data copy ops (pytorch#15164)

Summary:

References for data copies.

Reviewed By: skrtskrtfb

Differential Revision: D84674296

Author: Andrew Grebenisan

backends/cadence/aot/ops_registrations.py

@@ -55,8 +55,6 @@ def _validate_ref_impl_exists() -> None:
     _WARN_ONLY = {
         "cadence::quantized_w8a32_linear",
         "cadence::quantized_add",  # We should only support per_tensor variant, should remove
-        "cadence::idma_store",
-        "cadence::idma_load",
         "cadence::_softmax_f32_f32",
         "cadence::requantize",  # We should only support per_tensor variant, should remove
         "cadence::quantized_softmax.per_tensor",
@@ -70,13 +68,11 @@ def _validate_ref_impl_exists() -> None:
         "cadence::quantized_relu",  # We should only support per_tensor variant, should remove
         "cadence::linalg_svd",
         "cadence::quantized_conv2d_nhwc",  # We should only support per_tensor variant, should remove
-        "cadence::idma_copy",
         "cadence::quantize_per_tensor_asym16u",
         "cadence::dequantize_per_tensor_asym8s",
         "cadence::quantize_per_tensor_asym16s",
         "cadence::dequantize_per_tensor_asym16s",
         "cadence::quantized_softmax",
-        "cadence::idma_wait",
         "cadence::quantized_w8a32_gru",
         "cadence::quantized_layer_norm",  # We should only support per_tensor variant, should remove
     }
@@ -2003,6 +1999,7 @@ def im2row_per_tensor_meta(
     )
     return input.new_empty(output_size, dtype=input.dtype)
 
+
 @register_fake("cadence::linalg_svd")
 def linalg_svd_meta(
     A: torch.Tensor,

backends/cadence/aot/ref_implementations.py

@@ -1624,3 +1624,23 @@ def quantized_embedding_byte(
     )
 
     return weight[indices]
+
+
+@impl_tracked(m, "idma_copy")
+def idma_copy(src: torch.Tensor, task_num: int = 0, channel: int = 0) -> torch.Tensor:
+    return src.clone()
+
+
+@impl_tracked(m, "idma_store")
+def idma_store(src: torch.Tensor, task_num: int = 0, channel: int = 0) -> torch.Tensor:
+    return src.clone()
+
+
+@impl_tracked(m, "idma_load")
+def idma_load(src: torch.Tensor, task_num: int = 0, channel: int = 0) -> torch.Tensor:
+    return src.clone()
+
+
+@impl_tracked(m, "idma_wait")
+def idma_wait(src: torch.Tensor, task_num: int = 0, channel: int = 0) -> torch.Tensor:
+    return src.clone()