[Feature](bangc-ops): add transform binary oprator. (#839)

Co-authored-by: zhangtingyu <[email protected]>

bangc-ops/kernels/kernel_wrapper/wrapper.h

@@ -276,6 +276,16 @@
       const mluOpTensorDescriptor_t, const void *,               \
       const mluOpTensorDescriptor_t, void *diff_x
 
+#define TRANSFORM_PARAM_TYPE                                                  \
+    mluOpHandle_t,                                                            \
+    const mluOpPointerMode_t,                                                 \
+    const void *,                                                             \
+    const mluOpTensorDescriptor_t,                                            \
+    const void *,                                                             \
+    const void *,                                                             \
+    const mluOpTensorDescriptor_t,                                            \
+    void *
+
 #define STRIDEDSLICE_PARAM_TYPE                               \
   mluOpHandle_t, const mluOpTensorDescriptor_t, const void *, \
       const int *, const int *, const int *,      \
@@ -306,7 +316,6 @@ KERNEL_REGISTER(RoiAlignBackward, ROIALIGNBACKWARD_PARAM_TYPE);
 KERNEL_REGISTER(RoiAlignBackwardV2, ROIALIGNBACKWARD_V2_PARAM_TYPE);
 KERNEL_REGISTER(RoiPoolingForward, ROIPOOLINGFORWARD_PARAM_TYPE);
 KERNEL_REGISTER(RoiPoolingBackward, ROIPOOLINGBACKWARD_PARAM_TYPE);
-KERNEL_REGISTER(transform, TRANSFORM_PARAM_TYPE);
 KERNEL_REGISTER(SyncBatchNormStats, SYNCBATCHNORMSTATS_PARAM_TYPE);
 KERNEL_REGISTER(SyncBatchNormStatsV2, SYNCBATCHNORMSTATS_V2_PARAM_TYPE);
 KERNEL_REGISTER(SyncBatchNormGatherStatsWithCounts,
@@ -320,5 +329,6 @@ KERNEL_REGISTER(SyncBatchNormBackwardElemt,
         SYNCBATCHNORMBACKWARDELEMT_PARAM_TYPE);
 KERNEL_REGISTER(SyncBatchNormBackwardElemtV2,
         SYNCBATCHNORMBACKWARDELEMT_V2_PARAM_TYPE);
+KERNEL_REGISTER(transform, TRANSFORM_PARAM_TYPE);
 KERNEL_REGISTER(StridedSlice, STRIDEDSLICE_PARAM_TYPE);
 #endif  // KERNELS_KERNEL_WRAPPER_WRAPPER_H

bangc-ops/kernels/transform/transform.cpp

@@ -0,0 +1,40 @@
+/*************************************************************************
+ * Copyright (C) [2023] by Cambricon, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sublicense, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included
+ * in all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
+ * IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
+ * CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+ * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+ * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *************************************************************************/
+#include "kernels/kernel_wrapper/wrapper.h"
+
+mluOpStatus_t MLUOP_WIN_API mluOpTransform(
+    mluOpHandle_t handle,
+    const mluOpPointerMode_t pointer_mode,
+    const void *alpha,
+    const mluOpTensorDescriptor_t input_desc,
+    const void *input,
+    const void *beta,
+    const mluOpTensorDescriptor_t output_desc,
+    void *output) {
+  transformWrapper wrapper;
+  mluOpStatus_t ret = wrapper.invoke(
+      handle, pointer_mode, alpha, input_desc, input,
+      beta, output_desc, output);
+  return ret;
+}
+

bangc-ops/mlu_op.h

@@ -15096,6 +15096,101 @@ mluOpSyncBatchNormBackwardElemtV2(mluOpHandle_t handle,
                                   const mluOpTensorDescriptor_t diff_x_desc,
                                   void *diff_x);
 
