fix(cndrv): 增加一个计算类 kernel 的测试，改正 cnrtc 参数错误

Signed-off-by: YdrMaster <ydrml@hotmail.com>

Author: YdrMaster

cndrv/src/cnrtc/binary.rs

@@ -27,6 +27,7 @@ impl CnrtcBinary {
 
         let options = [
             CString::new("-O3").unwrap(),
+            CString::new("--bang-fatbin-only").unwrap(),
             CString::new(format!("--bang-mlu-arch=mtp_{isa}")).unwrap(),
         ];
         let options = options.iter().map(|s| s.as_ptr()).collect::<Vec<_>>();

cndrv/src/cnrtc/mod.rs

@@ -11,8 +11,7 @@ fn test_behavior() {
         ptr::{null, null_mut},
     };
 
-    const CODE: &str =
-        r#"extern "C" __mlu_global__ void kernel() { printf("Hello from MLU!\n"); }"#;
+    const CODE: &str = r#"extern "C" __mlu_entry__ void kernel() { printf("Hello from MLU!\n"); }"#;
 
     crate::init();
     let Some(dev) = crate::Device::fetch() else {
@@ -52,7 +51,6 @@ fn test_behavior() {
         bin
     };
 
-    crate::init();
     if let Some(dev) = crate::Device::fetch() {
         dev.context().apply(|ctx| {
             use crate::AsRaw;
@@ -82,3 +80,70 @@ fn test_behavior() {
         });
     };
 }
+
+#[test]
+fn test_add() {
+    use crate::memcpy_d2h;
+    use std::ffi::{c_void, CString};
+
+    const N: usize = 64;
+    let src = format!(
+        r#"
+extern "C" __mlu_entry__ void kernel(
+    float      * ans_,
+    float const* lhs_,
+    float const* rhs_
+) {{
+    __nram__ float lhs[{N}];
+    __nram__ float rhs[{N}];
+    __nram__ float ans[{N}];
+    __memcpy(lhs, lhs_, {N} * sizeof(float), GDRAM2NRAM);
+    __memcpy(rhs, rhs_, {N} * sizeof(float), GDRAM2NRAM);
+    __bang_add(ans, lhs, rhs, {N});
+    __memcpy(ans_, ans, {N} * sizeof(float), NRAM2GDRAM);
+}}"#
+    );
+
+    crate::init();
+    let Some(dev) = crate::Device::fetch() else {
+        return;
+    };
+
+    let (result, log) = CnrtcBinary::compile(src, dev.isa());
+    if !log.is_empty() {
+        eprintln!("{log}");
+    }
+    let bin = result.unwrap();
+
+    let a = vec![1.0f32; N];
+    let b = vec![2.0f32; N];
+    let mut c = vec![0.0f32; N];
+
+    if let Some(dev) = crate::Device::fetch() {
+        dev.context().apply(|ctx| {
+            let mut lhs = ctx.malloc::<f32>(N);
+            let mut rhs = ctx.malloc::<f32>(N);
+            let mut ans = ctx.malloc::<f32>(N);
+
+            let queue = ctx.queue();
+            queue.memcpy_h2d(&mut lhs, &a);
+            queue.memcpy_h2d(&mut rhs, &b);
+
+            let lhs_ptr = lhs.as_ptr();
+            let rhs_ptr = rhs.as_ptr();
+            let ans_ptr = ans.as_mut_ptr();
+            let params: [*const c_void; 3] = [
+                &ans_ptr as *const _ as _,
+                &lhs_ptr as *const _ as _,
+                &rhs_ptr as *const _ as _,
+            ];
+
+            ctx.load(&bin)
+                .get_kernel(&CString::new("kernel").unwrap())
+                .launch(1, 1, 1, params.as_ptr() as _, &queue);
+
+            memcpy_d2h(&mut c, &ans);
+        });
+        assert_eq!(c, &[3.0f32; N]);
+    };
+}