Merge pull request #162 from PSAL-POSTECH/ViT

ViT E2E model support

Author: YWHyuk

.github/workflows/pytorchsim_test.yml

@@ -42,7 +42,7 @@ RUN wget https://github.com/riscv-collab/riscv-gnu-toolchain/releases/download/2
     rm *.tar.gz
 
 # Install torchsim dependency
-RUN apt install ninja-build && pip install onnx matplotlib && pip install --user conan==1.56.0
+RUN apt install ninja-build && pip install onnx matplotlib && pip install --user conan==1.56.0 && pip install "transformers<4.44" && pip install diffusers==0.34.0
 
 ENV RISCV=/workspace/riscv
 ENV PATH=$RISCV/bin:$PATH

Dockerfile.base

@@ -11,6 +11,7 @@
 #include <ATen/native/cpu/Loops.h>
 #include <ATen/native/DispatchStub.h>
 #include <ATen/native/Resize.h>
+#include <ATen/native/TensorFactories.h>
 #include <ATen/EmptyTensor.h>
 #include <ATen/core/GeneratorForPrivateuseone.h>
 #include <ATen/NativeFunctions.h>
@@ -204,19 +205,25 @@ at::Tensor custom__copy_from(const at::Tensor& self, const at::Tensor& dst, bool
 
   // Some dummy asserts for the basic use case: inputs are the same size / dtype, all contiguous.
   TORCH_CHECK(self.sizes() == dst.sizes());
-  TORCH_CHECK(self.scalar_type() == dst.scalar_type());
 
-  if (self.is_contiguous() && dst.is_contiguous()) {
+  const bool same_dtype = (self.scalar_type() == dst.scalar_type());
+  const bool both_contig = self.is_contiguous() && dst.is_contiguous();
+
+  // 1) fast path
+  if (same_dtype && both_contig) {
     std::memcpy(dst.mutable_data_ptr(),
                 self.data_ptr(),
                 dst.storage().nbytes());
-  } else {
-    // Using cpu tensor to accomplishment stride copy.
-    at::Tensor cpu_self = unsafe_create_cpu_tensor_from_dummy_tensor(self);
-    at::Tensor cpu_dst = unsafe_create_cpu_tensor_from_dummy_tensor(dst);
-    cpu_dst.copy_(cpu_self);
+    return dst;
   }
 
+  // 2) slow path
+  at::Tensor cpu_self = unsafe_create_cpu_tensor_from_dummy_tensor(self);
+  at::Tensor cpu_dst  = unsafe_create_cpu_tensor_from_dummy_tensor(dst);
+  if (!same_dtype) {
+    cpu_self = cpu_self.to(cpu_dst.scalar_type(), /*non_blocking=*/false, /*copy=*/true);
+  }
+  cpu_dst.copy_(cpu_self);
   return dst;
 }
 
@@ -230,7 +237,6 @@ at::Tensor& custom_abs_out(const at::Tensor& self, at::Tensor& out) {
 
 at::Tensor custom_empty_strided(c10::IntArrayRef size, c10::IntArrayRef stride, c10::optional<at::ScalarType> dtype_opt, c10::optional<at::Layout> layout_opt, c10::optional<at::Device> device_opt, c10::optional<bool> pin_memory_opt) {
   op_counter += 1;
-
   constexpr c10::DispatchKeySet private_use_ks(c10::DispatchKey::PrivateUse1);
   auto dtype = c10::dtype_or_default(dtype_opt);
   return  at::detail::empty_strided_generic(size, stride, &global_custom_alloc, private_use_ks, dtype);
@@ -244,7 +250,23 @@ at::Tensor custom_empty(c10::IntArrayRef size, c10::optional<at::ScalarType> dty
   return  at::detail::empty_generic(size, &global_custom_alloc, private_use_ks, dtype, optional_memory_format);
 }
 
-// This macro does the heavy lifting.
+at::Tensor& custom_arange_start_out_impl(
+    const c10::Scalar& start,
+    const c10::Scalar& end,
+    const c10::Scalar& step,
+    at::Tensor& out) {
+    //const int64_t n = arange_len(start.toDouble(), end.toDouble(), step.toDouble());
+    //at::native::resize_output(out, {n});
+    return out;
+}
+
+static at::Tensor custom_to_dtype_impl(const at::Tensor& self,
+                                       c10::ScalarType dtype,
+                                       bool non_blocking, bool copy,
+                                       c10::optional<c10::MemoryFormat> memory_format) {
+  return at::native::to(self, dtype, non_blocking, copy, memory_format);
+}
+
 // With TORCH_LIBRARY_IMPL, you can register custom kernels for your backend.
 // For open registration, we're registering all of our kernels to the PrivateUse1 dispatch key.
 // Later in this file, we map a custom device to the PrivateUse1 device type,
@@ -255,21 +277,27 @@ at::Tensor custom_empty(c10::IntArrayRef size, c10::optional<at::ScalarType> dty
 // More details on the dispatcher can be found at http://blog.ezyang.com/2020/09/lets-talk-about-the-pytorch-dispatcher/.
 TORCH_LIBRARY_IMPL(aten, PrivateUse1, m) {
   m.impl("to.Device", &custom_to_device);
+  m.impl("to.dtype", &custom_to_dtype_impl);
   m.impl("fill_.Scalar", &custom_fill__scalar);
   m.impl("_copy_from", &custom__copy_from);
   m.impl("_copy_from_and_resize", &custom__copy_from_and_resize);
   m.impl("empty_strided", &custom_empty_strided);
   m.impl("empty.memory_format", &custom_empty);
   m.impl("as_strided", at::native::as_strided_tensorimpl);
   m.impl("view", at::native::view);
+  m.impl("arange.start_out", &custom_arange_start_out_impl);
+}
+
+TORCH_LIBRARY_IMPL(aten, AutogradPrivateUse1, m) {
+  m.impl("to.dtype", &custom_to_dtype_impl);
 }
 
