Merge branch 'develop' into ar/opt_pixelate

include/rppt_tensor_audio_augmentations.h

@@ -141,9 +141,9 @@ RppStatus rppt_pre_emphasis_filter_host(RppPtr_t srcPtr, RpptDescPtr srcDescPtr,
 /*! \brief Down Mixing augmentation on HOST backend
 * \details Down Mixing augmentation for audio data
 * \param [in] srcPtr source tensor in HOST memory
-* \param [in] srcDescPtr source tensor descriptor (Restrictions - numDims = 3, offsetInBytes >= 0, dataType = F32)
+* \param [in] srcDescPtr source tensor descriptor (Restrictions - numDims = 2 or 3 (for single-channel or multi-channel audio tensor), offsetInBytes >= 0, dataType = F32)
 * \param [out] dstPtr destination tensor in HOST memory
-* \param [in] dstDescPtr destination tensor descriptor (Restrictions - numDims = 3, offsetInBytes >= 0, dataType = F32)
+* \param [in] dstDescPtr destination tensor descriptor (Restrictions - numDims = 2, offsetInBytes >= 0, dataType = F32)
 * \param [in] srcDimsTensor source audio buffer length and number of channels (1D tensor in HOST memory, of size batchSize * 2)
 * \param [in] normalizeWeights bool flag to specify if normalization of weights is needed
 * \param [in] rppHandle RPP HOST handle created with <tt>\ref rppCreateWithBatchSize()</tt>
@@ -153,6 +153,23 @@ RppStatus rppt_pre_emphasis_filter_host(RppPtr_t srcPtr, RpptDescPtr srcDescPtr,
 */
 RppStatus rppt_down_mixing_host(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppPtr_t dstPtr, RpptDescPtr dstDescPtr, Rpp32s *srcDimsTensor, bool normalizeWeights, rppHandle_t rppHandle);
 
+#ifdef GPU_SUPPORT
+/*! \brief Down Mixing augmentation on HIP backend
+* \details Down Mixing augmentation for audio data
+* \param [in] srcPtr source tensor in HIP memory
+* \param [in] srcDescPtr source tensor descriptor (Restrictions - numDims = 2 or 3 (for single-channel or multi-channel audio tensor), offsetInBytes >= 0, dataType = F32)
+* \param [out] dstPtr destination tensor in HIP memory
+* \param [in] dstDescPtr destination tensor descriptor (Restrictions - numDims = 2, offsetInBytes >= 0, dataType = F32)
+* \param [in] srcDimsTensor source audio buffer length and number of channels (1D tensor in HIP/Pinned memory, of size batchSize * 2)
+* \param [in] normalizeWeights bool flag to specify if normalization of weights is needed
+* \param [in] rppHandle RPP HIP handle created with <tt>\ref rppCreateWithStreamAndBatchSize()</tt>
+* \return A <tt> \ref RppStatus</tt> enumeration.
+* \retval RPP_SUCCESS Successful completion.
+* \retval RPP_ERROR* Unsuccessful completion.
+*/
+RppStatus rppt_down_mixing_gpu(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppPtr_t dstPtr, RpptDescPtr dstDescPtr, Rpp32s *srcDimsTensor, bool normalizeWeights, rppHandle_t rppHandle);
+#endif // GPU_SUPPORT
+
 /*! \brief Produces a spectrogram from a 1D audio buffer on HOST backend
  * \details Spectrogram for 1D audio buffer
  * \param [in] srcPtr source tensor in HOST memory

src/modules/hip/hip_tensor_audio_augmentations.hpp

@@ -26,6 +26,7 @@ SOFTWARE.
 #define HIP_TENSOR_AUDIO_AUGMENTATIONS_HPP
 
 #include "kernel/non_silent_region_detection.hpp"
+#include "kernel/down_mixing.hpp"
 #include "kernel/to_decibels.hpp"
 