+// Group:Transform
+/*!
+ * @brief Linearly transforms an input tensor using the following formula:
+ *
+ * output = alpha * input + beta
+ *
+ *   Parameters \b input and \b output represent tensor input and output,
+ *  and \b alpha and \b beta represent scale factors.
+ *  \b mluOpTransform supports scale pointer
+ *  from both host and device for \b alpha and \b beta parameter.
+ * @param[in] handle
+ *   Handle to a Cambricon MLUOP context used to manage MLU
+ *  devices and queues.
+ *  For detailed information, please refer to ::mluOpHandle_t.
+ * @param[in] input_desc
+ *   Descriptor of the input tensor \b.
+ *  Refer to ::mluOpTensorDescriptor_t for further detail.
+ * @param[in] input
+ *   Device pointer to the MLU memory that stores the input tensor.
+ * @param[in] pointer_mode
+ *   An enum value that indicates which mode scalar values \b alpha
+ * and \b beta would be passed by on the host or device.
+ * Mode type is defined in ::mluOpPointerMode_t.
+ * @param[in] alpha
+ *   Pointer to scaling factor of tensor input.
+ *   If \b pointer_mode is \b MLUOP_POINTER_MODE_DEVICE, \b alpha
+ *  should be a device pointer.
+ *   If \b pointer_mode is \b MLUOP_POINTER_MODE_HOST, \b alpha
+ *  should be a host pointer.
+ * @param[in] beta
+ *   Pointer to scaling factor of tensor input.
+ *   If \b pointer_mode is \b MLUOP_POINTER_MODE_DEVICE, \b beta
+ *  should be a device pointer.
+ *   If \b pointer_mode is \b MLUOP_POINTER_MODE_HOST, \b beta
+ *  should be a host pointer.
+ * @param[in] output_desc
+ *   Descriptor of the output tensor \b output. For detailed
+ *  information, please refer to ::mluOpTensorDescriptor_t.
+ * @param[out] output
+ *   Descriptor of the output tensor. For detailed information, see
+ *   ::mluOpTensorDescriptor_t.
+ *
+ * @par Return
+ * - ::MLUOP_STATUS_SUCCESS, ::MLUOP_STATUS_BAD_PARAM
+ *
+ * @par Formula
+ * - See "Transform Operator" section in "Cambricon MLUOP User Guide" for
+ * details.
+ *
+ * @par Data Type
+ * - The combinations of the data type for input tensor \b input
+ *   and output tensor
+ *   \b output must be half-half, float-float or int32-int32.
+ * - \b alpha and \b beta: If data type of tensors is float or half, the data
+ *   type of \b alpha and \b beta should be float pointer.
+ *   If data type of tensors is
+ *   int32, the data type of \b alpha and \b beta should be int pointer.
+ *
+ * @par Scale Limitation
+ * - Tensor descriptors of input and output tensors must be the same.
+ * - Dimension number should not exceeds \p MLUOP_DIM_MAX.
+ *
+ * @note
+ * - None.
+ *
+ * @par Requirements
+ * - None.
+ *
+ * @par Example
+     @verbatim
+       Input tensor   :   [[1, 2, 3],
+                           [4, 5, 6],
+                           [7, 8, 9]]
+
+       alpha          :   2
+
+       beta           :   1
+
+       Output tensor  :   [[3,  5,  7],
+                           [9,  11, 13],
+                           [15, 17, 19]]
+     @endverbatim
+ *
+ */
+
+mluOpStatus_t MLUOP_WIN_API
+mluOpTransform(mluOpHandle_t handle,
+               const mluOpPointerMode_t pointer_mode,
+               const void *alpha,
+               const mluOpTensorDescriptor_t input_desc,
+               const void *input,
+               const void *beta,
+               const mluOpTensorDescriptor_t output_desc,
+               void *output);
+
 // Group:StridedSlice
 /*!
  * @brief Extracts a slice of size ``(end - begin) / stride`` from the given \p input tensor.

bangc-ops/test/mlu_op_gtest/pb_gtest/src/zoo/transform/test_case/case_0.prototxt

@@ -0,0 +1,49 @@
+op_name: "transform"
+op_type: "TRANSFORM"
+input {
+  id: "input1"
+  shape: {
+    dims: 1
+    dims: 1
+    dims: 1
+    dims: 1
+    dims: 1
+    dims: 1
+    dims: 1000
+    dims: 2000
+  }
+  layout: LAYOUT_ARRAY
+  dtype: DTYPE_FLOAT
+  random_data: {
+    seed: 23
+    upper_bound: 100
+    lower_bound: 0
+    distribution: UNIFORM
+  }
+}
+output {
+  id: "output"
+  shape: {
+    dims: 1
+    dims: 1
+    dims: 1
+    dims: 1
+    dims: 1
+    dims: 1
+    dims: 1000
+    dims: 2000
+  }
+  layout: LAYOUT_ARRAY
+  dtype: DTYPE_FLOAT
+}
+transform_param: {
+  alpha: 1
+  beta: 1
+}
+test_param: {
+  error_func: DIFF1
+  error_func: DIFF2
+  error_threshold: 0.003
+  error_threshold: 0.003
+  baseline_device: CPU
+}

