mirror of
https://github.com/facebookresearch/pytorch3d.git
synced 2025-12-20 22:30:35 +08:00
CPU implem for face areas normals
Summary: Added cpu implementation for face areas normals. Moved test and bm to separate functions. ``` Benchmark Avg Time(μs) Peak Time(μs) Iterations -------------------------------------------------------------------------------- FACE_AREAS_NORMALS_2_100_300_False 196 268 2550 FACE_AREAS_NORMALS_2_100_300_True 106 179 4733 FACE_AREAS_NORMALS_2_100_3000_False 1447 1630 346 FACE_AREAS_NORMALS_2_100_3000_True 107 178 4674 FACE_AREAS_NORMALS_2_1000_300_False 201 309 2486 FACE_AREAS_NORMALS_2_1000_300_True 107 186 4673 FACE_AREAS_NORMALS_2_1000_3000_False 1451 1636 345 FACE_AREAS_NORMALS_2_1000_3000_True 107 186 4655 FACE_AREAS_NORMALS_10_100_300_False 767 918 653 FACE_AREAS_NORMALS_10_100_300_True 106 167 4712 FACE_AREAS_NORMALS_10_100_3000_False 7036 7754 72 FACE_AREAS_NORMALS_10_100_3000_True 113 164 4445 FACE_AREAS_NORMALS_10_1000_300_False 748 947 669 FACE_AREAS_NORMALS_10_1000_300_True 108 169 4638 FACE_AREAS_NORMALS_10_1000_3000_False 7069 7783 71 FACE_AREAS_NORMALS_10_1000_3000_True 108 172 4646 FACE_AREAS_NORMALS_32_100_300_False 2286 2496 219 FACE_AREAS_NORMALS_32_100_300_True 108 180 4631 FACE_AREAS_NORMALS_32_100_3000_False 23184 24369 22 FACE_AREAS_NORMALS_32_100_3000_True 159 213 3147 FACE_AREAS_NORMALS_32_1000_300_False 2414 2645 208 FACE_AREAS_NORMALS_32_1000_300_True 112 197 4480 FACE_AREAS_NORMALS_32_1000_3000_False 21687 22964 24 FACE_AREAS_NORMALS_32_1000_3000_True 141 211 3540 -------------------------------------------------------------------------------- Benchmark Avg Time(μs) Peak Time(μs) Iterations -------------------------------------------------------------------------------- FACE_AREAS_NORMALS_TORCH_2_100_300_False 5465 5782 92 FACE_AREAS_NORMALS_TORCH_2_100_300_True 1198 1351 418 FACE_AREAS_NORMALS_TORCH_2_100_3000_False 48228 48869 11 FACE_AREAS_NORMALS_TORCH_2_100_3000_True 1186 1304 422 FACE_AREAS_NORMALS_TORCH_2_1000_300_False 5556 6097 90 FACE_AREAS_NORMALS_TORCH_2_1000_300_True 1200 1328 417 FACE_AREAS_NORMALS_TORCH_2_1000_3000_False 48683 50016 11 FACE_AREAS_NORMALS_TORCH_2_1000_3000_True 1185 1306 422 FACE_AREAS_NORMALS_TORCH_10_100_300_False 24215 25097 21 FACE_AREAS_NORMALS_TORCH_10_100_300_True 1150 1314 435 FACE_AREAS_NORMALS_TORCH_10_100_3000_False 232605 234952 3 FACE_AREAS_NORMALS_TORCH_10_100_3000_True 1193 1314 420 FACE_AREAS_NORMALS_TORCH_10_1000_300_False 24912 25343 21 FACE_AREAS_NORMALS_TORCH_10_1000_300_True 1216 1330 412 FACE_AREAS_NORMALS_TORCH_10_1000_3000_False 239907 241253 3 FACE_AREAS_NORMALS_TORCH_10_1000_3000_True 1226 1333 408 FACE_AREAS_NORMALS_TORCH_32_100_300_False 73991 75776 7 FACE_AREAS_NORMALS_TORCH_32_100_300_True 1193 1339 420 