diff --git a/hpvm/llvm_patches/include/IR/IntrinsicsApproxHPVM.td b/hpvm/llvm_patches/include/IR/IntrinsicsApproxHPVM.td
deleted file mode 100644
index 0287f083849d9385faa00cc2943834b5b58439b0..0000000000000000000000000000000000000000
--- a/hpvm/llvm_patches/include/IR/IntrinsicsApproxHPVM.td
+++ /dev/null
@@ -1,311 +0,0 @@
-//===- IntrinsicsHPVM.td - Defines HPVM intrinsics ---------*- tablegen -*-===//
-//
-// The LLVM Compiler Infrastructure
-//
-// This file is distributed under the University of Illinois Open Source
-// License. See LICENSE.TXT for details.
-//
-//===----------------------------------------------------------------------===//
-//
-// This file defines all of the HPVM-specific intrinsics.
-//
-//===----------------------------------------------------------------------===//
-
-let TargetPrefix = "hpvm" in {
- /* All intrinsics start with "llvm.hpvm."
- * As we do not want the compiler to mess with these intrinsics, we assume
- * worst memory behavior for all these intrinsics.
- */
-
- /* Initialization intrinsic -
- * i8* llvm.hpvm.setup(function*);
- */
- def int_hpvm_init : Intrinsic<[], [], []>;
-
- /* Launch intrinsic - with streaming argument
- * i8* llvm.hpvm.launch(i8*, ArgList*, i1);
- */
- def int_hpvm_launch : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
- llvm_ptr_ty, llvm_i1_ty], []>;
-
- /* Push intrinsic - push data on streaming pipeline
- * void llvm.hpvm.push(i8*, ArgList*);
- */
- def int_hpvm_push : Intrinsic<[], [llvm_ptr_ty, llvm_ptr_ty], []>;
-
- /* Pop intrinsic - pop data from streaming pipeline
- * i8* llvm.hpvm.pop(i8*);
- */
- def int_hpvm_pop : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty], []>;
-
- /* Cleanup intrinsic -
- * void llvm.hpvm.cleanup(i8*);
- */
- def int_hpvm_cleanup : Intrinsic<[], [], []>;
-
- /* Wait intrinsic -
- * void llvm.hpvm.wait(graphID*);
- */
- def int_hpvm_wait : Intrinsic<[], [llvm_ptr_ty], []>;
-
- /* Track memory intrinsic -
- * void llvm.hpvm.trackMemory(i8*, i64);
- */
- def int_hpvm_trackMemory : Intrinsic<[], [llvm_ptr_ty, llvm_i64_ty], []>;
-
- /* Track memory intrinsic -
- * void llvm.hpvm.untrackMemory(i8*);
- */
- def int_hpvm_untrackMemory : Intrinsic<[], [llvm_ptr_ty], []>;
-
- /* Request memory intrinsic -
- * void llvm.hpvm.requestMemory(i8*, i64);
- */
- def int_hpvm_requestMemory : Intrinsic<[], [llvm_ptr_ty, llvm_i64_ty], []>;
-
- /* Create Node intrinsic -
- * i8* llvm.hpvm.createNode(function*);
- */
- def int_hpvm_createNode : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty], []>;
-
- /* Create Node 1D array intrinsic -
- * i8* llvm.hpvm.createNode1D(function*, i64);
- */
- def int_hpvm_createNode1D : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
- llvm_i64_ty], []>;
-
- /* Create Node 2D array intrinsic -
- * i8* llvm.hpvm.createNode2D(function*, i64, i64);
- */
- def int_hpvm_createNode2D : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
- llvm_i64_ty, llvm_i64_ty], []>;
-
- /* Create Node 3D array intrinsic -
- * i8* llvm.hpvm.createNode2D(function*, i64, i64, i64);
- */
- def int_hpvm_createNode3D : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
- llvm_i64_ty, llvm_i64_ty, llvm_i64_ty],
- []>;
-
- /* Create dataflow edge intrinsic -
- * i8* llvm.hpvm.createEdge(i8*, i8*, i1, i32, i32, i1);
- */
- def int_hpvm_createEdge : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty, llvm_ptr_ty,
- llvm_i1_ty, llvm_i32_ty, llvm_i32_ty,
