Enable miopen transpose convolution and fp16 support

apache · Sep 13, 2019 · 05f4fcc · 05f4fcc
1 parent eb220d9
commit 05f4fcc
Show file tree

Hide file tree

Showing 4 changed files with 42 additions and 32 deletions.
diff --git a/python/tvm/contrib/miopen.py b/python/tvm/contrib/miopen.py
@@ -49,7 +49,8 @@ def conv2d_forward(x,
                    pad_w=0,
                    dilation_h=1,
                    dilation_w=1,
-                   conv_mode=0):
+                   conv_mode=0,
+                   data_type=1):
     """Create an extern op that compute 2D convolution with MIOpen
 
     Parameters
@@ -73,18 +74,22 @@ def conv2d_forward(x,
     conv_mode: int
         0: miopenConvolution
         1: miopenTranspose
+    data_type: int
+        0: miopenHalf (fp16)
+        1: miopenFloat (fp32)
 
     Returns
     -------
     y: Tensor
         The result tensor
     """
-    assert conv_mode == 0, "Transpose convolutions not supported yet."
+    assert (conv_mode == 0 or conv_mode == 1), "0: miopenConvolution / 1: miopenTranspose"
     oshape = np.zeros((len(x.shape)), dtype=np.int32)
     xshape = x.shape
     wshape = w.shape
     setup_func = _get_global_func("tvm.contrib.miopen.conv2d.setup")
     algo = setup_func(conv_mode,
+                      data_type,
                       pad_h,
                       pad_w,
                       stride_h,
@@ -106,6 +111,7 @@ def conv2d_forward(x,
         lambda ins, outs: _intrin.call_packed(
             "tvm.contrib.miopen.conv2d.forward",
             conv_mode,
+            data_type,
             pad_h,
             pad_w,
             stride_h,

diff --git a/src/contrib/miopen/conv_forward.cc b/src/contrib/miopen/conv_forward.cc
@@ -35,29 +35,30 @@ using namespace runtime;
 TVM_REGISTER_GLOBAL("tvm.contrib.miopen.conv2d.setup")
 .set_body([](TVMArgs args, TVMRetValue *ret) {
   const int mode = args[0];
-  const int pad_h = args[1];
-  const int pad_w = args[2];
-  const int stride_h = args[3];
-  const int stride_w = args[4];
-  const int dilation_h = args[5];
-  const int dilation_w = args[6];
-  const int x_dim0 = args[7];
-  const int x_dim1 = args[8];
-  const int x_dim2 = args[9];
-  const int x_dim3 = args[10];
-  const int w_dim0 = args[11];
-  const int w_dim1 = args[12];
-  const int w_dim2 = args[13];
-  const int w_dim3 = args[14];
-  void *out_shape = args[15];
+  const int dtype = args[1];
+  const int pad_h = args[2];
+  const int pad_w = args[3];
+  const int stride_h = args[4];
+  const int stride_w = args[5];
+  const int dilation_h = args[6];
+  const int dilation_w = args[7];
+  const int x_dim0 = args[8];
+  const int x_dim1 = args[9];
+  const int x_dim2 = args[10];
+  const int x_dim3 = args[11];
+  const int w_dim0 = args[12];
+  const int w_dim1 = args[13];
+  const int w_dim2 = args[14];
+  const int w_dim3 = args[15];
+  void *out_shape = args[16];
 