FACE_AREAS_NORMALS_TORCH_32_100_3000_False 728932 728932 1 FACE_AREAS_NORMALS_TORCH_32_100_3000_True 1186 1359 422 FACE_AREAS_NORMALS_TORCH_32_1000_300_False 76385 79129 7 FACE_AREAS_NORMALS_TORCH_32_1000_300_True 1165 1310 430 FACE_AREAS_NORMALS_TORCH_32_1000_3000_False 753276 753276 1 FACE_AREAS_NORMALS_TORCH_32_1000_3000_True 1205 1340 415 -------------------------------------------------------------------------------- ``` Reviewed By: bottler, jcjohnson Differential Revision: D19864385 fbshipit-source-id: 3a87ae41a8e3ab5560febcb94961798f2e09dfb8
This commit is contained in:
Georgia Gkioxari
committed by
Facebook Github Bot
Facebook Github Bot
parent
8fe65d5f56
commit
29cd181a83
@@ -9,7 +9,7 @@
|
||||
#include "rasterize_points/rasterize_points.h"
|
||||
|
||||
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
|
||||
m.def("face_areas_normals", &face_areas_normals);
|
||||
m.def("face_areas_normals", &FaceAreasNormals);
|
||||
m.def("packed_to_padded_tensor", &packed_to_padded_tensor);
|
||||
m.def("nn_points_idx", &NearestNeighborIdx);
|
||||
m.def("gather_scatter", &gather_scatter);
|
||||
|
||||
@@ -4,7 +4,7 @@
|
||||
#include <tuple>
|
||||
|
||||
template <typename scalar_t>
|
||||
__global__ void face_areas_kernel(
|
||||
__global__ void FaceAreasNormalsKernel(
|
||||
const scalar_t* __restrict__ verts,
|
||||
const long* __restrict__ faces,
|
||||
scalar_t* __restrict__ face_areas,
|
||||
@@ -55,7 +55,7 @@ __global__ void face_areas_kernel(
|
||||
}
|
||||
}
|
||||
|
||||
std::tuple<at::Tensor, at::Tensor> face_areas_cuda(
|
||||
std::tuple<at::Tensor, at::Tensor> FaceAreasNormalsCuda(
|
||||
at::Tensor verts,
|
||||
at::Tensor faces) {
|
||||
const auto V = verts.size(0);
|
||||
@@ -66,14 +66,15 @@ std::tuple<at::Tensor, at::Tensor> face_areas_cuda(
|
||||
|
||||
const int blocks = 64;
|
||||
const int threads = 512;
|
||||
AT_DISPATCH_FLOATING_TYPES(verts.type(), "face_areas_kernel", ([&] {
|
||||
face_areas_kernel<scalar_t><<<blocks, threads>>>(
|
||||
verts.data_ptr<scalar_t>(),
|
||||
faces.data_ptr<long>(),
|
||||
areas.data_ptr<scalar_t>(),
|
||||
normals.data_ptr<scalar_t>(),
|
||||
V,
|
||||
F);
|
||||
AT_DISPATCH_FLOATING_TYPES(verts.type(), "face_areas_normals_cuda", ([&] {
|
||||
FaceAreasNormalsKernel<scalar_t>
|
||||
<<<blocks, threads>>>(
|
||||
verts.data_ptr<scalar_t>(),
|
||||
faces.data_ptr<long>(),
|
||||
areas.data_ptr<scalar_t>(),
|
||||
normals.data_ptr<scalar_t>(),
|
||||
V,
|
||||
F);
|
||||
}));
|
||||
|
||||
return std::make_tuple(areas, normals);
|
||||
|
||||
@@ -16,21 +16,26 @@
|
||||
// faces[f]
|
||||
//
|
||||
|
||||
// Cpu implementation.
|
||||
std::tuple<at::Tensor, at::Tensor> FaceAreasNormalsCpu(
|
||||
at::Tensor verts,
|
||||
at::Tensor faces);
|
||||
|
||||
// Cuda implementation.