 #endif // HIP_TENSOR_AUDIO_AUGMENTATIONS_HPP

src/modules/hip/kernel/down_mixing.hpp

@@ -0,0 +1,72 @@
+#include <hip/hip_runtime.h>
+#include "rpp_hip_common.hpp"
+
+__global__ void down_mixing_hip_tensor(float *srcPtr,
+                                       uint srcStride,
+                                       float *dstPtr,
+                                       uint dstStride,
+                                       int2 *srcDimsTensor)
+
+{
+    int id_x = hipBlockIdx_x * hipBlockDim_x + hipThreadIdx_x;
+    int id_z = hipBlockIdx_z * hipBlockDim_z + hipThreadIdx_z;
+    int srcLength = srcDimsTensor[id_z].x;
+    int channels = srcDimsTensor[id_z].y;
+
+    if (id_x >= srcLength)
+        return;
+
+    float outVal = 0.0f;
+    uint srcIdx = id_z * srcStride + id_x * channels;
+    int i = 0;
+    int alignedChannels = (channels / 8) * 8;
+
+    // do 8 pixel load till alignedChannels value
+    if (alignedChannels)
+    {
+        d_float8 outVal_f8;
+        outVal_f8.f4[0] = static_cast<float4>(0.0f);
+        outVal_f8.f4[1] = outVal_f8.f4[0];
+        for(; i < alignedChannels; i += 8, srcIdx += 8)
+        {
+            d_float8 src_f8;
+            rpp_hip_load8_and_unpack_to_float8(srcPtr + srcIdx, &src_f8);
+            rpp_hip_math_add8(&outVal_f8, &src_f8, &outVal_f8);
+        }
+        outVal_f8.f4[0] += outVal_f8.f4[1];
+        outVal += (outVal_f8.f1[0] + outVal_f8.f1[1] + outVal_f8.f1[2] + outVal_f8.f1[3]);
+    }
+    // process remaining channels
+    for(; i < channels; i++, srcIdx++)
+        outVal += srcPtr[srcIdx];
+    outVal *= (1.f / channels);
+
+    uint dstIdx = id_z * dstStride + id_x;
+    dstPtr[dstIdx] = outVal;
+}
+
+RppStatus hip_exec_down_mixing_tensor(Rpp32f *srcPtr,
+                                      RpptDescPtr srcDescPtr,
+                                      Rpp32f *dstPtr,
+                                      RpptDescPtr dstDescPtr,
+                                      Rpp32s *srcDimsTensor,
+                                      bool normalizeWeights,
+                                      rpp::Handle& handle)
+{
+    Rpp32s globalThreads_x = dstDescPtr->strides.nStride;
+    Rpp32s globalThreads_y = 1;
+    Rpp32s globalThreads_z = dstDescPtr->n;
+
+    hipLaunchKernelGGL(down_mixing_hip_tensor,
+                       dim3(ceil((Rpp32f)globalThreads_x/LOCAL_THREADS_X_1DIM), ceil((Rpp32f)globalThreads_y/LOCAL_THREADS_Y_1DIM), ceil((Rpp32f)globalThreads_z/LOCAL_THREADS_Z_1DIM)),
+                       dim3(LOCAL_THREADS_X_1DIM, LOCAL_THREADS_Y_1DIM, LOCAL_THREADS_Z_1DIM),
+                       0,
+                       handle.GetStream(),
+                       srcPtr,
+                       srcDescPtr->strides.nStride,
+                       dstPtr,
+                       dstDescPtr->strides.nStride,
+                       reinterpret_cast<int2 *>(srcDimsTensor));
+
+    return RPP_SUCCESS;
+}

src/modules/rppt_tensor_audio_augmentations.cpp

@@ -362,5 +362,41 @@ RppStatus rppt_to_decibels_gpu(RppPtr_t srcPtr,
 #endif // backend
 }
 
+/******************** down_mixing ********************/
+
+RppStatus rppt_down_mixing_gpu(RppPtr_t srcPtr,
+                               RpptDescPtr srcDescPtr,
+                               RppPtr_t dstPtr,
+                               RpptDescPtr dstDescPtr,
+                               Rpp32s *srcDimsTensor,
+                               bool  normalizeWeights,
+                               rppHandle_t rppHandle)
+{
+#ifdef HIP_COMPILE
+    Rpp32u tensorDims = srcDescPtr->numDims - 1; // exclude batchsize from input dims
+    if (tensorDims != 1 && tensorDims != 2)
+        return RPP_ERROR_INVALID_SRC_DIMS;
+
+    if ((srcDescPtr->dataType == RpptDataType::F32) && (dstDescPtr->dataType == RpptDataType::F32))
+    {
+        hip_exec_down_mixing_tensor(static_cast<Rpp32f*>(srcPtr),
+                                    srcDescPtr,
+                                    static_cast<Rpp32f*>(dstPtr),
+                                    dstDescPtr,
+                                    srcDimsTensor,
+                                    normalizeWeights,
+                                    rpp::deref(rppHandle));
+    }
+    else
+    {
+        return RPP_ERROR_NOT_IMPLEMENTED;
+    }
+
+    return RPP_SUCCESS;
+#elif defined(OCL_COMPILE)
+    return RPP_ERROR_NOT_IMPLEMENTED;
+#endif // backend
+}
+
 #endif // GPU_SUPPORT
 #endif // AUDIO_SUPPORT

