From 12addc64a6a6787c224e4d96057a797ca7de1535 Mon Sep 17 00:00:00 2001
From: lcskrishna <lollachaitanya@gmail.com>
Date: Wed, 5 Dec 2018 23:52:42 -0800
Subject: [PATCH] Fixed MIOpen RNN Segfault issue and enabled RNN test (#14810)

Summary:
This pull request contains changes for:
1. Added MIOpen RNN API miopenGetRNNLayerBiasSize and miopenGetRNNLayerParamSize.
2. Fixed usage of API miopenGetRNNLayerParam.
3. Modifying the RNN test to run using MIOpen engine.

Differential Revision: D13355699

Pulled By: bddppq

fbshipit-source-id: 6f750657f8049c5446eca893880b397804120b69
---
 caffe2/operators/rnn/hip/recurrent_op_miopen.hip   | 42 ++++++++++++++++------
 .../python/operator_test/cudnn_recurrent_test.py   |  4 +--
 2 files changed, 34 insertions(+), 12 deletions(-)

diff --git a/caffe2/operators/rnn/hip/recurrent_op_miopen.hip b/caffe2/operators/rnn/hip/recurrent_op_miopen.hip
index 34c96a2..5900a08 100644
--- a/caffe2/operators/rnn/hip/recurrent_op_miopen.hip
+++ b/caffe2/operators/rnn/hip/recurrent_op_miopen.hip
@@ -379,8 +379,19 @@ bool RecurrentParamAccessOp<T, mode>::RunOnDevice() {
 
     miopenTensorDescriptor_t biasDesc;
     MIOPEN_ENFORCE(miopenCreateTensorDescriptor(&biasDesc));
-    void* bias;
 
+    size_t bias_size = 0;
+    MIOPEN_ENFORCE(miopenGetRNNLayerBiasSize(
+        miopen_wrapper_.inline_miopen_handle(),
+        rnnDesc_,
+        layer,
+        param_id,
+        &bias_size)); 
+
+    void* bias;
+    miopen_wrapper_.with_miopen_state(0, [&](MIOPENState* state) {
+        bias = state->workspace().get(bias_size);
+    });
     MIOPEN_ENFORCE(miopenGetRNNLayerBias(
         miopen_wrapper_.inline_miopen_handle(),
         rnnDesc_,
@@ -390,10 +401,9 @@ bool RecurrentParamAccessOp<T, mode>::RunOnDevice() {
         Input(1).template data<T>(),
         param_id, 
         biasDesc,
-        &bias));
-    int numBiasDims;
-    std::vector<int> biasDims;
-    std::vector<int> strideDims;
+        bias));
+    std::array<int,3> biasDims {1,1,1};
+    std::array<int,3> strideDims {1,1,1};
     miopenDataType_t dt;
 
     MIOPEN_ENFORCE(miopenGetTensorDescriptor(
@@ -419,7 +429,20 @@ bool RecurrentParamAccessOp<T, mode>::RunOnDevice() {
         weight_constants[param_type] + 4 * (input_type == "recurrent");
     miopenTensorDescriptor_t matrixParamDesc;
     MIOPEN_ENFORCE(miopenCreateTensorDescriptor(&matrixParamDesc));
+
+    size_t param_size = 0;
+    MIOPEN_ENFORCE(miopenGetRNNLayerParamSize(
+        miopen_wrapper_.inline_miopen_handle(),
+        rnnDesc_,
+        layer,
+        xDesc_->descs()[0],
+        param_id,
+        &param_size));
+    
     void* pmatrix;
+    miopen_wrapper_.with_miopen_state(0, [&](MIOPENState* state) {
+        pmatrix = state->workspace().get(param_size);
+    });
     MIOPEN_ENFORCE(miopenGetRNNLayerParam(
         miopen_wrapper_.inline_miopen_handle(),
         rnnDesc_,
@@ -429,15 +452,14 @@ bool RecurrentParamAccessOp<T, mode>::RunOnDevice() {
         Input(1).template data<T>(),
         param_id, 
         matrixParamDesc,
-        &pmatrix));
-    int numDims;
-    std::vector<int> matDims;
-    std::vector<int> strideDims;
+        pmatrix));
+    std::array<int,3> matDims {1,1,1};
+    std::array<int,3> strideDims {1,1,1};
     miopenDataType_t dt;
 
     MIOPEN_ENFORCE(miopenGetTensorDescriptor(
         matrixParamDesc, &dt, matDims.data(), strideDims.data()));
-    CAFFE_ENFORCE_EQ(numDims, 3);
+    CAFFE_ENFORCE_EQ(matDims.size(), 3);
     if (mode == SET_PARAM) {
       CAFFE_ENFORCE_EQ(matDims[0] * matDims[1] * matDims[2], Input(2).size());
       context_.template CopySameDevice<T>(
diff --git a/caffe2/python/operator_test/cudnn_recurrent_test.py b/caffe2/python/operator_test/cudnn_recurrent_test.py
index 40fe88b..46162f6 100644
--- a/caffe2/python/operator_test/cudnn_recurrent_test.py
+++ b/caffe2/python/operator_test/cudnn_recurrent_test.py
@@ -11,11 +11,11 @@ import numpy as np
 import unittest
 
 
-@unittest.skipIf(not workspace.has_gpu_support, "No gpu support.")
+@unittest.skipIf((not workspace.has_gpu_support) and (not workspace.has_hip_support), "No gpu support.")
 class TestLSTMs(unittest.TestCase):
 
     def testEqualToCudnn(self):
-        with core.DeviceScope(core.DeviceOption(caffe2_pb2.CUDA)):
+        with core.DeviceScope(core.DeviceOption(workspace.GpuDeviceType)):
             T = 8
             batch_size = 4
             input_dim = 8
-- 
2.7.4