|
||||
std::tuple<at::Tensor, at::Tensor> face_areas_cuda(
|
||||
std::tuple<at::Tensor, at::Tensor> FaceAreasNormalsCuda(
|
||||
at::Tensor verts,
|
||||
at::Tensor faces);
|
||||
|
||||
// Implementation which is exposed.
|
||||
std::tuple<at::Tensor, at::Tensor> face_areas_normals(
|
||||
std::tuple<at::Tensor, at::Tensor> FaceAreasNormals(
|
||||
at::Tensor verts,
|
||||
at::Tensor faces) {
|
||||
if (verts.type().is_cuda() && faces.type().is_cuda()) {
|
||||
#ifdef WITH_CUDA
|
||||
return face_areas_cuda(verts, faces);
|
||||
return FaceAreasNormalsCuda(verts, faces);
|
||||
#else
|
||||
AT_ERROR("Not compiled with GPU support.");
|
||||
#endif
|
||||
}
|
||||
AT_ERROR("Not implemented on the CPU.");
|
||||
return FaceAreasNormalsCpu(verts, faces);
|
||||
}
|
||||
|
||||
57
pytorch3d/csrc/face_areas_normals/face_areas_normals_cpu.cpp
Normal file
57
pytorch3d/csrc/face_areas_normals/face_areas_normals_cpu.cpp
Normal file
@@ -0,0 +1,57 @@
|
||||
// Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
|
||||
|
||||
#include <torch/extension.h>
|
||||
#include <tuple>
|
||||
|
||||
std::tuple<at::Tensor, at::Tensor> FaceAreasNormalsCpu(
|
||||
at::Tensor verts,
|
||||
at::Tensor faces) {
|
||||
const int V = verts.size(0);
|
||||
const int F = faces.size(0);
|
||||
|
||||
at::Tensor areas = at::empty({F}, verts.options());
|
||||
at::Tensor normals = at::empty({F, 3}, verts.options());
|
||||
|
||||
auto verts_a = verts.accessor<float, 2>();
|
||||
auto faces_a = faces.accessor<int64_t, 2>();
|
||||
auto areas_a = areas.accessor<float, 1>();
|
||||
auto normals_a = normals.accessor<float, 2>();
|
||||
|
||||
for (int f = 0; f < F; ++f) {
|
||||
const int64_t i0 = faces_a[f][0];
|
||||
const int64_t i1 = faces_a[f][1];
|
||||
const int64_t i2 = faces_a[f][2];
|
||||
|
||||
const float v0_x = verts_a[i0][0];
|
||||
const float v0_y = verts_a[i0][1];
|
||||
const float v0_z = verts_a[i0][2];
|
||||
|
||||
const float v1_x = verts_a[i1][0];
|
||||
const float v1_y = verts_a[i1][1];
|
||||
const float v1_z = verts_a[i1][2];
|
||||
|
||||
const float v2_x = verts_a[i2][0];
|
||||
const float v2_y = verts_a[i2][1];
|
||||
const float v2_z = verts_a[i2][2];
|
||||
|
||||
const float ax = v1_x - v0_x;
|
||||
const float ay = v1_y - v0_y;
|
||||
const float az = v1_z - v0_z;
|
||||
|
||||
const float bx = v2_x - v0_x;
|
||||
const float by = v2_y - v0_y;
|
||||
const float bz = v2_z - v0_z;
|
||||
|
||||
const float cx = ay * bz - az * by;
|
||||
const float cy = az * bx - ax * bz;
|
||||
const float cz = ax * by - ay * bx;
|
||||
|
||||
float norm = sqrt(cx * cx + cy * cy + cz * cz);
|
||||
areas_a[f] = norm / 2.0;
|
||||
norm = (norm < 1e-6) ? 1e-6 : norm; // max(norm, 1e-6)
|
||||
normals_a[f][0] = cx / norm;
|
||||
normals_a[f][1] = cy / norm;
|
||||
normals_a[f][2] = cz / norm;
|
||||
}
|
||||
return std::make_tuple(areas, normals);
|
||||
}
|
||||
Reference in New Issue
Block a user