37 #include <hip/hip_fp16.h>
39 #if __has_include(<half/half.hpp>)
40 #include <half/half.hpp>
44 using halfhpp = half_float::half;
45 typedef halfhpp Rpp16f;
54 #include <x86intrin.h>
55 #include <smmintrin.h>
56 #include <immintrin.h>
61 #define RPP_MIN_8U ( 0 )
63 #define RPP_MAX_8U ( 255 )
65 #define RPPT_MAX_DIMS ( 5 )
67 #define RPPT_MAX_AUDIO_CHANNELS ( 16 )
69 #define CHECK_RETURN_STATUS(x) do { \
72 fprintf(stderr, "Runtime error: %s returned %d at %s:%d", #x, retval, __FILE__, __LINE__); \
78 #include <hip/hip_runtime.h>
79 #define RPP_HOST_DEVICE __host__ __device__
81 #define RPP_HOST_DEVICE
84 const float ONE_OVER_6 = 1.0f / 6;
85 const float ONE_OVER_3 = 1.0f / 3;
86 const float ONE_OVER_255 = 1.0f / 255;
87 const uint MMS_MAX_SCRATCH_MEMORY = 115293120;
88 const uint SPECTROGRAM_MAX_SCRATCH_MEMORY = 372877312;
193 rppStatusSuccess = 0,
194 rppStatusBadParm = -1,
195 rppStatusUnknownError = -2,
196 rppStatusNotInitialized = -3,
197 rppStatusInvalidValue = -4,
198 rppStatusAllocFailed = -5,
199 rppStatusInternalError = -6,
200 rppStatusNotImplemented = -7,
201 rppStatusUnsupportedOp = -8,
204 #ifdef LEGACY_SUPPORT
210 RPP_SCALAR_OP_AND = 1,
215 RPP_SCALAR_OP_NOTEQUAL,
217 RPP_SCALAR_OP_LESSEQ,
218 RPP_SCALAR_OP_GREATER,
219 RPP_SCALAR_OP_GREATEREQ,
221 RPP_SCALAR_OP_SUBTRACT,
222 RPP_SCALAR_OP_MULTIPLY,
223 RPP_SCALAR_OP_DIVIDE,
224 RPP_SCALAR_OP_MODULUS,
236 } RppConvertBitDepthMode;
291 #ifdef LEGACY_SUPPORT
299 } RppiColorConvertMode;
316 RPPI_HORIZONTAL_AXIS,
389 unsigned int roiWidth;
390 unsigned int roiHeight;
453 NEAREST_NEIGHBOR = 0,
543 int roiWidth, roiHeight;
553 int roiWidth, roiHeight, roiDepth;
676 float boxMullerExtra;
711 switch(interpolationType)
713 case RpptInterpolationType::BICUBIC:
718 case RpptInterpolationType::LANCZOS:
720 if(in_size > out_size)
722 this->radius = 3.0f * scaleRatio;
723 this->scale = (1 / scaleRatio);
729 case RpptInterpolationType::GAUSSIAN:
731 if(in_size > out_size)
733 this->radius = scaleRatio;
734 this->scale = (1 / scaleRatio);
738 case RpptInterpolationType::TRIANGULAR:
740 if(in_size > out_size)
742 this->radius = scaleRatio;
743 this->scale = (1 / scaleRatio);
754 this->size = std::ceil(2 * this->radius);
772 Rpp32f locRaw = x * scale + center;
773 Rpp32s locFloor = std::floor(locRaw);
774 Rpp32f weight = locRaw - locFloor;
775 locFloor = std::max(std::min(locFloor, lookupSize - 2), 0);
776 Rpp32f current = lookup[locFloor];
777 Rpp32f next = lookup[locFloor + 1];
778 return current + weight * (next - current);
781 inline __m128 operator()(__m128 x)
783 __m128 pLocRaw = _mm_add_ps(_mm_mul_ps(x, pScale), pCenter);
784 __m128i pxLocFloor = _mm_cvttps_epi32(pLocRaw);
785 __m128 pLocFloor = _mm_cvtepi32_ps(pxLocFloor);
786 __m128 pWeight = _mm_sub_ps(pLocRaw, pLocFloor);
788 _mm_storeu_si128(
reinterpret_cast<__m128i*
>(idx), pxLocFloor);
789 __m128 pCurrent = _mm_setr_ps(lookup[idx[0]], lookup[idx[1]], lookup[idx[2]], lookup[idx[3]]);
790 __m128 pNext = _mm_setr_ps(lookup[idx[0] + 1], lookup[idx[1] + 1], lookup[idx[2] + 1], lookup[idx[3] + 1]);
791 return _mm_add_ps(pCurrent, _mm_mul_ps(pWeight, _mm_sub_ps(pNext, pCurrent)));
