Merge pull request #324 from HazarathKumarM/hk/mel_filter_bank_hip

Mel Filter Bank HIP - Address Review comments, Merge Develop changes and resolve conflicts

docs/sphinx/requirements.in

@@ -1 +1 @@
-rocm-docs-core[api_reference]==1.7.1
+rocm-docs-core[api_reference]==1.7.2

docs/sphinx/requirements.txt

@@ -92,7 +92,7 @@ requests==2.32.3
     # via
     #   pygithub
     #   sphinx
-rocm-docs-core[api-reference]==1.7.1
+rocm-docs-core[api-reference]==1.7.2
     # via -r requirements.in
 smmap==5.0.1
     # via gitdb

include/rppdefs.h

@@ -52,6 +52,8 @@ SOFTWARE.
 #define RPP_MAX_8U      ( 255 )
 /*! \brief RPP maximum dimensions in tensor \ingroup group_rppdefs \page subpage_rppt */
 #define RPPT_MAX_DIMS   ( 5 )
+/*! \brief RPP maximum channels in audio tensor \ingroup group_rppdefs \page subpage_rppt */
+#define RPPT_MAX_AUDIO_CHANNELS   ( 16 )
 
 #define CHECK_RETURN_STATUS(x) do { \
   int retval = (x); \
@@ -62,6 +64,7 @@ SOFTWARE.
 } while (0)
 
 #ifdef HIP_COMPILE
+#include <hip/hip_runtime.h>
 #define RPP_HOST_DEVICE __host__ __device__
 #else
 #define RPP_HOST_DEVICE
@@ -708,7 +711,7 @@ typedef struct GenericFilter
  */
 typedef struct RpptResamplingWindow
 {
-    inline void input_range(Rpp32f x, Rpp32s *loc0, Rpp32s *loc1)
+    inline RPP_HOST_DEVICE void input_range(Rpp32f x, Rpp32s *loc0, Rpp32s *loc1)
     {
         Rpp32s xc = std::ceil(x);
         *loc0 = xc - lobes;
@@ -742,7 +745,7 @@ typedef struct RpptResamplingWindow
     Rpp32f scale = 1, center = 1;
     Rpp32s lobes = 0, coeffs = 0;
     Rpp32s lookupSize = 0;
-    std::vector<Rpp32f> lookup;
+    Rpp32f *lookup = nullptr;
     __m128 pCenter, pScale;
 } RpptResamplingWindow;
 
@@ -806,6 +809,43 @@ struct SlaneyMelScale : public BaseMelScale
     public:
         ~SlaneyMelScale() {};
 };
+inline Rpp32f sinc(Rpp32f x)
+{
+    x *= M_PI;
+    return (std::abs(x) < 1e-5f) ? (1.f - (x * x * 0.16666667)) : std::sin(x) / x;
+}
+
+inline Rpp64f hann(Rpp64f x)
+{
+    return 0.5 * (1 + std::cos(x * M_PI));
+}
+
+// initialization function used for filling the values in Resampling window (RpptResamplingWindow)
+// using the coeffs and lobes value this function generates a LUT (look up table) which is further used in Resample audio augmentation
+inline void windowed_sinc(RpptResamplingWindow &window, Rpp32s coeffs, Rpp32s lobes)
+{
+    Rpp32f scale = 2.0f * lobes / (coeffs - 1);
+    Rpp32f scale_envelope = 2.0f / coeffs;
+    window.coeffs = coeffs;
+    window.lobes = lobes;
+    window.lookupSize = coeffs + 5;
+#ifdef GPU_SUPPORT
+    CHECK_RETURN_STATUS(hipHostMalloc(&(window.lookup), window.lookupSize * sizeof(Rpp32f)));
+#else
+    window.lookup = static_cast<Rpp32f *>(malloc(window.lookupSize * sizeof(Rpp32f)));
+#endif
+    Rpp32s center = (coeffs - 1) * 0.5f;
+    for (int i = 0; i < coeffs; i++) {
+        Rpp32f x = (i - center) * scale;
+        Rpp32f y = (i - center) * scale_envelope;
+        Rpp32f w = sinc(x) * hann(y);
+        window.lookup[i + 1] = w;
+    }
+    window.center = center + 1;
+    window.scale = 1 / scale;
+    window.pCenter = _mm_set1_ps(window.center);
+    window.pScale = _mm_set1_ps(window.scale);
+}
 
