After moving warp into subdir

Author: LegNeato

crates/cuda_std/src/lib.rs

@@ -24,8 +24,9 @@
 #![allow(internal_features)]
 #![cfg_attr(
     target_os = "cuda",
-    feature(alloc_error_handler, asm_experimental_arch, link_llvm_intrinsics)
+    feature(alloc_error_handler, asm_experimental_arch)
 )]
+#![feature(link_llvm_intrinsics)]
 
 extern crate alloc;
 

crates/cuda_std/src/warp/matrix.rs renamed to crates/cuda_std/src/warp/matrix/mod.rs

@@ -15,7 +15,6 @@ pub mod ops;
 
 // WMMA intrinsic declarations
 // These LLVM intrinsics return arrays by value which is their intended behavior
-#[cfg(target_arch = "nvptx64")]
 #[allow(improper_ctypes)]
 extern "C" {
     // ============= 16x16x16 intrinsics =============

crates/cuda_std/src/warp/matrix/ops.rs

@@ -6,7 +6,6 @@ use super::layout::{Col, Layout, Row};
 use super::*;
 
 // Import all the intrinsic functions from the parent module
-#[cfg(target_arch = "nvptx64")]
 use super::{
     wmma_load_a_bf16_col_m16n16k16,
     wmma_load_a_bf16_col_m16n8k16,