794 Rpp32f scale = 1, center = 1;
795 Rpp32s lobes = 0, coeffs = 0;
797 Rpp32f *lookupPinned =
nullptr;
798 std::vector<Rpp32f> lookup;
799 __m128 pCenter, pScale;
808 inline RPP_HOST_DEVICE
virtual Rpp32f hz_to_mel(
Rpp32f hz) = 0;
809 inline RPP_HOST_DEVICE
virtual Rpp32f mel_to_hz(
Rpp32f mel) = 0;
818 inline RPP_HOST_DEVICE
Rpp32f hz_to_mel(
Rpp32f hz) {
return 1127.0f * std::log(1.0f + (hz / 700.0f)); }
819 inline RPP_HOST_DEVICE
Rpp32f mel_to_hz(
Rpp32f mel) {
return 700.0f * (std::exp(mel / 1127.0f) - 1.0f); }
830 const Rpp32f fsp = 66.666667f;
831 const Rpp32f minLogHz = 1000.0;
832 const Rpp32f minLogMel = (minLogHz - freqLow) / fsp;
833 const Rpp32f stepLog = 0.068751777;
835 const Rpp32f invMinLogHz = 0.001f;
836 const Rpp32f invStepLog = 1.0f / stepLog;
837 const Rpp32f invFsp = 1.0f / fsp;
843 mel = minLogMel + std::log(hz * invMinLogHz) * invStepLog;
845 mel = (hz - freqLow) * invFsp;
853 if (mel >= minLogMel)
854 hz = minLogHz * std::exp(stepLog * (mel - minLogMel));
856 hz = freqLow + mel * fsp;
961 Rpp32f *scratchBufferHost;
1048 clmemSize maxSrcSize;
1049 clmemSize maxDstSize;
1051 clmemRpp32f floatArr[10];
1052 clmemRpp64f doubleArr[10];
1053 clmemRpp32u uintArr[10];
1054 clmemRpp32s intArr[10];
1055 clmemRpp8u ucharArr[10];
1056 clmemRpp8s charArr[10];
1057 cl_mem srcBatchIndex;
1058 cl_mem dstBatchIndex;
1072 #elif defined(HIP_COMPILE)
unsigned int Rpp32u
32 bit unsigned int
Definition: rppdefs.h:101
int Rpp32s
32 bit signed int
Definition: rppdefs.h:103
signed char Rpp8s
8 bit signed char
Definition: rppdefs.h:95
RppiChnFormat
RPPI Image channel format type enum.
Definition: rppdefs.h:368
RpptMelScaleFormula
RPPT Mel Scale Formula.
Definition: rppdefs.h:485
unsigned short Rpp16u
16 bit unsigned short
Definition: rppdefs.h:97
unsigned char Rpp8u
8 bit unsigned char
Definition: rppdefs.h:93
void * RppPtr_t
void pointer
Definition: rppdefs.h:113
RpptLayout
RPPT Tensor layout type enum.
Definition: rppdefs.h:411
RpptRoiType
RPPT Tensor 2D ROI type enum.
Definition: rppdefs.h:425
RpptDataType
RPPT Tensor datatype enum.
Definition: rppdefs.h:399
RpptImageBorderType
RPPT Image Border Type.
Definition: rppdefs.h:475
struct RpptResamplingWindow RpptResamplingWindow
RPPT Tensor RpptResamplingWindow type struct.
double Rpp64f
64 bit double
Definition: rppdefs.h:111
RpptAudioBorderType
RPPT Audio Border Type.
Definition: rppdefs.h:465
size_t RppSize_t
size_t
Definition: rppdefs.h:115
struct GenericFilter GenericFilter
RPPT Tensor GenericFilter type struct.
RppStatus
RPP RppStatus type enums.
Definition: rppdefs.h:121
RpptRoi3DType
RPPT Tensor 3D ROI type enum.
Definition: rppdefs.h:434
unsigned long long Rpp64u
64 bit unsigned long long
Definition: rppdefs.h:105
#define RPPT_MAX_DIMS
RPP maximum dimensions in tensor .
Definition: rppdefs.h:65
RpptInterpolationType
RPPT Tensor interpolation type enum.
Definition: rppdefs.h:452
RpptSubpixelLayout
RPPT Tensor subpixel layout type enum.
Definition: rppdefs.h:443
short Rpp16s
16 bit signed short
Definition: rppdefs.h:99
float Rpp32f
32 bit float
Definition: rppdefs.h:109
RppBackend
RPP RppBackend type enums.
