From d6a9dee3fe584921d24a264176468898d38a6202 Mon Sep 17 00:00:00 2001
From: Matthew Jackson <matthew.jackson@arm.com>
Date: Mon, 22 Jul 2019 13:53:24 +0100
Subject: IVGCVSW-3368 Add reference support for depthwise multiplier > 3

* Remove multiplier check in TfLite parser
* Add reference unit test for depthwise multipler of 64, as in DeepSpeaker

Signed-off-by: Matthew Jackson <matthew.jackson@arm.com>
Change-Id: I787339ab4c4d269333985353d191202d070906ba
---
 src/armnnTfLiteParser/TfLiteParser.cpp          |  3 +--
 src/backends/backendsCommon/test/LayerTests.cpp | 32 +++++++++++++++++++++++++
 src/backends/backendsCommon/test/LayerTests.hpp |  4 ++++
 src/backends/reference/test/RefLayerTests.cpp   |  2 ++
 4 files changed, 39 insertions(+), 2 deletions(-)

diff --git a/src/armnnTfLiteParser/TfLiteParser.cpp b/src/armnnTfLiteParser/TfLiteParser.cpp
index 04fa6b1947..f345d4a6e1 100644
--- a/src/armnnTfLiteParser/TfLiteParser.cpp
+++ b/src/armnnTfLiteParser/TfLiteParser.cpp
@@ -791,8 +791,7 @@ void TfLiteParser::ParseDepthwiseConv2D(size_t subgraphIndex, size_t operatorInd
     desc.m_StrideX = CHECKED_NON_NEGATIVE(options->stride_w);
     desc.m_StrideY = CHECKED_NON_NEGATIVE(options->stride_h);
     desc.m_DataLayout = armnn::DataLayout::NHWC;
-    // ACL only supports a depth (channel) multiplier of {1,2,3}, it is not currently stored in the descriptor
-    CHECK_VALID_SIZE(CHECKED_NON_NEGATIVE(options->depth_multiplier), 1,2,3 );
+    CHECKED_NON_NEGATIVE(options->depth_multiplier);
 
     auto inputs = GetInputs(m_Model, subgraphIndex, operatorIndex);
     CHECK_VALID_SIZE(inputs.size(), 2, 3);
diff --git a/src/backends/backendsCommon/test/LayerTests.cpp b/src/backends/backendsCommon/test/LayerTests.cpp
index 03ab37dd46..d6e0e879fd 100644
--- a/src/backends/backendsCommon/test/LayerTests.cpp
+++ b/src/backends/backendsCommon/test/LayerTests.cpp
@@ -1498,6 +1498,38 @@ LayerTestResult<float, 4> DepthwiseConvolution2dDepthMul1Test(
         workloadFactory, memoryManager, 0.0f, 0, biasEnabled, layout);
 }
 
+LayerTestResult<float, 4> DepthwiseConvolution2dDepthMul64Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    armnn::TensorInfo inputTensorInfo({ 1, 1, 2, 2 }, armnn::DataType::Float32);
+    auto input = MakeTensor<float, 4>(inputTensorInfo, { 1.f, 2.f, 3.f, 4.f });
+
+    std::vector<float> kernelData;
+    std::vector<float> singleDepthKernel{ 1.f, -1.f, -1.f, 1.f };
+    for (unsigned int i = 0; i < 64; ++i)
+    {
+        kernelData.insert(kernelData.end(), singleDepthKernel.begin(), singleDepthKernel.end());
+    }
+    armnn::TensorInfo kernelTensorInfo({ 64, 1, 2, 2 }, armnn::DataType::Float32);
+    auto kernel = MakeTensor<float, 4>(kernelTensorInfo, kernelData);
+
+    std::vector<float> expectedOutputData(64, 0.f);
+    armnn::TensorInfo outputTensorInfo({ 1, 64, 1, 1 }, armnn::DataType::Float32);
+    auto expectedOutput = MakeTensor<float, 4>(outputTensorInfo, expectedOutputData);
+
+    return DepthwiseConvolution2dTestImpl<armnn::DataType::Float32, armnn::DataType::Float32>(
+            workloadFactory,
+            memoryManager,
+            input,
+            kernel,
+            boost::multi_array<float, 1>(),
+            expectedOutput,
+            0.f,
+            0,
+            armnn::DataLayout::NCHW);
+}
+
 LayerTestResult<float, 4> DepthwiseConvolution2dAsymmetricTest(
     armnn::IWorkloadFactory& workloadFactory,
     const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
diff --git a/src/backends/backendsCommon/test/LayerTests.hpp b/src/backends/backendsCommon/test/LayerTests.hpp
index 3db826f379..d99e3b4f3a 100644
--- a/src/backends/backendsCommon/test/LayerTests.hpp
+++ b/src/backends/backendsCommon/test/LayerTests.hpp
@@ -159,6 +159,10 @@ LayerTestResult<float, 4> DepthwiseConvolution2dDepthMul1Test(
     bool biasEnabled,
     const armnn::DataLayout layout);
 
+LayerTestResult<float, 4> DepthwiseConvolution2dDepthMul64Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
 LayerTestResult<float, 4> DepthwiseConvolution2dAsymmetricTest(
     armnn::IWorkloadFactory& workloadFactory,
     const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
diff --git a/src/backends/reference/test/RefLayerTests.cpp b/src/backends/reference/test/RefLayerTests.cpp
index 4f46d27883..8af42eaba2 100644
--- a/src/backends/reference/test/RefLayerTests.cpp
+++ b/src/backends/reference/test/RefLayerTests.cpp
@@ -267,6 +267,8 @@ ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dAsymmetricNhwc,
 ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dAsymmetricNhwc,
                      DepthwiseConvolution2dAsymmetricTest, false, armnn::DataLayout::NHWC)
 
+ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dDepthMul64, DepthwiseConvolution2dDepthMul64Test);
+
 // Pooling
 //MaxPooling
 ARMNN_AUTO_TEST_CASE(SimpleMaxPooling2dSize2x2Stride2x2, SimpleMaxPooling2dSize2x2Stride2x2Test, false)
-- 
cgit v1.2.1