[FRONTEND] add python e2e launch empty kernel test (#68)

python/tests/test_compiler.py

@@ -2,18 +2,21 @@ import torch
 
 import triton
 import triton.language as tl
+import triton.runtime as runtime
 
 # trigger the torch.device implicitly to ensure cuda context initialization
 torch.zeros([10], device=torch.device('cuda'))
 
 
+@triton.jit
+def empty_kernel(X, stride_xm, BLOCK: tl.constexpr):
+    pass
+
+
 def test_empty_kernel_cubin_compile():
-    @triton.jit
-    def kernel(X, stride_xm, stride_xn, BLOCK: tl.constexpr):
-        pass
 
     device = torch.cuda.current_device()
-    cubin = triton.compile(kernel,
+    cubin = triton.compile(empty_kernel,
                            "*fp32,i32,i32",
                            device=device,
                            constants={"BLOCK": 256},
@@ -21,3 +24,25 @@ def test_empty_kernel_cubin_compile():
 
     print('cubin size:', len(cubin))
     assert len(cubin) > 0
+
+
+def test_empty_kernel_launch():
+    device = torch.cuda.current_device()
+    binary = runtime.build_kernel(empty_kernel, "*fp32,i32,i32",
+                                  device=device,
+                                  constants={"BLOCK": 256},
+                                  num_warps=4,
+                                  num_stages=3)
+    grid = lambda META: (
+        triton.cdiv(1024, META['BLOCK']) * triton.cdiv(1024, META['BLOCK']),
+    )
+
+    A = torch.zeros([1024], device="cuda")
+    runtime.launch_kernel(fn=empty_kernel,
+                          binary=binary,
+                          grid=grid,
+                          num_warps=4,
+                          num_stages=3,
+                          X=A,
+                          stride_xm=256,
+                          BLOCK=tl.constexpr(256))

python/tests/test_vecadd_no_scf.py

@@ -0,0 +1,27 @@
+import triton
+import triton.language as tl
+
+NUM_WARPS = 4
+
+# triton kernel
+
+
+def test_vecadd_no_scf():
+    @triton.jit
+    def kernel(x_ptr, stride_xn,
+               y_ptr, stride_yn,
+               z_ptr, stride_zn,
+               BLOCK_SIZE_N: tl.constexpr):
+        pid = tl.program_id(axis=0)
+        offset = pid * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+        x_ptrs = x_ptr + offset
+        y_ptrs = y_ptr + offset
+        x = tl.load(x_ptrs)
+        y = tl.load(y_ptrs)
+        z = x + y
+        z_ptrs = z_ptr + offset
+        tl.store(z_ptrs, z)
+
+    ret = triton.compile(kernel, "*fp32,i32,*fp32,i32,*fp32,i32", constants={"BLOCK_SIZE_N": 256}, num_warps=NUM_WARPS, device=0, output="ptx")
+
+    print(ret)