 /******************** HOST memory typedefs ********************/
 

include/rppt_tensor_audio_augmentations.h

@@ -49,7 +49,7 @@ extern "C" {
             \n Finds the starting index and length of non silent region in the audio buffer by comparing the
             calculated short-term power with cutoff value passed
  * \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, offsetInBytes >= 0, dataType = F32)
  * \param [in] srcLengthTensor source audio buffer length (1D tensor in HOST memory, of size batchSize)
  * \param [out] detectedIndexTensor beginning index of non silent region (1D tensor in HOST memory, of size batchSize)
  * \param [out] detectionLengthTensor length of non silent region  (1D tensor in HOST memory, of size batchSize)
@@ -70,7 +70,7 @@ RppStatus rppt_non_silent_region_detection_host(RppPtr_t srcPtr, RpptDescPtr src
             \n Finds the starting index and length of non silent region in the audio buffer by comparing the
             calculated short-term power with cutoff value passed
  * \param [in] srcPtr source tensor in HIP memory
- * \param [in] srcDescPtr source tensor descriptor (Restrictions - numDims = 3, offsetInBytes >= 0, dataType = F32)
+ * \param [in] srcDescPtr source tensor descriptor (Restrictions - numDims = 2, offsetInBytes >= 0, dataType = F32)
  * \param [in] srcLengthTensor source audio buffer length (1D tensor in Pinned/HIP memory, of size batchSize)
  * \param [out] detectedIndexTensor beginning index of non silent region (1D tensor in Pinned/HIP memory, of size batchSize)
  * \param [out] detectionLengthTensor length of non silent region  (1D tensor in Pinned/HIP memory, of size batchSize)
@@ -254,9 +254,9 @@ RppStatus rppt_mel_filter_bank_gpu(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppP
 /*! \brief Resample augmentation on HOST backend
 * \details Resample 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 or 3 (for single-channel or multi-channel audio tensor), offsetInBytes >= 0, dataType = F32)
 * \param [in] inRate Input sampling rate (1D tensor in HOST memory, of size batchSize)
 * \param [in] outRate Output sampling rate (1D tensor in HOST memory, of size batchSize)
 * \param [in] srcDimsTensor source audio buffer length and number of channels (1D tensor in HOST memory, of size batchSize * 2)
@@ -268,6 +268,25 @@ RppStatus rppt_mel_filter_bank_gpu(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppP
 */
 RppStatus rppt_resample_host(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppPtr_t dstPtr, RpptDescPtr dstDescPtr, Rpp32f *inRateTensor, Rpp32f *outRateTensor, Rpp32s *srcDimsTensor, RpptResamplingWindow &window, rppHandle_t rppHandle);
 
+#ifdef GPU_SUPPORT
+/*! \brief Resample augmentation on HIP backend
+* \details Resample 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 or 3 (for single-channel or multi-channel audio tensor), offsetInBytes >= 0, dataType = F32)
+* \param [in] inRate Input sampling rate (1D tensor in Pinned memory, of size batchSize)
+* \param [in] outRate Output sampling rate (1D tensor in Pinned memory, of size batchSize)
+* \param [in] srcDimsTensor source audio buffer length and number of channels (1D tensor in Pinned memory, of size batchSize * 2)
+* \param [in] window Resampling window (struct of type RpptRpptResamplingWindow in HIP/Pinned memory)
+* \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_resample_gpu(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppPtr_t dstPtr, RpptDescPtr dstDescPtr, Rpp32f *inRateTensor, Rpp32f *outRateTensor, Rpp32s *srcDimsTensor, RpptResamplingWindow &window, rppHandle_t rppHandle);
+#endif // GPU_SUPPORT
+
 /*! @}
  */
 

