How to install tensorflow for the new Mac M1 hardware

Prerequisite: Xcode 12.2 and Command Line Tools for Xcode 12.2

(1) # Download the archive for this repo from https://github.com/apple/tensorflow_macos/releases
cd $HOME/Downloads/
curl -fsSLO https://github.com/apple/tensorflow_macos/releases/download/v0.1alpha0/tensorflow_macos-0.1alpha0.tar.gz
tar xzvf tensorflow_macos-0.1alpha0.tar.gz
/bin/bash ./tensorflow_macos/install_venv.sh --help

(2) # Download Miniconda from https://conda-forge.org/blog/posts/2020-10-29-macos-arm64/

(3) # Install Miniconda and after installtion, exit shell and login again
/bin/bash -c "$(curl -fsSL https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-MacOSX-arm64.sh)"

(4) # Once it's installed, create a Python 3.8 env by running
conda create --name python38 python=3.8

(5) # Put a path to where the arm64 libraries are. For example...
libs="$HOME/Downloads/tensorflow_macos/arm64/"

(6) # Replace this with the path of your Conda environment
env="$HOME/miniforge3/envs/python38"

(7) # upgrade
conda upgrade -c conda-forge pip setuptools cached-property six

(8) # activate env
conda activate python38
# conda deactivate

(9) pip install --upgrade -t "$env/lib/python3.8/site-packages/" --no-dependencies --force "$libs/grpcio-1.33.2-cp38-cp38-macosx_11_0_arm64.whl"

(10) pip install --upgrade -t "$env/lib/python3.8/site-packages/" --no-dependencies --force "$libs/h5py-2.10.0-cp38-cp38-macosx_11_0_arm64.whl"

(11) pip install --upgrade -t "$env/lib/python3.8/site-packages/" --no-dependencies --force "$libs/numpy-1.18.5-cp38-cp38-macosx_11_0_arm64.whl"

(12) pip install --upgrade -t "$env/lib/python3.8/site-packages/" --no-dependencies --force "$libs/tensorflow_addons-0.11.2+mlcompute-cp38-cp38-macosx_11_0_arm64.whl"

(13) # install these

conda install -c conda-forge -y absl-py
conda install -c conda-forge -y astunparse
conda install -c conda-forge -y gast
conda install -c conda-forge -y opt_einsum
conda install -c conda-forge -y termcolor
conda install -c conda-forge -y typing_extensions
conda install -c conda-forge -y wheel
conda install -c conda-forge -y typeguard

pip install tensorboard

pip install wrapt flatbuffers tensorflow_estimator google_pasta keras_preprocessing protobuf

(14) pip install --upgrade -t "$env/lib/python3.8/site-packages/" --no-dependencies --force "$libs/tensorflow_macos-0.1a0-cp38-cp38-macosx_11_0_arm64.whl"

(15) # Run this to test

time python tftest.py

tftest.py    Select all

from datetime import datetime import numpy as np import tensorflow as tf from tensorflow.python.compiler.mlcompute import mlcompute mlcompute.set_mlc_device(device_name="cpu") # tensorflow:Eager mode on GPU is extremely slow. So use CPU instead print("Hello, Tensorflow! ", end='') print(tf.__version__) print("start" , datetime.now()) X_raw = np.array([2013, 2014, 2015, 2016, 2017, 2018], dtype=np.float32) y_raw = np.array([12000, 14000, 15000, 16500, 17500, 19000], dtype=np.float32) X = (X_raw - X_raw.min()) / (X_raw.max() - X_raw.min()) y = (y_raw - y_raw.min()) / (y_raw.max() - y_raw.min()) X = tf.constant(X) y = tf.constant(y) a = tf.Variable(initial_value=0.) b = tf.Variable(initial_value=0.) variables = [a, b] num_epoch = 10000 optimizer = tf.keras.optimizers.SGD(learning_rate=1e-3) for e in range(num_epoch): with tf.GradientTape() as tape: y_pred = a * X + b loss = 0.5 * tf.reduce_sum(tf.square(y_pred - y)) grads = tape.gradient(loss, variables) optimizer.apply_gradients(grads_and_vars=zip(grads, variables)) print(a, b) print("end" , datetime.now())

(16) Test run this cnn.py https://github.com/apple/tensorflow_macos/issues/25

How-to do ARM GCC Inline Assembler for iPhone

The following code demonstrates an example to write Inline Assembler in llvm-gcc for iPhone ARM

arithmetic_shift_right.c Select all


#include <stdio.h>
#include <stdlib.h>

int main(int argc, char *argv[]) {
int i = atoi(argv[1]);
asm volatile ("mov %0, %1, ASR #1" : "=r"(i) : "r"(i));
printf("arithmetic_shift_right is %d\n", i);
exit(0);
}


The general format (ref) is
asm volatile (assembler instructions : output operands (optional) : input operands (optional) : clobbered registers (optional) );

e.g.
asm volatile ("mul %0, %1, %2" : "=r" (result) : "r" (number1) , "r" (number2));


This version is for Assembler Macro

arithmetic_shift_right.c Select all


#include <stdio.h>
#include <stdlib.h>

inline int arithmetic_shift_right(int a) {
int y;
__asm__("mov %0, %1, ASR #1" : "=r" (y) : "r" (a));
// Register R0 will become the value of register R1 shifted to the right by 1 bit, with the sign maintained.
return y;
}

int main(int argc, char *argv[]) {
int i = atoi(argv[1]);
printf("arithmetic_shift_right %d is %d\n", i, arithmetic_shift_right(i));
exit(0);
}


Below is the otool output for Assembler Macro

otool -tV Select all


_arithmetic_shift_right:
00001eb4 e92d4080 stmdb sp!, {r7, lr}
00001eb8 e28d7000 add r7, sp, #0 ; 0x0
00001ebc e24dd008 sub sp, sp, #8 ; 0x8
00001ec0 e58d0000 str r0, [sp]
00001ec4 e59d3000 ldr r3, [sp]
00001ec8 e1a030c3 mov r3, r3, asr #1       @@ arithmetic shift right
00001ecc e58d3004 str r3, [sp, #4]
00001ed0 e59d3004 ldr r3, [sp, #4]
00001ed4 e1a00003 mov r0, r3
00001ed8 e247d000 sub sp, r7, #0 ; 0x0
00001edc e8bd8080 ldmia sp!, {r7, pc}


Below is the otool output for Assembler Macro (after full optimzation -O2)

otool -tV Select all


_arithmetic_shift_right:
00001f04 e1a000c0 mov r0, r0, asr #1
00001f08 e12fff1e bx lr


Assembler Macro for more than one assembler instruction
arithmetic shift right then perform 16 bit binary multiplication

Assembler Macros Select all


__asm__("mov %0, %1, ASR #1\n\t"
"mul %0, %0, %1"
: "=r" (y) : "r" (a));