• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright © 2014 Intel Corporation
3  *
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  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
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
13  * Software.
14  *
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 NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  *
23  * This code is based on original work by Ilia Mirkin.
24  */
25 
26 /**
27  * \file gfx6_gs_visitor.cpp
28  *
29  * Gfx6 geometry shader implementation
30  */
31 
32 #include "gfx6_gs_visitor.h"
33 #include "brw_eu.h"
34 
35 namespace brw {
36 
37 void
emit_prolog()38 gfx6_gs_visitor::emit_prolog()
39 {
40    vec4_gs_visitor::emit_prolog();
41 
42    /* Gfx6 geometry shaders require to allocate an initial VUE handle via
43     * FF_SYNC message, however the documentation remarks that only one thread
44     * can write to the URB simultaneously and the FF_SYNC message provides the
45     * synchronization mechanism for this, so using this message effectively
46     * stalls the thread until it is its turn to write to the URB. Because of
47     * this, the best way to implement geometry shader algorithms in gfx6 is to
48     * execute the algorithm before the FF_SYNC message to maximize parallelism.
49     *
50     * To achieve this we buffer the geometry shader outputs for each emitted
51     * vertex in vertex_output during operation. Then, when we have processed
52     * the last vertex (that is, at thread end time), we send the FF_SYNC
53     * message to allocate the initial VUE handle and write all buffered vertex
54     * data to the URB in one go.
55     *
56     * For each emitted vertex, vertex_output will hold vue_map.num_slots
57     * data items plus one additional item to hold required flags
58     * (PrimType, PrimStart, PrimEnd, as expected by the URB_WRITE message)
59     * which come right after the data items for that vertex. Vertex data and
60     * flags for the next vertex come right after the data items and flags for
61     * the previous vertex.
62     */
63    this->current_annotation = "gfx6 prolog";
64    this->vertex_output = src_reg(this,
65                                  glsl_type::uint_type,
66                                  (prog_data->vue_map.num_slots + 1) *
67                                  nir->info.gs.vertices_out);
68    this->vertex_output_offset = src_reg(this, glsl_type::uint_type);
69    emit(MOV(dst_reg(this->vertex_output_offset), brw_imm_ud(0u)));
70 
71    /* MRF 1 will be the header for all messages (FF_SYNC and URB_WRITES),
72     * so initialize it once to R0.
73     */
74    vec4_instruction *inst = emit(MOV(dst_reg(MRF, 1),
75                                      retype(brw_vec8_grf(0, 0),
76                                             BRW_REGISTER_TYPE_UD)));
77    inst->force_writemask_all = true;
78 
79    /* This will be used as a temporary to store writeback data of FF_SYNC
80     * and URB_WRITE messages.
81     */
82    this->temp = src_reg(this, glsl_type::uint_type);
83 
84    /* This will be used to know when we are processing the first vertex of
85     * a primitive. We will set this to URB_WRITE_PRIM_START only when we know
86     * that we are processing the first vertex in the primitive and to zero
87     * otherwise. This way we can use its value directly in the URB write
88     * headers.
89     */
90    this->first_vertex = src_reg(this, glsl_type::uint_type);
91    emit(MOV(dst_reg(this->first_vertex), brw_imm_ud(URB_WRITE_PRIM_START)));
92 
93    /* The FF_SYNC message requires to know the number of primitives generated,
94     * so keep a counter for this.
95     */
96    this->prim_count = src_reg(this, glsl_type::uint_type);
97    emit(MOV(dst_reg(this->prim_count), brw_imm_ud(0u)));
98 
99    if (gs_prog_data->num_transform_feedback_bindings) {
100       /* Create a virtual register to hold destination indices in SOL */
101       this->destination_indices = src_reg(this, glsl_type::uvec4_type);
102       /* Create a virtual register to hold number of written primitives */
103       this->sol_prim_written = src_reg(this, glsl_type::uint_type);
104       /* Create a virtual register to hold Streamed Vertex Buffer Indices */
105       this->svbi = src_reg(this, glsl_type::uvec4_type);
106       /* Create a virtual register to hold max values of SVBI */
107       this->max_svbi = src_reg(this, glsl_type::uvec4_type);
108       emit(MOV(dst_reg(this->max_svbi),
109                src_reg(retype(brw_vec1_grf(1, 4), BRW_REGISTER_TYPE_UD))));
110    }
111 
112    /* PrimitveID is delivered in r0.1 of the thread payload. If the program
113     * needs it we have to move it to a separate register where we can map
114     * the atttribute.