- llvm_i1_ty],
- []>;
-
- /* Create bind input intrinsic -
- * void llvm.hpvm.bind.input(i8*, i32, i32);
- */
- def int_hpvm_bind_input : Intrinsic<[], [llvm_ptr_ty, llvm_i32_ty,
- llvm_i32_ty, llvm_i1_ty], []>;
-
- /* Create bind output intrinsic -
- * void llvm.hpvm.bind.output(i8*, i32, i32);
- */
- def int_hpvm_bind_output : Intrinsic<[], [llvm_ptr_ty, llvm_i32_ty,
- llvm_i32_ty, llvm_i1_ty], []>;
-
- /* Find associated dataflow node intrinsic -
- * i8* llvm.hpvm.getNode();
- */
- def int_hpvm_getNode : Intrinsic<[llvm_ptr_ty], [], [IntrNoMem]>;
-
- /* Find parent dataflow node intrinsic -
- * i8* llvm.hpvm.getParentNode(i8*);
- */
- def int_hpvm_getParentNode : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty], [IntrNoMem]>;
-
- /* Find the number of dimensions of a dataflow node intrinsic -
- * i32 llvm.hpvm.getNumDims(i8*);
- */
- def int_hpvm_getNumDims : Intrinsic<[llvm_i32_ty], [llvm_ptr_ty], [IntrNoMem]>;
-
- /* Find the unique indentifier of a dataflow node (with respect to his parent
- * node) in the specified dimension intrinsic -
- */
-
- /* i64 llvm.hpvm.getNodeInstanceID.[xyz](i8*);
- */
- def int_hpvm_getNodeInstanceID_x : Intrinsic<[llvm_i64_ty], [llvm_ptr_ty],
- [IntrNoMem]>;
-
- def int_hpvm_getNodeInstanceID_y : Intrinsic<[llvm_i64_ty], [llvm_ptr_ty],
- [IntrNoMem]>;
-
- def int_hpvm_getNodeInstanceID_z : Intrinsic<[llvm_i64_ty], [llvm_ptr_ty],
- [IntrNoMem]>;
-
- /* Find the number of instances of a dataflow node in the specified dimension
- * intrinsic -
- */
-
- /* i64 llvm.hpvm.getNumNodeInstances.[xyz](i8*);
- */
- def int_hpvm_getNumNodeInstances_x : Intrinsic<[llvm_i64_ty], [llvm_ptr_ty],
- [IntrNoMem]>;
-
- def int_hpvm_getNumNodeInstances_y : Intrinsic<[llvm_i64_ty], [llvm_ptr_ty],
- [IntrNoMem]>;
-
- def int_hpvm_getNumNodeInstances_z : Intrinsic<[llvm_i64_ty], [llvm_ptr_ty],
- [IntrNoMem]>;
-
- /* Local Barrier
- * void llvm.hpvm.barrier();
- */
- def int_hpvm_barrier : Intrinsic<[], [], []>;
-
- /* Memory allocation inside the graph
- * i8* llvm.hpvm.malloc();
- */
- def int_hpvm_malloc : Intrinsic<[llvm_ptr_ty], [llvm_i64_ty], []>;
-
- /* Find the vector length supported by target architecture
- * intrinsic -
- * i32 llvm.hpvm.getVectorLength();
- */
- def int_hpvm_getVectorLength : Intrinsic<[llvm_i32_ty], [], []>;
-
- /* ============ Atomic intrinsics ============= */
- // Atomic arithmetic operations
-
- /* i32 llvm.hpvm.atomic.add(i32*, i32)*/
- def int_hpvm_atomic_add: Intrinsic<[llvm_i32_ty], [llvm_ptr_ty, llvm_i32_ty],
- []>;
-
- /* i32 llvm.hpvm.atomic.sub(i32*, i32)*/
- def int_hpvm_atomic_sub: Intrinsic<[llvm_i32_ty], [llvm_ptr_ty, llvm_i32_ty],
- []>;
-
- /* i32 llvm.hpvm.atomic.xchg(i32*, i32)*/
- def int_hpvm_atomic_xchg: Intrinsic<[llvm_i32_ty], [llvm_ptr_ty, llvm_i32_ty],
- []>;
-
- /* i32 llvm.hpvm.atomic.min(i32*, i32)*/
- def int_hpvm_atomic_min: Intrinsic<[llvm_i32_ty], [llvm_ptr_ty, llvm_i32_ty],
- []>;
-
- /* i32 llvm.hpvm.atomic.maxi32*, i32)*/
- def int_hpvm_atomic_max: Intrinsic<[llvm_i32_ty], [llvm_ptr_ty, llvm_i32_ty],
- []>;
-
- // Atomic bitwise operations
-
- /* i32 llvm.hpvm.atomic.and(i32*, i32)*/
- def int_hpvm_atomic_and: Intrinsic<[llvm_i32_ty], [llvm_ptr_ty, llvm_i32_ty],
- []>;
-
- /* i32 llvm.hpvm.atomic.or(i32*, i32)*/
- def int_hpvm_atomic_or: Intrinsic<[llvm_i32_ty], [llvm_ptr_ty, llvm_i32_ty],
- []>;
-
- /* i32 llvm.hpvm.atomic.xor(i32*, i32)*/
