sail-sg · peilinrao · Nov 1, 2022 · Nov 3, 2022 · Nov 11, 2022 · Nov 18, 2022
diff --git a/.bazelrc b/.bazelrc
@@ -1,7 +1,9 @@
+startup --windows_enable_symlinks
+build --enable_runfiles
 build --action_env=BAZEL_LINKLIBS=-l%:libstdc++.a:-lm
 build --action_env=BAZEL_LINKOPTS=-static-libgcc
-build --action_env=CUDA_DIR=/usr/local/cuda
-build --incompatible_strict_action_env --cxxopt=-std=c++17 --host_cxxopt=-std=c++17 --client_env=BAZEL_CXXOPTS=-std=c++17
+# build --action_env=CUDA_DIR=/usr/local/cuda
+build --incompatible_strict_action_env --cxxopt=/std:c++17 --host_cxxopt=/std:c++17 --client_env=BAZEL_CXXOPTS=/std:c++17
 build:debug --cxxopt=-DENVPOOL_TEST --compilation_mode=dbg -s
 build:test --cxxopt=-DENVPOOL_TEST --copt=-g0 --copt=-O3 --copt=-DNDEBUG --copt=-msse --copt=-msse2 --copt=-mmmx
 build:release --copt=-g0 --copt=-O3 --copt=-DNDEBUG --copt=-msse --copt=-msse2 --copt=-mmmx

diff --git a/BUILD b/BUILD
@@ -22,3 +22,4 @@ py_binary(
         requirement("wheel"),
     ],
 )
+
diff --git a/WINDOWS.md b/WINDOWS.md
@@ -0,0 +1,30 @@
+# Windows Migration
+
+Note that this merge request only intends to work on Windows. I haven't write code that switch support between linux and Windows. I'll update that feature after the whole thing can be build on Windows.
+
+## Setup Environment On Windows
+
+1. Install [Windows Powershell](https://learn.microsoft.com/en-us/powershell/scripting/install/installing-powershell-on-windows?view=powershell-7.2). Do step 2 and 3 in Powershell.
+2. Install [choco](https://chocolatey.org/install#individual)
+3. Install bazelisk/bazel ``` choco install bazelisk```
+4. Install [MinGW-w64](https://www.mingw-w64.org/)
+5. Install [git bash](https://gitforwindows.org/) and make sure it uses MINGW64 emulation. Do all the following steps in bash.
+6. Install [VC++ 2017](https://www.google.com/url?q=https://aka.ms/vs/15/release/vs_buildtools.exe&sa=D&source=docs&ust=1667504254128145&usg=AOvVaw3jOTmoHfuzAHgaScSgsmlE) on Windows
+7. Install python 3.10 (instead of 3.11 because pygame installation fails on Windows)
+8. Go where python.exe is installed, make a copy of python.exe and rename it python3.exe (it would bypass the bug "python interpreter cannot be found with requirement")
+9. Install [swig.exe](https://sourceforge.net/projects/swig/files/swigwin/swigwin-4.1.0/swigwin-4.1.0.zip/download?use_mirror=gigenet) and merge it into python folder and do [some hacks](https://stackoverflow.com/questions/44504899/installing-pocketsphinx-python-module-command-swig-exe-failed)
+10. Install [Clang](https://bazel.build/configure/windows#using) and follow this [link](https://github.com/llvm/llvm-project/releases/tag/llvmorg-15.0.2). Don't forget to set BAZEL_VC and BAZEL_LLVM in environmental variable manually.
+11. Now you can run some tests  
+```bazel build --test_output=all //envpool/dummy:dummy_envpool_test --config=debug --compiler=clang-cl```
+
+
+
+## Parts that built successfully
+- [x] Dummy
+- [x] Utils
+- [x] Box2d
+- [x] Classic Control
+- [x] Mujoco:mujoco_gym-env
+- [ ] Mujoco:mujoco_dmc_env:  no matching function for call to 'mj_loadXML', use of undeclared identifier 'M_PI'...
+- [ ] Atari:atari_env  #include <dlfcn.h> not found
+- [ ] Toy_text fatal error: 'cuda_runtime_api.h' file not found
diff --git a/docs/content/build.rst b/docs/content/build.rst
@@ -86,7 +86,7 @@ It also requires **Python version >= 3.7**:
     sudo apt install -y python3-dev python3-pip
     sudo ln -sf /usr/bin/python3 /usr/bin/python
 
-Install CUDA to enable XLA: see https://developer.nvidia.com/cuda-downloads
+Install CUDA to enable XLA: see https://developer.nvidia.com/-downloads
 
 Install other dependencies: see
 `Dockerfile <https://github.com/sail-sg/envpool/tree/main/docker>`_.

diff --git a/envpool/core/BUILD b/envpool/core/BUILD
@@ -171,6 +171,7 @@ cc_library(
     ],
 )
 
