From cded42ec99b1bccae276789ffc83ff6d3e615db9 Mon Sep 17 00:00:00 2001
From: Andreas Kloeckner <inform@tiker.net>
Date: Sat, 10 Nov 2018 16:36:00 -0600
Subject: [PATCH] Fix CUDA short vector codegen

---
 loopy/target/cuda.py |  2 +-
 test/test_target.py  | 17 +++++++++++++++++
 2 files changed, 18 insertions(+), 1 deletion(-)

diff --git a/loopy/target/cuda.py b/loopy/target/cuda.py
index 673d3b284..d6f55091a 100644
--- a/loopy/target/cuda.py
+++ b/loopy/target/cuda.py
@@ -290,7 +290,7 @@ class CUDACASTBuilder(CASTBuilder):
     _VEC_AXES = "xyzw"
 
     def add_vector_access(self, access_expr, index):
-        return access_expr.a(self._VEC_AXES[index])
+        return access_expr.attr(self._VEC_AXES[index])
 
     def emit_barrier(self, synchronization_kind, mem_kind, comment):
         """
diff --git a/test/test_target.py b/test/test_target.py
index 75b3c05ae..bcf85a340 100644
--- a/test/test_target.py
+++ b/test/test_target.py
@@ -358,6 +358,23 @@ def test_ispc_streaming_stores():
     lp.generate_code_v2(knl).all_code()
 
 
+def test_cuda_short_vector():
+    knl = lp.make_kernel(
+        "{ [i]: 0<=i<n }",
+        "out[i] = 2*a[i]",
+        target=lp.CudaTarget())
+
+    knl = lp.set_options(knl, write_code=True)
+    knl = lp.split_iname(knl, "i", 4, slabs=(0, 1), inner_tag="vec")
+    knl = lp.split_array_axis(knl, "a,out", axis_nr=0, count=4)
+    knl = lp.tag_array_axes(knl, "a,out", "C,vec")
+
+    knl = lp.set_options(knl, write_wrapper=True)
+    knl = lp.add_and_infer_dtypes(knl, {"a": np.float32})
+
+    print(lp.generate_code_v2(knl).device_code())
+
+
 if __name__ == "__main__":
     if len(sys.argv) > 1:
         exec(sys.argv[1])
-- 
GitLab