#10071: Merge abs_bw to TTNN

tenstorrent · Jul 11, 2024 · 3f24708 · 3f24708
1 parent 8bd3e5d
commit 3f24708
Show file tree

Hide file tree

Showing 11 changed files with 40 additions and 30 deletions.
diff --git a/docs/source/ttnn/ttnn/api.rst b/docs/source/ttnn/ttnn/api.rst
@@ -197,6 +197,7 @@ Pointwise Unary
    ttnn/round_bw
    ttnn/log_bw
    ttnn/relu6_bw
+   ttnn/abs_bw
 
 Pointwise Binary
 ================

diff --git a/docs/source/ttnn/ttnn/dependencies/tt_lib.rst b/docs/source/ttnn/ttnn/dependencies/tt_lib.rst
@@ -816,8 +816,6 @@ Backward Operations
 
 .. autofunction:: tt_lib.tensor.unary_sub_bw
 
-.. autofunction:: tt_lib.tensor.abs_bw
-
 .. autofunction:: tt_lib.tensor.complex_abs_bw
 
 .. autofunction:: tt_lib.tensor.lt_bw

diff --git a/docs/source/ttnn/ttnn/ttnn/abs_bw.rst b/docs/source/ttnn/ttnn/ttnn/abs_bw.rst
@@ -0,0 +1,6 @@
+.. _ttnn.abs_bw:
+
+  ttnn.abs_bw
+  ############
+
+  .. autofunction:: ttnn.abs_bw
diff --git a/tests/tt_eager/python_api_testing/unit_testing/backward_ops/test_backward_abs.py b/tests/tt_eager/python_api_testing/unit_testing/backward_ops/test_backward_abs.py
@@ -4,7 +4,7 @@
 
 import torch
 import pytest
-import tt_lib
+import ttnn
 from tests.tt_eager.python_api_testing.unit_testing.backward_ops.utility_funcs import data_gen_with_range, compare_pcc
 
 
@@ -22,7 +22,7 @@ def test_bw_abs(input_shapes, device):
 
     pyt_y = torch.abs(in_data)
 
-    tt_output_tensor_on_device = tt_lib.tensor.abs_bw(grad_tensor, input_tensor)
+    tt_output_tensor_on_device = ttnn.abs_bw(grad_tensor, input_tensor)
 
     in_data.retain_grad()
 

diff --git a/tt_eager/tt_dnn/op_library/backward/backward_ops.cpp b/tt_eager/tt_dnn/op_library/backward/backward_ops.cpp
@@ -280,14 +280,24 @@ std::vector<Tensor> ne_bw(const Tensor& grad, const MemoryConfig& output_mem_con
     return operation::decorate_as_composite(__func__, _ne_bw)(grad, output_mem_config);
 }
 
-std::vector<Tensor> _abs_bw(const Tensor& grad, const Tensor& input, const MemoryConfig& output_mem_config) {
+std::vector<Tensor> _rsqrt_bw(const Tensor& grad, const Tensor& input, const MemoryConfig& output_mem_config) {
     std::vector<Tensor> grad_tensor;
-    Tensor result = ttnn::multiply(grad, ttnn::sign(input, output_mem_config), std::nullopt, output_mem_config);
+    Tensor rsqrt_result = power(rsqrt(input, true, output_mem_config), 3, output_mem_config);
+    Tensor result = mul_unary(ttnn::multiply(grad, rsqrt_result, std::nullopt, output_mem_config), -0.5, output_mem_config);
+    float t_inf = std::numeric_limits<float>::infinity();
+    result = where(eqz(input, output_mem_config), t_inf, result, output_mem_config);
+    float t_nan = std::nanf("");
+    result = where(ltz(input, output_mem_config), t_nan, result, output_mem_config);
+    result = where(
+        ttnn::logical_and(eqz(input, output_mem_config), eqz(grad, output_mem_config), std::nullopt, output_mem_config),
+        t_nan,
+        result,
+        output_mem_config);
     grad_tensor.emplace_back(result);
     return grad_tensor;
 }
-std::vector<Tensor> abs_bw(const Tensor& grad, const Tensor& input, const MemoryConfig& output_mem_config) {
-    return operation::decorate_as_composite(__func__, _abs_bw)(grad, input, output_mem_config);
+std::vector<Tensor> rsqrt_bw(const Tensor& grad, const Tensor& input, const MemoryConfig& output_mem_config) {
+    return operation::decorate_as_composite(__func__, _rsqrt_bw)(grad, input, output_mem_config);
 }
 
 // bw(expm1) = grad * expm1(input) + 1

diff --git a/tt_eager/tt_dnn/op_library/backward/backward_ops.hpp b/tt_eager/tt_dnn/op_library/backward/backward_ops.hpp
@@ -111,11 +111,6 @@ std::vector<Tensor> binary_le_bw(
     const Tensor& input,
     const MemoryConfig& output_mem_config = operation::DEFAULT_OUTPUT_MEMORY_CONFIG);
 
-std::vector<Tensor> abs_bw(
-    const Tensor& grad,
-    const Tensor& input,
-    const MemoryConfig& output_mem_config = operation::DEFAULT_OUTPUT_MEMORY_CONFIG);
-
 std::vector<Tensor> complex_abs_bw(
     const Tensor& grad,
     const Tensor& input,

diff --git a/tt_eager/tt_lib/csrc/tt_lib_bindings_tensor_backward_ops.cpp b/tt_eager/tt_lib/csrc/tt_lib_bindings_tensor_backward_ops.cpp
@@ -328,22 +328,6 @@ namespace tt::tt_metal::detail{
                 "output_mem_config", "Layout of tensor in TT Accelerator device memory banks", "MemoryConfig", "Default is interleaved in DRAM", "No"
         )doc");
 