115     *
116     * Notice that we cannot use a virtual register for this, because we need to
117     * map all input attributes to hardware registers in setup_payload(),
118     * which happens before virtual registers are mapped to hardware registers.
119     * We could work around that issue if we were able to compute the first
120     * non-payload register here and move the PrimitiveID information to that
121     * register, but we can't because at this point we don't know the final
122     * number uniforms that will be included in the payload.
123     *
124     * So, what we do is to place PrimitiveID information in r1, which is always
125     * delivered as part of the payload, but its only populated with data
126     * relevant for transform feedback when we set GFX6_GS_SVBI_PAYLOAD_ENABLE
127     * in the 3DSTATE_GS state packet. That information can be obtained by other
128     * means though, so we can safely use r1 for this purpose.
129     */
130    if (gs_prog_data->include_primitive_id) {
131       this->primitive_id =
132          src_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD));
133       emit(GS_OPCODE_SET_PRIMITIVE_ID, dst_reg(this->primitive_id));
134    }
135 }
136 
137 void
gs_emit_vertex(int stream_id)138 gfx6_gs_visitor::gs_emit_vertex(int stream_id)
139 {
140    this->current_annotation = "gfx6 emit vertex";
141 
142    /* Buffer all output slots for this vertex in vertex_output */
143    for (int slot = 0; slot < prog_data->vue_map.num_slots; ++slot) {
144       int varying = prog_data->vue_map.slot_to_varying[slot];
145       if (varying != VARYING_SLOT_PSIZ) {
146          dst_reg dst(this->vertex_output);
147          dst.reladdr = ralloc(mem_ctx, src_reg);
148          memcpy(dst.reladdr, &this->vertex_output_offset, sizeof(src_reg));
149          emit_urb_slot(dst, varying);
150       } else {
151          /* The PSIZ slot can pack multiple varyings in different channels
152           * and emit_urb_slot() will produce a MOV instruction for each of
153           * them. Since we are writing to an array, that will translate to
154           * possibly multiple MOV instructions with an array destination and
155           * each will generate a scratch write with the same offset into
156           * scratch space (thus, each one overwriting the previous). This is
157           * not what we want. What we will do instead is emit PSIZ to a
158           * a regular temporary register, then move that resgister into the
159           * array. This way we only have one instruction with an array
160           * destination and we only produce a single scratch write.
161           */
162          dst_reg tmp = dst_reg(src_reg(this, glsl_type::uvec4_type));
163          emit_urb_slot(tmp, varying);
164          dst_reg dst(this->vertex_output);
165          dst.reladdr = ralloc(mem_ctx, src_reg);
166          memcpy(dst.reladdr, &this->vertex_output_offset, sizeof(src_reg));
167          vec4_instruction *inst = emit(MOV(dst, src_reg(tmp)));
168          inst->force_writemask_all = true;
169       }
170 
171       emit(ADD(dst_reg(this->vertex_output_offset),
172                this->vertex_output_offset, brw_imm_ud(1u)));
173    }
174 
175    /* Now buffer flags for this vertex */
176    dst_reg dst(this->vertex_output);
177    dst.reladdr = ralloc(mem_ctx, src_reg);
178    memcpy(dst.reladdr, &this->vertex_output_offset, sizeof(src_reg));
179    if (nir->info.gs.output_primitive == GL_POINTS) {
180       /* If we are outputting points, then every vertex has PrimStart and
181        * PrimEnd set.