include/rppt_tensor_effects_augmentations.h

@@ -470,7 +470,7 @@ RppStatus rppt_vignette_gpu(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppPtr_t ds
  * - srcPtr depth ranges - Rpp8u (0 to 255), Rpp16f (0 to 1), Rpp32f (0 to 1), Rpp8s (-128 to 127).
  * - dstPtr depth ranges - Will be same depth as srcPtr.
  * \image html img150x150.png Sample Input
- * \image html effects_augmentations_jitter_img150x150.png Sample Output
+ * \image html effects_augmentations_jitter_150x150.png Sample Output
  * \param [in] srcPtr source tensor in HOST memory
  * \param [in] srcDescPtr source tensor descriptor (Restrictions - numDims = 4, offsetInBytes >= 0, dataType = U8/F16/F32/I8, layout = NCHW/NHWC, c = 1/3)
  * \param [out] dstPtr destination tensor in HOST memory
@@ -491,7 +491,7 @@ RppStatus rppt_jitter_host(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppPtr_t dst
  * - srcPtr depth ranges - Rpp8u (0 to 255), Rpp16f (0 to 1), Rpp32f (0 to 1), Rpp8s (-128 to 127).
  * - dstPtr depth ranges - Will be same depth as srcPtr.
  * \image html img150x150.png Sample Input
- * \image html effects_augmentations_jitter_img150x150.png Sample Output
+ * \image html effects_augmentations_jitter_150x150.png Sample Output
  * \param [in] srcPtr source tensor in HIP memory
  * \param un[in] srcDescPtr source tensor descriptor (Restrictions - numDims = 4, offsetInBytes >= 0, dataType = U8/F16/F32/I8, layout = NCHW/NHWC, c = 1/3)
  * \param [out] dstPtr destination tensor in HIP memory
@@ -576,8 +576,8 @@ RppStatus rppt_erase_host(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppPtr_t dstP
  * \details This function erases one or more user defined regions from an image, for a batch of RGB(3 channel) / greyscale(1 channel) images with an NHWC/NCHW tensor layout.<br>
  *          srcPtr depth ranges - Rpp8u (0 to 255), Rpp16f (0 to 1), Rpp32f (0 to 1), Rpp8s (-128 to 127).
  *          dstPtr depth ranges - Will be same depth as srcPtr.
- * \image html img150x150.jpg Sample Input
- * \image html effects_augmentations_erase_img150x150.jpg Sample Output
+ * \image html img150x150.png Sample Input
+ * \image html effects_augmentations_erase_img150x150.png Sample Output
  * \param [in] srcPtr source tensor in HIP memory
  * \param [in] srcDescPtr source tensor descriptor (Restrictions - numDims = 4, offsetInBytes >= 0, dataType = U8/F16/F32/I8, layout = NCHW/NHWC, c = 1/3)
  * \param [out] dstPtr destination tensor in HIP memory
@@ -601,8 +601,8 @@ RppStatus rppt_erase_gpu(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppPtr_t dstPt
  * \details The glitch augmentation adds a glitch effect for a batch of RGB(3 channel) / greyscale(1 channel) images with an NHWC/NCHW tensor layout.<br>
  * - srcPtr depth ranges - Rpp8u (0 to 255), Rpp16f (0 to 1), Rpp32f (0 to 1), Rpp8s (-128 to 127).
  * - dstPtr depth ranges - Will be same depth as srcPtr.
- * \image html img150x150.jpg Sample Input
- * \image html effects_augmentations_glitch_img150x150.jpg Sample Output
+ * \image html img150x150.png Sample Input
+ * \image html effects_augmentations_glitch_img150x150.png Sample Output
  * \param [in] srcPtr source tensor in HOST memory
  * \param [in] srcDescPtr source tensor descriptor (Restrictions - numDims = 4, offsetInBytes >= 0, dataType = U8/F16/F32/I8, layout = NCHW/NHWC, c = 1/3)
  * \param [out] dstPtr destination tensor in HOST memory
@@ -623,8 +623,8 @@ RppStatus rppt_glitch_host(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppPtr_t dst
  * \details The glitch augmentation adds a glitch effect for a batch of RGB(3 channel) / greyscale(1 channel) images with an NHWC/NCHW tensor layout.<br>
  * - srcPtr depth ranges - Rpp8u (0 to 255), Rpp16f (0 to 1), Rpp32f (0 to 1), Rpp8s (-128 to 127).
  * - dstPtr depth ranges - Will be same depth as srcPtr.
- * \image html img150x150.jpg Sample Input
- * \image html effects_augmentations_glitch_img150x150.jpg Sample Output
+ * \image html img150x150.png Sample Input
+ * \image html effects_augmentations_glitch_img150x150.png Sample Output
  * \param [in] srcPtr source tensor in HIP memory
  * \param [in] srcDescPtr source tensor descriptor (Restrictions - numDims = 4, offsetInBytes >= 0, dataType = U8/F16/F32/I8, layout = NCHW/NHWC, c = 1/3)
  * \param [out] dstPtr destination tensor in HIP memory
@@ -645,8 +645,8 @@ RppStatus rppt_glitch_gpu(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppPtr_t dstP
  * \details The pixelate augmentation performs a pixelate transformation for a batch of RGB(3 channel) / greyscale(1 channel) images with an NHWC/NCHW tensor layout.<br>
  * - srcPtr depth ranges - Rpp8u (0 to 255), Rpp16f (0 to 1), Rpp32f (0 to 1), Rpp8s (-128 to 127).
  * - dstPtr depth ranges - Will be same depth as srcPtr.
- * \image html img150x150.jpg Sample Input
- * \image html effects_augmentations_pixelate_img150x150.jpg Sample Output
+ * \image html img150x150.png Sample Input
+ * \image html effects_augmentations_pixelate_img150x150.png Sample Output
  * \param [in] srcPtr source tensor in HOST memory
  * \param [in] srcDescPtr source tensor descriptor (Restrictions - numDims = 4, offsetInBytes >= 0, dataType = U8/F16/F32/I8, layout = NCHW/NHWC, c = 1/3)
  * \param [out] dstPtr destination tensor in HOST memory
@@ -667,8 +667,8 @@ RppStatus rppt_pixelate_host(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppPtr_t d
  * \details The pixelate augmentation performs a pixelate transformation for a batch of RGB(3 channel) / greyscale(1 channel) images with an NHWC/NCHW tensor layout.<br>
  * - srcPtr depth ranges - Rpp8u (0 to 255), Rpp16f (0 to 1), Rpp32f (0 to 1), Rpp8s (-128 to 127).
  * - dstPtr depth ranges - Will be same depth as srcPtr.
- * \image html img150x150.jpg Sample Input
- * \image html effects_augmentations_pixelate_img150x150.jpg Sample Output
+ * \image html img150x150.png Sample Input
+ * \image html effects_augmentations_pixelate_img150x150.png Sample Output
  * \param [in] srcPtr source tensor in HIP memory
  * \param [in] srcDescPtr source tensor descriptor (Restrictions - numDims = 4, offsetInBytes >= 0, dataType = U8/F16/F32/I8, layout = NCHW/NHWC, c = 1/3)
  * \param [out] dstPtr destination tensor in HIP memory