bangc-ops/test/mlu_op_gtest/pb_gtest/src/zoo/transform/test_case/case_1.prototxt

@@ -0,0 +1,41 @@
+op_name: "transform"
+op_type: "TRANSFORM"
+input {
+  id: "input1"
+  shape: {
+    dims: 13
+    dims: 17
+    dims: 100
+    dims: 70
+  }
+  layout: LAYOUT_ARRAY
+  dtype: DTYPE_FLOAT
+  random_data: {
+    seed: 23
+    upper_bound: 100
+    lower_bound: 0
+    distribution: UNIFORM
+  }
+}
+output {
+  id: "output"
+  shape: {
+    dims: 13
+    dims: 17
+    dims: 100
+    dims: 70
+  }
+  layout: LAYOUT_ARRAY
+  dtype: DTYPE_FLOAT
+}
+transform_param: {
+  alpha: 1
+  beta: 1
+}
+test_param: {
+  error_func: DIFF1
+  error_func: DIFF2
+  error_threshold: 0.003
+  error_threshold: 0.003
+  baseline_device: CPU
+}

bangc-ops/test/mlu_op_gtest/pb_gtest/src/zoo/transform/test_case/complex.prototxt

@@ -0,0 +1,37 @@
+op_name: "transform"
+op_type: "TRANSFORM"
+input {
+  id: "input1"
+  shape: {
+    dims: 1000
+  }
+  layout: LAYOUT_ARRAY
+  dtype: DTYPE_COMPLEX_FLOAT
+  random_data: {
+    seed: 23
+    upper_bound: 100
+    lower_bound: 0
+    distribution: UNIFORM
+  }
+}
+output {
+  id: "output"
+  shape: {
+    dims: 1000
+  }
+  layout: LAYOUT_ARRAY
+  dtype: DTYPE_COMPLEX_FLOAT
+}
+transform_param: {
+  alpha: 1.1
+  alpha_imag:1.1
+  beta: 1.2
+  beta_imag:1.1
+}
+test_param: {
+  error_func: DIFF1
+  error_func: DIFF2
+  error_threshold: 0.003
+  error_threshold: 0.003
+  baseline_device: CPU
+}

bangc-ops/test/mlu_op_gtest/pb_gtest/src/zoo/transform/test_case/half_0.prototxt

@@ -0,0 +1,37 @@
+op_name: "transform"
+op_type: "TRANSFORM"
+input {
+  id: "input1"
+  shape: {
+    dims: 1078
+    dims: 201
+  }
+  layout: LAYOUT_ARRAY
+  dtype: DTYPE_HALF
+  random_data: {
+    seed: 23
+    upper_bound: 100
+    lower_bound: 0
+    distribution: UNIFORM
+  }
+}
+output {
+  id: "output"
+  shape: {
+    dims: 1078
+    dims: 201
+  }
+  layout: LAYOUT_ARRAY
+  dtype: DTYPE_HALF
+}
+transform_param: {
+  alpha: 1.5
+  beta: 0.8
+}
+test_param: {
+  error_func: DIFF1
+  error_func: DIFF2
+  error_threshold: 0.003
+  error_threshold: 0.003
+  baseline_device: CPU
+}

bangc-ops/test/mlu_op_gtest/pb_gtest/src/zoo/transform/test_case/int32_0.prototxt

@@ -0,0 +1,37 @@
+op_name: "transform"
+op_type: "TRANSFORM"
+input {
+  id: "input1"
+  shape: {
+    dims: 1078
+    dims: 201
+  }
+  layout: LAYOUT_ARRAY
+  dtype: DTYPE_INT32
+  random_data: {
+    seed: 23
+    upper_bound: 100
+    lower_bound: 0
+    distribution: UNIFORM
+  }
+}
+output {
+  id: "output"
+  shape: {
+    dims: 1078
+    dims: 201
+  }
+  layout: LAYOUT_ARRAY
+  dtype: DTYPE_INT32
+}
+transform_param: {
+  alpha: 1.5
+  beta: 0.8
+}
+test_param: {
+  error_func: DIFF1
+  error_func: DIFF2
+  error_threshold: 0.003
+  error_threshold: 0.003
+  baseline_device: CPU
+}