182        */
183       emit(MOV(dst, brw_imm_d((_3DPRIM_POINTLIST << URB_WRITE_PRIM_TYPE_SHIFT) |
184                               URB_WRITE_PRIM_START | URB_WRITE_PRIM_END)));
185       emit(ADD(dst_reg(this->prim_count), this->prim_count, brw_imm_ud(1u)));
186    } else {
187       /* Otherwise, we can only set the PrimStart flag, which we have stored
188        * in the first_vertex register. We will have to wait until we execute
189        * EndPrimitive() or we end the thread to set the PrimEnd flag on a
190        * vertex.
191        */
192       emit(OR(dst, this->first_vertex,
193               brw_imm_ud(gs_prog_data->output_topology <<
194                          URB_WRITE_PRIM_TYPE_SHIFT)));
195       emit(MOV(dst_reg(this->first_vertex), brw_imm_ud(0u)));
196    }
197    emit(ADD(dst_reg(this->vertex_output_offset),
198             this->vertex_output_offset, brw_imm_ud(1u)));
199 }
200 
201 void
gs_end_primitive()202 gfx6_gs_visitor::gs_end_primitive()
203 {
204    this->current_annotation = "gfx6 end primitive";
205    /* Calling EndPrimitive() is optional for point output. In this case we set
206     * the PrimEnd flag when we process EmitVertex().
207     */
208    if (nir->info.gs.output_primitive == GL_POINTS)
209       return;
210 
211    /* Otherwise we know that the last vertex we have processed was the last
212     * vertex in the primitive and we need to set its PrimEnd flag, so do this
213     * unless we haven't emitted that vertex at all (vertex_count != 0).
214     *
215     * Notice that we have already incremented vertex_count when we processed
216     * the last emit_vertex, so we need to take that into account in the
217     * comparison below (hence the num_output_vertices + 1 in the comparison
218     * below).
219     */
220    unsigned num_output_vertices = nir->info.gs.vertices_out;
221    emit(CMP(dst_null_ud(), this->vertex_count,
222             brw_imm_ud(num_output_vertices + 1), BRW_CONDITIONAL_L));
223    vec4_instruction *inst = emit(CMP(dst_null_ud(),
224                                      this->vertex_count, brw_imm_ud(0u),
225                                      BRW_CONDITIONAL_NEQ));
226    inst->predicate = BRW_PREDICATE_NORMAL;
227    emit(IF(BRW_PREDICATE_NORMAL));
228    {
229       /* vertex_output_offset is already pointing at the first entry of the
230        * next vertex. So subtract 1 to modify the flags for the previous
231        * vertex.
232        */
233       src_reg offset(this, glsl_type::uint_type);
234       emit(ADD(dst_reg(offset), this->vertex_output_offset, brw_imm_d(-1)));
235 
236       src_reg dst(this->vertex_output);
237       dst.reladdr = ralloc(mem_ctx, src_reg);
238       memcpy(dst.reladdr, &offset, sizeof(src_reg));
239 
240       emit(OR(dst_reg(dst), dst, brw_imm_d(URB_WRITE_PRIM_END)));
241       emit(ADD(dst_reg(this->prim_count), this->prim_count, brw_imm_ud(1u)));
242 
243       /* Set the first vertex flag to indicate that the next vertex will start
244        * a primitive.
245        */
246       emit(MOV(dst_reg(this->first_vertex), brw_imm_d(URB_WRITE_PRIM_START)));
247    }
248    emit(BRW_OPCODE_ENDIF);
249 }
250 
251 void
emit_urb_write_header(int mrf)252 gfx6_gs_visitor::emit_urb_write_header(int mrf)
253 {
254    this->current_annotation = "gfx6 urb header";
255    /* Compute offset of the flags for the current vertex in vertex_output and
256     * write them in dw2 of the message header.
