blender
/
blender
136
407
Fork 591
Code Issues 6.3k Pull Requests 777 Projects 19 Wiki Activity

WIP: Improved parallelization of subdivision #113382

Closed
Eugene-Kuznetsov wants to merge 1 commits from Eugene-Kuznetsov/blender:ek_mt_subdiv into main
pull from: Eugene-Kuznetsov/blender:ek_mt_subdiv
merge into: blender:main

When changing the target branch, be careful to rebase the branch in your fork to match. See documentation.
Conversation 5 Commits 1 Files Changed 2 +51 -24
2 changed files with 51 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
intern/opensubdiv/internal/evaluator/eval_output_cpu.h
View File

@ -9,12 +9,12 @@
#include "internal/evaluator/eval_output.h"
#include <opensubdiv/osd/cpuEvaluator.h>
#include <opensubdiv/osd/tbbEvaluator.h>
#include <opensubdiv/osd/cpuPatchTable.h>
#include <opensubdiv/osd/cpuVertexBuffer.h>
using OpenSubdiv::Far::StencilTable;
using OpenSubdiv::Osd::CpuEvaluator;
using CpuEvaluator = OpenSubdiv::Osd::TbbEvaluator;
using OpenSubdiv::Osd::CpuVertexBuffer;
namespace blender {

71
source/blender/blenkernel/intern/subdiv_foreach.cc
View File

@ -395,14 +395,22 @@ static void subdiv_foreach_every_corner_vertices(SubdivForeachTaskContext *ctx,
return;
}
const Mesh *coarse_mesh = ctx->coarse_mesh;
for (int face_index = 0; face_index < coarse_mesh->faces_num; face_index++) {
if (ctx->coarse_faces[face_index].size() == 4) {
subdiv_foreach_every_corner_vertices_regular(ctx, tls, face_index);
}
else {
subdiv_foreach_every_corner_vertices_special(ctx, tls, face_index);
}
}
TaskParallelSettings parallel_range_settings;
BLI_parallel_range_settings_defaults(&parallel_range_settings);
parallel_range_settings.userdata_chunk = tls;
parallel_range_settings.userdata_chunk_size = ctx->foreach_context->user_data_tls_size;
BLI_task_parallel_range(
0, coarse_mesh->faces_num, ctx,
[](void* _ctx, int face_index, const TaskParallelTLS* _tls) {
SubdivForeachTaskContext *ctx = (SubdivForeachTaskContext*) _ctx;
void* tls = const_cast<TaskParallelTLS*>(_tls)->userdata_chunk;
if (ctx->coarse_faces[face_index].size() == 4) {
subdiv_foreach_every_corner_vertices_regular(ctx, tls, face_index);
}
else {
subdiv_foreach_every_corner_vertices_special(ctx, tls, face_index);
}
}, &parallel_range_settings);
}
/* Traverse of edge vertices. They are coming from coarse edges. */
@ -574,14 +582,24 @@ static void subdiv_foreach_every_edge_vertices(SubdivForeachTaskContext *ctx, vo
return;
}
const Mesh *coarse_mesh = ctx->coarse_mesh;
for (int face_index = 0; face_index < coarse_mesh->faces_num; face_index++) {
if (ctx->coarse_faces[face_index].size() == 4) {
subdiv_foreach_every_edge_vertices_regular(ctx, tls, face_index);
}
else {
subdiv_foreach_every_edge_vertices_special(ctx, tls, face_index);
}
}
TaskParallelSettings parallel_range_settings;
BLI_parallel_range_settings_defaults(&parallel_range_settings);
parallel_range_settings.userdata_chunk = tls;
parallel_range_settings.userdata_chunk_size = ctx->foreach_context->user_data_tls_size;
BLI_task_parallel_range(
0, coarse_mesh->faces_num, ctx,
[](void* _ctx, int face_index, const TaskParallelTLS* _tls) {
SubdivForeachTaskContext *ctx = (SubdivForeachTaskContext*) _ctx;
void* tls = const_cast<TaskParallelTLS*>(_tls)->userdata_chunk;
if (ctx->coarse_faces[face_index].size() == 4) {
subdiv_foreach_every_edge_vertices_regular(ctx, tls, face_index);
}
else {
subdiv_foreach_every_edge_vertices_special(ctx, tls, face_index);
}
},
&parallel_range_settings);
}
/* Traversal of inner vertices, they are coming from ptex patches. */
@ -1725,10 +1743,20 @@ static void subdiv_foreach_single_geometry_vertices(SubdivForeachTaskContext *ct
return;
}
const Mesh *coarse_mesh = ctx->coarse_mesh;
for (int face_index = 0; face_index < coarse_mesh->faces_num; face_index++) {
subdiv_foreach_corner_vertices(ctx, tls, face_index);
subdiv_foreach_edge_vertices(ctx, tls, face_index);
}
TaskParallelSettings parallel_range_settings;
BLI_parallel_range_settings_defaults(&parallel_range_settings);
parallel_range_settings.userdata_chunk = tls;
parallel_range_settings.userdata_chunk_size = ctx->foreach_context->user_data_tls_size;
BLI_task_parallel_range(
0, coarse_mesh->faces_num, ctx,
[](void* _ctx, int face_index, const TaskParallelTLS* _tls) {
SubdivForeachTaskContext *ctx = (SubdivForeachTaskContext*) _ctx;
void* tls = const_cast<TaskParallelTLS*>(_tls)->userdata_chunk;
subdiv_foreach_corner_vertices(ctx, tls, face_index);
subdiv_foreach_edge_vertices(ctx, tls, face_index);
},
&parallel_range_settings);
}
static void subdiv_foreach_mark_non_loose_geometry(SubdivForeachTaskContext *ctx)
@ -1746,14 +1774,13 @@ static void subdiv_foreach_single_thread_tasks(SubdivForeachTaskContext *ctx)
/* NOTE: In theory, we can try to skip allocation of TLS here, but in
* practice if the callbacks used here are not specified then TLS will not
* be requested anyway. */
void *tls = subdiv_foreach_tls_alloc(ctx);
void *tls = ctx->foreach_context->user_data_tls;
/* Passes to average displacement on the corner vertices
* and boundary edges. */
subdiv_foreach_every_corner_vertices(ctx, tls);
subdiv_foreach_every_edge_vertices(ctx, tls);
/* Run callbacks which are supposed to be run once per shared geometry. */
subdiv_foreach_single_geometry_vertices(ctx, tls);
subdiv_foreach_tls_free(ctx, tls);
const SubdivForeachContext *foreach_context = ctx->foreach_context;
const bool is_loose_geometry_tagged = (foreach_context->vertex_every_edge != nullptr &&