include/rppt_tensor_geometric_augmentations.h

@@ -613,8 +613,8 @@ RppStatus rppt_remap_host(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppPtr_t dstP
  * \details Performs a remap operation using user specified remap tables for a batch of RGB(3 channel) / greyscale(1 channel) images with an NHWC/NCHW tensor layout. For each image, the output(x,y) = input(mapx(x, y), mapy(x, y)) for every (x,y) in the destination image.<br>
  * - srcPtr depth ranges - Rpp8u (0 to 255), Rpp16f (0 to 1), Rpp32f (0 to 1), Rpp8s (-128 to 127).
  * - dstPtr depth ranges - Will be same depth as srcPtr.
- * \image html img150x150.jpg Sample Input
- * \image html geometric_augmentations_remap_img150x150.jpg Sample Output
+ * \image html img150x150.png Sample Input
+ * \image html geometric_augmentations_remap_img150x150.png Sample Output
  * \param [in] srcPtr source tensor in HIP memory
  * \param [in] srcDescPtr source tensor descriptor (Restrictions - numDims = 4, offsetInBytes >= 0, dataType = U8/F16/F32/I8, layout = NCHW/NHWC, c = 1/3)
  * \param [out] dstPtr destination tensor in HIP memory

src/include/cpu/rpp_cpu_common.hpp

@@ -5452,12 +5452,6 @@ inline void compute_bicubic_coefficient(Rpp32f weight, Rpp32f &coeff)
     coeff = (x >= 2) ? 0 : ((x > 1) ? (x * x * (-0.5f * x + 2.5f) - 4.0f * x + 2.0f) : (x * x * (1.5f * x - 2.5f) + 1.0f));
 }
 
-inline Rpp32f sinc(Rpp32f x)
-{
-    x *= M_PI;
-    return (std::abs(x) < 1e-5f) ? (1.0f - x * x * ONE_OVER_6) : std::sin(x) / x;
-}
-
 inline void compute_lanczos3_coefficient(Rpp32f weight, Rpp32f &coeff)
 {
     coeff = fabs(weight) >= 3 ? 0.0f : (sinc(weight) * sinc(weight * 0.333333f));

src/include/hip/rpp_hip_common.hpp

@@ -156,6 +156,11 @@ struct RPPTensorFunctionMetaData
   (byte & 0x02 ? '1' : '0'), \
   (byte & 0x01 ? '1' : '0')
 
+// float4 floor
+
+#define FLOOR4(src, dst) \
+dst = make_int4(floorf(src.x), floorf(src.y), floorf(src.z), floorf(src.w));
+
 /******************** HOST FUNCTIONS ********************/
 
 inline int getplnpkdind(RppiChnFormat &format)

src/modules/hip/handlehip.cpp

@@ -245,7 +245,7 @@ struct HandleImpl
            - 293 is the size required for storing reduction outputs for 600000 size sample
            - 128 is the size required for storing cutOffDB values for batch size 128 */
         hipMalloc(&(this->initHandle->mem.mgpu.scratchBufferHip.floatmem), sizeof(Rpp32f) * 76853888);
-        hipHostMalloc(&(this->initHandle->mem.mgpu.scratchBufferPinned.floatmem), sizeof(Rpp32f) * 8294400);
+        hipHostMalloc(&(this->initHandle->mem.mgpu.scratchBufferPinned.floatmem), sizeof(Rpp32f) * 8294400);    // 3840 x 2160
     }
 };
 

src/modules/hip/hip_tensor_audio_augmentations.hpp

@@ -30,5 +30,6 @@ SOFTWARE.
 #include "kernel/mel_filter_bank.hpp"
 #include "kernel/pre_emphasis_filter.hpp"
 #include "kernel/to_decibels.hpp"
+#include "kernel/resample.hpp"
 
 #endif // HIP_TENSOR_AUDIO_AUGMENTATIONS_HPP