-    m_tensor.def("abs_bw", &tt::tt_metal::abs_bw,
-            py::arg("grad").noconvert(), py::arg("input").noconvert(), py::arg("output_mem_config").noconvert() = operation::DEFAULT_OUTPUT_MEMORY_CONFIG, R"doc(
-            Performs backward operations for abs of ``input`` tensors with given ``grad``.
-
-            Input tensors must have BFLOAT16 data type.
-
-            Output tensor will have BFLOAT16 data type.
-
-            .. csv-table::
-                :header: "Argument", "Description", "Data type", "Valid range", "Required"
-
-                "grad", "Gradient tensor", "Tensor", "Tensor of shape [W, Z, Y, X]", "Yes"
-                "input", "Tensor add is applied to", "Tensor", "Tensor of shape [W, Z, Y, X]", "Yes"
-                "output_mem_config", "Layout of tensor in TT Accelerator device memory banks", "MemoryConfig", "Default is interleaved in DRAM", "No"
-        )doc");
-
     m_tensor.def("complex_abs_bw", py::overload_cast<const Tensor&, const Tensor&, const MemoryConfig&>(&complex_abs_bw),
             py::arg("grad").noconvert(), py::arg("input").noconvert(), py::arg("output_mem_config").noconvert() = operation::DEFAULT_OUTPUT_MEMORY_CONFIG, R"doc(
             Performs backward operations for abs of complex ``input`` tensor with given ``grad``.

diff --git a/ttnn/cpp/ttnn/operations/eltwise/unary_backward/device/unary_backward_op.cpp b/ttnn/cpp/ttnn/operations/eltwise/unary_backward/device/unary_backward_op.cpp
@@ -411,6 +411,13 @@ std::vector<Tensor> _relu6_bw(const Tensor& grad, const Tensor& input, const Mem
     return grad_tensor;
 }
 
+std::vector<Tensor> _abs_bw(const Tensor& grad, const Tensor& input, const MemoryConfig& output_mem_config) {
+    std::vector<Tensor> grad_tensor;
+    Tensor result = ttnn::multiply(grad, ttnn::sign(input, output_mem_config), std::nullopt, output_mem_config);
+    grad_tensor.emplace_back(result);
+    return grad_tensor;
+}
+
 std::function<std::vector<ttnn::Tensor>(const Tensor&, const Tensor&, const MemoryConfig&)> UnaryBackwardFunction::get_function_type1(UnaryBackwardOpType OpType){
     switch (OpType) {
         case UnaryBackwardOpType::ASSIGN_BW:
@@ -449,6 +456,8 @@ std::function<std::vector<ttnn::Tensor>(const Tensor&, const Tensor&, const Memo
             return _log_bw;
         case UnaryBackwardOpType::RELU6_BW:
             return _relu6_bw;
+        case UnaryBackwardOpType::ABS_BW:
+            return _abs_bw;
         default:
             TT_ASSERT(false && "Undefined op type");
             return 0;

diff --git a/ttnn/cpp/ttnn/operations/eltwise/unary_backward/device/unary_backward_op.hpp b/ttnn/cpp/ttnn/operations/eltwise/unary_backward/device/unary_backward_op.hpp
@@ -43,7 +43,8 @@ enum class UnaryBackwardOpType {
     FLOOR_BW,
     ROUND_BW,
     LOG_BW,
-    RELU6_BW
+    RELU6_BW,
+    ABS_BW,
 };
 
 struct UnaryBackwardFunction{

diff --git a/ttnn/cpp/ttnn/operations/eltwise/unary_backward/unary_backward.hpp b/ttnn/cpp/ttnn/operations/eltwise/unary_backward/unary_backward.hpp
@@ -96,6 +96,7 @@ constexpr auto floor_bw = ttnn::register_operation<operations::unary_backward::E
 constexpr auto round_bw = ttnn::register_operation<operations::unary_backward::ExecuteUnaryBackward<operations::unary_backward::UnaryBackwardOpType::ROUND_BW>>("ttnn::round_bw");
 constexpr auto log_bw = ttnn::register_operation<operations::unary_backward::ExecuteUnaryBackward<operations::unary_backward::UnaryBackwardOpType::LOG_BW>>("ttnn::log_bw");
 constexpr auto relu6_bw = ttnn::register_operation<operations::unary_backward::ExecuteUnaryBackward<operations::unary_backward::UnaryBackwardOpType::RELU6_BW>>("ttnn::relu6_bw");
+constexpr auto abs_bw = ttnn::register_operation<operations::unary_backward::ExecuteUnaryBackward<operations::unary_backward::UnaryBackwardOpType::ABS_BW>>("ttnn::abs_bw");
 
 
 }  // namespace ttnn
diff --git a/ttnn/cpp/ttnn/operations/eltwise/unary_backward/unary_backward_pybind.hpp b/ttnn/cpp/ttnn/operations/eltwise/unary_backward/unary_backward_pybind.hpp
@@ -318,6 +318,11 @@ void py_module(py::module& module) {
         ttnn::relu6_bw,
         R"doc(Performs backward operations for relu6 on :attr:`input_tensor` with given :attr:`grad_tensor`)doc");
 
+    detail::bind_unary_backward(
+        module,
+        ttnn::abs_bw,
+        R"doc(Performs backward operations for abs on :attr:`input_tensor` with given :attr:`grad_tensor`)doc");
+
 }
 
 }  // namespace binary_backward