns[AC_MAX_ARGS];
557   LLVMValueRef func, ret;
558
559   memset(&ctx->args, 0, sizeof(ctx->args));
560
561   if (ctx->screen->info.chip_class >= GFX9) {
562      /* Other user SGPRs are not needed by GS. */
563      num_sgprs = 8 + SI_NUM_VS_STATE_RESOURCE_SGPRS;
564      num_vgprs = 5; /* ES inputs are not needed by GS */
565   } else {
566      num_sgprs = GFX6_GS_NUM_USER_SGPR + 2;
567      num_vgprs = 8;
568   }
569
570   for (unsigned i = 0; i < num_sgprs; ++i) {
571      ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
572      returns[i] = ctx->ac.i32;
573   }
574
575   for (unsigned i = 0; i < num_vgprs; ++i) {
576      ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL);
577      returns[num_sgprs + i] = ctx->ac.f32;
578   }
579
580   /* Create the function. */
581   si_llvm_create_func(ctx, "gs_prolog", returns, num_sgprs + num_vgprs, 0);
582   func = ctx->main_fn;
583
584   /* Copy inputs to outputs. This should be no-op, as the registers match,
585    * but it will prevent the compiler from overwriting them unintentionally.
586    */
587   ret = ctx->return_value;
588   for (unsigned i = 0; i < num_sgprs; i++) {
589      LLVMValueRef p = LLVMGetParam(func, i);
590      ret = LLVMBuildInsertValue(builder, ret, p, i, "");
591   }
592   for (unsigned i = 0; i < num_vgprs; i++) {
593      LLVMValueRef p = LLVMGetParam(func, num_sgprs + i);
594      p = ac_to_float(&ctx->ac, p);
595      ret = LLVMBuildInsertValue(builder, ret, p, num_sgprs + i, "");
596   }
597
598   if (key->gs_prolog.states.tri_strip_adj_fix) {
599      /* Remap the input vertices for every other primitive. */
600      const struct ac_arg gfx6_vtx_params[6] = {
601         {.used = true, .arg_index = num_sgprs},     {.used = true, .arg_index = num_sgprs + 1},
602         {.used = true, .arg_index = num_sgprs + 3}, {.used = true, .arg_index = num_sgprs + 4},
603         {.used = true, .arg_index = num_sgprs + 5}, {.used = true, .arg_index = num_sgprs + 6},
604      };
605      const struct ac_arg gfx9_vtx_params[3] = {
606         {.used = true, .arg_index = num_sgprs},
607         {.used = true, .arg_index = num_sgprs + 1},
608         {.used = true, .arg_index = num_sgprs + 4},
609      };
610      LLVMValueRef vtx_in[6], vtx_out[6];
611      LLVMValueRef prim_id, rotate;
612
613      if (ctx->screen->info.chip_class >= GFX9) {
614         for (unsigned i = 0; i < 3; i++) {
615            vtx_in[i * 2] = si_unpack_param(ctx, gfx9_vtx_params[i], 0, 16);
616            vtx_in[i * 2 + 1] = si_unpack_param(ctx, gfx9_vtx_params[i], 16, 16);
617         }
618      } else {
619         for (unsigned i = 0; i < 6; i++)
620            vtx_in[i] = ac_get_arg(&ctx->ac, gfx6_vtx_params[i]);
621      }
622
623      prim_id = LLVMGetParam(func, num_sgprs + 2);
624      rotate = LLVMBuildTrunc(builder, prim_id, ctx->ac.i1, "");
625
626      for (unsigned i = 0; i < 6; ++i) {
627         LLVMValueRef base, rotated;
628         base = vtx_in[i];
629         rotated = vtx_in[(i + 4) % 6];
630         vtx_out[i] = LLVMBuildSelect(builder, rotate, rotated, base, "");
631      }
632
633      if (ctx->screen->info.chip_class >= GFX9) {
634         for (unsigned i = 0; i < 3; i++) {
635            LLVMValueRef hi, out;
636
637            hi = LLVMBuildShl(builder, vtx_out[i * 2 + 1], LLVMConstInt(ctx->ac.i32, 16, 0), "");
638            out = LLVMBuildOr(builder, vtx_out[i * 2], hi, "");
639            out = ac_to_float(&ctx->ac, out);
640            ret = LLVMBuildInsertValue(builder, ret, out, gfx9_vtx_params[i].arg_index, "");
641         }
642      } else {
643         for (unsigned i = 0; i < 6; i++) {
644            LLVMValueRef out;
645
646            out = ac_to_float(&ctx->ac, vtx_out[i]);
647            ret = LLVMBuildInsertValue(builder, ret, out, gfx6_vtx_params[i].arg_index, "");
648         }
649      }
650   }
651
652   LLVMBuildRet(builder, ret);
653}
654
655void si_llvm_init_gs_callbacks(struct si_shader_context *ctx)
656{
657   ctx->abi.load_inputs = si_nir_load_input_gs;
658   ctx->abi.emit_vertex = si_llvm_emit_vertex;
659   ctx->abi.emit_primitive = si_llvm_emit_primitive;
660   ctx->abi.emit_outputs = si_llvm_emit_gs_epilogue;
661}
662