misc/pytorch: Tensors and dynamic neural networks in Python (C++ library)

misc/Makefile

@@ -445,6 +445,7 @@
     SUBDIR += py-wurlitzer
     SUBDIR += py-xgboost
     SUBDIR += pyobd
+    SUBDIR += pytorch
     SUBDIR += qbrew
     SUBDIR += qlas
     SUBDIR += qt5-doc

misc/pytorch/Makefile

@@ -0,0 +1,63 @@
+PORTNAME=	pytorch
+DISTVERSIONPREFIX=	v
+DISTVERSION=	1.13.1
+CATEGORIES=	misc # machine-learning
+MASTER_SITES=	https://github.com/pytorch/pytorch/releases/download/v1.13.1/
+DIST_SUBDIR=	${PORTNAME}
+
+MAINTAINER=	yuri@FreeBSD.org
+COMMENT=	Tensors and dynamic neural networks in Python (C++ library)
+WWW=		https://pytorch.org/
+
+LICENSE=	BSD3CLAUSE
+
+BUILD_DEPENDS=	gmake:devel/gmake \
+		${LOCALBASE}/include/fxdiv.h:devel/fxdiv
+LIB_DEPENDS=	libopenblas.so:math/openblas \
+		libmpi.so:net/openmpi \
+		libpthreadpool.so:devel/pthreadpool \
+		libonnx.so:misc/onnx \
+		libprotobuf.so:devel/protobuf \
+		libsleef.so:math/sleef
+
+USES=		compiler:c++14-lang cmake localbase:ldflags python # requires python even with PYTHON=off
+
+CMAKE_OFF=	BUILD_CUSTOM_PROTOBUF USE_CUDA USE_ROCM  USE_NNPACK USE_QNNPACK USE_PYTORCH_QNNPACK \
+		USE_FBGEMM # workaround recommended by the upstream here: https://github.com/pytorch/pytorch/issues/28337
+CMAKE_ON=	USE_SYSTEM_PYBIND11 \
+		USE_SYSTEM_SLEEF \
+		USE_SYSTEM_ONNX # see other USE_SYSTEM_xx in CMakeLists.txt
+CMAKE_ARGS=	-DPSIMD_SOURCE_DIR=${WRKSRC}/third_party/psimd \
+		-DFREEBSD_PYTHON_VER=${PYTHON_VER} \
+		-DPYTHON_EXECUTABLE:STRING=${PYTHON_CMD}
+
+MAKE_ENV=	USE_NINJA=no # ninja breaks for some reason
+LDFLAGS+=	-lexecinfo
+
+BINARY_ALIAS=	make=${GMAKE} git=true
+
+OPTIONS_DEFINE=		PYTHON # PYTHON only provides python support for the py-PyTorch port - it's not a python binding
+OPTIONS_DEFAULT=	PYTHON
+OPTIONS_SUB=		yes
+
+PYTHON_CMAKE_BOOL=	BUILD_PYTHON
+PYTHON_PY_DEPENDS=	${PYNUMPY} \
+			${PYTHON_PKGNAMEPREFIX}typing-extensions>0:devel/py-typing-extensions@${PY_FLAVOR} \
+			${PYTHON_PKGNAMEPREFIX}yaml>0:devel/py-yaml@${PY_FLAVOR}
+PYTHON_BUILD_DEPENDS=	${PYTHON_PY_DEPENDS} \
+			pybind11>0:devel/pybind11
+PYTHON_RUN_DEPENDS=	${PYTHON_PY_DEPENDS}
+
+post-patch:
+	# replace malloc.h -> stdlib.h in many places
+	@cd ${WRKSRC} && for f in $$(${GREP} -rl "include <malloc\.h>" * | ${GREP} -E "\.(c|cpp|h|hpp)$$"); do \
+	  ${REINPLACE_CMD} -i'' -E "s|include <malloc\.h>|include <stdlib.h>|" $$f; \
+	done
+
+post-install:
+	# functorch.so is installed back into the source directory, see https://github.com/pytorch/pytorch/issues/91524
+	@${RM} $$(${FIND} ${STAGEDIR} -name functorch.so)
+	# https://github.com/pytorch/pytorch/issues/24417
+	@${FIND} ${STAGEDIR} -type d -empty -delete
+
+.include <bsd.port.mk>

misc/pytorch/distinfo

@@ -0,0 +1,3 @@
+TIMESTAMP = 1672353185
+SHA256 (pytorch/pytorch-v1.13.1.tar.gz) = dbc229ee9750b02b514937d017744443a269ea0241ed3f32b9af0703589d25d4
+SIZE (pytorch/pytorch-v1.13.1.tar.gz) = 234057741

misc/pytorch/files/patch-aten_src_ATen_native_SoftMax.cpp

@@ -0,0 +1,11 @@
+--- aten/src/ATen/native/SoftMax.cpp.orig	2022-12-29 23:05:30 UTC
++++ aten/src/ATen/native/SoftMax.cpp
+@@ -132,7 +132,7 @@ void host_softmax(
+     const Tensor& input,
+     const int64_t dim,
+     bool* mask = nullptr,
+-    const c10::optional<int64_t> mask_type_ = NULL) {
++    const c10::optional<int64_t> mask_type_ = 0) {
+ 
+   if (MaskedSoftMax) {
+     TORCH_CHECK(mask_type_.has_value(), "Mask Type should be defined");

misc/pytorch/files/patch-aten_src_ATen_native_sparse_ValidateCompressedIndicesCommon.h

