Fix conv2d_nchw for opencl intel graphics (#8201)

apache · Jun 7, 2021 · 51bbd63 · 51bbd63
1 parent 3ab4a6b
commit 51bbd63
Show file tree

Hide file tree

Showing 2 changed files with 3 additions and 1 deletion.
diff --git a/python/tvm/topi/intel_graphics/conv2d.py b/python/tvm/topi/intel_graphics/conv2d.py
@@ -580,7 +580,7 @@ def _schedule_cl_spatialpack(s, op):
     temp = s[conv].op.input_tensors[0]
     kernel_vec = s[conv].op.input_tensors[1]
     kernel = s[kernel_vec].op.input_tensors[0]
-    temp_W = s.cache_read(temp, "warp", [conv])
+    temp_W = s.cache_read(temp, "shared", [conv])
     conv_L = s.cache_write(conv, "local")
 
     kernel_L = s.cache_read(kernel_vec, "local", [conv_L])

diff --git a/tests/python/topi/python/test_topi_conv2d_nchw.py b/tests/python/topi/python/test_topi_conv2d_nchw.py
@@ -148,6 +148,8 @@ def check_target(target):
 
     if use_cudnn:
         check_target("cuda -model=unknown -libs=cudnn")
+    if ("opencl", tvm.device("opencl")) in tvm.testing.enabled_targets():
+        check_target("opencl -device=intel_graphics")
 
 
 @tvm.testing.uses_gpu