 TORCH_LIBRARY_FRAGMENT(aten, m) {
-  m.def(
-    "_reinterpret_tensor(Tensor self, int[] size, int[] stride, int offset_increment=0) -> Tensor",
-    torch::dispatch(
-        c10::DispatchKey::AutogradPrivateUse1, _reinterpret_tensor),
-    {at::Tag::pt2_compliant_tag});
+m.def(
+  "_reinterpret_tensor(Tensor self, int[] size, int[] stride, int offset_increment=0) -> Tensor",
+  torch::dispatch(
+      c10::DispatchKey::AutogradPrivateUse1, _reinterpret_tensor),
+  {at::Tag::pt2_compliant_tag});
 }
 
 void custom_cpu_fallback(const c10::OperatorHandle& op, torch::jit::Stack* stack) {
@@ -307,6 +335,9 @@ TORCH_LIBRARY_IMPL(aten, PrivateUse1, m) {
   m.impl("_foreach_addcdiv_.ScalarList", torch::CppFunction::makeFromBoxedFunction<&custom_cpu_fallback>());
   m.impl("_foreach_add_.List", torch::CppFunction::makeFromBoxedFunction<&custom_cpu_fallback>());
   m.impl("cat.out", torch::CppFunction::makeFromBoxedFunction<&custom_cpu_fallback>());
+  m.impl("_native_multi_head_attention", torch::CppFunction::makeFromBoxedFunction<&custom_cpu_fallback>());
+  m.impl("resize_", torch::CppFunction::makeFromBoxedFunction<&custom_cpu_fallback>());
+  m.impl("exp.out", torch::CppFunction::makeFromBoxedFunction<&custom_cpu_fallback>());
 }
 
 // This basic implementation doesn't bother dealing with different device indices

PyTorchSimFrontend/extension_device.cpp

@@ -173,9 +173,9 @@ def render(self,
 
         W_tensor =  empty_strided(W.layout.size, W.layout.stride)
         X_tensor =  empty_strided(X.layout.size, X.layout.stride)
-        if len(W_tensor.size()) > 3:
+        if len(W_tensor.size()) > 3 or len(W_tensor.size()) == 2:
           W_tensor = W_tensor.view([-1, W_tensor.shape[-2], W_tensor.shape[-1]])
-        if len(X_tensor.size()) > 3:
+        if len(X_tensor.size()) > 3 or len(X_tensor.size()) == 2:
           X_tensor = X_tensor.view([-1, X_tensor.shape[-2], X_tensor.shape[-1]])
         B, M, N, K = X_tensor.size()[0], X_tensor.size()[1], W_tensor.size()[2], X_tensor.size()[2]
 
@@ -217,6 +217,7 @@ def render(self,
         X_tile_desc = mlir_common.MLIRMultiDimTile(X_tile_size, kernel.vector_lane, vlane_split_axis, vlane_stride)
         X_tile_desc.set_tile_size_stride(X_tile_size, X_tile_stride)
         X_tile_desc.set_name("X_buffer")
+        X_tile_desc.offset = X.get_layout().offset
         X_stride = X_tensor.stride()
         X_idx = [loop_dim[0]*X_stride[0], loop_dim[1]*X_stride[1], loop_dim[3]*X_stride[2]] # To keep index arguemnt order, we used index_list
 
@@ -225,6 +226,7 @@ def render(self,
         W_tile_desc = mlir_common.MLIRMultiDimTile(X_tile_size, kernel.vector_lane, vlane_split_axis, vlane_stride)
         W_tile_desc.set_tile_size_stride(W_tile_size, W_tile_stride)
         W_tile_desc.set_name("W_buffer")
+        W_tile_desc.offset = W.get_layout().offset
         W_stride = W_tensor.stride()
         W_idx = [loop_dim[0]*W_stride[0], loop_dim[3]*W_stride[1], loop_dim[2]*W_stride[2]]
 
@@ -241,8 +243,12 @@ def render(self,
           Y_idx = [loop_dim[0]*Y_stride[0], loop_dim[2]*Y_stride[2], loop_dim[1]*Y_stride[1]]
 
         # Extract Bias info
+        Bias_tile_desc = mlir_common.MLIRMultiDimTile(Y_tile_size, kernel.vector_lane, vlane_split_axis, vlane_stride)
+        Bias_tile_desc.set_tile_size_stride(Y_tile_size, Y_tile_stride)
+        Bias_tile_desc.set_name("Y_buffer")
         if Bias is not None:
           Bias_stride = Bias.get_layout().stride
+          Bias_tile_desc.offset = Bias.get_layout().offset
           if nr_rdim == 0:
             Bias_idx = [loop_dim[0]*Bias_stride[0], loop_dim[1]*Bias_stride[1], loop_dim[2]*Bias_stride[2]]
           else: