relu in python

ane/2_compile/simple/neuron.plist

@@ -31,11 +31,11 @@
 			<integer>1</integer>
 
 			<key>InputChannels</key>
-			<integer>3</integer>
+			<integer>1</integer>
 			<key>InputHeight</key>
 			<integer>1</integer>
 			<key>InputWidth</key>
-			<integer>1</integer>
+			<integer>16</integer>
 
 			<key>InputPlaneStride</key>
 			<integer>64</integer>
@@ -56,7 +56,7 @@
 			<key>Params</key>
 			<dict>
 				<key>Type</key>
-				<string>Sign</string>
+				<string>ReLU</string>
 			</dict>
 			<key>Type</key>
 			<string>Neuron</string>

ane/lib/.gitignore

@@ -1 +1 @@
-ane.dylib
+libane.dylib

ane/lib/ane.mm

@@ -43,6 +43,8 @@ int ANE_Open() {
 
   ret = dev->ANE_IsPowered();
   printf("powered? %d\n", ret);
+
+  return 0;
 }
 
 int stride_for_width(int width) {
@@ -51,7 +53,7 @@ int stride_for_width(int width) {
   return ret;
 }
 
-void *ANE_CreateTensor(int width, int height) {
+void *ANE_TensorCreate(int width, int height) {
   // all float16
   // input buffer
 
@@ -73,9 +75,10 @@ void* ANE_TensorData(void *out_surf) {
 
 uint64_t ANE_Compile(char *prog, int sz) {
   int ret;
+  printf("ANE_Compile %p with size %d\n", prog, sz);
   H11ANEProgramCreateArgsStruct mprog = {0};
   mprog.program = prog;
-  mprog.program_length = 0x8000;
+  mprog.program_length = sz;
 
   H11ANEProgramCreateArgsStructOutput *out = new H11ANEProgramCreateArgsStructOutput;
   memset(out, 0, sizeof(H11ANEProgramCreateArgsStructOutput));
@@ -93,7 +96,7 @@ uint64_t ANE_Compile(char *prog, int sz) {
   return program_handle;
 }
 
-void ANE_Run(uint64_t program_handle, void *in_surf, void *out_surf) {
+int ANE_Run(uint64_t program_handle, void *in_surf, void *out_surf) {
   int ret;
   H11ANEProgramRequestArgsStruct *pras = new H11ANEProgramRequestArgsStruct;
   memset(pras, 0, sizeof(H11ANEProgramRequestArgsStruct));
@@ -135,6 +138,8 @@ void ANE_Run(uint64_t program_handle, void *in_surf, void *out_surf) {
           MACH_PORT_NULL);
   printf("got message: %d sz %d\n", ret, message.header.msgh_size);
   delete pras;
+
+  return 0;
 }
 
 }

ane/lib/ane.py

@@ -0,0 +1,60 @@
+#!/usr/bin/env python3
+from ctypes import *
+import numpy as np
+import faulthandler
+faulthandler.enable()
+
+libane = cdll.LoadLibrary("libane.dylib")
+
+libane.ANE_Compile.argtypes = [c_char_p, c_int]
+libane.ANE_Compile.restype = c_void_p
+
+libane.ANE_TensorCreate.restype = c_void_p
+
+libane.ANE_TensorData.argtypes = [c_void_p]
+libane.ANE_TensorData.restype = POINTER(c_uint16)
+
+libane.ANE_Run.argtypes = [c_void_p]*3
+libane.ANE_Run.restype = c_int
+
+class ANETensor:
+  def __init__(self, *shape):
+    self.shape = shape
+    self.dtype = np.float16
+    self.sz = int(np.prod(shape))
+    self.tt = libane.ANE_TensorCreate(self.sz, 1)
+
+  def data(self):
+    data = libane.ANE_TensorData(self.tt)
+    buf = np.ctypeslib.as_array(data, shape=(self.sz,))
+    return np.frombuffer(buf, dtype=self.dtype)
+
+class ANE:
+  def __init__(self):
+    libane.ANE_Open()
+
+  def compile(self, dat):
+    ret = libane.ANE_Compile(create_string_buffer(dat), len(dat))
+    assert(ret is not None)
+    return ret
+
+  def run(self, prog, tin, tout):
+    libane.ANE_Run(prog, tin.tt, tout.tt)
+
+if __name__ == "__main__":
+  ane = ANE()
+
+  tin = ANETensor(16)
+  tout = ANETensor(16)
+
+  tind = tin.data()
+  toutd = tout.data()
+
+  tind[0:4] = [-1,1,-2,2]
+  print(tind)
+
+  comp = ane.compile(open("../2_compile/model.hwx", "rb").read())
+  ret = ane.run(comp, tin, tout)
+
+  print(toutd)
+

ane/lib/build.sh

@@ -1,4 +1,3 @@
 #!/bin/bash
-clang++ ane.mm --shared -F /System/Library/PrivateFrameworks/ -framework ANEServices -framework IOSurface -framework Foundation -framework IOKit -o ane.dylib
-codesign --entitlements entitlements.xml -s "Taylor Swift Child" ane.dylib
+clang++ ane.mm --shared -F /System/Library/PrivateFrameworks/ -framework ANEServices -framework IOSurface -framework Foundation -framework IOKit -o libane.dylib
 

ane/lib/lib.py

@@ -0,0 +1,29 @@
+#!/usr/bin/env python3
+from ctypes import *
+libane = cdll.LoadLibrary("libane.dylib")
+
+libane.ANE_Compile.argtypes = [c_char_p, c_int]
+libane.ANE_Compile.restype = c_void_p
+
+libane.ANE_TensorCreate.restype = c_void_p
+
+libane.ANE_TensorData.argtypes = [c_void_p]
+libane.ANE_TensorData.restype = c_void_p
+
+#libane.ANE_TensorRun.argtypes = [c_void_p]*3
+
+libane.ANE_Open()
+
+dat = open("../2_compile/model.hwx", "rb").read()
+comp = libane.ANE_Compile(create_string_buffer(dat), len(dat))
+print("compile", comp)
+
+tin = libane.ANE_TensorCreate(16,16)
+tout = libane.ANE_TensorCreate(16,16)
+
+#addr = libane.ANE_TensorData(dat)
+#print(dat, addr)
+
+
+
+

ane/lib/sign_python.sh

@@ -0,0 +1,3 @@
+#!/bin/bash
+codesign --entitlements entitlements.xml -s "Taylor Swift Child" /opt/homebrew/Cellar/python@3.9/3.9.0_2/Frameworks/Python.framework/Versions/3.9/Resources/Python.app/Contents/MacOS/Python
+