@@ -0,0 +1,80 @@
+--- aten/src/ATen/native/sparse/ValidateCompressedIndicesCommon.h.orig	2022-12-29 23:09:42 UTC
++++ aten/src/ATen/native/sparse/ValidateCompressedIndicesCommon.h
+@@ -38,7 +38,7 @@ namespace {
+ // respectively.
+ 
+ INVARIANT_CHECK_FUNC_API
+-_assert(const bool cond, const char* const message) {
++__assert__(const bool cond, const char* const message) {
+ #ifdef GPUCC
+   CUDA_KERNEL_ASSERT(cond && message);
+ #else
+@@ -58,10 +58,10 @@ INVARIANT_CHECK_FUNC_API
+ _check_first_cidx_is_zero(const index_t& cidx, const index_t& zero) {
+   const bool invariant = cidx == zero;
+   if (cdim_name == CDimName::CRow) {
+-    _assert(invariant, "`crow_indices[..., 0] == 0` is not satisfied.");
++    __assert__(invariant, "`crow_indices[..., 0] == 0` is not satisfied.");
+   }
+   else {
+-    _assert(invariant, "`ccol_indices[..., 0] == 0` is not satisfied.");
++    __assert__(invariant, "`ccol_indices[..., 0] == 0` is not satisfied.");
+   }
+ }
+ 
+@@ -72,10 +72,10 @@ INVARIANT_CHECK_FUNC_API
+ _check_last_cidx_is_nnz(const index_t& cidx, const index_t& nnz) {
+   const bool invariant = cidx == nnz;
+   if (cdim_name == CDimName::CRow) {
+-    _assert(invariant, "`crow_indices[..., -1] == nnz` is not satisfied.");
++    __assert__(invariant, "`crow_indices[..., -1] == nnz` is not satisfied.");
+   }
+   else {
+-    _assert(invariant, "`ccol_indices[..., -1] == nnz` is not satisfied.");
++    __assert__(invariant, "`ccol_indices[..., -1] == nnz` is not satisfied.");
+   }
+ }
+ 
+@@ -91,11 +91,11 @@ _check_cidx_nondecreasing_locally_bounded_sequence(
+   const auto s_cidx = cidx_next - cidx;
+   const bool invariant = zero <= s_cidx && s_cidx <= dim;
+   if (cdim_name == CDimName::CRow) {
+-    _assert(invariant,
++    __assert__(invariant,
+         "`0 <= crow_indices[..., 1:] - crow_indices[..., :-1] <= ncols` is not satisfied.");
+   }
+   else {
+-    _assert(invariant,
++    __assert__(invariant,
+         "`0 <= ccol_indices[..., 1:] - ccol_indices[..., :-1] <= nrows` is not satisfied.");
+   }
+ }
+@@ -110,10 +110,10 @@ _check_idx_bounds(
+     const index_t& dim) {
+   const bool invariant = zero <= idx && idx < dim;
+   if (cdim_name == CDimName::CRow) {
+-    _assert(invariant, "`0 <= col_indices < ncols` is not satisfied.");
++    __assert__(invariant, "`0 <= col_indices < ncols` is not satisfied.");
+   }
+   else {
+-    _assert(invariant, "`0 <= row_indices < nrows` is not satisfied.");
++    __assert__(invariant, "`0 <= row_indices < nrows` is not satisfied.");
+   }
+ }
+ 
+@@ -133,13 +133,13 @@ _check_idx_sorted_distinct_vals_slices_with_cidx(
+   for (auto* RESTRICT curr = slice_begin + 1; curr < slice_end; ++curr) {
+     const auto invariant = *(curr - 1) < *curr;
+     if (cdim_name == CDimName::CRow) {
+-      _assert(invariant, "`col_indices[..., crow_indices[..., i - 1]:crow_indices[..., i]] "
++      __assert__(invariant, "`col_indices[..., crow_indices[..., i - 1]:crow_indices[..., i]] "
+                          "for all i = 1, ..., nrows "
+                          "are sorted and distinct along the last dimension values` "
+                          "is not satisfied.");
+     }
+     else {
+-      _assert(invariant, "`row_indices[..., ccol_indices[..., i - 1]:ccol_indices[..., i]] "
++      __assert__(invariant, "`row_indices[..., ccol_indices[..., i - 1]:ccol_indices[..., i]] "
+                          "for all i = 1, ..., ncols "
+                          "are sorted and distinct along the last dimension values` "
+                          "is not satisfied.");