257     *
258     * Notice that by the time that emit_thread_end() calls here
259     * vertex_output_offset should point to the first data item of the current
260     * vertex in vertex_output, thus we only need to add the number of output
261     * slots per vertex to that offset to obtain the flags data offset.
262     */
263    src_reg flags_offset(this, glsl_type::uint_type);
264    emit(ADD(dst_reg(flags_offset),
265             this->vertex_output_offset,
266             brw_imm_d(prog_data->vue_map.num_slots)));
267 
268    src_reg flags_data(this->vertex_output);
269    flags_data.reladdr = ralloc(mem_ctx, src_reg);
270    memcpy(flags_data.reladdr, &flags_offset, sizeof(src_reg));
271 
272    emit(GS_OPCODE_SET_DWORD_2, dst_reg(MRF, mrf), flags_data);
273 }
274 
275 static unsigned
align_interleaved_urb_mlen(unsigned mlen)276 align_interleaved_urb_mlen(unsigned mlen)
277 {
278    /* URB data written (does not include the message header reg) must
279     * be a multiple of 256 bits, or 2 VS registers.  See vol5c.5,
280     * section 5.4.3.2.2: URB_INTERLEAVED.
281     */
282    if ((mlen % 2) != 1)
283       mlen++;
284    return mlen;
285 }
286 
287 void
emit_snb_gs_urb_write_opcode(bool complete,int base_mrf,int last_mrf,int urb_offset)288 gfx6_gs_visitor::emit_snb_gs_urb_write_opcode(bool complete, int base_mrf,
289                                               int last_mrf, int urb_offset)
290 {
291    vec4_instruction *inst = NULL;
292 
293    if (!complete) {
294       /* If the vertex is not complete we don't have to do anything special */
295       inst = emit(GS_OPCODE_URB_WRITE);
296       inst->urb_write_flags = BRW_URB_WRITE_NO_FLAGS;
297    } else {
298       /* Otherwise we always request to allocate a new VUE handle. If this is
299        * the last write before the EOT message and the new handle never gets
300        * used it will be dereferenced when we send the EOT message. This is
301        * necessary to avoid different setups for the EOT message (one for the
302        * case when there is no output and another for the case when there is)
303        * which would require to end the program with an IF/ELSE/ENDIF block,
304        * something we do not want.
305        */
306       inst = emit(GS_OPCODE_URB_WRITE_ALLOCATE);
307       inst->urb_write_flags = BRW_URB_WRITE_COMPLETE;
308       inst->dst = dst_reg(MRF, base_mrf);
309       inst->src[0] = this->temp;
310    }
311 
312    inst->base_mrf = base_mrf;
313    inst->mlen = align_interleaved_urb_mlen(last_mrf - base_mrf);
314    inst->offset = urb_offset;
315 }
316 
317 void
emit_thread_end()318 gfx6_gs_visitor::emit_thread_end()
319 {
320    /* Make sure the current primitive is ended: we know it is not ended when
321     * first_vertex is not zero. This is only relevant for outputs other than
322     * points because in the point case we set PrimEnd on all vertices.
323     */
324    if (nir->info.gs.output_primitive != GL_POINTS) {
325       emit(CMP(dst_null_ud(), this->first_vertex, brw_imm_ud(0u), BRW_CONDITIONAL_Z));
326       emit(IF(BRW_PREDICATE_NORMAL));
327       gs_end_primitive();
328       emit(BRW_OPCODE_ENDIF);
329    }
330 
331    /* Here we have to:
332     * 1) Emit an FF_SYNC messsage to obtain an initial VUE handle.
333     * 2) Loop over all buffered vertex data and write it to corresponding
334     *    URB entries.
335     * 3) Allocate new VUE handles for all vertices other than the first.
336     * 4) Send a final EOT message.
337     */
338 
339    /* MRF 0 is reserved for the debugger, so start with message header
340     * in MRF 1.
