[WIP] adding DP4A support to rocm

Author: masahi

python/tvm/relay/op/strategy/cuda.py

@@ -145,7 +145,7 @@ def conv2d_strategy_cuda(attrs, inputs, out_type, target):
         if layout == "NCHW":
             assert kernel_layout == "OIHW"
             if (
-                (target.kind.name in ["cuda", "vulkan"])
+                (target.kind.name in ["cuda", "vulkan", "rocm"])
                 and data.dtype in ("int8", "uint8")
                 and kernel.dtype in ("int8", "uint8")
             ):
@@ -297,7 +297,7 @@ def conv2d_strategy_cuda(attrs, inputs, out_type, target):
                                     Need to satisfy tensor core schedule."
                 )
         elif (
-            (target.kind.name in ["cuda", "vulkan"])
+            (target.kind.name in ["cuda", "vulkan", "rocm"])
             and layout == "NCHW4c"
             and data.dtype in ["int8", "uint8"]
         ):
@@ -376,7 +376,7 @@ def conv2d_strategy_cuda(attrs, inputs, out_type, target):
             ic_chunk = in_channels // 4
 
             if (
-                (target.kind.name in ["cuda", "vulkan"])
+                (target.kind.name in ["cuda", "vulkan", "rocm"])
                 and data.dtype in ["int8", "uint8"]
                 and kernel.dtype in ["int8", "uint8"]
                 and channels % groups == 0
@@ -836,7 +836,7 @@ def dense_strategy_cuda(attrs, inputs, out_type, target):
     b, i = get_const_tuple(data.shape)
     o, _ = get_const_tuple(weights.shape)
     if (
-        target.kind.name in ["cuda", "vulkan"]
+        target.kind.name in ["cuda", "vulkan", "rocm"]
         and data.dtype == "int8"
         and weights.dtype == "int8"
         and out_type.dtype == "int32"

python/tvm/relay/qnn/op/legalizations.py

@@ -387,6 +387,12 @@ def is_aarch64_arm():
     return "aarch64" in target.attrs.get("mtriple", "")
 
 
+def is_rocm():
+    """Checks whether we are compiling for a rocm/spirv target."""
+    target = tvm.target.Target.current(allow_none=False)
+    return "rocm" in target.keys
+
+
 def is_vulkan():
     """Checks whether we are compiling for a vulkan/spirv target."""
     target = tvm.target.Target.current(allow_none=False)
@@ -456,7 +462,7 @@ def _qnn_dense_legalize_intel_cpu(attrs, inputs, types):
 
 @qnn_conv2d_legalize.register(["cuda", "gpu"])
 def _qnn_conv2d_legalize_cuda(attrs, inputs, types):
-    if is_vulkan():
+    if is_vulkan() or is_rocm():
         # prefers the dtypes to be same. Mixed type is not yet supported.
         return helper_change_dtypes_to_be_same(attrs, inputs, types, relay.qnn.op.conv2d)
     if is_cuda():
@@ -467,7 +473,7 @@ def _qnn_conv2d_legalize_cuda(attrs, inputs, types):
 
 @qnn_dense_legalize.register(["cuda", "gpu"])
 def _qnn_dense_legalize_cuda(attrs, inputs, types):
-    if is_vulkan():
+    if is_vulkan() or is_rocm():
         # prefers the dtypes to be same. Mixed type is not yet supported.
         return helper_change_dtypes_to_be_same(attrs, inputs, types, relay.qnn.op.dense)
     if is_cuda():

python/tvm/topi/cuda/batch_matmul.py

@@ -372,7 +372,7 @@ def _schedule_batch_matmul_int8(cfg, s, output):
     target = tvm.target.Target.current(allow_none=False)
     do_tensorize = True
 
-    if "vulkan" in target.keys:
+    if "vulkan" in target.keys or "rocm" in target.keys:
         do_tensorize = "+dotprod" in target.mattr or target.supports_integer_dot_product
 
     if do_tensorize:

python/tvm/topi/cuda/conv2d_alter_op.py

@@ -34,7 +34,7 @@
 @nn.conv2d_alter_layout.register(["cuda", "gpu"])
 def _alter_conv2d_layout(attrs, inputs, tinfos, out_type):
     target = tvm.target.Target.current(allow_none=False)
-    doit = "vulkan" in target.keys or "cuda" in target.keys
+    doit = "vulkan" in target.keys or "cuda" in target.keys or "rocm" in target.keys
     if not doit:
         return None
     dispatch_ctx = autotvm.task.DispatchContext.current
@@ -87,7 +87,7 @@ def _alter_conv2d_layout(attrs, inputs, tinfos, out_type):
     if cfg.is_fallback:  # if is fallback, clear query cache and return None
         autotvm.task.clear_fallback_cache(target, workload)
         do_new_layout = False
-        if "vulkan" in target.keys:
+        if "vulkan" in target.keys or "rocm" in target.keys:
             do_new_layout = "+dotprod" in target.mattr or target.supports_integer_dot_product
         if not do_new_layout:
             return None
@@ -351,7 +351,7 @@ def _conv2d_legalize(attrs, inputs, arg_types):
     """
 
     target = tvm.target.Target.current(allow_none=False)
-    doit = "vulkan" in target.keys or "cuda" in target.keys
+    doit = "vulkan" in target.keys or "cuda" in target.keys or "rocm" in target.keys
     if not doit:
         return None
     # Dilation not supported yet. Return None if dilation is not (1, 1)

python/tvm/topi/cuda/conv2d_int8.py

@@ -312,7 +312,7 @@ def _schedule_conv2d_NCHWc_int8(cfg, s, output):
     _, rc_block = s[conv].split(rc_block, factor=4)
     target = tvm.target.Target.current(allow_none=False)
     do_tensorize = True
-    if "vulkan" in target.keys:
+    if "vulkan" in target.keys or "rocm" in target.keys:
         do_tensorize = "+dotprod" in target.mattr or target.supports_integer_dot_product
     if do_tensorize:
         dtypes = (pad_data.dtype, packed_kernel.dtype)

python/tvm/topi/cuda/dense.py

@@ -173,7 +173,7 @@ def _schedule_dense_int8(cfg, s, output):
     ko, kt = cfg["tile_k"].apply(s, CC, ko)
     target = tvm.target.Target.current(allow_none=False)
     do_tensorize = True
-    if "vulkan" in target.keys:
+    if "vulkan" in target.keys or "rocm" in target.keys:
         do_tensorize = "+dotprod" in target.mattr or target.supports_integer_dot_product
     if do_tensorize:
         dtypes = (data.dtype, weight.dtype)