Definition: rppdefs.h:182
long long Rpp64s
64 bit long long
Definition: rppdefs.h:107
rppStatus_t
RPP rppStatus_t type enums.
Definition: rppdefs.h:192
@ RPP_ERROR_NOT_IMPLEMENTED
Function variant requested is not implemented / unsupported.
Definition: rppdefs.h:135
@ RPP_ERROR_INVALID_DST_DATATYPE
Invalid dst tensor datatype. (Needs to adhere to function specification.)
Definition: rppdefs.h:147
@ RPP_ERROR_INVALID_OUTPUT_TILE_LENGTH
Invalid output tile length (Needs to adhere to function specification.)
Definition: rppdefs.h:163
@ RPP_ERROR_INVALID_DIM_LENGTHS
Length in some dimensions are invalid. (Needs to adhere to function specification....
Definition: rppdefs.h:173
@ RPP_ERROR_OUT_OF_BOUND_SRC_ROI
Out of bound source ROI.
Definition: rppdefs.h:157
@ RPP_ERROR_OUT_OF_BOUND_SHARED_MEMORY_SIZE
Shared memory size needed is beyond the bounds (Needs to adhere to function specification....
Definition: rppdefs.h:165
@ RPP_ERROR_ZERO_DIVISION
Arguments provided will result in zero division error.
Definition: rppdefs.h:131
@ RPP_ERROR_INVALID_ARGUMENTS
One or more arguments invalid. (Needs to adhere to function specification.)
Definition: rppdefs.h:127
@ RPP_ERROR_LAYOUT_MISMATCH
src and dst layout mismatch
Definition: rppdefs.h:159
@ RPP_ERROR_INVALID_CHANNELS
Number of channels is invalid. (Needs to adhere to function specification.)
Definition: rppdefs.h:161
@ RPP_ERROR_INVALID_SRC_CHANNELS
Invalid src tensor number of channels. (Needs to adhere to function specification....
Definition: rppdefs.h:137
@ RPP_ERROR_INVALID_SRC_OR_DST_DATATYPE
Invalid src/dst tensor datatype. (Needs to adhere to function specification.)
Definition: rppdefs.h:149
@ RPP_ERROR_INVALID_PARAMETER_DATATYPE
Invalid datatype.
Definition: rppdefs.h:153
@ RPP_ERROR_NOT_ENOUGH_MEMORY
Not enough memory to write outputs, as per dim-lengths and strides set in descriptor.
Definition: rppdefs.h:155
@ RPP_ERROR
Unspecified error.
Definition: rppdefs.h:125
@ RPP_ERROR_INVALID_SRC_DIMS
Number of src dims is invalid. (Needs to adhere to function specification.)
Definition: rppdefs.h:169
@ RPP_ERROR_OUT_OF_BOUND_SCRATCH_MEMORY_SIZE
Scratch memory size needed is beyond the bounds (Needs to adhere to function specification....
Definition: rppdefs.h:167
@ RPP_ERROR_INVALID_DST_LAYOUT
Invalid dst tensor layout. (Needs to adhere to function specification.)
Definition: rppdefs.h:143
@ RPP_ERROR_INVALID_AXIS
The specified axis is invalid or out of range. (Needs to adhere to function specification....
Definition: rppdefs.h:175
@ RPP_ERROR_INVALID_DST_CHANNELS
Invalid dst tensor number of channels. (Needs to adhere to function specification....
Definition: rppdefs.h:139
@ RPP_ERROR_INVALID_SRC_DATATYPE
Invalid src tensor datatype. (Needs to adhere to function specification.)
Definition: rppdefs.h:145
@ RPP_ERROR_INVALID_SRC_LAYOUT
Invalid src tensor layout. (Needs to adhere to function specification.)
Definition: rppdefs.h:141
@ RPP_ERROR_HIGH_SRC_DIMENSION
Src tensor / src ROI dimension too high. (Needs to adhere to function specification....
Definition: rppdefs.h:133
@ RPP_SUCCESS
No error.
Definition: rppdefs.h:123
@ RPP_ERROR_INSUFFICIENT_DST_BUFFER_LENGTH
Insufficient dst buffer length provided. (Needs to adhere to function specification....
Definition: rppdefs.h:151
@ RPP_ERROR_INVALID_DST_DIMS
Number of dst dims is invalid. (Needs to adhere to function specification.)
Definition: rppdefs.h:171
@ RPP_ERROR_LOW_OFFSET
Low tensor offsetInBytes provided for src/dst tensor.