341     */
342    int base_mrf = 1;
343 
344    /* In the process of generating our URB write message contents, we
345     * may need to unspill a register or load from an array.  Those
346     * reads would use MRFs 21..23
347     */
348    int max_usable_mrf = FIRST_SPILL_MRF(devinfo->ver);
349 
350    /* Issue the FF_SYNC message and obtain the initial VUE handle. */
351    this->current_annotation = "gfx6 thread end: ff_sync";
352 
353    vec4_instruction *inst = NULL;
354    if (gs_prog_data->num_transform_feedback_bindings) {
355       src_reg sol_temp(this, glsl_type::uvec4_type);
356       emit(GS_OPCODE_FF_SYNC_SET_PRIMITIVES,
357            dst_reg(this->svbi),
358            this->vertex_count,
359            this->prim_count,
360            sol_temp);
361       inst = emit(GS_OPCODE_FF_SYNC,
362                   dst_reg(this->temp), this->prim_count, this->svbi);
363    } else {
364       inst = emit(GS_OPCODE_FF_SYNC,
365                   dst_reg(this->temp), this->prim_count, brw_imm_ud(0u));
366    }
367    inst->base_mrf = base_mrf;
368 
369    emit(CMP(dst_null_ud(), this->vertex_count, brw_imm_ud(0u), BRW_CONDITIONAL_G));
370    emit(IF(BRW_PREDICATE_NORMAL));
371    {
372       /* Loop over all buffered vertices and emit URB write messages */
373       this->current_annotation = "gfx6 thread end: urb writes init";
374       src_reg vertex(this, glsl_type::uint_type);
375       emit(MOV(dst_reg(vertex), brw_imm_ud(0u)));
376       emit(MOV(dst_reg(this->vertex_output_offset), brw_imm_ud(0u)));
377 
378       this->current_annotation = "gfx6 thread end: urb writes";
379       emit(BRW_OPCODE_DO);
380       {
381          emit(CMP(dst_null_d(), vertex, this->vertex_count, BRW_CONDITIONAL_GE));
382          inst = emit(BRW_OPCODE_BREAK);
383          inst->predicate = BRW_PREDICATE_NORMAL;
384 
385          /* First we prepare the message header */
386          emit_urb_write_header(base_mrf);
387 
388          /* Then add vertex data to the message in interleaved fashion */
389          int slot = 0;
390          bool complete = false;
391          do {
392             int mrf = base_mrf + 1;
393 
394             /* URB offset is in URB row increments, and each of our MRFs is half
395              * of one of those, since we're doing interleaved writes.
396              */
397             int urb_offset = slot / 2;
398 
399             for (; slot < prog_data->vue_map.num_slots; ++slot) {
400                int varying = prog_data->vue_map.slot_to_varying[slot];
401                current_annotation = output_reg_annotation[varying];
402 
403                /* Compute offset of this slot for the current vertex
404                 * in vertex_output
405                 */
406                src_reg data(this->vertex_output);
407                data.reladdr = ralloc(mem_ctx, src_reg);
408                memcpy(data.reladdr, &this->vertex_output_offset,
409                       sizeof(src_reg));
410 
411                /* Copy this slot to the appropriate message register */
412                dst_reg reg = dst_reg(MRF, mrf);
413                reg.type = output_reg[varying][0].type;
414                data.type = reg.type;
415                inst = emit(MOV(reg, data));
416                inst->force_writemask_all = true;
417 
418                mrf++;
419                emit(ADD(dst_reg(this->vertex_output_offset),
420                         this->vertex_output_offset, brw_imm_ud(1u)));
421 
422                /* If this was max_usable_mrf, we can't fit anything more into
423                 * this URB WRITE. Same if we reached the max. message length.
424                 */
425                if (mrf > max_usable_mrf ||
426                    align_interleaved_urb_mlen(mrf - base_mrf + 1) > BRW_MAX_MSG_LENGTH) {
427                   slot++;
428                   break;
429                }
430             }
431 
432             complete = slot >= prog_data->vue_map.num_slots;
433             emit_snb_gs_urb_write_opcode(complete, base_mrf, mrf, urb_offset);
434          } while (!complete);
435 
436          /* Skip over the flags data item so that vertex_output_offset points
437           * to the first data item of the next vertex, so that we can start
438           * writing the next vertex.
