[WIP] OpenBLAS and Numpy

[meow464]

During the last ~2 months I got to what seems to be a working port of flang for iOS, but I still have to figure out how to make it work with numpy and kivy.

Here are the steps to get where I'm at, first build flang and openblas:

1. Build flang for iOS whit this instructions. Build the version with embedded bitcode, feel free to ask questions here if it's not clear how.
2. Build OpenBLAS with lapack for iOS: Clone openblas, save this script and run from within the repo. Everything will be installed in this same directory under the "INSTALL" folder.

Make a kivy project:

1. mkdir kivy-OpenBLAS && cd kivy-OpenBLAS
2. python -m venv venv && . venv/bin/activate
3. git clone https://github.com/kivy/kivy-ios.git && pip install -e kivy-ios

You can make modification in the kivy-ios/ folder and it will affect the installed kivy-ios.

Create a new project as per usual:

1. Create a test-app folder and put a hello world main.py in it;
2. toolchain build python3 kivy

Copy the openblas files to `dist/hostpython3/:

1. Copy INSTALL/lib/libopenblas_armv8p-r0.3.19.dev.a from the openblas folder to dist/hostpython3/lib/libopenblas.a. Note that you need to rename it.
2. Copy everything in INSTALL/include to dist/hostpython3/include

So far openblas is only built for iOS therefore you will need to use toolchain build numpy --arch arm64 to test numpy. To make changes to the numpy recipe simply edit kivy-ios/kivy_ios/recipes/numpy/__init__.py.

My currently not working numpy recipe

``` from kivy_ios.toolchain import CythonRecipe from os.path import join import sh import shutil class NumpyRecipe(CythonRecipe): version = "1.20.2" url = "https://pypi.python.org/packages/source/n/numpy/numpy-{version}.zip" library = "libnumpy.a" libraries = ["libnpymath.a", "libnpyrandom.a"] include_dir = "numpy/core/include" depends = ["python"] hostpython_prerequisites = ["Cython"] cythonize = False def prebuild_arch(self, arch): if self.has_marker("patched"): return self.apply_patch("duplicated_symbols.patch") self.set_marker("patched") def get_recipe_env(self, arch): env = super().get_recipe_env(arch) # CC must have the CFLAGS with arm arch, because numpy tries first to # compile and execute an empty C to see if the compiler works. This is # obviously not working when crosscompiling env["CC"] = "{} {}".format(env["CC"], env["CFLAGS"]) env["FC"] = "/ur/local/bin/flang -L/usr/local/flang-iOS" # Disable Accelerate.framework by disabling the optimized BLAS and LAPACK libraries cause it's now unsupported env["NPY_BLAS_ORDER"] = "openblas" env["NPY_LAPACK_ORDER"] = "openblas" return env def build_arch(self, arch): super().build_arch(arch) sh.cp(sh.glob(join(self.build_dir, "build", "temp.*", "libnpy*.a")), self.build_dir) def reduce_python_package(self): dest_dir = join(self.ctx.site_packages_dir, "numpy") shutil.rmtree(join(dest_dir, "core", "include")) shutil.rmtree(join(dest_dir, "core", "tests")) shutil.rmtree(join(dest_dir, "distutils")) shutil.rmtree(join(dest_dir, "doc")) shutil.rmtree(join(dest_dir, "f2py", "tests")) shutil.rmtree(join(dest_dir, "fft", "tests")) shutil.rmtree(join(dest_dir, "lib", "tests")) shutil.rmtree(join(dest_dir, "linalg", "tests")) shutil.rmtree(join(dest_dir, "ma", "tests")) shutil.rmtree(join(dest_dir, "matrixlib", "tests")) shutil.rmtree(join(dest_dir, "polynomial", "tests")) shutil.rmtree(join(dest_dir, "random", "tests")) shutil.rmtree(join(dest_dir, "tests")) recipe = NumpyRecipe() ```

All input is welcome!

[meow464]

This recipe seem to build successfully but there are still some things to be worked out.