- def int_hpvm_atomic_xor: Intrinsic<[llvm_i32_ty], [llvm_ptr_ty, llvm_i32_ty],
- []>;
-
-
- /***************************************************************************/
- /* ApproxHPVM intrinsics */
- /***************************************************************************/
-
- /* Tensor add intrinsic
- * i8* llvm.hpvm.tensor.add(i8*, i8*);
- */
- def int_hpvm_tensor_add : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
- llvm_ptr_ty], []>;
-
- /* Tensor mul intrinsic
- * i8* llvm.hpvm.tensor.mul(i8*, i8*);
- */
- def int_hpvm_tensor_mul : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
- llvm_ptr_ty], []>;
-
- /* Tensor relu intrinsic
- * i8* llvm.hpvm.tensor.relu(i8*);
- */
- def int_hpvm_tensor_relu : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty], []>;
-
- /* Tensor clipped relu intrinsic
- * i8* llvm.hpvm.tensor.clipped.relu(i8*);
- */
- def int_hpvm_tensor_clipped_relu : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty], []>;
-
- /* Tensor tanh intrinsic
- * i8* llvm.hpvm.tensor.tanh(i8*);
- */
- def int_hpvm_tensor_tanh : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty], []>;
-
- /* Tensor sigmoid intrinsic
- * i8* llvm.hpvm.tensor.sigmoid(i8*);
- */
- def int_hpvm_tensor_sigmoid : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty], []>;
-
- /* Tensor softmax intrinsic
- * i8* llvm.hpvm.tensor.softmax(i8*);
- */
- def int_hpvm_tensor_softmax : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty], []>;
-
- /* Tensor convolution intrinsic
- * i8* llvm.hpvm.tensor.convolution(i8*, i8*, i32, i32, i32, i32);
- */
- def int_hpvm_tensor_convolution : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
- llvm_ptr_ty,
- llvm_i32_ty,
- llvm_i32_ty,
- llvm_i32_ty,
- llvm_i32_ty], []>;
-
- /* Tensor group convolution intrinsic
- * i8* llvm.hpvm.tensor.group.convolution(i8*, i8*, i32, i32, i32, i32, i32, i32);
- */
- def int_hpvm_tensor_group_convolution : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
- llvm_ptr_ty,
- llvm_i32_ty,
- llvm_i32_ty,
- llvm_i32_ty,
- llvm_i32_ty,
- llvm_i32_ty,
- llvm_i32_ty], []>;
-
- /* Tensor BatchNorm intrinsic
- * i8* llvm.hpvm.tensor.batchnorm(i8*, i8*, i8*, i8*, i8*, double);
- */
- def int_hpvm_tensor_batchnorm : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
- llvm_ptr_ty,
- llvm_ptr_ty,
- llvm_ptr_ty,
- llvm_ptr_ty,
- llvm_double_ty], []>;
-
-
- /* Tensor pool intrinsics: max, min, average
- * i8* llvm.hpvm.tensor.pool.max(i8*, i32, i32, i32, i32, i32, i32);
- * i8* llvm.hpvm.tensor.pool.min(i8*, i32, i32, i32, i32, i32, i32);
- * i8* llvm.hpvm.tensor.pool.average(i8*, i32, i32, i32, i32, i32, i32);
- */
- def int_hpvm_tensor_pool_max : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
- llvm_i32_ty,
- llvm_i32_ty,
- llvm_i32_ty,
- llvm_i32_ty,
- llvm_i32_ty,
- llvm_i32_ty], []>;
- def int_hpvm_tensor_pool_min : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
- llvm_i32_ty,
- llvm_i32_ty,
- llvm_i32_ty,
- llvm_i32_ty,
- llvm_i32_ty,
- llvm_i32_ty], []>;
- def int_hpvm_tensor_pool_mean : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
- llvm_i32_ty,
- llvm_i32_ty,
- llvm_i32_ty,
- llvm_i32_ty,
- llvm_i32_ty,
- llvm_i32_ty], []>;
-
- def int_hpvm_node_id : Intrinsic<[llvm_ptr_ty], [llvm_i32_ty], []>;
-}
diff --git a/hpvm/llvm_patches/include/IR/IntrinsicsHPVM.td b/hpvm/llvm_patches/include/IR/IntrinsicsHPVM.td
index 410e9c8d3345e67df9614e0d518e5e596a4368e1..0287f083849d9385faa00cc2943834b5b58439b0 100644