439           */
440          emit(ADD(dst_reg(this->vertex_output_offset),
441                   this->vertex_output_offset, brw_imm_ud(1u)));
442 
443          emit(ADD(dst_reg(vertex), vertex, brw_imm_ud(1u)));
444       }
445       emit(BRW_OPCODE_WHILE);
446 
447       if (gs_prog_data->num_transform_feedback_bindings)
448          xfb_write();
449    }
450    emit(BRW_OPCODE_ENDIF);
451 
452    /* Finally, emit EOT message.
453     *
454     * In gfx6 we need to end the thread differently depending on whether we have
455     * emitted at least one vertex or not. In case we did, the EOT message must
456     * always include the COMPLETE flag or else the GPU hangs. If we have not
457     * produced any output we can't use the COMPLETE flag.
458     *
459     * However, this would lead us to end the program with an ENDIF opcode,
460     * which we want to avoid, so what we do is that we always request a new
461     * VUE handle every time, even if GS produces no output.
462     * With this we make sure that whether we have emitted at least one vertex
463     * or none at all, we have to finish the thread without writing to the URB,
464     * which works for both cases by setting the COMPLETE and UNUSED flags in
465     * the EOT message.
466     */
467    this->current_annotation = "gfx6 thread end: EOT";
468 
469    if (gs_prog_data->num_transform_feedback_bindings) {
470       /* When emitting EOT, set SONumPrimsWritten Increment Value. */
471       src_reg data(this, glsl_type::uint_type);
472       emit(AND(dst_reg(data), this->sol_prim_written, brw_imm_ud(0xffffu)));
473       emit(SHL(dst_reg(data), data, brw_imm_ud(16u)));
474       emit(GS_OPCODE_SET_DWORD_2, dst_reg(MRF, base_mrf), data);
475    }
476 
477    inst = emit(GS_OPCODE_THREAD_END);
478    inst->urb_write_flags = BRW_URB_WRITE_COMPLETE | BRW_URB_WRITE_UNUSED;
479    inst->base_mrf = base_mrf;
480    inst->mlen = 1;
481 }
482 
483 void
setup_payload()484 gfx6_gs_visitor::setup_payload()
485 {
486    int attribute_map[BRW_VARYING_SLOT_COUNT * MAX_GS_INPUT_VERTICES];
487 
488    /* Attributes are going to be interleaved, so one register contains two
489     * attribute slots.
490     */
491    int attributes_per_reg = 2;
492 
493    /* If a geometry shader tries to read from an input that wasn't written by
494     * the vertex shader, that produces undefined results, but it shouldn't
495     * crash anything.  So initialize attribute_map to zeros--that ensures that
496     * these undefined results are read from r0.
497     */
498    memset(attribute_map, 0, sizeof(attribute_map));
499 
500    int reg = 0;
501 
502    /* The payload always contains important data in r0. */
503    reg++;
504 
505    /* r1 is always part of the payload and it holds information relevant
506     * for transform feedback when we set the GFX6_GS_SVBI_PAYLOAD_ENABLE bit in
507     * the 3DSTATE_GS packet. We will overwrite it with the PrimitiveID
508     * information (and move the original value to a virtual register if
509     * necessary).