   MIOpenThreadEntry* entry_ptr = MIOpenThreadEntry::ThreadLocal();
   // Set Mode
   entry_ptr->conv_entry.mode = static_cast<miopenConvolutionMode_t>(mode);
   // Set Ctx
   entry_ptr->conv_entry.ctx = TVMContext{kDLROCM, 0};
   // Set Data Type
-  entry_ptr->conv_entry.data_type = miopenFloat;  // MIOpen only suppports fp32
+  entry_ptr->conv_entry.data_type = static_cast<miopenDataType_t>(dtype);  // MIOpen suppports fp32(miopenFloat), fp16(miopenHalf) at this moment.
   // Set Desc
   MIOPEN_CALL(miopenInitConvolutionDescriptor(entry_ptr->conv_entry.conv_desc,
                                               entry_ptr->conv_entry.mode,
@@ -170,16 +171,17 @@ TVM_REGISTER_GLOBAL("tvm.contrib.miopen.conv2d.setup")
 TVM_REGISTER_GLOBAL("tvm.contrib.miopen.conv2d.forward")
 .set_body([](TVMArgs args, TVMRetValue *ret) {
   const int mode = args[0];
-  const int pad_h = args[1];
-  const int pad_w = args[2];
-  const int stride_h = args[3];
-  const int stride_w = args[4];
-  const int dilation_h = args[5];
-  const int dilation_w = args[6];
-  const int algo = args[7];
-  const DLTensor *x = args[8];
-  const DLTensor *w = args[9];
-  const DLTensor *y = args[10];
+  const int dtype = args[1];
+  const int pad_h = args[2];
+  const int pad_w = args[3];
+  const int stride_h = args[4];
+  const int stride_w = args[5];
+  const int dilation_h = args[6];
+  const int dilation_w = args[7];
+  const int algo = args[8];
+  const DLTensor *x = args[9];
+  const DLTensor *w = args[10];
+  const DLTensor *y = args[11];
 
   MIOpenThreadEntry* entry_ptr = MIOpenThreadEntry::ThreadLocal();
   entry_ptr->conv_entry.fwd_algo = static_cast<miopenConvFwdAlgorithm_t>(algo);
@@ -188,7 +190,7 @@ TVM_REGISTER_GLOBAL("tvm.contrib.miopen.conv2d.forward")
   // Set Ctx
   entry_ptr->conv_entry.ctx = x->ctx;
   // Set Data Type
-  entry_ptr->conv_entry.data_type = miopenFloat;  // MIOpen only suppports fp32
+  entry_ptr->conv_entry.data_type = static_cast<miopenDataType_t>(dtype);  // MIOpen suppports fp32(miopenFloat), fp16(miopenHalf) at this moment.
   // Set Desc
   MIOPEN_CALL(miopenInitConvolutionDescriptor(entry_ptr->conv_entry.conv_desc,
                                               entry_ptr->conv_entry.mode,

diff --git a/tests/python/contrib/test_miopen.py b/tests/python/contrib/test_miopen.py
@@ -50,7 +50,8 @@ def test_conv2d():
                               pad_w,
                               dilation_h,
                               dilation_w,
-                              conv_mode=0)
+                              conv_mode=0,
+                              data_type=1)
 
     yshape = [x.value for x in Y.shape]
     import topi
@@ -65,7 +66,7 @@ def verify():
         y = tvm.nd.array(np.random.uniform(-1, 1, yshape).astype(np.float32), ctx)
         f(x, w, y)
 
-        Y_ref = topi.nn.conv2d_nchw(X, W, (stride_h, stride_w), (pad_h, pad_w))
+        Y_ref = topi.nn.conv2d_nchw(X, W, (stride_h, stride_w), (pad_h, pad_w), (dilation_h, dilation_w))
         with tvm.target.rocm():
             s_ref = topi.generic.schedule_conv2d_nchw([Y_ref])
         f_ref = tvm.build(s_ref, [X, W, Y_ref], "rocm")

diff --git a/topi/python/topi/rocm/conv2d.py b/topi/python/topi/rocm/conv2d.py
@@ -78,7 +78,8 @@ def conv2d_rocm(cfg, data, kernel, strides, padding, dilation, layout='NCHW', ou
                                      pad_w,
                                      dilation_h,
                                      dilation_w,
-                                     conv_mode=0)
+                                     conv_mode=0,
+                                     data_type=1)
 
     return conv2d_cuda(cfg, data, kernel, strides, padding, dilation, layout, out_dtype)