Updated main-cpu branch to the latest

.gitignore

@@ -0,0 +1 @@
+*.egg-info/

difflogic/compiled_model.py

@@ -284,7 +284,7 @@ def compile(self, opt_level=1, save_lib_path=None, verbose=False):
                             "-shared",
                             "-fPIC",
                             "-O{}".format(opt_level),
-                            "-march=native",
+                            # "-march=native",  # removed for compatibility with Apple Silicon: https://stackoverflow.com/questions/65966969/why-does-march-native-not-work-on-apple-m1
                             "-o",
                             lib_file.name,
                             c_file.name,

difflogic/cuda/difflogic_kernel.cu

@@ -557,7 +557,7 @@ __global__ void tensor_packbits_cuda_kernel(
     ) {
         for (  // batch in b
             auto col = blockIdx.x * blockDim.x + threadIdx.x;
-            col < t.size(1);
+            col < b.size(1);
             col += blockDim.x * gridDim.x
         ) {
 
@@ -569,8 +569,11 @@ __global__ void tensor_packbits_cuda_kernel(
             constexpr int bit_count = std::numeric_limits<unsigned_scalar_t>::digits;
             val.signed_scalar = b[row][col];
             for (unsigned int i = 0; i < bit_count; ++i) {
-                const unsigned_scalar_t bit_mask = static_cast<unsigned_scalar_t>(t[row][bit_count * col + i]) << i;
-                val.unsigned_scalar = val.unsigned_scalar | bit_mask;
+                const auto t_col = bit_count * col + i;
+                if (t_col < t.size(1)) {    
+                    const unsigned_scalar_t bit_mask = static_cast<unsigned_scalar_t>(t[row][t_col]) << i;
+                    val.unsigned_scalar = val.unsigned_scalar | bit_mask;
+                }
             }
             b[row][col] = val.signed_scalar;
         }