510     */
511    if (gs_prog_data->include_primitive_id)
512       attribute_map[VARYING_SLOT_PRIMITIVE_ID] = attributes_per_reg * reg;
513    reg++;
514 
515    reg = setup_uniforms(reg);
516 
517    reg = setup_varying_inputs(reg, attributes_per_reg);
518 
519    this->first_non_payload_grf = reg;
520 }
521 
522 void
xfb_write()523 gfx6_gs_visitor::xfb_write()
524 {
525    unsigned num_verts;
526 
527    switch (gs_prog_data->output_topology) {
528    case _3DPRIM_POINTLIST:
529       num_verts = 1;
530       break;
531    case _3DPRIM_LINELIST:
532    case _3DPRIM_LINESTRIP:
533    case _3DPRIM_LINELOOP:
534       num_verts = 2;
535       break;
536    case _3DPRIM_TRILIST:
537    case _3DPRIM_TRIFAN:
538    case _3DPRIM_TRISTRIP:
539    case _3DPRIM_RECTLIST:
540       num_verts = 3;
541       break;
542    case _3DPRIM_QUADLIST:
543    case _3DPRIM_QUADSTRIP:
544    case _3DPRIM_POLYGON:
545       num_verts = 3;
546       break;
547    default:
548       unreachable("Unexpected primitive type in Gfx6 SOL program.");
549    }
550 
551    this->current_annotation = "gfx6 thread end: svb writes init";
552 
553    emit(MOV(dst_reg(this->vertex_output_offset), brw_imm_ud(0u)));
554    emit(MOV(dst_reg(this->sol_prim_written), brw_imm_ud(0u)));
555 
556    /* Check that at least one primitive can be written
557     *
558     * Note: since we use the binding table to keep track of buffer offsets
559     * and stride, the GS doesn't need to keep track of a separate pointer
560     * into each buffer; it uses a single pointer which increments by 1 for
561     * each vertex.  So we use SVBI0 for this pointer, regardless of whether
562     * transform feedback is in interleaved or separate attribs mode.
563     */
564    src_reg sol_temp(this, glsl_type::uvec4_type);
565    emit(ADD(dst_reg(sol_temp), this->svbi, brw_imm_ud(num_verts)));
566 
567    /* Compare SVBI calculated number with the maximum value, which is
568     * in R1.4 (previously saved in this->max_svbi) for gfx6.
569     */
570    emit(CMP(dst_null_d(), sol_temp, this->max_svbi, BRW_CONDITIONAL_LE));
571    emit(IF(BRW_PREDICATE_NORMAL));
572    {
573       vec4_instruction *inst = emit(MOV(dst_reg(destination_indices),
574                                         brw_imm_vf4(brw_float_to_vf(0.0),
575                                                     brw_float_to_vf(1.0),
576                                                     brw_float_to_vf(2.0),
577                                                     brw_float_to_vf(0.0))));
578       inst->force_writemask_all = true;
579 
580       emit(ADD(dst_reg(this->destination_indices),
581                this->destination_indices,
582                this->svbi));
583    }
584    emit(BRW_OPCODE_ENDIF);
585 
586    /* Write transform feedback data for all processed vertices. */
587    for (int i = 0; i < (int)nir->info.gs.vertices_out; i++) {
588       emit(MOV(dst_reg(sol_temp), brw_imm_d(i)));
589       emit(CMP(dst_null_d(), sol_temp, this->vertex_count,
590                BRW_CONDITIONAL_L));
591       emit(IF(BRW_PREDICATE_NORMAL));
592       {
593          xfb_program(i, num_verts);
594       }
595       emit(BRW_OPCODE_ENDIF);
596    }
597 }
598 
599 void
xfb_program(unsigned vertex,unsigned num_verts)600 gfx6_gs_visitor::xfb_program(unsigned vertex, unsigned num_verts)
601 {
602    unsigned binding;
603    unsigned num_bindings = gs_prog_data->num_transform_feedback_bindings;
604    src_reg sol_temp(this, glsl_type::uvec4_type);
605 
606    /* Check for buffer overflow: we need room to write the complete primitive
607     * (all vertices). Otherwise, avoid writing any vertices for it
608     */
609    emit(ADD(dst_reg(sol_temp), this->sol_prim_written, brw_imm_ud(1u)));
610    emit(MUL(dst_reg(sol_temp), sol_temp, brw_imm_ud(num_verts)));
611    emit(ADD(dst_reg(sol_temp), sol_temp, this->svbi));
612    emit(CMP(dst_null_d(), sol_temp, this->max_svbi, BRW_CONDITIONAL_LE));
613    emit(IF(BRW_PREDICATE_NORMAL));
614    {
615       /* Avoid overwriting MRF 1 as it is used as URB write message header */
616       dst_reg mrf_reg(MRF, 2);
617 
618       this->current_annotation = "gfx6: emit SOL vertex data";
619       /* For each vertex, generate code to output each varying using the
620        * appropriate binding table entry.