--- a/hpvm/llvm_patches/include/IR/IntrinsicsHPVM.td
+++ b/hpvm/llvm_patches/include/IR/IntrinsicsHPVM.td
@@ -205,4 +205,107 @@ let TargetPrefix = "hpvm" in {
def int_hpvm_atomic_xor: Intrinsic<[llvm_i32_ty], [llvm_ptr_ty, llvm_i32_ty],
[]>;
+
+ /***************************************************************************/
+ /* ApproxHPVM intrinsics */
+ /***************************************************************************/
+
+ /* Tensor add intrinsic
+ * i8* llvm.hpvm.tensor.add(i8*, i8*);
+ */
+ def int_hpvm_tensor_add : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
+ llvm_ptr_ty], []>;
+
+ /* Tensor mul intrinsic
+ * i8* llvm.hpvm.tensor.mul(i8*, i8*);
+ */
+ def int_hpvm_tensor_mul : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
+ llvm_ptr_ty], []>;
+
+ /* Tensor relu intrinsic
+ * i8* llvm.hpvm.tensor.relu(i8*);
+ */
+ def int_hpvm_tensor_relu : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty], []>;
+
+ /* Tensor clipped relu intrinsic
+ * i8* llvm.hpvm.tensor.clipped.relu(i8*);
+ */
+ def int_hpvm_tensor_clipped_relu : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty], []>;
+
+ /* Tensor tanh intrinsic
+ * i8* llvm.hpvm.tensor.tanh(i8*);
+ */
+ def int_hpvm_tensor_tanh : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty], []>;
+
+ /* Tensor sigmoid intrinsic
+ * i8* llvm.hpvm.tensor.sigmoid(i8*);
+ */
+ def int_hpvm_tensor_sigmoid : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty], []>;
+
+ /* Tensor softmax intrinsic
+ * i8* llvm.hpvm.tensor.softmax(i8*);
+ */
+ def int_hpvm_tensor_softmax : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty], []>;
+
+ /* Tensor convolution intrinsic
+ * i8* llvm.hpvm.tensor.convolution(i8*, i8*, i32, i32, i32, i32);
+ */
+ def int_hpvm_tensor_convolution : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
+ llvm_ptr_ty,
+ llvm_i32_ty,
+ llvm_i32_ty,
+ llvm_i32_ty,
+ llvm_i32_ty], []>;
+
+ /* Tensor group convolution intrinsic
+ * i8* llvm.hpvm.tensor.group.convolution(i8*, i8*, i32, i32, i32, i32, i32, i32);
+ */
+ def int_hpvm_tensor_group_convolution : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
+ llvm_ptr_ty,
+ llvm_i32_ty,
+ llvm_i32_ty,
+ llvm_i32_ty,
+ llvm_i32_ty,
+ llvm_i32_ty,
+ llvm_i32_ty], []>;
+
+ /* Tensor BatchNorm intrinsic
+ * i8* llvm.hpvm.tensor.batchnorm(i8*, i8*, i8*, i8*, i8*, double);
+ */
+ def int_hpvm_tensor_batchnorm : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
+ llvm_ptr_ty,
+ llvm_ptr_ty,
+ llvm_ptr_ty,
+ llvm_ptr_ty,
+ llvm_double_ty], []>;
+
+
+ /* Tensor pool intrinsics: max, min, average
+ * i8* llvm.hpvm.tensor.pool.max(i8*, i32, i32, i32, i32, i32, i32);
+ * i8* llvm.hpvm.tensor.pool.min(i8*, i32, i32, i32, i32, i32, i32);
+ * i8* llvm.hpvm.tensor.pool.average(i8*, i32, i32, i32, i32, i32, i32);
+ */
+ def int_hpvm_tensor_pool_max : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
+ llvm_i32_ty,
+ llvm_i32_ty,
+ llvm_i32_ty,
+ llvm_i32_ty,
+ llvm_i32_ty,
+ llvm_i32_ty], []>;
+ def int_hpvm_tensor_pool_min : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
+ llvm_i32_ty,
+ llvm_i32_ty,
+ llvm_i32_ty,
+ llvm_i32_ty,
+ llvm_i32_ty,
+ llvm_i32_ty], []>;
+ def int_hpvm_tensor_pool_mean : Intrinsic<[llvm_ptr_ty], [llvm_ptr_ty,
+ llvm_i32_ty,
+ llvm_i32_ty,
+ llvm_i32_ty,
+ llvm_i32_ty,
+ llvm_i32_ty,
+ llvm_i32_ty], []>;
+
+ def int_hpvm_node_id : Intrinsic<[llvm_ptr_ty], [llvm_i32_ty], []>;
}