+# if windows, makes up a dummy xla without cuda dependency
 cc_library(
     name = "xla_template",
     hdrs = ["xla_template.h"],
@@ -183,6 +184,7 @@ cc_library(
     ],
 )
 
+
 cc_library(
     name = "xla",
     hdrs = ["xla.h"],
@@ -198,6 +200,6 @@ pybind_library(
     hdrs = ["py_envpool.h"],
     deps = [
         ":envpool",
-        ":xla",
+        ":xla"
     ],
 )
diff --git a/envpool/core/async_envpool.h b/envpool/core/async_envpool.h
@@ -106,14 +106,16 @@ class AsyncEnvPool : public EnvPool<typename Env::Spec> {
     if (spec.config["thread_affinity_offset"_] >= 0) {
       std::size_t thread_affinity_offset =
           spec.config["thread_affinity_offset"_];
-      for (std::size_t tid = 0; tid < num_threads_; ++tid) {
-        cpu_set_t cpuset;
-        CPU_ZERO(&cpuset);
-        std::size_t cid = (thread_affinity_offset + tid) % processor_count;
-        CPU_SET(cid, &cpuset);
-        pthread_setaffinity_np(workers_[tid].native_handle(), sizeof(cpu_set_t),
-                               &cpuset);
-      }
+
+      // The following does not work under Windows. Need to find alternatives.
+      // for (std::size_t tid = 0; tid < num_threads_; ++tid) {
+      //   cpu_set_t cpuset;
+      //   CPU_ZERO(&cpuset);
+      //   std::size_t cid = (thread_affinity_offset + tid) % processor_count;
+      //   CPU_SET(cid, &cpuset);
+      //   pthread_setaffinity_np(workers_[tid].native_handle(), sizeof(cpu_set_t),
+      //                          &cpuset);
+      // }
     }
   }
 

diff --git a/envpool/core/state_buffer_queue.h b/envpool/core/state_buffer_queue.h
@@ -83,7 +83,10 @@ class StateBufferQueue {
     }
     std::size_t processor_count = std::thread::hardware_concurrency();
     // hardcode here :(
-    std::size_t create_buffer_thread_num = std::max(1UL, processor_count / 64);
+    std::size_t create_buffer_thread_num = 1;
+    if (processor_count / 64 > create_buffer_thread_num){
+      create_buffer_thread_num = processor_count / 64;
+    }
     for (std::size_t i = 0; i < create_buffer_thread_num; ++i) {
       create_buffer_thread_.emplace_back(std::thread([&]() {
         while (true) {

diff --git a/envpool/core/xla.h b/envpool/core/xla.h
@@ -17,7 +17,9 @@
 #ifndef ENVPOOL_CORE_XLA_H_
 #define ENVPOOL_CORE_XLA_H_
 
-#include <cuda_runtime_api.h>
+#include "envpool/core/array.h"
+#include "envpool/core/xla_template.h"
+#include "cuda_runtime_api.h"
 
 #include <cstdint>
 #include <memory>
@@ -26,8 +28,8 @@
 #include <tuple>
 #include <vector>
 
-#include "envpool/core/array.h"
-#include "envpool/core/xla_template.h"
+
+
 
 template <typename D>
 constexpr bool is_container_v = false;  // NOLINT

diff --git a/envpool/core/xla_template.h b/envpool/core/xla_template.h
@@ -17,7 +17,7 @@
 #ifndef ENVPOOL_CORE_XLA_TEMPLATE_H_
 #define ENVPOOL_CORE_XLA_TEMPLATE_H_
 
-#include <cuda_runtime_api.h>
+#include "cuda_runtime_api.h"
 #include <pybind11/functional.h>
 #include <pybind11/numpy.h>
 #include <pybind11/pybind11.h>
@@ -109,6 +109,7 @@ struct CustomCall {
     return std::make_tuple(
         py::capsule(reinterpret_cast<void*>(Cpu), "xla._CUSTOM_CALL_TARGET"),
         py::capsule(reinterpret_cast<void*>(Gpu), "xla._CUSTOM_CALL_TARGET"));
+
   }
 
   static auto Xla(Class* obj) {

diff --git a/envpool/mujoco/dmc/mujoco_env.h b/envpool/mujoco/dmc/mujoco_env.h
@@ -23,6 +23,7 @@
 #include <memory>
 #include <random>
 #include <string>
+#include <array>
 
 #include "envpool/mujoco/dmc/utils.h"
 

diff --git a/envpool/toy_text/BUILD b/envpool/toy_text/BUILD
@@ -26,6 +26,7 @@ pybind_extension(
     deps = [
         ":toy_text_env",
         "//envpool/core:py_envpool",
+        "//envpool/core:xla",
     ],
 )
 

diff --git a/envpool/workspace0.bzl b/envpool/workspace0.bzl
@@ -18,6 +18,7 @@ load("@bazel_tools//tools/build_defs/repo:http.bzl", "http_archive")
 load("@bazel_tools//tools/build_defs/repo:utils.bzl", "maybe")
 load("//third_party/cuda:cuda.bzl", "cuda_configure")
 