misc/pytorch/files/patch-aten_src_ATen_native_transformers_attention.cpp

@@ -0,0 +1,11 @@
+--- aten/src/ATen/native/transformers/attention.cpp.orig	2022-12-29 23:18:12 UTC
++++ aten/src/ATen/native/transformers/attention.cpp
+@@ -118,7 +118,7 @@ Tensor masked_softmax(
+     Tensor& attn_scores,
+     c10::optional<Tensor> attn_mask,
+     const Tensor& query,
+-    c10::optional<int64_t> mask_type = NULL) {
++    c10::optional<int64_t> mask_type = 0) {
+   if (query.is_nested() && !attn_mask) {
+     return at::_nested_tensor_softmax_with_shape(attn_scores, query);
+   }

misc/pytorch/files/patch-cmake_Dependencies.cmake

@@ -0,0 +1,11 @@
+--- cmake/Dependencies.cmake.orig	2022-12-16 00:23:46 UTC
++++ cmake/Dependencies.cmake
+@@ -339,7 +339,7 @@ if(USE_NNPACK OR USE_QNNPACK OR USE_PYTORCH_QNNPACK OR
+       set(DISABLE_NNPACK_AND_FAMILY ON)
+     endif()
+   else()
+-    if(NOT IOS AND NOT (CMAKE_SYSTEM_NAME MATCHES "^(Android|Linux|Darwin|Windows)$"))
++    if(NOT IOS AND NOT (CMAKE_SYSTEM_NAME MATCHES "^(Android|Linux|FreeBSD|Darwin|Windows)$"))
+       message(WARNING
+         "Target platform \"${CMAKE_SYSTEM_NAME}\" is not supported in {Q/X}NNPACK. "
+         "Supported platforms are Android, iOS, Linux, and macOS. "

misc/pytorch/files/patch-cmake_public_mkldnn.cmake

@@ -0,0 +1,11 @@
+--- cmake/public/mkldnn.cmake.orig	2022-12-16 00:23:46 UTC
++++ cmake/public/mkldnn.cmake
+@@ -4,7 +4,7 @@ if(CPU_AARCH64)
+   include(${CMAKE_CURRENT_LIST_DIR}/ComputeLibrary.cmake)
+ endif()
+ 
+-find_package(MKLDNN QUIET)
++find_package(MKLDNN REQUIRED)
+ 
+ if(NOT TARGET caffe2::mkldnn)
+   add_library(caffe2::mkldnn INTERFACE IMPORTED)

misc/pytorch/files/patch-third__party_cpuinfo_CMakeLists.txt

@@ -0,0 +1,11 @@
+--- third_party/cpuinfo/CMakeLists.txt.orig	2022-12-16 00:23:47 UTC
++++ third_party/cpuinfo/CMakeLists.txt
+@@ -77,7 +77,7 @@ IF(NOT CMAKE_SYSTEM_NAME)
+       "Target operating system is not specified. "
+       "cpuinfo will compile, but cpuinfo_initialize() will always fail.")
+   SET(CPUINFO_SUPPORTED_PLATFORM FALSE)
+-ELSEIF(NOT CMAKE_SYSTEM_NAME MATCHES "^(Windows|CYGWIN|MSYS|Darwin|Linux|Android)$")
++ELSEIF(NOT CMAKE_SYSTEM_NAME MATCHES "^(Windows|CYGWIN|MSYS|Darwin|Linux|FreeBSD|Android)$")
+   IF(${CMAKE_VERSION} VERSION_GREATER_EQUAL "3.14" AND NOT CMAKE_SYSTEM_NAME STREQUAL "iOS")
+     MESSAGE(WARNING
+       "Target operating system \"${CMAKE_SYSTEM_NAME}\" is not supported in cpuinfo. "

misc/pytorch/files/patch-third__party_fbgemm_third__party_asmjit_src_asmjit_core_virtmem.cpp

@@ -0,0 +1,10 @@
+--- third_party/fbgemm/third_party/asmjit/src/asmjit/core/virtmem.cpp.orig	2022-12-16 00:23:48 UTC
++++ third_party/fbgemm/third_party/asmjit/src/asmjit/core/virtmem.cpp
+@@ -45,6 +45,7 @@
+ #endif
+ 
+ #include <atomic>
++#include <sys/stat.h>
+ 
+ #if defined(__APPLE__) || defined(__BIONIC__)
+   #define ASMJIT_VM_SHM_DETECT 0

misc/pytorch/files/patch-third__party_kineto_libkineto_src_ThreadUtil.cpp

@@ -0,0 +1,11 @@
+--- third_party/kineto/libkineto/src/ThreadUtil.cpp.orig	2022-12-29 22:41:51 UTC
++++ third_party/kineto/libkineto/src/ThreadUtil.cpp
+@@ -49,7 +49,7 @@ int32_t systemThreadId() {
+ #elif defined _MSC_VER
+     _sysTid = (int32_t)GetCurrentThreadId();
+ #else
+-    _sysTid = (int32_t)syscall(SYS_gettid);
++    _sysTid = (int32_t)syscall(SYS_getpid);
+ #endif
+   }
+   return _sysTid;

misc/pytorch/pkg-descr

@@ -0,0 +1,3 @@
+PyTorch is a Python package that provides two high-level features:
+* Tensor computation (like NumPy) with strong GPU acceleration
+* Deep neural networks built on a tape-based autograd system