utilities/test_suite/HIP/Tensor_audio_hip.cpp

@@ -106,7 +106,10 @@ int main(int argc, char **argv)
     set_audio_descriptor_dims_and_strides(srcDescPtr, batchSize, maxSrcHeight, maxSrcWidth, maxSrcChannels, offsetInBytes);
     int maxDstChannels = maxSrcChannels;
     if(testCase == 3)
+    {
+        srcDescPtr->numDims = 3;
         maxDstChannels = 1;
+    }
     set_audio_descriptor_dims_and_strides(dstDescPtr, batchSize, maxDstHeight, maxDstWidth, maxDstChannels, offsetInBytes);
 
     // set buffer sizes for src/dst
@@ -131,6 +134,11 @@ int main(int argc, char **argv)
     CHECK_RETURN_STATUS(hipHostMalloc(&srcDims, batchSize * sizeof(RpptImagePatch)));
     CHECK_RETURN_STATUS(hipHostMalloc(&dstDims, batchSize * sizeof(RpptImagePatch)));
 
+    // allocate the buffer for srcDimsTensor
+    Rpp32s *srcDimsTensor;
+    if(testCase == 3)
+        CHECK_RETURN_STATUS(hipHostMalloc(&srcDimsTensor, batchSize * 2 * sizeof(Rpp32s)));
+
     Rpp32s *detectedIndex = nullptr, *detectionLength = nullptr;
     if(testCase == 0)
     {
@@ -190,6 +198,24 @@ int main(int argc, char **argv)
 
                     break;
                 }
+                case 3:
+                {
+                    testCaseName = "down_mixing";
+                    bool normalizeWeights = false;
+
+                    for (int i = 0, j = 0; i < batchSize; i++, j += 2)
+                    {
+                        srcDimsTensor[j] = srcLengthTensor[i];
+                        srcDimsTensor[j + 1] = channelsTensor[i];
+                        dstDims[i].height = srcLengthTensor[i];
+                        dstDims[i].width = 1;
+                    }
+
+                    startWallTime = omp_get_wtime();
+                    rppt_down_mixing_gpu(d_inputf32, srcDescPtr, d_outputf32, dstDescPtr, srcDimsTensor, normalizeWeights, handle);
+
+                    break;
+                }
                 default:
                 {
                     missingFuncFlag = 1;
@@ -263,6 +289,8 @@ int main(int argc, char **argv)
     CHECK_RETURN_STATUS(hipHostFree(channelsTensor));
     CHECK_RETURN_STATUS(hipHostFree(srcDims));
     CHECK_RETURN_STATUS(hipHostFree(dstDims));
+    if(testCase == 3)
+        CHECK_RETURN_STATUS(hipHostFree(srcDimsTensor));
     if (detectedIndex != nullptr)
         CHECK_RETURN_STATUS(hipHostFree(detectedIndex));
     if (detectionLength != nullptr)

utilities/test_suite/HIP/runAudioTests.py

@@ -36,7 +36,7 @@
 outFolderPath = os.getcwd()
 buildFolderPath = os.getcwd()
 caseMin = 0
-caseMax = 1
+caseMax = 3
 
 
 # Get a list of log files based on a flag for preserving output
@@ -224,7 +224,7 @@ def rpp_test_suite_parser_and_validator():
 subprocess.call(["make", "-j16"], cwd=".")    # nosec
 
 # List of cases supported
-supportedCaseList = ['0', '1']
+supportedCaseList = ['0', '1', '3']
 if qaMode and batchSize != 3:
     print("QA tests can only run with a batch size of 3.")
     exit(0)

utilities/test_suite/HOST/Tensor_audio_host.cpp

@@ -106,7 +106,10 @@ int main(int argc, char **argv)
     set_audio_descriptor_dims_and_strides(srcDescPtr, batchSize, maxSrcHeight, maxSrcWidth, maxSrcChannels, offsetInBytes);
     int maxDstChannels = maxSrcChannels;
     if(testCase == 3)
+    {
+        srcDescPtr->numDims = 3;
         maxDstChannels = 1;
+    }
     set_audio_descriptor_dims_and_strides(dstDescPtr, batchSize, maxDstHeight, maxDstWidth, maxDstChannels, offsetInBytes);
 
     // create generic descriptor in case of slice