+
 def workspace():
     """Load requested packages."""
     maybe(
@@ -318,13 +319,21 @@ def workspace():
 
     maybe(
         http_archive,
-        name = "mujoco",
+        name = "mujoco_linux",
         sha256 = "d1cb3a720546240d894cd315b7fd358a2b96013a1f59b6d718036eca6f6edac2",
         strip_prefix = "mujoco-2.2.1",
         urls = [
-            "https://github.com/deepmind/mujoco/releases/download/2.2.1/mujoco-2.2.1-linux-x86_64.tar.gz",
-            "https://ml.cs.tsinghua.edu.cn/~jiayi/envpool/deepmind/mujoco/mujoco-2.2.1-linux-x86_64.tar.gz",
-        ],
+                "https://github.com/deepmind/mujoco/releases/download/2.2.1/mujoco-2.2.1-linux-x86_64.tar.gz",
+                "https://ml.cs.tsinghua.edu.cn/~jiayi/envpool/deepmind/mujoco/mujoco-2.2.1-linux-x86_64.tar.gz",
+            ],
+        build_file = "//third_party/mujoco:mujoco.BUILD",
+    )
+
+    maybe(
+        http_archive,
+        name = "mujoco",
-        name = "mujoco",
+        name = "mujoco_windows",
-        name = "mujoco",
+        name = "mujoco_windows",
+        sha256 = "e0068d25fdb00ec3afa53e429b5eff438d222455603bafa080086abaf11fe16f",
+        urls = ["https://github.com/deepmind/mujoco/releases/download/2.2.1/mujoco-2.2.1-windows-x86_64.zip"],
         build_file = "//third_party/mujoco:mujoco.BUILD",
     )
 

diff --git a/third_party/cuda/cuda.bzl b/third_party/cuda/cuda.bzl
@@ -17,16 +17,18 @@
 _CUDA_DIR = "CUDA_DIR"
 
 def _impl(rctx):
-    cuda_dir = rctx.os.environ.get(_CUDA_DIR, default = "/usr/local/cuda")
+    cuda_dir = "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.7"
+    print("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!")
+    # rctx.os.environ.get(_CUDA_DIR, default = "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.7")
     rctx.symlink("{}/include".format(cuda_dir), "include")
-    rctx.symlink("{}/lib64".format(cuda_dir), "lib64")
+    rctx.symlink("{}/lib".format(cuda_dir), "lib")
     rctx.file("WORKSPACE")
     rctx.file("BUILD", content = """
 package(default_visibility = ["//visibility:public"])
 
 cc_library(
     name = "cudart_static",
-    srcs = ["lib64/libcudart_static.a"],
+    srcs = ["lib/x64/cudart_static.lib"],
     hdrs = glob([
         "include/*.h",
         "include/**/*.h",

diff --git a/third_party/vizdoom/vizdoom.BUILD b/third_party/vizdoom/vizdoom.BUILD
@@ -52,7 +52,6 @@ cc_library(
     hdrs = glob(["gdtoa/*.h"]),
     copts = [
         "-Wall",
-        "-Wextra",
         "-DINFNAN_CHECK",
         "-DMULTIPLE_THREADS",
         "-include $(execpath @glibc_version_header//:glibc_2_17)",
@@ -78,7 +77,6 @@ cc_library(
     copts = [
         "-DBZ_NO_STDIO",
         "-Wall",
-        "-Wextra",
         "-fomit-frame-pointer",
         "-include $(execpath @glibc_version_header//:glibc_2_17)",
     ],
@@ -120,7 +118,6 @@ cc_library(
     copts = [
         "-D_7ZIP_ST",
         "-Wall",
-        "-Wextra",
         "-fomit-frame-pointer",
         "-D_7ZIP_PPMD_SUPPPORT",
         "-include $(execpath @glibc_version_header//:glibc_2_17)",
@@ -211,7 +208,6 @@ cc_library(
     copts = [
         "-DNEED_ITOA=1",
         "-Wall",
-        "-Wno-pointer-sign",
         "-Wno-uninitialized",
         "-Wno-unused-but-set-variable",
         "-include $(execpath @glibc_version_header//:glibc_2_17)",

diff --git a/third_party/zlib/zlib.BUILD b/third_party/zlib/zlib.BUILD
@@ -33,7 +33,6 @@ cc_library(
     ],
     hdrs = ["zlib.h"],
     copts = [
-        "-Wno-shift-negative-value",
         "-DZ_HAVE_UNISTD_H",
     ],
     includes = ["."],
Original file line number	Diff line number	Diff line change
Expand Up		@@ -22,3 +22,4 @@ py_binary(
		requirement("wheel"),
		],
		)