``` from kivy_ios.toolchain import CythonRecipe from os.path import join import sh import shutil class NumpyRecipe(CythonRecipe): version = "1.20.2" url = "https://pypi.python.org/packages/source/n/numpy/numpy-{version}.zip" library = "libnumpy.a" libraries = ["libnpymath.a", "libnpyrandom.a"] include_dir = "numpy/core/include" depends = ["python"] hostpython_prerequisites = ["Cython"] cythonize = False def prebuild_arch(self, arch): if self.has_marker("patched"): return self.apply_patch("duplicated_symbols.patch") self.set_marker("patched") def get_recipe_env(self, arch): env = super().get_recipe_env(arch) # CC must have the CFLAGS with arm arch, because numpy tries first to # compile and execute an empty C to see if the compiler works. This is # obviously not working when crosscompiling env["CC"] = "{} {} -L/usr/local/flang-iOS/lib -lflang -lflangrti -lpgmath -lompstub -lm -lpthread -lSystem".format(env["CC"], env["CFLAGS"]) # Disable Accelerate.framework by disabling the optimized BLAS and LAPACK libraries cause it's now unsupported env["NPY_BLAS_ORDER"] = "openblas" env["NPY_LAPACK_ORDER"] = "openblas" return env def build_arch(self, arch): super().build_arch(arch) sh.cp(sh.glob(join(self.build_dir, "build", "temp.*", "libnpy*.a")), self.build_dir) def reduce_python_package(self): dest_dir = join(self.ctx.site_packages_dir, "numpy") shutil.rmtree(join(dest_dir, "core", "include")) shutil.rmtree(join(dest_dir, "core", "tests")) shutil.rmtree(join(dest_dir, "distutils")) shutil.rmtree(join(dest_dir, "doc")) shutil.rmtree(join(dest_dir, "f2py", "tests")) shutil.rmtree(join(dest_dir, "fft", "tests")) shutil.rmtree(join(dest_dir, "lib", "tests")) shutil.rmtree(join(dest_dir, "linalg", "tests")) shutil.rmtree(join(dest_dir, "ma", "tests")) shutil.rmtree(join(dest_dir, "matrixlib", "tests")) shutil.rmtree(join(dest_dir, "polynomial", "tests")) shutil.rmtree(join(dest_dir, "random", "tests")) shutil.rmtree(join(dest_dir, "tests")) recipe = NumpyRecipe() ```

The flags -lflang -lflangrti -lpgmath -lompstub -lm -lpthread -lSystem are the default linker flags for flang. Problems:

1. The vast majority of linkings don't need those flags so we see a lot of warnings.
2. I'm passing the directory -L/usr/local/flang-iOS/lib but the flang libs should be probably somewhere under dist/. I did try copying them to dist/hostpython3/lib/ and dist/lib but it didn't work.

[meow464]

It seems the numpy setup.py or cython were trying to build a test binary but the linker was not receiving /usr/local/flang-iOS or dist/hostpython3/lib. Adding an -L flag solved this.

Now I'm getting a duplicated symbol _xerbla_ in the files: libnumpy.a (_multiarray_umath.cpython-39-darwin.so.o) and libnumpy.a (lapack_lite.cpython-39-darwin.so.o). But lapack_lite shouldn't be build if openblas is available? Or should it?

[meow464]

I think I made progress:

duplicated_symbols.patch (numpy)

``` diff -Naur numpy-1.20.2.orig/numpy/linalg/setup.py numpy-1.20.2/numpy/linalg/setup.py --- numpy-1.20.2.orig/numpy/linalg/setup.py 2021-04-04 11:04:17.000000000 +0200 +++ numpy-1.20.2/numpy/linalg/setup.py 2021-04-04 11:04:54.000000000 +0200 @@ -75,7 +75,7 @@ # umath_linalg module config.add_extension( '_umath_linalg', - sources=['umath_linalg.c.src', get_lapack_lite_sources], + sources=['umath_linalg.c.src'], depends=['lapack_lite/f2c.h'], extra_info=lapack_info, libraries=['npymath'], diff -Naur numpy-1.20.2.orig/numpy/random/setup.py numpy-1.20.2/numpy/random/setup.py --- numpy-1.20.2.orig/numpy/random/setup.py 2021-04-04 11:04:17.000000000 +0200 +++ numpy-1.20.2/numpy/random/setup.py 2021-04-04 11:05:22.000000000 +0200 @@ -127,7 +127,6 @@ config.add_extension('mtrand', sources=['mtrand.c', 'src/legacy/legacy-distributions.c', - 'src/distributions/distributions.c', ], include_dirs=['.', 'src', 'src/legacy'], libraries=['m'] if os.name != 'nt' else [],diff --git a/numpy/linalg/setup.py b/numpy/linalg/setup.py diff --git a/numpy/linalg/setup.py b/numpy/linalg/setup.py index 94536bb2c..c5b0d5fba 100644 --- a/numpy/linalg/setup.py +++ b/numpy/linalg/setup.py @@ -61,7 +61,7 @@ def get_lapack_lite_sources(ext, build_dir): config.add_extension( 'lapack_lite', - sources=['lapack_litemodule.c', get_lapack_lite_sources], + sources=['lapack_litemodule.c'], depends=['lapack_lite/f2c.h'], extra_info=lapack_info, ) diff --git a/numpy/core/setup.py b/numpy/core/setup.py index f6b31075d..f12c8b9e6 100644 --- a/numpy/core/setup.py +++ b/numpy/core/setup.py @@ -818,7 +818,7 @@ def gl_if_msvc(build_cmd): # These files are also in MANIFEST.in so that they are always in # the source distribution independently of HAVE_CBLAS. common_src.extend([join('src', 'common', 'cblasfuncs.c'), - join('src', 'common', 'python_xerbla.c'), + # join('src', 'common', 'python_xerbla.c'), ]) else: extra_info = {} ```

__init__.py (numpy)

