Conda build recipe (#288)

* Typofix.

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

* Probe for nvrtc in lib directory as well.

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

* Conda build recipe for TVM.

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

CMakeLists.txt

@@ -90,7 +90,8 @@ find_package(CUDA)
     list(APPEND TVM_LINKER_LIBS ${CUDA_NVRTC_LIB})
   else(MSVC)
     find_library(CUDA_NVRTC_LIB nvrtc
-      ${CUDA_TOOLKIT_ROOT_DIR}/lib64)
+      ${CUDA_TOOLKIT_ROOT_DIR}/lib64
+      ${CUDA_TOOLKIT_ROOT_DIR}/lib)
     list(APPEND TVM_LINKER_LIBS ${CUDA_NVRTC_LIB})
   endif(MSVC)
   add_definitions(-DTVM_CUDA_RUNTIME=1)

docs/faq.md

@@ -17,6 +17,6 @@ See also [This Issue](https://github.com/dmlc/tvm/issues/151)
 
 TVM's relation to libDNN cuDNN
 ------------------------------
-TVM can incoporate these library as external calls. One goal of TVM is to be able to
+TVM can incorporate these library as external calls. One goal of TVM is to be able to
 generate high performing kernels. We will evolve TVM an incremental manner as
 we learn from the technics of manual kernel crafting and add these as primitives in DSL.

include/tvm/ir.h

@@ -348,7 +348,7 @@ constexpr const char* tvm_thread_context = "tvm_thread_context";
  */
 constexpr const char* tvm_call_packed_lowered = "tvm_call_packed_lowered";
 /*!
- * \brief See pesudo code
+ * \brief See pseudo code
  *
  *  int tvm_storage_sync(std::string storage_scope) {
  *     __sync(storage_scope);

python/conda/build.sh

+#!/bin/bash
+
+set -e
+
+# See Note [CUDA_TOOLKIT_ROOT_DIR versus CUDA_BIN_PATH]
+if [ -z "$CONDA_CUDA_HOME" ]; then
+  CUDA_ARGS=""
+else
+  # See Note [Bash argument quoting]
+  CUDA_ARGS="-DCUDA_TOOLKIT_ROOT_DIR=$(printf %q "$CONDA_CUDA_HOME")"
+fi
+
+if [ -z "$PREFIX" ]; then
+  PREFIX="$CONDA_PREFIX"
+fi
+
+if [ "$(uname)" = 'Darwin' ]
+then
+    # Without this, Apple's default shipped clang will refuse to see any
+    # headers like mutex.
+    export MACOSX_DEPLOYMENT_TARGET=10.9
+fi
+
+rm -rf build || true
+mkdir -p build
+cd build
+# Enable static-libstdc++ to make it easier to link this library with
+# other C++ compilers
+CXXFLAGS=-static-libstdc++ cmake3 -DCMAKE_PREFIX_PATH=${PREFIX} -DCMAKE_INSTALL_PREFIX=${PREFIX} -DUSE_CUDA=1 -DUSE_LLVM=1 -DINSTALL_DEV=1 $CUDA_ARGS ..
+make -j20 VERBOSE=1
+make install/fast
+cd ..
+
+# Also install the headers for libraries that TVM vendored
+mkdir -p "$PREFIX/include"
+# TODO: arguably dlpack and dmlc-core should get its own packaging and
+# install their headers themselves
+cp -R dlpack/include/. "$PREFIX/include"
+cp -R dmlc-core/include/. "$PREFIX/include"
+# TODO: HalideIR's includes could conflict, but TVM currently assumes they
+# are installed here, awfully enough
+cp -R HalideIR/src/. "$PREFIX/include"
+
+cd python
+$PYTHON setup.py install
+cd ..

python/conda/meta.yaml

+{% set version = "0.1.dev" %}
+
+package:
+  name: tvm
+  version: {{ version }}
+
+source:
+  path: ../..
+
+build:
+  number: 1
+  skip: True  # [win]
+  script_env:
+    - CONDA_CUDA_HOME
+
+requirements:
+  build:
+    - llvmdev ==4.0.0
+    - python >=3
+    - numpy
+    - setuptools
+    - nose
+    - decorator
+  run:
+    - python >=3
+    - numpy
+    - decorator
+
+about:
+  home: https://github.com/dmlc/tvm
+  license: Apache2
+  summary: a low level domain specific language for compiling tensor computation pipelines

tutorials/python/opt_gemm.py

@@ -95,11 +95,11 @@ print('Opt1: %f' % evaluator(a, b, c).mean)
 # Vectorization
 # -------------
 # Another important trick is vectorization. When the memory access pattern is uniform, the compiler
-# can dectect this pattern and pass the continuous memory to vector processor. In TVM, we can use
+# can detect this pattern and pass the continuous memory to vector processor. In TVM, we can use
 # `vectorize` interface to hint the compiler this pattern, so that we can accelerate it vastly.
 #
 
-# After trying different schedule, we finally found that we can benefit from vectorizing 
+# After trying different schedule, we finally found that we can benefit from vectorizing
 # the row loop most, i.e. yi.
 s[C].vectorize(yi)
 func = tvm.build(s, [A, B, C], name = 'mmult')