621        */
622       for (binding = 0; binding < num_bindings; ++binding) {
623          unsigned char varying =
624             gs_prog_data->transform_feedback_bindings[binding];
625 
626          /* Set up the correct destination index for this vertex */
627          vec4_instruction *inst = emit(GS_OPCODE_SVB_SET_DST_INDEX,
628                                        mrf_reg,
629                                        this->destination_indices);
630          inst->sol_vertex = vertex % num_verts;
631 
632          /* From the Sandybridge PRM, Volume 2, Part 1, Section 4.5.1:
633           *
634           *   "Prior to End of Thread with a URB_WRITE, the kernel must
635           *   ensure that all writes are complete by sending the final
636           *   write as a committed write."
637           */
638          bool final_write = binding == (unsigned) num_bindings - 1 &&
639                             inst->sol_vertex == num_verts - 1;
640 
641          /* Compute offset of this varying for the current vertex
642           * in vertex_output
643           */
644          this->current_annotation = output_reg_annotation[varying];
645          src_reg data(this->vertex_output);
646          data.reladdr = ralloc(mem_ctx, src_reg);
647          int offset = get_vertex_output_offset_for_varying(vertex, varying);
648          emit(MOV(dst_reg(this->vertex_output_offset), brw_imm_d(offset)));
649          memcpy(data.reladdr, &this->vertex_output_offset, sizeof(src_reg));
650          data.type = output_reg[varying][0].type;
651          data.swizzle = gs_prog_data->transform_feedback_swizzles[binding];
652 
653          /* Write data */
654          inst = emit(GS_OPCODE_SVB_WRITE, mrf_reg, data, sol_temp);
655          inst->sol_binding = binding;
656          inst->sol_final_write = final_write;
657 
658          if (final_write) {
659             /* This is the last vertex of the primitive, then increment
660              * SO num primitive counter and destination indices.
661              */
662             emit(ADD(dst_reg(this->destination_indices),
663                      this->destination_indices,
664                      brw_imm_ud(num_verts)));
665             emit(ADD(dst_reg(this->sol_prim_written),
666                      this->sol_prim_written, brw_imm_ud(1u)));
667          }
668 
669       }
670       this->current_annotation = NULL;
671    }
672    emit(BRW_OPCODE_ENDIF);
673 }
674 
675 int
get_vertex_output_offset_for_varying(int vertex,int varying)676 gfx6_gs_visitor::get_vertex_output_offset_for_varying(int vertex, int varying)
677 {
678    /* Find the output slot assigned to this varying.
679     *
680     * VARYING_SLOT_LAYER and VARYING_SLOT_VIEWPORT are packed in the same slot
681     * as VARYING_SLOT_PSIZ.
682     */
683    if (varying == VARYING_SLOT_LAYER || varying == VARYING_SLOT_VIEWPORT)
684       varying = VARYING_SLOT_PSIZ;
685    int slot = prog_data->vue_map.varying_to_slot[varying];
686 
687    if (slot < 0) {
688       /* This varying does not exist in the VUE so we are not writing to it
689        * and its value is undefined. We still want to return a valid offset
690        * into vertex_output though, to prevent any out-of-bound accesses into
691        * the vertex_output array. Since the value for this varying is undefined
692        * we don't really care for the value we assign to it, so any offset
693        * within the limits of vertex_output will do.
694        */
695       slot = 0;
696    }
697 
698    return vertex * (prog_data->vue_map.num_slots + 1) + slot;
699 }
700 
701 } /* namespace brw */
702