``` from kivy_ios.toolchain import CythonRecipe from os.path import join import sh import shutil class NumpyRecipe(CythonRecipe): version = "1.20.2" url = "https://pypi.python.org/packages/source/n/numpy/numpy-{version}.zip" library = "libnumpy.a" libraries = ["libnpymath.a", "libnpyrandom.a"] include_dir = "numpy/core/include" depends = ["python"] hostpython_prerequisites = ["Cython"] cythonize = False def prebuild_arch(self, arch): if self.has_marker("patched"): return self.apply_patch("duplicated_symbols.patch") self.set_marker("patched") def get_recipe_env(self, arch): env = super().get_recipe_env(arch) # CC must have the CFLAGS with arm arch, because numpy tries first to # compile and execute an empty C to see if the compiler works. This is # obviously not working when crosscompiling env["CC"] = "{} {} -lflang -lpgmath -lflangrti -lompstub -lm -lpthread -lSystem -v -L/Users/tuco/Projects/kivy-OpenBLAS/dist/hostpython3/lib".format(env["CC"], env["CFLAGS"]) # -L/usr/local/flang-iOS/lib # Disable Accelerate.framework by disabling the optimized BLAS and LAPACK libraries cause it's now unsupported env["NPY_BLAS_ORDER"] = "openblas" env["NPY_LAPACK_ORDER"] = "openblas" return env def build_arch(self, arch): super().build_arch(arch) sh.cp(sh.glob(join(self.build_dir, "build", "temp.*", "libnpy*.a")), self.build_dir) def reduce_python_package(self): dest_dir = join(self.ctx.site_packages_dir, "numpy") shutil.rmtree(join(dest_dir, "core", "include")) shutil.rmtree(join(dest_dir, "core", "tests")) shutil.rmtree(join(dest_dir, "distutils")) shutil.rmtree(join(dest_dir, "doc")) shutil.rmtree(join(dest_dir, "f2py", "tests")) shutil.rmtree(join(dest_dir, "fft", "tests")) shutil.rmtree(join(dest_dir, "lib", "tests")) shutil.rmtree(join(dest_dir, "linalg", "tests")) shutil.rmtree(join(dest_dir, "ma", "tests")) shutil.rmtree(join(dest_dir, "matrixlib", "tests")) shutil.rmtree(join(dest_dir, "polynomial", "tests")) shutil.rmtree(join(dest_dir, "random", "tests")) shutil.rmtree(join(dest_dir, "tests")) recipe = NumpyRecipe() ```

But I don't know if just commenting out the source of _xerbla_ like that is the right approach? Ideally numpy should detect openblas and disable lapack_lite, right? (I think _xerbla_ comes from lapack_lite).

In any case numpy still can't find openblas when deployed to a phone.

I'm using this code to test.

``` import kivy from kivy.app import App from kivy.uix.label import Label import numpy as np # Replace this with your # current version kivy.require('1.11.1') # Defining a class class MyFirstKivyApp(App): # Function that returns # the root widget def build(self): # Label with text Hello World is # returned as root widget version = np.show_config() return Label(text = str(version)) # Here our class is initialized # and its run() method is called. # This initializes and starts # our Kivy application. MyFirstKivyApp().run() ```

[meow464]

It seems numpy can find openblas but it shows the host's path /Users/tuco/Projects/kivy-OpenBLAS/dist/hostpython3/lib. I'm able to run some examples but I still don't know for sure if it can use openblas.

[meow464]

I don't yet have a recipe, I just drop the flang and openblas .a files under dist/hostpython3/lib and dist/lib. If I just do that and run the project the benchmark finishes in ~1.5 seconds but if I drop the .a files in xcode the run time goes up to ~5.5 seconds. Very weird.

benchmark

``` import kivy from kivy.app import App from kivy.uix.label import Label from random import random import numpy as np from time import time # Replace this with your # current version kivy.require('1.11.1') # Defining a class class MyFirstKivyApp(App): # Function that returns # the root widget def build(self): np.show_config() # Let's take the randomness out of random numbers (for reproducibility) np.random.seed(0) size = 4096 A, B = np.random.random((size, size)), np.random.random((size, size)) # C, D = np.random.random((size * 128,)), np.random.random((size * 128,)) # E = np.random.random((int(size / 2), int(size / 4))) # F = np.random.random((int(size / 2), int(size / 2))) # F = np.dot(F, F.T) # G = np.random.random((int(size / 2), int(size / 2))) # Matrix multiplication N = 20 t = time() for i in range(N): np.dot(A, B) delta = time() - t result = 'Dotted two %dx%d matrices in %0.2f s.' % (size, size, delta / N) return Label(text = result) # Here our class is initialized # and its run() method is called. # This initializes and starts # our Kivy application. MyFirstKivyApp().run() ```

[meow464]

I tried with pure numpy, size=1536 and it took ~8 seconds. (4096 took over 20 minutes and didn't finish). The 5 seconds result from above was openblas and the 1.5 second must have been Accelerate.

Accelerate is included in the default kivy-ios project and even if numpy doesn't support it some stuff works and the symbols are there.

[meow464]

Calling linalg.svd and linalg.eig causes a EXC_BAD_ACCESS. I'm using the examples from here. I will post more information later.

I will try the examples on macos with openblas built with flang and hope I can replicate the crash.