Definition: rppdefs.h:129
Base class for Mel scale conversions.
Definition: rppdefs.h:806
RPPT Tensor GenericFilter type struct.
Definition: rppdefs.h:705
Derived class for HTK Mel scale conversions.
Definition: rppdefs.h:817
RPP initialize handle.
Definition: rppdefs.h:1208
RPP 24 float vector.
Definition: rppdefs.h:285
RPP 6 float vector.
Definition: rppdefs.h:261
RPP 9 float vector.
Definition: rppdefs.h:269
RPP 24 signed int vector.
Definition: rppdefs.h:277
RPP layout params.
Definition: rppdefs.h:252
RPPI Image 2D ROI (XYWH format) type struct.
Definition: rppdefs.h:386
RPPI Image size(Width/Height dimensions) type struct.
Definition: rppdefs.h:377
RPPT Tensor 2D bilinear neighborhood 32-bit signed int 8-length-vectors type struct.
Definition: rppdefs.h:683
RPPT Tensor 2D bilinear neighborhood 32-bit float 8-length-vectors type struct.
Definition: rppdefs.h:694
RPPT Tensor Channel Offsets struct.
Definition: rppdefs.h:522
RPPT Tensor descriptor type struct.
Definition: rppdefs.h:592
RPPT Tensor 32-bit float RGB type struct.
Definition: rppdefs.h:628
RPPT Tensor 2D 32-bit float vector type struct.
Definition: rppdefs.h:647
RPPT Tensor Generic descriptor type struct.
Definition: rppdefs.h:605
RPPT Tensor 2D image patch dimensions type struct.
Definition: rppdefs.h:656
RPPT Image 2D cartesian point type struct.
Definition: rppdefs.h:494
RPPT Image 3D point type struct.
Definition: rppdefs.h:503
RPPT Tensor 8-bit uchar RGB type struct.
Definition: rppdefs.h:618
RPPT Tensor RpptResamplingWindow type struct.
Definition: rppdefs.h:762
RPPT Tensor 3D ROI LTFRBB struct.
Definition: rppdefs.h:532
RPPT Tensor 2D ROI LTRB struct.
Definition: rppdefs.h:513
RPPT Tensor 2D ROI XYWH struct.
Definition: rppdefs.h:541
RPPT Tensor 3D ROI XYZWHD struct.
Definition: rppdefs.h:551
RPPT Tensor strides type struct.
Definition: rppdefs.h:581
RPPT Tensor 2D 32-bit uint vector type struct.
Definition: rppdefs.h:638
RPPT Tensor random number generator state (xorwow box muller state) type struct.
Definition: rppdefs.h:673
RPPT Tensor random number generator state (xorwow state) type struct.
Definition: rppdefs.h:665
Derived class for Slaney Mel scale conversions.
Definition: rppdefs.h:828
RPP HIP 2D ROI memory.
Definition: rppdefs.h:1145
RPP HIP 32-bit float memory.
Definition: rppdefs.h:1080
RPP HIP 32-bit signed int memory.
Definition: rppdefs.h:1104
RPP HIP 32-bit unsigned int memory.
Definition: rppdefs.h:1096
RPP HIP 64-bit double memory.
Definition: rppdefs.h:1088
RPP HIP 8-bit signed char memory.
Definition: rppdefs.h:1120
RPP HIP 8-bit unsigned char memory.
Definition: rppdefs.h:1112
RPP HIP RGB memory.
Definition: rppdefs.h:1128
RPP HIP 2D dimensions memory.
Definition: rppdefs.h:1136
RPP HOST memory type struct.
Definition: rppdefs.h:944
RPP OCL memory management type struct.
Definition: rppdefs.h:1156
RPP HIP-HOST memory management.
Definition: rppdefs.h:1187
RPP HOST 2D ROI memory.
Definition: rppdefs.h:934
RPP HOST 32-bit float memory.
Definition: rppdefs.h:869
RPP HOST 32-bit signed int memory.
Definition: rppdefs.h:893
RPP HOST 32-bit unsigned int memory.
Definition: rppdefs.h:885
RPP HOST 64-bit double memory.
Definition: rppdefs.h:877
RPP HOST 8-bit signed char memory.
Definition: rppdefs.h:909
RPP HOST 8-bit unsigned char memory.
Definition: rppdefs.h:901
RPP HOST RGB memory.
Definition: rppdefs.h:917
RPP HOST 2D dimensions memory.
Definition: rppdefs.h:925
RPPT Tensor 3D ROI union.
Definition: rppdefs.h:571
RPPT Tensor 2D ROI union.
Definition: rppdefs.h:561